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