mm: filemap_fault: unique path for locking page
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  *
108  *  (code doesn't rely on that order, so you could switch it around)
109  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
110  *    ->i_mmap_lock
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         if (PageSwapBacked(page))
127                 __dec_zone_page_state(page, NR_SHMEM);
128         BUG_ON(page_mapped(page));
129
130         /*
131          * Some filesystems seem to re-dirty the page even after
132          * the VM has canceled the dirty bit (eg ext3 journaling).
133          *
134          * Fix it up by doing a final dirty accounting check after
135          * having removed the page entirely.
136          */
137         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
138                 dec_zone_page_state(page, NR_FILE_DIRTY);
139                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
140         }
141 }
142
143 void remove_from_page_cache(struct page *page)
144 {
145         struct address_space *mapping = page->mapping;
146
147         BUG_ON(!PageLocked(page));
148
149         spin_lock_irq(&mapping->tree_lock);
150         __remove_from_page_cache(page);
151         spin_unlock_irq(&mapping->tree_lock);
152         mem_cgroup_uncharge_cache_page(page);
153 }
154 EXPORT_SYMBOL(remove_from_page_cache);
155
156 static int sync_page(void *word)
157 {
158         struct address_space *mapping;
159         struct page *page;
160
161         page = container_of((unsigned long *)word, struct page, flags);
162
163         /*
164          * page_mapping() is being called without PG_locked held.
165          * Some knowledge of the state and use of the page is used to
166          * reduce the requirements down to a memory barrier.
167          * The danger here is of a stale page_mapping() return value
168          * indicating a struct address_space different from the one it's
169          * associated with when it is associated with one.
170          * After smp_mb(), it's either the correct page_mapping() for
171          * the page, or an old page_mapping() and the page's own
172          * page_mapping() has gone NULL.
173          * The ->sync_page() address_space operation must tolerate
174          * page_mapping() going NULL. By an amazing coincidence,
175          * this comes about because none of the users of the page
176          * in the ->sync_page() methods make essential use of the
177          * page_mapping(), merely passing the page down to the backing
178          * device's unplug functions when it's non-NULL, which in turn
179          * ignore it for all cases but swap, where only page_private(page) is
180          * of interest. When page_mapping() does go NULL, the entire
181          * call stack gracefully ignores the page and returns.
182          * -- wli
183          */
184         smp_mb();
185         mapping = page_mapping(page);
186         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
187                 mapping->a_ops->sync_page(page);
188         io_schedule();
189         return 0;
190 }
191
192 static int sync_page_killable(void *word)
193 {
194         sync_page(word);
195         return fatal_signal_pending(current) ? -EINTR : 0;
196 }
197
198 /**
199  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
200  * @mapping:    address space structure to write
201  * @start:      offset in bytes where the range starts
202  * @end:        offset in bytes where the range ends (inclusive)
203  * @sync_mode:  enable synchronous operation
204  *
205  * Start writeback against all of a mapping's dirty pages that lie
206  * within the byte offsets <start, end> inclusive.
207  *
208  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
209  * opposed to a regular memory cleansing writeback.  The difference between
210  * these two operations is that if a dirty page/buffer is encountered, it must
211  * be waited upon, and not just skipped over.
212  */
213 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
214                                 loff_t end, int sync_mode)
215 {
216         int ret;
217         struct writeback_control wbc = {
218                 .sync_mode = sync_mode,
219                 .nr_to_write = LONG_MAX,
220                 .range_start = start,
221                 .range_end = end,
222         };
223
224         if (!mapping_cap_writeback_dirty(mapping))
225                 return 0;
226
227         ret = do_writepages(mapping, &wbc);
228         return ret;
229 }
230
231 static inline int __filemap_fdatawrite(struct address_space *mapping,
232         int sync_mode)
233 {
234         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
235 }
236
237 int filemap_fdatawrite(struct address_space *mapping)
238 {
239         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite);
242
243 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
244                                 loff_t end)
245 {
246         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
247 }
248 EXPORT_SYMBOL(filemap_fdatawrite_range);
249
250 /**
251  * filemap_flush - mostly a non-blocking flush
252  * @mapping:    target address_space
253  *
254  * This is a mostly non-blocking flush.  Not suitable for data-integrity
255  * purposes - I/O may not be started against all dirty pages.
256  */
257 int filemap_flush(struct address_space *mapping)
258 {
259         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
260 }
261 EXPORT_SYMBOL(filemap_flush);
262
263 /**
264  * filemap_fdatawait_range - wait for writeback to complete
265  * @mapping:            address space structure to wait for
266  * @start_byte:         offset in bytes where the range starts
267  * @end_byte:           offset in bytes where the range ends (inclusive)
268  *
269  * Walk the list of under-writeback pages of the given address space
270  * in the given range and wait for all of them.
271  */
272 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
273                             loff_t end_byte)
274 {
275         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
276         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
277         struct pagevec pvec;
278         int nr_pages;
279         int ret = 0;
280
281         if (end_byte < start_byte)
282                 return 0;
283
284         pagevec_init(&pvec, 0);
285         while ((index <= end) &&
286                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
287                         PAGECACHE_TAG_WRITEBACK,
288                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
289                 unsigned i;
290
291                 for (i = 0; i < nr_pages; i++) {
292                         struct page *page = pvec.pages[i];
293
294                         /* until radix tree lookup accepts end_index */
295                         if (page->index > end)
296                                 continue;
297
298                         wait_on_page_writeback(page);
299                         if (PageError(page))
300                                 ret = -EIO;
301                 }
302                 pagevec_release(&pvec);
303                 cond_resched();
304         }
305
306         /* Check for outstanding write errors */
307         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
308                 ret = -ENOSPC;
309         if (test_and_clear_bit(AS_EIO, &mapping->flags))
310                 ret = -EIO;
311
312         return ret;
313 }
314 EXPORT_SYMBOL(filemap_fdatawait_range);
315
316 /**
317  * filemap_fdatawait - wait for all under-writeback pages to complete
318  * @mapping: address space structure to wait for
319  *
320  * Walk the list of under-writeback pages of the given address space
321  * and wait for all of them.
322  */
323 int filemap_fdatawait(struct address_space *mapping)
324 {
325         loff_t i_size = i_size_read(mapping->host);
326
327         if (i_size == 0)
328                 return 0;
329
330         return filemap_fdatawait_range(mapping, 0, i_size - 1);
331 }
332 EXPORT_SYMBOL(filemap_fdatawait);
333
334 int filemap_write_and_wait(struct address_space *mapping)
335 {
336         int err = 0;
337
338         if (mapping->nrpages) {
339                 err = filemap_fdatawrite(mapping);
340                 /*
341                  * Even if the above returned error, the pages may be
342                  * written partially (e.g. -ENOSPC), so we wait for it.
343                  * But the -EIO is special case, it may indicate the worst
344                  * thing (e.g. bug) happened, so we avoid waiting for it.
345                  */
346                 if (err != -EIO) {
347                         int err2 = filemap_fdatawait(mapping);
348                         if (!err)
349                                 err = err2;
350                 }
351         }
352         return err;
353 }
354 EXPORT_SYMBOL(filemap_write_and_wait);
355
356 /**
357  * filemap_write_and_wait_range - write out & wait on a file range
358  * @mapping:    the address_space for the pages
359  * @lstart:     offset in bytes where the range starts
360  * @lend:       offset in bytes where the range ends (inclusive)
361  *
362  * Write out and wait upon file offsets lstart->lend, inclusive.
363  *
364  * Note that `lend' is inclusive (describes the last byte to be written) so
365  * that this function can be used to write to the very end-of-file (end = -1).
366  */
367 int filemap_write_and_wait_range(struct address_space *mapping,
368                                  loff_t lstart, loff_t lend)
369 {
370         int err = 0;
371
372         if (mapping->nrpages) {
373                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
374                                                  WB_SYNC_ALL);
375                 /* See comment of filemap_write_and_wait() */
376                 if (err != -EIO) {
377                         int err2 = filemap_fdatawait_range(mapping,
378                                                 lstart, lend);
379                         if (!err)
380                                 err = err2;
381                 }
382         }
383         return err;
384 }
385 EXPORT_SYMBOL(filemap_write_and_wait_range);
386
387 /**
388  * add_to_page_cache_locked - add a locked page to the pagecache
389  * @page:       page to add
390  * @mapping:    the page's address_space
391  * @offset:     page index
392  * @gfp_mask:   page allocation mode
393  *
394  * This function is used to add a page to the pagecache. It must be locked.
395  * This function does not add the page to the LRU.  The caller must do that.
396  */
397 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
398                 pgoff_t offset, gfp_t gfp_mask)
399 {
400         int error;
401
402         VM_BUG_ON(!PageLocked(page));
403
404         error = mem_cgroup_cache_charge(page, current->mm,
405                                         gfp_mask & GFP_RECLAIM_MASK);
406         if (error)
407                 goto out;
408
409         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
410         if (error == 0) {
411                 page_cache_get(page);
412                 page->mapping = mapping;
413                 page->index = offset;
414
415                 spin_lock_irq(&mapping->tree_lock);
416                 error = radix_tree_insert(&mapping->page_tree, offset, page);
417                 if (likely(!error)) {
418                         mapping->nrpages++;
419                         __inc_zone_page_state(page, NR_FILE_PAGES);
420                         if (PageSwapBacked(page))
421                                 __inc_zone_page_state(page, NR_SHMEM);
422                         spin_unlock_irq(&mapping->tree_lock);
423                 } else {
424                         page->mapping = NULL;
425                         spin_unlock_irq(&mapping->tree_lock);
426                         mem_cgroup_uncharge_cache_page(page);
427                         page_cache_release(page);
428                 }
429                 radix_tree_preload_end();
430         } else
431                 mem_cgroup_uncharge_cache_page(page);
432 out:
433         return error;
434 }
435 EXPORT_SYMBOL(add_to_page_cache_locked);
436
437 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
438                                 pgoff_t offset, gfp_t gfp_mask)
439 {
440         int ret;
441
442         /*
443          * Splice_read and readahead add shmem/tmpfs pages into the page cache
444          * before shmem_readpage has a chance to mark them as SwapBacked: they
445          * need to go on the anon lru below, and mem_cgroup_cache_charge
446          * (called in add_to_page_cache) needs to know where they're going too.
447          */
448         if (mapping_cap_swap_backed(mapping))
449                 SetPageSwapBacked(page);
450
451         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
452         if (ret == 0) {
453                 if (page_is_file_cache(page))
454                         lru_cache_add_file(page);
455                 else
456                         lru_cache_add_anon(page);
457         }
458         return ret;
459 }
460 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
461
462 #ifdef CONFIG_NUMA
463 struct page *__page_cache_alloc(gfp_t gfp)
464 {
465         int n;
466         struct page *page;
467
468         if (cpuset_do_page_mem_spread()) {
469                 get_mems_allowed();
470                 n = cpuset_mem_spread_node();
471                 page = alloc_pages_exact_node(n, gfp, 0);
472                 put_mems_allowed();
473                 return page;
474         }
475         return alloc_pages(gfp, 0);
476 }
477 EXPORT_SYMBOL(__page_cache_alloc);
478 #endif
479
480 static int __sleep_on_page_lock(void *word)
481 {
482         io_schedule();
483         return 0;
484 }
485
486 /*
487  * In order to wait for pages to become available there must be
488  * waitqueues associated with pages. By using a hash table of
489  * waitqueues where the bucket discipline is to maintain all
490  * waiters on the same queue and wake all when any of the pages
491  * become available, and for the woken contexts to check to be
492  * sure the appropriate page became available, this saves space
493  * at a cost of "thundering herd" phenomena during rare hash
494  * collisions.
495  */
496 static wait_queue_head_t *page_waitqueue(struct page *page)
497 {
498         const struct zone *zone = page_zone(page);
499
500         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
501 }
502
503 static inline void wake_up_page(struct page *page, int bit)
504 {
505         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
506 }
507
508 void wait_on_page_bit(struct page *page, int bit_nr)
509 {
510         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
511
512         if (test_bit(bit_nr, &page->flags))
513                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
514                                                         TASK_UNINTERRUPTIBLE);
515 }
516 EXPORT_SYMBOL(wait_on_page_bit);
517
518 /**
519  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
520  * @page: Page defining the wait queue of interest
521  * @waiter: Waiter to add to the queue
522  *
523  * Add an arbitrary @waiter to the wait queue for the nominated @page.
524  */
525 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
526 {
527         wait_queue_head_t *q = page_waitqueue(page);
528         unsigned long flags;
529
530         spin_lock_irqsave(&q->lock, flags);
531         __add_wait_queue(q, waiter);
532         spin_unlock_irqrestore(&q->lock, flags);
533 }
534 EXPORT_SYMBOL_GPL(add_page_wait_queue);
535
536 /**
537  * unlock_page - unlock a locked page
538  * @page: the page
539  *
540  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
541  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
542  * mechananism between PageLocked pages and PageWriteback pages is shared.
543  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
544  *
545  * The mb is necessary to enforce ordering between the clear_bit and the read
546  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
547  */
548 void unlock_page(struct page *page)
549 {
550         VM_BUG_ON(!PageLocked(page));
551         clear_bit_unlock(PG_locked, &page->flags);
552         smp_mb__after_clear_bit();
553         wake_up_page(page, PG_locked);
554 }
555 EXPORT_SYMBOL(unlock_page);
556
557 /**
558  * end_page_writeback - end writeback against a page
559  * @page: the page
560  */
561 void end_page_writeback(struct page *page)
562 {
563         if (TestClearPageReclaim(page))
564                 rotate_reclaimable_page(page);
565
566         if (!test_clear_page_writeback(page))
567                 BUG();
568
569         smp_mb__after_clear_bit();
570         wake_up_page(page, PG_writeback);
571 }
572 EXPORT_SYMBOL(end_page_writeback);
573
574 /**
575  * __lock_page - get a lock on the page, assuming we need to sleep to get it
576  * @page: the page to lock
577  *
578  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
579  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
580  * chances are that on the second loop, the block layer's plug list is empty,
581  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
582  */
583 void __lock_page(struct page *page)
584 {
585         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
586
587         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
588                                                         TASK_UNINTERRUPTIBLE);
589 }
590 EXPORT_SYMBOL(__lock_page);
591
592 int __lock_page_killable(struct page *page)
593 {
594         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
595
596         return __wait_on_bit_lock(page_waitqueue(page), &wait,
597                                         sync_page_killable, TASK_KILLABLE);
598 }
599 EXPORT_SYMBOL_GPL(__lock_page_killable);
600
601 /**
602  * __lock_page_nosync - get a lock on the page, without calling sync_page()
603  * @page: the page to lock
604  *
605  * Variant of lock_page that does not require the caller to hold a reference
606  * on the page's mapping.
607  */
608 void __lock_page_nosync(struct page *page)
609 {
610         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
611         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
612                                                         TASK_UNINTERRUPTIBLE);
613 }
614
615 /**
616  * find_get_page - find and get a page reference
617  * @mapping: the address_space to search
618  * @offset: the page index
619  *
620  * Is there a pagecache struct page at the given (mapping, offset) tuple?
621  * If yes, increment its refcount and return it; if no, return NULL.
622  */
623 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
624 {
625         void **pagep;
626         struct page *page;
627
628         rcu_read_lock();
629 repeat:
630         page = NULL;
631         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
632         if (pagep) {
633                 page = radix_tree_deref_slot(pagep);
634                 if (unlikely(!page || page == RADIX_TREE_RETRY))
635                         goto repeat;
636
637                 if (!page_cache_get_speculative(page))
638                         goto repeat;
639
640                 /*
641                  * Has the page moved?
642                  * This is part of the lockless pagecache protocol. See
643                  * include/linux/pagemap.h for details.
644                  */
645                 if (unlikely(page != *pagep)) {
646                         page_cache_release(page);
647                         goto repeat;
648                 }
649         }
650         rcu_read_unlock();
651
652         return page;
653 }
654 EXPORT_SYMBOL(find_get_page);
655
656 /**
657  * find_lock_page - locate, pin and lock a pagecache page
658  * @mapping: the address_space to search
659  * @offset: the page index
660  *
661  * Locates the desired pagecache page, locks it, increments its reference
662  * count and returns its address.
663  *
664  * Returns zero if the page was not present. find_lock_page() may sleep.
665  */
666 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
667 {
668         struct page *page;
669
670 repeat:
671         page = find_get_page(mapping, offset);
672         if (page) {
673                 lock_page(page);
674                 /* Has the page been truncated? */
675                 if (unlikely(page->mapping != mapping)) {
676                         unlock_page(page);
677                         page_cache_release(page);
678                         goto repeat;
679                 }
680                 VM_BUG_ON(page->index != offset);
681         }
682         return page;
683 }
684 EXPORT_SYMBOL(find_lock_page);
685
686 /**
687  * find_or_create_page - locate or add a pagecache page
688  * @mapping: the page's address_space
689  * @index: the page's index into the mapping
690  * @gfp_mask: page allocation mode
691  *
692  * Locates a page in the pagecache.  If the page is not present, a new page
693  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
694  * LRU list.  The returned page is locked and has its reference count
695  * incremented.
696  *
697  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
698  * allocation!
699  *
700  * find_or_create_page() returns the desired page's address, or zero on
701  * memory exhaustion.
702  */
703 struct page *find_or_create_page(struct address_space *mapping,
704                 pgoff_t index, gfp_t gfp_mask)
705 {
706         struct page *page;
707         int err;
708 repeat:
709         page = find_lock_page(mapping, index);
710         if (!page) {
711                 page = __page_cache_alloc(gfp_mask);
712                 if (!page)
713                         return NULL;
714                 /*
715                  * We want a regular kernel memory (not highmem or DMA etc)
716                  * allocation for the radix tree nodes, but we need to honour
717                  * the context-specific requirements the caller has asked for.
718                  * GFP_RECLAIM_MASK collects those requirements.
719                  */
720                 err = add_to_page_cache_lru(page, mapping, index,
721                         (gfp_mask & GFP_RECLAIM_MASK));
722                 if (unlikely(err)) {
723                         page_cache_release(page);
724                         page = NULL;
725                         if (err == -EEXIST)
726                                 goto repeat;
727                 }
728         }
729         return page;
730 }
731 EXPORT_SYMBOL(find_or_create_page);
732
733 /**
734  * find_get_pages - gang pagecache lookup
735  * @mapping:    The address_space to search
736  * @start:      The starting page index
737  * @nr_pages:   The maximum number of pages
738  * @pages:      Where the resulting pages are placed
739  *
740  * find_get_pages() will search for and return a group of up to
741  * @nr_pages pages in the mapping.  The pages are placed at @pages.
742  * find_get_pages() takes a reference against the returned pages.
743  *
744  * The search returns a group of mapping-contiguous pages with ascending
745  * indexes.  There may be holes in the indices due to not-present pages.
746  *
747  * find_get_pages() returns the number of pages which were found.
748  */
749 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
750                             unsigned int nr_pages, struct page **pages)
751 {
752         unsigned int i;
753         unsigned int ret;
754         unsigned int nr_found;
755
756         rcu_read_lock();
757 restart:
758         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
759                                 (void ***)pages, start, nr_pages);
760         ret = 0;
761         for (i = 0; i < nr_found; i++) {
762                 struct page *page;
763 repeat:
764                 page = radix_tree_deref_slot((void **)pages[i]);
765                 if (unlikely(!page))
766                         continue;
767                 /*
768                  * this can only trigger if nr_found == 1, making livelock
769                  * a non issue.
770                  */
771                 if (unlikely(page == RADIX_TREE_RETRY))
772                         goto restart;
773
774                 if (!page_cache_get_speculative(page))
775                         goto repeat;
776
777                 /* Has the page moved? */
778                 if (unlikely(page != *((void **)pages[i]))) {
779                         page_cache_release(page);
780                         goto repeat;
781                 }
782
783                 pages[ret] = page;
784                 ret++;
785         }
786         rcu_read_unlock();
787         return ret;
788 }
789
790 /**
791  * find_get_pages_contig - gang contiguous pagecache lookup
792  * @mapping:    The address_space to search
793  * @index:      The starting page index
794  * @nr_pages:   The maximum number of pages
795  * @pages:      Where the resulting pages are placed
796  *
797  * find_get_pages_contig() works exactly like find_get_pages(), except
798  * that the returned number of pages are guaranteed to be contiguous.
799  *
800  * find_get_pages_contig() returns the number of pages which were found.
801  */
802 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
803                                unsigned int nr_pages, struct page **pages)
804 {
805         unsigned int i;
806         unsigned int ret;
807         unsigned int nr_found;
808
809         rcu_read_lock();
810 restart:
811         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
812                                 (void ***)pages, index, nr_pages);
813         ret = 0;
814         for (i = 0; i < nr_found; i++) {
815                 struct page *page;
816 repeat:
817                 page = radix_tree_deref_slot((void **)pages[i]);
818                 if (unlikely(!page))
819                         continue;
820                 /*
821                  * this can only trigger if nr_found == 1, making livelock
822                  * a non issue.
823                  */
824                 if (unlikely(page == RADIX_TREE_RETRY))
825                         goto restart;
826
827                 if (page->mapping == NULL || page->index != index)
828                         break;
829
830                 if (!page_cache_get_speculative(page))
831                         goto repeat;
832
833                 /* Has the page moved? */
834                 if (unlikely(page != *((void **)pages[i]))) {
835                         page_cache_release(page);
836                         goto repeat;
837                 }
838
839                 pages[ret] = page;
840                 ret++;
841                 index++;
842         }
843         rcu_read_unlock();
844         return ret;
845 }
846 EXPORT_SYMBOL(find_get_pages_contig);
847
848 /**
849  * find_get_pages_tag - find and return pages that match @tag
850  * @mapping:    the address_space to search
851  * @index:      the starting page index
852  * @tag:        the tag index
853  * @nr_pages:   the maximum number of pages
854  * @pages:      where the resulting pages are placed
855  *
856  * Like find_get_pages, except we only return pages which are tagged with
857  * @tag.   We update @index to index the next page for the traversal.
858  */
859 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
860                         int tag, unsigned int nr_pages, struct page **pages)
861 {
862         unsigned int i;
863         unsigned int ret;
864         unsigned int nr_found;
865
866         rcu_read_lock();
867 restart:
868         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
869                                 (void ***)pages, *index, nr_pages, tag);
870         ret = 0;
871         for (i = 0; i < nr_found; i++) {
872                 struct page *page;
873 repeat:
874                 page = radix_tree_deref_slot((void **)pages[i]);
875                 if (unlikely(!page))
876                         continue;
877                 /*
878                  * this can only trigger if nr_found == 1, making livelock
879                  * a non issue.
880                  */
881                 if (unlikely(page == RADIX_TREE_RETRY))
882                         goto restart;
883
884                 if (!page_cache_get_speculative(page))
885                         goto repeat;
886
887                 /* Has the page moved? */
888                 if (unlikely(page != *((void **)pages[i]))) {
889                         page_cache_release(page);
890                         goto repeat;
891                 }
892
893                 pages[ret] = page;
894                 ret++;
895         }
896         rcu_read_unlock();
897
898         if (ret)
899                 *index = pages[ret - 1]->index + 1;
900
901         return ret;
902 }
903 EXPORT_SYMBOL(find_get_pages_tag);
904
905 /**
906  * grab_cache_page_nowait - returns locked page at given index in given cache
907  * @mapping: target address_space
908  * @index: the page index
909  *
910  * Same as grab_cache_page(), but do not wait if the page is unavailable.
911  * This is intended for speculative data generators, where the data can
912  * be regenerated if the page couldn't be grabbed.  This routine should
913  * be safe to call while holding the lock for another page.
914  *
915  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
916  * and deadlock against the caller's locked page.
917  */
918 struct page *
919 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
920 {
921         struct page *page = find_get_page(mapping, index);
922
923         if (page) {
924                 if (trylock_page(page))
925                         return page;
926                 page_cache_release(page);
927                 return NULL;
928         }
929         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
930         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
931                 page_cache_release(page);
932                 page = NULL;
933         }
934         return page;
935 }
936 EXPORT_SYMBOL(grab_cache_page_nowait);
937
938 /*
939  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
940  * a _large_ part of the i/o request. Imagine the worst scenario:
941  *
942  *      ---R__________________________________________B__________
943  *         ^ reading here                             ^ bad block(assume 4k)
944  *
945  * read(R) => miss => readahead(R...B) => media error => frustrating retries
946  * => failing the whole request => read(R) => read(R+1) =>
947  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
948  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
949  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
950  *
951  * It is going insane. Fix it by quickly scaling down the readahead size.
952  */
953 static void shrink_readahead_size_eio(struct file *filp,
954                                         struct file_ra_state *ra)
955 {
956         ra->ra_pages /= 4;
957 }
958
959 /**
960  * do_generic_file_read - generic file read routine
961  * @filp:       the file to read
962  * @ppos:       current file position
963  * @desc:       read_descriptor
964  * @actor:      read method
965  *
966  * This is a generic file read routine, and uses the
967  * mapping->a_ops->readpage() function for the actual low-level stuff.
968  *
969  * This is really ugly. But the goto's actually try to clarify some
970  * of the logic when it comes to error handling etc.
971  */
972 static void do_generic_file_read(struct file *filp, loff_t *ppos,
973                 read_descriptor_t *desc, read_actor_t actor)
974 {
975         struct address_space *mapping = filp->f_mapping;
976         struct inode *inode = mapping->host;
977         struct file_ra_state *ra = &filp->f_ra;
978         pgoff_t index;
979         pgoff_t last_index;
980         pgoff_t prev_index;
981         unsigned long offset;      /* offset into pagecache page */
982         unsigned int prev_offset;
983         int error;
984
985         index = *ppos >> PAGE_CACHE_SHIFT;
986         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
987         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
988         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
989         offset = *ppos & ~PAGE_CACHE_MASK;
990
991         for (;;) {
992                 struct page *page;
993                 pgoff_t end_index;
994                 loff_t isize;
995                 unsigned long nr, ret;
996
997                 cond_resched();
998 find_page:
999                 page = find_get_page(mapping, index);
1000                 if (!page) {
1001                         page_cache_sync_readahead(mapping,
1002                                         ra, filp,
1003                                         index, last_index - index);
1004                         page = find_get_page(mapping, index);
1005                         if (unlikely(page == NULL))
1006                                 goto no_cached_page;
1007                 }
1008                 if (PageReadahead(page)) {
1009                         page_cache_async_readahead(mapping,
1010                                         ra, filp, page,
1011                                         index, last_index - index);
1012                 }
1013                 if (!PageUptodate(page)) {
1014                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1015                                         !mapping->a_ops->is_partially_uptodate)
1016                                 goto page_not_up_to_date;
1017                         if (!trylock_page(page))
1018                                 goto page_not_up_to_date;
1019                         if (!mapping->a_ops->is_partially_uptodate(page,
1020                                                                 desc, offset))
1021                                 goto page_not_up_to_date_locked;
1022                         unlock_page(page);
1023                 }
1024 page_ok:
1025                 /*
1026                  * i_size must be checked after we know the page is Uptodate.
1027                  *
1028                  * Checking i_size after the check allows us to calculate
1029                  * the correct value for "nr", which means the zero-filled
1030                  * part of the page is not copied back to userspace (unless
1031                  * another truncate extends the file - this is desired though).
1032                  */
1033
1034                 isize = i_size_read(inode);
1035                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1036                 if (unlikely(!isize || index > end_index)) {
1037                         page_cache_release(page);
1038                         goto out;
1039                 }
1040
1041                 /* nr is the maximum number of bytes to copy from this page */
1042                 nr = PAGE_CACHE_SIZE;
1043                 if (index == end_index) {
1044                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1045                         if (nr <= offset) {
1046                                 page_cache_release(page);
1047                                 goto out;
1048                         }
1049                 }
1050                 nr = nr - offset;
1051
1052                 /* If users can be writing to this page using arbitrary
1053                  * virtual addresses, take care about potential aliasing
1054                  * before reading the page on the kernel side.
1055                  */
1056                 if (mapping_writably_mapped(mapping))
1057                         flush_dcache_page(page);
1058
1059                 /*
1060                  * When a sequential read accesses a page several times,
1061                  * only mark it as accessed the first time.
1062                  */
1063                 if (prev_index != index || offset != prev_offset)
1064                         mark_page_accessed(page);
1065                 prev_index = index;
1066
1067                 /*
1068                  * Ok, we have the page, and it's up-to-date, so
1069                  * now we can copy it to user space...
1070                  *
1071                  * The actor routine returns how many bytes were actually used..
1072                  * NOTE! This may not be the same as how much of a user buffer
1073                  * we filled up (we may be padding etc), so we can only update
1074                  * "pos" here (the actor routine has to update the user buffer
1075                  * pointers and the remaining count).
1076                  */
1077                 ret = actor(desc, page, offset, nr);
1078                 offset += ret;
1079                 index += offset >> PAGE_CACHE_SHIFT;
1080                 offset &= ~PAGE_CACHE_MASK;
1081                 prev_offset = offset;
1082
1083                 page_cache_release(page);
1084                 if (ret == nr && desc->count)
1085                         continue;
1086                 goto out;
1087
1088 page_not_up_to_date:
1089                 /* Get exclusive access to the page ... */
1090                 error = lock_page_killable(page);
1091                 if (unlikely(error))
1092                         goto readpage_error;
1093
1094 page_not_up_to_date_locked:
1095                 /* Did it get truncated before we got the lock? */
1096                 if (!page->mapping) {
1097                         unlock_page(page);
1098                         page_cache_release(page);
1099                         continue;
1100                 }
1101
1102                 /* Did somebody else fill it already? */
1103                 if (PageUptodate(page)) {
1104                         unlock_page(page);
1105                         goto page_ok;
1106                 }
1107
1108 readpage:
1109                 /*
1110                  * A previous I/O error may have been due to temporary
1111                  * failures, eg. multipath errors.
1112                  * PG_error will be set again if readpage fails.
1113                  */
1114                 ClearPageError(page);
1115                 /* Start the actual read. The read will unlock the page. */
1116                 error = mapping->a_ops->readpage(filp, page);
1117
1118                 if (unlikely(error)) {
1119                         if (error == AOP_TRUNCATED_PAGE) {
1120                                 page_cache_release(page);
1121                                 goto find_page;
1122                         }
1123                         goto readpage_error;
1124                 }
1125
1126                 if (!PageUptodate(page)) {
1127                         error = lock_page_killable(page);
1128                         if (unlikely(error))
1129                                 goto readpage_error;
1130                         if (!PageUptodate(page)) {
1131                                 if (page->mapping == NULL) {
1132                                         /*
1133                                          * invalidate_mapping_pages got it
1134                                          */
1135                                         unlock_page(page);
1136                                         page_cache_release(page);
1137                                         goto find_page;
1138                                 }
1139                                 unlock_page(page);
1140                                 shrink_readahead_size_eio(filp, ra);
1141                                 error = -EIO;
1142                                 goto readpage_error;
1143                         }
1144                         unlock_page(page);
1145                 }
1146
1147                 goto page_ok;
1148
1149 readpage_error:
1150                 /* UHHUH! A synchronous read error occurred. Report it */
1151                 desc->error = error;
1152                 page_cache_release(page);
1153                 goto out;
1154
1155 no_cached_page:
1156                 /*
1157                  * Ok, it wasn't cached, so we need to create a new
1158                  * page..
1159                  */
1160                 page = page_cache_alloc_cold(mapping);
1161                 if (!page) {
1162                         desc->error = -ENOMEM;
1163                         goto out;
1164                 }
1165                 error = add_to_page_cache_lru(page, mapping,
1166                                                 index, GFP_KERNEL);
1167                 if (error) {
1168                         page_cache_release(page);
1169                         if (error == -EEXIST)
1170                                 goto find_page;
1171                         desc->error = error;
1172                         goto out;
1173                 }
1174                 goto readpage;
1175         }
1176
1177 out:
1178         ra->prev_pos = prev_index;
1179         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1180         ra->prev_pos |= prev_offset;
1181
1182         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1183         file_accessed(filp);
1184 }
1185
1186 int file_read_actor(read_descriptor_t *desc, struct page *page,
1187                         unsigned long offset, unsigned long size)
1188 {
1189         char *kaddr;
1190         unsigned long left, count = desc->count;
1191
1192         if (size > count)
1193                 size = count;
1194
1195         /*
1196          * Faults on the destination of a read are common, so do it before
1197          * taking the kmap.
1198          */
1199         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1200                 kaddr = kmap_atomic(page, KM_USER0);
1201                 left = __copy_to_user_inatomic(desc->arg.buf,
1202                                                 kaddr + offset, size);
1203                 kunmap_atomic(kaddr, KM_USER0);
1204                 if (left == 0)
1205                         goto success;
1206         }
1207
1208         /* Do it the slow way */
1209         kaddr = kmap(page);
1210         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1211         kunmap(page);
1212
1213         if (left) {
1214                 size -= left;
1215                 desc->error = -EFAULT;
1216         }
1217 success:
1218         desc->count = count - size;
1219         desc->written += size;
1220         desc->arg.buf += size;
1221         return size;
1222 }
1223
1224 /*
1225  * Performs necessary checks before doing a write
1226  * @iov:        io vector request
1227  * @nr_segs:    number of segments in the iovec
1228  * @count:      number of bytes to write
1229  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1230  *
1231  * Adjust number of segments and amount of bytes to write (nr_segs should be
1232  * properly initialized first). Returns appropriate error code that caller
1233  * should return or zero in case that write should be allowed.
1234  */
1235 int generic_segment_checks(const struct iovec *iov,
1236                         unsigned long *nr_segs, size_t *count, int access_flags)
1237 {
1238         unsigned long   seg;
1239         size_t cnt = 0;
1240         for (seg = 0; seg < *nr_segs; seg++) {
1241                 const struct iovec *iv = &iov[seg];
1242
1243                 /*
1244                  * If any segment has a negative length, or the cumulative
1245                  * length ever wraps negative then return -EINVAL.
1246                  */
1247                 cnt += iv->iov_len;
1248                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1249                         return -EINVAL;
1250                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1251                         continue;
1252                 if (seg == 0)
1253                         return -EFAULT;
1254                 *nr_segs = seg;
1255                 cnt -= iv->iov_len;     /* This segment is no good */
1256                 break;
1257         }
1258         *count = cnt;
1259         return 0;
1260 }
1261 EXPORT_SYMBOL(generic_segment_checks);
1262
1263 /**
1264  * generic_file_aio_read - generic filesystem read routine
1265  * @iocb:       kernel I/O control block
1266  * @iov:        io vector request
1267  * @nr_segs:    number of segments in the iovec
1268  * @pos:        current file position
1269  *
1270  * This is the "read()" routine for all filesystems
1271  * that can use the page cache directly.
1272  */
1273 ssize_t
1274 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1275                 unsigned long nr_segs, loff_t pos)
1276 {
1277         struct file *filp = iocb->ki_filp;
1278         ssize_t retval;
1279         unsigned long seg = 0;
1280         size_t count;
1281         loff_t *ppos = &iocb->ki_pos;
1282
1283         count = 0;
1284         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1285         if (retval)
1286                 return retval;
1287
1288         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1289         if (filp->f_flags & O_DIRECT) {
1290                 loff_t size;
1291                 struct address_space *mapping;
1292                 struct inode *inode;
1293
1294                 mapping = filp->f_mapping;
1295                 inode = mapping->host;
1296                 if (!count)
1297                         goto out; /* skip atime */
1298                 size = i_size_read(inode);
1299                 if (pos < size) {
1300                         retval = filemap_write_and_wait_range(mapping, pos,
1301                                         pos + iov_length(iov, nr_segs) - 1);
1302                         if (!retval) {
1303                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1304                                                         iov, pos, nr_segs);
1305                         }
1306                         if (retval > 0) {
1307                                 *ppos = pos + retval;
1308                                 count -= retval;
1309                         }
1310
1311                         /*
1312                          * Btrfs can have a short DIO read if we encounter
1313                          * compressed extents, so if there was an error, or if
1314                          * we've already read everything we wanted to, or if
1315                          * there was a short read because we hit EOF, go ahead
1316                          * and return.  Otherwise fallthrough to buffered io for
1317                          * the rest of the read.
1318                          */
1319                         if (retval < 0 || !count || *ppos >= size) {
1320                                 file_accessed(filp);
1321                                 goto out;
1322                         }
1323                 }
1324         }
1325
1326         count = retval;
1327         for (seg = 0; seg < nr_segs; seg++) {
1328                 read_descriptor_t desc;
1329                 loff_t offset = 0;
1330
1331                 /*
1332                  * If we did a short DIO read we need to skip the section of the
1333                  * iov that we've already read data into.
1334                  */
1335                 if (count) {
1336                         if (count > iov[seg].iov_len) {
1337                                 count -= iov[seg].iov_len;
1338                                 continue;
1339                         }
1340                         offset = count;
1341                         count = 0;
1342                 }
1343
1344                 desc.written = 0;
1345                 desc.arg.buf = iov[seg].iov_base + offset;
1346                 desc.count = iov[seg].iov_len - offset;
1347                 if (desc.count == 0)
1348                         continue;
1349                 desc.error = 0;
1350                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1351                 retval += desc.written;
1352                 if (desc.error) {
1353                         retval = retval ?: desc.error;
1354                         break;
1355                 }
1356                 if (desc.count > 0)
1357                         break;
1358         }
1359 out:
1360         return retval;
1361 }
1362 EXPORT_SYMBOL(generic_file_aio_read);
1363
1364 static ssize_t
1365 do_readahead(struct address_space *mapping, struct file *filp,
1366              pgoff_t index, unsigned long nr)
1367 {
1368         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1369                 return -EINVAL;
1370
1371         force_page_cache_readahead(mapping, filp, index, nr);
1372         return 0;
1373 }
1374
1375 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1376 {
1377         ssize_t ret;
1378         struct file *file;
1379
1380         ret = -EBADF;
1381         file = fget(fd);
1382         if (file) {
1383                 if (file->f_mode & FMODE_READ) {
1384                         struct address_space *mapping = file->f_mapping;
1385                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1386                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1387                         unsigned long len = end - start + 1;
1388                         ret = do_readahead(mapping, file, start, len);
1389                 }
1390                 fput(file);
1391         }
1392         return ret;
1393 }
1394 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1395 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1396 {
1397         return SYSC_readahead((int) fd, offset, (size_t) count);
1398 }
1399 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1400 #endif
1401
1402 #ifdef CONFIG_MMU
1403 /**
1404  * page_cache_read - adds requested page to the page cache if not already there
1405  * @file:       file to read
1406  * @offset:     page index
1407  *
1408  * This adds the requested page to the page cache if it isn't already there,
1409  * and schedules an I/O to read in its contents from disk.
1410  */
1411 static int page_cache_read(struct file *file, pgoff_t offset)
1412 {
1413         struct address_space *mapping = file->f_mapping;
1414         struct page *page; 
1415         int ret;
1416
1417         do {
1418                 page = page_cache_alloc_cold(mapping);
1419                 if (!page)
1420                         return -ENOMEM;
1421
1422                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1423                 if (ret == 0)
1424                         ret = mapping->a_ops->readpage(file, page);
1425                 else if (ret == -EEXIST)
1426                         ret = 0; /* losing race to add is OK */
1427
1428                 page_cache_release(page);
1429
1430         } while (ret == AOP_TRUNCATED_PAGE);
1431                 
1432         return ret;
1433 }
1434
1435 #define MMAP_LOTSAMISS  (100)
1436
1437 /*
1438  * Synchronous readahead happens when we don't even find
1439  * a page in the page cache at all.
1440  */
1441 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1442                                    struct file_ra_state *ra,
1443                                    struct file *file,
1444                                    pgoff_t offset)
1445 {
1446         unsigned long ra_pages;
1447         struct address_space *mapping = file->f_mapping;
1448
1449         /* If we don't want any read-ahead, don't bother */
1450         if (VM_RandomReadHint(vma))
1451                 return;
1452
1453         if (VM_SequentialReadHint(vma) ||
1454                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1455                 page_cache_sync_readahead(mapping, ra, file, offset,
1456                                           ra->ra_pages);
1457                 return;
1458         }
1459
1460         if (ra->mmap_miss < INT_MAX)
1461                 ra->mmap_miss++;
1462
1463         /*
1464          * Do we miss much more than hit in this file? If so,
1465          * stop bothering with read-ahead. It will only hurt.
1466          */
1467         if (ra->mmap_miss > MMAP_LOTSAMISS)
1468                 return;
1469
1470         /*
1471          * mmap read-around
1472          */
1473         ra_pages = max_sane_readahead(ra->ra_pages);
1474         if (ra_pages) {
1475                 ra->start = max_t(long, 0, offset - ra_pages/2);
1476                 ra->size = ra_pages;
1477                 ra->async_size = 0;
1478                 ra_submit(ra, mapping, file);
1479         }
1480 }
1481
1482 /*
1483  * Asynchronous readahead happens when we find the page and PG_readahead,
1484  * so we want to possibly extend the readahead further..
1485  */
1486 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1487                                     struct file_ra_state *ra,
1488                                     struct file *file,
1489                                     struct page *page,
1490                                     pgoff_t offset)
1491 {
1492         struct address_space *mapping = file->f_mapping;
1493
1494         /* If we don't want any read-ahead, don't bother */
1495         if (VM_RandomReadHint(vma))
1496                 return;
1497         if (ra->mmap_miss > 0)
1498                 ra->mmap_miss--;
1499         if (PageReadahead(page))
1500                 page_cache_async_readahead(mapping, ra, file,
1501                                            page, offset, ra->ra_pages);
1502 }
1503
1504 /**
1505  * filemap_fault - read in file data for page fault handling
1506  * @vma:        vma in which the fault was taken
1507  * @vmf:        struct vm_fault containing details of the fault
1508  *
1509  * filemap_fault() is invoked via the vma operations vector for a
1510  * mapped memory region to read in file data during a page fault.
1511  *
1512  * The goto's are kind of ugly, but this streamlines the normal case of having
1513  * it in the page cache, and handles the special cases reasonably without
1514  * having a lot of duplicated code.
1515  */
1516 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1517 {
1518         int error;
1519         struct file *file = vma->vm_file;
1520         struct address_space *mapping = file->f_mapping;
1521         struct file_ra_state *ra = &file->f_ra;
1522         struct inode *inode = mapping->host;
1523         pgoff_t offset = vmf->pgoff;
1524         struct page *page;
1525         pgoff_t size;
1526         int ret = 0;
1527
1528         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1529         if (offset >= size)
1530                 return VM_FAULT_SIGBUS;
1531
1532         /*
1533          * Do we have something in the page cache already?
1534          */
1535         page = find_get_page(mapping, offset);
1536         if (likely(page)) {
1537                 /*
1538                  * We found the page, so try async readahead before
1539                  * waiting for the lock.
1540                  */
1541                 do_async_mmap_readahead(vma, ra, file, page, offset);
1542         } else {
1543                 /* No page in the page cache at all */
1544                 do_sync_mmap_readahead(vma, ra, file, offset);
1545                 count_vm_event(PGMAJFAULT);
1546                 ret = VM_FAULT_MAJOR;
1547 retry_find:
1548                 page = find_get_page(mapping, offset);
1549                 if (!page)
1550                         goto no_cached_page;
1551         }
1552
1553         lock_page(page);
1554
1555         /* Did it get truncated? */
1556         if (unlikely(page->mapping != mapping)) {
1557                 unlock_page(page);
1558                 put_page(page);
1559                 goto retry_find;
1560         }
1561         VM_BUG_ON(page->index != offset);
1562
1563         /*
1564          * We have a locked page in the page cache, now we need to check
1565          * that it's up-to-date. If not, it is going to be due to an error.
1566          */
1567         if (unlikely(!PageUptodate(page)))
1568                 goto page_not_uptodate;
1569
1570         /*
1571          * Found the page and have a reference on it.
1572          * We must recheck i_size under page lock.
1573          */
1574         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1575         if (unlikely(offset >= size)) {
1576                 unlock_page(page);
1577                 page_cache_release(page);
1578                 return VM_FAULT_SIGBUS;
1579         }
1580
1581         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1582         vmf->page = page;
1583         return ret | VM_FAULT_LOCKED;
1584
1585 no_cached_page:
1586         /*
1587          * We're only likely to ever get here if MADV_RANDOM is in
1588          * effect.
1589          */
1590         error = page_cache_read(file, offset);
1591
1592         /*
1593          * The page we want has now been added to the page cache.
1594          * In the unlikely event that someone removed it in the
1595          * meantime, we'll just come back here and read it again.
1596          */
1597         if (error >= 0)
1598                 goto retry_find;
1599
1600         /*
1601          * An error return from page_cache_read can result if the
1602          * system is low on memory, or a problem occurs while trying
1603          * to schedule I/O.
1604          */
1605         if (error == -ENOMEM)
1606                 return VM_FAULT_OOM;
1607         return VM_FAULT_SIGBUS;
1608
1609 page_not_uptodate:
1610         /*
1611          * Umm, take care of errors if the page isn't up-to-date.
1612          * Try to re-read it _once_. We do this synchronously,
1613          * because there really aren't any performance issues here
1614          * and we need to check for errors.
1615          */
1616         ClearPageError(page);
1617         error = mapping->a_ops->readpage(file, page);
1618         if (!error) {
1619                 wait_on_page_locked(page);
1620                 if (!PageUptodate(page))
1621                         error = -EIO;
1622         }
1623         page_cache_release(page);
1624
1625         if (!error || error == AOP_TRUNCATED_PAGE)
1626                 goto retry_find;
1627
1628         /* Things didn't work out. Return zero to tell the mm layer so. */
1629         shrink_readahead_size_eio(file, ra);
1630         return VM_FAULT_SIGBUS;
1631 }
1632 EXPORT_SYMBOL(filemap_fault);
1633
1634 const struct vm_operations_struct generic_file_vm_ops = {
1635         .fault          = filemap_fault,
1636 };
1637
1638 /* This is used for a general mmap of a disk file */
1639
1640 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1641 {
1642         struct address_space *mapping = file->f_mapping;
1643
1644         if (!mapping->a_ops->readpage)
1645                 return -ENOEXEC;
1646         file_accessed(file);
1647         vma->vm_ops = &generic_file_vm_ops;
1648         vma->vm_flags |= VM_CAN_NONLINEAR;
1649         return 0;
1650 }
1651
1652 /*
1653  * This is for filesystems which do not implement ->writepage.
1654  */
1655 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1656 {
1657         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1658                 return -EINVAL;
1659         return generic_file_mmap(file, vma);
1660 }
1661 #else
1662 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1663 {
1664         return -ENOSYS;
1665 }
1666 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1667 {
1668         return -ENOSYS;
1669 }
1670 #endif /* CONFIG_MMU */
1671
1672 EXPORT_SYMBOL(generic_file_mmap);
1673 EXPORT_SYMBOL(generic_file_readonly_mmap);
1674
1675 static struct page *__read_cache_page(struct address_space *mapping,
1676                                 pgoff_t index,
1677                                 int (*filler)(void *,struct page*),
1678                                 void *data,
1679                                 gfp_t gfp)
1680 {
1681         struct page *page;
1682         int err;
1683 repeat:
1684         page = find_get_page(mapping, index);
1685         if (!page) {
1686                 page = __page_cache_alloc(gfp | __GFP_COLD);
1687                 if (!page)
1688                         return ERR_PTR(-ENOMEM);
1689                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1690                 if (unlikely(err)) {
1691                         page_cache_release(page);
1692                         if (err == -EEXIST)
1693                                 goto repeat;
1694                         /* Presumably ENOMEM for radix tree node */
1695                         return ERR_PTR(err);
1696                 }
1697                 err = filler(data, page);
1698                 if (err < 0) {
1699                         page_cache_release(page);
1700                         page = ERR_PTR(err);
1701                 }
1702         }
1703         return page;
1704 }
1705
1706 static struct page *do_read_cache_page(struct address_space *mapping,
1707                                 pgoff_t index,
1708                                 int (*filler)(void *,struct page*),
1709                                 void *data,
1710                                 gfp_t gfp)
1711
1712 {
1713         struct page *page;
1714         int err;
1715
1716 retry:
1717         page = __read_cache_page(mapping, index, filler, data, gfp);
1718         if (IS_ERR(page))
1719                 return page;
1720         if (PageUptodate(page))
1721                 goto out;
1722
1723         lock_page(page);
1724         if (!page->mapping) {
1725                 unlock_page(page);
1726                 page_cache_release(page);
1727                 goto retry;
1728         }
1729         if (PageUptodate(page)) {
1730                 unlock_page(page);
1731                 goto out;
1732         }
1733         err = filler(data, page);
1734         if (err < 0) {
1735                 page_cache_release(page);
1736                 return ERR_PTR(err);
1737         }
1738 out:
1739         mark_page_accessed(page);
1740         return page;
1741 }
1742
1743 /**
1744  * read_cache_page_async - read into page cache, fill it if needed
1745  * @mapping:    the page's address_space
1746  * @index:      the page index
1747  * @filler:     function to perform the read
1748  * @data:       destination for read data
1749  *
1750  * Same as read_cache_page, but don't wait for page to become unlocked
1751  * after submitting it to the filler.
1752  *
1753  * Read into the page cache. If a page already exists, and PageUptodate() is
1754  * not set, try to fill the page but don't wait for it to become unlocked.
1755  *
1756  * If the page does not get brought uptodate, return -EIO.
1757  */
1758 struct page *read_cache_page_async(struct address_space *mapping,
1759                                 pgoff_t index,
1760                                 int (*filler)(void *,struct page*),
1761                                 void *data)
1762 {
1763         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1764 }
1765 EXPORT_SYMBOL(read_cache_page_async);
1766
1767 static struct page *wait_on_page_read(struct page *page)
1768 {
1769         if (!IS_ERR(page)) {
1770                 wait_on_page_locked(page);
1771                 if (!PageUptodate(page)) {
1772                         page_cache_release(page);
1773                         page = ERR_PTR(-EIO);
1774                 }
1775         }
1776         return page;
1777 }
1778
1779 /**
1780  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1781  * @mapping:    the page's address_space
1782  * @index:      the page index
1783  * @gfp:        the page allocator flags to use if allocating
1784  *
1785  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1786  * any new page allocations done using the specified allocation flags. Note
1787  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1788  * expect to do this atomically or anything like that - but you can pass in
1789  * other page requirements.
1790  *
1791  * If the page does not get brought uptodate, return -EIO.
1792  */
1793 struct page *read_cache_page_gfp(struct address_space *mapping,
1794                                 pgoff_t index,
1795                                 gfp_t gfp)
1796 {
1797         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1798
1799         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1800 }
1801 EXPORT_SYMBOL(read_cache_page_gfp);
1802
1803 /**
1804  * read_cache_page - read into page cache, fill it if needed
1805  * @mapping:    the page's address_space
1806  * @index:      the page index
1807  * @filler:     function to perform the read
1808  * @data:       destination for read data
1809  *
1810  * Read into the page cache. If a page already exists, and PageUptodate() is
1811  * not set, try to fill the page then wait for it to become unlocked.
1812  *
1813  * If the page does not get brought uptodate, return -EIO.
1814  */
1815 struct page *read_cache_page(struct address_space *mapping,
1816                                 pgoff_t index,
1817                                 int (*filler)(void *,struct page*),
1818                                 void *data)
1819 {
1820         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1821 }
1822 EXPORT_SYMBOL(read_cache_page);
1823
1824 /*
1825  * The logic we want is
1826  *
1827  *      if suid or (sgid and xgrp)
1828  *              remove privs
1829  */
1830 int should_remove_suid(struct dentry *dentry)
1831 {
1832         mode_t mode = dentry->d_inode->i_mode;
1833         int kill = 0;
1834
1835         /* suid always must be killed */
1836         if (unlikely(mode & S_ISUID))
1837                 kill = ATTR_KILL_SUID;
1838
1839         /*
1840          * sgid without any exec bits is just a mandatory locking mark; leave
1841          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1842          */
1843         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1844                 kill |= ATTR_KILL_SGID;
1845
1846         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1847                 return kill;
1848
1849         return 0;
1850 }
1851 EXPORT_SYMBOL(should_remove_suid);
1852
1853 static int __remove_suid(struct dentry *dentry, int kill)
1854 {
1855         struct iattr newattrs;
1856
1857         newattrs.ia_valid = ATTR_FORCE | kill;
1858         return notify_change(dentry, &newattrs);
1859 }
1860
1861 int file_remove_suid(struct file *file)
1862 {
1863         struct dentry *dentry = file->f_path.dentry;
1864         int killsuid = should_remove_suid(dentry);
1865         int killpriv = security_inode_need_killpriv(dentry);
1866         int error = 0;
1867
1868         if (killpriv < 0)
1869                 return killpriv;
1870         if (killpriv)
1871                 error = security_inode_killpriv(dentry);
1872         if (!error && killsuid)
1873                 error = __remove_suid(dentry, killsuid);
1874
1875         return error;
1876 }
1877 EXPORT_SYMBOL(file_remove_suid);
1878
1879 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1880                         const struct iovec *iov, size_t base, size_t bytes)
1881 {
1882         size_t copied = 0, left = 0;
1883
1884         while (bytes) {
1885                 char __user *buf = iov->iov_base + base;
1886                 int copy = min(bytes, iov->iov_len - base);
1887
1888                 base = 0;
1889                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1890                 copied += copy;
1891                 bytes -= copy;
1892                 vaddr += copy;
1893                 iov++;
1894
1895                 if (unlikely(left))
1896                         break;
1897         }
1898         return copied - left;
1899 }
1900
1901 /*
1902  * Copy as much as we can into the page and return the number of bytes which
1903  * were successfully copied.  If a fault is encountered then return the number of
1904  * bytes which were copied.
1905  */
1906 size_t iov_iter_copy_from_user_atomic(struct page *page,
1907                 struct iov_iter *i, unsigned long offset, size_t bytes)
1908 {
1909         char *kaddr;
1910         size_t copied;
1911
1912         BUG_ON(!in_atomic());
1913         kaddr = kmap_atomic(page, KM_USER0);
1914         if (likely(i->nr_segs == 1)) {
1915                 int left;
1916                 char __user *buf = i->iov->iov_base + i->iov_offset;
1917                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1918                 copied = bytes - left;
1919         } else {
1920                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1921                                                 i->iov, i->iov_offset, bytes);
1922         }
1923         kunmap_atomic(kaddr, KM_USER0);
1924
1925         return copied;
1926 }
1927 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1928
1929 /*
1930  * This has the same sideeffects and return value as
1931  * iov_iter_copy_from_user_atomic().
1932  * The difference is that it attempts to resolve faults.
1933  * Page must not be locked.
1934  */
1935 size_t iov_iter_copy_from_user(struct page *page,
1936                 struct iov_iter *i, unsigned long offset, size_t bytes)
1937 {
1938         char *kaddr;
1939         size_t copied;
1940
1941         kaddr = kmap(page);
1942         if (likely(i->nr_segs == 1)) {
1943                 int left;
1944                 char __user *buf = i->iov->iov_base + i->iov_offset;
1945                 left = __copy_from_user(kaddr + offset, buf, bytes);
1946                 copied = bytes - left;
1947         } else {
1948                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1949                                                 i->iov, i->iov_offset, bytes);
1950         }
1951         kunmap(page);
1952         return copied;
1953 }
1954 EXPORT_SYMBOL(iov_iter_copy_from_user);
1955
1956 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1957 {
1958         BUG_ON(i->count < bytes);
1959
1960         if (likely(i->nr_segs == 1)) {
1961                 i->iov_offset += bytes;
1962                 i->count -= bytes;
1963         } else {
1964                 const struct iovec *iov = i->iov;
1965                 size_t base = i->iov_offset;
1966
1967                 /*
1968                  * The !iov->iov_len check ensures we skip over unlikely
1969                  * zero-length segments (without overruning the iovec).
1970                  */
1971                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1972                         int copy;
1973
1974                         copy = min(bytes, iov->iov_len - base);
1975                         BUG_ON(!i->count || i->count < copy);
1976                         i->count -= copy;
1977                         bytes -= copy;
1978                         base += copy;
1979                         if (iov->iov_len == base) {
1980                                 iov++;
1981                                 base = 0;
1982                         }
1983                 }
1984                 i->iov = iov;
1985                 i->iov_offset = base;
1986         }
1987 }
1988 EXPORT_SYMBOL(iov_iter_advance);
1989
1990 /*
1991  * Fault in the first iovec of the given iov_iter, to a maximum length
1992  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1993  * accessed (ie. because it is an invalid address).
1994  *
1995  * writev-intensive code may want this to prefault several iovecs -- that
1996  * would be possible (callers must not rely on the fact that _only_ the
1997  * first iovec will be faulted with the current implementation).
1998  */
1999 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2000 {
2001         char __user *buf = i->iov->iov_base + i->iov_offset;
2002         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2003         return fault_in_pages_readable(buf, bytes);
2004 }
2005 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2006
2007 /*
2008  * Return the count of just the current iov_iter segment.
2009  */
2010 size_t iov_iter_single_seg_count(struct iov_iter *i)
2011 {
2012         const struct iovec *iov = i->iov;
2013         if (i->nr_segs == 1)
2014                 return i->count;
2015         else
2016                 return min(i->count, iov->iov_len - i->iov_offset);
2017 }
2018 EXPORT_SYMBOL(iov_iter_single_seg_count);
2019
2020 /*
2021  * Performs necessary checks before doing a write
2022  *
2023  * Can adjust writing position or amount of bytes to write.
2024  * Returns appropriate error code that caller should return or
2025  * zero in case that write should be allowed.
2026  */
2027 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2028 {
2029         struct inode *inode = file->f_mapping->host;
2030         unsigned long limit = rlimit(RLIMIT_FSIZE);
2031
2032         if (unlikely(*pos < 0))
2033                 return -EINVAL;
2034
2035         if (!isblk) {
2036                 /* FIXME: this is for backwards compatibility with 2.4 */
2037                 if (file->f_flags & O_APPEND)
2038                         *pos = i_size_read(inode);
2039
2040                 if (limit != RLIM_INFINITY) {
2041                         if (*pos >= limit) {
2042                                 send_sig(SIGXFSZ, current, 0);
2043                                 return -EFBIG;
2044                         }
2045                         if (*count > limit - (typeof(limit))*pos) {
2046                                 *count = limit - (typeof(limit))*pos;
2047                         }
2048                 }
2049         }
2050
2051         /*
2052          * LFS rule
2053          */
2054         if (unlikely(*pos + *count > MAX_NON_LFS &&
2055                                 !(file->f_flags & O_LARGEFILE))) {
2056                 if (*pos >= MAX_NON_LFS) {
2057                         return -EFBIG;
2058                 }
2059                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2060                         *count = MAX_NON_LFS - (unsigned long)*pos;
2061                 }
2062         }
2063
2064         /*
2065          * Are we about to exceed the fs block limit ?
2066          *
2067          * If we have written data it becomes a short write.  If we have
2068          * exceeded without writing data we send a signal and return EFBIG.
2069          * Linus frestrict idea will clean these up nicely..
2070          */
2071         if (likely(!isblk)) {
2072                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2073                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2074                                 return -EFBIG;
2075                         }
2076                         /* zero-length writes at ->s_maxbytes are OK */
2077                 }
2078
2079                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2080                         *count = inode->i_sb->s_maxbytes - *pos;
2081         } else {
2082 #ifdef CONFIG_BLOCK
2083                 loff_t isize;
2084                 if (bdev_read_only(I_BDEV(inode)))
2085                         return -EPERM;
2086                 isize = i_size_read(inode);
2087                 if (*pos >= isize) {
2088                         if (*count || *pos > isize)
2089                                 return -ENOSPC;
2090                 }
2091
2092                 if (*pos + *count > isize)
2093                         *count = isize - *pos;
2094 #else
2095                 return -EPERM;
2096 #endif
2097         }
2098         return 0;
2099 }
2100 EXPORT_SYMBOL(generic_write_checks);
2101
2102 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2103                                 loff_t pos, unsigned len, unsigned flags,
2104                                 struct page **pagep, void **fsdata)
2105 {
2106         const struct address_space_operations *aops = mapping->a_ops;
2107
2108         return aops->write_begin(file, mapping, pos, len, flags,
2109                                                         pagep, fsdata);
2110 }
2111 EXPORT_SYMBOL(pagecache_write_begin);
2112
2113 int pagecache_write_end(struct file *file, struct address_space *mapping,
2114                                 loff_t pos, unsigned len, unsigned copied,
2115                                 struct page *page, void *fsdata)
2116 {
2117         const struct address_space_operations *aops = mapping->a_ops;
2118
2119         mark_page_accessed(page);
2120         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2121 }
2122 EXPORT_SYMBOL(pagecache_write_end);
2123
2124 ssize_t
2125 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2126                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2127                 size_t count, size_t ocount)
2128 {
2129         struct file     *file = iocb->ki_filp;
2130         struct address_space *mapping = file->f_mapping;
2131         struct inode    *inode = mapping->host;
2132         ssize_t         written;
2133         size_t          write_len;
2134         pgoff_t         end;
2135
2136         if (count != ocount)
2137                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2138
2139         write_len = iov_length(iov, *nr_segs);
2140         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2141
2142         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2143         if (written)
2144                 goto out;
2145
2146         /*
2147          * After a write we want buffered reads to be sure to go to disk to get
2148          * the new data.  We invalidate clean cached page from the region we're
2149          * about to write.  We do this *before* the write so that we can return
2150          * without clobbering -EIOCBQUEUED from ->direct_IO().
2151          */
2152         if (mapping->nrpages) {
2153                 written = invalidate_inode_pages2_range(mapping,
2154                                         pos >> PAGE_CACHE_SHIFT, end);
2155                 /*
2156                  * If a page can not be invalidated, return 0 to fall back
2157                  * to buffered write.
2158                  */
2159                 if (written) {
2160                         if (written == -EBUSY)
2161                                 return 0;
2162                         goto out;
2163                 }
2164         }
2165
2166         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2167
2168         /*
2169          * Finally, try again to invalidate clean pages which might have been
2170          * cached by non-direct readahead, or faulted in by get_user_pages()
2171          * if the source of the write was an mmap'ed region of the file
2172          * we're writing.  Either one is a pretty crazy thing to do,
2173          * so we don't support it 100%.  If this invalidation
2174          * fails, tough, the write still worked...
2175          */
2176         if (mapping->nrpages) {
2177                 invalidate_inode_pages2_range(mapping,
2178                                               pos >> PAGE_CACHE_SHIFT, end);
2179         }
2180
2181         if (written > 0) {
2182                 loff_t end = pos + written;
2183                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2184                         i_size_write(inode,  end);
2185                         mark_inode_dirty(inode);
2186                 }
2187                 *ppos = end;
2188         }
2189 out:
2190         return written;
2191 }
2192 EXPORT_SYMBOL(generic_file_direct_write);
2193
2194 /*
2195  * Find or create a page at the given pagecache position. Return the locked
2196  * page. This function is specifically for buffered writes.
2197  */
2198 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2199                                         pgoff_t index, unsigned flags)
2200 {
2201         int status;
2202         struct page *page;
2203         gfp_t gfp_notmask = 0;
2204         if (flags & AOP_FLAG_NOFS)
2205                 gfp_notmask = __GFP_FS;
2206 repeat:
2207         page = find_lock_page(mapping, index);
2208         if (likely(page))
2209                 return page;
2210
2211         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2212         if (!page)
2213                 return NULL;
2214         status = add_to_page_cache_lru(page, mapping, index,
2215                                                 GFP_KERNEL & ~gfp_notmask);
2216         if (unlikely(status)) {
2217                 page_cache_release(page);
2218                 if (status == -EEXIST)
2219                         goto repeat;
2220                 return NULL;
2221         }
2222         return page;
2223 }
2224 EXPORT_SYMBOL(grab_cache_page_write_begin);
2225
2226 static ssize_t generic_perform_write(struct file *file,
2227                                 struct iov_iter *i, loff_t pos)
2228 {
2229         struct address_space *mapping = file->f_mapping;
2230         const struct address_space_operations *a_ops = mapping->a_ops;
2231         long status = 0;
2232         ssize_t written = 0;
2233         unsigned int flags = 0;
2234
2235         /*
2236          * Copies from kernel address space cannot fail (NFSD is a big user).
2237          */
2238         if (segment_eq(get_fs(), KERNEL_DS))
2239                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2240
2241         do {
2242                 struct page *page;
2243                 unsigned long offset;   /* Offset into pagecache page */
2244                 unsigned long bytes;    /* Bytes to write to page */
2245                 size_t copied;          /* Bytes copied from user */
2246                 void *fsdata;
2247
2248                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2249                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2250                                                 iov_iter_count(i));
2251
2252 again:
2253
2254                 /*
2255                  * Bring in the user page that we will copy from _first_.
2256                  * Otherwise there's a nasty deadlock on copying from the
2257                  * same page as we're writing to, without it being marked
2258                  * up-to-date.
2259                  *
2260                  * Not only is this an optimisation, but it is also required
2261                  * to check that the address is actually valid, when atomic
2262                  * usercopies are used, below.
2263                  */
2264                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2265                         status = -EFAULT;
2266                         break;
2267                 }
2268
2269                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2270                                                 &page, &fsdata);
2271                 if (unlikely(status))
2272                         break;
2273
2274                 if (mapping_writably_mapped(mapping))
2275                         flush_dcache_page(page);
2276
2277                 pagefault_disable();
2278                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2279                 pagefault_enable();
2280                 flush_dcache_page(page);
2281
2282                 mark_page_accessed(page);
2283                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2284                                                 page, fsdata);
2285                 if (unlikely(status < 0))
2286                         break;
2287                 copied = status;
2288
2289                 cond_resched();
2290
2291                 iov_iter_advance(i, copied);
2292                 if (unlikely(copied == 0)) {
2293                         /*
2294                          * If we were unable to copy any data at all, we must
2295                          * fall back to a single segment length write.
2296                          *
2297                          * If we didn't fallback here, we could livelock
2298                          * because not all segments in the iov can be copied at
2299                          * once without a pagefault.
2300                          */
2301                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2302                                                 iov_iter_single_seg_count(i));
2303                         goto again;
2304                 }
2305                 pos += copied;
2306                 written += copied;
2307
2308                 balance_dirty_pages_ratelimited(mapping);
2309
2310         } while (iov_iter_count(i));
2311
2312         return written ? written : status;
2313 }
2314
2315 ssize_t
2316 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2317                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2318                 size_t count, ssize_t written)
2319 {
2320         struct file *file = iocb->ki_filp;
2321         ssize_t status;
2322         struct iov_iter i;
2323
2324         iov_iter_init(&i, iov, nr_segs, count, written);
2325         status = generic_perform_write(file, &i, pos);
2326
2327         if (likely(status >= 0)) {
2328                 written += status;
2329                 *ppos = pos + status;
2330         }
2331         
2332         return written ? written : status;
2333 }
2334 EXPORT_SYMBOL(generic_file_buffered_write);
2335
2336 /**
2337  * __generic_file_aio_write - write data to a file
2338  * @iocb:       IO state structure (file, offset, etc.)
2339  * @iov:        vector with data to write
2340  * @nr_segs:    number of segments in the vector
2341  * @ppos:       position where to write
2342  *
2343  * This function does all the work needed for actually writing data to a
2344  * file. It does all basic checks, removes SUID from the file, updates
2345  * modification times and calls proper subroutines depending on whether we
2346  * do direct IO or a standard buffered write.
2347  *
2348  * It expects i_mutex to be grabbed unless we work on a block device or similar
2349  * object which does not need locking at all.
2350  *
2351  * This function does *not* take care of syncing data in case of O_SYNC write.
2352  * A caller has to handle it. This is mainly due to the fact that we want to
2353  * avoid syncing under i_mutex.
2354  */
2355 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2356                                  unsigned long nr_segs, loff_t *ppos)
2357 {
2358         struct file *file = iocb->ki_filp;
2359         struct address_space * mapping = file->f_mapping;
2360         size_t ocount;          /* original count */
2361         size_t count;           /* after file limit checks */
2362         struct inode    *inode = mapping->host;
2363         loff_t          pos;
2364         ssize_t         written;
2365         ssize_t         err;
2366
2367         ocount = 0;
2368         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2369         if (err)
2370                 return err;
2371
2372         count = ocount;
2373         pos = *ppos;
2374
2375         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2376
2377         /* We can write back this queue in page reclaim */
2378         current->backing_dev_info = mapping->backing_dev_info;
2379         written = 0;
2380
2381         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2382         if (err)
2383                 goto out;
2384
2385         if (count == 0)
2386                 goto out;
2387
2388         err = file_remove_suid(file);
2389         if (err)
2390                 goto out;
2391
2392         file_update_time(file);
2393
2394         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2395         if (unlikely(file->f_flags & O_DIRECT)) {
2396                 loff_t endbyte;
2397                 ssize_t written_buffered;
2398
2399                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2400                                                         ppos, count, ocount);
2401                 if (written < 0 || written == count)
2402                         goto out;
2403                 /*
2404                  * direct-io write to a hole: fall through to buffered I/O
2405                  * for completing the rest of the request.
2406                  */
2407                 pos += written;
2408                 count -= written;
2409                 written_buffered = generic_file_buffered_write(iocb, iov,
2410                                                 nr_segs, pos, ppos, count,
2411                                                 written);
2412                 /*
2413                  * If generic_file_buffered_write() retuned a synchronous error
2414                  * then we want to return the number of bytes which were
2415                  * direct-written, or the error code if that was zero.  Note
2416                  * that this differs from normal direct-io semantics, which
2417                  * will return -EFOO even if some bytes were written.
2418                  */
2419                 if (written_buffered < 0) {
2420                         err = written_buffered;
2421                         goto out;
2422                 }
2423
2424                 /*
2425                  * We need to ensure that the page cache pages are written to
2426                  * disk and invalidated to preserve the expected O_DIRECT
2427                  * semantics.
2428                  */
2429                 endbyte = pos + written_buffered - written - 1;
2430                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2431                 if (err == 0) {
2432                         written = written_buffered;
2433                         invalidate_mapping_pages(mapping,
2434                                                  pos >> PAGE_CACHE_SHIFT,
2435                                                  endbyte >> PAGE_CACHE_SHIFT);
2436                 } else {
2437                         /*
2438                          * We don't know how much we wrote, so just return
2439                          * the number of bytes which were direct-written
2440                          */
2441                 }
2442         } else {
2443                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2444                                 pos, ppos, count, written);
2445         }
2446 out:
2447         current->backing_dev_info = NULL;
2448         return written ? written : err;
2449 }
2450 EXPORT_SYMBOL(__generic_file_aio_write);
2451
2452 /**
2453  * generic_file_aio_write - write data to a file
2454  * @iocb:       IO state structure
2455  * @iov:        vector with data to write
2456  * @nr_segs:    number of segments in the vector
2457  * @pos:        position in file where to write
2458  *
2459  * This is a wrapper around __generic_file_aio_write() to be used by most
2460  * filesystems. It takes care of syncing the file in case of O_SYNC file
2461  * and acquires i_mutex as needed.
2462  */
2463 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2464                 unsigned long nr_segs, loff_t pos)
2465 {
2466         struct file *file = iocb->ki_filp;
2467         struct inode *inode = file->f_mapping->host;
2468         ssize_t ret;
2469
2470         BUG_ON(iocb->ki_pos != pos);
2471
2472         mutex_lock(&inode->i_mutex);
2473         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2474         mutex_unlock(&inode->i_mutex);
2475
2476         if (ret > 0 || ret == -EIOCBQUEUED) {
2477                 ssize_t err;
2478
2479                 err = generic_write_sync(file, pos, ret);
2480                 if (err < 0 && ret > 0)
2481                         ret = err;
2482         }
2483         return ret;
2484 }
2485 EXPORT_SYMBOL(generic_file_aio_write);
2486
2487 /**
2488  * try_to_release_page() - release old fs-specific metadata on a page
2489  *
2490  * @page: the page which the kernel is trying to free
2491  * @gfp_mask: memory allocation flags (and I/O mode)
2492  *
2493  * The address_space is to try to release any data against the page
2494  * (presumably at page->private).  If the release was successful, return `1'.
2495  * Otherwise return zero.
2496  *
2497  * This may also be called if PG_fscache is set on a page, indicating that the
2498  * page is known to the local caching routines.
2499  *
2500  * The @gfp_mask argument specifies whether I/O may be performed to release
2501  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2502  *
2503  */
2504 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2505 {
2506         struct address_space * const mapping = page->mapping;
2507
2508         BUG_ON(!PageLocked(page));
2509         if (PageWriteback(page))
2510                 return 0;
2511
2512         if (mapping && mapping->a_ops->releasepage)
2513                 return mapping->a_ops->releasepage(page, gfp_mask);
2514         return try_to_free_buffers(page);
2515 }
2516
2517 EXPORT_SYMBOL(try_to_release_page);