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