Release page reference during page fault retry
[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  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  *
108  *  (code doesn't rely on that order, so you could switch it around)
109  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
110  *    ->i_mmap_lock
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         if (PageSwapBacked(page))
127                 __dec_zone_page_state(page, NR_SHMEM);
128         BUG_ON(page_mapped(page));
129
130         /*
131          * Some filesystems seem to re-dirty the page even after
132          * the VM has canceled the dirty bit (eg ext3 journaling).
133          *
134          * Fix it up by doing a final dirty accounting check after
135          * having removed the page entirely.
136          */
137         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
138                 dec_zone_page_state(page, NR_FILE_DIRTY);
139                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
140         }
141 }
142
143 void remove_from_page_cache(struct page *page)
144 {
145         struct address_space *mapping = page->mapping;
146
147         BUG_ON(!PageLocked(page));
148
149         spin_lock_irq(&mapping->tree_lock);
150         __remove_from_page_cache(page);
151         spin_unlock_irq(&mapping->tree_lock);
152         mem_cgroup_uncharge_cache_page(page);
153 }
154 EXPORT_SYMBOL(remove_from_page_cache);
155
156 static int sync_page(void *word)
157 {
158         struct address_space *mapping;
159         struct page *page;
160
161         page = container_of((unsigned long *)word, struct page, flags);
162
163         /*
164          * page_mapping() is being called without PG_locked held.
165          * Some knowledge of the state and use of the page is used to
166          * reduce the requirements down to a memory barrier.
167          * The danger here is of a stale page_mapping() return value
168          * indicating a struct address_space different from the one it's
169          * associated with when it is associated with one.
170          * After smp_mb(), it's either the correct page_mapping() for
171          * the page, or an old page_mapping() and the page's own
172          * page_mapping() has gone NULL.
173          * The ->sync_page() address_space operation must tolerate
174          * page_mapping() going NULL. By an amazing coincidence,
175          * this comes about because none of the users of the page
176          * in the ->sync_page() methods make essential use of the
177          * page_mapping(), merely passing the page down to the backing
178          * device's unplug functions when it's non-NULL, which in turn
179          * ignore it for all cases but swap, where only page_private(page) is
180          * of interest. When page_mapping() does go NULL, the entire
181          * call stack gracefully ignores the page and returns.
182          * -- wli
183          */
184         smp_mb();
185         mapping = page_mapping(page);
186         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
187                 mapping->a_ops->sync_page(page);
188         io_schedule();
189         return 0;
190 }
191
192 static int sync_page_killable(void *word)
193 {
194         sync_page(word);
195         return fatal_signal_pending(current) ? -EINTR : 0;
196 }
197
198 /**
199  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
200  * @mapping:    address space structure to write
201  * @start:      offset in bytes where the range starts
202  * @end:        offset in bytes where the range ends (inclusive)
203  * @sync_mode:  enable synchronous operation
204  *
205  * Start writeback against all of a mapping's dirty pages that lie
206  * within the byte offsets <start, end> inclusive.
207  *
208  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
209  * opposed to a regular memory cleansing writeback.  The difference between
210  * these two operations is that if a dirty page/buffer is encountered, it must
211  * be waited upon, and not just skipped over.
212  */
213 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
214                                 loff_t end, int sync_mode)
215 {
216         int ret;
217         struct writeback_control wbc = {
218                 .sync_mode = sync_mode,
219                 .nr_to_write = LONG_MAX,
220                 .range_start = start,
221                 .range_end = end,
222         };
223
224         if (!mapping_cap_writeback_dirty(mapping))
225                 return 0;
226
227         ret = do_writepages(mapping, &wbc);
228         return ret;
229 }
230
231 static inline int __filemap_fdatawrite(struct address_space *mapping,
232         int sync_mode)
233 {
234         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
235 }
236
237 int filemap_fdatawrite(struct address_space *mapping)
238 {
239         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite);
242
243 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
244                                 loff_t end)
245 {
246         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
247 }
248 EXPORT_SYMBOL(filemap_fdatawrite_range);
249
250 /**
251  * filemap_flush - mostly a non-blocking flush
252  * @mapping:    target address_space
253  *
254  * This is a mostly non-blocking flush.  Not suitable for data-integrity
255  * purposes - I/O may not be started against all dirty pages.
256  */
257 int filemap_flush(struct address_space *mapping)
258 {
259         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
260 }
261 EXPORT_SYMBOL(filemap_flush);
262
263 /**
264  * filemap_fdatawait_range - wait for writeback to complete
265  * @mapping:            address space structure to wait for
266  * @start_byte:         offset in bytes where the range starts
267  * @end_byte:           offset in bytes where the range ends (inclusive)
268  *
269  * Walk the list of under-writeback pages of the given address space
270  * in the given range and wait for all of them.
271  */
272 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
273                             loff_t end_byte)
274 {
275         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
276         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
277         struct pagevec pvec;
278         int nr_pages;
279         int ret = 0;
280
281         if (end_byte < start_byte)
282                 return 0;
283
284         pagevec_init(&pvec, 0);
285         while ((index <= end) &&
286                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
287                         PAGECACHE_TAG_WRITEBACK,
288                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
289                 unsigned i;
290
291                 for (i = 0; i < nr_pages; i++) {
292                         struct page *page = pvec.pages[i];
293
294                         /* until radix tree lookup accepts end_index */
295                         if (page->index > end)
296                                 continue;
297
298                         wait_on_page_writeback(page);
299                         if (PageError(page))
300                                 ret = -EIO;
301                 }
302                 pagevec_release(&pvec);
303                 cond_resched();
304         }
305
306         /* Check for outstanding write errors */
307         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
308                 ret = -ENOSPC;
309         if (test_and_clear_bit(AS_EIO, &mapping->flags))
310                 ret = -EIO;
311
312         return ret;
313 }
314 EXPORT_SYMBOL(filemap_fdatawait_range);
315
316 /**
317  * filemap_fdatawait - wait for all under-writeback pages to complete
318  * @mapping: address space structure to wait for
319  *
320  * Walk the list of under-writeback pages of the given address space
321  * and wait for all of them.
322  */
323 int filemap_fdatawait(struct address_space *mapping)
324 {
325         loff_t i_size = i_size_read(mapping->host);
326
327         if (i_size == 0)
328                 return 0;
329
330         return filemap_fdatawait_range(mapping, 0, i_size - 1);
331 }
332 EXPORT_SYMBOL(filemap_fdatawait);
333
334 int filemap_write_and_wait(struct address_space *mapping)
335 {
336         int err = 0;
337
338         if (mapping->nrpages) {
339                 err = filemap_fdatawrite(mapping);
340                 /*
341                  * Even if the above returned error, the pages may be
342                  * written partially (e.g. -ENOSPC), so we wait for it.
343                  * But the -EIO is special case, it may indicate the worst
344                  * thing (e.g. bug) happened, so we avoid waiting for it.
345                  */
346                 if (err != -EIO) {
347                         int err2 = filemap_fdatawait(mapping);
348                         if (!err)
349                                 err = err2;
350                 }
351         }
352         return err;
353 }
354 EXPORT_SYMBOL(filemap_write_and_wait);
355
356 /**
357  * filemap_write_and_wait_range - write out & wait on a file range
358  * @mapping:    the address_space for the pages
359  * @lstart:     offset in bytes where the range starts
360  * @lend:       offset in bytes where the range ends (inclusive)
361  *
362  * Write out and wait upon file offsets lstart->lend, inclusive.
363  *
364  * Note that `lend' is inclusive (describes the last byte to be written) so
365  * that this function can be used to write to the very end-of-file (end = -1).
366  */
367 int filemap_write_and_wait_range(struct address_space *mapping,
368                                  loff_t lstart, loff_t lend)
369 {
370         int err = 0;
371
372         if (mapping->nrpages) {
373                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
374                                                  WB_SYNC_ALL);
375                 /* See comment of filemap_write_and_wait() */
376                 if (err != -EIO) {
377                         int err2 = filemap_fdatawait_range(mapping,
378                                                 lstart, lend);
379                         if (!err)
380                                 err = err2;
381                 }
382         }
383         return err;
384 }
385 EXPORT_SYMBOL(filemap_write_and_wait_range);
386
387 /**
388  * add_to_page_cache_locked - add a locked page to the pagecache
389  * @page:       page to add
390  * @mapping:    the page's address_space
391  * @offset:     page index
392  * @gfp_mask:   page allocation mode
393  *
394  * This function is used to add a page to the pagecache. It must be locked.
395  * This function does not add the page to the LRU.  The caller must do that.
396  */
397 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
398                 pgoff_t offset, gfp_t gfp_mask)
399 {
400         int error;
401
402         VM_BUG_ON(!PageLocked(page));
403
404         error = mem_cgroup_cache_charge(page, current->mm,
405                                         gfp_mask & GFP_RECLAIM_MASK);
406         if (error)
407                 goto out;
408
409         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
410         if (error == 0) {
411                 page_cache_get(page);
412                 page->mapping = mapping;
413                 page->index = offset;
414
415                 spin_lock_irq(&mapping->tree_lock);
416                 error = radix_tree_insert(&mapping->page_tree, offset, page);
417                 if (likely(!error)) {
418                         mapping->nrpages++;
419                         __inc_zone_page_state(page, NR_FILE_PAGES);
420                         if (PageSwapBacked(page))
421                                 __inc_zone_page_state(page, NR_SHMEM);
422                         spin_unlock_irq(&mapping->tree_lock);
423                 } else {
424                         page->mapping = NULL;
425                         spin_unlock_irq(&mapping->tree_lock);
426                         mem_cgroup_uncharge_cache_page(page);
427                         page_cache_release(page);
428                 }
429                 radix_tree_preload_end();
430         } else
431                 mem_cgroup_uncharge_cache_page(page);
432 out:
433         return error;
434 }
435 EXPORT_SYMBOL(add_to_page_cache_locked);
436
437 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
438                                 pgoff_t offset, gfp_t gfp_mask)
439 {
440         int ret;
441
442         /*
443          * Splice_read and readahead add shmem/tmpfs pages into the page cache
444          * before shmem_readpage has a chance to mark them as SwapBacked: they
445          * need to go on the anon lru below, and mem_cgroup_cache_charge
446          * (called in add_to_page_cache) needs to know where they're going too.
447          */
448         if (mapping_cap_swap_backed(mapping))
449                 SetPageSwapBacked(page);
450
451         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
452         if (ret == 0) {
453                 if (page_is_file_cache(page))
454                         lru_cache_add_file(page);
455                 else
456                         lru_cache_add_anon(page);
457         }
458         return ret;
459 }
460 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
461
462 #ifdef CONFIG_NUMA
463 struct page *__page_cache_alloc(gfp_t gfp)
464 {
465         int n;
466         struct page *page;
467
468         if (cpuset_do_page_mem_spread()) {
469                 get_mems_allowed();
470                 n = cpuset_mem_spread_node();
471                 page = alloc_pages_exact_node(n, gfp, 0);
472                 put_mems_allowed();
473                 return page;
474         }
475         return alloc_pages(gfp, 0);
476 }
477 EXPORT_SYMBOL(__page_cache_alloc);
478 #endif
479
480 static int __sleep_on_page_lock(void *word)
481 {
482         io_schedule();
483         return 0;
484 }
485
486 /*
487  * In order to wait for pages to become available there must be
488  * waitqueues associated with pages. By using a hash table of
489  * waitqueues where the bucket discipline is to maintain all
490  * waiters on the same queue and wake all when any of the pages
491  * become available, and for the woken contexts to check to be
492  * sure the appropriate page became available, this saves space
493  * at a cost of "thundering herd" phenomena during rare hash
494  * collisions.
495  */
496 static wait_queue_head_t *page_waitqueue(struct page *page)
497 {
498         const struct zone *zone = page_zone(page);
499
500         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
501 }
502
503 static inline void wake_up_page(struct page *page, int bit)
504 {
505         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
506 }
507
508 void wait_on_page_bit(struct page *page, int bit_nr)
509 {
510         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
511
512         if (test_bit(bit_nr, &page->flags))
513                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
514                                                         TASK_UNINTERRUPTIBLE);
515 }
516 EXPORT_SYMBOL(wait_on_page_bit);
517
518 /**
519  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
520  * @page: Page defining the wait queue of interest
521  * @waiter: Waiter to add to the queue
522  *
523  * Add an arbitrary @waiter to the wait queue for the nominated @page.
524  */
525 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
526 {
527         wait_queue_head_t *q = page_waitqueue(page);
528         unsigned long flags;
529
530         spin_lock_irqsave(&q->lock, flags);
531         __add_wait_queue(q, waiter);
532         spin_unlock_irqrestore(&q->lock, flags);
533 }
534 EXPORT_SYMBOL_GPL(add_page_wait_queue);
535
536 /**
537  * unlock_page - unlock a locked page
538  * @page: the page
539  *
540  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
541  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
542  * mechananism between PageLocked pages and PageWriteback pages is shared.
543  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
544  *
545  * The mb is necessary to enforce ordering between the clear_bit and the read
546  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
547  */
548 void unlock_page(struct page *page)
549 {
550         VM_BUG_ON(!PageLocked(page));
551         clear_bit_unlock(PG_locked, &page->flags);
552         smp_mb__after_clear_bit();
553         wake_up_page(page, PG_locked);
554 }
555 EXPORT_SYMBOL(unlock_page);
556
557 /**
558  * end_page_writeback - end writeback against a page
559  * @page: the page
560  */
561 void end_page_writeback(struct page *page)
562 {
563         if (TestClearPageReclaim(page))
564                 rotate_reclaimable_page(page);
565
566         if (!test_clear_page_writeback(page))
567                 BUG();
568
569         smp_mb__after_clear_bit();
570         wake_up_page(page, PG_writeback);
571 }
572 EXPORT_SYMBOL(end_page_writeback);
573
574 /**
575  * __lock_page - get a lock on the page, assuming we need to sleep to get it
576  * @page: the page to lock
577  *
578  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
579  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
580  * chances are that on the second loop, the block layer's plug list is empty,
581  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
582  */
583 void __lock_page(struct page *page)
584 {
585         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
586
587         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
588                                                         TASK_UNINTERRUPTIBLE);
589 }
590 EXPORT_SYMBOL(__lock_page);
591
592 int __lock_page_killable(struct page *page)
593 {
594         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
595
596         return __wait_on_bit_lock(page_waitqueue(page), &wait,
597                                         sync_page_killable, TASK_KILLABLE);
598 }
599 EXPORT_SYMBOL_GPL(__lock_page_killable);
600
601 /**
602  * __lock_page_nosync - get a lock on the page, without calling sync_page()
603  * @page: the page to lock
604  *
605  * Variant of lock_page that does not require the caller to hold a reference
606  * on the page's mapping.
607  */
608 void __lock_page_nosync(struct page *page)
609 {
610         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
611         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
612                                                         TASK_UNINTERRUPTIBLE);
613 }
614
615 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
616                          unsigned int flags)
617 {
618         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
619                 __lock_page(page);
620                 return 1;
621         } else {
622                 up_read(&mm->mmap_sem);
623                 wait_on_page_locked(page);
624                 return 0;
625         }
626 }
627
628 /**
629  * find_get_page - find and get a page reference
630  * @mapping: the address_space to search
631  * @offset: the page index
632  *
633  * Is there a pagecache struct page at the given (mapping, offset) tuple?
634  * If yes, increment its refcount and return it; if no, return NULL.
635  */
636 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
637 {
638         void **pagep;
639         struct page *page;
640
641         rcu_read_lock();
642 repeat:
643         page = NULL;
644         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
645         if (pagep) {
646                 page = radix_tree_deref_slot(pagep);
647                 if (unlikely(!page || page == RADIX_TREE_RETRY))
648                         goto repeat;
649
650                 if (!page_cache_get_speculative(page))
651                         goto repeat;
652
653                 /*
654                  * Has the page moved?
655                  * This is part of the lockless pagecache protocol. See
656                  * include/linux/pagemap.h for details.
657                  */
658                 if (unlikely(page != *pagep)) {
659                         page_cache_release(page);
660                         goto repeat;
661                 }
662         }
663         rcu_read_unlock();
664
665         return page;
666 }
667 EXPORT_SYMBOL(find_get_page);
668
669 /**
670  * find_lock_page - locate, pin and lock a pagecache page
671  * @mapping: the address_space to search
672  * @offset: the page index
673  *
674  * Locates the desired pagecache page, locks it, increments its reference
675  * count and returns its address.
676  *
677  * Returns zero if the page was not present. find_lock_page() may sleep.
678  */
679 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
680 {
681         struct page *page;
682
683 repeat:
684         page = find_get_page(mapping, offset);
685         if (page) {
686                 lock_page(page);
687                 /* Has the page been truncated? */
688                 if (unlikely(page->mapping != mapping)) {
689                         unlock_page(page);
690                         page_cache_release(page);
691                         goto repeat;
692                 }
693                 VM_BUG_ON(page->index != offset);
694         }
695         return page;
696 }
697 EXPORT_SYMBOL(find_lock_page);
698
699 /**
700  * find_or_create_page - locate or add a pagecache page
701  * @mapping: the page's address_space
702  * @index: the page's index into the mapping
703  * @gfp_mask: page allocation mode
704  *
705  * Locates a page in the pagecache.  If the page is not present, a new page
706  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
707  * LRU list.  The returned page is locked and has its reference count
708  * incremented.
709  *
710  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
711  * allocation!
712  *
713  * find_or_create_page() returns the desired page's address, or zero on
714  * memory exhaustion.
715  */
716 struct page *find_or_create_page(struct address_space *mapping,
717                 pgoff_t index, gfp_t gfp_mask)
718 {
719         struct page *page;
720         int err;
721 repeat:
722         page = find_lock_page(mapping, index);
723         if (!page) {
724                 page = __page_cache_alloc(gfp_mask);
725                 if (!page)
726                         return NULL;
727                 /*
728                  * We want a regular kernel memory (not highmem or DMA etc)
729                  * allocation for the radix tree nodes, but we need to honour
730                  * the context-specific requirements the caller has asked for.
731                  * GFP_RECLAIM_MASK collects those requirements.
732                  */
733                 err = add_to_page_cache_lru(page, mapping, index,
734                         (gfp_mask & GFP_RECLAIM_MASK));
735                 if (unlikely(err)) {
736                         page_cache_release(page);
737                         page = NULL;
738                         if (err == -EEXIST)
739                                 goto repeat;
740                 }
741         }
742         return page;
743 }
744 EXPORT_SYMBOL(find_or_create_page);
745
746 /**
747  * find_get_pages - gang pagecache lookup
748  * @mapping:    The address_space to search
749  * @start:      The starting page index
750  * @nr_pages:   The maximum number of pages
751  * @pages:      Where the resulting pages are placed
752  *
753  * find_get_pages() will search for and return a group of up to
754  * @nr_pages pages in the mapping.  The pages are placed at @pages.
755  * find_get_pages() takes a reference against the returned pages.
756  *
757  * The search returns a group of mapping-contiguous pages with ascending
758  * indexes.  There may be holes in the indices due to not-present pages.
759  *
760  * find_get_pages() returns the number of pages which were found.
761  */
762 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
763                             unsigned int nr_pages, struct page **pages)
764 {
765         unsigned int i;
766         unsigned int ret;
767         unsigned int nr_found;
768
769         rcu_read_lock();
770 restart:
771         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
772                                 (void ***)pages, start, nr_pages);
773         ret = 0;
774         for (i = 0; i < nr_found; i++) {
775                 struct page *page;
776 repeat:
777                 page = radix_tree_deref_slot((void **)pages[i]);
778                 if (unlikely(!page))
779                         continue;
780                 /*
781                  * this can only trigger if nr_found == 1, making livelock
782                  * a non issue.
783                  */
784                 if (unlikely(page == RADIX_TREE_RETRY))
785                         goto restart;
786
787                 if (!page_cache_get_speculative(page))
788                         goto repeat;
789
790                 /* Has the page moved? */
791                 if (unlikely(page != *((void **)pages[i]))) {
792                         page_cache_release(page);
793                         goto repeat;
794                 }
795
796                 pages[ret] = page;
797                 ret++;
798         }
799         rcu_read_unlock();
800         return ret;
801 }
802
803 /**
804  * find_get_pages_contig - gang contiguous pagecache lookup
805  * @mapping:    The address_space to search
806  * @index:      The starting page index
807  * @nr_pages:   The maximum number of pages
808  * @pages:      Where the resulting pages are placed
809  *
810  * find_get_pages_contig() works exactly like find_get_pages(), except
811  * that the returned number of pages are guaranteed to be contiguous.
812  *
813  * find_get_pages_contig() returns the number of pages which were found.
814  */
815 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
816                                unsigned int nr_pages, struct page **pages)
817 {
818         unsigned int i;
819         unsigned int ret;
820         unsigned int nr_found;
821
822         rcu_read_lock();
823 restart:
824         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
825                                 (void ***)pages, index, nr_pages);
826         ret = 0;
827         for (i = 0; i < nr_found; i++) {
828                 struct page *page;
829 repeat:
830                 page = radix_tree_deref_slot((void **)pages[i]);
831                 if (unlikely(!page))
832                         continue;
833                 /*
834                  * this can only trigger if nr_found == 1, making livelock
835                  * a non issue.
836                  */
837                 if (unlikely(page == RADIX_TREE_RETRY))
838                         goto restart;
839
840                 if (page->mapping == NULL || page->index != index)
841                         break;
842
843                 if (!page_cache_get_speculative(page))
844                         goto repeat;
845
846                 /* Has the page moved? */
847                 if (unlikely(page != *((void **)pages[i]))) {
848                         page_cache_release(page);
849                         goto repeat;
850                 }
851
852                 pages[ret] = page;
853                 ret++;
854                 index++;
855         }
856         rcu_read_unlock();
857         return ret;
858 }
859 EXPORT_SYMBOL(find_get_pages_contig);
860
861 /**
862  * find_get_pages_tag - find and return pages that match @tag
863  * @mapping:    the address_space to search
864  * @index:      the starting page index
865  * @tag:        the tag index
866  * @nr_pages:   the maximum number of pages
867  * @pages:      where the resulting pages are placed
868  *
869  * Like find_get_pages, except we only return pages which are tagged with
870  * @tag.   We update @index to index the next page for the traversal.
871  */
872 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
873                         int tag, unsigned int nr_pages, struct page **pages)
874 {
875         unsigned int i;
876         unsigned int ret;
877         unsigned int nr_found;
878
879         rcu_read_lock();
880 restart:
881         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
882                                 (void ***)pages, *index, nr_pages, tag);
883         ret = 0;
884         for (i = 0; i < nr_found; i++) {
885                 struct page *page;
886 repeat:
887                 page = radix_tree_deref_slot((void **)pages[i]);
888                 if (unlikely(!page))
889                         continue;
890                 /*
891                  * this can only trigger if nr_found == 1, making livelock
892                  * a non issue.
893                  */
894                 if (unlikely(page == RADIX_TREE_RETRY))
895                         goto restart;
896
897                 if (!page_cache_get_speculative(page))
898                         goto repeat;
899
900                 /* Has the page moved? */
901                 if (unlikely(page != *((void **)pages[i]))) {
902                         page_cache_release(page);
903                         goto repeat;
904                 }
905
906                 pages[ret] = page;
907                 ret++;
908         }
909         rcu_read_unlock();
910
911         if (ret)
912                 *index = pages[ret - 1]->index + 1;
913
914         return ret;
915 }
916 EXPORT_SYMBOL(find_get_pages_tag);
917
918 /**
919  * grab_cache_page_nowait - returns locked page at given index in given cache
920  * @mapping: target address_space
921  * @index: the page index
922  *
923  * Same as grab_cache_page(), but do not wait if the page is unavailable.
924  * This is intended for speculative data generators, where the data can
925  * be regenerated if the page couldn't be grabbed.  This routine should
926  * be safe to call while holding the lock for another page.
927  *
928  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
929  * and deadlock against the caller's locked page.
930  */
931 struct page *
932 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
933 {
934         struct page *page = find_get_page(mapping, index);
935
936         if (page) {
937                 if (trylock_page(page))
938                         return page;
939                 page_cache_release(page);
940                 return NULL;
941         }
942         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
943         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
944                 page_cache_release(page);
945                 page = NULL;
946         }
947         return page;
948 }
949 EXPORT_SYMBOL(grab_cache_page_nowait);
950
951 /*
952  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
953  * a _large_ part of the i/o request. Imagine the worst scenario:
954  *
955  *      ---R__________________________________________B__________
956  *         ^ reading here                             ^ bad block(assume 4k)
957  *
958  * read(R) => miss => readahead(R...B) => media error => frustrating retries
959  * => failing the whole request => read(R) => read(R+1) =>
960  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
961  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
962  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
963  *
964  * It is going insane. Fix it by quickly scaling down the readahead size.
965  */
966 static void shrink_readahead_size_eio(struct file *filp,
967                                         struct file_ra_state *ra)
968 {
969         ra->ra_pages /= 4;
970 }
971
972 /**
973  * do_generic_file_read - generic file read routine
974  * @filp:       the file to read
975  * @ppos:       current file position
976  * @desc:       read_descriptor
977  * @actor:      read method
978  *
979  * This is a generic file read routine, and uses the
980  * mapping->a_ops->readpage() function for the actual low-level stuff.
981  *
982  * This is really ugly. But the goto's actually try to clarify some
983  * of the logic when it comes to error handling etc.
984  */
985 static void do_generic_file_read(struct file *filp, loff_t *ppos,
986                 read_descriptor_t *desc, read_actor_t actor)
987 {
988         struct address_space *mapping = filp->f_mapping;
989         struct inode *inode = mapping->host;
990         struct file_ra_state *ra = &filp->f_ra;
991         pgoff_t index;
992         pgoff_t last_index;
993         pgoff_t prev_index;
994         unsigned long offset;      /* offset into pagecache page */
995         unsigned int prev_offset;
996         int error;
997
998         index = *ppos >> PAGE_CACHE_SHIFT;
999         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1000         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1001         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1002         offset = *ppos & ~PAGE_CACHE_MASK;
1003
1004         for (;;) {
1005                 struct page *page;
1006                 pgoff_t end_index;
1007                 loff_t isize;
1008                 unsigned long nr, ret;
1009
1010                 cond_resched();
1011 find_page:
1012                 page = find_get_page(mapping, index);
1013                 if (!page) {
1014                         page_cache_sync_readahead(mapping,
1015                                         ra, filp,
1016                                         index, last_index - index);
1017                         page = find_get_page(mapping, index);
1018                         if (unlikely(page == NULL))
1019                                 goto no_cached_page;
1020                 }
1021                 if (PageReadahead(page)) {
1022                         page_cache_async_readahead(mapping,
1023                                         ra, filp, page,
1024                                         index, last_index - index);
1025                 }
1026                 if (!PageUptodate(page)) {
1027                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1028                                         !mapping->a_ops->is_partially_uptodate)
1029                                 goto page_not_up_to_date;
1030                         if (!trylock_page(page))
1031                                 goto page_not_up_to_date;
1032                         if (!mapping->a_ops->is_partially_uptodate(page,
1033                                                                 desc, offset))
1034                                 goto page_not_up_to_date_locked;
1035                         unlock_page(page);
1036                 }
1037 page_ok:
1038                 /*
1039                  * i_size must be checked after we know the page is Uptodate.
1040                  *
1041                  * Checking i_size after the check allows us to calculate
1042                  * the correct value for "nr", which means the zero-filled
1043                  * part of the page is not copied back to userspace (unless
1044                  * another truncate extends the file - this is desired though).
1045                  */
1046
1047                 isize = i_size_read(inode);
1048                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1049                 if (unlikely(!isize || index > end_index)) {
1050                         page_cache_release(page);
1051                         goto out;
1052                 }
1053
1054                 /* nr is the maximum number of bytes to copy from this page */
1055                 nr = PAGE_CACHE_SIZE;
1056                 if (index == end_index) {
1057                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1058                         if (nr <= offset) {
1059                                 page_cache_release(page);
1060                                 goto out;
1061                         }
1062                 }
1063                 nr = nr - offset;
1064
1065                 /* If users can be writing to this page using arbitrary
1066                  * virtual addresses, take care about potential aliasing
1067                  * before reading the page on the kernel side.
1068                  */
1069                 if (mapping_writably_mapped(mapping))
1070                         flush_dcache_page(page);
1071
1072                 /*
1073                  * When a sequential read accesses a page several times,
1074                  * only mark it as accessed the first time.
1075                  */
1076                 if (prev_index != index || offset != prev_offset)
1077                         mark_page_accessed(page);
1078                 prev_index = index;
1079
1080                 /*
1081                  * Ok, we have the page, and it's up-to-date, so
1082                  * now we can copy it to user space...
1083                  *
1084                  * The actor routine returns how many bytes were actually used..
1085                  * NOTE! This may not be the same as how much of a user buffer
1086                  * we filled up (we may be padding etc), so we can only update
1087                  * "pos" here (the actor routine has to update the user buffer
1088                  * pointers and the remaining count).
1089                  */
1090                 ret = actor(desc, page, offset, nr);
1091                 offset += ret;
1092                 index += offset >> PAGE_CACHE_SHIFT;
1093                 offset &= ~PAGE_CACHE_MASK;
1094                 prev_offset = offset;
1095
1096                 page_cache_release(page);
1097                 if (ret == nr && desc->count)
1098                         continue;
1099                 goto out;
1100
1101 page_not_up_to_date:
1102                 /* Get exclusive access to the page ... */
1103                 error = lock_page_killable(page);
1104                 if (unlikely(error))
1105                         goto readpage_error;
1106
1107 page_not_up_to_date_locked:
1108                 /* Did it get truncated before we got the lock? */
1109                 if (!page->mapping) {
1110                         unlock_page(page);
1111                         page_cache_release(page);
1112                         continue;
1113                 }
1114
1115                 /* Did somebody else fill it already? */
1116                 if (PageUptodate(page)) {
1117                         unlock_page(page);
1118                         goto page_ok;
1119                 }
1120
1121 readpage:
1122                 /*
1123                  * A previous I/O error may have been due to temporary
1124                  * failures, eg. multipath errors.
1125                  * PG_error will be set again if readpage fails.
1126                  */
1127                 ClearPageError(page);
1128                 /* Start the actual read. The read will unlock the page. */
1129                 error = mapping->a_ops->readpage(filp, page);
1130
1131                 if (unlikely(error)) {
1132                         if (error == AOP_TRUNCATED_PAGE) {
1133                                 page_cache_release(page);
1134                                 goto find_page;
1135                         }
1136                         goto readpage_error;
1137                 }
1138
1139                 if (!PageUptodate(page)) {
1140                         error = lock_page_killable(page);
1141                         if (unlikely(error))
1142                                 goto readpage_error;
1143                         if (!PageUptodate(page)) {
1144                                 if (page->mapping == NULL) {
1145                                         /*
1146                                          * invalidate_mapping_pages got it
1147                                          */
1148                                         unlock_page(page);
1149                                         page_cache_release(page);
1150                                         goto find_page;
1151                                 }
1152                                 unlock_page(page);
1153                                 shrink_readahead_size_eio(filp, ra);
1154                                 error = -EIO;
1155                                 goto readpage_error;
1156                         }
1157                         unlock_page(page);
1158                 }
1159
1160                 goto page_ok;
1161
1162 readpage_error:
1163                 /* UHHUH! A synchronous read error occurred. Report it */
1164                 desc->error = error;
1165                 page_cache_release(page);
1166                 goto out;
1167
1168 no_cached_page:
1169                 /*
1170                  * Ok, it wasn't cached, so we need to create a new
1171                  * page..
1172                  */
1173                 page = page_cache_alloc_cold(mapping);
1174                 if (!page) {
1175                         desc->error = -ENOMEM;
1176                         goto out;
1177                 }
1178                 error = add_to_page_cache_lru(page, mapping,
1179                                                 index, GFP_KERNEL);
1180                 if (error) {
1181                         page_cache_release(page);
1182                         if (error == -EEXIST)
1183                                 goto find_page;
1184                         desc->error = error;
1185                         goto out;
1186                 }
1187                 goto readpage;
1188         }
1189
1190 out:
1191         ra->prev_pos = prev_index;
1192         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1193         ra->prev_pos |= prev_offset;
1194
1195         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1196         file_accessed(filp);
1197 }
1198
1199 int file_read_actor(read_descriptor_t *desc, struct page *page,
1200                         unsigned long offset, unsigned long size)
1201 {
1202         char *kaddr;
1203         unsigned long left, count = desc->count;
1204
1205         if (size > count)
1206                 size = count;
1207
1208         /*
1209          * Faults on the destination of a read are common, so do it before
1210          * taking the kmap.
1211          */
1212         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1213                 kaddr = kmap_atomic(page, KM_USER0);
1214                 left = __copy_to_user_inatomic(desc->arg.buf,
1215                                                 kaddr + offset, size);
1216                 kunmap_atomic(kaddr, KM_USER0);
1217                 if (left == 0)
1218                         goto success;
1219         }
1220
1221         /* Do it the slow way */
1222         kaddr = kmap(page);
1223         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1224         kunmap(page);
1225
1226         if (left) {
1227                 size -= left;
1228                 desc->error = -EFAULT;
1229         }
1230 success:
1231         desc->count = count - size;
1232         desc->written += size;
1233         desc->arg.buf += size;
1234         return size;
1235 }
1236
1237 /*
1238  * Performs necessary checks before doing a write
1239  * @iov:        io vector request
1240  * @nr_segs:    number of segments in the iovec
1241  * @count:      number of bytes to write
1242  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1243  *
1244  * Adjust number of segments and amount of bytes to write (nr_segs should be
1245  * properly initialized first). Returns appropriate error code that caller
1246  * should return or zero in case that write should be allowed.
1247  */
1248 int generic_segment_checks(const struct iovec *iov,
1249                         unsigned long *nr_segs, size_t *count, int access_flags)
1250 {
1251         unsigned long   seg;
1252         size_t cnt = 0;
1253         for (seg = 0; seg < *nr_segs; seg++) {
1254                 const struct iovec *iv = &iov[seg];
1255
1256                 /*
1257                  * If any segment has a negative length, or the cumulative
1258                  * length ever wraps negative then return -EINVAL.
1259                  */
1260                 cnt += iv->iov_len;
1261                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1262                         return -EINVAL;
1263                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1264                         continue;
1265                 if (seg == 0)
1266                         return -EFAULT;
1267                 *nr_segs = seg;
1268                 cnt -= iv->iov_len;     /* This segment is no good */
1269                 break;
1270         }
1271         *count = cnt;
1272         return 0;
1273 }
1274 EXPORT_SYMBOL(generic_segment_checks);
1275
1276 /**
1277  * generic_file_aio_read - generic filesystem read routine
1278  * @iocb:       kernel I/O control block
1279  * @iov:        io vector request
1280  * @nr_segs:    number of segments in the iovec
1281  * @pos:        current file position
1282  *
1283  * This is the "read()" routine for all filesystems
1284  * that can use the page cache directly.
1285  */
1286 ssize_t
1287 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1288                 unsigned long nr_segs, loff_t pos)
1289 {
1290         struct file *filp = iocb->ki_filp;
1291         ssize_t retval;
1292         unsigned long seg = 0;
1293         size_t count;
1294         loff_t *ppos = &iocb->ki_pos;
1295
1296         count = 0;
1297         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1298         if (retval)
1299                 return retval;
1300
1301         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1302         if (filp->f_flags & O_DIRECT) {
1303                 loff_t size;
1304                 struct address_space *mapping;
1305                 struct inode *inode;
1306
1307                 mapping = filp->f_mapping;
1308                 inode = mapping->host;
1309                 if (!count)
1310                         goto out; /* skip atime */
1311                 size = i_size_read(inode);
1312                 if (pos < size) {
1313                         retval = filemap_write_and_wait_range(mapping, pos,
1314                                         pos + iov_length(iov, nr_segs) - 1);
1315                         if (!retval) {
1316                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1317                                                         iov, pos, nr_segs);
1318                         }
1319                         if (retval > 0) {
1320                                 *ppos = pos + retval;
1321                                 count -= retval;
1322                         }
1323
1324                         /*
1325                          * Btrfs can have a short DIO read if we encounter
1326                          * compressed extents, so if there was an error, or if
1327                          * we've already read everything we wanted to, or if
1328                          * there was a short read because we hit EOF, go ahead
1329                          * and return.  Otherwise fallthrough to buffered io for
1330                          * the rest of the read.
1331                          */
1332                         if (retval < 0 || !count || *ppos >= size) {
1333                                 file_accessed(filp);
1334                                 goto out;
1335                         }
1336                 }
1337         }
1338
1339         count = retval;
1340         for (seg = 0; seg < nr_segs; seg++) {
1341                 read_descriptor_t desc;
1342                 loff_t offset = 0;
1343
1344                 /*
1345                  * If we did a short DIO read we need to skip the section of the
1346                  * iov that we've already read data into.
1347                  */
1348                 if (count) {
1349                         if (count > iov[seg].iov_len) {
1350                                 count -= iov[seg].iov_len;
1351                                 continue;
1352                         }
1353                         offset = count;
1354                         count = 0;
1355                 }
1356
1357                 desc.written = 0;
1358                 desc.arg.buf = iov[seg].iov_base + offset;
1359                 desc.count = iov[seg].iov_len - offset;
1360                 if (desc.count == 0)
1361                         continue;
1362                 desc.error = 0;
1363                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1364                 retval += desc.written;
1365                 if (desc.error) {
1366                         retval = retval ?: desc.error;
1367                         break;
1368                 }
1369                 if (desc.count > 0)
1370                         break;
1371         }
1372 out:
1373         return retval;
1374 }
1375 EXPORT_SYMBOL(generic_file_aio_read);
1376
1377 static ssize_t
1378 do_readahead(struct address_space *mapping, struct file *filp,
1379              pgoff_t index, unsigned long nr)
1380 {
1381         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1382                 return -EINVAL;
1383
1384         force_page_cache_readahead(mapping, filp, index, nr);
1385         return 0;
1386 }
1387
1388 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1389 {
1390         ssize_t ret;
1391         struct file *file;
1392
1393         ret = -EBADF;
1394         file = fget(fd);
1395         if (file) {
1396                 if (file->f_mode & FMODE_READ) {
1397                         struct address_space *mapping = file->f_mapping;
1398                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1399                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1400                         unsigned long len = end - start + 1;
1401                         ret = do_readahead(mapping, file, start, len);
1402                 }
1403                 fput(file);
1404         }
1405         return ret;
1406 }
1407 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1408 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1409 {
1410         return SYSC_readahead((int) fd, offset, (size_t) count);
1411 }
1412 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1413 #endif
1414
1415 #ifdef CONFIG_MMU
1416 /**
1417  * page_cache_read - adds requested page to the page cache if not already there
1418  * @file:       file to read
1419  * @offset:     page index
1420  *
1421  * This adds the requested page to the page cache if it isn't already there,
1422  * and schedules an I/O to read in its contents from disk.
1423  */
1424 static int page_cache_read(struct file *file, pgoff_t offset)
1425 {
1426         struct address_space *mapping = file->f_mapping;
1427         struct page *page; 
1428         int ret;
1429
1430         do {
1431                 page = page_cache_alloc_cold(mapping);
1432                 if (!page)
1433                         return -ENOMEM;
1434
1435                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1436                 if (ret == 0)
1437                         ret = mapping->a_ops->readpage(file, page);
1438                 else if (ret == -EEXIST)
1439                         ret = 0; /* losing race to add is OK */
1440
1441                 page_cache_release(page);
1442
1443         } while (ret == AOP_TRUNCATED_PAGE);
1444                 
1445         return ret;
1446 }
1447
1448 #define MMAP_LOTSAMISS  (100)
1449
1450 /*
1451  * Synchronous readahead happens when we don't even find
1452  * a page in the page cache at all.
1453  */
1454 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1455                                    struct file_ra_state *ra,
1456                                    struct file *file,
1457                                    pgoff_t offset)
1458 {
1459         unsigned long ra_pages;
1460         struct address_space *mapping = file->f_mapping;
1461
1462         /* If we don't want any read-ahead, don't bother */
1463         if (VM_RandomReadHint(vma))
1464                 return;
1465
1466         if (VM_SequentialReadHint(vma) ||
1467                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1468                 page_cache_sync_readahead(mapping, ra, file, offset,
1469                                           ra->ra_pages);
1470                 return;
1471         }
1472
1473         if (ra->mmap_miss < INT_MAX)
1474                 ra->mmap_miss++;
1475
1476         /*
1477          * Do we miss much more than hit in this file? If so,
1478          * stop bothering with read-ahead. It will only hurt.
1479          */
1480         if (ra->mmap_miss > MMAP_LOTSAMISS)
1481                 return;
1482
1483         /*
1484          * mmap read-around
1485          */
1486         ra_pages = max_sane_readahead(ra->ra_pages);
1487         if (ra_pages) {
1488                 ra->start = max_t(long, 0, offset - ra_pages/2);
1489                 ra->size = ra_pages;
1490                 ra->async_size = 0;
1491                 ra_submit(ra, mapping, file);
1492         }
1493 }
1494
1495 /*
1496  * Asynchronous readahead happens when we find the page and PG_readahead,
1497  * so we want to possibly extend the readahead further..
1498  */
1499 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1500                                     struct file_ra_state *ra,
1501                                     struct file *file,
1502                                     struct page *page,
1503                                     pgoff_t offset)
1504 {
1505         struct address_space *mapping = file->f_mapping;
1506
1507         /* If we don't want any read-ahead, don't bother */
1508         if (VM_RandomReadHint(vma))
1509                 return;
1510         if (ra->mmap_miss > 0)
1511                 ra->mmap_miss--;
1512         if (PageReadahead(page))
1513                 page_cache_async_readahead(mapping, ra, file,
1514                                            page, offset, ra->ra_pages);
1515 }
1516
1517 /**
1518  * filemap_fault - read in file data for page fault handling
1519  * @vma:        vma in which the fault was taken
1520  * @vmf:        struct vm_fault containing details of the fault
1521  *
1522  * filemap_fault() is invoked via the vma operations vector for a
1523  * mapped memory region to read in file data during a page fault.
1524  *
1525  * The goto's are kind of ugly, but this streamlines the normal case of having
1526  * it in the page cache, and handles the special cases reasonably without
1527  * having a lot of duplicated code.
1528  */
1529 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1530 {
1531         int error;
1532         struct file *file = vma->vm_file;
1533         struct address_space *mapping = file->f_mapping;
1534         struct file_ra_state *ra = &file->f_ra;
1535         struct inode *inode = mapping->host;
1536         pgoff_t offset = vmf->pgoff;
1537         struct page *page;
1538         pgoff_t size;
1539         int ret = 0;
1540
1541         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1542         if (offset >= size)
1543                 return VM_FAULT_SIGBUS;
1544
1545         /*
1546          * Do we have something in the page cache already?
1547          */
1548         page = find_get_page(mapping, offset);
1549         if (likely(page)) {
1550                 /*
1551                  * We found the page, so try async readahead before
1552                  * waiting for the lock.
1553                  */
1554                 do_async_mmap_readahead(vma, ra, file, page, offset);
1555         } else {
1556                 /* No page in the page cache at all */
1557                 do_sync_mmap_readahead(vma, ra, file, offset);
1558                 count_vm_event(PGMAJFAULT);
1559                 ret = VM_FAULT_MAJOR;
1560 retry_find:
1561                 page = find_get_page(mapping, offset);
1562                 if (!page)
1563                         goto no_cached_page;
1564         }
1565
1566         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1567                 page_cache_release(page);
1568                 return ret | VM_FAULT_RETRY;
1569         }
1570
1571         /* Did it get truncated? */
1572         if (unlikely(page->mapping != mapping)) {
1573                 unlock_page(page);
1574                 put_page(page);
1575                 goto retry_find;
1576         }
1577         VM_BUG_ON(page->index != offset);
1578
1579         /*
1580          * We have a locked page in the page cache, now we need to check
1581          * that it's up-to-date. If not, it is going to be due to an error.
1582          */
1583         if (unlikely(!PageUptodate(page)))
1584                 goto page_not_uptodate;
1585
1586         /*
1587          * Found the page and have a reference on it.
1588          * We must recheck i_size under page lock.
1589          */
1590         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1591         if (unlikely(offset >= size)) {
1592                 unlock_page(page);
1593                 page_cache_release(page);
1594                 return VM_FAULT_SIGBUS;
1595         }
1596
1597         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1598         vmf->page = page;
1599         return ret | VM_FAULT_LOCKED;
1600
1601 no_cached_page:
1602         /*
1603          * We're only likely to ever get here if MADV_RANDOM is in
1604          * effect.
1605          */
1606         error = page_cache_read(file, offset);
1607
1608         /*
1609          * The page we want has now been added to the page cache.
1610          * In the unlikely event that someone removed it in the
1611          * meantime, we'll just come back here and read it again.
1612          */
1613         if (error >= 0)
1614                 goto retry_find;
1615
1616         /*
1617          * An error return from page_cache_read can result if the
1618          * system is low on memory, or a problem occurs while trying
1619          * to schedule I/O.
1620          */
1621         if (error == -ENOMEM)
1622                 return VM_FAULT_OOM;
1623         return VM_FAULT_SIGBUS;
1624
1625 page_not_uptodate:
1626         /*
1627          * Umm, take care of errors if the page isn't up-to-date.
1628          * Try to re-read it _once_. We do this synchronously,
1629          * because there really aren't any performance issues here
1630          * and we need to check for errors.
1631          */
1632         ClearPageError(page);
1633         error = mapping->a_ops->readpage(file, page);
1634         if (!error) {
1635                 wait_on_page_locked(page);
1636                 if (!PageUptodate(page))
1637                         error = -EIO;
1638         }
1639         page_cache_release(page);
1640
1641         if (!error || error == AOP_TRUNCATED_PAGE)
1642                 goto retry_find;
1643
1644         /* Things didn't work out. Return zero to tell the mm layer so. */
1645         shrink_readahead_size_eio(file, ra);
1646         return VM_FAULT_SIGBUS;
1647 }
1648 EXPORT_SYMBOL(filemap_fault);
1649
1650 const struct vm_operations_struct generic_file_vm_ops = {
1651         .fault          = filemap_fault,
1652 };
1653
1654 /* This is used for a general mmap of a disk file */
1655
1656 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1657 {
1658         struct address_space *mapping = file->f_mapping;
1659
1660         if (!mapping->a_ops->readpage)
1661                 return -ENOEXEC;
1662         file_accessed(file);
1663         vma->vm_ops = &generic_file_vm_ops;
1664         vma->vm_flags |= VM_CAN_NONLINEAR;
1665         return 0;
1666 }
1667
1668 /*
1669  * This is for filesystems which do not implement ->writepage.
1670  */
1671 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1672 {
1673         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1674                 return -EINVAL;
1675         return generic_file_mmap(file, vma);
1676 }
1677 #else
1678 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1679 {
1680         return -ENOSYS;
1681 }
1682 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1683 {
1684         return -ENOSYS;
1685 }
1686 #endif /* CONFIG_MMU */
1687
1688 EXPORT_SYMBOL(generic_file_mmap);
1689 EXPORT_SYMBOL(generic_file_readonly_mmap);
1690
1691 static struct page *__read_cache_page(struct address_space *mapping,
1692                                 pgoff_t index,
1693                                 int (*filler)(void *,struct page*),
1694                                 void *data,
1695                                 gfp_t gfp)
1696 {
1697         struct page *page;
1698         int err;
1699 repeat:
1700         page = find_get_page(mapping, index);
1701         if (!page) {
1702                 page = __page_cache_alloc(gfp | __GFP_COLD);
1703                 if (!page)
1704                         return ERR_PTR(-ENOMEM);
1705                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1706                 if (unlikely(err)) {
1707                         page_cache_release(page);
1708                         if (err == -EEXIST)
1709                                 goto repeat;
1710                         /* Presumably ENOMEM for radix tree node */
1711                         return ERR_PTR(err);
1712                 }
1713                 err = filler(data, page);
1714                 if (err < 0) {
1715                         page_cache_release(page);
1716                         page = ERR_PTR(err);
1717                 }
1718         }
1719         return page;
1720 }
1721
1722 static struct page *do_read_cache_page(struct address_space *mapping,
1723                                 pgoff_t index,
1724                                 int (*filler)(void *,struct page*),
1725                                 void *data,
1726                                 gfp_t gfp)
1727
1728 {
1729         struct page *page;
1730         int err;
1731
1732 retry:
1733         page = __read_cache_page(mapping, index, filler, data, gfp);
1734         if (IS_ERR(page))
1735                 return page;
1736         if (PageUptodate(page))
1737                 goto out;
1738
1739         lock_page(page);
1740         if (!page->mapping) {
1741                 unlock_page(page);
1742                 page_cache_release(page);
1743                 goto retry;
1744         }
1745         if (PageUptodate(page)) {
1746                 unlock_page(page);
1747                 goto out;
1748         }
1749         err = filler(data, page);
1750         if (err < 0) {
1751                 page_cache_release(page);
1752                 return ERR_PTR(err);
1753         }
1754 out:
1755         mark_page_accessed(page);
1756         return page;
1757 }
1758
1759 /**
1760  * read_cache_page_async - read into page cache, fill it if needed
1761  * @mapping:    the page's address_space
1762  * @index:      the page index
1763  * @filler:     function to perform the read
1764  * @data:       destination for read data
1765  *
1766  * Same as read_cache_page, but don't wait for page to become unlocked
1767  * after submitting it to the filler.
1768  *
1769  * Read into the page cache. If a page already exists, and PageUptodate() is
1770  * not set, try to fill the page but don't wait for it to become unlocked.
1771  *
1772  * If the page does not get brought uptodate, return -EIO.
1773  */
1774 struct page *read_cache_page_async(struct address_space *mapping,
1775                                 pgoff_t index,
1776                                 int (*filler)(void *,struct page*),
1777                                 void *data)
1778 {
1779         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1780 }
1781 EXPORT_SYMBOL(read_cache_page_async);
1782
1783 static struct page *wait_on_page_read(struct page *page)
1784 {
1785         if (!IS_ERR(page)) {
1786                 wait_on_page_locked(page);
1787                 if (!PageUptodate(page)) {
1788                         page_cache_release(page);
1789                         page = ERR_PTR(-EIO);
1790                 }
1791         }
1792         return page;
1793 }
1794
1795 /**
1796  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1797  * @mapping:    the page's address_space
1798  * @index:      the page index
1799  * @gfp:        the page allocator flags to use if allocating
1800  *
1801  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1802  * any new page allocations done using the specified allocation flags. Note
1803  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1804  * expect to do this atomically or anything like that - but you can pass in
1805  * other page requirements.
1806  *
1807  * If the page does not get brought uptodate, return -EIO.
1808  */
1809 struct page *read_cache_page_gfp(struct address_space *mapping,
1810                                 pgoff_t index,
1811                                 gfp_t gfp)
1812 {
1813         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1814
1815         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1816 }
1817 EXPORT_SYMBOL(read_cache_page_gfp);
1818
1819 /**
1820  * read_cache_page - read into page cache, fill it if needed
1821  * @mapping:    the page's address_space
1822  * @index:      the page index
1823  * @filler:     function to perform the read
1824  * @data:       destination for read data
1825  *
1826  * Read into the page cache. If a page already exists, and PageUptodate() is
1827  * not set, try to fill the page then wait for it to become unlocked.
1828  *
1829  * If the page does not get brought uptodate, return -EIO.
1830  */
1831 struct page *read_cache_page(struct address_space *mapping,
1832                                 pgoff_t index,
1833                                 int (*filler)(void *,struct page*),
1834                                 void *data)
1835 {
1836         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1837 }
1838 EXPORT_SYMBOL(read_cache_page);
1839
1840 /*
1841  * The logic we want is
1842  *
1843  *      if suid or (sgid and xgrp)
1844  *              remove privs
1845  */
1846 int should_remove_suid(struct dentry *dentry)
1847 {
1848         mode_t mode = dentry->d_inode->i_mode;
1849         int kill = 0;
1850
1851         /* suid always must be killed */
1852         if (unlikely(mode & S_ISUID))
1853                 kill = ATTR_KILL_SUID;
1854
1855         /*
1856          * sgid without any exec bits is just a mandatory locking mark; leave
1857          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1858          */
1859         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1860                 kill |= ATTR_KILL_SGID;
1861
1862         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1863                 return kill;
1864
1865         return 0;
1866 }
1867 EXPORT_SYMBOL(should_remove_suid);
1868
1869 static int __remove_suid(struct dentry *dentry, int kill)
1870 {
1871         struct iattr newattrs;
1872
1873         newattrs.ia_valid = ATTR_FORCE | kill;
1874         return notify_change(dentry, &newattrs);
1875 }
1876
1877 int file_remove_suid(struct file *file)
1878 {
1879         struct dentry *dentry = file->f_path.dentry;
1880         int killsuid = should_remove_suid(dentry);
1881         int killpriv = security_inode_need_killpriv(dentry);
1882         int error = 0;
1883
1884         if (killpriv < 0)
1885                 return killpriv;
1886         if (killpriv)
1887                 error = security_inode_killpriv(dentry);
1888         if (!error && killsuid)
1889                 error = __remove_suid(dentry, killsuid);
1890
1891         return error;
1892 }
1893 EXPORT_SYMBOL(file_remove_suid);
1894
1895 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1896                         const struct iovec *iov, size_t base, size_t bytes)
1897 {
1898         size_t copied = 0, left = 0;
1899
1900         while (bytes) {
1901                 char __user *buf = iov->iov_base + base;
1902                 int copy = min(bytes, iov->iov_len - base);
1903
1904                 base = 0;
1905                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1906                 copied += copy;
1907                 bytes -= copy;
1908                 vaddr += copy;
1909                 iov++;
1910
1911                 if (unlikely(left))
1912                         break;
1913         }
1914         return copied - left;
1915 }
1916
1917 /*
1918  * Copy as much as we can into the page and return the number of bytes which
1919  * were successfully copied.  If a fault is encountered then return the number of
1920  * bytes which were copied.
1921  */
1922 size_t iov_iter_copy_from_user_atomic(struct page *page,
1923                 struct iov_iter *i, unsigned long offset, size_t bytes)
1924 {
1925         char *kaddr;
1926         size_t copied;
1927
1928         BUG_ON(!in_atomic());
1929         kaddr = kmap_atomic(page, KM_USER0);
1930         if (likely(i->nr_segs == 1)) {
1931                 int left;
1932                 char __user *buf = i->iov->iov_base + i->iov_offset;
1933                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1934                 copied = bytes - left;
1935         } else {
1936                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1937                                                 i->iov, i->iov_offset, bytes);
1938         }
1939         kunmap_atomic(kaddr, KM_USER0);
1940
1941         return copied;
1942 }
1943 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1944
1945 /*
1946  * This has the same sideeffects and return value as
1947  * iov_iter_copy_from_user_atomic().
1948  * The difference is that it attempts to resolve faults.
1949  * Page must not be locked.
1950  */
1951 size_t iov_iter_copy_from_user(struct page *page,
1952                 struct iov_iter *i, unsigned long offset, size_t bytes)
1953 {
1954         char *kaddr;
1955         size_t copied;
1956
1957         kaddr = kmap(page);
1958         if (likely(i->nr_segs == 1)) {
1959                 int left;
1960                 char __user *buf = i->iov->iov_base + i->iov_offset;
1961                 left = __copy_from_user(kaddr + offset, buf, bytes);
1962                 copied = bytes - left;
1963         } else {
1964                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1965                                                 i->iov, i->iov_offset, bytes);
1966         }
1967         kunmap(page);
1968         return copied;
1969 }
1970 EXPORT_SYMBOL(iov_iter_copy_from_user);
1971
1972 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1973 {
1974         BUG_ON(i->count < bytes);
1975
1976         if (likely(i->nr_segs == 1)) {
1977                 i->iov_offset += bytes;
1978                 i->count -= bytes;
1979         } else {
1980                 const struct iovec *iov = i->iov;
1981                 size_t base = i->iov_offset;
1982
1983                 /*
1984                  * The !iov->iov_len check ensures we skip over unlikely
1985                  * zero-length segments (without overruning the iovec).
1986                  */
1987                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1988                         int copy;
1989
1990                         copy = min(bytes, iov->iov_len - base);
1991                         BUG_ON(!i->count || i->count < copy);
1992                         i->count -= copy;
1993                         bytes -= copy;
1994                         base += copy;
1995                         if (iov->iov_len == base) {
1996                                 iov++;
1997                                 base = 0;
1998                         }
1999                 }
2000                 i->iov = iov;
2001                 i->iov_offset = base;
2002         }
2003 }
2004 EXPORT_SYMBOL(iov_iter_advance);
2005
2006 /*
2007  * Fault in the first iovec of the given iov_iter, to a maximum length
2008  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2009  * accessed (ie. because it is an invalid address).
2010  *
2011  * writev-intensive code may want this to prefault several iovecs -- that
2012  * would be possible (callers must not rely on the fact that _only_ the
2013  * first iovec will be faulted with the current implementation).
2014  */
2015 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2016 {
2017         char __user *buf = i->iov->iov_base + i->iov_offset;
2018         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2019         return fault_in_pages_readable(buf, bytes);
2020 }
2021 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2022
2023 /*
2024  * Return the count of just the current iov_iter segment.
2025  */
2026 size_t iov_iter_single_seg_count(struct iov_iter *i)
2027 {
2028         const struct iovec *iov = i->iov;
2029         if (i->nr_segs == 1)
2030                 return i->count;
2031         else
2032                 return min(i->count, iov->iov_len - i->iov_offset);
2033 }
2034 EXPORT_SYMBOL(iov_iter_single_seg_count);
2035
2036 /*
2037  * Performs necessary checks before doing a write
2038  *
2039  * Can adjust writing position or amount of bytes to write.
2040  * Returns appropriate error code that caller should return or
2041  * zero in case that write should be allowed.
2042  */
2043 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2044 {
2045         struct inode *inode = file->f_mapping->host;
2046         unsigned long limit = rlimit(RLIMIT_FSIZE);
2047
2048         if (unlikely(*pos < 0))
2049                 return -EINVAL;
2050
2051         if (!isblk) {
2052                 /* FIXME: this is for backwards compatibility with 2.4 */
2053                 if (file->f_flags & O_APPEND)
2054                         *pos = i_size_read(inode);
2055
2056                 if (limit != RLIM_INFINITY) {
2057                         if (*pos >= limit) {
2058                                 send_sig(SIGXFSZ, current, 0);
2059                                 return -EFBIG;
2060                         }
2061                         if (*count > limit - (typeof(limit))*pos) {
2062                                 *count = limit - (typeof(limit))*pos;
2063                         }
2064                 }
2065         }
2066
2067         /*
2068          * LFS rule
2069          */
2070         if (unlikely(*pos + *count > MAX_NON_LFS &&
2071                                 !(file->f_flags & O_LARGEFILE))) {
2072                 if (*pos >= MAX_NON_LFS) {
2073                         return -EFBIG;
2074                 }
2075                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2076                         *count = MAX_NON_LFS - (unsigned long)*pos;
2077                 }
2078         }
2079
2080         /*
2081          * Are we about to exceed the fs block limit ?
2082          *
2083          * If we have written data it becomes a short write.  If we have
2084          * exceeded without writing data we send a signal and return EFBIG.
2085          * Linus frestrict idea will clean these up nicely..
2086          */
2087         if (likely(!isblk)) {
2088                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2089                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2090                                 return -EFBIG;
2091                         }
2092                         /* zero-length writes at ->s_maxbytes are OK */
2093                 }
2094
2095                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2096                         *count = inode->i_sb->s_maxbytes - *pos;
2097         } else {
2098 #ifdef CONFIG_BLOCK
2099                 loff_t isize;
2100                 if (bdev_read_only(I_BDEV(inode)))
2101                         return -EPERM;
2102                 isize = i_size_read(inode);
2103                 if (*pos >= isize) {
2104                         if (*count || *pos > isize)
2105                                 return -ENOSPC;
2106                 }
2107
2108                 if (*pos + *count > isize)
2109                         *count = isize - *pos;
2110 #else
2111                 return -EPERM;
2112 #endif
2113         }
2114         return 0;
2115 }
2116 EXPORT_SYMBOL(generic_write_checks);
2117
2118 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2119                                 loff_t pos, unsigned len, unsigned flags,
2120                                 struct page **pagep, void **fsdata)
2121 {
2122         const struct address_space_operations *aops = mapping->a_ops;
2123
2124         return aops->write_begin(file, mapping, pos, len, flags,
2125                                                         pagep, fsdata);
2126 }
2127 EXPORT_SYMBOL(pagecache_write_begin);
2128
2129 int pagecache_write_end(struct file *file, struct address_space *mapping,
2130                                 loff_t pos, unsigned len, unsigned copied,
2131                                 struct page *page, void *fsdata)
2132 {
2133         const struct address_space_operations *aops = mapping->a_ops;
2134
2135         mark_page_accessed(page);
2136         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2137 }
2138 EXPORT_SYMBOL(pagecache_write_end);
2139
2140 ssize_t
2141 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2142                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2143                 size_t count, size_t ocount)
2144 {
2145         struct file     *file = iocb->ki_filp;
2146         struct address_space *mapping = file->f_mapping;
2147         struct inode    *inode = mapping->host;
2148         ssize_t         written;
2149         size_t          write_len;
2150         pgoff_t         end;
2151
2152         if (count != ocount)
2153                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2154
2155         write_len = iov_length(iov, *nr_segs);
2156         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2157
2158         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2159         if (written)
2160                 goto out;
2161
2162         /*
2163          * After a write we want buffered reads to be sure to go to disk to get
2164          * the new data.  We invalidate clean cached page from the region we're
2165          * about to write.  We do this *before* the write so that we can return
2166          * without clobbering -EIOCBQUEUED from ->direct_IO().
2167          */
2168         if (mapping->nrpages) {
2169                 written = invalidate_inode_pages2_range(mapping,
2170                                         pos >> PAGE_CACHE_SHIFT, end);
2171                 /*
2172                  * If a page can not be invalidated, return 0 to fall back
2173                  * to buffered write.
2174                  */
2175                 if (written) {
2176                         if (written == -EBUSY)
2177                                 return 0;
2178                         goto out;
2179                 }
2180         }
2181
2182         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2183
2184         /*
2185          * Finally, try again to invalidate clean pages which might have been
2186          * cached by non-direct readahead, or faulted in by get_user_pages()
2187          * if the source of the write was an mmap'ed region of the file
2188          * we're writing.  Either one is a pretty crazy thing to do,
2189          * so we don't support it 100%.  If this invalidation
2190          * fails, tough, the write still worked...
2191          */
2192         if (mapping->nrpages) {
2193                 invalidate_inode_pages2_range(mapping,
2194                                               pos >> PAGE_CACHE_SHIFT, end);
2195         }
2196
2197         if (written > 0) {
2198                 pos += written;
2199                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2200                         i_size_write(inode, pos);
2201                         mark_inode_dirty(inode);
2202                 }
2203                 *ppos = pos;
2204         }
2205 out:
2206         return written;
2207 }
2208 EXPORT_SYMBOL(generic_file_direct_write);
2209
2210 /*
2211  * Find or create a page at the given pagecache position. Return the locked
2212  * page. This function is specifically for buffered writes.
2213  */
2214 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2215                                         pgoff_t index, unsigned flags)
2216 {
2217         int status;
2218         struct page *page;
2219         gfp_t gfp_notmask = 0;
2220         if (flags & AOP_FLAG_NOFS)
2221                 gfp_notmask = __GFP_FS;
2222 repeat:
2223         page = find_lock_page(mapping, index);
2224         if (likely(page))
2225                 return page;
2226
2227         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2228         if (!page)
2229                 return NULL;
2230         status = add_to_page_cache_lru(page, mapping, index,
2231                                                 GFP_KERNEL & ~gfp_notmask);
2232         if (unlikely(status)) {
2233                 page_cache_release(page);
2234                 if (status == -EEXIST)
2235                         goto repeat;
2236                 return NULL;
2237         }
2238         return page;
2239 }
2240 EXPORT_SYMBOL(grab_cache_page_write_begin);
2241
2242 static ssize_t generic_perform_write(struct file *file,
2243                                 struct iov_iter *i, loff_t pos)
2244 {
2245         struct address_space *mapping = file->f_mapping;
2246         const struct address_space_operations *a_ops = mapping->a_ops;
2247         long status = 0;
2248         ssize_t written = 0;
2249         unsigned int flags = 0;
2250
2251         /*
2252          * Copies from kernel address space cannot fail (NFSD is a big user).
2253          */
2254         if (segment_eq(get_fs(), KERNEL_DS))
2255                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2256
2257         do {
2258                 struct page *page;
2259                 unsigned long offset;   /* Offset into pagecache page */
2260                 unsigned long bytes;    /* Bytes to write to page */
2261                 size_t copied;          /* Bytes copied from user */
2262                 void *fsdata;
2263
2264                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2265                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2266                                                 iov_iter_count(i));
2267
2268 again:
2269
2270                 /*
2271                  * Bring in the user page that we will copy from _first_.
2272                  * Otherwise there's a nasty deadlock on copying from the
2273                  * same page as we're writing to, without it being marked
2274                  * up-to-date.
2275                  *
2276                  * Not only is this an optimisation, but it is also required
2277                  * to check that the address is actually valid, when atomic
2278                  * usercopies are used, below.
2279                  */
2280                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2281                         status = -EFAULT;
2282                         break;
2283                 }
2284
2285                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2286                                                 &page, &fsdata);
2287                 if (unlikely(status))
2288                         break;
2289
2290                 if (mapping_writably_mapped(mapping))
2291                         flush_dcache_page(page);
2292
2293                 pagefault_disable();
2294                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2295                 pagefault_enable();
2296                 flush_dcache_page(page);
2297
2298                 mark_page_accessed(page);
2299                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2300                                                 page, fsdata);
2301                 if (unlikely(status < 0))
2302                         break;
2303                 copied = status;
2304
2305                 cond_resched();
2306
2307                 iov_iter_advance(i, copied);
2308                 if (unlikely(copied == 0)) {
2309                         /*
2310                          * If we were unable to copy any data at all, we must
2311                          * fall back to a single segment length write.
2312                          *
2313                          * If we didn't fallback here, we could livelock
2314                          * because not all segments in the iov can be copied at
2315                          * once without a pagefault.
2316                          */
2317                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2318                                                 iov_iter_single_seg_count(i));
2319                         goto again;
2320                 }
2321                 pos += copied;
2322                 written += copied;
2323
2324                 balance_dirty_pages_ratelimited(mapping);
2325
2326         } while (iov_iter_count(i));
2327
2328         return written ? written : status;
2329 }
2330
2331 ssize_t
2332 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2333                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2334                 size_t count, ssize_t written)
2335 {
2336         struct file *file = iocb->ki_filp;
2337         ssize_t status;
2338         struct iov_iter i;
2339
2340         iov_iter_init(&i, iov, nr_segs, count, written);
2341         status = generic_perform_write(file, &i, pos);
2342
2343         if (likely(status >= 0)) {
2344                 written += status;
2345                 *ppos = pos + status;
2346         }
2347         
2348         return written ? written : status;
2349 }
2350 EXPORT_SYMBOL(generic_file_buffered_write);
2351
2352 /**
2353  * __generic_file_aio_write - write data to a file
2354  * @iocb:       IO state structure (file, offset, etc.)
2355  * @iov:        vector with data to write
2356  * @nr_segs:    number of segments in the vector
2357  * @ppos:       position where to write
2358  *
2359  * This function does all the work needed for actually writing data to a
2360  * file. It does all basic checks, removes SUID from the file, updates
2361  * modification times and calls proper subroutines depending on whether we
2362  * do direct IO or a standard buffered write.
2363  *
2364  * It expects i_mutex to be grabbed unless we work on a block device or similar
2365  * object which does not need locking at all.
2366  *
2367  * This function does *not* take care of syncing data in case of O_SYNC write.
2368  * A caller has to handle it. This is mainly due to the fact that we want to
2369  * avoid syncing under i_mutex.
2370  */
2371 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2372                                  unsigned long nr_segs, loff_t *ppos)
2373 {
2374         struct file *file = iocb->ki_filp;
2375         struct address_space * mapping = file->f_mapping;
2376         size_t ocount;          /* original count */
2377         size_t count;           /* after file limit checks */
2378         struct inode    *inode = mapping->host;
2379         loff_t          pos;
2380         ssize_t         written;
2381         ssize_t         err;
2382
2383         ocount = 0;
2384         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2385         if (err)
2386                 return err;
2387
2388         count = ocount;
2389         pos = *ppos;
2390
2391         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2392
2393         /* We can write back this queue in page reclaim */
2394         current->backing_dev_info = mapping->backing_dev_info;
2395         written = 0;
2396
2397         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2398         if (err)
2399                 goto out;
2400
2401         if (count == 0)
2402                 goto out;
2403
2404         err = file_remove_suid(file);
2405         if (err)
2406                 goto out;
2407
2408         file_update_time(file);
2409
2410         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2411         if (unlikely(file->f_flags & O_DIRECT)) {
2412                 loff_t endbyte;
2413                 ssize_t written_buffered;
2414
2415                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2416                                                         ppos, count, ocount);
2417                 if (written < 0 || written == count)
2418                         goto out;
2419                 /*
2420                  * direct-io write to a hole: fall through to buffered I/O
2421                  * for completing the rest of the request.
2422                  */
2423                 pos += written;
2424                 count -= written;
2425                 written_buffered = generic_file_buffered_write(iocb, iov,
2426                                                 nr_segs, pos, ppos, count,
2427                                                 written);
2428                 /*
2429                  * If generic_file_buffered_write() retuned a synchronous error
2430                  * then we want to return the number of bytes which were
2431                  * direct-written, or the error code if that was zero.  Note
2432                  * that this differs from normal direct-io semantics, which
2433                  * will return -EFOO even if some bytes were written.
2434                  */
2435                 if (written_buffered < 0) {
2436                         err = written_buffered;
2437                         goto out;
2438                 }
2439
2440                 /*
2441                  * We need to ensure that the page cache pages are written to
2442                  * disk and invalidated to preserve the expected O_DIRECT
2443                  * semantics.
2444                  */
2445                 endbyte = pos + written_buffered - written - 1;
2446                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2447                 if (err == 0) {
2448                         written = written_buffered;
2449                         invalidate_mapping_pages(mapping,
2450                                                  pos >> PAGE_CACHE_SHIFT,
2451                                                  endbyte >> PAGE_CACHE_SHIFT);
2452                 } else {
2453                         /*
2454                          * We don't know how much we wrote, so just return
2455                          * the number of bytes which were direct-written
2456                          */
2457                 }
2458         } else {
2459                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2460                                 pos, ppos, count, written);
2461         }
2462 out:
2463         current->backing_dev_info = NULL;
2464         return written ? written : err;
2465 }
2466 EXPORT_SYMBOL(__generic_file_aio_write);
2467
2468 /**
2469  * generic_file_aio_write - write data to a file
2470  * @iocb:       IO state structure
2471  * @iov:        vector with data to write
2472  * @nr_segs:    number of segments in the vector
2473  * @pos:        position in file where to write
2474  *
2475  * This is a wrapper around __generic_file_aio_write() to be used by most
2476  * filesystems. It takes care of syncing the file in case of O_SYNC file
2477  * and acquires i_mutex as needed.
2478  */
2479 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2480                 unsigned long nr_segs, loff_t pos)
2481 {
2482         struct file *file = iocb->ki_filp;
2483         struct inode *inode = file->f_mapping->host;
2484         ssize_t ret;
2485
2486         BUG_ON(iocb->ki_pos != pos);
2487
2488         mutex_lock(&inode->i_mutex);
2489         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2490         mutex_unlock(&inode->i_mutex);
2491
2492         if (ret > 0 || ret == -EIOCBQUEUED) {
2493                 ssize_t err;
2494
2495                 err = generic_write_sync(file, pos, ret);
2496                 if (err < 0 && ret > 0)
2497                         ret = err;
2498         }
2499         return ret;
2500 }
2501 EXPORT_SYMBOL(generic_file_aio_write);
2502
2503 /**
2504  * try_to_release_page() - release old fs-specific metadata on a page
2505  *
2506  * @page: the page which the kernel is trying to free
2507  * @gfp_mask: memory allocation flags (and I/O mode)
2508  *
2509  * The address_space is to try to release any data against the page
2510  * (presumably at page->private).  If the release was successful, return `1'.
2511  * Otherwise return zero.
2512  *
2513  * This may also be called if PG_fscache is set on a page, indicating that the
2514  * page is known to the local caching routines.
2515  *
2516  * The @gfp_mask argument specifies whether I/O may be performed to release
2517  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2518  *
2519  */
2520 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2521 {
2522         struct address_space * const mapping = page->mapping;
2523
2524         BUG_ON(!PageLocked(page));
2525         if (PageWriteback(page))
2526                 return 0;
2527
2528         if (mapping && mapping->a_ops->releasepage)
2529                 return mapping->a_ops->releasepage(page, gfp_mask);
2530         return try_to_free_buffers(page);
2531 }
2532
2533 EXPORT_SYMBOL(try_to_release_page);