cpuset,mm: fix no node to alloc memory when changing cpuset's mems
[linux-3.10.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
155 static int sync_page(void *word)
156 {
157         struct address_space *mapping;
158         struct page *page;
159
160         page = container_of((unsigned long *)word, struct page, flags);
161
162         /*
163          * page_mapping() is being called without PG_locked held.
164          * Some knowledge of the state and use of the page is used to
165          * reduce the requirements down to a memory barrier.
166          * The danger here is of a stale page_mapping() return value
167          * indicating a struct address_space different from the one it's
168          * associated with when it is associated with one.
169          * After smp_mb(), it's either the correct page_mapping() for
170          * the page, or an old page_mapping() and the page's own
171          * page_mapping() has gone NULL.
172          * The ->sync_page() address_space operation must tolerate
173          * page_mapping() going NULL. By an amazing coincidence,
174          * this comes about because none of the users of the page
175          * in the ->sync_page() methods make essential use of the
176          * page_mapping(), merely passing the page down to the backing
177          * device's unplug functions when it's non-NULL, which in turn
178          * ignore it for all cases but swap, where only page_private(page) is
179          * of interest. When page_mapping() does go NULL, the entire
180          * call stack gracefully ignores the page and returns.
181          * -- wli
182          */
183         smp_mb();
184         mapping = page_mapping(page);
185         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
186                 mapping->a_ops->sync_page(page);
187         io_schedule();
188         return 0;
189 }
190
191 static int sync_page_killable(void *word)
192 {
193         sync_page(word);
194         return fatal_signal_pending(current) ? -EINTR : 0;
195 }
196
197 /**
198  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
199  * @mapping:    address space structure to write
200  * @start:      offset in bytes where the range starts
201  * @end:        offset in bytes where the range ends (inclusive)
202  * @sync_mode:  enable synchronous operation
203  *
204  * Start writeback against all of a mapping's dirty pages that lie
205  * within the byte offsets <start, end> inclusive.
206  *
207  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
208  * opposed to a regular memory cleansing writeback.  The difference between
209  * these two operations is that if a dirty page/buffer is encountered, it must
210  * be waited upon, and not just skipped over.
211  */
212 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
213                                 loff_t end, int sync_mode)
214 {
215         int ret;
216         struct writeback_control wbc = {
217                 .sync_mode = sync_mode,
218                 .nr_to_write = LONG_MAX,
219                 .range_start = start,
220                 .range_end = end,
221         };
222
223         if (!mapping_cap_writeback_dirty(mapping))
224                 return 0;
225
226         ret = do_writepages(mapping, &wbc);
227         return ret;
228 }
229
230 static inline int __filemap_fdatawrite(struct address_space *mapping,
231         int sync_mode)
232 {
233         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
234 }
235
236 int filemap_fdatawrite(struct address_space *mapping)
237 {
238         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
239 }
240 EXPORT_SYMBOL(filemap_fdatawrite);
241
242 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
243                                 loff_t end)
244 {
245         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
246 }
247 EXPORT_SYMBOL(filemap_fdatawrite_range);
248
249 /**
250  * filemap_flush - mostly a non-blocking flush
251  * @mapping:    target address_space
252  *
253  * This is a mostly non-blocking flush.  Not suitable for data-integrity
254  * purposes - I/O may not be started against all dirty pages.
255  */
256 int filemap_flush(struct address_space *mapping)
257 {
258         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
259 }
260 EXPORT_SYMBOL(filemap_flush);
261
262 /**
263  * filemap_fdatawait_range - wait for writeback to complete
264  * @mapping:            address space structure to wait for
265  * @start_byte:         offset in bytes where the range starts
266  * @end_byte:           offset in bytes where the range ends (inclusive)
267  *
268  * Walk the list of under-writeback pages of the given address space
269  * in the given range and wait for all of them.
270  */
271 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
272                             loff_t end_byte)
273 {
274         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
275         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
276         struct pagevec pvec;
277         int nr_pages;
278         int ret = 0;
279
280         if (end_byte < start_byte)
281                 return 0;
282
283         pagevec_init(&pvec, 0);
284         while ((index <= end) &&
285                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
286                         PAGECACHE_TAG_WRITEBACK,
287                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
288                 unsigned i;
289
290                 for (i = 0; i < nr_pages; i++) {
291                         struct page *page = pvec.pages[i];
292
293                         /* until radix tree lookup accepts end_index */
294                         if (page->index > end)
295                                 continue;
296
297                         wait_on_page_writeback(page);
298                         if (PageError(page))
299                                 ret = -EIO;
300                 }
301                 pagevec_release(&pvec);
302                 cond_resched();
303         }
304
305         /* Check for outstanding write errors */
306         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
307                 ret = -ENOSPC;
308         if (test_and_clear_bit(AS_EIO, &mapping->flags))
309                 ret = -EIO;
310
311         return ret;
312 }
313 EXPORT_SYMBOL(filemap_fdatawait_range);
314
315 /**
316  * filemap_fdatawait - wait for all under-writeback pages to complete
317  * @mapping: address space structure to wait for
318  *
319  * Walk the list of under-writeback pages of the given address space
320  * and wait for all of them.
321  */
322 int filemap_fdatawait(struct address_space *mapping)
323 {
324         loff_t i_size = i_size_read(mapping->host);
325
326         if (i_size == 0)
327                 return 0;
328
329         return filemap_fdatawait_range(mapping, 0, i_size - 1);
330 }
331 EXPORT_SYMBOL(filemap_fdatawait);
332
333 int filemap_write_and_wait(struct address_space *mapping)
334 {
335         int err = 0;
336
337         if (mapping->nrpages) {
338                 err = filemap_fdatawrite(mapping);
339                 /*
340                  * Even if the above returned error, the pages may be
341                  * written partially (e.g. -ENOSPC), so we wait for it.
342                  * But the -EIO is special case, it may indicate the worst
343                  * thing (e.g. bug) happened, so we avoid waiting for it.
344                  */
345                 if (err != -EIO) {
346                         int err2 = filemap_fdatawait(mapping);
347                         if (!err)
348                                 err = err2;
349                 }
350         }
351         return err;
352 }
353 EXPORT_SYMBOL(filemap_write_and_wait);
354
355 /**
356  * filemap_write_and_wait_range - write out & wait on a file range
357  * @mapping:    the address_space for the pages
358  * @lstart:     offset in bytes where the range starts
359  * @lend:       offset in bytes where the range ends (inclusive)
360  *
361  * Write out and wait upon file offsets lstart->lend, inclusive.
362  *
363  * Note that `lend' is inclusive (describes the last byte to be written) so
364  * that this function can be used to write to the very end-of-file (end = -1).
365  */
366 int filemap_write_and_wait_range(struct address_space *mapping,
367                                  loff_t lstart, loff_t lend)
368 {
369         int err = 0;
370
371         if (mapping->nrpages) {
372                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
373                                                  WB_SYNC_ALL);
374                 /* See comment of filemap_write_and_wait() */
375                 if (err != -EIO) {
376                         int err2 = filemap_fdatawait_range(mapping,
377                                                 lstart, lend);
378                         if (!err)
379                                 err = err2;
380                 }
381         }
382         return err;
383 }
384 EXPORT_SYMBOL(filemap_write_and_wait_range);
385
386 /**
387  * add_to_page_cache_locked - add a locked page to the pagecache
388  * @page:       page to add
389  * @mapping:    the page's address_space
390  * @offset:     page index
391  * @gfp_mask:   page allocation mode
392  *
393  * This function is used to add a page to the pagecache. It must be locked.
394  * This function does not add the page to the LRU.  The caller must do that.
395  */
396 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
397                 pgoff_t offset, gfp_t gfp_mask)
398 {
399         int error;
400
401         VM_BUG_ON(!PageLocked(page));
402
403         error = mem_cgroup_cache_charge(page, current->mm,
404                                         gfp_mask & GFP_RECLAIM_MASK);
405         if (error)
406                 goto out;
407
408         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
409         if (error == 0) {
410                 page_cache_get(page);
411                 page->mapping = mapping;
412                 page->index = offset;
413
414                 spin_lock_irq(&mapping->tree_lock);
415                 error = radix_tree_insert(&mapping->page_tree, offset, page);
416                 if (likely(!error)) {
417                         mapping->nrpages++;
418                         __inc_zone_page_state(page, NR_FILE_PAGES);
419                         if (PageSwapBacked(page))
420                                 __inc_zone_page_state(page, NR_SHMEM);
421                         spin_unlock_irq(&mapping->tree_lock);
422                 } else {
423                         page->mapping = NULL;
424                         spin_unlock_irq(&mapping->tree_lock);
425                         mem_cgroup_uncharge_cache_page(page);
426                         page_cache_release(page);
427                 }
428                 radix_tree_preload_end();
429         } else
430                 mem_cgroup_uncharge_cache_page(page);
431 out:
432         return error;
433 }
434 EXPORT_SYMBOL(add_to_page_cache_locked);
435
436 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
437                                 pgoff_t offset, gfp_t gfp_mask)
438 {
439         int ret;
440
441         /*
442          * Splice_read and readahead add shmem/tmpfs pages into the page cache
443          * before shmem_readpage has a chance to mark them as SwapBacked: they
444          * need to go on the anon lru below, and mem_cgroup_cache_charge
445          * (called in add_to_page_cache) needs to know where they're going too.
446          */
447         if (mapping_cap_swap_backed(mapping))
448                 SetPageSwapBacked(page);
449
450         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
451         if (ret == 0) {
452                 if (page_is_file_cache(page))
453                         lru_cache_add_file(page);
454                 else
455                         lru_cache_add_anon(page);
456         }
457         return ret;
458 }
459 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
460
461 #ifdef CONFIG_NUMA
462 struct page *__page_cache_alloc(gfp_t gfp)
463 {
464         int n;
465         struct page *page;
466
467         if (cpuset_do_page_mem_spread()) {
468                 get_mems_allowed();
469                 n = cpuset_mem_spread_node();
470                 page = alloc_pages_exact_node(n, gfp, 0);
471                 put_mems_allowed();
472                 return page;
473         }
474         return alloc_pages(gfp, 0);
475 }
476 EXPORT_SYMBOL(__page_cache_alloc);
477 #endif
478
479 static int __sleep_on_page_lock(void *word)
480 {
481         io_schedule();
482         return 0;
483 }
484
485 /*
486  * In order to wait for pages to become available there must be
487  * waitqueues associated with pages. By using a hash table of
488  * waitqueues where the bucket discipline is to maintain all
489  * waiters on the same queue and wake all when any of the pages
490  * become available, and for the woken contexts to check to be
491  * sure the appropriate page became available, this saves space
492  * at a cost of "thundering herd" phenomena during rare hash
493  * collisions.
494  */
495 static wait_queue_head_t *page_waitqueue(struct page *page)
496 {
497         const struct zone *zone = page_zone(page);
498
499         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
500 }
501
502 static inline void wake_up_page(struct page *page, int bit)
503 {
504         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
505 }
506
507 void wait_on_page_bit(struct page *page, int bit_nr)
508 {
509         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
510
511         if (test_bit(bit_nr, &page->flags))
512                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
513                                                         TASK_UNINTERRUPTIBLE);
514 }
515 EXPORT_SYMBOL(wait_on_page_bit);
516
517 /**
518  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
519  * @page: Page defining the wait queue of interest
520  * @waiter: Waiter to add to the queue
521  *
522  * Add an arbitrary @waiter to the wait queue for the nominated @page.
523  */
524 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
525 {
526         wait_queue_head_t *q = page_waitqueue(page);
527         unsigned long flags;
528
529         spin_lock_irqsave(&q->lock, flags);
530         __add_wait_queue(q, waiter);
531         spin_unlock_irqrestore(&q->lock, flags);
532 }
533 EXPORT_SYMBOL_GPL(add_page_wait_queue);
534
535 /**
536  * unlock_page - unlock a locked page
537  * @page: the page
538  *
539  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
540  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
541  * mechananism between PageLocked pages and PageWriteback pages is shared.
542  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
543  *
544  * The mb is necessary to enforce ordering between the clear_bit and the read
545  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
546  */
547 void unlock_page(struct page *page)
548 {
549         VM_BUG_ON(!PageLocked(page));
550         clear_bit_unlock(PG_locked, &page->flags);
551         smp_mb__after_clear_bit();
552         wake_up_page(page, PG_locked);
553 }
554 EXPORT_SYMBOL(unlock_page);
555
556 /**
557  * end_page_writeback - end writeback against a page
558  * @page: the page
559  */
560 void end_page_writeback(struct page *page)
561 {
562         if (TestClearPageReclaim(page))
563                 rotate_reclaimable_page(page);
564
565         if (!test_clear_page_writeback(page))
566                 BUG();
567
568         smp_mb__after_clear_bit();
569         wake_up_page(page, PG_writeback);
570 }
571 EXPORT_SYMBOL(end_page_writeback);
572
573 /**
574  * __lock_page - get a lock on the page, assuming we need to sleep to get it
575  * @page: the page to lock
576  *
577  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
578  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
579  * chances are that on the second loop, the block layer's plug list is empty,
580  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
581  */
582 void __lock_page(struct page *page)
583 {
584         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
585
586         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
587                                                         TASK_UNINTERRUPTIBLE);
588 }
589 EXPORT_SYMBOL(__lock_page);
590
591 int __lock_page_killable(struct page *page)
592 {
593         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
594
595         return __wait_on_bit_lock(page_waitqueue(page), &wait,
596                                         sync_page_killable, TASK_KILLABLE);
597 }
598 EXPORT_SYMBOL_GPL(__lock_page_killable);
599
600 /**
601  * __lock_page_nosync - get a lock on the page, without calling sync_page()
602  * @page: the page to lock
603  *
604  * Variant of lock_page that does not require the caller to hold a reference
605  * on the page's mapping.
606  */
607 void __lock_page_nosync(struct page *page)
608 {
609         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
610         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
611                                                         TASK_UNINTERRUPTIBLE);
612 }
613
614 /**
615  * find_get_page - find and get a page reference
616  * @mapping: the address_space to search
617  * @offset: the page index
618  *
619  * Is there a pagecache struct page at the given (mapping, offset) tuple?
620  * If yes, increment its refcount and return it; if no, return NULL.
621  */
622 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
623 {
624         void **pagep;
625         struct page *page;
626
627         rcu_read_lock();
628 repeat:
629         page = NULL;
630         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
631         if (pagep) {
632                 page = radix_tree_deref_slot(pagep);
633                 if (unlikely(!page || page == RADIX_TREE_RETRY))
634                         goto repeat;
635
636                 if (!page_cache_get_speculative(page))
637                         goto repeat;
638
639                 /*
640                  * Has the page moved?
641                  * This is part of the lockless pagecache protocol. See
642                  * include/linux/pagemap.h for details.
643                  */
644                 if (unlikely(page != *pagep)) {
645                         page_cache_release(page);
646                         goto repeat;
647                 }
648         }
649         rcu_read_unlock();
650
651         return page;
652 }
653 EXPORT_SYMBOL(find_get_page);
654
655 /**
656  * find_lock_page - locate, pin and lock a pagecache page
657  * @mapping: the address_space to search
658  * @offset: the page index
659  *
660  * Locates the desired pagecache page, locks it, increments its reference
661  * count and returns its address.
662  *
663  * Returns zero if the page was not present. find_lock_page() may sleep.
664  */
665 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
666 {
667         struct page *page;
668
669 repeat:
670         page = find_get_page(mapping, offset);
671         if (page) {
672                 lock_page(page);
673                 /* Has the page been truncated? */
674                 if (unlikely(page->mapping != mapping)) {
675                         unlock_page(page);
676                         page_cache_release(page);
677                         goto repeat;
678                 }
679                 VM_BUG_ON(page->index != offset);
680         }
681         return page;
682 }
683 EXPORT_SYMBOL(find_lock_page);
684
685 /**
686  * find_or_create_page - locate or add a pagecache page
687  * @mapping: the page's address_space
688  * @index: the page's index into the mapping
689  * @gfp_mask: page allocation mode
690  *
691  * Locates a page in the pagecache.  If the page is not present, a new page
692  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
693  * LRU list.  The returned page is locked and has its reference count
694  * incremented.
695  *
696  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
697  * allocation!
698  *
699  * find_or_create_page() returns the desired page's address, or zero on
700  * memory exhaustion.
701  */
702 struct page *find_or_create_page(struct address_space *mapping,
703                 pgoff_t index, gfp_t gfp_mask)
704 {
705         struct page *page;
706         int err;
707 repeat:
708         page = find_lock_page(mapping, index);
709         if (!page) {
710                 page = __page_cache_alloc(gfp_mask);
711                 if (!page)
712                         return NULL;
713                 /*
714                  * We want a regular kernel memory (not highmem or DMA etc)
715                  * allocation for the radix tree nodes, but we need to honour
716                  * the context-specific requirements the caller has asked for.
717                  * GFP_RECLAIM_MASK collects those requirements.
718                  */
719                 err = add_to_page_cache_lru(page, mapping, index,
720                         (gfp_mask & GFP_RECLAIM_MASK));
721                 if (unlikely(err)) {
722                         page_cache_release(page);
723                         page = NULL;
724                         if (err == -EEXIST)
725                                 goto repeat;
726                 }
727         }
728         return page;
729 }
730 EXPORT_SYMBOL(find_or_create_page);
731
732 /**
733  * find_get_pages - gang pagecache lookup
734  * @mapping:    The address_space to search
735  * @start:      The starting page index
736  * @nr_pages:   The maximum number of pages
737  * @pages:      Where the resulting pages are placed
738  *
739  * find_get_pages() will search for and return a group of up to
740  * @nr_pages pages in the mapping.  The pages are placed at @pages.
741  * find_get_pages() takes a reference against the returned pages.
742  *
743  * The search returns a group of mapping-contiguous pages with ascending
744  * indexes.  There may be holes in the indices due to not-present pages.
745  *
746  * find_get_pages() returns the number of pages which were found.
747  */
748 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
749                             unsigned int nr_pages, struct page **pages)
750 {
751         unsigned int i;
752         unsigned int ret;
753         unsigned int nr_found;
754
755         rcu_read_lock();
756 restart:
757         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
758                                 (void ***)pages, start, nr_pages);
759         ret = 0;
760         for (i = 0; i < nr_found; i++) {
761                 struct page *page;
762 repeat:
763                 page = radix_tree_deref_slot((void **)pages[i]);
764                 if (unlikely(!page))
765                         continue;
766                 /*
767                  * this can only trigger if nr_found == 1, making livelock
768                  * a non issue.
769                  */
770                 if (unlikely(page == RADIX_TREE_RETRY))
771                         goto restart;
772
773                 if (!page_cache_get_speculative(page))
774                         goto repeat;
775
776                 /* Has the page moved? */
777                 if (unlikely(page != *((void **)pages[i]))) {
778                         page_cache_release(page);
779                         goto repeat;
780                 }
781
782                 pages[ret] = page;
783                 ret++;
784         }
785         rcu_read_unlock();
786         return ret;
787 }
788
789 /**
790  * find_get_pages_contig - gang contiguous pagecache lookup
791  * @mapping:    The address_space to search
792  * @index:      The starting page index
793  * @nr_pages:   The maximum number of pages
794  * @pages:      Where the resulting pages are placed
795  *
796  * find_get_pages_contig() works exactly like find_get_pages(), except
797  * that the returned number of pages are guaranteed to be contiguous.
798  *
799  * find_get_pages_contig() returns the number of pages which were found.
800  */
801 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
802                                unsigned int nr_pages, struct page **pages)
803 {
804         unsigned int i;
805         unsigned int ret;
806         unsigned int nr_found;
807
808         rcu_read_lock();
809 restart:
810         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
811                                 (void ***)pages, index, nr_pages);
812         ret = 0;
813         for (i = 0; i < nr_found; i++) {
814                 struct page *page;
815 repeat:
816                 page = radix_tree_deref_slot((void **)pages[i]);
817                 if (unlikely(!page))
818                         continue;
819                 /*
820                  * this can only trigger if nr_found == 1, making livelock
821                  * a non issue.
822                  */
823                 if (unlikely(page == RADIX_TREE_RETRY))
824                         goto restart;
825
826                 if (page->mapping == NULL || page->index != index)
827                         break;
828
829                 if (!page_cache_get_speculative(page))
830                         goto repeat;
831
832                 /* Has the page moved? */
833                 if (unlikely(page != *((void **)pages[i]))) {
834                         page_cache_release(page);
835                         goto repeat;
836                 }
837
838                 pages[ret] = page;
839                 ret++;
840                 index++;
841         }
842         rcu_read_unlock();
843         return ret;
844 }
845 EXPORT_SYMBOL(find_get_pages_contig);
846
847 /**
848  * find_get_pages_tag - find and return pages that match @tag
849  * @mapping:    the address_space to search
850  * @index:      the starting page index
851  * @tag:        the tag index
852  * @nr_pages:   the maximum number of pages
853  * @pages:      where the resulting pages are placed
854  *
855  * Like find_get_pages, except we only return pages which are tagged with
856  * @tag.   We update @index to index the next page for the traversal.
857  */
858 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
859                         int tag, unsigned int nr_pages, struct page **pages)
860 {
861         unsigned int i;
862         unsigned int ret;
863         unsigned int nr_found;
864
865         rcu_read_lock();
866 restart:
867         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
868                                 (void ***)pages, *index, nr_pages, tag);
869         ret = 0;
870         for (i = 0; i < nr_found; i++) {
871                 struct page *page;
872 repeat:
873                 page = radix_tree_deref_slot((void **)pages[i]);
874                 if (unlikely(!page))
875                         continue;
876                 /*
877                  * this can only trigger if nr_found == 1, making livelock
878                  * a non issue.
879                  */
880                 if (unlikely(page == RADIX_TREE_RETRY))
881                         goto restart;
882
883                 if (!page_cache_get_speculative(page))
884                         goto repeat;
885
886                 /* Has the page moved? */
887                 if (unlikely(page != *((void **)pages[i]))) {
888                         page_cache_release(page);
889                         goto repeat;
890                 }
891
892                 pages[ret] = page;
893                 ret++;
894         }
895         rcu_read_unlock();
896
897         if (ret)
898                 *index = pages[ret - 1]->index + 1;
899
900         return ret;
901 }
902 EXPORT_SYMBOL(find_get_pages_tag);
903
904 /**
905  * grab_cache_page_nowait - returns locked page at given index in given cache
906  * @mapping: target address_space
907  * @index: the page index
908  *
909  * Same as grab_cache_page(), but do not wait if the page is unavailable.
910  * This is intended for speculative data generators, where the data can
911  * be regenerated if the page couldn't be grabbed.  This routine should
912  * be safe to call while holding the lock for another page.
913  *
914  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
915  * and deadlock against the caller's locked page.
916  */
917 struct page *
918 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
919 {
920         struct page *page = find_get_page(mapping, index);
921
922         if (page) {
923                 if (trylock_page(page))
924                         return page;
925                 page_cache_release(page);
926                 return NULL;
927         }
928         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
929         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
930                 page_cache_release(page);
931                 page = NULL;
932         }
933         return page;
934 }
935 EXPORT_SYMBOL(grab_cache_page_nowait);
936
937 /*
938  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
939  * a _large_ part of the i/o request. Imagine the worst scenario:
940  *
941  *      ---R__________________________________________B__________
942  *         ^ reading here                             ^ bad block(assume 4k)
943  *
944  * read(R) => miss => readahead(R...B) => media error => frustrating retries
945  * => failing the whole request => read(R) => read(R+1) =>
946  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
947  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
948  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
949  *
950  * It is going insane. Fix it by quickly scaling down the readahead size.
951  */
952 static void shrink_readahead_size_eio(struct file *filp,
953                                         struct file_ra_state *ra)
954 {
955         ra->ra_pages /= 4;
956 }
957
958 /**
959  * do_generic_file_read - generic file read routine
960  * @filp:       the file to read
961  * @ppos:       current file position
962  * @desc:       read_descriptor
963  * @actor:      read method
964  *
965  * This is a generic file read routine, and uses the
966  * mapping->a_ops->readpage() function for the actual low-level stuff.
967  *
968  * This is really ugly. But the goto's actually try to clarify some
969  * of the logic when it comes to error handling etc.
970  */
971 static void do_generic_file_read(struct file *filp, loff_t *ppos,
972                 read_descriptor_t *desc, read_actor_t actor)
973 {
974         struct address_space *mapping = filp->f_mapping;
975         struct inode *inode = mapping->host;
976         struct file_ra_state *ra = &filp->f_ra;
977         pgoff_t index;
978         pgoff_t last_index;
979         pgoff_t prev_index;
980         unsigned long offset;      /* offset into pagecache page */
981         unsigned int prev_offset;
982         int error;
983
984         index = *ppos >> PAGE_CACHE_SHIFT;
985         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
986         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
987         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
988         offset = *ppos & ~PAGE_CACHE_MASK;
989
990         for (;;) {
991                 struct page *page;
992                 pgoff_t end_index;
993                 loff_t isize;
994                 unsigned long nr, ret;
995
996                 cond_resched();
997 find_page:
998                 page = find_get_page(mapping, index);
999                 if (!page) {
1000                         page_cache_sync_readahead(mapping,
1001                                         ra, filp,
1002                                         index, last_index - index);
1003                         page = find_get_page(mapping, index);
1004                         if (unlikely(page == NULL))
1005                                 goto no_cached_page;
1006                 }
1007                 if (PageReadahead(page)) {
1008                         page_cache_async_readahead(mapping,
1009                                         ra, filp, page,
1010                                         index, last_index - index);
1011                 }
1012                 if (!PageUptodate(page)) {
1013                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1014                                         !mapping->a_ops->is_partially_uptodate)
1015                                 goto page_not_up_to_date;
1016                         if (!trylock_page(page))
1017                                 goto page_not_up_to_date;
1018                         if (!mapping->a_ops->is_partially_uptodate(page,
1019                                                                 desc, offset))
1020                                 goto page_not_up_to_date_locked;
1021                         unlock_page(page);
1022                 }
1023 page_ok:
1024                 /*
1025                  * i_size must be checked after we know the page is Uptodate.
1026                  *
1027                  * Checking i_size after the check allows us to calculate
1028                  * the correct value for "nr", which means the zero-filled
1029                  * part of the page is not copied back to userspace (unless
1030                  * another truncate extends the file - this is desired though).
1031                  */
1032
1033                 isize = i_size_read(inode);
1034                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1035                 if (unlikely(!isize || index > end_index)) {
1036                         page_cache_release(page);
1037                         goto out;
1038                 }
1039
1040                 /* nr is the maximum number of bytes to copy from this page */
1041                 nr = PAGE_CACHE_SIZE;
1042                 if (index == end_index) {
1043                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1044                         if (nr <= offset) {
1045                                 page_cache_release(page);
1046                                 goto out;
1047                         }
1048                 }
1049                 nr = nr - offset;
1050
1051                 /* If users can be writing to this page using arbitrary
1052                  * virtual addresses, take care about potential aliasing
1053                  * before reading the page on the kernel side.
1054                  */
1055                 if (mapping_writably_mapped(mapping))
1056                         flush_dcache_page(page);
1057
1058                 /*
1059                  * When a sequential read accesses a page several times,
1060                  * only mark it as accessed the first time.
1061                  */
1062                 if (prev_index != index || offset != prev_offset)
1063                         mark_page_accessed(page);
1064                 prev_index = index;
1065
1066                 /*
1067                  * Ok, we have the page, and it's up-to-date, so
1068                  * now we can copy it to user space...
1069                  *
1070                  * The actor routine returns how many bytes were actually used..
1071                  * NOTE! This may not be the same as how much of a user buffer
1072                  * we filled up (we may be padding etc), so we can only update
1073                  * "pos" here (the actor routine has to update the user buffer
1074                  * pointers and the remaining count).
1075                  */
1076                 ret = actor(desc, page, offset, nr);
1077                 offset += ret;
1078                 index += offset >> PAGE_CACHE_SHIFT;
1079                 offset &= ~PAGE_CACHE_MASK;
1080                 prev_offset = offset;
1081
1082                 page_cache_release(page);
1083                 if (ret == nr && desc->count)
1084                         continue;
1085                 goto out;
1086
1087 page_not_up_to_date:
1088                 /* Get exclusive access to the page ... */
1089                 error = lock_page_killable(page);
1090                 if (unlikely(error))
1091                         goto readpage_error;
1092
1093 page_not_up_to_date_locked:
1094                 /* Did it get truncated before we got the lock? */
1095                 if (!page->mapping) {
1096                         unlock_page(page);
1097                         page_cache_release(page);
1098                         continue;
1099                 }
1100
1101                 /* Did somebody else fill it already? */
1102                 if (PageUptodate(page)) {
1103                         unlock_page(page);
1104                         goto page_ok;
1105                 }
1106
1107 readpage:
1108                 /* Start the actual read. The read will unlock the page. */
1109                 error = mapping->a_ops->readpage(filp, page);
1110
1111                 if (unlikely(error)) {
1112                         if (error == AOP_TRUNCATED_PAGE) {
1113                                 page_cache_release(page);
1114                                 goto find_page;
1115                         }
1116                         goto readpage_error;
1117                 }
1118
1119                 if (!PageUptodate(page)) {
1120                         error = lock_page_killable(page);
1121                         if (unlikely(error))
1122                                 goto readpage_error;
1123                         if (!PageUptodate(page)) {
1124                                 if (page->mapping == NULL) {
1125                                         /*
1126                                          * invalidate_mapping_pages got it
1127                                          */
1128                                         unlock_page(page);
1129                                         page_cache_release(page);
1130                                         goto find_page;
1131                                 }
1132                                 unlock_page(page);
1133                                 shrink_readahead_size_eio(filp, ra);
1134                                 error = -EIO;
1135                                 goto readpage_error;
1136                         }
1137                         unlock_page(page);
1138                 }
1139
1140                 goto page_ok;
1141
1142 readpage_error:
1143                 /* UHHUH! A synchronous read error occurred. Report it */
1144                 desc->error = error;
1145                 page_cache_release(page);
1146                 goto out;
1147
1148 no_cached_page:
1149                 /*
1150                  * Ok, it wasn't cached, so we need to create a new
1151                  * page..
1152                  */
1153                 page = page_cache_alloc_cold(mapping);
1154                 if (!page) {
1155                         desc->error = -ENOMEM;
1156                         goto out;
1157                 }
1158                 error = add_to_page_cache_lru(page, mapping,
1159                                                 index, GFP_KERNEL);
1160                 if (error) {
1161                         page_cache_release(page);
1162                         if (error == -EEXIST)
1163                                 goto find_page;
1164                         desc->error = error;
1165                         goto out;
1166                 }
1167                 goto readpage;
1168         }
1169
1170 out:
1171         ra->prev_pos = prev_index;
1172         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1173         ra->prev_pos |= prev_offset;
1174
1175         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1176         file_accessed(filp);
1177 }
1178
1179 int file_read_actor(read_descriptor_t *desc, struct page *page,
1180                         unsigned long offset, unsigned long size)
1181 {
1182         char *kaddr;
1183         unsigned long left, count = desc->count;
1184
1185         if (size > count)
1186                 size = count;
1187
1188         /*
1189          * Faults on the destination of a read are common, so do it before
1190          * taking the kmap.
1191          */
1192         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1193                 kaddr = kmap_atomic(page, KM_USER0);
1194                 left = __copy_to_user_inatomic(desc->arg.buf,
1195                                                 kaddr + offset, size);
1196                 kunmap_atomic(kaddr, KM_USER0);
1197                 if (left == 0)
1198                         goto success;
1199         }
1200
1201         /* Do it the slow way */
1202         kaddr = kmap(page);
1203         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1204         kunmap(page);
1205
1206         if (left) {
1207                 size -= left;
1208                 desc->error = -EFAULT;
1209         }
1210 success:
1211         desc->count = count - size;
1212         desc->written += size;
1213         desc->arg.buf += size;
1214         return size;
1215 }
1216
1217 /*
1218  * Performs necessary checks before doing a write
1219  * @iov:        io vector request
1220  * @nr_segs:    number of segments in the iovec
1221  * @count:      number of bytes to write
1222  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1223  *
1224  * Adjust number of segments and amount of bytes to write (nr_segs should be
1225  * properly initialized first). Returns appropriate error code that caller
1226  * should return or zero in case that write should be allowed.
1227  */
1228 int generic_segment_checks(const struct iovec *iov,
1229                         unsigned long *nr_segs, size_t *count, int access_flags)
1230 {
1231         unsigned long   seg;
1232         size_t cnt = 0;
1233         for (seg = 0; seg < *nr_segs; seg++) {
1234                 const struct iovec *iv = &iov[seg];
1235
1236                 /*
1237                  * If any segment has a negative length, or the cumulative
1238                  * length ever wraps negative then return -EINVAL.
1239                  */
1240                 cnt += iv->iov_len;
1241                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1242                         return -EINVAL;
1243                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1244                         continue;
1245                 if (seg == 0)
1246                         return -EFAULT;
1247                 *nr_segs = seg;
1248                 cnt -= iv->iov_len;     /* This segment is no good */
1249                 break;
1250         }
1251         *count = cnt;
1252         return 0;
1253 }
1254 EXPORT_SYMBOL(generic_segment_checks);
1255
1256 /**
1257  * generic_file_aio_read - generic filesystem read routine
1258  * @iocb:       kernel I/O control block
1259  * @iov:        io vector request
1260  * @nr_segs:    number of segments in the iovec
1261  * @pos:        current file position
1262  *
1263  * This is the "read()" routine for all filesystems
1264  * that can use the page cache directly.
1265  */
1266 ssize_t
1267 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1268                 unsigned long nr_segs, loff_t pos)
1269 {
1270         struct file *filp = iocb->ki_filp;
1271         ssize_t retval;
1272         unsigned long seg;
1273         size_t count;
1274         loff_t *ppos = &iocb->ki_pos;
1275
1276         count = 0;
1277         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1278         if (retval)
1279                 return retval;
1280
1281         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1282         if (filp->f_flags & O_DIRECT) {
1283                 loff_t size;
1284                 struct address_space *mapping;
1285                 struct inode *inode;
1286
1287                 mapping = filp->f_mapping;
1288                 inode = mapping->host;
1289                 if (!count)
1290                         goto out; /* skip atime */
1291                 size = i_size_read(inode);
1292                 if (pos < size) {
1293                         retval = filemap_write_and_wait_range(mapping, pos,
1294                                         pos + iov_length(iov, nr_segs) - 1);
1295                         if (!retval) {
1296                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1297                                                         iov, pos, nr_segs);
1298                         }
1299                         if (retval > 0)
1300                                 *ppos = pos + retval;
1301                         if (retval) {
1302                                 file_accessed(filp);
1303                                 goto out;
1304                         }
1305                 }
1306         }
1307
1308         for (seg = 0; seg < nr_segs; seg++) {
1309                 read_descriptor_t desc;
1310
1311                 desc.written = 0;
1312                 desc.arg.buf = iov[seg].iov_base;
1313                 desc.count = iov[seg].iov_len;
1314                 if (desc.count == 0)
1315                         continue;
1316                 desc.error = 0;
1317                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1318                 retval += desc.written;
1319                 if (desc.error) {
1320                         retval = retval ?: desc.error;
1321                         break;
1322                 }
1323                 if (desc.count > 0)
1324                         break;
1325         }
1326 out:
1327         return retval;
1328 }
1329 EXPORT_SYMBOL(generic_file_aio_read);
1330
1331 static ssize_t
1332 do_readahead(struct address_space *mapping, struct file *filp,
1333              pgoff_t index, unsigned long nr)
1334 {
1335         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1336                 return -EINVAL;
1337
1338         force_page_cache_readahead(mapping, filp, index, nr);
1339         return 0;
1340 }
1341
1342 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1343 {
1344         ssize_t ret;
1345         struct file *file;
1346
1347         ret = -EBADF;
1348         file = fget(fd);
1349         if (file) {
1350                 if (file->f_mode & FMODE_READ) {
1351                         struct address_space *mapping = file->f_mapping;
1352                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1353                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1354                         unsigned long len = end - start + 1;
1355                         ret = do_readahead(mapping, file, start, len);
1356                 }
1357                 fput(file);
1358         }
1359         return ret;
1360 }
1361 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1362 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1363 {
1364         return SYSC_readahead((int) fd, offset, (size_t) count);
1365 }
1366 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1367 #endif
1368
1369 #ifdef CONFIG_MMU
1370 /**
1371  * page_cache_read - adds requested page to the page cache if not already there
1372  * @file:       file to read
1373  * @offset:     page index
1374  *
1375  * This adds the requested page to the page cache if it isn't already there,
1376  * and schedules an I/O to read in its contents from disk.
1377  */
1378 static int page_cache_read(struct file *file, pgoff_t offset)
1379 {
1380         struct address_space *mapping = file->f_mapping;
1381         struct page *page; 
1382         int ret;
1383
1384         do {
1385                 page = page_cache_alloc_cold(mapping);
1386                 if (!page)
1387                         return -ENOMEM;
1388
1389                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1390                 if (ret == 0)
1391                         ret = mapping->a_ops->readpage(file, page);
1392                 else if (ret == -EEXIST)
1393                         ret = 0; /* losing race to add is OK */
1394
1395                 page_cache_release(page);
1396
1397         } while (ret == AOP_TRUNCATED_PAGE);
1398                 
1399         return ret;
1400 }
1401
1402 #define MMAP_LOTSAMISS  (100)
1403
1404 /*
1405  * Synchronous readahead happens when we don't even find
1406  * a page in the page cache at all.
1407  */
1408 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1409                                    struct file_ra_state *ra,
1410                                    struct file *file,
1411                                    pgoff_t offset)
1412 {
1413         unsigned long ra_pages;
1414         struct address_space *mapping = file->f_mapping;
1415
1416         /* If we don't want any read-ahead, don't bother */
1417         if (VM_RandomReadHint(vma))
1418                 return;
1419
1420         if (VM_SequentialReadHint(vma) ||
1421                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1422                 page_cache_sync_readahead(mapping, ra, file, offset,
1423                                           ra->ra_pages);
1424                 return;
1425         }
1426
1427         if (ra->mmap_miss < INT_MAX)
1428                 ra->mmap_miss++;
1429
1430         /*
1431          * Do we miss much more than hit in this file? If so,
1432          * stop bothering with read-ahead. It will only hurt.
1433          */
1434         if (ra->mmap_miss > MMAP_LOTSAMISS)
1435                 return;
1436
1437         /*
1438          * mmap read-around
1439          */
1440         ra_pages = max_sane_readahead(ra->ra_pages);
1441         if (ra_pages) {
1442                 ra->start = max_t(long, 0, offset - ra_pages/2);
1443                 ra->size = ra_pages;
1444                 ra->async_size = 0;
1445                 ra_submit(ra, mapping, file);
1446         }
1447 }
1448
1449 /*
1450  * Asynchronous readahead happens when we find the page and PG_readahead,
1451  * so we want to possibly extend the readahead further..
1452  */
1453 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1454                                     struct file_ra_state *ra,
1455                                     struct file *file,
1456                                     struct page *page,
1457                                     pgoff_t offset)
1458 {
1459         struct address_space *mapping = file->f_mapping;
1460
1461         /* If we don't want any read-ahead, don't bother */
1462         if (VM_RandomReadHint(vma))
1463                 return;
1464         if (ra->mmap_miss > 0)
1465                 ra->mmap_miss--;
1466         if (PageReadahead(page))
1467                 page_cache_async_readahead(mapping, ra, file,
1468                                            page, offset, ra->ra_pages);
1469 }
1470
1471 /**
1472  * filemap_fault - read in file data for page fault handling
1473  * @vma:        vma in which the fault was taken
1474  * @vmf:        struct vm_fault containing details of the fault
1475  *
1476  * filemap_fault() is invoked via the vma operations vector for a
1477  * mapped memory region to read in file data during a page fault.
1478  *
1479  * The goto's are kind of ugly, but this streamlines the normal case of having
1480  * it in the page cache, and handles the special cases reasonably without
1481  * having a lot of duplicated code.
1482  */
1483 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1484 {
1485         int error;
1486         struct file *file = vma->vm_file;
1487         struct address_space *mapping = file->f_mapping;
1488         struct file_ra_state *ra = &file->f_ra;
1489         struct inode *inode = mapping->host;
1490         pgoff_t offset = vmf->pgoff;
1491         struct page *page;
1492         pgoff_t size;
1493         int ret = 0;
1494
1495         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1496         if (offset >= size)
1497                 return VM_FAULT_SIGBUS;
1498
1499         /*
1500          * Do we have something in the page cache already?
1501          */
1502         page = find_get_page(mapping, offset);
1503         if (likely(page)) {
1504                 /*
1505                  * We found the page, so try async readahead before
1506                  * waiting for the lock.
1507                  */
1508                 do_async_mmap_readahead(vma, ra, file, page, offset);
1509                 lock_page(page);
1510
1511                 /* Did it get truncated? */
1512                 if (unlikely(page->mapping != mapping)) {
1513                         unlock_page(page);
1514                         put_page(page);
1515                         goto no_cached_page;
1516                 }
1517         } else {
1518                 /* No page in the page cache at all */
1519                 do_sync_mmap_readahead(vma, ra, file, offset);
1520                 count_vm_event(PGMAJFAULT);
1521                 ret = VM_FAULT_MAJOR;
1522 retry_find:
1523                 page = find_lock_page(mapping, offset);
1524                 if (!page)
1525                         goto no_cached_page;
1526         }
1527
1528         /*
1529          * We have a locked page in the page cache, now we need to check
1530          * that it's up-to-date. If not, it is going to be due to an error.
1531          */
1532         if (unlikely(!PageUptodate(page)))
1533                 goto page_not_uptodate;
1534
1535         /*
1536          * Found the page and have a reference on it.
1537          * We must recheck i_size under page lock.
1538          */
1539         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1540         if (unlikely(offset >= size)) {
1541                 unlock_page(page);
1542                 page_cache_release(page);
1543                 return VM_FAULT_SIGBUS;
1544         }
1545
1546         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1547         vmf->page = page;
1548         return ret | VM_FAULT_LOCKED;
1549
1550 no_cached_page:
1551         /*
1552          * We're only likely to ever get here if MADV_RANDOM is in
1553          * effect.
1554          */
1555         error = page_cache_read(file, offset);
1556
1557         /*
1558          * The page we want has now been added to the page cache.
1559          * In the unlikely event that someone removed it in the
1560          * meantime, we'll just come back here and read it again.
1561          */
1562         if (error >= 0)
1563                 goto retry_find;
1564
1565         /*
1566          * An error return from page_cache_read can result if the
1567          * system is low on memory, or a problem occurs while trying
1568          * to schedule I/O.
1569          */
1570         if (error == -ENOMEM)
1571                 return VM_FAULT_OOM;
1572         return VM_FAULT_SIGBUS;
1573
1574 page_not_uptodate:
1575         /*
1576          * Umm, take care of errors if the page isn't up-to-date.
1577          * Try to re-read it _once_. We do this synchronously,
1578          * because there really aren't any performance issues here
1579          * and we need to check for errors.
1580          */
1581         ClearPageError(page);
1582         error = mapping->a_ops->readpage(file, page);
1583         if (!error) {
1584                 wait_on_page_locked(page);
1585                 if (!PageUptodate(page))
1586                         error = -EIO;
1587         }
1588         page_cache_release(page);
1589
1590         if (!error || error == AOP_TRUNCATED_PAGE)
1591                 goto retry_find;
1592
1593         /* Things didn't work out. Return zero to tell the mm layer so. */
1594         shrink_readahead_size_eio(file, ra);
1595         return VM_FAULT_SIGBUS;
1596 }
1597 EXPORT_SYMBOL(filemap_fault);
1598
1599 const struct vm_operations_struct generic_file_vm_ops = {
1600         .fault          = filemap_fault,
1601 };
1602
1603 /* This is used for a general mmap of a disk file */
1604
1605 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1606 {
1607         struct address_space *mapping = file->f_mapping;
1608
1609         if (!mapping->a_ops->readpage)
1610                 return -ENOEXEC;
1611         file_accessed(file);
1612         vma->vm_ops = &generic_file_vm_ops;
1613         vma->vm_flags |= VM_CAN_NONLINEAR;
1614         return 0;
1615 }
1616
1617 /*
1618  * This is for filesystems which do not implement ->writepage.
1619  */
1620 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1621 {
1622         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1623                 return -EINVAL;
1624         return generic_file_mmap(file, vma);
1625 }
1626 #else
1627 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1628 {
1629         return -ENOSYS;
1630 }
1631 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1632 {
1633         return -ENOSYS;
1634 }
1635 #endif /* CONFIG_MMU */
1636
1637 EXPORT_SYMBOL(generic_file_mmap);
1638 EXPORT_SYMBOL(generic_file_readonly_mmap);
1639
1640 static struct page *__read_cache_page(struct address_space *mapping,
1641                                 pgoff_t index,
1642                                 int (*filler)(void *,struct page*),
1643                                 void *data,
1644                                 gfp_t gfp)
1645 {
1646         struct page *page;
1647         int err;
1648 repeat:
1649         page = find_get_page(mapping, index);
1650         if (!page) {
1651                 page = __page_cache_alloc(gfp | __GFP_COLD);
1652                 if (!page)
1653                         return ERR_PTR(-ENOMEM);
1654                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1655                 if (unlikely(err)) {
1656                         page_cache_release(page);
1657                         if (err == -EEXIST)
1658                                 goto repeat;
1659                         /* Presumably ENOMEM for radix tree node */
1660                         return ERR_PTR(err);
1661                 }
1662                 err = filler(data, page);
1663                 if (err < 0) {
1664                         page_cache_release(page);
1665                         page = ERR_PTR(err);
1666                 }
1667         }
1668         return page;
1669 }
1670
1671 static struct page *do_read_cache_page(struct address_space *mapping,
1672                                 pgoff_t index,
1673                                 int (*filler)(void *,struct page*),
1674                                 void *data,
1675                                 gfp_t gfp)
1676
1677 {
1678         struct page *page;
1679         int err;
1680
1681 retry:
1682         page = __read_cache_page(mapping, index, filler, data, gfp);
1683         if (IS_ERR(page))
1684                 return page;
1685         if (PageUptodate(page))
1686                 goto out;
1687
1688         lock_page(page);
1689         if (!page->mapping) {
1690                 unlock_page(page);
1691                 page_cache_release(page);
1692                 goto retry;
1693         }
1694         if (PageUptodate(page)) {
1695                 unlock_page(page);
1696                 goto out;
1697         }
1698         err = filler(data, page);
1699         if (err < 0) {
1700                 page_cache_release(page);
1701                 return ERR_PTR(err);
1702         }
1703 out:
1704         mark_page_accessed(page);
1705         return page;
1706 }
1707
1708 /**
1709  * read_cache_page_async - read into page cache, fill it if needed
1710  * @mapping:    the page's address_space
1711  * @index:      the page index
1712  * @filler:     function to perform the read
1713  * @data:       destination for read data
1714  *
1715  * Same as read_cache_page, but don't wait for page to become unlocked
1716  * after submitting it to the filler.
1717  *
1718  * Read into the page cache. If a page already exists, and PageUptodate() is
1719  * not set, try to fill the page but don't wait for it to become unlocked.
1720  *
1721  * If the page does not get brought uptodate, return -EIO.
1722  */
1723 struct page *read_cache_page_async(struct address_space *mapping,
1724                                 pgoff_t index,
1725                                 int (*filler)(void *,struct page*),
1726                                 void *data)
1727 {
1728         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1729 }
1730 EXPORT_SYMBOL(read_cache_page_async);
1731
1732 static struct page *wait_on_page_read(struct page *page)
1733 {
1734         if (!IS_ERR(page)) {
1735                 wait_on_page_locked(page);
1736                 if (!PageUptodate(page)) {
1737                         page_cache_release(page);
1738                         page = ERR_PTR(-EIO);
1739                 }
1740         }
1741         return page;
1742 }
1743
1744 /**
1745  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1746  * @mapping:    the page's address_space
1747  * @index:      the page index
1748  * @gfp:        the page allocator flags to use if allocating
1749  *
1750  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1751  * any new page allocations done using the specified allocation flags. Note
1752  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1753  * expect to do this atomically or anything like that - but you can pass in
1754  * other page requirements.
1755  *
1756  * If the page does not get brought uptodate, return -EIO.
1757  */
1758 struct page *read_cache_page_gfp(struct address_space *mapping,
1759                                 pgoff_t index,
1760                                 gfp_t gfp)
1761 {
1762         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1763
1764         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1765 }
1766 EXPORT_SYMBOL(read_cache_page_gfp);
1767
1768 /**
1769  * read_cache_page - read into page cache, fill it if needed
1770  * @mapping:    the page's address_space
1771  * @index:      the page index
1772  * @filler:     function to perform the read
1773  * @data:       destination for read data
1774  *
1775  * Read into the page cache. If a page already exists, and PageUptodate() is
1776  * not set, try to fill the page then wait for it to become unlocked.
1777  *
1778  * If the page does not get brought uptodate, return -EIO.
1779  */
1780 struct page *read_cache_page(struct address_space *mapping,
1781                                 pgoff_t index,
1782                                 int (*filler)(void *,struct page*),
1783                                 void *data)
1784 {
1785         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1786 }
1787 EXPORT_SYMBOL(read_cache_page);
1788
1789 /*
1790  * The logic we want is
1791  *
1792  *      if suid or (sgid and xgrp)
1793  *              remove privs
1794  */
1795 int should_remove_suid(struct dentry *dentry)
1796 {
1797         mode_t mode = dentry->d_inode->i_mode;
1798         int kill = 0;
1799
1800         /* suid always must be killed */
1801         if (unlikely(mode & S_ISUID))
1802                 kill = ATTR_KILL_SUID;
1803
1804         /*
1805          * sgid without any exec bits is just a mandatory locking mark; leave
1806          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1807          */
1808         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1809                 kill |= ATTR_KILL_SGID;
1810
1811         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1812                 return kill;
1813
1814         return 0;
1815 }
1816 EXPORT_SYMBOL(should_remove_suid);
1817
1818 static int __remove_suid(struct dentry *dentry, int kill)
1819 {
1820         struct iattr newattrs;
1821
1822         newattrs.ia_valid = ATTR_FORCE | kill;
1823         return notify_change(dentry, &newattrs);
1824 }
1825
1826 int file_remove_suid(struct file *file)
1827 {
1828         struct dentry *dentry = file->f_path.dentry;
1829         int killsuid = should_remove_suid(dentry);
1830         int killpriv = security_inode_need_killpriv(dentry);
1831         int error = 0;
1832
1833         if (killpriv < 0)
1834                 return killpriv;
1835         if (killpriv)
1836                 error = security_inode_killpriv(dentry);
1837         if (!error && killsuid)
1838                 error = __remove_suid(dentry, killsuid);
1839
1840         return error;
1841 }
1842 EXPORT_SYMBOL(file_remove_suid);
1843
1844 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1845                         const struct iovec *iov, size_t base, size_t bytes)
1846 {
1847         size_t copied = 0, left = 0;
1848
1849         while (bytes) {
1850                 char __user *buf = iov->iov_base + base;
1851                 int copy = min(bytes, iov->iov_len - base);
1852
1853                 base = 0;
1854                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1855                 copied += copy;
1856                 bytes -= copy;
1857                 vaddr += copy;
1858                 iov++;
1859
1860                 if (unlikely(left))
1861                         break;
1862         }
1863         return copied - left;
1864 }
1865
1866 /*
1867  * Copy as much as we can into the page and return the number of bytes which
1868  * were successfully copied.  If a fault is encountered then return the number of
1869  * bytes which were copied.
1870  */
1871 size_t iov_iter_copy_from_user_atomic(struct page *page,
1872                 struct iov_iter *i, unsigned long offset, size_t bytes)
1873 {
1874         char *kaddr;
1875         size_t copied;
1876
1877         BUG_ON(!in_atomic());
1878         kaddr = kmap_atomic(page, KM_USER0);
1879         if (likely(i->nr_segs == 1)) {
1880                 int left;
1881                 char __user *buf = i->iov->iov_base + i->iov_offset;
1882                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1883                 copied = bytes - left;
1884         } else {
1885                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1886                                                 i->iov, i->iov_offset, bytes);
1887         }
1888         kunmap_atomic(kaddr, KM_USER0);
1889
1890         return copied;
1891 }
1892 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1893
1894 /*
1895  * This has the same sideeffects and return value as
1896  * iov_iter_copy_from_user_atomic().
1897  * The difference is that it attempts to resolve faults.
1898  * Page must not be locked.
1899  */
1900 size_t iov_iter_copy_from_user(struct page *page,
1901                 struct iov_iter *i, unsigned long offset, size_t bytes)
1902 {
1903         char *kaddr;
1904         size_t copied;
1905
1906         kaddr = kmap(page);
1907         if (likely(i->nr_segs == 1)) {
1908                 int left;
1909                 char __user *buf = i->iov->iov_base + i->iov_offset;
1910                 left = __copy_from_user(kaddr + offset, buf, bytes);
1911                 copied = bytes - left;
1912         } else {
1913                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1914                                                 i->iov, i->iov_offset, bytes);
1915         }
1916         kunmap(page);
1917         return copied;
1918 }
1919 EXPORT_SYMBOL(iov_iter_copy_from_user);
1920
1921 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1922 {
1923         BUG_ON(i->count < bytes);
1924
1925         if (likely(i->nr_segs == 1)) {
1926                 i->iov_offset += bytes;
1927                 i->count -= bytes;
1928         } else {
1929                 const struct iovec *iov = i->iov;
1930                 size_t base = i->iov_offset;
1931
1932                 /*
1933                  * The !iov->iov_len check ensures we skip over unlikely
1934                  * zero-length segments (without overruning the iovec).
1935                  */
1936                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1937                         int copy;
1938
1939                         copy = min(bytes, iov->iov_len - base);
1940                         BUG_ON(!i->count || i->count < copy);
1941                         i->count -= copy;
1942                         bytes -= copy;
1943                         base += copy;
1944                         if (iov->iov_len == base) {
1945                                 iov++;
1946                                 base = 0;
1947                         }
1948                 }
1949                 i->iov = iov;
1950                 i->iov_offset = base;
1951         }
1952 }
1953 EXPORT_SYMBOL(iov_iter_advance);
1954
1955 /*
1956  * Fault in the first iovec of the given iov_iter, to a maximum length
1957  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1958  * accessed (ie. because it is an invalid address).
1959  *
1960  * writev-intensive code may want this to prefault several iovecs -- that
1961  * would be possible (callers must not rely on the fact that _only_ the
1962  * first iovec will be faulted with the current implementation).
1963  */
1964 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1965 {
1966         char __user *buf = i->iov->iov_base + i->iov_offset;
1967         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1968         return fault_in_pages_readable(buf, bytes);
1969 }
1970 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1971
1972 /*
1973  * Return the count of just the current iov_iter segment.
1974  */
1975 size_t iov_iter_single_seg_count(struct iov_iter *i)
1976 {
1977         const struct iovec *iov = i->iov;
1978         if (i->nr_segs == 1)
1979                 return i->count;
1980         else
1981                 return min(i->count, iov->iov_len - i->iov_offset);
1982 }
1983 EXPORT_SYMBOL(iov_iter_single_seg_count);
1984
1985 /*
1986  * Performs necessary checks before doing a write
1987  *
1988  * Can adjust writing position or amount of bytes to write.
1989  * Returns appropriate error code that caller should return or
1990  * zero in case that write should be allowed.
1991  */
1992 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1993 {
1994         struct inode *inode = file->f_mapping->host;
1995         unsigned long limit = rlimit(RLIMIT_FSIZE);
1996
1997         if (unlikely(*pos < 0))
1998                 return -EINVAL;
1999
2000         if (!isblk) {
2001                 /* FIXME: this is for backwards compatibility with 2.4 */
2002                 if (file->f_flags & O_APPEND)
2003                         *pos = i_size_read(inode);
2004
2005                 if (limit != RLIM_INFINITY) {
2006                         if (*pos >= limit) {
2007                                 send_sig(SIGXFSZ, current, 0);
2008                                 return -EFBIG;
2009                         }
2010                         if (*count > limit - (typeof(limit))*pos) {
2011                                 *count = limit - (typeof(limit))*pos;
2012                         }
2013                 }
2014         }
2015
2016         /*
2017          * LFS rule
2018          */
2019         if (unlikely(*pos + *count > MAX_NON_LFS &&
2020                                 !(file->f_flags & O_LARGEFILE))) {
2021                 if (*pos >= MAX_NON_LFS) {
2022                         return -EFBIG;
2023                 }
2024                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2025                         *count = MAX_NON_LFS - (unsigned long)*pos;
2026                 }
2027         }
2028
2029         /*
2030          * Are we about to exceed the fs block limit ?
2031          *
2032          * If we have written data it becomes a short write.  If we have
2033          * exceeded without writing data we send a signal and return EFBIG.
2034          * Linus frestrict idea will clean these up nicely..
2035          */
2036         if (likely(!isblk)) {
2037                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2038                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2039                                 return -EFBIG;
2040                         }
2041                         /* zero-length writes at ->s_maxbytes are OK */
2042                 }
2043
2044                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2045                         *count = inode->i_sb->s_maxbytes - *pos;
2046         } else {
2047 #ifdef CONFIG_BLOCK
2048                 loff_t isize;
2049                 if (bdev_read_only(I_BDEV(inode)))
2050                         return -EPERM;
2051                 isize = i_size_read(inode);
2052                 if (*pos >= isize) {
2053                         if (*count || *pos > isize)
2054                                 return -ENOSPC;
2055                 }
2056
2057                 if (*pos + *count > isize)
2058                         *count = isize - *pos;
2059 #else
2060                 return -EPERM;
2061 #endif
2062         }
2063         return 0;
2064 }
2065 EXPORT_SYMBOL(generic_write_checks);
2066
2067 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2068                                 loff_t pos, unsigned len, unsigned flags,
2069                                 struct page **pagep, void **fsdata)
2070 {
2071         const struct address_space_operations *aops = mapping->a_ops;
2072
2073         return aops->write_begin(file, mapping, pos, len, flags,
2074                                                         pagep, fsdata);
2075 }
2076 EXPORT_SYMBOL(pagecache_write_begin);
2077
2078 int pagecache_write_end(struct file *file, struct address_space *mapping,
2079                                 loff_t pos, unsigned len, unsigned copied,
2080                                 struct page *page, void *fsdata)
2081 {
2082         const struct address_space_operations *aops = mapping->a_ops;
2083
2084         mark_page_accessed(page);
2085         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2086 }
2087 EXPORT_SYMBOL(pagecache_write_end);
2088
2089 ssize_t
2090 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2091                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2092                 size_t count, size_t ocount)
2093 {
2094         struct file     *file = iocb->ki_filp;
2095         struct address_space *mapping = file->f_mapping;
2096         struct inode    *inode = mapping->host;
2097         ssize_t         written;
2098         size_t          write_len;
2099         pgoff_t         end;
2100
2101         if (count != ocount)
2102                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2103
2104         write_len = iov_length(iov, *nr_segs);
2105         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2106
2107         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2108         if (written)
2109                 goto out;
2110
2111         /*
2112          * After a write we want buffered reads to be sure to go to disk to get
2113          * the new data.  We invalidate clean cached page from the region we're
2114          * about to write.  We do this *before* the write so that we can return
2115          * without clobbering -EIOCBQUEUED from ->direct_IO().
2116          */
2117         if (mapping->nrpages) {
2118                 written = invalidate_inode_pages2_range(mapping,
2119                                         pos >> PAGE_CACHE_SHIFT, end);
2120                 /*
2121                  * If a page can not be invalidated, return 0 to fall back
2122                  * to buffered write.
2123                  */
2124                 if (written) {
2125                         if (written == -EBUSY)
2126                                 return 0;
2127                         goto out;
2128                 }
2129         }
2130
2131         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2132
2133         /*
2134          * Finally, try again to invalidate clean pages which might have been
2135          * cached by non-direct readahead, or faulted in by get_user_pages()
2136          * if the source of the write was an mmap'ed region of the file
2137          * we're writing.  Either one is a pretty crazy thing to do,
2138          * so we don't support it 100%.  If this invalidation
2139          * fails, tough, the write still worked...
2140          */
2141         if (mapping->nrpages) {
2142                 invalidate_inode_pages2_range(mapping,
2143                                               pos >> PAGE_CACHE_SHIFT, end);
2144         }
2145
2146         if (written > 0) {
2147                 loff_t end = pos + written;
2148                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2149                         i_size_write(inode,  end);
2150                         mark_inode_dirty(inode);
2151                 }
2152                 *ppos = end;
2153         }
2154 out:
2155         return written;
2156 }
2157 EXPORT_SYMBOL(generic_file_direct_write);
2158
2159 /*
2160  * Find or create a page at the given pagecache position. Return the locked
2161  * page. This function is specifically for buffered writes.
2162  */
2163 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2164                                         pgoff_t index, unsigned flags)
2165 {
2166         int status;
2167         struct page *page;
2168         gfp_t gfp_notmask = 0;
2169         if (flags & AOP_FLAG_NOFS)
2170                 gfp_notmask = __GFP_FS;
2171 repeat:
2172         page = find_lock_page(mapping, index);
2173         if (likely(page))
2174                 return page;
2175
2176         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2177         if (!page)
2178                 return NULL;
2179         status = add_to_page_cache_lru(page, mapping, index,
2180                                                 GFP_KERNEL & ~gfp_notmask);
2181         if (unlikely(status)) {
2182                 page_cache_release(page);
2183                 if (status == -EEXIST)
2184                         goto repeat;
2185                 return NULL;
2186         }
2187         return page;
2188 }
2189 EXPORT_SYMBOL(grab_cache_page_write_begin);
2190
2191 static ssize_t generic_perform_write(struct file *file,
2192                                 struct iov_iter *i, loff_t pos)
2193 {
2194         struct address_space *mapping = file->f_mapping;
2195         const struct address_space_operations *a_ops = mapping->a_ops;
2196         long status = 0;
2197         ssize_t written = 0;
2198         unsigned int flags = 0;
2199
2200         /*
2201          * Copies from kernel address space cannot fail (NFSD is a big user).
2202          */
2203         if (segment_eq(get_fs(), KERNEL_DS))
2204                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2205
2206         do {
2207                 struct page *page;
2208                 pgoff_t index;          /* Pagecache index for current page */
2209                 unsigned long offset;   /* Offset into pagecache page */
2210                 unsigned long bytes;    /* Bytes to write to page */
2211                 size_t copied;          /* Bytes copied from user */
2212                 void *fsdata;
2213
2214                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2215                 index = pos >> PAGE_CACHE_SHIFT;
2216                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2217                                                 iov_iter_count(i));
2218
2219 again:
2220
2221                 /*
2222                  * Bring in the user page that we will copy from _first_.
2223                  * Otherwise there's a nasty deadlock on copying from the
2224                  * same page as we're writing to, without it being marked
2225                  * up-to-date.
2226                  *
2227                  * Not only is this an optimisation, but it is also required
2228                  * to check that the address is actually valid, when atomic
2229                  * usercopies are used, below.
2230                  */
2231                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2232                         status = -EFAULT;
2233                         break;
2234                 }
2235
2236                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2237                                                 &page, &fsdata);
2238                 if (unlikely(status))
2239                         break;
2240
2241                 if (mapping_writably_mapped(mapping))
2242                         flush_dcache_page(page);
2243
2244                 pagefault_disable();
2245                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2246                 pagefault_enable();
2247                 flush_dcache_page(page);
2248
2249                 mark_page_accessed(page);
2250                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2251                                                 page, fsdata);
2252                 if (unlikely(status < 0))
2253                         break;
2254                 copied = status;
2255
2256                 cond_resched();
2257
2258                 iov_iter_advance(i, copied);
2259                 if (unlikely(copied == 0)) {
2260                         /*
2261                          * If we were unable to copy any data at all, we must
2262                          * fall back to a single segment length write.
2263                          *
2264                          * If we didn't fallback here, we could livelock
2265                          * because not all segments in the iov can be copied at
2266                          * once without a pagefault.
2267                          */
2268                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2269                                                 iov_iter_single_seg_count(i));
2270                         goto again;
2271                 }
2272                 pos += copied;
2273                 written += copied;
2274
2275                 balance_dirty_pages_ratelimited(mapping);
2276
2277         } while (iov_iter_count(i));
2278
2279         return written ? written : status;
2280 }
2281
2282 ssize_t
2283 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2284                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2285                 size_t count, ssize_t written)
2286 {
2287         struct file *file = iocb->ki_filp;
2288         ssize_t status;
2289         struct iov_iter i;
2290
2291         iov_iter_init(&i, iov, nr_segs, count, written);
2292         status = generic_perform_write(file, &i, pos);
2293
2294         if (likely(status >= 0)) {
2295                 written += status;
2296                 *ppos = pos + status;
2297         }
2298         
2299         return written ? written : status;
2300 }
2301 EXPORT_SYMBOL(generic_file_buffered_write);
2302
2303 /**
2304  * __generic_file_aio_write - write data to a file
2305  * @iocb:       IO state structure (file, offset, etc.)
2306  * @iov:        vector with data to write
2307  * @nr_segs:    number of segments in the vector
2308  * @ppos:       position where to write
2309  *
2310  * This function does all the work needed for actually writing data to a
2311  * file. It does all basic checks, removes SUID from the file, updates
2312  * modification times and calls proper subroutines depending on whether we
2313  * do direct IO or a standard buffered write.
2314  *
2315  * It expects i_mutex to be grabbed unless we work on a block device or similar
2316  * object which does not need locking at all.
2317  *
2318  * This function does *not* take care of syncing data in case of O_SYNC write.
2319  * A caller has to handle it. This is mainly due to the fact that we want to
2320  * avoid syncing under i_mutex.
2321  */
2322 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2323                                  unsigned long nr_segs, loff_t *ppos)
2324 {
2325         struct file *file = iocb->ki_filp;
2326         struct address_space * mapping = file->f_mapping;
2327         size_t ocount;          /* original count */
2328         size_t count;           /* after file limit checks */
2329         struct inode    *inode = mapping->host;
2330         loff_t          pos;
2331         ssize_t         written;
2332         ssize_t         err;
2333
2334         ocount = 0;
2335         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2336         if (err)
2337                 return err;
2338
2339         count = ocount;
2340         pos = *ppos;
2341
2342         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2343
2344         /* We can write back this queue in page reclaim */
2345         current->backing_dev_info = mapping->backing_dev_info;
2346         written = 0;
2347
2348         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2349         if (err)
2350                 goto out;
2351
2352         if (count == 0)
2353                 goto out;
2354
2355         err = file_remove_suid(file);
2356         if (err)
2357                 goto out;
2358
2359         file_update_time(file);
2360
2361         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2362         if (unlikely(file->f_flags & O_DIRECT)) {
2363                 loff_t endbyte;
2364                 ssize_t written_buffered;
2365
2366                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2367                                                         ppos, count, ocount);
2368                 if (written < 0 || written == count)
2369                         goto out;
2370                 /*
2371                  * direct-io write to a hole: fall through to buffered I/O
2372                  * for completing the rest of the request.
2373                  */
2374                 pos += written;
2375                 count -= written;
2376                 written_buffered = generic_file_buffered_write(iocb, iov,
2377                                                 nr_segs, pos, ppos, count,
2378                                                 written);
2379                 /*
2380                  * If generic_file_buffered_write() retuned a synchronous error
2381                  * then we want to return the number of bytes which were
2382                  * direct-written, or the error code if that was zero.  Note
2383                  * that this differs from normal direct-io semantics, which
2384                  * will return -EFOO even if some bytes were written.
2385                  */
2386                 if (written_buffered < 0) {
2387                         err = written_buffered;
2388                         goto out;
2389                 }
2390
2391                 /*
2392                  * We need to ensure that the page cache pages are written to
2393                  * disk and invalidated to preserve the expected O_DIRECT
2394                  * semantics.
2395                  */
2396                 endbyte = pos + written_buffered - written - 1;
2397                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2398                 if (err == 0) {
2399                         written = written_buffered;
2400                         invalidate_mapping_pages(mapping,
2401                                                  pos >> PAGE_CACHE_SHIFT,
2402                                                  endbyte >> PAGE_CACHE_SHIFT);
2403                 } else {
2404                         /*
2405                          * We don't know how much we wrote, so just return
2406                          * the number of bytes which were direct-written
2407                          */
2408                 }
2409         } else {
2410                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2411                                 pos, ppos, count, written);
2412         }
2413 out:
2414         current->backing_dev_info = NULL;
2415         return written ? written : err;
2416 }
2417 EXPORT_SYMBOL(__generic_file_aio_write);
2418
2419 /**
2420  * generic_file_aio_write - write data to a file
2421  * @iocb:       IO state structure
2422  * @iov:        vector with data to write
2423  * @nr_segs:    number of segments in the vector
2424  * @pos:        position in file where to write
2425  *
2426  * This is a wrapper around __generic_file_aio_write() to be used by most
2427  * filesystems. It takes care of syncing the file in case of O_SYNC file
2428  * and acquires i_mutex as needed.
2429  */
2430 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2431                 unsigned long nr_segs, loff_t pos)
2432 {
2433         struct file *file = iocb->ki_filp;
2434         struct inode *inode = file->f_mapping->host;
2435         ssize_t ret;
2436
2437         BUG_ON(iocb->ki_pos != pos);
2438
2439         mutex_lock(&inode->i_mutex);
2440         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2441         mutex_unlock(&inode->i_mutex);
2442
2443         if (ret > 0 || ret == -EIOCBQUEUED) {
2444                 ssize_t err;
2445
2446                 err = generic_write_sync(file, pos, ret);
2447                 if (err < 0 && ret > 0)
2448                         ret = err;
2449         }
2450         return ret;
2451 }
2452 EXPORT_SYMBOL(generic_file_aio_write);
2453
2454 /**
2455  * try_to_release_page() - release old fs-specific metadata on a page
2456  *
2457  * @page: the page which the kernel is trying to free
2458  * @gfp_mask: memory allocation flags (and I/O mode)
2459  *
2460  * The address_space is to try to release any data against the page
2461  * (presumably at page->private).  If the release was successful, return `1'.
2462  * Otherwise return zero.
2463  *
2464  * This may also be called if PG_fscache is set on a page, indicating that the
2465  * page is known to the local caching routines.
2466  *
2467  * The @gfp_mask argument specifies whether I/O may be performed to release
2468  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2469  *
2470  */
2471 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2472 {
2473         struct address_space * const mapping = page->mapping;
2474
2475         BUG_ON(!PageLocked(page));
2476         if (PageWriteback(page))
2477                 return 0;
2478
2479         if (mapping && mapping->a_ops->releasepage)
2480                 return mapping->a_ops->releasepage(page, gfp_mask);
2481         return try_to_free_buffers(page);
2482 }
2483
2484 EXPORT_SYMBOL(try_to_release_page);