mm: goodbye remove_from_page_cache()
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.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/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  (code doesn't rely on that order, so you could switch it around)
106  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
107  *    ->i_mmap_lock
108  */
109
110 /*
111  * Remove 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 __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         if (PageSwapBacked(page))
124                 __dec_zone_page_state(page, NR_SHMEM);
125         BUG_ON(page_mapped(page));
126
127         /*
128          * Some filesystems seem to re-dirty the page even after
129          * the VM has canceled the dirty bit (eg ext3 journaling).
130          *
131          * Fix it up by doing a final dirty accounting check after
132          * having removed the page entirely.
133          */
134         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
135                 dec_zone_page_state(page, NR_FILE_DIRTY);
136                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
137         }
138 }
139
140 /**
141  * delete_from_page_cache - delete page from page cache
142  * @page: the page which the kernel is trying to remove from page cache
143  *
144  * This must be called only on pages that have been verified to be in the page
145  * cache and locked.  It will never put the page into the free list, the caller
146  * has a reference on the page.
147  */
148 void delete_from_page_cache(struct page *page)
149 {
150         struct address_space *mapping = page->mapping;
151         void (*freepage)(struct page *);
152
153         BUG_ON(!PageLocked(page));
154
155         freepage = mapping->a_ops->freepage;
156         spin_lock_irq(&mapping->tree_lock);
157         __remove_from_page_cache(page);
158         spin_unlock_irq(&mapping->tree_lock);
159         mem_cgroup_uncharge_cache_page(page);
160
161         if (freepage)
162                 freepage(page);
163         page_cache_release(page);
164 }
165 EXPORT_SYMBOL(delete_from_page_cache);
166
167 static int sync_page(void *word)
168 {
169         struct address_space *mapping;
170         struct page *page;
171
172         page = container_of((unsigned long *)word, struct page, flags);
173
174         /*
175          * page_mapping() is being called without PG_locked held.
176          * Some knowledge of the state and use of the page is used to
177          * reduce the requirements down to a memory barrier.
178          * The danger here is of a stale page_mapping() return value
179          * indicating a struct address_space different from the one it's
180          * associated with when it is associated with one.
181          * After smp_mb(), it's either the correct page_mapping() for
182          * the page, or an old page_mapping() and the page's own
183          * page_mapping() has gone NULL.
184          * The ->sync_page() address_space operation must tolerate
185          * page_mapping() going NULL. By an amazing coincidence,
186          * this comes about because none of the users of the page
187          * in the ->sync_page() methods make essential use of the
188          * page_mapping(), merely passing the page down to the backing
189          * device's unplug functions when it's non-NULL, which in turn
190          * ignore it for all cases but swap, where only page_private(page) is
191          * of interest. When page_mapping() does go NULL, the entire
192          * call stack gracefully ignores the page and returns.
193          * -- wli
194          */
195         smp_mb();
196         mapping = page_mapping(page);
197         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
198                 mapping->a_ops->sync_page(page);
199         io_schedule();
200         return 0;
201 }
202
203 static int sync_page_killable(void *word)
204 {
205         sync_page(word);
206         return fatal_signal_pending(current) ? -EINTR : 0;
207 }
208
209 /**
210  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
211  * @mapping:    address space structure to write
212  * @start:      offset in bytes where the range starts
213  * @end:        offset in bytes where the range ends (inclusive)
214  * @sync_mode:  enable synchronous operation
215  *
216  * Start writeback against all of a mapping's dirty pages that lie
217  * within the byte offsets <start, end> inclusive.
218  *
219  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
220  * opposed to a regular memory cleansing writeback.  The difference between
221  * these two operations is that if a dirty page/buffer is encountered, it must
222  * be waited upon, and not just skipped over.
223  */
224 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
225                                 loff_t end, int sync_mode)
226 {
227         int ret;
228         struct writeback_control wbc = {
229                 .sync_mode = sync_mode,
230                 .nr_to_write = LONG_MAX,
231                 .range_start = start,
232                 .range_end = end,
233         };
234
235         if (!mapping_cap_writeback_dirty(mapping))
236                 return 0;
237
238         ret = do_writepages(mapping, &wbc);
239         return ret;
240 }
241
242 static inline int __filemap_fdatawrite(struct address_space *mapping,
243         int sync_mode)
244 {
245         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
246 }
247
248 int filemap_fdatawrite(struct address_space *mapping)
249 {
250         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
251 }
252 EXPORT_SYMBOL(filemap_fdatawrite);
253
254 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
255                                 loff_t end)
256 {
257         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
258 }
259 EXPORT_SYMBOL(filemap_fdatawrite_range);
260
261 /**
262  * filemap_flush - mostly a non-blocking flush
263  * @mapping:    target address_space
264  *
265  * This is a mostly non-blocking flush.  Not suitable for data-integrity
266  * purposes - I/O may not be started against all dirty pages.
267  */
268 int filemap_flush(struct address_space *mapping)
269 {
270         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
271 }
272 EXPORT_SYMBOL(filemap_flush);
273
274 /**
275  * filemap_fdatawait_range - wait for writeback to complete
276  * @mapping:            address space structure to wait for
277  * @start_byte:         offset in bytes where the range starts
278  * @end_byte:           offset in bytes where the range ends (inclusive)
279  *
280  * Walk the list of under-writeback pages of the given address space
281  * in the given range and wait for all of them.
282  */
283 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
284                             loff_t end_byte)
285 {
286         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
287         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
288         struct pagevec pvec;
289         int nr_pages;
290         int ret = 0;
291
292         if (end_byte < start_byte)
293                 return 0;
294
295         pagevec_init(&pvec, 0);
296         while ((index <= end) &&
297                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
298                         PAGECACHE_TAG_WRITEBACK,
299                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
300                 unsigned i;
301
302                 for (i = 0; i < nr_pages; i++) {
303                         struct page *page = pvec.pages[i];
304
305                         /* until radix tree lookup accepts end_index */
306                         if (page->index > end)
307                                 continue;
308
309                         wait_on_page_writeback(page);
310                         if (TestClearPageError(page))
311                                 ret = -EIO;
312                 }
313                 pagevec_release(&pvec);
314                 cond_resched();
315         }
316
317         /* Check for outstanding write errors */
318         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
319                 ret = -ENOSPC;
320         if (test_and_clear_bit(AS_EIO, &mapping->flags))
321                 ret = -EIO;
322
323         return ret;
324 }
325 EXPORT_SYMBOL(filemap_fdatawait_range);
326
327 /**
328  * filemap_fdatawait - wait for all under-writeback pages to complete
329  * @mapping: address space structure to wait for
330  *
331  * Walk the list of under-writeback pages of the given address space
332  * and wait for all of them.
333  */
334 int filemap_fdatawait(struct address_space *mapping)
335 {
336         loff_t i_size = i_size_read(mapping->host);
337
338         if (i_size == 0)
339                 return 0;
340
341         return filemap_fdatawait_range(mapping, 0, i_size - 1);
342 }
343 EXPORT_SYMBOL(filemap_fdatawait);
344
345 int filemap_write_and_wait(struct address_space *mapping)
346 {
347         int err = 0;
348
349         if (mapping->nrpages) {
350                 err = filemap_fdatawrite(mapping);
351                 /*
352                  * Even if the above returned error, the pages may be
353                  * written partially (e.g. -ENOSPC), so we wait for it.
354                  * But the -EIO is special case, it may indicate the worst
355                  * thing (e.g. bug) happened, so we avoid waiting for it.
356                  */
357                 if (err != -EIO) {
358                         int err2 = filemap_fdatawait(mapping);
359                         if (!err)
360                                 err = err2;
361                 }
362         }
363         return err;
364 }
365 EXPORT_SYMBOL(filemap_write_and_wait);
366
367 /**
368  * filemap_write_and_wait_range - write out & wait on a file range
369  * @mapping:    the address_space for the pages
370  * @lstart:     offset in bytes where the range starts
371  * @lend:       offset in bytes where the range ends (inclusive)
372  *
373  * Write out and wait upon file offsets lstart->lend, inclusive.
374  *
375  * Note that `lend' is inclusive (describes the last byte to be written) so
376  * that this function can be used to write to the very end-of-file (end = -1).
377  */
378 int filemap_write_and_wait_range(struct address_space *mapping,
379                                  loff_t lstart, loff_t lend)
380 {
381         int err = 0;
382
383         if (mapping->nrpages) {
384                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
385                                                  WB_SYNC_ALL);
386                 /* See comment of filemap_write_and_wait() */
387                 if (err != -EIO) {
388                         int err2 = filemap_fdatawait_range(mapping,
389                                                 lstart, lend);
390                         if (!err)
391                                 err = err2;
392                 }
393         }
394         return err;
395 }
396 EXPORT_SYMBOL(filemap_write_and_wait_range);
397
398 /**
399  * replace_page_cache_page - replace a pagecache page with a new one
400  * @old:        page to be replaced
401  * @new:        page to replace with
402  * @gfp_mask:   allocation mode
403  *
404  * This function replaces a page in the pagecache with a new one.  On
405  * success it acquires the pagecache reference for the new page and
406  * drops it for the old page.  Both the old and new pages must be
407  * locked.  This function does not add the new page to the LRU, the
408  * caller must do that.
409  *
410  * The remove + add is atomic.  The only way this function can fail is
411  * memory allocation failure.
412  */
413 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
414 {
415         int error;
416         struct mem_cgroup *memcg = NULL;
417
418         VM_BUG_ON(!PageLocked(old));
419         VM_BUG_ON(!PageLocked(new));
420         VM_BUG_ON(new->mapping);
421
422         /*
423          * This is not page migration, but prepare_migration and
424          * end_migration does enough work for charge replacement.
425          *
426          * In the longer term we probably want a specialized function
427          * for moving the charge from old to new in a more efficient
428          * manner.
429          */
430         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
431         if (error)
432                 return error;
433
434         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
435         if (!error) {
436                 struct address_space *mapping = old->mapping;
437                 void (*freepage)(struct page *);
438
439                 pgoff_t offset = old->index;
440                 freepage = mapping->a_ops->freepage;
441
442                 page_cache_get(new);
443                 new->mapping = mapping;
444                 new->index = offset;
445
446                 spin_lock_irq(&mapping->tree_lock);
447                 __remove_from_page_cache(old);
448                 error = radix_tree_insert(&mapping->page_tree, offset, new);
449                 BUG_ON(error);
450                 mapping->nrpages++;
451                 __inc_zone_page_state(new, NR_FILE_PAGES);
452                 if (PageSwapBacked(new))
453                         __inc_zone_page_state(new, NR_SHMEM);
454                 spin_unlock_irq(&mapping->tree_lock);
455                 radix_tree_preload_end();
456                 if (freepage)
457                         freepage(old);
458                 page_cache_release(old);
459                 mem_cgroup_end_migration(memcg, old, new, true);
460         } else {
461                 mem_cgroup_end_migration(memcg, old, new, false);
462         }
463
464         return error;
465 }
466 EXPORT_SYMBOL_GPL(replace_page_cache_page);
467
468 /**
469  * add_to_page_cache_locked - add a locked page to the pagecache
470  * @page:       page to add
471  * @mapping:    the page's address_space
472  * @offset:     page index
473  * @gfp_mask:   page allocation mode
474  *
475  * This function is used to add a page to the pagecache. It must be locked.
476  * This function does not add the page to the LRU.  The caller must do that.
477  */
478 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
479                 pgoff_t offset, gfp_t gfp_mask)
480 {
481         int error;
482
483         VM_BUG_ON(!PageLocked(page));
484
485         error = mem_cgroup_cache_charge(page, current->mm,
486                                         gfp_mask & GFP_RECLAIM_MASK);
487         if (error)
488                 goto out;
489
490         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
491         if (error == 0) {
492                 page_cache_get(page);
493                 page->mapping = mapping;
494                 page->index = offset;
495
496                 spin_lock_irq(&mapping->tree_lock);
497                 error = radix_tree_insert(&mapping->page_tree, offset, page);
498                 if (likely(!error)) {
499                         mapping->nrpages++;
500                         __inc_zone_page_state(page, NR_FILE_PAGES);
501                         if (PageSwapBacked(page))
502                                 __inc_zone_page_state(page, NR_SHMEM);
503                         spin_unlock_irq(&mapping->tree_lock);
504                 } else {
505                         page->mapping = NULL;
506                         spin_unlock_irq(&mapping->tree_lock);
507                         mem_cgroup_uncharge_cache_page(page);
508                         page_cache_release(page);
509                 }
510                 radix_tree_preload_end();
511         } else
512                 mem_cgroup_uncharge_cache_page(page);
513 out:
514         return error;
515 }
516 EXPORT_SYMBOL(add_to_page_cache_locked);
517
518 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
519                                 pgoff_t offset, gfp_t gfp_mask)
520 {
521         int ret;
522
523         /*
524          * Splice_read and readahead add shmem/tmpfs pages into the page cache
525          * before shmem_readpage has a chance to mark them as SwapBacked: they
526          * need to go on the anon lru below, and mem_cgroup_cache_charge
527          * (called in add_to_page_cache) needs to know where they're going too.
528          */
529         if (mapping_cap_swap_backed(mapping))
530                 SetPageSwapBacked(page);
531
532         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
533         if (ret == 0) {
534                 if (page_is_file_cache(page))
535                         lru_cache_add_file(page);
536                 else
537                         lru_cache_add_anon(page);
538         }
539         return ret;
540 }
541 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
542
543 #ifdef CONFIG_NUMA
544 struct page *__page_cache_alloc(gfp_t gfp)
545 {
546         int n;
547         struct page *page;
548
549         if (cpuset_do_page_mem_spread()) {
550                 get_mems_allowed();
551                 n = cpuset_mem_spread_node();
552                 page = alloc_pages_exact_node(n, gfp, 0);
553                 put_mems_allowed();
554                 return page;
555         }
556         return alloc_pages(gfp, 0);
557 }
558 EXPORT_SYMBOL(__page_cache_alloc);
559 #endif
560
561 static int __sleep_on_page_lock(void *word)
562 {
563         io_schedule();
564         return 0;
565 }
566
567 /*
568  * In order to wait for pages to become available there must be
569  * waitqueues associated with pages. By using a hash table of
570  * waitqueues where the bucket discipline is to maintain all
571  * waiters on the same queue and wake all when any of the pages
572  * become available, and for the woken contexts to check to be
573  * sure the appropriate page became available, this saves space
574  * at a cost of "thundering herd" phenomena during rare hash
575  * collisions.
576  */
577 static wait_queue_head_t *page_waitqueue(struct page *page)
578 {
579         const struct zone *zone = page_zone(page);
580
581         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
582 }
583
584 static inline void wake_up_page(struct page *page, int bit)
585 {
586         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
587 }
588
589 void wait_on_page_bit(struct page *page, int bit_nr)
590 {
591         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
592
593         if (test_bit(bit_nr, &page->flags))
594                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
595                                                         TASK_UNINTERRUPTIBLE);
596 }
597 EXPORT_SYMBOL(wait_on_page_bit);
598
599 /**
600  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
601  * @page: Page defining the wait queue of interest
602  * @waiter: Waiter to add to the queue
603  *
604  * Add an arbitrary @waiter to the wait queue for the nominated @page.
605  */
606 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
607 {
608         wait_queue_head_t *q = page_waitqueue(page);
609         unsigned long flags;
610
611         spin_lock_irqsave(&q->lock, flags);
612         __add_wait_queue(q, waiter);
613         spin_unlock_irqrestore(&q->lock, flags);
614 }
615 EXPORT_SYMBOL_GPL(add_page_wait_queue);
616
617 /**
618  * unlock_page - unlock a locked page
619  * @page: the page
620  *
621  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
622  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
623  * mechananism between PageLocked pages and PageWriteback pages is shared.
624  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
625  *
626  * The mb is necessary to enforce ordering between the clear_bit and the read
627  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
628  */
629 void unlock_page(struct page *page)
630 {
631         VM_BUG_ON(!PageLocked(page));
632         clear_bit_unlock(PG_locked, &page->flags);
633         smp_mb__after_clear_bit();
634         wake_up_page(page, PG_locked);
635 }
636 EXPORT_SYMBOL(unlock_page);
637
638 /**
639  * end_page_writeback - end writeback against a page
640  * @page: the page
641  */
642 void end_page_writeback(struct page *page)
643 {
644         if (TestClearPageReclaim(page))
645                 rotate_reclaimable_page(page);
646
647         if (!test_clear_page_writeback(page))
648                 BUG();
649
650         smp_mb__after_clear_bit();
651         wake_up_page(page, PG_writeback);
652 }
653 EXPORT_SYMBOL(end_page_writeback);
654
655 /**
656  * __lock_page - get a lock on the page, assuming we need to sleep to get it
657  * @page: the page to lock
658  *
659  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
660  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
661  * chances are that on the second loop, the block layer's plug list is empty,
662  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
663  */
664 void __lock_page(struct page *page)
665 {
666         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
667
668         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
669                                                         TASK_UNINTERRUPTIBLE);
670 }
671 EXPORT_SYMBOL(__lock_page);
672
673 int __lock_page_killable(struct page *page)
674 {
675         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
676
677         return __wait_on_bit_lock(page_waitqueue(page), &wait,
678                                         sync_page_killable, TASK_KILLABLE);
679 }
680 EXPORT_SYMBOL_GPL(__lock_page_killable);
681
682 /**
683  * __lock_page_nosync - get a lock on the page, without calling sync_page()
684  * @page: the page to lock
685  *
686  * Variant of lock_page that does not require the caller to hold a reference
687  * on the page's mapping.
688  */
689 void __lock_page_nosync(struct page *page)
690 {
691         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
692         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
693                                                         TASK_UNINTERRUPTIBLE);
694 }
695
696 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
697                          unsigned int flags)
698 {
699         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
700                 __lock_page(page);
701                 return 1;
702         } else {
703                 if (!(flags & FAULT_FLAG_RETRY_NOWAIT)) {
704                         up_read(&mm->mmap_sem);
705                         wait_on_page_locked(page);
706                 }
707                 return 0;
708         }
709 }
710
711 /**
712  * find_get_page - find and get a page reference
713  * @mapping: the address_space to search
714  * @offset: the page index
715  *
716  * Is there a pagecache struct page at the given (mapping, offset) tuple?
717  * If yes, increment its refcount and return it; if no, return NULL.
718  */
719 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
720 {
721         void **pagep;
722         struct page *page;
723
724         rcu_read_lock();
725 repeat:
726         page = NULL;
727         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
728         if (pagep) {
729                 page = radix_tree_deref_slot(pagep);
730                 if (unlikely(!page))
731                         goto out;
732                 if (radix_tree_deref_retry(page))
733                         goto repeat;
734
735                 if (!page_cache_get_speculative(page))
736                         goto repeat;
737
738                 /*
739                  * Has the page moved?
740                  * This is part of the lockless pagecache protocol. See
741                  * include/linux/pagemap.h for details.
742                  */
743                 if (unlikely(page != *pagep)) {
744                         page_cache_release(page);
745                         goto repeat;
746                 }
747         }
748 out:
749         rcu_read_unlock();
750
751         return page;
752 }
753 EXPORT_SYMBOL(find_get_page);
754
755 /**
756  * find_lock_page - locate, pin and lock a pagecache page
757  * @mapping: the address_space to search
758  * @offset: the page index
759  *
760  * Locates the desired pagecache page, locks it, increments its reference
761  * count and returns its address.
762  *
763  * Returns zero if the page was not present. find_lock_page() may sleep.
764  */
765 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
766 {
767         struct page *page;
768
769 repeat:
770         page = find_get_page(mapping, offset);
771         if (page) {
772                 lock_page(page);
773                 /* Has the page been truncated? */
774                 if (unlikely(page->mapping != mapping)) {
775                         unlock_page(page);
776                         page_cache_release(page);
777                         goto repeat;
778                 }
779                 VM_BUG_ON(page->index != offset);
780         }
781         return page;
782 }
783 EXPORT_SYMBOL(find_lock_page);
784
785 /**
786  * find_or_create_page - locate or add a pagecache page
787  * @mapping: the page's address_space
788  * @index: the page's index into the mapping
789  * @gfp_mask: page allocation mode
790  *
791  * Locates a page in the pagecache.  If the page is not present, a new page
792  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
793  * LRU list.  The returned page is locked and has its reference count
794  * incremented.
795  *
796  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
797  * allocation!
798  *
799  * find_or_create_page() returns the desired page's address, or zero on
800  * memory exhaustion.
801  */
802 struct page *find_or_create_page(struct address_space *mapping,
803                 pgoff_t index, gfp_t gfp_mask)
804 {
805         struct page *page;
806         int err;
807 repeat:
808         page = find_lock_page(mapping, index);
809         if (!page) {
810                 page = __page_cache_alloc(gfp_mask);
811                 if (!page)
812                         return NULL;
813                 /*
814                  * We want a regular kernel memory (not highmem or DMA etc)
815                  * allocation for the radix tree nodes, but we need to honour
816                  * the context-specific requirements the caller has asked for.
817                  * GFP_RECLAIM_MASK collects those requirements.
818                  */
819                 err = add_to_page_cache_lru(page, mapping, index,
820                         (gfp_mask & GFP_RECLAIM_MASK));
821                 if (unlikely(err)) {
822                         page_cache_release(page);
823                         page = NULL;
824                         if (err == -EEXIST)
825                                 goto repeat;
826                 }
827         }
828         return page;
829 }
830 EXPORT_SYMBOL(find_or_create_page);
831
832 /**
833  * find_get_pages - gang pagecache lookup
834  * @mapping:    The address_space to search
835  * @start:      The starting page index
836  * @nr_pages:   The maximum number of pages
837  * @pages:      Where the resulting pages are placed
838  *
839  * find_get_pages() will search for and return a group of up to
840  * @nr_pages pages in the mapping.  The pages are placed at @pages.
841  * find_get_pages() takes a reference against the returned pages.
842  *
843  * The search returns a group of mapping-contiguous pages with ascending
844  * indexes.  There may be holes in the indices due to not-present pages.
845  *
846  * find_get_pages() returns the number of pages which were found.
847  */
848 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
849                             unsigned int nr_pages, struct page **pages)
850 {
851         unsigned int i;
852         unsigned int ret;
853         unsigned int nr_found;
854
855         rcu_read_lock();
856 restart:
857         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
858                                 (void ***)pages, start, nr_pages);
859         ret = 0;
860         for (i = 0; i < nr_found; i++) {
861                 struct page *page;
862 repeat:
863                 page = radix_tree_deref_slot((void **)pages[i]);
864                 if (unlikely(!page))
865                         continue;
866                 if (radix_tree_deref_retry(page)) {
867                         if (ret)
868                                 start = pages[ret-1]->index;
869                         goto restart;
870                 }
871
872                 if (!page_cache_get_speculative(page))
873                         goto repeat;
874
875                 /* Has the page moved? */
876                 if (unlikely(page != *((void **)pages[i]))) {
877                         page_cache_release(page);
878                         goto repeat;
879                 }
880
881                 pages[ret] = page;
882                 ret++;
883         }
884         rcu_read_unlock();
885         return ret;
886 }
887
888 /**
889  * find_get_pages_contig - gang contiguous pagecache lookup
890  * @mapping:    The address_space to search
891  * @index:      The starting page index
892  * @nr_pages:   The maximum number of pages
893  * @pages:      Where the resulting pages are placed
894  *
895  * find_get_pages_contig() works exactly like find_get_pages(), except
896  * that the returned number of pages are guaranteed to be contiguous.
897  *
898  * find_get_pages_contig() returns the number of pages which were found.
899  */
900 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
901                                unsigned int nr_pages, struct page **pages)
902 {
903         unsigned int i;
904         unsigned int ret;
905         unsigned int nr_found;
906
907         rcu_read_lock();
908 restart:
909         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
910                                 (void ***)pages, index, nr_pages);
911         ret = 0;
912         for (i = 0; i < nr_found; i++) {
913                 struct page *page;
914 repeat:
915                 page = radix_tree_deref_slot((void **)pages[i]);
916                 if (unlikely(!page))
917                         continue;
918                 if (radix_tree_deref_retry(page))
919                         goto restart;
920
921                 if (!page_cache_get_speculative(page))
922                         goto repeat;
923
924                 /* Has the page moved? */
925                 if (unlikely(page != *((void **)pages[i]))) {
926                         page_cache_release(page);
927                         goto repeat;
928                 }
929
930                 /*
931                  * must check mapping and index after taking the ref.
932                  * otherwise we can get both false positives and false
933                  * negatives, which is just confusing to the caller.
934                  */
935                 if (page->mapping == NULL || page->index != index) {
936                         page_cache_release(page);
937                         break;
938                 }
939
940                 pages[ret] = page;
941                 ret++;
942                 index++;
943         }
944         rcu_read_unlock();
945         return ret;
946 }
947 EXPORT_SYMBOL(find_get_pages_contig);
948
949 /**
950  * find_get_pages_tag - find and return pages that match @tag
951  * @mapping:    the address_space to search
952  * @index:      the starting page index
953  * @tag:        the tag index
954  * @nr_pages:   the maximum number of pages
955  * @pages:      where the resulting pages are placed
956  *
957  * Like find_get_pages, except we only return pages which are tagged with
958  * @tag.   We update @index to index the next page for the traversal.
959  */
960 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
961                         int tag, unsigned int nr_pages, struct page **pages)
962 {
963         unsigned int i;
964         unsigned int ret;
965         unsigned int nr_found;
966
967         rcu_read_lock();
968 restart:
969         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
970                                 (void ***)pages, *index, nr_pages, tag);
971         ret = 0;
972         for (i = 0; i < nr_found; i++) {
973                 struct page *page;
974 repeat:
975                 page = radix_tree_deref_slot((void **)pages[i]);
976                 if (unlikely(!page))
977                         continue;
978                 if (radix_tree_deref_retry(page))
979                         goto restart;
980
981                 if (!page_cache_get_speculative(page))
982                         goto repeat;
983
984                 /* Has the page moved? */
985                 if (unlikely(page != *((void **)pages[i]))) {
986                         page_cache_release(page);
987                         goto repeat;
988                 }
989
990                 pages[ret] = page;
991                 ret++;
992         }
993         rcu_read_unlock();
994
995         if (ret)
996                 *index = pages[ret - 1]->index + 1;
997
998         return ret;
999 }
1000 EXPORT_SYMBOL(find_get_pages_tag);
1001
1002 /**
1003  * grab_cache_page_nowait - returns locked page at given index in given cache
1004  * @mapping: target address_space
1005  * @index: the page index
1006  *
1007  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1008  * This is intended for speculative data generators, where the data can
1009  * be regenerated if the page couldn't be grabbed.  This routine should
1010  * be safe to call while holding the lock for another page.
1011  *
1012  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1013  * and deadlock against the caller's locked page.
1014  */
1015 struct page *
1016 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1017 {
1018         struct page *page = find_get_page(mapping, index);
1019
1020         if (page) {
1021                 if (trylock_page(page))
1022                         return page;
1023                 page_cache_release(page);
1024                 return NULL;
1025         }
1026         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1027         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1028                 page_cache_release(page);
1029                 page = NULL;
1030         }
1031         return page;
1032 }
1033 EXPORT_SYMBOL(grab_cache_page_nowait);
1034
1035 /*
1036  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1037  * a _large_ part of the i/o request. Imagine the worst scenario:
1038  *
1039  *      ---R__________________________________________B__________
1040  *         ^ reading here                             ^ bad block(assume 4k)
1041  *
1042  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1043  * => failing the whole request => read(R) => read(R+1) =>
1044  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1045  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1046  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1047  *
1048  * It is going insane. Fix it by quickly scaling down the readahead size.
1049  */
1050 static void shrink_readahead_size_eio(struct file *filp,
1051                                         struct file_ra_state *ra)
1052 {
1053         ra->ra_pages /= 4;
1054 }
1055
1056 /**
1057  * do_generic_file_read - generic file read routine
1058  * @filp:       the file to read
1059  * @ppos:       current file position
1060  * @desc:       read_descriptor
1061  * @actor:      read method
1062  *
1063  * This is a generic file read routine, and uses the
1064  * mapping->a_ops->readpage() function for the actual low-level stuff.
1065  *
1066  * This is really ugly. But the goto's actually try to clarify some
1067  * of the logic when it comes to error handling etc.
1068  */
1069 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1070                 read_descriptor_t *desc, read_actor_t actor)
1071 {
1072         struct address_space *mapping = filp->f_mapping;
1073         struct inode *inode = mapping->host;
1074         struct file_ra_state *ra = &filp->f_ra;
1075         pgoff_t index;
1076         pgoff_t last_index;
1077         pgoff_t prev_index;
1078         unsigned long offset;      /* offset into pagecache page */
1079         unsigned int prev_offset;
1080         int error;
1081
1082         index = *ppos >> PAGE_CACHE_SHIFT;
1083         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1084         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1085         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1086         offset = *ppos & ~PAGE_CACHE_MASK;
1087
1088         for (;;) {
1089                 struct page *page;
1090                 pgoff_t end_index;
1091                 loff_t isize;
1092                 unsigned long nr, ret;
1093
1094                 cond_resched();
1095 find_page:
1096                 page = find_get_page(mapping, index);
1097                 if (!page) {
1098                         page_cache_sync_readahead(mapping,
1099                                         ra, filp,
1100                                         index, last_index - index);
1101                         page = find_get_page(mapping, index);
1102                         if (unlikely(page == NULL))
1103                                 goto no_cached_page;
1104                 }
1105                 if (PageReadahead(page)) {
1106                         page_cache_async_readahead(mapping,
1107                                         ra, filp, page,
1108                                         index, last_index - index);
1109                 }
1110                 if (!PageUptodate(page)) {
1111                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1112                                         !mapping->a_ops->is_partially_uptodate)
1113                                 goto page_not_up_to_date;
1114                         if (!trylock_page(page))
1115                                 goto page_not_up_to_date;
1116                         /* Did it get truncated before we got the lock? */
1117                         if (!page->mapping)
1118                                 goto page_not_up_to_date_locked;
1119                         if (!mapping->a_ops->is_partially_uptodate(page,
1120                                                                 desc, offset))
1121                                 goto page_not_up_to_date_locked;
1122                         unlock_page(page);
1123                 }
1124 page_ok:
1125                 /*
1126                  * i_size must be checked after we know the page is Uptodate.
1127                  *
1128                  * Checking i_size after the check allows us to calculate
1129                  * the correct value for "nr", which means the zero-filled
1130                  * part of the page is not copied back to userspace (unless
1131                  * another truncate extends the file - this is desired though).
1132                  */
1133
1134                 isize = i_size_read(inode);
1135                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1136                 if (unlikely(!isize || index > end_index)) {
1137                         page_cache_release(page);
1138                         goto out;
1139                 }
1140
1141                 /* nr is the maximum number of bytes to copy from this page */
1142                 nr = PAGE_CACHE_SIZE;
1143                 if (index == end_index) {
1144                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1145                         if (nr <= offset) {
1146                                 page_cache_release(page);
1147                                 goto out;
1148                         }
1149                 }
1150                 nr = nr - offset;
1151
1152                 /* If users can be writing to this page using arbitrary
1153                  * virtual addresses, take care about potential aliasing
1154                  * before reading the page on the kernel side.
1155                  */
1156                 if (mapping_writably_mapped(mapping))
1157                         flush_dcache_page(page);
1158
1159                 /*
1160                  * When a sequential read accesses a page several times,
1161                  * only mark it as accessed the first time.
1162                  */
1163                 if (prev_index != index || offset != prev_offset)
1164                         mark_page_accessed(page);
1165                 prev_index = index;
1166
1167                 /*
1168                  * Ok, we have the page, and it's up-to-date, so
1169                  * now we can copy it to user space...
1170                  *
1171                  * The actor routine returns how many bytes were actually used..
1172                  * NOTE! This may not be the same as how much of a user buffer
1173                  * we filled up (we may be padding etc), so we can only update
1174                  * "pos" here (the actor routine has to update the user buffer
1175                  * pointers and the remaining count).
1176                  */
1177                 ret = actor(desc, page, offset, nr);
1178                 offset += ret;
1179                 index += offset >> PAGE_CACHE_SHIFT;
1180                 offset &= ~PAGE_CACHE_MASK;
1181                 prev_offset = offset;
1182
1183                 page_cache_release(page);
1184                 if (ret == nr && desc->count)
1185                         continue;
1186                 goto out;
1187
1188 page_not_up_to_date:
1189                 /* Get exclusive access to the page ... */
1190                 error = lock_page_killable(page);
1191                 if (unlikely(error))
1192                         goto readpage_error;
1193
1194 page_not_up_to_date_locked:
1195                 /* Did it get truncated before we got the lock? */
1196                 if (!page->mapping) {
1197                         unlock_page(page);
1198                         page_cache_release(page);
1199                         continue;
1200                 }
1201
1202                 /* Did somebody else fill it already? */
1203                 if (PageUptodate(page)) {
1204                         unlock_page(page);
1205                         goto page_ok;
1206                 }
1207
1208 readpage:
1209                 /*
1210                  * A previous I/O error may have been due to temporary
1211                  * failures, eg. multipath errors.
1212                  * PG_error will be set again if readpage fails.
1213                  */
1214                 ClearPageError(page);
1215                 /* Start the actual read. The read will unlock the page. */
1216                 error = mapping->a_ops->readpage(filp, page);
1217
1218                 if (unlikely(error)) {
1219                         if (error == AOP_TRUNCATED_PAGE) {
1220                                 page_cache_release(page);
1221                                 goto find_page;
1222                         }
1223                         goto readpage_error;
1224                 }
1225
1226                 if (!PageUptodate(page)) {
1227                         error = lock_page_killable(page);
1228                         if (unlikely(error))
1229                                 goto readpage_error;
1230                         if (!PageUptodate(page)) {
1231                                 if (page->mapping == NULL) {
1232                                         /*
1233                                          * invalidate_mapping_pages got it
1234                                          */
1235                                         unlock_page(page);
1236                                         page_cache_release(page);
1237                                         goto find_page;
1238                                 }
1239                                 unlock_page(page);
1240                                 shrink_readahead_size_eio(filp, ra);
1241                                 error = -EIO;
1242                                 goto readpage_error;
1243                         }
1244                         unlock_page(page);
1245                 }
1246
1247                 goto page_ok;
1248
1249 readpage_error:
1250                 /* UHHUH! A synchronous read error occurred. Report it */
1251                 desc->error = error;
1252                 page_cache_release(page);
1253                 goto out;
1254
1255 no_cached_page:
1256                 /*
1257                  * Ok, it wasn't cached, so we need to create a new
1258                  * page..
1259                  */
1260                 page = page_cache_alloc_cold(mapping);
1261                 if (!page) {
1262                         desc->error = -ENOMEM;
1263                         goto out;
1264                 }
1265                 error = add_to_page_cache_lru(page, mapping,
1266                                                 index, GFP_KERNEL);
1267                 if (error) {
1268                         page_cache_release(page);
1269                         if (error == -EEXIST)
1270                                 goto find_page;
1271                         desc->error = error;
1272                         goto out;
1273                 }
1274                 goto readpage;
1275         }
1276
1277 out:
1278         ra->prev_pos = prev_index;
1279         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1280         ra->prev_pos |= prev_offset;
1281
1282         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1283         file_accessed(filp);
1284 }
1285
1286 int file_read_actor(read_descriptor_t *desc, struct page *page,
1287                         unsigned long offset, unsigned long size)
1288 {
1289         char *kaddr;
1290         unsigned long left, count = desc->count;
1291
1292         if (size > count)
1293                 size = count;
1294
1295         /*
1296          * Faults on the destination of a read are common, so do it before
1297          * taking the kmap.
1298          */
1299         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1300                 kaddr = kmap_atomic(page, KM_USER0);
1301                 left = __copy_to_user_inatomic(desc->arg.buf,
1302                                                 kaddr + offset, size);
1303                 kunmap_atomic(kaddr, KM_USER0);
1304                 if (left == 0)
1305                         goto success;
1306         }
1307
1308         /* Do it the slow way */
1309         kaddr = kmap(page);
1310         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1311         kunmap(page);
1312
1313         if (left) {
1314                 size -= left;
1315                 desc->error = -EFAULT;
1316         }
1317 success:
1318         desc->count = count - size;
1319         desc->written += size;
1320         desc->arg.buf += size;
1321         return size;
1322 }
1323
1324 /*
1325  * Performs necessary checks before doing a write
1326  * @iov:        io vector request
1327  * @nr_segs:    number of segments in the iovec
1328  * @count:      number of bytes to write
1329  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1330  *
1331  * Adjust number of segments and amount of bytes to write (nr_segs should be
1332  * properly initialized first). Returns appropriate error code that caller
1333  * should return or zero in case that write should be allowed.
1334  */
1335 int generic_segment_checks(const struct iovec *iov,
1336                         unsigned long *nr_segs, size_t *count, int access_flags)
1337 {
1338         unsigned long   seg;
1339         size_t cnt = 0;
1340         for (seg = 0; seg < *nr_segs; seg++) {
1341                 const struct iovec *iv = &iov[seg];
1342
1343                 /*
1344                  * If any segment has a negative length, or the cumulative
1345                  * length ever wraps negative then return -EINVAL.
1346                  */
1347                 cnt += iv->iov_len;
1348                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1349                         return -EINVAL;
1350                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1351                         continue;
1352                 if (seg == 0)
1353                         return -EFAULT;
1354                 *nr_segs = seg;
1355                 cnt -= iv->iov_len;     /* This segment is no good */
1356                 break;
1357         }
1358         *count = cnt;
1359         return 0;
1360 }
1361 EXPORT_SYMBOL(generic_segment_checks);
1362
1363 /**
1364  * generic_file_aio_read - generic filesystem read routine
1365  * @iocb:       kernel I/O control block
1366  * @iov:        io vector request
1367  * @nr_segs:    number of segments in the iovec
1368  * @pos:        current file position
1369  *
1370  * This is the "read()" routine for all filesystems
1371  * that can use the page cache directly.
1372  */
1373 ssize_t
1374 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1375                 unsigned long nr_segs, loff_t pos)
1376 {
1377         struct file *filp = iocb->ki_filp;
1378         ssize_t retval;
1379         unsigned long seg = 0;
1380         size_t count;
1381         loff_t *ppos = &iocb->ki_pos;
1382
1383         count = 0;
1384         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1385         if (retval)
1386                 return retval;
1387
1388         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1389         if (filp->f_flags & O_DIRECT) {
1390                 loff_t size;
1391                 struct address_space *mapping;
1392                 struct inode *inode;
1393
1394                 mapping = filp->f_mapping;
1395                 inode = mapping->host;
1396                 if (!count)
1397                         goto out; /* skip atime */
1398                 size = i_size_read(inode);
1399                 if (pos < size) {
1400                         retval = filemap_write_and_wait_range(mapping, pos,
1401                                         pos + iov_length(iov, nr_segs) - 1);
1402                         if (!retval) {
1403                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1404                                                         iov, pos, nr_segs);
1405                         }
1406                         if (retval > 0) {
1407                                 *ppos = pos + retval;
1408                                 count -= retval;
1409                         }
1410
1411                         /*
1412                          * Btrfs can have a short DIO read if we encounter
1413                          * compressed extents, so if there was an error, or if
1414                          * we've already read everything we wanted to, or if
1415                          * there was a short read because we hit EOF, go ahead
1416                          * and return.  Otherwise fallthrough to buffered io for
1417                          * the rest of the read.
1418                          */
1419                         if (retval < 0 || !count || *ppos >= size) {
1420                                 file_accessed(filp);
1421                                 goto out;
1422                         }
1423                 }
1424         }
1425
1426         count = retval;
1427         for (seg = 0; seg < nr_segs; seg++) {
1428                 read_descriptor_t desc;
1429                 loff_t offset = 0;
1430
1431                 /*
1432                  * If we did a short DIO read we need to skip the section of the
1433                  * iov that we've already read data into.
1434                  */
1435                 if (count) {
1436                         if (count > iov[seg].iov_len) {
1437                                 count -= iov[seg].iov_len;
1438                                 continue;
1439                         }
1440                         offset = count;
1441                         count = 0;
1442                 }
1443
1444                 desc.written = 0;
1445                 desc.arg.buf = iov[seg].iov_base + offset;
1446                 desc.count = iov[seg].iov_len - offset;
1447                 if (desc.count == 0)
1448                         continue;
1449                 desc.error = 0;
1450                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1451                 retval += desc.written;
1452                 if (desc.error) {
1453                         retval = retval ?: desc.error;
1454                         break;
1455                 }
1456                 if (desc.count > 0)
1457                         break;
1458         }
1459 out:
1460         return retval;
1461 }
1462 EXPORT_SYMBOL(generic_file_aio_read);
1463
1464 static ssize_t
1465 do_readahead(struct address_space *mapping, struct file *filp,
1466              pgoff_t index, unsigned long nr)
1467 {
1468         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1469                 return -EINVAL;
1470
1471         force_page_cache_readahead(mapping, filp, index, nr);
1472         return 0;
1473 }
1474
1475 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1476 {
1477         ssize_t ret;
1478         struct file *file;
1479
1480         ret = -EBADF;
1481         file = fget(fd);
1482         if (file) {
1483                 if (file->f_mode & FMODE_READ) {
1484                         struct address_space *mapping = file->f_mapping;
1485                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1486                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1487                         unsigned long len = end - start + 1;
1488                         ret = do_readahead(mapping, file, start, len);
1489                 }
1490                 fput(file);
1491         }
1492         return ret;
1493 }
1494 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1495 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1496 {
1497         return SYSC_readahead((int) fd, offset, (size_t) count);
1498 }
1499 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1500 #endif
1501
1502 #ifdef CONFIG_MMU
1503 /**
1504  * page_cache_read - adds requested page to the page cache if not already there
1505  * @file:       file to read
1506  * @offset:     page index
1507  *
1508  * This adds the requested page to the page cache if it isn't already there,
1509  * and schedules an I/O to read in its contents from disk.
1510  */
1511 static int page_cache_read(struct file *file, pgoff_t offset)
1512 {
1513         struct address_space *mapping = file->f_mapping;
1514         struct page *page; 
1515         int ret;
1516
1517         do {
1518                 page = page_cache_alloc_cold(mapping);
1519                 if (!page)
1520                         return -ENOMEM;
1521
1522                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1523                 if (ret == 0)
1524                         ret = mapping->a_ops->readpage(file, page);
1525                 else if (ret == -EEXIST)
1526                         ret = 0; /* losing race to add is OK */
1527
1528                 page_cache_release(page);
1529
1530         } while (ret == AOP_TRUNCATED_PAGE);
1531                 
1532         return ret;
1533 }
1534
1535 #define MMAP_LOTSAMISS  (100)
1536
1537 /*
1538  * Synchronous readahead happens when we don't even find
1539  * a page in the page cache at all.
1540  */
1541 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1542                                    struct file_ra_state *ra,
1543                                    struct file *file,
1544                                    pgoff_t offset)
1545 {
1546         unsigned long ra_pages;
1547         struct address_space *mapping = file->f_mapping;
1548
1549         /* If we don't want any read-ahead, don't bother */
1550         if (VM_RandomReadHint(vma))
1551                 return;
1552
1553         if (VM_SequentialReadHint(vma) ||
1554                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1555                 page_cache_sync_readahead(mapping, ra, file, offset,
1556                                           ra->ra_pages);
1557                 return;
1558         }
1559
1560         if (ra->mmap_miss < INT_MAX)
1561                 ra->mmap_miss++;
1562
1563         /*
1564          * Do we miss much more than hit in this file? If so,
1565          * stop bothering with read-ahead. It will only hurt.
1566          */
1567         if (ra->mmap_miss > MMAP_LOTSAMISS)
1568                 return;
1569
1570         /*
1571          * mmap read-around
1572          */
1573         ra_pages = max_sane_readahead(ra->ra_pages);
1574         if (ra_pages) {
1575                 ra->start = max_t(long, 0, offset - ra_pages/2);
1576                 ra->size = ra_pages;
1577                 ra->async_size = 0;
1578                 ra_submit(ra, mapping, file);
1579         }
1580 }
1581
1582 /*
1583  * Asynchronous readahead happens when we find the page and PG_readahead,
1584  * so we want to possibly extend the readahead further..
1585  */
1586 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1587                                     struct file_ra_state *ra,
1588                                     struct file *file,
1589                                     struct page *page,
1590                                     pgoff_t offset)
1591 {
1592         struct address_space *mapping = file->f_mapping;
1593
1594         /* If we don't want any read-ahead, don't bother */
1595         if (VM_RandomReadHint(vma))
1596                 return;
1597         if (ra->mmap_miss > 0)
1598                 ra->mmap_miss--;
1599         if (PageReadahead(page))
1600                 page_cache_async_readahead(mapping, ra, file,
1601                                            page, offset, ra->ra_pages);
1602 }
1603
1604 /**
1605  * filemap_fault - read in file data for page fault handling
1606  * @vma:        vma in which the fault was taken
1607  * @vmf:        struct vm_fault containing details of the fault
1608  *
1609  * filemap_fault() is invoked via the vma operations vector for a
1610  * mapped memory region to read in file data during a page fault.
1611  *
1612  * The goto's are kind of ugly, but this streamlines the normal case of having
1613  * it in the page cache, and handles the special cases reasonably without
1614  * having a lot of duplicated code.
1615  */
1616 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1617 {
1618         int error;
1619         struct file *file = vma->vm_file;
1620         struct address_space *mapping = file->f_mapping;
1621         struct file_ra_state *ra = &file->f_ra;
1622         struct inode *inode = mapping->host;
1623         pgoff_t offset = vmf->pgoff;
1624         struct page *page;
1625         pgoff_t size;
1626         int ret = 0;
1627
1628         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1629         if (offset >= size)
1630                 return VM_FAULT_SIGBUS;
1631
1632         /*
1633          * Do we have something in the page cache already?
1634          */
1635         page = find_get_page(mapping, offset);
1636         if (likely(page)) {
1637                 /*
1638                  * We found the page, so try async readahead before
1639                  * waiting for the lock.
1640                  */
1641                 do_async_mmap_readahead(vma, ra, file, page, offset);
1642         } else {
1643                 /* No page in the page cache at all */
1644                 do_sync_mmap_readahead(vma, ra, file, offset);
1645                 count_vm_event(PGMAJFAULT);
1646                 ret = VM_FAULT_MAJOR;
1647 retry_find:
1648                 page = find_get_page(mapping, offset);
1649                 if (!page)
1650                         goto no_cached_page;
1651         }
1652
1653         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1654                 page_cache_release(page);
1655                 return ret | VM_FAULT_RETRY;
1656         }
1657
1658         /* Did it get truncated? */
1659         if (unlikely(page->mapping != mapping)) {
1660                 unlock_page(page);
1661                 put_page(page);
1662                 goto retry_find;
1663         }
1664         VM_BUG_ON(page->index != offset);
1665
1666         /*
1667          * We have a locked page in the page cache, now we need to check
1668          * that it's up-to-date. If not, it is going to be due to an error.
1669          */
1670         if (unlikely(!PageUptodate(page)))
1671                 goto page_not_uptodate;
1672
1673         /*
1674          * Found the page and have a reference on it.
1675          * We must recheck i_size under page lock.
1676          */
1677         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1678         if (unlikely(offset >= size)) {
1679                 unlock_page(page);
1680                 page_cache_release(page);
1681                 return VM_FAULT_SIGBUS;
1682         }
1683
1684         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1685         vmf->page = page;
1686         return ret | VM_FAULT_LOCKED;
1687
1688 no_cached_page:
1689         /*
1690          * We're only likely to ever get here if MADV_RANDOM is in
1691          * effect.
1692          */
1693         error = page_cache_read(file, offset);
1694
1695         /*
1696          * The page we want has now been added to the page cache.
1697          * In the unlikely event that someone removed it in the
1698          * meantime, we'll just come back here and read it again.
1699          */
1700         if (error >= 0)
1701                 goto retry_find;
1702
1703         /*
1704          * An error return from page_cache_read can result if the
1705          * system is low on memory, or a problem occurs while trying
1706          * to schedule I/O.
1707          */
1708         if (error == -ENOMEM)
1709                 return VM_FAULT_OOM;
1710         return VM_FAULT_SIGBUS;
1711
1712 page_not_uptodate:
1713         /*
1714          * Umm, take care of errors if the page isn't up-to-date.
1715          * Try to re-read it _once_. We do this synchronously,
1716          * because there really aren't any performance issues here
1717          * and we need to check for errors.
1718          */
1719         ClearPageError(page);
1720         error = mapping->a_ops->readpage(file, page);
1721         if (!error) {
1722                 wait_on_page_locked(page);
1723                 if (!PageUptodate(page))
1724                         error = -EIO;
1725         }
1726         page_cache_release(page);
1727
1728         if (!error || error == AOP_TRUNCATED_PAGE)
1729                 goto retry_find;
1730
1731         /* Things didn't work out. Return zero to tell the mm layer so. */
1732         shrink_readahead_size_eio(file, ra);
1733         return VM_FAULT_SIGBUS;
1734 }
1735 EXPORT_SYMBOL(filemap_fault);
1736
1737 const struct vm_operations_struct generic_file_vm_ops = {
1738         .fault          = filemap_fault,
1739 };
1740
1741 /* This is used for a general mmap of a disk file */
1742
1743 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1744 {
1745         struct address_space *mapping = file->f_mapping;
1746
1747         if (!mapping->a_ops->readpage)
1748                 return -ENOEXEC;
1749         file_accessed(file);
1750         vma->vm_ops = &generic_file_vm_ops;
1751         vma->vm_flags |= VM_CAN_NONLINEAR;
1752         return 0;
1753 }
1754
1755 /*
1756  * This is for filesystems which do not implement ->writepage.
1757  */
1758 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1759 {
1760         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1761                 return -EINVAL;
1762         return generic_file_mmap(file, vma);
1763 }
1764 #else
1765 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1766 {
1767         return -ENOSYS;
1768 }
1769 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1770 {
1771         return -ENOSYS;
1772 }
1773 #endif /* CONFIG_MMU */
1774
1775 EXPORT_SYMBOL(generic_file_mmap);
1776 EXPORT_SYMBOL(generic_file_readonly_mmap);
1777
1778 static struct page *__read_cache_page(struct address_space *mapping,
1779                                 pgoff_t index,
1780                                 int (*filler)(void *,struct page*),
1781                                 void *data,
1782                                 gfp_t gfp)
1783 {
1784         struct page *page;
1785         int err;
1786 repeat:
1787         page = find_get_page(mapping, index);
1788         if (!page) {
1789                 page = __page_cache_alloc(gfp | __GFP_COLD);
1790                 if (!page)
1791                         return ERR_PTR(-ENOMEM);
1792                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1793                 if (unlikely(err)) {
1794                         page_cache_release(page);
1795                         if (err == -EEXIST)
1796                                 goto repeat;
1797                         /* Presumably ENOMEM for radix tree node */
1798                         return ERR_PTR(err);
1799                 }
1800                 err = filler(data, page);
1801                 if (err < 0) {
1802                         page_cache_release(page);
1803                         page = ERR_PTR(err);
1804                 }
1805         }
1806         return page;
1807 }
1808
1809 static struct page *do_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 {
1816         struct page *page;
1817         int err;
1818
1819 retry:
1820         page = __read_cache_page(mapping, index, filler, data, gfp);
1821         if (IS_ERR(page))
1822                 return page;
1823         if (PageUptodate(page))
1824                 goto out;
1825
1826         lock_page(page);
1827         if (!page->mapping) {
1828                 unlock_page(page);
1829                 page_cache_release(page);
1830                 goto retry;
1831         }
1832         if (PageUptodate(page)) {
1833                 unlock_page(page);
1834                 goto out;
1835         }
1836         err = filler(data, page);
1837         if (err < 0) {
1838                 page_cache_release(page);
1839                 return ERR_PTR(err);
1840         }
1841 out:
1842         mark_page_accessed(page);
1843         return page;
1844 }
1845
1846 /**
1847  * read_cache_page_async - read into page cache, fill it if needed
1848  * @mapping:    the page's address_space
1849  * @index:      the page index
1850  * @filler:     function to perform the read
1851  * @data:       destination for read data
1852  *
1853  * Same as read_cache_page, but don't wait for page to become unlocked
1854  * after submitting it to the filler.
1855  *
1856  * Read into the page cache. If a page already exists, and PageUptodate() is
1857  * not set, try to fill the page but don't wait for it to become unlocked.
1858  *
1859  * If the page does not get brought uptodate, return -EIO.
1860  */
1861 struct page *read_cache_page_async(struct address_space *mapping,
1862                                 pgoff_t index,
1863                                 int (*filler)(void *,struct page*),
1864                                 void *data)
1865 {
1866         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1867 }
1868 EXPORT_SYMBOL(read_cache_page_async);
1869
1870 static struct page *wait_on_page_read(struct page *page)
1871 {
1872         if (!IS_ERR(page)) {
1873                 wait_on_page_locked(page);
1874                 if (!PageUptodate(page)) {
1875                         page_cache_release(page);
1876                         page = ERR_PTR(-EIO);
1877                 }
1878         }
1879         return page;
1880 }
1881
1882 /**
1883  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1884  * @mapping:    the page's address_space
1885  * @index:      the page index
1886  * @gfp:        the page allocator flags to use if allocating
1887  *
1888  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1889  * any new page allocations done using the specified allocation flags. Note
1890  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1891  * expect to do this atomically or anything like that - but you can pass in
1892  * other page requirements.
1893  *
1894  * If the page does not get brought uptodate, return -EIO.
1895  */
1896 struct page *read_cache_page_gfp(struct address_space *mapping,
1897                                 pgoff_t index,
1898                                 gfp_t gfp)
1899 {
1900         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1901
1902         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1903 }
1904 EXPORT_SYMBOL(read_cache_page_gfp);
1905
1906 /**
1907  * read_cache_page - read into page cache, fill it if needed
1908  * @mapping:    the page's address_space
1909  * @index:      the page index
1910  * @filler:     function to perform the read
1911  * @data:       destination for read data
1912  *
1913  * Read into the page cache. If a page already exists, and PageUptodate() is
1914  * not set, try to fill the page then wait for it to become unlocked.
1915  *
1916  * If the page does not get brought uptodate, return -EIO.
1917  */
1918 struct page *read_cache_page(struct address_space *mapping,
1919                                 pgoff_t index,
1920                                 int (*filler)(void *,struct page*),
1921                                 void *data)
1922 {
1923         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1924 }
1925 EXPORT_SYMBOL(read_cache_page);
1926
1927 /*
1928  * The logic we want is
1929  *
1930  *      if suid or (sgid and xgrp)
1931  *              remove privs
1932  */
1933 int should_remove_suid(struct dentry *dentry)
1934 {
1935         mode_t mode = dentry->d_inode->i_mode;
1936         int kill = 0;
1937
1938         /* suid always must be killed */
1939         if (unlikely(mode & S_ISUID))
1940                 kill = ATTR_KILL_SUID;
1941
1942         /*
1943          * sgid without any exec bits is just a mandatory locking mark; leave
1944          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1945          */
1946         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1947                 kill |= ATTR_KILL_SGID;
1948
1949         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1950                 return kill;
1951
1952         return 0;
1953 }
1954 EXPORT_SYMBOL(should_remove_suid);
1955
1956 static int __remove_suid(struct dentry *dentry, int kill)
1957 {
1958         struct iattr newattrs;
1959
1960         newattrs.ia_valid = ATTR_FORCE | kill;
1961         return notify_change(dentry, &newattrs);
1962 }
1963
1964 int file_remove_suid(struct file *file)
1965 {
1966         struct dentry *dentry = file->f_path.dentry;
1967         int killsuid = should_remove_suid(dentry);
1968         int killpriv = security_inode_need_killpriv(dentry);
1969         int error = 0;
1970
1971         if (killpriv < 0)
1972                 return killpriv;
1973         if (killpriv)
1974                 error = security_inode_killpriv(dentry);
1975         if (!error && killsuid)
1976                 error = __remove_suid(dentry, killsuid);
1977
1978         return error;
1979 }
1980 EXPORT_SYMBOL(file_remove_suid);
1981
1982 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1983                         const struct iovec *iov, size_t base, size_t bytes)
1984 {
1985         size_t copied = 0, left = 0;
1986
1987         while (bytes) {
1988                 char __user *buf = iov->iov_base + base;
1989                 int copy = min(bytes, iov->iov_len - base);
1990
1991                 base = 0;
1992                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1993                 copied += copy;
1994                 bytes -= copy;
1995                 vaddr += copy;
1996                 iov++;
1997
1998                 if (unlikely(left))
1999                         break;
2000         }
2001         return copied - left;
2002 }
2003
2004 /*
2005  * Copy as much as we can into the page and return the number of bytes which
2006  * were successfully copied.  If a fault is encountered then return the number of
2007  * bytes which were copied.
2008  */
2009 size_t iov_iter_copy_from_user_atomic(struct page *page,
2010                 struct iov_iter *i, unsigned long offset, size_t bytes)
2011 {
2012         char *kaddr;
2013         size_t copied;
2014
2015         BUG_ON(!in_atomic());
2016         kaddr = kmap_atomic(page, KM_USER0);
2017         if (likely(i->nr_segs == 1)) {
2018                 int left;
2019                 char __user *buf = i->iov->iov_base + i->iov_offset;
2020                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2021                 copied = bytes - left;
2022         } else {
2023                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2024                                                 i->iov, i->iov_offset, bytes);
2025         }
2026         kunmap_atomic(kaddr, KM_USER0);
2027
2028         return copied;
2029 }
2030 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2031
2032 /*
2033  * This has the same sideeffects and return value as
2034  * iov_iter_copy_from_user_atomic().
2035  * The difference is that it attempts to resolve faults.
2036  * Page must not be locked.
2037  */
2038 size_t iov_iter_copy_from_user(struct page *page,
2039                 struct iov_iter *i, unsigned long offset, size_t bytes)
2040 {
2041         char *kaddr;
2042         size_t copied;
2043
2044         kaddr = kmap(page);
2045         if (likely(i->nr_segs == 1)) {
2046                 int left;
2047                 char __user *buf = i->iov->iov_base + i->iov_offset;
2048                 left = __copy_from_user(kaddr + offset, buf, bytes);
2049                 copied = bytes - left;
2050         } else {
2051                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2052                                                 i->iov, i->iov_offset, bytes);
2053         }
2054         kunmap(page);
2055         return copied;
2056 }
2057 EXPORT_SYMBOL(iov_iter_copy_from_user);
2058
2059 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2060 {
2061         BUG_ON(i->count < bytes);
2062
2063         if (likely(i->nr_segs == 1)) {
2064                 i->iov_offset += bytes;
2065                 i->count -= bytes;
2066         } else {
2067                 const struct iovec *iov = i->iov;
2068                 size_t base = i->iov_offset;
2069
2070                 /*
2071                  * The !iov->iov_len check ensures we skip over unlikely
2072                  * zero-length segments (without overruning the iovec).
2073                  */
2074                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2075                         int copy;
2076
2077                         copy = min(bytes, iov->iov_len - base);
2078                         BUG_ON(!i->count || i->count < copy);
2079                         i->count -= copy;
2080                         bytes -= copy;
2081                         base += copy;
2082                         if (iov->iov_len == base) {
2083                                 iov++;
2084                                 base = 0;
2085                         }
2086                 }
2087                 i->iov = iov;
2088                 i->iov_offset = base;
2089         }
2090 }
2091 EXPORT_SYMBOL(iov_iter_advance);
2092
2093 /*
2094  * Fault in the first iovec of the given iov_iter, to a maximum length
2095  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2096  * accessed (ie. because it is an invalid address).
2097  *
2098  * writev-intensive code may want this to prefault several iovecs -- that
2099  * would be possible (callers must not rely on the fact that _only_ the
2100  * first iovec will be faulted with the current implementation).
2101  */
2102 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2103 {
2104         char __user *buf = i->iov->iov_base + i->iov_offset;
2105         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2106         return fault_in_pages_readable(buf, bytes);
2107 }
2108 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2109
2110 /*
2111  * Return the count of just the current iov_iter segment.
2112  */
2113 size_t iov_iter_single_seg_count(struct iov_iter *i)
2114 {
2115         const struct iovec *iov = i->iov;
2116         if (i->nr_segs == 1)
2117                 return i->count;
2118         else
2119                 return min(i->count, iov->iov_len - i->iov_offset);
2120 }
2121 EXPORT_SYMBOL(iov_iter_single_seg_count);
2122
2123 /*
2124  * Performs necessary checks before doing a write
2125  *
2126  * Can adjust writing position or amount of bytes to write.
2127  * Returns appropriate error code that caller should return or
2128  * zero in case that write should be allowed.
2129  */
2130 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2131 {
2132         struct inode *inode = file->f_mapping->host;
2133         unsigned long limit = rlimit(RLIMIT_FSIZE);
2134
2135         if (unlikely(*pos < 0))
2136                 return -EINVAL;
2137
2138         if (!isblk) {
2139                 /* FIXME: this is for backwards compatibility with 2.4 */
2140                 if (file->f_flags & O_APPEND)
2141                         *pos = i_size_read(inode);
2142
2143                 if (limit != RLIM_INFINITY) {
2144                         if (*pos >= limit) {
2145                                 send_sig(SIGXFSZ, current, 0);
2146                                 return -EFBIG;
2147                         }
2148                         if (*count > limit - (typeof(limit))*pos) {
2149                                 *count = limit - (typeof(limit))*pos;
2150                         }
2151                 }
2152         }
2153
2154         /*
2155          * LFS rule
2156          */
2157         if (unlikely(*pos + *count > MAX_NON_LFS &&
2158                                 !(file->f_flags & O_LARGEFILE))) {
2159                 if (*pos >= MAX_NON_LFS) {
2160                         return -EFBIG;
2161                 }
2162                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2163                         *count = MAX_NON_LFS - (unsigned long)*pos;
2164                 }
2165         }
2166
2167         /*
2168          * Are we about to exceed the fs block limit ?
2169          *
2170          * If we have written data it becomes a short write.  If we have
2171          * exceeded without writing data we send a signal and return EFBIG.
2172          * Linus frestrict idea will clean these up nicely..
2173          */
2174         if (likely(!isblk)) {
2175                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2176                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2177                                 return -EFBIG;
2178                         }
2179                         /* zero-length writes at ->s_maxbytes are OK */
2180                 }
2181
2182                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2183                         *count = inode->i_sb->s_maxbytes - *pos;
2184         } else {
2185 #ifdef CONFIG_BLOCK
2186                 loff_t isize;
2187                 if (bdev_read_only(I_BDEV(inode)))
2188                         return -EPERM;
2189                 isize = i_size_read(inode);
2190                 if (*pos >= isize) {
2191                         if (*count || *pos > isize)
2192                                 return -ENOSPC;
2193                 }
2194
2195                 if (*pos + *count > isize)
2196                         *count = isize - *pos;
2197 #else
2198                 return -EPERM;
2199 #endif
2200         }
2201         return 0;
2202 }
2203 EXPORT_SYMBOL(generic_write_checks);
2204
2205 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2206                                 loff_t pos, unsigned len, unsigned flags,
2207                                 struct page **pagep, void **fsdata)
2208 {
2209         const struct address_space_operations *aops = mapping->a_ops;
2210
2211         return aops->write_begin(file, mapping, pos, len, flags,
2212                                                         pagep, fsdata);
2213 }
2214 EXPORT_SYMBOL(pagecache_write_begin);
2215
2216 int pagecache_write_end(struct file *file, struct address_space *mapping,
2217                                 loff_t pos, unsigned len, unsigned copied,
2218                                 struct page *page, void *fsdata)
2219 {
2220         const struct address_space_operations *aops = mapping->a_ops;
2221
2222         mark_page_accessed(page);
2223         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2224 }
2225 EXPORT_SYMBOL(pagecache_write_end);
2226
2227 ssize_t
2228 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2229                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2230                 size_t count, size_t ocount)
2231 {
2232         struct file     *file = iocb->ki_filp;
2233         struct address_space *mapping = file->f_mapping;
2234         struct inode    *inode = mapping->host;
2235         ssize_t         written;
2236         size_t          write_len;
2237         pgoff_t         end;
2238
2239         if (count != ocount)
2240                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2241
2242         write_len = iov_length(iov, *nr_segs);
2243         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2244
2245         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2246         if (written)
2247                 goto out;
2248
2249         /*
2250          * After a write we want buffered reads to be sure to go to disk to get
2251          * the new data.  We invalidate clean cached page from the region we're
2252          * about to write.  We do this *before* the write so that we can return
2253          * without clobbering -EIOCBQUEUED from ->direct_IO().
2254          */
2255         if (mapping->nrpages) {
2256                 written = invalidate_inode_pages2_range(mapping,
2257                                         pos >> PAGE_CACHE_SHIFT, end);
2258                 /*
2259                  * If a page can not be invalidated, return 0 to fall back
2260                  * to buffered write.
2261                  */
2262                 if (written) {
2263                         if (written == -EBUSY)
2264                                 return 0;
2265                         goto out;
2266                 }
2267         }
2268
2269         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2270
2271         /*
2272          * Finally, try again to invalidate clean pages which might have been
2273          * cached by non-direct readahead, or faulted in by get_user_pages()
2274          * if the source of the write was an mmap'ed region of the file
2275          * we're writing.  Either one is a pretty crazy thing to do,
2276          * so we don't support it 100%.  If this invalidation
2277          * fails, tough, the write still worked...
2278          */
2279         if (mapping->nrpages) {
2280                 invalidate_inode_pages2_range(mapping,
2281                                               pos >> PAGE_CACHE_SHIFT, end);
2282         }
2283
2284         if (written > 0) {
2285                 pos += written;
2286                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2287                         i_size_write(inode, pos);
2288                         mark_inode_dirty(inode);
2289                 }
2290                 *ppos = pos;
2291         }
2292 out:
2293         return written;
2294 }
2295 EXPORT_SYMBOL(generic_file_direct_write);
2296
2297 /*
2298  * Find or create a page at the given pagecache position. Return the locked
2299  * page. This function is specifically for buffered writes.
2300  */
2301 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2302                                         pgoff_t index, unsigned flags)
2303 {
2304         int status;
2305         struct page *page;
2306         gfp_t gfp_notmask = 0;
2307         if (flags & AOP_FLAG_NOFS)
2308                 gfp_notmask = __GFP_FS;
2309 repeat:
2310         page = find_lock_page(mapping, index);
2311         if (page)
2312                 return page;
2313
2314         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2315         if (!page)
2316                 return NULL;
2317         status = add_to_page_cache_lru(page, mapping, index,
2318                                                 GFP_KERNEL & ~gfp_notmask);
2319         if (unlikely(status)) {
2320                 page_cache_release(page);
2321                 if (status == -EEXIST)
2322                         goto repeat;
2323                 return NULL;
2324         }
2325         return page;
2326 }
2327 EXPORT_SYMBOL(grab_cache_page_write_begin);
2328
2329 static ssize_t generic_perform_write(struct file *file,
2330                                 struct iov_iter *i, loff_t pos)
2331 {
2332         struct address_space *mapping = file->f_mapping;
2333         const struct address_space_operations *a_ops = mapping->a_ops;
2334         long status = 0;
2335         ssize_t written = 0;
2336         unsigned int flags = 0;
2337
2338         /*
2339          * Copies from kernel address space cannot fail (NFSD is a big user).
2340          */
2341         if (segment_eq(get_fs(), KERNEL_DS))
2342                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2343
2344         do {
2345                 struct page *page;
2346                 unsigned long offset;   /* Offset into pagecache page */
2347                 unsigned long bytes;    /* Bytes to write to page */
2348                 size_t copied;          /* Bytes copied from user */
2349                 void *fsdata;
2350
2351                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2352                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2353                                                 iov_iter_count(i));
2354
2355 again:
2356
2357                 /*
2358                  * Bring in the user page that we will copy from _first_.
2359                  * Otherwise there's a nasty deadlock on copying from the
2360                  * same page as we're writing to, without it being marked
2361                  * up-to-date.
2362                  *
2363                  * Not only is this an optimisation, but it is also required
2364                  * to check that the address is actually valid, when atomic
2365                  * usercopies are used, below.
2366                  */
2367                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2368                         status = -EFAULT;
2369                         break;
2370                 }
2371
2372                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2373                                                 &page, &fsdata);
2374                 if (unlikely(status))
2375                         break;
2376
2377                 if (mapping_writably_mapped(mapping))
2378                         flush_dcache_page(page);
2379
2380                 pagefault_disable();
2381                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2382                 pagefault_enable();
2383                 flush_dcache_page(page);
2384
2385                 mark_page_accessed(page);
2386                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2387                                                 page, fsdata);
2388                 if (unlikely(status < 0))
2389                         break;
2390                 copied = status;
2391
2392                 cond_resched();
2393
2394                 iov_iter_advance(i, copied);
2395                 if (unlikely(copied == 0)) {
2396                         /*
2397                          * If we were unable to copy any data at all, we must
2398                          * fall back to a single segment length write.
2399                          *
2400                          * If we didn't fallback here, we could livelock
2401                          * because not all segments in the iov can be copied at
2402                          * once without a pagefault.
2403                          */
2404                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2405                                                 iov_iter_single_seg_count(i));
2406                         goto again;
2407                 }
2408                 pos += copied;
2409                 written += copied;
2410
2411                 balance_dirty_pages_ratelimited(mapping);
2412
2413         } while (iov_iter_count(i));
2414
2415         return written ? written : status;
2416 }
2417
2418 ssize_t
2419 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2420                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2421                 size_t count, ssize_t written)
2422 {
2423         struct file *file = iocb->ki_filp;
2424         ssize_t status;
2425         struct iov_iter i;
2426
2427         iov_iter_init(&i, iov, nr_segs, count, written);
2428         status = generic_perform_write(file, &i, pos);
2429
2430         if (likely(status >= 0)) {
2431                 written += status;
2432                 *ppos = pos + status;
2433         }
2434         
2435         return written ? written : status;
2436 }
2437 EXPORT_SYMBOL(generic_file_buffered_write);
2438
2439 /**
2440  * __generic_file_aio_write - write data to a file
2441  * @iocb:       IO state structure (file, offset, etc.)
2442  * @iov:        vector with data to write
2443  * @nr_segs:    number of segments in the vector
2444  * @ppos:       position where to write
2445  *
2446  * This function does all the work needed for actually writing data to a
2447  * file. It does all basic checks, removes SUID from the file, updates
2448  * modification times and calls proper subroutines depending on whether we
2449  * do direct IO or a standard buffered write.
2450  *
2451  * It expects i_mutex to be grabbed unless we work on a block device or similar
2452  * object which does not need locking at all.
2453  *
2454  * This function does *not* take care of syncing data in case of O_SYNC write.
2455  * A caller has to handle it. This is mainly due to the fact that we want to
2456  * avoid syncing under i_mutex.
2457  */
2458 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2459                                  unsigned long nr_segs, loff_t *ppos)
2460 {
2461         struct file *file = iocb->ki_filp;
2462         struct address_space * mapping = file->f_mapping;
2463         size_t ocount;          /* original count */
2464         size_t count;           /* after file limit checks */
2465         struct inode    *inode = mapping->host;
2466         loff_t          pos;
2467         ssize_t         written;
2468         ssize_t         err;
2469
2470         ocount = 0;
2471         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2472         if (err)
2473                 return err;
2474
2475         count = ocount;
2476         pos = *ppos;
2477
2478         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2479
2480         /* We can write back this queue in page reclaim */
2481         current->backing_dev_info = mapping->backing_dev_info;
2482         written = 0;
2483
2484         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2485         if (err)
2486                 goto out;
2487
2488         if (count == 0)
2489                 goto out;
2490
2491         err = file_remove_suid(file);
2492         if (err)
2493                 goto out;
2494
2495         file_update_time(file);
2496
2497         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2498         if (unlikely(file->f_flags & O_DIRECT)) {
2499                 loff_t endbyte;
2500                 ssize_t written_buffered;
2501
2502                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2503                                                         ppos, count, ocount);
2504                 if (written < 0 || written == count)
2505                         goto out;
2506                 /*
2507                  * direct-io write to a hole: fall through to buffered I/O
2508                  * for completing the rest of the request.
2509                  */
2510                 pos += written;
2511                 count -= written;
2512                 written_buffered = generic_file_buffered_write(iocb, iov,
2513                                                 nr_segs, pos, ppos, count,
2514                                                 written);
2515                 /*
2516                  * If generic_file_buffered_write() retuned a synchronous error
2517                  * then we want to return the number of bytes which were
2518                  * direct-written, or the error code if that was zero.  Note
2519                  * that this differs from normal direct-io semantics, which
2520                  * will return -EFOO even if some bytes were written.
2521                  */
2522                 if (written_buffered < 0) {
2523                         err = written_buffered;
2524                         goto out;
2525                 }
2526
2527                 /*
2528                  * We need to ensure that the page cache pages are written to
2529                  * disk and invalidated to preserve the expected O_DIRECT
2530                  * semantics.
2531                  */
2532                 endbyte = pos + written_buffered - written - 1;
2533                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2534                 if (err == 0) {
2535                         written = written_buffered;
2536                         invalidate_mapping_pages(mapping,
2537                                                  pos >> PAGE_CACHE_SHIFT,
2538                                                  endbyte >> PAGE_CACHE_SHIFT);
2539                 } else {
2540                         /*
2541                          * We don't know how much we wrote, so just return
2542                          * the number of bytes which were direct-written
2543                          */
2544                 }
2545         } else {
2546                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2547                                 pos, ppos, count, written);
2548         }
2549 out:
2550         current->backing_dev_info = NULL;
2551         return written ? written : err;
2552 }
2553 EXPORT_SYMBOL(__generic_file_aio_write);
2554
2555 /**
2556  * generic_file_aio_write - write data to a file
2557  * @iocb:       IO state structure
2558  * @iov:        vector with data to write
2559  * @nr_segs:    number of segments in the vector
2560  * @pos:        position in file where to write
2561  *
2562  * This is a wrapper around __generic_file_aio_write() to be used by most
2563  * filesystems. It takes care of syncing the file in case of O_SYNC file
2564  * and acquires i_mutex as needed.
2565  */
2566 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2567                 unsigned long nr_segs, loff_t pos)
2568 {
2569         struct file *file = iocb->ki_filp;
2570         struct inode *inode = file->f_mapping->host;
2571         ssize_t ret;
2572
2573         BUG_ON(iocb->ki_pos != pos);
2574
2575         mutex_lock(&inode->i_mutex);
2576         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2577         mutex_unlock(&inode->i_mutex);
2578
2579         if (ret > 0 || ret == -EIOCBQUEUED) {
2580                 ssize_t err;
2581
2582                 err = generic_write_sync(file, pos, ret);
2583                 if (err < 0 && ret > 0)
2584                         ret = err;
2585         }
2586         return ret;
2587 }
2588 EXPORT_SYMBOL(generic_file_aio_write);
2589
2590 /**
2591  * try_to_release_page() - release old fs-specific metadata on a page
2592  *
2593  * @page: the page which the kernel is trying to free
2594  * @gfp_mask: memory allocation flags (and I/O mode)
2595  *
2596  * The address_space is to try to release any data against the page
2597  * (presumably at page->private).  If the release was successful, return `1'.
2598  * Otherwise return zero.
2599  *
2600  * This may also be called if PG_fscache is set on a page, indicating that the
2601  * page is known to the local caching routines.
2602  *
2603  * The @gfp_mask argument specifies whether I/O may be performed to release
2604  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2605  *
2606  */
2607 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2608 {
2609         struct address_space * const mapping = page->mapping;
2610
2611         BUG_ON(!PageLocked(page));
2612         if (PageWriteback(page))
2613                 return 0;
2614
2615         if (mapping && mapping->a_ops->releasepage)
2616                 return mapping->a_ops->releasepage(page, gfp_mask);
2617         return try_to_free_buffers(page);
2618 }
2619
2620 EXPORT_SYMBOL(try_to_release_page);