mm: remove likely() from grab_cache_page_write_begin()
[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 sync_page(void *word)
159 {
160         struct address_space *mapping;
161         struct page *page;
162
163         page = container_of((unsigned long *)word, struct page, flags);
164
165         /*
166          * page_mapping() is being called without PG_locked held.
167          * Some knowledge of the state and use of the page is used to
168          * reduce the requirements down to a memory barrier.
169          * The danger here is of a stale page_mapping() return value
170          * indicating a struct address_space different from the one it's
171          * associated with when it is associated with one.
172          * After smp_mb(), it's either the correct page_mapping() for
173          * the page, or an old page_mapping() and the page's own
174          * page_mapping() has gone NULL.
175          * The ->sync_page() address_space operation must tolerate
176          * page_mapping() going NULL. By an amazing coincidence,
177          * this comes about because none of the users of the page
178          * in the ->sync_page() methods make essential use of the
179          * page_mapping(), merely passing the page down to the backing
180          * device's unplug functions when it's non-NULL, which in turn
181          * ignore it for all cases but swap, where only page_private(page) is
182          * of interest. When page_mapping() does go NULL, the entire
183          * call stack gracefully ignores the page and returns.
184          * -- wli
185          */
186         smp_mb();
187         mapping = page_mapping(page);
188         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
189                 mapping->a_ops->sync_page(page);
190         io_schedule();
191         return 0;
192 }
193
194 static int sync_page_killable(void *word)
195 {
196         sync_page(word);
197         return fatal_signal_pending(current) ? -EINTR : 0;
198 }
199
200 /**
201  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
202  * @mapping:    address space structure to write
203  * @start:      offset in bytes where the range starts
204  * @end:        offset in bytes where the range ends (inclusive)
205  * @sync_mode:  enable synchronous operation
206  *
207  * Start writeback against all of a mapping's dirty pages that lie
208  * within the byte offsets <start, end> inclusive.
209  *
210  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
211  * opposed to a regular memory cleansing writeback.  The difference between
212  * these two operations is that if a dirty page/buffer is encountered, it must
213  * be waited upon, and not just skipped over.
214  */
215 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
216                                 loff_t end, int sync_mode)
217 {
218         int ret;
219         struct writeback_control wbc = {
220                 .sync_mode = sync_mode,
221                 .nr_to_write = LONG_MAX,
222                 .range_start = start,
223                 .range_end = end,
224         };
225
226         if (!mapping_cap_writeback_dirty(mapping))
227                 return 0;
228
229         ret = do_writepages(mapping, &wbc);
230         return ret;
231 }
232
233 static inline int __filemap_fdatawrite(struct address_space *mapping,
234         int sync_mode)
235 {
236         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
237 }
238
239 int filemap_fdatawrite(struct address_space *mapping)
240 {
241         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
242 }
243 EXPORT_SYMBOL(filemap_fdatawrite);
244
245 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
246                                 loff_t end)
247 {
248         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
249 }
250 EXPORT_SYMBOL(filemap_fdatawrite_range);
251
252 /**
253  * filemap_flush - mostly a non-blocking flush
254  * @mapping:    target address_space
255  *
256  * This is a mostly non-blocking flush.  Not suitable for data-integrity
257  * purposes - I/O may not be started against all dirty pages.
258  */
259 int filemap_flush(struct address_space *mapping)
260 {
261         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
262 }
263 EXPORT_SYMBOL(filemap_flush);
264
265 /**
266  * filemap_fdatawait_range - wait for writeback to complete
267  * @mapping:            address space structure to wait for
268  * @start_byte:         offset in bytes where the range starts
269  * @end_byte:           offset in bytes where the range ends (inclusive)
270  *
271  * Walk the list of under-writeback pages of the given address space
272  * in the given range and wait for all of them.
273  */
274 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
275                             loff_t end_byte)
276 {
277         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
278         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
279         struct pagevec pvec;
280         int nr_pages;
281         int ret = 0;
282
283         if (end_byte < start_byte)
284                 return 0;
285
286         pagevec_init(&pvec, 0);
287         while ((index <= end) &&
288                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
289                         PAGECACHE_TAG_WRITEBACK,
290                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
291                 unsigned i;
292
293                 for (i = 0; i < nr_pages; i++) {
294                         struct page *page = pvec.pages[i];
295
296                         /* until radix tree lookup accepts end_index */
297                         if (page->index > end)
298                                 continue;
299
300                         wait_on_page_writeback(page);
301                         if (TestClearPageError(page))
302                                 ret = -EIO;
303                 }
304                 pagevec_release(&pvec);
305                 cond_resched();
306         }
307
308         /* Check for outstanding write errors */
309         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
310                 ret = -ENOSPC;
311         if (test_and_clear_bit(AS_EIO, &mapping->flags))
312                 ret = -EIO;
313
314         return ret;
315 }
316 EXPORT_SYMBOL(filemap_fdatawait_range);
317
318 /**
319  * filemap_fdatawait - wait for all under-writeback pages to complete
320  * @mapping: address space structure to wait for
321  *
322  * Walk the list of under-writeback pages of the given address space
323  * and wait for all of them.
324  */
325 int filemap_fdatawait(struct address_space *mapping)
326 {
327         loff_t i_size = i_size_read(mapping->host);
328
329         if (i_size == 0)
330                 return 0;
331
332         return filemap_fdatawait_range(mapping, 0, i_size - 1);
333 }
334 EXPORT_SYMBOL(filemap_fdatawait);
335
336 int filemap_write_and_wait(struct address_space *mapping)
337 {
338         int err = 0;
339
340         if (mapping->nrpages) {
341                 err = filemap_fdatawrite(mapping);
342                 /*
343                  * Even if the above returned error, the pages may be
344                  * written partially (e.g. -ENOSPC), so we wait for it.
345                  * But the -EIO is special case, it may indicate the worst
346                  * thing (e.g. bug) happened, so we avoid waiting for it.
347                  */
348                 if (err != -EIO) {
349                         int err2 = filemap_fdatawait(mapping);
350                         if (!err)
351                                 err = err2;
352                 }
353         }
354         return err;
355 }
356 EXPORT_SYMBOL(filemap_write_and_wait);
357
358 /**
359  * filemap_write_and_wait_range - write out & wait on a file range
360  * @mapping:    the address_space for the pages
361  * @lstart:     offset in bytes where the range starts
362  * @lend:       offset in bytes where the range ends (inclusive)
363  *
364  * Write out and wait upon file offsets lstart->lend, inclusive.
365  *
366  * Note that `lend' is inclusive (describes the last byte to be written) so
367  * that this function can be used to write to the very end-of-file (end = -1).
368  */
369 int filemap_write_and_wait_range(struct address_space *mapping,
370                                  loff_t lstart, loff_t lend)
371 {
372         int err = 0;
373
374         if (mapping->nrpages) {
375                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
376                                                  WB_SYNC_ALL);
377                 /* See comment of filemap_write_and_wait() */
378                 if (err != -EIO) {
379                         int err2 = filemap_fdatawait_range(mapping,
380                                                 lstart, lend);
381                         if (!err)
382                                 err = err2;
383                 }
384         }
385         return err;
386 }
387 EXPORT_SYMBOL(filemap_write_and_wait_range);
388
389 /**
390  * add_to_page_cache_locked - add a locked page to the pagecache
391  * @page:       page to add
392  * @mapping:    the page's address_space
393  * @offset:     page index
394  * @gfp_mask:   page allocation mode
395  *
396  * This function is used to add a page to the pagecache. It must be locked.
397  * This function does not add the page to the LRU.  The caller must do that.
398  */
399 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
400                 pgoff_t offset, gfp_t gfp_mask)
401 {
402         int error;
403
404         VM_BUG_ON(!PageLocked(page));
405
406         error = mem_cgroup_cache_charge(page, current->mm,
407                                         gfp_mask & GFP_RECLAIM_MASK);
408         if (error)
409                 goto out;
410
411         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
412         if (error == 0) {
413                 page_cache_get(page);
414                 page->mapping = mapping;
415                 page->index = offset;
416
417                 spin_lock_irq(&mapping->tree_lock);
418                 error = radix_tree_insert(&mapping->page_tree, offset, page);
419                 if (likely(!error)) {
420                         mapping->nrpages++;
421                         __inc_zone_page_state(page, NR_FILE_PAGES);
422                         if (PageSwapBacked(page))
423                                 __inc_zone_page_state(page, NR_SHMEM);
424                         spin_unlock_irq(&mapping->tree_lock);
425                 } else {
426                         page->mapping = NULL;
427                         spin_unlock_irq(&mapping->tree_lock);
428                         mem_cgroup_uncharge_cache_page(page);
429                         page_cache_release(page);
430                 }
431                 radix_tree_preload_end();
432         } else
433                 mem_cgroup_uncharge_cache_page(page);
434 out:
435         return error;
436 }
437 EXPORT_SYMBOL(add_to_page_cache_locked);
438
439 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
440                                 pgoff_t offset, gfp_t gfp_mask)
441 {
442         int ret;
443
444         /*
445          * Splice_read and readahead add shmem/tmpfs pages into the page cache
446          * before shmem_readpage has a chance to mark them as SwapBacked: they
447          * need to go on the anon lru below, and mem_cgroup_cache_charge
448          * (called in add_to_page_cache) needs to know where they're going too.
449          */
450         if (mapping_cap_swap_backed(mapping))
451                 SetPageSwapBacked(page);
452
453         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
454         if (ret == 0) {
455                 if (page_is_file_cache(page))
456                         lru_cache_add_file(page);
457                 else
458                         lru_cache_add_anon(page);
459         }
460         return ret;
461 }
462 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
463
464 #ifdef CONFIG_NUMA
465 struct page *__page_cache_alloc(gfp_t gfp)
466 {
467         int n;
468         struct page *page;
469
470         if (cpuset_do_page_mem_spread()) {
471                 get_mems_allowed();
472                 n = cpuset_mem_spread_node();
473                 page = alloc_pages_exact_node(n, gfp, 0);
474                 put_mems_allowed();
475                 return page;
476         }
477         return alloc_pages(gfp, 0);
478 }
479 EXPORT_SYMBOL(__page_cache_alloc);
480 #endif
481
482 static int __sleep_on_page_lock(void *word)
483 {
484         io_schedule();
485         return 0;
486 }
487
488 /*
489  * In order to wait for pages to become available there must be
490  * waitqueues associated with pages. By using a hash table of
491  * waitqueues where the bucket discipline is to maintain all
492  * waiters on the same queue and wake all when any of the pages
493  * become available, and for the woken contexts to check to be
494  * sure the appropriate page became available, this saves space
495  * at a cost of "thundering herd" phenomena during rare hash
496  * collisions.
497  */
498 static wait_queue_head_t *page_waitqueue(struct page *page)
499 {
500         const struct zone *zone = page_zone(page);
501
502         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
503 }
504
505 static inline void wake_up_page(struct page *page, int bit)
506 {
507         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
508 }
509
510 void wait_on_page_bit(struct page *page, int bit_nr)
511 {
512         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
513
514         if (test_bit(bit_nr, &page->flags))
515                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
516                                                         TASK_UNINTERRUPTIBLE);
517 }
518 EXPORT_SYMBOL(wait_on_page_bit);
519
520 /**
521  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
522  * @page: Page defining the wait queue of interest
523  * @waiter: Waiter to add to the queue
524  *
525  * Add an arbitrary @waiter to the wait queue for the nominated @page.
526  */
527 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
528 {
529         wait_queue_head_t *q = page_waitqueue(page);
530         unsigned long flags;
531
532         spin_lock_irqsave(&q->lock, flags);
533         __add_wait_queue(q, waiter);
534         spin_unlock_irqrestore(&q->lock, flags);
535 }
536 EXPORT_SYMBOL_GPL(add_page_wait_queue);
537
538 /**
539  * unlock_page - unlock a locked page
540  * @page: the page
541  *
542  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
543  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
544  * mechananism between PageLocked pages and PageWriteback pages is shared.
545  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
546  *
547  * The mb is necessary to enforce ordering between the clear_bit and the read
548  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
549  */
550 void unlock_page(struct page *page)
551 {
552         VM_BUG_ON(!PageLocked(page));
553         clear_bit_unlock(PG_locked, &page->flags);
554         smp_mb__after_clear_bit();
555         wake_up_page(page, PG_locked);
556 }
557 EXPORT_SYMBOL(unlock_page);
558
559 /**
560  * end_page_writeback - end writeback against a page
561  * @page: the page
562  */
563 void end_page_writeback(struct page *page)
564 {
565         if (TestClearPageReclaim(page))
566                 rotate_reclaimable_page(page);
567
568         if (!test_clear_page_writeback(page))
569                 BUG();
570
571         smp_mb__after_clear_bit();
572         wake_up_page(page, PG_writeback);
573 }
574 EXPORT_SYMBOL(end_page_writeback);
575
576 /**
577  * __lock_page - get a lock on the page, assuming we need to sleep to get it
578  * @page: the page to lock
579  *
580  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
581  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
582  * chances are that on the second loop, the block layer's plug list is empty,
583  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
584  */
585 void __lock_page(struct page *page)
586 {
587         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
588
589         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
590                                                         TASK_UNINTERRUPTIBLE);
591 }
592 EXPORT_SYMBOL(__lock_page);
593
594 int __lock_page_killable(struct page *page)
595 {
596         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
597
598         return __wait_on_bit_lock(page_waitqueue(page), &wait,
599                                         sync_page_killable, TASK_KILLABLE);
600 }
601 EXPORT_SYMBOL_GPL(__lock_page_killable);
602
603 /**
604  * __lock_page_nosync - get a lock on the page, without calling sync_page()
605  * @page: the page to lock
606  *
607  * Variant of lock_page that does not require the caller to hold a reference
608  * on the page's mapping.
609  */
610 void __lock_page_nosync(struct page *page)
611 {
612         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
613         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
614                                                         TASK_UNINTERRUPTIBLE);
615 }
616
617 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
618                          unsigned int flags)
619 {
620         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
621                 __lock_page(page);
622                 return 1;
623         } else {
624                 up_read(&mm->mmap_sem);
625                 wait_on_page_locked(page);
626                 return 0;
627         }
628 }
629
630 /**
631  * find_get_page - find and get a page reference
632  * @mapping: the address_space to search
633  * @offset: the page index
634  *
635  * Is there a pagecache struct page at the given (mapping, offset) tuple?
636  * If yes, increment its refcount and return it; if no, return NULL.
637  */
638 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
639 {
640         void **pagep;
641         struct page *page;
642
643         rcu_read_lock();
644 repeat:
645         page = NULL;
646         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
647         if (pagep) {
648                 page = radix_tree_deref_slot(pagep);
649                 if (unlikely(!page))
650                         goto out;
651                 if (radix_tree_deref_retry(page))
652                         goto repeat;
653
654                 if (!page_cache_get_speculative(page))
655                         goto repeat;
656
657                 /*
658                  * Has the page moved?
659                  * This is part of the lockless pagecache protocol. See
660                  * include/linux/pagemap.h for details.
661                  */
662                 if (unlikely(page != *pagep)) {
663                         page_cache_release(page);
664                         goto repeat;
665                 }
666         }
667 out:
668         rcu_read_unlock();
669
670         return page;
671 }
672 EXPORT_SYMBOL(find_get_page);
673
674 /**
675  * find_lock_page - locate, pin and lock a pagecache page
676  * @mapping: the address_space to search
677  * @offset: the page index
678  *
679  * Locates the desired pagecache page, locks it, increments its reference
680  * count and returns its address.
681  *
682  * Returns zero if the page was not present. find_lock_page() may sleep.
683  */
684 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
685 {
686         struct page *page;
687
688 repeat:
689         page = find_get_page(mapping, offset);
690         if (page) {
691                 lock_page(page);
692                 /* Has the page been truncated? */
693                 if (unlikely(page->mapping != mapping)) {
694                         unlock_page(page);
695                         page_cache_release(page);
696                         goto repeat;
697                 }
698                 VM_BUG_ON(page->index != offset);
699         }
700         return page;
701 }
702 EXPORT_SYMBOL(find_lock_page);
703
704 /**
705  * find_or_create_page - locate or add a pagecache page
706  * @mapping: the page's address_space
707  * @index: the page's index into the mapping
708  * @gfp_mask: page allocation mode
709  *
710  * Locates a page in the pagecache.  If the page is not present, a new page
711  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
712  * LRU list.  The returned page is locked and has its reference count
713  * incremented.
714  *
715  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
716  * allocation!
717  *
718  * find_or_create_page() returns the desired page's address, or zero on
719  * memory exhaustion.
720  */
721 struct page *find_or_create_page(struct address_space *mapping,
722                 pgoff_t index, gfp_t gfp_mask)
723 {
724         struct page *page;
725         int err;
726 repeat:
727         page = find_lock_page(mapping, index);
728         if (!page) {
729                 page = __page_cache_alloc(gfp_mask);
730                 if (!page)
731                         return NULL;
732                 /*
733                  * We want a regular kernel memory (not highmem or DMA etc)
734                  * allocation for the radix tree nodes, but we need to honour
735                  * the context-specific requirements the caller has asked for.
736                  * GFP_RECLAIM_MASK collects those requirements.
737                  */
738                 err = add_to_page_cache_lru(page, mapping, index,
739                         (gfp_mask & GFP_RECLAIM_MASK));
740                 if (unlikely(err)) {
741                         page_cache_release(page);
742                         page = NULL;
743                         if (err == -EEXIST)
744                                 goto repeat;
745                 }
746         }
747         return page;
748 }
749 EXPORT_SYMBOL(find_or_create_page);
750
751 /**
752  * find_get_pages - gang pagecache lookup
753  * @mapping:    The address_space to search
754  * @start:      The starting page index
755  * @nr_pages:   The maximum number of pages
756  * @pages:      Where the resulting pages are placed
757  *
758  * find_get_pages() will search for and return a group of up to
759  * @nr_pages pages in the mapping.  The pages are placed at @pages.
760  * find_get_pages() takes a reference against the returned pages.
761  *
762  * The search returns a group of mapping-contiguous pages with ascending
763  * indexes.  There may be holes in the indices due to not-present pages.
764  *
765  * find_get_pages() returns the number of pages which were found.
766  */
767 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
768                             unsigned int nr_pages, struct page **pages)
769 {
770         unsigned int i;
771         unsigned int ret;
772         unsigned int nr_found;
773
774         rcu_read_lock();
775 restart:
776         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
777                                 (void ***)pages, start, nr_pages);
778         ret = 0;
779         for (i = 0; i < nr_found; i++) {
780                 struct page *page;
781 repeat:
782                 page = radix_tree_deref_slot((void **)pages[i]);
783                 if (unlikely(!page))
784                         continue;
785                 if (radix_tree_deref_retry(page)) {
786                         if (ret)
787                                 start = pages[ret-1]->index;
788                         goto restart;
789                 }
790
791                 if (!page_cache_get_speculative(page))
792                         goto repeat;
793
794                 /* Has the page moved? */
795                 if (unlikely(page != *((void **)pages[i]))) {
796                         page_cache_release(page);
797                         goto repeat;
798                 }
799
800                 pages[ret] = page;
801                 ret++;
802         }
803         rcu_read_unlock();
804         return ret;
805 }
806
807 /**
808  * find_get_pages_contig - gang contiguous pagecache lookup
809  * @mapping:    The address_space to search
810  * @index:      The starting page index
811  * @nr_pages:   The maximum number of pages
812  * @pages:      Where the resulting pages are placed
813  *
814  * find_get_pages_contig() works exactly like find_get_pages(), except
815  * that the returned number of pages are guaranteed to be contiguous.
816  *
817  * find_get_pages_contig() returns the number of pages which were found.
818  */
819 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
820                                unsigned int nr_pages, struct page **pages)
821 {
822         unsigned int i;
823         unsigned int ret;
824         unsigned int nr_found;
825
826         rcu_read_lock();
827 restart:
828         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
829                                 (void ***)pages, index, nr_pages);
830         ret = 0;
831         for (i = 0; i < nr_found; i++) {
832                 struct page *page;
833 repeat:
834                 page = radix_tree_deref_slot((void **)pages[i]);
835                 if (unlikely(!page))
836                         continue;
837                 if (radix_tree_deref_retry(page))
838                         goto restart;
839
840                 if (!page_cache_get_speculative(page))
841                         goto repeat;
842
843                 /* Has the page moved? */
844                 if (unlikely(page != *((void **)pages[i]))) {
845                         page_cache_release(page);
846                         goto repeat;
847                 }
848
849                 /*
850                  * must check mapping and index after taking the ref.
851                  * otherwise we can get both false positives and false
852                  * negatives, which is just confusing to the caller.
853                  */
854                 if (page->mapping == NULL || page->index != index) {
855                         page_cache_release(page);
856                         break;
857                 }
858
859                 pages[ret] = page;
860                 ret++;
861                 index++;
862         }
863         rcu_read_unlock();
864         return ret;
865 }
866 EXPORT_SYMBOL(find_get_pages_contig);
867
868 /**
869  * find_get_pages_tag - find and return pages that match @tag
870  * @mapping:    the address_space to search
871  * @index:      the starting page index
872  * @tag:        the tag index
873  * @nr_pages:   the maximum number of pages
874  * @pages:      where the resulting pages are placed
875  *
876  * Like find_get_pages, except we only return pages which are tagged with
877  * @tag.   We update @index to index the next page for the traversal.
878  */
879 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
880                         int tag, unsigned int nr_pages, struct page **pages)
881 {
882         unsigned int i;
883         unsigned int ret;
884         unsigned int nr_found;
885
886         rcu_read_lock();
887 restart:
888         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
889                                 (void ***)pages, *index, nr_pages, tag);
890         ret = 0;
891         for (i = 0; i < nr_found; i++) {
892                 struct page *page;
893 repeat:
894                 page = radix_tree_deref_slot((void **)pages[i]);
895                 if (unlikely(!page))
896                         continue;
897                 if (radix_tree_deref_retry(page))
898                         goto restart;
899
900                 if (!page_cache_get_speculative(page))
901                         goto repeat;
902
903                 /* Has the page moved? */
904                 if (unlikely(page != *((void **)pages[i]))) {
905                         page_cache_release(page);
906                         goto repeat;
907                 }
908
909                 pages[ret] = page;
910                 ret++;
911         }
912         rcu_read_unlock();
913
914         if (ret)
915                 *index = pages[ret - 1]->index + 1;
916
917         return ret;
918 }
919 EXPORT_SYMBOL(find_get_pages_tag);
920
921 /**
922  * grab_cache_page_nowait - returns locked page at given index in given cache
923  * @mapping: target address_space
924  * @index: the page index
925  *
926  * Same as grab_cache_page(), but do not wait if the page is unavailable.
927  * This is intended for speculative data generators, where the data can
928  * be regenerated if the page couldn't be grabbed.  This routine should
929  * be safe to call while holding the lock for another page.
930  *
931  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
932  * and deadlock against the caller's locked page.
933  */
934 struct page *
935 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
936 {
937         struct page *page = find_get_page(mapping, index);
938
939         if (page) {
940                 if (trylock_page(page))
941                         return page;
942                 page_cache_release(page);
943                 return NULL;
944         }
945         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
946         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
947                 page_cache_release(page);
948                 page = NULL;
949         }
950         return page;
951 }
952 EXPORT_SYMBOL(grab_cache_page_nowait);
953
954 /*
955  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
956  * a _large_ part of the i/o request. Imagine the worst scenario:
957  *
958  *      ---R__________________________________________B__________
959  *         ^ reading here                             ^ bad block(assume 4k)
960  *
961  * read(R) => miss => readahead(R...B) => media error => frustrating retries
962  * => failing the whole request => read(R) => read(R+1) =>
963  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
964  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
965  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
966  *
967  * It is going insane. Fix it by quickly scaling down the readahead size.
968  */
969 static void shrink_readahead_size_eio(struct file *filp,
970                                         struct file_ra_state *ra)
971 {
972         ra->ra_pages /= 4;
973 }
974
975 /**
976  * do_generic_file_read - generic file read routine
977  * @filp:       the file to read
978  * @ppos:       current file position
979  * @desc:       read_descriptor
980  * @actor:      read method
981  *
982  * This is a generic file read routine, and uses the
983  * mapping->a_ops->readpage() function for the actual low-level stuff.
984  *
985  * This is really ugly. But the goto's actually try to clarify some
986  * of the logic when it comes to error handling etc.
987  */
988 static void do_generic_file_read(struct file *filp, loff_t *ppos,
989                 read_descriptor_t *desc, read_actor_t actor)
990 {
991         struct address_space *mapping = filp->f_mapping;
992         struct inode *inode = mapping->host;
993         struct file_ra_state *ra = &filp->f_ra;
994         pgoff_t index;
995         pgoff_t last_index;
996         pgoff_t prev_index;
997         unsigned long offset;      /* offset into pagecache page */
998         unsigned int prev_offset;
999         int error;
1000
1001         index = *ppos >> PAGE_CACHE_SHIFT;
1002         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1003         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1004         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1005         offset = *ppos & ~PAGE_CACHE_MASK;
1006
1007         for (;;) {
1008                 struct page *page;
1009                 pgoff_t end_index;
1010                 loff_t isize;
1011                 unsigned long nr, ret;
1012
1013                 cond_resched();
1014 find_page:
1015                 page = find_get_page(mapping, index);
1016                 if (!page) {
1017                         page_cache_sync_readahead(mapping,
1018                                         ra, filp,
1019                                         index, last_index - index);
1020                         page = find_get_page(mapping, index);
1021                         if (unlikely(page == NULL))
1022                                 goto no_cached_page;
1023                 }
1024                 if (PageReadahead(page)) {
1025                         page_cache_async_readahead(mapping,
1026                                         ra, filp, page,
1027                                         index, last_index - index);
1028                 }
1029                 if (!PageUptodate(page)) {
1030                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1031                                         !mapping->a_ops->is_partially_uptodate)
1032                                 goto page_not_up_to_date;
1033                         if (!trylock_page(page))
1034                                 goto page_not_up_to_date;
1035                         /* Did it get truncated before we got the lock? */
1036                         if (!page->mapping)
1037                                 goto page_not_up_to_date_locked;
1038                         if (!mapping->a_ops->is_partially_uptodate(page,
1039                                                                 desc, offset))
1040                                 goto page_not_up_to_date_locked;
1041                         unlock_page(page);
1042                 }
1043 page_ok:
1044                 /*
1045                  * i_size must be checked after we know the page is Uptodate.
1046                  *
1047                  * Checking i_size after the check allows us to calculate
1048                  * the correct value for "nr", which means the zero-filled
1049                  * part of the page is not copied back to userspace (unless
1050                  * another truncate extends the file - this is desired though).
1051                  */
1052
1053                 isize = i_size_read(inode);
1054                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1055                 if (unlikely(!isize || index > end_index)) {
1056                         page_cache_release(page);
1057                         goto out;
1058                 }
1059
1060                 /* nr is the maximum number of bytes to copy from this page */
1061                 nr = PAGE_CACHE_SIZE;
1062                 if (index == end_index) {
1063                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1064                         if (nr <= offset) {
1065                                 page_cache_release(page);
1066                                 goto out;
1067                         }
1068                 }
1069                 nr = nr - offset;
1070
1071                 /* If users can be writing to this page using arbitrary
1072                  * virtual addresses, take care about potential aliasing
1073                  * before reading the page on the kernel side.
1074                  */
1075                 if (mapping_writably_mapped(mapping))
1076                         flush_dcache_page(page);
1077
1078                 /*
1079                  * When a sequential read accesses a page several times,
1080                  * only mark it as accessed the first time.
1081                  */
1082                 if (prev_index != index || offset != prev_offset)
1083                         mark_page_accessed(page);
1084                 prev_index = index;
1085
1086                 /*
1087                  * Ok, we have the page, and it's up-to-date, so
1088                  * now we can copy it to user space...
1089                  *
1090                  * The actor routine returns how many bytes were actually used..
1091                  * NOTE! This may not be the same as how much of a user buffer
1092                  * we filled up (we may be padding etc), so we can only update
1093                  * "pos" here (the actor routine has to update the user buffer
1094                  * pointers and the remaining count).
1095                  */
1096                 ret = actor(desc, page, offset, nr);
1097                 offset += ret;
1098                 index += offset >> PAGE_CACHE_SHIFT;
1099                 offset &= ~PAGE_CACHE_MASK;
1100                 prev_offset = offset;
1101
1102                 page_cache_release(page);
1103                 if (ret == nr && desc->count)
1104                         continue;
1105                 goto out;
1106
1107 page_not_up_to_date:
1108                 /* Get exclusive access to the page ... */
1109                 error = lock_page_killable(page);
1110                 if (unlikely(error))
1111                         goto readpage_error;
1112
1113 page_not_up_to_date_locked:
1114                 /* Did it get truncated before we got the lock? */
1115                 if (!page->mapping) {
1116                         unlock_page(page);
1117                         page_cache_release(page);
1118                         continue;
1119                 }
1120
1121                 /* Did somebody else fill it already? */
1122                 if (PageUptodate(page)) {
1123                         unlock_page(page);
1124                         goto page_ok;
1125                 }
1126
1127 readpage:
1128                 /*
1129                  * A previous I/O error may have been due to temporary
1130                  * failures, eg. multipath errors.
1131                  * PG_error will be set again if readpage fails.
1132                  */
1133                 ClearPageError(page);
1134                 /* Start the actual read. The read will unlock the page. */
1135                 error = mapping->a_ops->readpage(filp, page);
1136
1137                 if (unlikely(error)) {
1138                         if (error == AOP_TRUNCATED_PAGE) {
1139                                 page_cache_release(page);
1140                                 goto find_page;
1141                         }
1142                         goto readpage_error;
1143                 }
1144
1145                 if (!PageUptodate(page)) {
1146                         error = lock_page_killable(page);
1147                         if (unlikely(error))
1148                                 goto readpage_error;
1149                         if (!PageUptodate(page)) {
1150                                 if (page->mapping == NULL) {
1151                                         /*
1152                                          * invalidate_mapping_pages got it
1153                                          */
1154                                         unlock_page(page);
1155                                         page_cache_release(page);
1156                                         goto find_page;
1157                                 }
1158                                 unlock_page(page);
1159                                 shrink_readahead_size_eio(filp, ra);
1160                                 error = -EIO;
1161                                 goto readpage_error;
1162                         }
1163                         unlock_page(page);
1164                 }
1165
1166                 goto page_ok;
1167
1168 readpage_error:
1169                 /* UHHUH! A synchronous read error occurred. Report it */
1170                 desc->error = error;
1171                 page_cache_release(page);
1172                 goto out;
1173
1174 no_cached_page:
1175                 /*
1176                  * Ok, it wasn't cached, so we need to create a new
1177                  * page..
1178                  */
1179                 page = page_cache_alloc_cold(mapping);
1180                 if (!page) {
1181                         desc->error = -ENOMEM;
1182                         goto out;
1183                 }
1184                 error = add_to_page_cache_lru(page, mapping,
1185                                                 index, GFP_KERNEL);
1186                 if (error) {
1187                         page_cache_release(page);
1188                         if (error == -EEXIST)
1189                                 goto find_page;
1190                         desc->error = error;
1191                         goto out;
1192                 }
1193                 goto readpage;
1194         }
1195
1196 out:
1197         ra->prev_pos = prev_index;
1198         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1199         ra->prev_pos |= prev_offset;
1200
1201         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1202         file_accessed(filp);
1203 }
1204
1205 int file_read_actor(read_descriptor_t *desc, struct page *page,
1206                         unsigned long offset, unsigned long size)
1207 {
1208         char *kaddr;
1209         unsigned long left, count = desc->count;
1210
1211         if (size > count)
1212                 size = count;
1213
1214         /*
1215          * Faults on the destination of a read are common, so do it before
1216          * taking the kmap.
1217          */
1218         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1219                 kaddr = kmap_atomic(page, KM_USER0);
1220                 left = __copy_to_user_inatomic(desc->arg.buf,
1221                                                 kaddr + offset, size);
1222                 kunmap_atomic(kaddr, KM_USER0);
1223                 if (left == 0)
1224                         goto success;
1225         }
1226
1227         /* Do it the slow way */
1228         kaddr = kmap(page);
1229         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1230         kunmap(page);
1231
1232         if (left) {
1233                 size -= left;
1234                 desc->error = -EFAULT;
1235         }
1236 success:
1237         desc->count = count - size;
1238         desc->written += size;
1239         desc->arg.buf += size;
1240         return size;
1241 }
1242
1243 /*
1244  * Performs necessary checks before doing a write
1245  * @iov:        io vector request
1246  * @nr_segs:    number of segments in the iovec
1247  * @count:      number of bytes to write
1248  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1249  *
1250  * Adjust number of segments and amount of bytes to write (nr_segs should be
1251  * properly initialized first). Returns appropriate error code that caller
1252  * should return or zero in case that write should be allowed.
1253  */
1254 int generic_segment_checks(const struct iovec *iov,
1255                         unsigned long *nr_segs, size_t *count, int access_flags)
1256 {
1257         unsigned long   seg;
1258         size_t cnt = 0;
1259         for (seg = 0; seg < *nr_segs; seg++) {
1260                 const struct iovec *iv = &iov[seg];
1261
1262                 /*
1263                  * If any segment has a negative length, or the cumulative
1264                  * length ever wraps negative then return -EINVAL.
1265                  */
1266                 cnt += iv->iov_len;
1267                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1268                         return -EINVAL;
1269                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1270                         continue;
1271                 if (seg == 0)
1272                         return -EFAULT;
1273                 *nr_segs = seg;
1274                 cnt -= iv->iov_len;     /* This segment is no good */
1275                 break;
1276         }
1277         *count = cnt;
1278         return 0;
1279 }
1280 EXPORT_SYMBOL(generic_segment_checks);
1281
1282 /**
1283  * generic_file_aio_read - generic filesystem read routine
1284  * @iocb:       kernel I/O control block
1285  * @iov:        io vector request
1286  * @nr_segs:    number of segments in the iovec
1287  * @pos:        current file position
1288  *
1289  * This is the "read()" routine for all filesystems
1290  * that can use the page cache directly.
1291  */
1292 ssize_t
1293 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1294                 unsigned long nr_segs, loff_t pos)
1295 {
1296         struct file *filp = iocb->ki_filp;
1297         ssize_t retval;
1298         unsigned long seg = 0;
1299         size_t count;
1300         loff_t *ppos = &iocb->ki_pos;
1301
1302         count = 0;
1303         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1304         if (retval)
1305                 return retval;
1306
1307         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1308         if (filp->f_flags & O_DIRECT) {
1309                 loff_t size;
1310                 struct address_space *mapping;
1311                 struct inode *inode;
1312
1313                 mapping = filp->f_mapping;
1314                 inode = mapping->host;
1315                 if (!count)
1316                         goto out; /* skip atime */
1317                 size = i_size_read(inode);
1318                 if (pos < size) {
1319                         retval = filemap_write_and_wait_range(mapping, pos,
1320                                         pos + iov_length(iov, nr_segs) - 1);
1321                         if (!retval) {
1322                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1323                                                         iov, pos, nr_segs);
1324                         }
1325                         if (retval > 0) {
1326                                 *ppos = pos + retval;
1327                                 count -= retval;
1328                         }
1329
1330                         /*
1331                          * Btrfs can have a short DIO read if we encounter
1332                          * compressed extents, so if there was an error, or if
1333                          * we've already read everything we wanted to, or if
1334                          * there was a short read because we hit EOF, go ahead
1335                          * and return.  Otherwise fallthrough to buffered io for
1336                          * the rest of the read.
1337                          */
1338                         if (retval < 0 || !count || *ppos >= size) {
1339                                 file_accessed(filp);
1340                                 goto out;
1341                         }
1342                 }
1343         }
1344
1345         count = retval;
1346         for (seg = 0; seg < nr_segs; seg++) {
1347                 read_descriptor_t desc;
1348                 loff_t offset = 0;
1349
1350                 /*
1351                  * If we did a short DIO read we need to skip the section of the
1352                  * iov that we've already read data into.
1353                  */
1354                 if (count) {
1355                         if (count > iov[seg].iov_len) {
1356                                 count -= iov[seg].iov_len;
1357                                 continue;
1358                         }
1359                         offset = count;
1360                         count = 0;
1361                 }
1362
1363                 desc.written = 0;
1364                 desc.arg.buf = iov[seg].iov_base + offset;
1365                 desc.count = iov[seg].iov_len - offset;
1366                 if (desc.count == 0)
1367                         continue;
1368                 desc.error = 0;
1369                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1370                 retval += desc.written;
1371                 if (desc.error) {
1372                         retval = retval ?: desc.error;
1373                         break;
1374                 }
1375                 if (desc.count > 0)
1376                         break;
1377         }
1378 out:
1379         return retval;
1380 }
1381 EXPORT_SYMBOL(generic_file_aio_read);
1382
1383 static ssize_t
1384 do_readahead(struct address_space *mapping, struct file *filp,
1385              pgoff_t index, unsigned long nr)
1386 {
1387         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1388                 return -EINVAL;
1389
1390         force_page_cache_readahead(mapping, filp, index, nr);
1391         return 0;
1392 }
1393
1394 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1395 {
1396         ssize_t ret;
1397         struct file *file;
1398
1399         ret = -EBADF;
1400         file = fget(fd);
1401         if (file) {
1402                 if (file->f_mode & FMODE_READ) {
1403                         struct address_space *mapping = file->f_mapping;
1404                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1405                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1406                         unsigned long len = end - start + 1;
1407                         ret = do_readahead(mapping, file, start, len);
1408                 }
1409                 fput(file);
1410         }
1411         return ret;
1412 }
1413 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1414 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1415 {
1416         return SYSC_readahead((int) fd, offset, (size_t) count);
1417 }
1418 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1419 #endif
1420
1421 #ifdef CONFIG_MMU
1422 /**
1423  * page_cache_read - adds requested page to the page cache if not already there
1424  * @file:       file to read
1425  * @offset:     page index
1426  *
1427  * This adds the requested page to the page cache if it isn't already there,
1428  * and schedules an I/O to read in its contents from disk.
1429  */
1430 static int page_cache_read(struct file *file, pgoff_t offset)
1431 {
1432         struct address_space *mapping = file->f_mapping;
1433         struct page *page; 
1434         int ret;
1435
1436         do {
1437                 page = page_cache_alloc_cold(mapping);
1438                 if (!page)
1439                         return -ENOMEM;
1440
1441                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1442                 if (ret == 0)
1443                         ret = mapping->a_ops->readpage(file, page);
1444                 else if (ret == -EEXIST)
1445                         ret = 0; /* losing race to add is OK */
1446
1447                 page_cache_release(page);
1448
1449         } while (ret == AOP_TRUNCATED_PAGE);
1450                 
1451         return ret;
1452 }
1453
1454 #define MMAP_LOTSAMISS  (100)
1455
1456 /*
1457  * Synchronous readahead happens when we don't even find
1458  * a page in the page cache at all.
1459  */
1460 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1461                                    struct file_ra_state *ra,
1462                                    struct file *file,
1463                                    pgoff_t offset)
1464 {
1465         unsigned long ra_pages;
1466         struct address_space *mapping = file->f_mapping;
1467
1468         /* If we don't want any read-ahead, don't bother */
1469         if (VM_RandomReadHint(vma))
1470                 return;
1471
1472         if (VM_SequentialReadHint(vma) ||
1473                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1474                 page_cache_sync_readahead(mapping, ra, file, offset,
1475                                           ra->ra_pages);
1476                 return;
1477         }
1478
1479         if (ra->mmap_miss < INT_MAX)
1480                 ra->mmap_miss++;
1481
1482         /*
1483          * Do we miss much more than hit in this file? If so,
1484          * stop bothering with read-ahead. It will only hurt.
1485          */
1486         if (ra->mmap_miss > MMAP_LOTSAMISS)
1487                 return;
1488
1489         /*
1490          * mmap read-around
1491          */
1492         ra_pages = max_sane_readahead(ra->ra_pages);
1493         if (ra_pages) {
1494                 ra->start = max_t(long, 0, offset - ra_pages/2);
1495                 ra->size = ra_pages;
1496                 ra->async_size = 0;
1497                 ra_submit(ra, mapping, file);
1498         }
1499 }
1500
1501 /*
1502  * Asynchronous readahead happens when we find the page and PG_readahead,
1503  * so we want to possibly extend the readahead further..
1504  */
1505 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1506                                     struct file_ra_state *ra,
1507                                     struct file *file,
1508                                     struct page *page,
1509                                     pgoff_t offset)
1510 {
1511         struct address_space *mapping = file->f_mapping;
1512
1513         /* If we don't want any read-ahead, don't bother */
1514         if (VM_RandomReadHint(vma))
1515                 return;
1516         if (ra->mmap_miss > 0)
1517                 ra->mmap_miss--;
1518         if (PageReadahead(page))
1519                 page_cache_async_readahead(mapping, ra, file,
1520                                            page, offset, ra->ra_pages);
1521 }
1522
1523 /**
1524  * filemap_fault - read in file data for page fault handling
1525  * @vma:        vma in which the fault was taken
1526  * @vmf:        struct vm_fault containing details of the fault
1527  *
1528  * filemap_fault() is invoked via the vma operations vector for a
1529  * mapped memory region to read in file data during a page fault.
1530  *
1531  * The goto's are kind of ugly, but this streamlines the normal case of having
1532  * it in the page cache, and handles the special cases reasonably without
1533  * having a lot of duplicated code.
1534  */
1535 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1536 {
1537         int error;
1538         struct file *file = vma->vm_file;
1539         struct address_space *mapping = file->f_mapping;
1540         struct file_ra_state *ra = &file->f_ra;
1541         struct inode *inode = mapping->host;
1542         pgoff_t offset = vmf->pgoff;
1543         struct page *page;
1544         pgoff_t size;
1545         int ret = 0;
1546
1547         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1548         if (offset >= size)
1549                 return VM_FAULT_SIGBUS;
1550
1551         /*
1552          * Do we have something in the page cache already?
1553          */
1554         page = find_get_page(mapping, offset);
1555         if (likely(page)) {
1556                 /*
1557                  * We found the page, so try async readahead before
1558                  * waiting for the lock.
1559                  */
1560                 do_async_mmap_readahead(vma, ra, file, page, offset);
1561         } else {
1562                 /* No page in the page cache at all */
1563                 do_sync_mmap_readahead(vma, ra, file, offset);
1564                 count_vm_event(PGMAJFAULT);
1565                 ret = VM_FAULT_MAJOR;
1566 retry_find:
1567                 page = find_get_page(mapping, offset);
1568                 if (!page)
1569                         goto no_cached_page;
1570         }
1571
1572         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1573                 page_cache_release(page);
1574                 return ret | VM_FAULT_RETRY;
1575         }
1576
1577         /* Did it get truncated? */
1578         if (unlikely(page->mapping != mapping)) {
1579                 unlock_page(page);
1580                 put_page(page);
1581                 goto retry_find;
1582         }
1583         VM_BUG_ON(page->index != offset);
1584
1585         /*
1586          * We have a locked page in the page cache, now we need to check
1587          * that it's up-to-date. If not, it is going to be due to an error.
1588          */
1589         if (unlikely(!PageUptodate(page)))
1590                 goto page_not_uptodate;
1591
1592         /*
1593          * Found the page and have a reference on it.
1594          * We must recheck i_size under page lock.
1595          */
1596         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1597         if (unlikely(offset >= size)) {
1598                 unlock_page(page);
1599                 page_cache_release(page);
1600                 return VM_FAULT_SIGBUS;
1601         }
1602
1603         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1604         vmf->page = page;
1605         return ret | VM_FAULT_LOCKED;
1606
1607 no_cached_page:
1608         /*
1609          * We're only likely to ever get here if MADV_RANDOM is in
1610          * effect.
1611          */
1612         error = page_cache_read(file, offset);
1613
1614         /*
1615          * The page we want has now been added to the page cache.
1616          * In the unlikely event that someone removed it in the
1617          * meantime, we'll just come back here and read it again.
1618          */
1619         if (error >= 0)
1620                 goto retry_find;
1621
1622         /*
1623          * An error return from page_cache_read can result if the
1624          * system is low on memory, or a problem occurs while trying
1625          * to schedule I/O.
1626          */
1627         if (error == -ENOMEM)
1628                 return VM_FAULT_OOM;
1629         return VM_FAULT_SIGBUS;
1630
1631 page_not_uptodate:
1632         /*
1633          * Umm, take care of errors if the page isn't up-to-date.
1634          * Try to re-read it _once_. We do this synchronously,
1635          * because there really aren't any performance issues here
1636          * and we need to check for errors.
1637          */
1638         ClearPageError(page);
1639         error = mapping->a_ops->readpage(file, page);
1640         if (!error) {
1641                 wait_on_page_locked(page);
1642                 if (!PageUptodate(page))
1643                         error = -EIO;
1644         }
1645         page_cache_release(page);
1646
1647         if (!error || error == AOP_TRUNCATED_PAGE)
1648                 goto retry_find;
1649
1650         /* Things didn't work out. Return zero to tell the mm layer so. */
1651         shrink_readahead_size_eio(file, ra);
1652         return VM_FAULT_SIGBUS;
1653 }
1654 EXPORT_SYMBOL(filemap_fault);
1655
1656 const struct vm_operations_struct generic_file_vm_ops = {
1657         .fault          = filemap_fault,
1658 };
1659
1660 /* This is used for a general mmap of a disk file */
1661
1662 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1663 {
1664         struct address_space *mapping = file->f_mapping;
1665
1666         if (!mapping->a_ops->readpage)
1667                 return -ENOEXEC;
1668         file_accessed(file);
1669         vma->vm_ops = &generic_file_vm_ops;
1670         vma->vm_flags |= VM_CAN_NONLINEAR;
1671         return 0;
1672 }
1673
1674 /*
1675  * This is for filesystems which do not implement ->writepage.
1676  */
1677 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1678 {
1679         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1680                 return -EINVAL;
1681         return generic_file_mmap(file, vma);
1682 }
1683 #else
1684 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1685 {
1686         return -ENOSYS;
1687 }
1688 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1689 {
1690         return -ENOSYS;
1691 }
1692 #endif /* CONFIG_MMU */
1693
1694 EXPORT_SYMBOL(generic_file_mmap);
1695 EXPORT_SYMBOL(generic_file_readonly_mmap);
1696
1697 static struct page *__read_cache_page(struct address_space *mapping,
1698                                 pgoff_t index,
1699                                 int (*filler)(void *,struct page*),
1700                                 void *data,
1701                                 gfp_t gfp)
1702 {
1703         struct page *page;
1704         int err;
1705 repeat:
1706         page = find_get_page(mapping, index);
1707         if (!page) {
1708                 page = __page_cache_alloc(gfp | __GFP_COLD);
1709                 if (!page)
1710                         return ERR_PTR(-ENOMEM);
1711                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1712                 if (unlikely(err)) {
1713                         page_cache_release(page);
1714                         if (err == -EEXIST)
1715                                 goto repeat;
1716                         /* Presumably ENOMEM for radix tree node */
1717                         return ERR_PTR(err);
1718                 }
1719                 err = filler(data, page);
1720                 if (err < 0) {
1721                         page_cache_release(page);
1722                         page = ERR_PTR(err);
1723                 }
1724         }
1725         return page;
1726 }
1727
1728 static struct page *do_read_cache_page(struct address_space *mapping,
1729                                 pgoff_t index,
1730                                 int (*filler)(void *,struct page*),
1731                                 void *data,
1732                                 gfp_t gfp)
1733
1734 {
1735         struct page *page;
1736         int err;
1737
1738 retry:
1739         page = __read_cache_page(mapping, index, filler, data, gfp);
1740         if (IS_ERR(page))
1741                 return page;
1742         if (PageUptodate(page))
1743                 goto out;
1744
1745         lock_page(page);
1746         if (!page->mapping) {
1747                 unlock_page(page);
1748                 page_cache_release(page);
1749                 goto retry;
1750         }
1751         if (PageUptodate(page)) {
1752                 unlock_page(page);
1753                 goto out;
1754         }
1755         err = filler(data, page);
1756         if (err < 0) {
1757                 page_cache_release(page);
1758                 return ERR_PTR(err);
1759         }
1760 out:
1761         mark_page_accessed(page);
1762         return page;
1763 }
1764
1765 /**
1766  * read_cache_page_async - read into page cache, fill it if needed
1767  * @mapping:    the page's address_space
1768  * @index:      the page index
1769  * @filler:     function to perform the read
1770  * @data:       destination for read data
1771  *
1772  * Same as read_cache_page, but don't wait for page to become unlocked
1773  * after submitting it to the filler.
1774  *
1775  * Read into the page cache. If a page already exists, and PageUptodate() is
1776  * not set, try to fill the page but don't wait for it to become unlocked.
1777  *
1778  * If the page does not get brought uptodate, return -EIO.
1779  */
1780 struct page *read_cache_page_async(struct address_space *mapping,
1781                                 pgoff_t index,
1782                                 int (*filler)(void *,struct page*),
1783                                 void *data)
1784 {
1785         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1786 }
1787 EXPORT_SYMBOL(read_cache_page_async);
1788
1789 static struct page *wait_on_page_read(struct page *page)
1790 {
1791         if (!IS_ERR(page)) {
1792                 wait_on_page_locked(page);
1793                 if (!PageUptodate(page)) {
1794                         page_cache_release(page);
1795                         page = ERR_PTR(-EIO);
1796                 }
1797         }
1798         return page;
1799 }
1800
1801 /**
1802  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1803  * @mapping:    the page's address_space
1804  * @index:      the page index
1805  * @gfp:        the page allocator flags to use if allocating
1806  *
1807  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1808  * any new page allocations done using the specified allocation flags. Note
1809  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1810  * expect to do this atomically or anything like that - but you can pass in
1811  * other page requirements.
1812  *
1813  * If the page does not get brought uptodate, return -EIO.
1814  */
1815 struct page *read_cache_page_gfp(struct address_space *mapping,
1816                                 pgoff_t index,
1817                                 gfp_t gfp)
1818 {
1819         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1820
1821         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1822 }
1823 EXPORT_SYMBOL(read_cache_page_gfp);
1824
1825 /**
1826  * read_cache_page - read into page cache, fill it if needed
1827  * @mapping:    the page's address_space
1828  * @index:      the page index
1829  * @filler:     function to perform the read
1830  * @data:       destination for read data
1831  *
1832  * Read into the page cache. If a page already exists, and PageUptodate() is
1833  * not set, try to fill the page then wait for it to become unlocked.
1834  *
1835  * If the page does not get brought uptodate, return -EIO.
1836  */
1837 struct page *read_cache_page(struct address_space *mapping,
1838                                 pgoff_t index,
1839                                 int (*filler)(void *,struct page*),
1840                                 void *data)
1841 {
1842         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1843 }
1844 EXPORT_SYMBOL(read_cache_page);
1845
1846 /*
1847  * The logic we want is
1848  *
1849  *      if suid or (sgid and xgrp)
1850  *              remove privs
1851  */
1852 int should_remove_suid(struct dentry *dentry)
1853 {
1854         mode_t mode = dentry->d_inode->i_mode;
1855         int kill = 0;
1856
1857         /* suid always must be killed */
1858         if (unlikely(mode & S_ISUID))
1859                 kill = ATTR_KILL_SUID;
1860
1861         /*
1862          * sgid without any exec bits is just a mandatory locking mark; leave
1863          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1864          */
1865         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1866                 kill |= ATTR_KILL_SGID;
1867
1868         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1869                 return kill;
1870
1871         return 0;
1872 }
1873 EXPORT_SYMBOL(should_remove_suid);
1874
1875 static int __remove_suid(struct dentry *dentry, int kill)
1876 {
1877         struct iattr newattrs;
1878
1879         newattrs.ia_valid = ATTR_FORCE | kill;
1880         return notify_change(dentry, &newattrs);
1881 }
1882
1883 int file_remove_suid(struct file *file)
1884 {
1885         struct dentry *dentry = file->f_path.dentry;
1886         int killsuid = should_remove_suid(dentry);
1887         int killpriv = security_inode_need_killpriv(dentry);
1888         int error = 0;
1889
1890         if (killpriv < 0)
1891                 return killpriv;
1892         if (killpriv)
1893                 error = security_inode_killpriv(dentry);
1894         if (!error && killsuid)
1895                 error = __remove_suid(dentry, killsuid);
1896
1897         return error;
1898 }
1899 EXPORT_SYMBOL(file_remove_suid);
1900
1901 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1902                         const struct iovec *iov, size_t base, size_t bytes)
1903 {
1904         size_t copied = 0, left = 0;
1905
1906         while (bytes) {
1907                 char __user *buf = iov->iov_base + base;
1908                 int copy = min(bytes, iov->iov_len - base);
1909
1910                 base = 0;
1911                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1912                 copied += copy;
1913                 bytes -= copy;
1914                 vaddr += copy;
1915                 iov++;
1916
1917                 if (unlikely(left))
1918                         break;
1919         }
1920         return copied - left;
1921 }
1922
1923 /*
1924  * Copy as much as we can into the page and return the number of bytes which
1925  * were successfully copied.  If a fault is encountered then return the number of
1926  * bytes which were copied.
1927  */
1928 size_t iov_iter_copy_from_user_atomic(struct page *page,
1929                 struct iov_iter *i, unsigned long offset, size_t bytes)
1930 {
1931         char *kaddr;
1932         size_t copied;
1933
1934         BUG_ON(!in_atomic());
1935         kaddr = kmap_atomic(page, KM_USER0);
1936         if (likely(i->nr_segs == 1)) {
1937                 int left;
1938                 char __user *buf = i->iov->iov_base + i->iov_offset;
1939                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1940                 copied = bytes - left;
1941         } else {
1942                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1943                                                 i->iov, i->iov_offset, bytes);
1944         }
1945         kunmap_atomic(kaddr, KM_USER0);
1946
1947         return copied;
1948 }
1949 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1950
1951 /*
1952  * This has the same sideeffects and return value as
1953  * iov_iter_copy_from_user_atomic().
1954  * The difference is that it attempts to resolve faults.
1955  * Page must not be locked.
1956  */
1957 size_t iov_iter_copy_from_user(struct page *page,
1958                 struct iov_iter *i, unsigned long offset, size_t bytes)
1959 {
1960         char *kaddr;
1961         size_t copied;
1962
1963         kaddr = kmap(page);
1964         if (likely(i->nr_segs == 1)) {
1965                 int left;
1966                 char __user *buf = i->iov->iov_base + i->iov_offset;
1967                 left = __copy_from_user(kaddr + offset, buf, bytes);
1968                 copied = bytes - left;
1969         } else {
1970                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1971                                                 i->iov, i->iov_offset, bytes);
1972         }
1973         kunmap(page);
1974         return copied;
1975 }
1976 EXPORT_SYMBOL(iov_iter_copy_from_user);
1977
1978 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1979 {
1980         BUG_ON(i->count < bytes);
1981
1982         if (likely(i->nr_segs == 1)) {
1983                 i->iov_offset += bytes;
1984                 i->count -= bytes;
1985         } else {
1986                 const struct iovec *iov = i->iov;
1987                 size_t base = i->iov_offset;
1988
1989                 /*
1990                  * The !iov->iov_len check ensures we skip over unlikely
1991                  * zero-length segments (without overruning the iovec).
1992                  */
1993                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1994                         int copy;
1995
1996                         copy = min(bytes, iov->iov_len - base);
1997                         BUG_ON(!i->count || i->count < copy);
1998                         i->count -= copy;
1999                         bytes -= copy;
2000                         base += copy;
2001                         if (iov->iov_len == base) {
2002                                 iov++;
2003                                 base = 0;
2004                         }
2005                 }
2006                 i->iov = iov;
2007                 i->iov_offset = base;
2008         }
2009 }
2010 EXPORT_SYMBOL(iov_iter_advance);
2011
2012 /*
2013  * Fault in the first iovec of the given iov_iter, to a maximum length
2014  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2015  * accessed (ie. because it is an invalid address).
2016  *
2017  * writev-intensive code may want this to prefault several iovecs -- that
2018  * would be possible (callers must not rely on the fact that _only_ the
2019  * first iovec will be faulted with the current implementation).
2020  */
2021 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2022 {
2023         char __user *buf = i->iov->iov_base + i->iov_offset;
2024         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2025         return fault_in_pages_readable(buf, bytes);
2026 }
2027 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2028
2029 /*
2030  * Return the count of just the current iov_iter segment.
2031  */
2032 size_t iov_iter_single_seg_count(struct iov_iter *i)
2033 {
2034         const struct iovec *iov = i->iov;
2035         if (i->nr_segs == 1)
2036                 return i->count;
2037         else
2038                 return min(i->count, iov->iov_len - i->iov_offset);
2039 }
2040 EXPORT_SYMBOL(iov_iter_single_seg_count);
2041
2042 /*
2043  * Performs necessary checks before doing a write
2044  *
2045  * Can adjust writing position or amount of bytes to write.
2046  * Returns appropriate error code that caller should return or
2047  * zero in case that write should be allowed.
2048  */
2049 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2050 {
2051         struct inode *inode = file->f_mapping->host;
2052         unsigned long limit = rlimit(RLIMIT_FSIZE);
2053
2054         if (unlikely(*pos < 0))
2055                 return -EINVAL;
2056
2057         if (!isblk) {
2058                 /* FIXME: this is for backwards compatibility with 2.4 */
2059                 if (file->f_flags & O_APPEND)
2060                         *pos = i_size_read(inode);
2061
2062                 if (limit != RLIM_INFINITY) {
2063                         if (*pos >= limit) {
2064                                 send_sig(SIGXFSZ, current, 0);
2065                                 return -EFBIG;
2066                         }
2067                         if (*count > limit - (typeof(limit))*pos) {
2068                                 *count = limit - (typeof(limit))*pos;
2069                         }
2070                 }
2071         }
2072
2073         /*
2074          * LFS rule
2075          */
2076         if (unlikely(*pos + *count > MAX_NON_LFS &&
2077                                 !(file->f_flags & O_LARGEFILE))) {
2078                 if (*pos >= MAX_NON_LFS) {
2079                         return -EFBIG;
2080                 }
2081                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2082                         *count = MAX_NON_LFS - (unsigned long)*pos;
2083                 }
2084         }
2085
2086         /*
2087          * Are we about to exceed the fs block limit ?
2088          *
2089          * If we have written data it becomes a short write.  If we have
2090          * exceeded without writing data we send a signal and return EFBIG.
2091          * Linus frestrict idea will clean these up nicely..
2092          */
2093         if (likely(!isblk)) {
2094                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2095                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2096                                 return -EFBIG;
2097                         }
2098                         /* zero-length writes at ->s_maxbytes are OK */
2099                 }
2100
2101                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2102                         *count = inode->i_sb->s_maxbytes - *pos;
2103         } else {
2104 #ifdef CONFIG_BLOCK
2105                 loff_t isize;
2106                 if (bdev_read_only(I_BDEV(inode)))
2107                         return -EPERM;
2108                 isize = i_size_read(inode);
2109                 if (*pos >= isize) {
2110                         if (*count || *pos > isize)
2111                                 return -ENOSPC;
2112                 }
2113
2114                 if (*pos + *count > isize)
2115                         *count = isize - *pos;
2116 #else
2117                 return -EPERM;
2118 #endif
2119         }
2120         return 0;
2121 }
2122 EXPORT_SYMBOL(generic_write_checks);
2123
2124 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2125                                 loff_t pos, unsigned len, unsigned flags,
2126                                 struct page **pagep, void **fsdata)
2127 {
2128         const struct address_space_operations *aops = mapping->a_ops;
2129
2130         return aops->write_begin(file, mapping, pos, len, flags,
2131                                                         pagep, fsdata);
2132 }
2133 EXPORT_SYMBOL(pagecache_write_begin);
2134
2135 int pagecache_write_end(struct file *file, struct address_space *mapping,
2136                                 loff_t pos, unsigned len, unsigned copied,
2137                                 struct page *page, void *fsdata)
2138 {
2139         const struct address_space_operations *aops = mapping->a_ops;
2140
2141         mark_page_accessed(page);
2142         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2143 }
2144 EXPORT_SYMBOL(pagecache_write_end);
2145
2146 ssize_t
2147 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2148                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2149                 size_t count, size_t ocount)
2150 {
2151         struct file     *file = iocb->ki_filp;
2152         struct address_space *mapping = file->f_mapping;
2153         struct inode    *inode = mapping->host;
2154         ssize_t         written;
2155         size_t          write_len;
2156         pgoff_t         end;
2157
2158         if (count != ocount)
2159                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2160
2161         write_len = iov_length(iov, *nr_segs);
2162         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2163
2164         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2165         if (written)
2166                 goto out;
2167
2168         /*
2169          * After a write we want buffered reads to be sure to go to disk to get
2170          * the new data.  We invalidate clean cached page from the region we're
2171          * about to write.  We do this *before* the write so that we can return
2172          * without clobbering -EIOCBQUEUED from ->direct_IO().
2173          */
2174         if (mapping->nrpages) {
2175                 written = invalidate_inode_pages2_range(mapping,
2176                                         pos >> PAGE_CACHE_SHIFT, end);
2177                 /*
2178                  * If a page can not be invalidated, return 0 to fall back
2179                  * to buffered write.
2180                  */
2181                 if (written) {
2182                         if (written == -EBUSY)
2183                                 return 0;
2184                         goto out;
2185                 }
2186         }
2187
2188         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2189
2190         /*
2191          * Finally, try again to invalidate clean pages which might have been
2192          * cached by non-direct readahead, or faulted in by get_user_pages()
2193          * if the source of the write was an mmap'ed region of the file
2194          * we're writing.  Either one is a pretty crazy thing to do,
2195          * so we don't support it 100%.  If this invalidation
2196          * fails, tough, the write still worked...
2197          */
2198         if (mapping->nrpages) {
2199                 invalidate_inode_pages2_range(mapping,
2200                                               pos >> PAGE_CACHE_SHIFT, end);
2201         }
2202
2203         if (written > 0) {
2204                 pos += written;
2205                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2206                         i_size_write(inode, pos);
2207                         mark_inode_dirty(inode);
2208                 }
2209                 *ppos = pos;
2210         }
2211 out:
2212         return written;
2213 }
2214 EXPORT_SYMBOL(generic_file_direct_write);
2215
2216 /*
2217  * Find or create a page at the given pagecache position. Return the locked
2218  * page. This function is specifically for buffered writes.
2219  */
2220 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2221                                         pgoff_t index, unsigned flags)
2222 {
2223         int status;
2224         struct page *page;
2225         gfp_t gfp_notmask = 0;
2226         if (flags & AOP_FLAG_NOFS)
2227                 gfp_notmask = __GFP_FS;
2228 repeat:
2229         page = find_lock_page(mapping, index);
2230         if (page)
2231                 return page;
2232
2233         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2234         if (!page)
2235                 return NULL;
2236         status = add_to_page_cache_lru(page, mapping, index,
2237                                                 GFP_KERNEL & ~gfp_notmask);
2238         if (unlikely(status)) {
2239                 page_cache_release(page);
2240                 if (status == -EEXIST)
2241                         goto repeat;
2242                 return NULL;
2243         }
2244         return page;
2245 }
2246 EXPORT_SYMBOL(grab_cache_page_write_begin);
2247
2248 static ssize_t generic_perform_write(struct file *file,
2249                                 struct iov_iter *i, loff_t pos)
2250 {
2251         struct address_space *mapping = file->f_mapping;
2252         const struct address_space_operations *a_ops = mapping->a_ops;
2253         long status = 0;
2254         ssize_t written = 0;
2255         unsigned int flags = 0;
2256
2257         /*
2258          * Copies from kernel address space cannot fail (NFSD is a big user).
2259          */
2260         if (segment_eq(get_fs(), KERNEL_DS))
2261                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2262
2263         do {
2264                 struct page *page;
2265                 unsigned long offset;   /* Offset into pagecache page */
2266                 unsigned long bytes;    /* Bytes to write to page */
2267                 size_t copied;          /* Bytes copied from user */
2268                 void *fsdata;
2269
2270                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2271                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2272                                                 iov_iter_count(i));
2273
2274 again:
2275
2276                 /*
2277                  * Bring in the user page that we will copy from _first_.
2278                  * Otherwise there's a nasty deadlock on copying from the
2279                  * same page as we're writing to, without it being marked
2280                  * up-to-date.
2281                  *
2282                  * Not only is this an optimisation, but it is also required
2283                  * to check that the address is actually valid, when atomic
2284                  * usercopies are used, below.
2285                  */
2286                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2287                         status = -EFAULT;
2288                         break;
2289                 }
2290
2291                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2292                                                 &page, &fsdata);
2293                 if (unlikely(status))
2294                         break;
2295
2296                 if (mapping_writably_mapped(mapping))
2297                         flush_dcache_page(page);
2298
2299                 pagefault_disable();
2300                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2301                 pagefault_enable();
2302                 flush_dcache_page(page);
2303
2304                 mark_page_accessed(page);
2305                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2306                                                 page, fsdata);
2307                 if (unlikely(status < 0))
2308                         break;
2309                 copied = status;
2310
2311                 cond_resched();
2312
2313                 iov_iter_advance(i, copied);
2314                 if (unlikely(copied == 0)) {
2315                         /*
2316                          * If we were unable to copy any data at all, we must
2317                          * fall back to a single segment length write.
2318                          *
2319                          * If we didn't fallback here, we could livelock
2320                          * because not all segments in the iov can be copied at
2321                          * once without a pagefault.
2322                          */
2323                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2324                                                 iov_iter_single_seg_count(i));
2325                         goto again;
2326                 }
2327                 pos += copied;
2328                 written += copied;
2329
2330                 balance_dirty_pages_ratelimited(mapping);
2331
2332         } while (iov_iter_count(i));
2333
2334         return written ? written : status;
2335 }
2336
2337 ssize_t
2338 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2339                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2340                 size_t count, ssize_t written)
2341 {
2342         struct file *file = iocb->ki_filp;
2343         ssize_t status;
2344         struct iov_iter i;
2345
2346         iov_iter_init(&i, iov, nr_segs, count, written);
2347         status = generic_perform_write(file, &i, pos);
2348
2349         if (likely(status >= 0)) {
2350                 written += status;
2351                 *ppos = pos + status;
2352         }
2353         
2354         return written ? written : status;
2355 }
2356 EXPORT_SYMBOL(generic_file_buffered_write);
2357
2358 /**
2359  * __generic_file_aio_write - write data to a file
2360  * @iocb:       IO state structure (file, offset, etc.)
2361  * @iov:        vector with data to write
2362  * @nr_segs:    number of segments in the vector
2363  * @ppos:       position where to write
2364  *
2365  * This function does all the work needed for actually writing data to a
2366  * file. It does all basic checks, removes SUID from the file, updates
2367  * modification times and calls proper subroutines depending on whether we
2368  * do direct IO or a standard buffered write.
2369  *
2370  * It expects i_mutex to be grabbed unless we work on a block device or similar
2371  * object which does not need locking at all.
2372  *
2373  * This function does *not* take care of syncing data in case of O_SYNC write.
2374  * A caller has to handle it. This is mainly due to the fact that we want to
2375  * avoid syncing under i_mutex.
2376  */
2377 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2378                                  unsigned long nr_segs, loff_t *ppos)
2379 {
2380         struct file *file = iocb->ki_filp;
2381         struct address_space * mapping = file->f_mapping;
2382         size_t ocount;          /* original count */
2383         size_t count;           /* after file limit checks */
2384         struct inode    *inode = mapping->host;
2385         loff_t          pos;
2386         ssize_t         written;
2387         ssize_t         err;
2388
2389         ocount = 0;
2390         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2391         if (err)
2392                 return err;
2393
2394         count = ocount;
2395         pos = *ppos;
2396
2397         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2398
2399         /* We can write back this queue in page reclaim */
2400         current->backing_dev_info = mapping->backing_dev_info;
2401         written = 0;
2402
2403         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2404         if (err)
2405                 goto out;
2406
2407         if (count == 0)
2408                 goto out;
2409
2410         err = file_remove_suid(file);
2411         if (err)
2412                 goto out;
2413
2414         file_update_time(file);
2415
2416         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2417         if (unlikely(file->f_flags & O_DIRECT)) {
2418                 loff_t endbyte;
2419                 ssize_t written_buffered;
2420
2421                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2422                                                         ppos, count, ocount);
2423                 if (written < 0 || written == count)
2424                         goto out;
2425                 /*
2426                  * direct-io write to a hole: fall through to buffered I/O
2427                  * for completing the rest of the request.
2428                  */
2429                 pos += written;
2430                 count -= written;
2431                 written_buffered = generic_file_buffered_write(iocb, iov,
2432                                                 nr_segs, pos, ppos, count,
2433                                                 written);
2434                 /*
2435                  * If generic_file_buffered_write() retuned a synchronous error
2436                  * then we want to return the number of bytes which were
2437                  * direct-written, or the error code if that was zero.  Note
2438                  * that this differs from normal direct-io semantics, which
2439                  * will return -EFOO even if some bytes were written.
2440                  */
2441                 if (written_buffered < 0) {
2442                         err = written_buffered;
2443                         goto out;
2444                 }
2445
2446                 /*
2447                  * We need to ensure that the page cache pages are written to
2448                  * disk and invalidated to preserve the expected O_DIRECT
2449                  * semantics.
2450                  */
2451                 endbyte = pos + written_buffered - written - 1;
2452                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2453                 if (err == 0) {
2454                         written = written_buffered;
2455                         invalidate_mapping_pages(mapping,
2456                                                  pos >> PAGE_CACHE_SHIFT,
2457                                                  endbyte >> PAGE_CACHE_SHIFT);
2458                 } else {
2459                         /*
2460                          * We don't know how much we wrote, so just return
2461                          * the number of bytes which were direct-written
2462                          */
2463                 }
2464         } else {
2465                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2466                                 pos, ppos, count, written);
2467         }
2468 out:
2469         current->backing_dev_info = NULL;
2470         return written ? written : err;
2471 }
2472 EXPORT_SYMBOL(__generic_file_aio_write);
2473
2474 /**
2475  * generic_file_aio_write - write data to a file
2476  * @iocb:       IO state structure
2477  * @iov:        vector with data to write
2478  * @nr_segs:    number of segments in the vector
2479  * @pos:        position in file where to write
2480  *
2481  * This is a wrapper around __generic_file_aio_write() to be used by most
2482  * filesystems. It takes care of syncing the file in case of O_SYNC file
2483  * and acquires i_mutex as needed.
2484  */
2485 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2486                 unsigned long nr_segs, loff_t pos)
2487 {
2488         struct file *file = iocb->ki_filp;
2489         struct inode *inode = file->f_mapping->host;
2490         ssize_t ret;
2491
2492         BUG_ON(iocb->ki_pos != pos);
2493
2494         mutex_lock(&inode->i_mutex);
2495         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2496         mutex_unlock(&inode->i_mutex);
2497
2498         if (ret > 0 || ret == -EIOCBQUEUED) {
2499                 ssize_t err;
2500
2501                 err = generic_write_sync(file, pos, ret);
2502                 if (err < 0 && ret > 0)
2503                         ret = err;
2504         }
2505         return ret;
2506 }
2507 EXPORT_SYMBOL(generic_file_aio_write);
2508
2509 /**
2510  * try_to_release_page() - release old fs-specific metadata on a page
2511  *
2512  * @page: the page which the kernel is trying to free
2513  * @gfp_mask: memory allocation flags (and I/O mode)
2514  *
2515  * The address_space is to try to release any data against the page
2516  * (presumably at page->private).  If the release was successful, return `1'.
2517  * Otherwise return zero.
2518  *
2519  * This may also be called if PG_fscache is set on a page, indicating that the
2520  * page is known to the local caching routines.
2521  *
2522  * The @gfp_mask argument specifies whether I/O may be performed to release
2523  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2524  *
2525  */
2526 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2527 {
2528         struct address_space * const mapping = page->mapping;
2529
2530         BUG_ON(!PageLocked(page));
2531         if (PageWriteback(page))
2532                 return 0;
2533
2534         if (mapping && mapping->a_ops->releasepage)
2535                 return mapping->a_ops->releasepage(page, gfp_mask);
2536         return try_to_free_buffers(page);
2537 }
2538
2539 EXPORT_SYMBOL(try_to_release_page);