block: remove per-queue plugging
[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 void remove_from_page_cache(struct page *page)
141 {
142         struct address_space *mapping = page->mapping;
143         void (*freepage)(struct page *);
144
145         BUG_ON(!PageLocked(page));
146
147         freepage = mapping->a_ops->freepage;
148         spin_lock_irq(&mapping->tree_lock);
149         __remove_from_page_cache(page);
150         spin_unlock_irq(&mapping->tree_lock);
151         mem_cgroup_uncharge_cache_page(page);
152
153         if (freepage)
154                 freepage(page);
155 }
156 EXPORT_SYMBOL(remove_from_page_cache);
157
158 static int sleep_on_page(void *word)
159 {
160         io_schedule();
161         return 0;
162 }
163
164 static int sleep_on_page_killable(void *word)
165 {
166         sleep_on_page(word);
167         return fatal_signal_pending(current) ? -EINTR : 0;
168 }
169
170 /**
171  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
172  * @mapping:    address space structure to write
173  * @start:      offset in bytes where the range starts
174  * @end:        offset in bytes where the range ends (inclusive)
175  * @sync_mode:  enable synchronous operation
176  *
177  * Start writeback against all of a mapping's dirty pages that lie
178  * within the byte offsets <start, end> inclusive.
179  *
180  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
181  * opposed to a regular memory cleansing writeback.  The difference between
182  * these two operations is that if a dirty page/buffer is encountered, it must
183  * be waited upon, and not just skipped over.
184  */
185 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
186                                 loff_t end, int sync_mode)
187 {
188         int ret;
189         struct writeback_control wbc = {
190                 .sync_mode = sync_mode,
191                 .nr_to_write = LONG_MAX,
192                 .range_start = start,
193                 .range_end = end,
194         };
195
196         if (!mapping_cap_writeback_dirty(mapping))
197                 return 0;
198
199         ret = do_writepages(mapping, &wbc);
200         return ret;
201 }
202
203 static inline int __filemap_fdatawrite(struct address_space *mapping,
204         int sync_mode)
205 {
206         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
207 }
208
209 int filemap_fdatawrite(struct address_space *mapping)
210 {
211         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
212 }
213 EXPORT_SYMBOL(filemap_fdatawrite);
214
215 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
216                                 loff_t end)
217 {
218         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
219 }
220 EXPORT_SYMBOL(filemap_fdatawrite_range);
221
222 /**
223  * filemap_flush - mostly a non-blocking flush
224  * @mapping:    target address_space
225  *
226  * This is a mostly non-blocking flush.  Not suitable for data-integrity
227  * purposes - I/O may not be started against all dirty pages.
228  */
229 int filemap_flush(struct address_space *mapping)
230 {
231         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
232 }
233 EXPORT_SYMBOL(filemap_flush);
234
235 /**
236  * filemap_fdatawait_range - wait for writeback to complete
237  * @mapping:            address space structure to wait for
238  * @start_byte:         offset in bytes where the range starts
239  * @end_byte:           offset in bytes where the range ends (inclusive)
240  *
241  * Walk the list of under-writeback pages of the given address space
242  * in the given range and wait for all of them.
243  */
244 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
245                             loff_t end_byte)
246 {
247         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
248         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
249         struct pagevec pvec;
250         int nr_pages;
251         int ret = 0;
252
253         if (end_byte < start_byte)
254                 return 0;
255
256         pagevec_init(&pvec, 0);
257         while ((index <= end) &&
258                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
259                         PAGECACHE_TAG_WRITEBACK,
260                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
261                 unsigned i;
262
263                 for (i = 0; i < nr_pages; i++) {
264                         struct page *page = pvec.pages[i];
265
266                         /* until radix tree lookup accepts end_index */
267                         if (page->index > end)
268                                 continue;
269
270                         wait_on_page_writeback(page);
271                         if (TestClearPageError(page))
272                                 ret = -EIO;
273                 }
274                 pagevec_release(&pvec);
275                 cond_resched();
276         }
277
278         /* Check for outstanding write errors */
279         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
280                 ret = -ENOSPC;
281         if (test_and_clear_bit(AS_EIO, &mapping->flags))
282                 ret = -EIO;
283
284         return ret;
285 }
286 EXPORT_SYMBOL(filemap_fdatawait_range);
287
288 /**
289  * filemap_fdatawait - wait for all under-writeback pages to complete
290  * @mapping: address space structure to wait for
291  *
292  * Walk the list of under-writeback pages of the given address space
293  * and wait for all of them.
294  */
295 int filemap_fdatawait(struct address_space *mapping)
296 {
297         loff_t i_size = i_size_read(mapping->host);
298
299         if (i_size == 0)
300                 return 0;
301
302         return filemap_fdatawait_range(mapping, 0, i_size - 1);
303 }
304 EXPORT_SYMBOL(filemap_fdatawait);
305
306 int filemap_write_and_wait(struct address_space *mapping)
307 {
308         int err = 0;
309
310         if (mapping->nrpages) {
311                 err = filemap_fdatawrite(mapping);
312                 /*
313                  * Even if the above returned error, the pages may be
314                  * written partially (e.g. -ENOSPC), so we wait for it.
315                  * But the -EIO is special case, it may indicate the worst
316                  * thing (e.g. bug) happened, so we avoid waiting for it.
317                  */
318                 if (err != -EIO) {
319                         int err2 = filemap_fdatawait(mapping);
320                         if (!err)
321                                 err = err2;
322                 }
323         }
324         return err;
325 }
326 EXPORT_SYMBOL(filemap_write_and_wait);
327
328 /**
329  * filemap_write_and_wait_range - write out & wait on a file range
330  * @mapping:    the address_space for the pages
331  * @lstart:     offset in bytes where the range starts
332  * @lend:       offset in bytes where the range ends (inclusive)
333  *
334  * Write out and wait upon file offsets lstart->lend, inclusive.
335  *
336  * Note that `lend' is inclusive (describes the last byte to be written) so
337  * that this function can be used to write to the very end-of-file (end = -1).
338  */
339 int filemap_write_and_wait_range(struct address_space *mapping,
340                                  loff_t lstart, loff_t lend)
341 {
342         int err = 0;
343
344         if (mapping->nrpages) {
345                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
346                                                  WB_SYNC_ALL);
347                 /* See comment of filemap_write_and_wait() */
348                 if (err != -EIO) {
349                         int err2 = filemap_fdatawait_range(mapping,
350                                                 lstart, lend);
351                         if (!err)
352                                 err = err2;
353                 }
354         }
355         return err;
356 }
357 EXPORT_SYMBOL(filemap_write_and_wait_range);
358
359 /**
360  * add_to_page_cache_locked - add a locked page to the pagecache
361  * @page:       page to add
362  * @mapping:    the page's address_space
363  * @offset:     page index
364  * @gfp_mask:   page allocation mode
365  *
366  * This function is used to add a page to the pagecache. It must be locked.
367  * This function does not add the page to the LRU.  The caller must do that.
368  */
369 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
370                 pgoff_t offset, gfp_t gfp_mask)
371 {
372         int error;
373
374         VM_BUG_ON(!PageLocked(page));
375
376         error = mem_cgroup_cache_charge(page, current->mm,
377                                         gfp_mask & GFP_RECLAIM_MASK);
378         if (error)
379                 goto out;
380
381         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
382         if (error == 0) {
383                 page_cache_get(page);
384                 page->mapping = mapping;
385                 page->index = offset;
386
387                 spin_lock_irq(&mapping->tree_lock);
388                 error = radix_tree_insert(&mapping->page_tree, offset, page);
389                 if (likely(!error)) {
390                         mapping->nrpages++;
391                         __inc_zone_page_state(page, NR_FILE_PAGES);
392                         if (PageSwapBacked(page))
393                                 __inc_zone_page_state(page, NR_SHMEM);
394                         spin_unlock_irq(&mapping->tree_lock);
395                 } else {
396                         page->mapping = NULL;
397                         spin_unlock_irq(&mapping->tree_lock);
398                         mem_cgroup_uncharge_cache_page(page);
399                         page_cache_release(page);
400                 }
401                 radix_tree_preload_end();
402         } else
403                 mem_cgroup_uncharge_cache_page(page);
404 out:
405         return error;
406 }
407 EXPORT_SYMBOL(add_to_page_cache_locked);
408
409 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
410                                 pgoff_t offset, gfp_t gfp_mask)
411 {
412         int ret;
413
414         /*
415          * Splice_read and readahead add shmem/tmpfs pages into the page cache
416          * before shmem_readpage has a chance to mark them as SwapBacked: they
417          * need to go on the anon lru below, and mem_cgroup_cache_charge
418          * (called in add_to_page_cache) needs to know where they're going too.
419          */
420         if (mapping_cap_swap_backed(mapping))
421                 SetPageSwapBacked(page);
422
423         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
424         if (ret == 0) {
425                 if (page_is_file_cache(page))
426                         lru_cache_add_file(page);
427                 else
428                         lru_cache_add_anon(page);
429         }
430         return ret;
431 }
432 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
433
434 #ifdef CONFIG_NUMA
435 struct page *__page_cache_alloc(gfp_t gfp)
436 {
437         int n;
438         struct page *page;
439
440         if (cpuset_do_page_mem_spread()) {
441                 get_mems_allowed();
442                 n = cpuset_mem_spread_node();
443                 page = alloc_pages_exact_node(n, gfp, 0);
444                 put_mems_allowed();
445                 return page;
446         }
447         return alloc_pages(gfp, 0);
448 }
449 EXPORT_SYMBOL(__page_cache_alloc);
450 #endif
451
452 /*
453  * In order to wait for pages to become available there must be
454  * waitqueues associated with pages. By using a hash table of
455  * waitqueues where the bucket discipline is to maintain all
456  * waiters on the same queue and wake all when any of the pages
457  * become available, and for the woken contexts to check to be
458  * sure the appropriate page became available, this saves space
459  * at a cost of "thundering herd" phenomena during rare hash
460  * collisions.
461  */
462 static wait_queue_head_t *page_waitqueue(struct page *page)
463 {
464         const struct zone *zone = page_zone(page);
465
466         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
467 }
468
469 static inline void wake_up_page(struct page *page, int bit)
470 {
471         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
472 }
473
474 void wait_on_page_bit(struct page *page, int bit_nr)
475 {
476         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
477
478         if (test_bit(bit_nr, &page->flags))
479                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
480                                                         TASK_UNINTERRUPTIBLE);
481 }
482 EXPORT_SYMBOL(wait_on_page_bit);
483
484 /**
485  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
486  * @page: Page defining the wait queue of interest
487  * @waiter: Waiter to add to the queue
488  *
489  * Add an arbitrary @waiter to the wait queue for the nominated @page.
490  */
491 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
492 {
493         wait_queue_head_t *q = page_waitqueue(page);
494         unsigned long flags;
495
496         spin_lock_irqsave(&q->lock, flags);
497         __add_wait_queue(q, waiter);
498         spin_unlock_irqrestore(&q->lock, flags);
499 }
500 EXPORT_SYMBOL_GPL(add_page_wait_queue);
501
502 /**
503  * unlock_page - unlock a locked page
504  * @page: the page
505  *
506  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
507  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
508  * mechananism between PageLocked pages and PageWriteback pages is shared.
509  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
510  *
511  * The mb is necessary to enforce ordering between the clear_bit and the read
512  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
513  */
514 void unlock_page(struct page *page)
515 {
516         VM_BUG_ON(!PageLocked(page));
517         clear_bit_unlock(PG_locked, &page->flags);
518         smp_mb__after_clear_bit();
519         wake_up_page(page, PG_locked);
520 }
521 EXPORT_SYMBOL(unlock_page);
522
523 /**
524  * end_page_writeback - end writeback against a page
525  * @page: the page
526  */
527 void end_page_writeback(struct page *page)
528 {
529         if (TestClearPageReclaim(page))
530                 rotate_reclaimable_page(page);
531
532         if (!test_clear_page_writeback(page))
533                 BUG();
534
535         smp_mb__after_clear_bit();
536         wake_up_page(page, PG_writeback);
537 }
538 EXPORT_SYMBOL(end_page_writeback);
539
540 /**
541  * __lock_page - get a lock on the page, assuming we need to sleep to get it
542  * @page: the page to lock
543  */
544 void __lock_page(struct page *page)
545 {
546         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
547
548         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
549                                                         TASK_UNINTERRUPTIBLE);
550 }
551 EXPORT_SYMBOL(__lock_page);
552
553 int __lock_page_killable(struct page *page)
554 {
555         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
556
557         return __wait_on_bit_lock(page_waitqueue(page), &wait,
558                                         sleep_on_page_killable, TASK_KILLABLE);
559 }
560 EXPORT_SYMBOL_GPL(__lock_page_killable);
561
562 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
563                          unsigned int flags)
564 {
565         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
566                 __lock_page(page);
567                 return 1;
568         } else {
569                 up_read(&mm->mmap_sem);
570                 wait_on_page_locked(page);
571                 return 0;
572         }
573 }
574
575 /**
576  * find_get_page - find and get a page reference
577  * @mapping: the address_space to search
578  * @offset: the page index
579  *
580  * Is there a pagecache struct page at the given (mapping, offset) tuple?
581  * If yes, increment its refcount and return it; if no, return NULL.
582  */
583 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
584 {
585         void **pagep;
586         struct page *page;
587
588         rcu_read_lock();
589 repeat:
590         page = NULL;
591         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
592         if (pagep) {
593                 page = radix_tree_deref_slot(pagep);
594                 if (unlikely(!page))
595                         goto out;
596                 if (radix_tree_deref_retry(page))
597                         goto repeat;
598
599                 if (!page_cache_get_speculative(page))
600                         goto repeat;
601
602                 /*
603                  * Has the page moved?
604                  * This is part of the lockless pagecache protocol. See
605                  * include/linux/pagemap.h for details.
606                  */
607                 if (unlikely(page != *pagep)) {
608                         page_cache_release(page);
609                         goto repeat;
610                 }
611         }
612 out:
613         rcu_read_unlock();
614
615         return page;
616 }
617 EXPORT_SYMBOL(find_get_page);
618
619 /**
620  * find_lock_page - locate, pin and lock a pagecache page
621  * @mapping: the address_space to search
622  * @offset: the page index
623  *
624  * Locates the desired pagecache page, locks it, increments its reference
625  * count and returns its address.
626  *
627  * Returns zero if the page was not present. find_lock_page() may sleep.
628  */
629 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
630 {
631         struct page *page;
632
633 repeat:
634         page = find_get_page(mapping, offset);
635         if (page) {
636                 lock_page(page);
637                 /* Has the page been truncated? */
638                 if (unlikely(page->mapping != mapping)) {
639                         unlock_page(page);
640                         page_cache_release(page);
641                         goto repeat;
642                 }
643                 VM_BUG_ON(page->index != offset);
644         }
645         return page;
646 }
647 EXPORT_SYMBOL(find_lock_page);
648
649 /**
650  * find_or_create_page - locate or add a pagecache page
651  * @mapping: the page's address_space
652  * @index: the page's index into the mapping
653  * @gfp_mask: page allocation mode
654  *
655  * Locates a page in the pagecache.  If the page is not present, a new page
656  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
657  * LRU list.  The returned page is locked and has its reference count
658  * incremented.
659  *
660  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
661  * allocation!
662  *
663  * find_or_create_page() returns the desired page's address, or zero on
664  * memory exhaustion.
665  */
666 struct page *find_or_create_page(struct address_space *mapping,
667                 pgoff_t index, gfp_t gfp_mask)
668 {
669         struct page *page;
670         int err;
671 repeat:
672         page = find_lock_page(mapping, index);
673         if (!page) {
674                 page = __page_cache_alloc(gfp_mask);
675                 if (!page)
676                         return NULL;
677                 /*
678                  * We want a regular kernel memory (not highmem or DMA etc)
679                  * allocation for the radix tree nodes, but we need to honour
680                  * the context-specific requirements the caller has asked for.
681                  * GFP_RECLAIM_MASK collects those requirements.
682                  */
683                 err = add_to_page_cache_lru(page, mapping, index,
684                         (gfp_mask & GFP_RECLAIM_MASK));
685                 if (unlikely(err)) {
686                         page_cache_release(page);
687                         page = NULL;
688                         if (err == -EEXIST)
689                                 goto repeat;
690                 }
691         }
692         return page;
693 }
694 EXPORT_SYMBOL(find_or_create_page);
695
696 /**
697  * find_get_pages - gang pagecache lookup
698  * @mapping:    The address_space to search
699  * @start:      The starting page index
700  * @nr_pages:   The maximum number of pages
701  * @pages:      Where the resulting pages are placed
702  *
703  * find_get_pages() will search for and return a group of up to
704  * @nr_pages pages in the mapping.  The pages are placed at @pages.
705  * find_get_pages() takes a reference against the returned pages.
706  *
707  * The search returns a group of mapping-contiguous pages with ascending
708  * indexes.  There may be holes in the indices due to not-present pages.
709  *
710  * find_get_pages() returns the number of pages which were found.
711  */
712 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
713                             unsigned int nr_pages, struct page **pages)
714 {
715         unsigned int i;
716         unsigned int ret;
717         unsigned int nr_found;
718
719         rcu_read_lock();
720 restart:
721         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
722                                 (void ***)pages, start, nr_pages);
723         ret = 0;
724         for (i = 0; i < nr_found; i++) {
725                 struct page *page;
726 repeat:
727                 page = radix_tree_deref_slot((void **)pages[i]);
728                 if (unlikely(!page))
729                         continue;
730                 if (radix_tree_deref_retry(page)) {
731                         if (ret)
732                                 start = pages[ret-1]->index;
733                         goto restart;
734                 }
735
736                 if (!page_cache_get_speculative(page))
737                         goto repeat;
738
739                 /* Has the page moved? */
740                 if (unlikely(page != *((void **)pages[i]))) {
741                         page_cache_release(page);
742                         goto repeat;
743                 }
744
745                 pages[ret] = page;
746                 ret++;
747         }
748         rcu_read_unlock();
749         return ret;
750 }
751
752 /**
753  * find_get_pages_contig - gang contiguous pagecache lookup
754  * @mapping:    The address_space to search
755  * @index:      The starting page index
756  * @nr_pages:   The maximum number of pages
757  * @pages:      Where the resulting pages are placed
758  *
759  * find_get_pages_contig() works exactly like find_get_pages(), except
760  * that the returned number of pages are guaranteed to be contiguous.
761  *
762  * find_get_pages_contig() returns the number of pages which were found.
763  */
764 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
765                                unsigned int nr_pages, struct page **pages)
766 {
767         unsigned int i;
768         unsigned int ret;
769         unsigned int nr_found;
770
771         rcu_read_lock();
772 restart:
773         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
774                                 (void ***)pages, index, nr_pages);
775         ret = 0;
776         for (i = 0; i < nr_found; i++) {
777                 struct page *page;
778 repeat:
779                 page = radix_tree_deref_slot((void **)pages[i]);
780                 if (unlikely(!page))
781                         continue;
782                 if (radix_tree_deref_retry(page))
783                         goto restart;
784
785                 if (!page_cache_get_speculative(page))
786                         goto repeat;
787
788                 /* Has the page moved? */
789                 if (unlikely(page != *((void **)pages[i]))) {
790                         page_cache_release(page);
791                         goto repeat;
792                 }
793
794                 /*
795                  * must check mapping and index after taking the ref.
796                  * otherwise we can get both false positives and false
797                  * negatives, which is just confusing to the caller.
798                  */
799                 if (page->mapping == NULL || page->index != index) {
800                         page_cache_release(page);
801                         break;
802                 }
803
804                 pages[ret] = page;
805                 ret++;
806                 index++;
807         }
808         rcu_read_unlock();
809         return ret;
810 }
811 EXPORT_SYMBOL(find_get_pages_contig);
812
813 /**
814  * find_get_pages_tag - find and return pages that match @tag
815  * @mapping:    the address_space to search
816  * @index:      the starting page index
817  * @tag:        the tag index
818  * @nr_pages:   the maximum number of pages
819  * @pages:      where the resulting pages are placed
820  *
821  * Like find_get_pages, except we only return pages which are tagged with
822  * @tag.   We update @index to index the next page for the traversal.
823  */
824 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
825                         int tag, unsigned int nr_pages, struct page **pages)
826 {
827         unsigned int i;
828         unsigned int ret;
829         unsigned int nr_found;
830
831         rcu_read_lock();
832 restart:
833         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
834                                 (void ***)pages, *index, nr_pages, tag);
835         ret = 0;
836         for (i = 0; i < nr_found; i++) {
837                 struct page *page;
838 repeat:
839                 page = radix_tree_deref_slot((void **)pages[i]);
840                 if (unlikely(!page))
841                         continue;
842                 if (radix_tree_deref_retry(page))
843                         goto restart;
844
845                 if (!page_cache_get_speculative(page))
846                         goto repeat;
847
848                 /* Has the page moved? */
849                 if (unlikely(page != *((void **)pages[i]))) {
850                         page_cache_release(page);
851                         goto repeat;
852                 }
853
854                 pages[ret] = page;
855                 ret++;
856         }
857         rcu_read_unlock();
858
859         if (ret)
860                 *index = pages[ret - 1]->index + 1;
861
862         return ret;
863 }
864 EXPORT_SYMBOL(find_get_pages_tag);
865
866 /**
867  * grab_cache_page_nowait - returns locked page at given index in given cache
868  * @mapping: target address_space
869  * @index: the page index
870  *
871  * Same as grab_cache_page(), but do not wait if the page is unavailable.
872  * This is intended for speculative data generators, where the data can
873  * be regenerated if the page couldn't be grabbed.  This routine should
874  * be safe to call while holding the lock for another page.
875  *
876  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
877  * and deadlock against the caller's locked page.
878  */
879 struct page *
880 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
881 {
882         struct page *page = find_get_page(mapping, index);
883
884         if (page) {
885                 if (trylock_page(page))
886                         return page;
887                 page_cache_release(page);
888                 return NULL;
889         }
890         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
891         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
892                 page_cache_release(page);
893                 page = NULL;
894         }
895         return page;
896 }
897 EXPORT_SYMBOL(grab_cache_page_nowait);
898
899 /*
900  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
901  * a _large_ part of the i/o request. Imagine the worst scenario:
902  *
903  *      ---R__________________________________________B__________
904  *         ^ reading here                             ^ bad block(assume 4k)
905  *
906  * read(R) => miss => readahead(R...B) => media error => frustrating retries
907  * => failing the whole request => read(R) => read(R+1) =>
908  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
909  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
910  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
911  *
912  * It is going insane. Fix it by quickly scaling down the readahead size.
913  */
914 static void shrink_readahead_size_eio(struct file *filp,
915                                         struct file_ra_state *ra)
916 {
917         ra->ra_pages /= 4;
918 }
919
920 /**
921  * do_generic_file_read - generic file read routine
922  * @filp:       the file to read
923  * @ppos:       current file position
924  * @desc:       read_descriptor
925  * @actor:      read method
926  *
927  * This is a generic file read routine, and uses the
928  * mapping->a_ops->readpage() function for the actual low-level stuff.
929  *
930  * This is really ugly. But the goto's actually try to clarify some
931  * of the logic when it comes to error handling etc.
932  */
933 static void do_generic_file_read(struct file *filp, loff_t *ppos,
934                 read_descriptor_t *desc, read_actor_t actor)
935 {
936         struct address_space *mapping = filp->f_mapping;
937         struct inode *inode = mapping->host;
938         struct file_ra_state *ra = &filp->f_ra;
939         pgoff_t index;
940         pgoff_t last_index;
941         pgoff_t prev_index;
942         unsigned long offset;      /* offset into pagecache page */
943         unsigned int prev_offset;
944         int error;
945
946         index = *ppos >> PAGE_CACHE_SHIFT;
947         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
948         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
949         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
950         offset = *ppos & ~PAGE_CACHE_MASK;
951
952         for (;;) {
953                 struct page *page;
954                 pgoff_t end_index;
955                 loff_t isize;
956                 unsigned long nr, ret;
957
958                 cond_resched();
959 find_page:
960                 page = find_get_page(mapping, index);
961                 if (!page) {
962                         page_cache_sync_readahead(mapping,
963                                         ra, filp,
964                                         index, last_index - index);
965                         page = find_get_page(mapping, index);
966                         if (unlikely(page == NULL))
967                                 goto no_cached_page;
968                 }
969                 if (PageReadahead(page)) {
970                         page_cache_async_readahead(mapping,
971                                         ra, filp, page,
972                                         index, last_index - index);
973                 }
974                 if (!PageUptodate(page)) {
975                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
976                                         !mapping->a_ops->is_partially_uptodate)
977                                 goto page_not_up_to_date;
978                         if (!trylock_page(page))
979                                 goto page_not_up_to_date;
980                         /* Did it get truncated before we got the lock? */
981                         if (!page->mapping)
982                                 goto page_not_up_to_date_locked;
983                         if (!mapping->a_ops->is_partially_uptodate(page,
984                                                                 desc, offset))
985                                 goto page_not_up_to_date_locked;
986                         unlock_page(page);
987                 }
988 page_ok:
989                 /*
990                  * i_size must be checked after we know the page is Uptodate.
991                  *
992                  * Checking i_size after the check allows us to calculate
993                  * the correct value for "nr", which means the zero-filled
994                  * part of the page is not copied back to userspace (unless
995                  * another truncate extends the file - this is desired though).
996                  */
997
998                 isize = i_size_read(inode);
999                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1000                 if (unlikely(!isize || index > end_index)) {
1001                         page_cache_release(page);
1002                         goto out;
1003                 }
1004
1005                 /* nr is the maximum number of bytes to copy from this page */
1006                 nr = PAGE_CACHE_SIZE;
1007                 if (index == end_index) {
1008                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1009                         if (nr <= offset) {
1010                                 page_cache_release(page);
1011                                 goto out;
1012                         }
1013                 }
1014                 nr = nr - offset;
1015
1016                 /* If users can be writing to this page using arbitrary
1017                  * virtual addresses, take care about potential aliasing
1018                  * before reading the page on the kernel side.
1019                  */
1020                 if (mapping_writably_mapped(mapping))
1021                         flush_dcache_page(page);
1022
1023                 /*
1024                  * When a sequential read accesses a page several times,
1025                  * only mark it as accessed the first time.
1026                  */
1027                 if (prev_index != index || offset != prev_offset)
1028                         mark_page_accessed(page);
1029                 prev_index = index;
1030
1031                 /*
1032                  * Ok, we have the page, and it's up-to-date, so
1033                  * now we can copy it to user space...
1034                  *
1035                  * The actor routine returns how many bytes were actually used..
1036                  * NOTE! This may not be the same as how much of a user buffer
1037                  * we filled up (we may be padding etc), so we can only update
1038                  * "pos" here (the actor routine has to update the user buffer
1039                  * pointers and the remaining count).
1040                  */
1041                 ret = actor(desc, page, offset, nr);
1042                 offset += ret;
1043                 index += offset >> PAGE_CACHE_SHIFT;
1044                 offset &= ~PAGE_CACHE_MASK;
1045                 prev_offset = offset;
1046
1047                 page_cache_release(page);
1048                 if (ret == nr && desc->count)
1049                         continue;
1050                 goto out;
1051
1052 page_not_up_to_date:
1053                 /* Get exclusive access to the page ... */
1054                 error = lock_page_killable(page);
1055                 if (unlikely(error))
1056                         goto readpage_error;
1057
1058 page_not_up_to_date_locked:
1059                 /* Did it get truncated before we got the lock? */
1060                 if (!page->mapping) {
1061                         unlock_page(page);
1062                         page_cache_release(page);
1063                         continue;
1064                 }
1065
1066                 /* Did somebody else fill it already? */
1067                 if (PageUptodate(page)) {
1068                         unlock_page(page);
1069                         goto page_ok;
1070                 }
1071
1072 readpage:
1073                 /*
1074                  * A previous I/O error may have been due to temporary
1075                  * failures, eg. multipath errors.
1076                  * PG_error will be set again if readpage fails.
1077                  */
1078                 ClearPageError(page);
1079                 /* Start the actual read. The read will unlock the page. */
1080                 error = mapping->a_ops->readpage(filp, page);
1081
1082                 if (unlikely(error)) {
1083                         if (error == AOP_TRUNCATED_PAGE) {
1084                                 page_cache_release(page);
1085                                 goto find_page;
1086                         }
1087                         goto readpage_error;
1088                 }
1089
1090                 if (!PageUptodate(page)) {
1091                         error = lock_page_killable(page);
1092                         if (unlikely(error))
1093                                 goto readpage_error;
1094                         if (!PageUptodate(page)) {
1095                                 if (page->mapping == NULL) {
1096                                         /*
1097                                          * invalidate_mapping_pages got it
1098                                          */
1099                                         unlock_page(page);
1100                                         page_cache_release(page);
1101                                         goto find_page;
1102                                 }
1103                                 unlock_page(page);
1104                                 shrink_readahead_size_eio(filp, ra);
1105                                 error = -EIO;
1106                                 goto readpage_error;
1107                         }
1108                         unlock_page(page);
1109                 }
1110
1111                 goto page_ok;
1112
1113 readpage_error:
1114                 /* UHHUH! A synchronous read error occurred. Report it */
1115                 desc->error = error;
1116                 page_cache_release(page);
1117                 goto out;
1118
1119 no_cached_page:
1120                 /*
1121                  * Ok, it wasn't cached, so we need to create a new
1122                  * page..
1123                  */
1124                 page = page_cache_alloc_cold(mapping);
1125                 if (!page) {
1126                         desc->error = -ENOMEM;
1127                         goto out;
1128                 }
1129                 error = add_to_page_cache_lru(page, mapping,
1130                                                 index, GFP_KERNEL);
1131                 if (error) {
1132                         page_cache_release(page);
1133                         if (error == -EEXIST)
1134                                 goto find_page;
1135                         desc->error = error;
1136                         goto out;
1137                 }
1138                 goto readpage;
1139         }
1140
1141 out:
1142         ra->prev_pos = prev_index;
1143         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1144         ra->prev_pos |= prev_offset;
1145
1146         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1147         file_accessed(filp);
1148 }
1149
1150 int file_read_actor(read_descriptor_t *desc, struct page *page,
1151                         unsigned long offset, unsigned long size)
1152 {
1153         char *kaddr;
1154         unsigned long left, count = desc->count;
1155
1156         if (size > count)
1157                 size = count;
1158
1159         /*
1160          * Faults on the destination of a read are common, so do it before
1161          * taking the kmap.
1162          */
1163         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1164                 kaddr = kmap_atomic(page, KM_USER0);
1165                 left = __copy_to_user_inatomic(desc->arg.buf,
1166                                                 kaddr + offset, size);
1167                 kunmap_atomic(kaddr, KM_USER0);
1168                 if (left == 0)
1169                         goto success;
1170         }
1171
1172         /* Do it the slow way */
1173         kaddr = kmap(page);
1174         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1175         kunmap(page);
1176
1177         if (left) {
1178                 size -= left;
1179                 desc->error = -EFAULT;
1180         }
1181 success:
1182         desc->count = count - size;
1183         desc->written += size;
1184         desc->arg.buf += size;
1185         return size;
1186 }
1187
1188 /*
1189  * Performs necessary checks before doing a write
1190  * @iov:        io vector request
1191  * @nr_segs:    number of segments in the iovec
1192  * @count:      number of bytes to write
1193  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1194  *
1195  * Adjust number of segments and amount of bytes to write (nr_segs should be
1196  * properly initialized first). Returns appropriate error code that caller
1197  * should return or zero in case that write should be allowed.
1198  */
1199 int generic_segment_checks(const struct iovec *iov,
1200                         unsigned long *nr_segs, size_t *count, int access_flags)
1201 {
1202         unsigned long   seg;
1203         size_t cnt = 0;
1204         for (seg = 0; seg < *nr_segs; seg++) {
1205                 const struct iovec *iv = &iov[seg];
1206
1207                 /*
1208                  * If any segment has a negative length, or the cumulative
1209                  * length ever wraps negative then return -EINVAL.
1210                  */
1211                 cnt += iv->iov_len;
1212                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1213                         return -EINVAL;
1214                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1215                         continue;
1216                 if (seg == 0)
1217                         return -EFAULT;
1218                 *nr_segs = seg;
1219                 cnt -= iv->iov_len;     /* This segment is no good */
1220                 break;
1221         }
1222         *count = cnt;
1223         return 0;
1224 }
1225 EXPORT_SYMBOL(generic_segment_checks);
1226
1227 /**
1228  * generic_file_aio_read - generic filesystem read routine
1229  * @iocb:       kernel I/O control block
1230  * @iov:        io vector request
1231  * @nr_segs:    number of segments in the iovec
1232  * @pos:        current file position
1233  *
1234  * This is the "read()" routine for all filesystems
1235  * that can use the page cache directly.
1236  */
1237 ssize_t
1238 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1239                 unsigned long nr_segs, loff_t pos)
1240 {
1241         struct file *filp = iocb->ki_filp;
1242         ssize_t retval;
1243         unsigned long seg = 0;
1244         size_t count;
1245         loff_t *ppos = &iocb->ki_pos;
1246
1247         count = 0;
1248         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1249         if (retval)
1250                 return retval;
1251
1252         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1253         if (filp->f_flags & O_DIRECT) {
1254                 loff_t size;
1255                 struct address_space *mapping;
1256                 struct inode *inode;
1257
1258                 mapping = filp->f_mapping;
1259                 inode = mapping->host;
1260                 if (!count)
1261                         goto out; /* skip atime */
1262                 size = i_size_read(inode);
1263                 if (pos < size) {
1264                         retval = filemap_write_and_wait_range(mapping, pos,
1265                                         pos + iov_length(iov, nr_segs) - 1);
1266                         if (!retval) {
1267                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1268                                                         iov, pos, nr_segs);
1269                         }
1270                         if (retval > 0) {
1271                                 *ppos = pos + retval;
1272                                 count -= retval;
1273                         }
1274
1275                         /*
1276                          * Btrfs can have a short DIO read if we encounter
1277                          * compressed extents, so if there was an error, or if
1278                          * we've already read everything we wanted to, or if
1279                          * there was a short read because we hit EOF, go ahead
1280                          * and return.  Otherwise fallthrough to buffered io for
1281                          * the rest of the read.
1282                          */
1283                         if (retval < 0 || !count || *ppos >= size) {
1284                                 file_accessed(filp);
1285                                 goto out;
1286                         }
1287                 }
1288         }
1289
1290         count = retval;
1291         for (seg = 0; seg < nr_segs; seg++) {
1292                 read_descriptor_t desc;
1293                 loff_t offset = 0;
1294
1295                 /*
1296                  * If we did a short DIO read we need to skip the section of the
1297                  * iov that we've already read data into.
1298                  */
1299                 if (count) {
1300                         if (count > iov[seg].iov_len) {
1301                                 count -= iov[seg].iov_len;
1302                                 continue;
1303                         }
1304                         offset = count;
1305                         count = 0;
1306                 }
1307
1308                 desc.written = 0;
1309                 desc.arg.buf = iov[seg].iov_base + offset;
1310                 desc.count = iov[seg].iov_len - offset;
1311                 if (desc.count == 0)
1312                         continue;
1313                 desc.error = 0;
1314                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1315                 retval += desc.written;
1316                 if (desc.error) {
1317                         retval = retval ?: desc.error;
1318                         break;
1319                 }
1320                 if (desc.count > 0)
1321                         break;
1322         }
1323 out:
1324         return retval;
1325 }
1326 EXPORT_SYMBOL(generic_file_aio_read);
1327
1328 static ssize_t
1329 do_readahead(struct address_space *mapping, struct file *filp,
1330              pgoff_t index, unsigned long nr)
1331 {
1332         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1333                 return -EINVAL;
1334
1335         force_page_cache_readahead(mapping, filp, index, nr);
1336         return 0;
1337 }
1338
1339 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1340 {
1341         ssize_t ret;
1342         struct file *file;
1343
1344         ret = -EBADF;
1345         file = fget(fd);
1346         if (file) {
1347                 if (file->f_mode & FMODE_READ) {
1348                         struct address_space *mapping = file->f_mapping;
1349                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1350                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1351                         unsigned long len = end - start + 1;
1352                         ret = do_readahead(mapping, file, start, len);
1353                 }
1354                 fput(file);
1355         }
1356         return ret;
1357 }
1358 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1359 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1360 {
1361         return SYSC_readahead((int) fd, offset, (size_t) count);
1362 }
1363 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1364 #endif
1365
1366 #ifdef CONFIG_MMU
1367 /**
1368  * page_cache_read - adds requested page to the page cache if not already there
1369  * @file:       file to read
1370  * @offset:     page index
1371  *
1372  * This adds the requested page to the page cache if it isn't already there,
1373  * and schedules an I/O to read in its contents from disk.
1374  */
1375 static int page_cache_read(struct file *file, pgoff_t offset)
1376 {
1377         struct address_space *mapping = file->f_mapping;
1378         struct page *page; 
1379         int ret;
1380
1381         do {
1382                 page = page_cache_alloc_cold(mapping);
1383                 if (!page)
1384                         return -ENOMEM;
1385
1386                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1387                 if (ret == 0)
1388                         ret = mapping->a_ops->readpage(file, page);
1389                 else if (ret == -EEXIST)
1390                         ret = 0; /* losing race to add is OK */
1391
1392                 page_cache_release(page);
1393
1394         } while (ret == AOP_TRUNCATED_PAGE);
1395                 
1396         return ret;
1397 }
1398
1399 #define MMAP_LOTSAMISS  (100)
1400
1401 /*
1402  * Synchronous readahead happens when we don't even find
1403  * a page in the page cache at all.
1404  */
1405 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1406                                    struct file_ra_state *ra,
1407                                    struct file *file,
1408                                    pgoff_t offset)
1409 {
1410         unsigned long ra_pages;
1411         struct address_space *mapping = file->f_mapping;
1412
1413         /* If we don't want any read-ahead, don't bother */
1414         if (VM_RandomReadHint(vma))
1415                 return;
1416
1417         if (VM_SequentialReadHint(vma) ||
1418                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1419                 page_cache_sync_readahead(mapping, ra, file, offset,
1420                                           ra->ra_pages);
1421                 return;
1422         }
1423
1424         if (ra->mmap_miss < INT_MAX)
1425                 ra->mmap_miss++;
1426
1427         /*
1428          * Do we miss much more than hit in this file? If so,
1429          * stop bothering with read-ahead. It will only hurt.
1430          */
1431         if (ra->mmap_miss > MMAP_LOTSAMISS)
1432                 return;
1433
1434         /*
1435          * mmap read-around
1436          */
1437         ra_pages = max_sane_readahead(ra->ra_pages);
1438         if (ra_pages) {
1439                 ra->start = max_t(long, 0, offset - ra_pages/2);
1440                 ra->size = ra_pages;
1441                 ra->async_size = 0;
1442                 ra_submit(ra, mapping, file);
1443         }
1444 }
1445
1446 /*
1447  * Asynchronous readahead happens when we find the page and PG_readahead,
1448  * so we want to possibly extend the readahead further..
1449  */
1450 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1451                                     struct file_ra_state *ra,
1452                                     struct file *file,
1453                                     struct page *page,
1454                                     pgoff_t offset)
1455 {
1456         struct address_space *mapping = file->f_mapping;
1457
1458         /* If we don't want any read-ahead, don't bother */
1459         if (VM_RandomReadHint(vma))
1460                 return;
1461         if (ra->mmap_miss > 0)
1462                 ra->mmap_miss--;
1463         if (PageReadahead(page))
1464                 page_cache_async_readahead(mapping, ra, file,
1465                                            page, offset, ra->ra_pages);
1466 }
1467
1468 /**
1469  * filemap_fault - read in file data for page fault handling
1470  * @vma:        vma in which the fault was taken
1471  * @vmf:        struct vm_fault containing details of the fault
1472  *
1473  * filemap_fault() is invoked via the vma operations vector for a
1474  * mapped memory region to read in file data during a page fault.
1475  *
1476  * The goto's are kind of ugly, but this streamlines the normal case of having
1477  * it in the page cache, and handles the special cases reasonably without
1478  * having a lot of duplicated code.
1479  */
1480 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1481 {
1482         int error;
1483         struct file *file = vma->vm_file;
1484         struct address_space *mapping = file->f_mapping;
1485         struct file_ra_state *ra = &file->f_ra;
1486         struct inode *inode = mapping->host;
1487         pgoff_t offset = vmf->pgoff;
1488         struct page *page;
1489         pgoff_t size;
1490         int ret = 0;
1491
1492         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1493         if (offset >= size)
1494                 return VM_FAULT_SIGBUS;
1495
1496         /*
1497          * Do we have something in the page cache already?
1498          */
1499         page = find_get_page(mapping, offset);
1500         if (likely(page)) {
1501                 /*
1502                  * We found the page, so try async readahead before
1503                  * waiting for the lock.
1504                  */
1505                 do_async_mmap_readahead(vma, ra, file, page, offset);
1506         } else {
1507                 /* No page in the page cache at all */
1508                 do_sync_mmap_readahead(vma, ra, file, offset);
1509                 count_vm_event(PGMAJFAULT);
1510                 ret = VM_FAULT_MAJOR;
1511 retry_find:
1512                 page = find_get_page(mapping, offset);
1513                 if (!page)
1514                         goto no_cached_page;
1515         }
1516
1517         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1518                 page_cache_release(page);
1519                 return ret | VM_FAULT_RETRY;
1520         }
1521
1522         /* Did it get truncated? */
1523         if (unlikely(page->mapping != mapping)) {
1524                 unlock_page(page);
1525                 put_page(page);
1526                 goto retry_find;
1527         }
1528         VM_BUG_ON(page->index != offset);
1529
1530         /*
1531          * We have a locked page in the page cache, now we need to check
1532          * that it's up-to-date. If not, it is going to be due to an error.
1533          */
1534         if (unlikely(!PageUptodate(page)))
1535                 goto page_not_uptodate;
1536
1537         /*
1538          * Found the page and have a reference on it.
1539          * We must recheck i_size under page lock.
1540          */
1541         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1542         if (unlikely(offset >= size)) {
1543                 unlock_page(page);
1544                 page_cache_release(page);
1545                 return VM_FAULT_SIGBUS;
1546         }
1547
1548         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1549         vmf->page = page;
1550         return ret | VM_FAULT_LOCKED;
1551
1552 no_cached_page:
1553         /*
1554          * We're only likely to ever get here if MADV_RANDOM is in
1555          * effect.
1556          */
1557         error = page_cache_read(file, offset);
1558
1559         /*
1560          * The page we want has now been added to the page cache.
1561          * In the unlikely event that someone removed it in the
1562          * meantime, we'll just come back here and read it again.
1563          */
1564         if (error >= 0)
1565                 goto retry_find;
1566
1567         /*
1568          * An error return from page_cache_read can result if the
1569          * system is low on memory, or a problem occurs while trying
1570          * to schedule I/O.
1571          */
1572         if (error == -ENOMEM)
1573                 return VM_FAULT_OOM;
1574         return VM_FAULT_SIGBUS;
1575
1576 page_not_uptodate:
1577         /*
1578          * Umm, take care of errors if the page isn't up-to-date.
1579          * Try to re-read it _once_. We do this synchronously,
1580          * because there really aren't any performance issues here
1581          * and we need to check for errors.
1582          */
1583         ClearPageError(page);
1584         error = mapping->a_ops->readpage(file, page);
1585         if (!error) {
1586                 wait_on_page_locked(page);
1587                 if (!PageUptodate(page))
1588                         error = -EIO;
1589         }
1590         page_cache_release(page);
1591
1592         if (!error || error == AOP_TRUNCATED_PAGE)
1593                 goto retry_find;
1594
1595         /* Things didn't work out. Return zero to tell the mm layer so. */
1596         shrink_readahead_size_eio(file, ra);
1597         return VM_FAULT_SIGBUS;
1598 }
1599 EXPORT_SYMBOL(filemap_fault);
1600
1601 const struct vm_operations_struct generic_file_vm_ops = {
1602         .fault          = filemap_fault,
1603 };
1604
1605 /* This is used for a general mmap of a disk file */
1606
1607 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1608 {
1609         struct address_space *mapping = file->f_mapping;
1610
1611         if (!mapping->a_ops->readpage)
1612                 return -ENOEXEC;
1613         file_accessed(file);
1614         vma->vm_ops = &generic_file_vm_ops;
1615         vma->vm_flags |= VM_CAN_NONLINEAR;
1616         return 0;
1617 }
1618
1619 /*
1620  * This is for filesystems which do not implement ->writepage.
1621  */
1622 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1623 {
1624         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1625                 return -EINVAL;
1626         return generic_file_mmap(file, vma);
1627 }
1628 #else
1629 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1630 {
1631         return -ENOSYS;
1632 }
1633 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1634 {
1635         return -ENOSYS;
1636 }
1637 #endif /* CONFIG_MMU */
1638
1639 EXPORT_SYMBOL(generic_file_mmap);
1640 EXPORT_SYMBOL(generic_file_readonly_mmap);
1641
1642 static struct page *__read_cache_page(struct address_space *mapping,
1643                                 pgoff_t index,
1644                                 int (*filler)(void *,struct page*),
1645                                 void *data,
1646                                 gfp_t gfp)
1647 {
1648         struct page *page;
1649         int err;
1650 repeat:
1651         page = find_get_page(mapping, index);
1652         if (!page) {
1653                 page = __page_cache_alloc(gfp | __GFP_COLD);
1654                 if (!page)
1655                         return ERR_PTR(-ENOMEM);
1656                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1657                 if (unlikely(err)) {
1658                         page_cache_release(page);
1659                         if (err == -EEXIST)
1660                                 goto repeat;
1661                         /* Presumably ENOMEM for radix tree node */
1662                         return ERR_PTR(err);
1663                 }
1664                 err = filler(data, page);
1665                 if (err < 0) {
1666                         page_cache_release(page);
1667                         page = ERR_PTR(err);
1668                 }
1669         }
1670         return page;
1671 }
1672
1673 static struct page *do_read_cache_page(struct address_space *mapping,
1674                                 pgoff_t index,
1675                                 int (*filler)(void *,struct page*),
1676                                 void *data,
1677                                 gfp_t gfp)
1678
1679 {
1680         struct page *page;
1681         int err;
1682
1683 retry:
1684         page = __read_cache_page(mapping, index, filler, data, gfp);
1685         if (IS_ERR(page))
1686                 return page;
1687         if (PageUptodate(page))
1688                 goto out;
1689
1690         lock_page(page);
1691         if (!page->mapping) {
1692                 unlock_page(page);
1693                 page_cache_release(page);
1694                 goto retry;
1695         }
1696         if (PageUptodate(page)) {
1697                 unlock_page(page);
1698                 goto out;
1699         }
1700         err = filler(data, page);
1701         if (err < 0) {
1702                 page_cache_release(page);
1703                 return ERR_PTR(err);
1704         }
1705 out:
1706         mark_page_accessed(page);
1707         return page;
1708 }
1709
1710 /**
1711  * read_cache_page_async - read into page cache, fill it if needed
1712  * @mapping:    the page's address_space
1713  * @index:      the page index
1714  * @filler:     function to perform the read
1715  * @data:       destination for read data
1716  *
1717  * Same as read_cache_page, but don't wait for page to become unlocked
1718  * after submitting it to the filler.
1719  *
1720  * Read into the page cache. If a page already exists, and PageUptodate() is
1721  * not set, try to fill the page but don't wait for it to become unlocked.
1722  *
1723  * If the page does not get brought uptodate, return -EIO.
1724  */
1725 struct page *read_cache_page_async(struct address_space *mapping,
1726                                 pgoff_t index,
1727                                 int (*filler)(void *,struct page*),
1728                                 void *data)
1729 {
1730         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1731 }
1732 EXPORT_SYMBOL(read_cache_page_async);
1733
1734 static struct page *wait_on_page_read(struct page *page)
1735 {
1736         if (!IS_ERR(page)) {
1737                 wait_on_page_locked(page);
1738                 if (!PageUptodate(page)) {
1739                         page_cache_release(page);
1740                         page = ERR_PTR(-EIO);
1741                 }
1742         }
1743         return page;
1744 }
1745
1746 /**
1747  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1748  * @mapping:    the page's address_space
1749  * @index:      the page index
1750  * @gfp:        the page allocator flags to use if allocating
1751  *
1752  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1753  * any new page allocations done using the specified allocation flags. Note
1754  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1755  * expect to do this atomically or anything like that - but you can pass in
1756  * other page requirements.
1757  *
1758  * If the page does not get brought uptodate, return -EIO.
1759  */
1760 struct page *read_cache_page_gfp(struct address_space *mapping,
1761                                 pgoff_t index,
1762                                 gfp_t gfp)
1763 {
1764         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1765
1766         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1767 }
1768 EXPORT_SYMBOL(read_cache_page_gfp);
1769
1770 /**
1771  * read_cache_page - read into page cache, fill it if needed
1772  * @mapping:    the page's address_space
1773  * @index:      the page index
1774  * @filler:     function to perform the read
1775  * @data:       destination for read data
1776  *
1777  * Read into the page cache. If a page already exists, and PageUptodate() is
1778  * not set, try to fill the page then wait for it to become unlocked.
1779  *
1780  * If the page does not get brought uptodate, return -EIO.
1781  */
1782 struct page *read_cache_page(struct address_space *mapping,
1783                                 pgoff_t index,
1784                                 int (*filler)(void *,struct page*),
1785                                 void *data)
1786 {
1787         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1788 }
1789 EXPORT_SYMBOL(read_cache_page);
1790
1791 /*
1792  * The logic we want is
1793  *
1794  *      if suid or (sgid and xgrp)
1795  *              remove privs
1796  */
1797 int should_remove_suid(struct dentry *dentry)
1798 {
1799         mode_t mode = dentry->d_inode->i_mode;
1800         int kill = 0;
1801
1802         /* suid always must be killed */
1803         if (unlikely(mode & S_ISUID))
1804                 kill = ATTR_KILL_SUID;
1805
1806         /*
1807          * sgid without any exec bits is just a mandatory locking mark; leave
1808          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1809          */
1810         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1811                 kill |= ATTR_KILL_SGID;
1812
1813         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1814                 return kill;
1815
1816         return 0;
1817 }
1818 EXPORT_SYMBOL(should_remove_suid);
1819
1820 static int __remove_suid(struct dentry *dentry, int kill)
1821 {
1822         struct iattr newattrs;
1823
1824         newattrs.ia_valid = ATTR_FORCE | kill;
1825         return notify_change(dentry, &newattrs);
1826 }
1827
1828 int file_remove_suid(struct file *file)
1829 {
1830         struct dentry *dentry = file->f_path.dentry;
1831         int killsuid = should_remove_suid(dentry);
1832         int killpriv = security_inode_need_killpriv(dentry);
1833         int error = 0;
1834
1835         if (killpriv < 0)
1836                 return killpriv;
1837         if (killpriv)
1838                 error = security_inode_killpriv(dentry);
1839         if (!error && killsuid)
1840                 error = __remove_suid(dentry, killsuid);
1841
1842         return error;
1843 }
1844 EXPORT_SYMBOL(file_remove_suid);
1845
1846 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1847                         const struct iovec *iov, size_t base, size_t bytes)
1848 {
1849         size_t copied = 0, left = 0;
1850
1851         while (bytes) {
1852                 char __user *buf = iov->iov_base + base;
1853                 int copy = min(bytes, iov->iov_len - base);
1854
1855                 base = 0;
1856                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1857                 copied += copy;
1858                 bytes -= copy;
1859                 vaddr += copy;
1860                 iov++;
1861
1862                 if (unlikely(left))
1863                         break;
1864         }
1865         return copied - left;
1866 }
1867
1868 /*
1869  * Copy as much as we can into the page and return the number of bytes which
1870  * were successfully copied.  If a fault is encountered then return the number of
1871  * bytes which were copied.
1872  */
1873 size_t iov_iter_copy_from_user_atomic(struct page *page,
1874                 struct iov_iter *i, unsigned long offset, size_t bytes)
1875 {
1876         char *kaddr;
1877         size_t copied;
1878
1879         BUG_ON(!in_atomic());
1880         kaddr = kmap_atomic(page, KM_USER0);
1881         if (likely(i->nr_segs == 1)) {
1882                 int left;
1883                 char __user *buf = i->iov->iov_base + i->iov_offset;
1884                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1885                 copied = bytes - left;
1886         } else {
1887                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1888                                                 i->iov, i->iov_offset, bytes);
1889         }
1890         kunmap_atomic(kaddr, KM_USER0);
1891
1892         return copied;
1893 }
1894 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1895
1896 /*
1897  * This has the same sideeffects and return value as
1898  * iov_iter_copy_from_user_atomic().
1899  * The difference is that it attempts to resolve faults.
1900  * Page must not be locked.
1901  */
1902 size_t iov_iter_copy_from_user(struct page *page,
1903                 struct iov_iter *i, unsigned long offset, size_t bytes)
1904 {
1905         char *kaddr;
1906         size_t copied;
1907
1908         kaddr = kmap(page);
1909         if (likely(i->nr_segs == 1)) {
1910                 int left;
1911                 char __user *buf = i->iov->iov_base + i->iov_offset;
1912                 left = __copy_from_user(kaddr + offset, buf, bytes);
1913                 copied = bytes - left;
1914         } else {
1915                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1916                                                 i->iov, i->iov_offset, bytes);
1917         }
1918         kunmap(page);
1919         return copied;
1920 }
1921 EXPORT_SYMBOL(iov_iter_copy_from_user);
1922
1923 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1924 {
1925         BUG_ON(i->count < bytes);
1926
1927         if (likely(i->nr_segs == 1)) {
1928                 i->iov_offset += bytes;
1929                 i->count -= bytes;
1930         } else {
1931                 const struct iovec *iov = i->iov;
1932                 size_t base = i->iov_offset;
1933
1934                 /*
1935                  * The !iov->iov_len check ensures we skip over unlikely
1936                  * zero-length segments (without overruning the iovec).
1937                  */
1938                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1939                         int copy;
1940
1941                         copy = min(bytes, iov->iov_len - base);
1942                         BUG_ON(!i->count || i->count < copy);
1943                         i->count -= copy;
1944                         bytes -= copy;
1945                         base += copy;
1946                         if (iov->iov_len == base) {
1947                                 iov++;
1948                                 base = 0;
1949                         }
1950                 }
1951                 i->iov = iov;
1952                 i->iov_offset = base;
1953         }
1954 }
1955 EXPORT_SYMBOL(iov_iter_advance);
1956
1957 /*
1958  * Fault in the first iovec of the given iov_iter, to a maximum length
1959  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1960  * accessed (ie. because it is an invalid address).
1961  *
1962  * writev-intensive code may want this to prefault several iovecs -- that
1963  * would be possible (callers must not rely on the fact that _only_ the
1964  * first iovec will be faulted with the current implementation).
1965  */
1966 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1967 {
1968         char __user *buf = i->iov->iov_base + i->iov_offset;
1969         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1970         return fault_in_pages_readable(buf, bytes);
1971 }
1972 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1973
1974 /*
1975  * Return the count of just the current iov_iter segment.
1976  */
1977 size_t iov_iter_single_seg_count(struct iov_iter *i)
1978 {
1979         const struct iovec *iov = i->iov;
1980         if (i->nr_segs == 1)
1981                 return i->count;
1982         else
1983                 return min(i->count, iov->iov_len - i->iov_offset);
1984 }
1985 EXPORT_SYMBOL(iov_iter_single_seg_count);
1986
1987 /*
1988  * Performs necessary checks before doing a write
1989  *
1990  * Can adjust writing position or amount of bytes to write.
1991  * Returns appropriate error code that caller should return or
1992  * zero in case that write should be allowed.
1993  */
1994 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1995 {
1996         struct inode *inode = file->f_mapping->host;
1997         unsigned long limit = rlimit(RLIMIT_FSIZE);
1998
1999         if (unlikely(*pos < 0))
2000                 return -EINVAL;
2001
2002         if (!isblk) {
2003                 /* FIXME: this is for backwards compatibility with 2.4 */
2004                 if (file->f_flags & O_APPEND)
2005                         *pos = i_size_read(inode);
2006
2007                 if (limit != RLIM_INFINITY) {
2008                         if (*pos >= limit) {
2009                                 send_sig(SIGXFSZ, current, 0);
2010                                 return -EFBIG;
2011                         }
2012                         if (*count > limit - (typeof(limit))*pos) {
2013                                 *count = limit - (typeof(limit))*pos;
2014                         }
2015                 }
2016         }
2017
2018         /*
2019          * LFS rule
2020          */
2021         if (unlikely(*pos + *count > MAX_NON_LFS &&
2022                                 !(file->f_flags & O_LARGEFILE))) {
2023                 if (*pos >= MAX_NON_LFS) {
2024                         return -EFBIG;
2025                 }
2026                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2027                         *count = MAX_NON_LFS - (unsigned long)*pos;
2028                 }
2029         }
2030
2031         /*
2032          * Are we about to exceed the fs block limit ?
2033          *
2034          * If we have written data it becomes a short write.  If we have
2035          * exceeded without writing data we send a signal and return EFBIG.
2036          * Linus frestrict idea will clean these up nicely..
2037          */
2038         if (likely(!isblk)) {
2039                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2040                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2041                                 return -EFBIG;
2042                         }
2043                         /* zero-length writes at ->s_maxbytes are OK */
2044                 }
2045
2046                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2047                         *count = inode->i_sb->s_maxbytes - *pos;
2048         } else {
2049 #ifdef CONFIG_BLOCK
2050                 loff_t isize;
2051                 if (bdev_read_only(I_BDEV(inode)))
2052                         return -EPERM;
2053                 isize = i_size_read(inode);
2054                 if (*pos >= isize) {
2055                         if (*count || *pos > isize)
2056                                 return -ENOSPC;
2057                 }
2058
2059                 if (*pos + *count > isize)
2060                         *count = isize - *pos;
2061 #else
2062                 return -EPERM;
2063 #endif
2064         }
2065         return 0;
2066 }
2067 EXPORT_SYMBOL(generic_write_checks);
2068
2069 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2070                                 loff_t pos, unsigned len, unsigned flags,
2071                                 struct page **pagep, void **fsdata)
2072 {
2073         const struct address_space_operations *aops = mapping->a_ops;
2074
2075         return aops->write_begin(file, mapping, pos, len, flags,
2076                                                         pagep, fsdata);
2077 }
2078 EXPORT_SYMBOL(pagecache_write_begin);
2079
2080 int pagecache_write_end(struct file *file, struct address_space *mapping,
2081                                 loff_t pos, unsigned len, unsigned copied,
2082                                 struct page *page, void *fsdata)
2083 {
2084         const struct address_space_operations *aops = mapping->a_ops;
2085
2086         mark_page_accessed(page);
2087         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2088 }
2089 EXPORT_SYMBOL(pagecache_write_end);
2090
2091 ssize_t
2092 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2093                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2094                 size_t count, size_t ocount)
2095 {
2096         struct file     *file = iocb->ki_filp;
2097         struct address_space *mapping = file->f_mapping;
2098         struct inode    *inode = mapping->host;
2099         ssize_t         written;
2100         size_t          write_len;
2101         pgoff_t         end;
2102
2103         if (count != ocount)
2104                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2105
2106         write_len = iov_length(iov, *nr_segs);
2107         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2108
2109         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2110         if (written)
2111                 goto out;
2112
2113         /*
2114          * After a write we want buffered reads to be sure to go to disk to get
2115          * the new data.  We invalidate clean cached page from the region we're
2116          * about to write.  We do this *before* the write so that we can return
2117          * without clobbering -EIOCBQUEUED from ->direct_IO().
2118          */
2119         if (mapping->nrpages) {
2120                 written = invalidate_inode_pages2_range(mapping,
2121                                         pos >> PAGE_CACHE_SHIFT, end);
2122                 /*
2123                  * If a page can not be invalidated, return 0 to fall back
2124                  * to buffered write.
2125                  */
2126                 if (written) {
2127                         if (written == -EBUSY)
2128                                 return 0;
2129                         goto out;
2130                 }
2131         }
2132
2133         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2134
2135         /*
2136          * Finally, try again to invalidate clean pages which might have been
2137          * cached by non-direct readahead, or faulted in by get_user_pages()
2138          * if the source of the write was an mmap'ed region of the file
2139          * we're writing.  Either one is a pretty crazy thing to do,
2140          * so we don't support it 100%.  If this invalidation
2141          * fails, tough, the write still worked...
2142          */
2143         if (mapping->nrpages) {
2144                 invalidate_inode_pages2_range(mapping,
2145                                               pos >> PAGE_CACHE_SHIFT, end);
2146         }
2147
2148         if (written > 0) {
2149                 pos += written;
2150                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2151                         i_size_write(inode, pos);
2152                         mark_inode_dirty(inode);
2153                 }
2154                 *ppos = pos;
2155         }
2156 out:
2157         return written;
2158 }
2159 EXPORT_SYMBOL(generic_file_direct_write);
2160
2161 /*
2162  * Find or create a page at the given pagecache position. Return the locked
2163  * page. This function is specifically for buffered writes.
2164  */
2165 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2166                                         pgoff_t index, unsigned flags)
2167 {
2168         int status;
2169         struct page *page;
2170         gfp_t gfp_notmask = 0;
2171         if (flags & AOP_FLAG_NOFS)
2172                 gfp_notmask = __GFP_FS;
2173 repeat:
2174         page = find_lock_page(mapping, index);
2175         if (page)
2176                 return page;
2177
2178         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2179         if (!page)
2180                 return NULL;
2181         status = add_to_page_cache_lru(page, mapping, index,
2182                                                 GFP_KERNEL & ~gfp_notmask);
2183         if (unlikely(status)) {
2184                 page_cache_release(page);
2185                 if (status == -EEXIST)
2186                         goto repeat;
2187                 return NULL;
2188         }
2189         return page;
2190 }
2191 EXPORT_SYMBOL(grab_cache_page_write_begin);
2192
2193 static ssize_t generic_perform_write(struct file *file,
2194                                 struct iov_iter *i, loff_t pos)
2195 {
2196         struct address_space *mapping = file->f_mapping;
2197         const struct address_space_operations *a_ops = mapping->a_ops;
2198         long status = 0;
2199         ssize_t written = 0;
2200         unsigned int flags = 0;
2201
2202         /*
2203          * Copies from kernel address space cannot fail (NFSD is a big user).
2204          */
2205         if (segment_eq(get_fs(), KERNEL_DS))
2206                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2207
2208         do {
2209                 struct page *page;
2210                 unsigned long offset;   /* Offset into pagecache page */
2211                 unsigned long bytes;    /* Bytes to write to page */
2212                 size_t copied;          /* Bytes copied from user */
2213                 void *fsdata;
2214
2215                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2216                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2217                                                 iov_iter_count(i));
2218
2219 again:
2220
2221                 /*
2222                  * Bring in the user page that we will copy from _first_.
2223                  * Otherwise there's a nasty deadlock on copying from the
2224                  * same page as we're writing to, without it being marked
2225                  * up-to-date.
2226                  *
2227                  * Not only is this an optimisation, but it is also required
2228                  * to check that the address is actually valid, when atomic
2229                  * usercopies are used, below.
2230                  */
2231                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2232                         status = -EFAULT;
2233                         break;
2234                 }
2235
2236                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2237                                                 &page, &fsdata);
2238                 if (unlikely(status))
2239                         break;
2240
2241                 if (mapping_writably_mapped(mapping))
2242                         flush_dcache_page(page);
2243
2244                 pagefault_disable();
2245                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2246                 pagefault_enable();
2247                 flush_dcache_page(page);
2248
2249                 mark_page_accessed(page);
2250                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2251                                                 page, fsdata);
2252                 if (unlikely(status < 0))
2253                         break;
2254                 copied = status;
2255
2256                 cond_resched();
2257
2258                 iov_iter_advance(i, copied);
2259                 if (unlikely(copied == 0)) {
2260                         /*
2261                          * If we were unable to copy any data at all, we must
2262                          * fall back to a single segment length write.
2263                          *
2264                          * If we didn't fallback here, we could livelock
2265                          * because not all segments in the iov can be copied at
2266                          * once without a pagefault.
2267                          */
2268                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2269                                                 iov_iter_single_seg_count(i));
2270                         goto again;
2271                 }
2272                 pos += copied;
2273                 written += copied;
2274
2275                 balance_dirty_pages_ratelimited(mapping);
2276
2277         } while (iov_iter_count(i));
2278
2279         return written ? written : status;
2280 }
2281
2282 ssize_t
2283 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2284                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2285                 size_t count, ssize_t written)
2286 {
2287         struct file *file = iocb->ki_filp;
2288         ssize_t status;
2289         struct iov_iter i;
2290
2291         iov_iter_init(&i, iov, nr_segs, count, written);
2292         status = generic_perform_write(file, &i, pos);
2293
2294         if (likely(status >= 0)) {
2295                 written += status;
2296                 *ppos = pos + status;
2297         }
2298         
2299         return written ? written : status;
2300 }
2301 EXPORT_SYMBOL(generic_file_buffered_write);
2302
2303 /**
2304  * __generic_file_aio_write - write data to a file
2305  * @iocb:       IO state structure (file, offset, etc.)
2306  * @iov:        vector with data to write
2307  * @nr_segs:    number of segments in the vector
2308  * @ppos:       position where to write
2309  *
2310  * This function does all the work needed for actually writing data to a
2311  * file. It does all basic checks, removes SUID from the file, updates
2312  * modification times and calls proper subroutines depending on whether we
2313  * do direct IO or a standard buffered write.
2314  *
2315  * It expects i_mutex to be grabbed unless we work on a block device or similar
2316  * object which does not need locking at all.
2317  *
2318  * This function does *not* take care of syncing data in case of O_SYNC write.
2319  * A caller has to handle it. This is mainly due to the fact that we want to
2320  * avoid syncing under i_mutex.
2321  */
2322 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2323                                  unsigned long nr_segs, loff_t *ppos)
2324 {
2325         struct file *file = iocb->ki_filp;
2326         struct address_space * mapping = file->f_mapping;
2327         size_t ocount;          /* original count */
2328         size_t count;           /* after file limit checks */
2329         struct inode    *inode = mapping->host;
2330         loff_t          pos;
2331         ssize_t         written;
2332         ssize_t         err;
2333
2334         ocount = 0;
2335         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2336         if (err)
2337                 return err;
2338
2339         count = ocount;
2340         pos = *ppos;
2341
2342         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2343
2344         /* We can write back this queue in page reclaim */
2345         current->backing_dev_info = mapping->backing_dev_info;
2346         written = 0;
2347
2348         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2349         if (err)
2350                 goto out;
2351
2352         if (count == 0)
2353                 goto out;
2354
2355         err = file_remove_suid(file);
2356         if (err)
2357                 goto out;
2358
2359         file_update_time(file);
2360
2361         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2362         if (unlikely(file->f_flags & O_DIRECT)) {
2363                 loff_t endbyte;
2364                 ssize_t written_buffered;
2365
2366                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2367                                                         ppos, count, ocount);
2368                 if (written < 0 || written == count)
2369                         goto out;
2370                 /*
2371                  * direct-io write to a hole: fall through to buffered I/O
2372                  * for completing the rest of the request.
2373                  */
2374                 pos += written;
2375                 count -= written;
2376                 written_buffered = generic_file_buffered_write(iocb, iov,
2377                                                 nr_segs, pos, ppos, count,
2378                                                 written);
2379                 /*
2380                  * If generic_file_buffered_write() retuned a synchronous error
2381                  * then we want to return the number of bytes which were
2382                  * direct-written, or the error code if that was zero.  Note
2383                  * that this differs from normal direct-io semantics, which
2384                  * will return -EFOO even if some bytes were written.
2385                  */
2386                 if (written_buffered < 0) {
2387                         err = written_buffered;
2388                         goto out;
2389                 }
2390
2391                 /*
2392                  * We need to ensure that the page cache pages are written to
2393                  * disk and invalidated to preserve the expected O_DIRECT
2394                  * semantics.
2395                  */
2396                 endbyte = pos + written_buffered - written - 1;
2397                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2398                 if (err == 0) {
2399                         written = written_buffered;
2400                         invalidate_mapping_pages(mapping,
2401                                                  pos >> PAGE_CACHE_SHIFT,
2402                                                  endbyte >> PAGE_CACHE_SHIFT);
2403                 } else {
2404                         /*
2405                          * We don't know how much we wrote, so just return
2406                          * the number of bytes which were direct-written
2407                          */
2408                 }
2409         } else {
2410                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2411                                 pos, ppos, count, written);
2412         }
2413 out:
2414         current->backing_dev_info = NULL;
2415         return written ? written : err;
2416 }
2417 EXPORT_SYMBOL(__generic_file_aio_write);
2418
2419 /**
2420  * generic_file_aio_write - write data to a file
2421  * @iocb:       IO state structure
2422  * @iov:        vector with data to write
2423  * @nr_segs:    number of segments in the vector
2424  * @pos:        position in file where to write
2425  *
2426  * This is a wrapper around __generic_file_aio_write() to be used by most
2427  * filesystems. It takes care of syncing the file in case of O_SYNC file
2428  * and acquires i_mutex as needed.
2429  */
2430 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2431                 unsigned long nr_segs, loff_t pos)
2432 {
2433         struct file *file = iocb->ki_filp;
2434         struct inode *inode = file->f_mapping->host;
2435         ssize_t ret;
2436
2437         BUG_ON(iocb->ki_pos != pos);
2438
2439         mutex_lock(&inode->i_mutex);
2440         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2441         mutex_unlock(&inode->i_mutex);
2442
2443         if (ret > 0 || ret == -EIOCBQUEUED) {
2444                 ssize_t err;
2445
2446                 err = generic_write_sync(file, pos, ret);
2447                 if (err < 0 && ret > 0)
2448                         ret = err;
2449         }
2450         return ret;
2451 }
2452 EXPORT_SYMBOL(generic_file_aio_write);
2453
2454 /**
2455  * try_to_release_page() - release old fs-specific metadata on a page
2456  *
2457  * @page: the page which the kernel is trying to free
2458  * @gfp_mask: memory allocation flags (and I/O mode)
2459  *
2460  * The address_space is to try to release any data against the page
2461  * (presumably at page->private).  If the release was successful, return `1'.
2462  * Otherwise return zero.
2463  *
2464  * This may also be called if PG_fscache is set on a page, indicating that the
2465  * page is known to the local caching routines.
2466  *
2467  * The @gfp_mask argument specifies whether I/O may be performed to release
2468  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2469  *
2470  */
2471 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2472 {
2473         struct address_space * const mapping = page->mapping;
2474
2475         BUG_ON(!PageLocked(page));
2476         if (PageWriteback(page))
2477                 return 0;
2478
2479         if (mapping && mapping->a_ops->releasepage)
2480                 return mapping->a_ops->releasepage(page, gfp_mask);
2481         return try_to_free_buffers(page);
2482 }
2483
2484 EXPORT_SYMBOL(try_to_release_page);