readahead: clean up and simplify the code for filemap page fault readahead
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for generic_osync_inode */
43
44 #include <asm/mman.h>
45
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  ->inode_lock
85  *    ->sb_lock                 (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_lock
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  *  ->task->proc_lock
107  *    ->dcache_lock             (proc_pid_lookup)
108  */
109
110 /*
111  * Remove a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold the mapping's tree_lock.
114  */
115 void __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124
125         /*
126          * Some filesystems seem to re-dirty the page even after
127          * the VM has canceled the dirty bit (eg ext3 journaling).
128          *
129          * Fix it up by doing a final dirty accounting check after
130          * having removed the page entirely.
131          */
132         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
133                 dec_zone_page_state(page, NR_FILE_DIRTY);
134                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
135         }
136 }
137
138 void remove_from_page_cache(struct page *page)
139 {
140         struct address_space *mapping = page->mapping;
141
142         BUG_ON(!PageLocked(page));
143
144         spin_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         spin_unlock_irq(&mapping->tree_lock);
147         mem_cgroup_uncharge_cache_page(page);
148 }
149
150 static int sync_page(void *word)
151 {
152         struct address_space *mapping;
153         struct page *page;
154
155         page = container_of((unsigned long *)word, struct page, flags);
156
157         /*
158          * page_mapping() is being called without PG_locked held.
159          * Some knowledge of the state and use of the page is used to
160          * reduce the requirements down to a memory barrier.
161          * The danger here is of a stale page_mapping() return value
162          * indicating a struct address_space different from the one it's
163          * associated with when it is associated with one.
164          * After smp_mb(), it's either the correct page_mapping() for
165          * the page, or an old page_mapping() and the page's own
166          * page_mapping() has gone NULL.
167          * The ->sync_page() address_space operation must tolerate
168          * page_mapping() going NULL. By an amazing coincidence,
169          * this comes about because none of the users of the page
170          * in the ->sync_page() methods make essential use of the
171          * page_mapping(), merely passing the page down to the backing
172          * device's unplug functions when it's non-NULL, which in turn
173          * ignore it for all cases but swap, where only page_private(page) is
174          * of interest. When page_mapping() does go NULL, the entire
175          * call stack gracefully ignores the page and returns.
176          * -- wli
177          */
178         smp_mb();
179         mapping = page_mapping(page);
180         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
181                 mapping->a_ops->sync_page(page);
182         io_schedule();
183         return 0;
184 }
185
186 static int sync_page_killable(void *word)
187 {
188         sync_page(word);
189         return fatal_signal_pending(current) ? -EINTR : 0;
190 }
191
192 /**
193  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
194  * @mapping:    address space structure to write
195  * @start:      offset in bytes where the range starts
196  * @end:        offset in bytes where the range ends (inclusive)
197  * @sync_mode:  enable synchronous operation
198  *
199  * Start writeback against all of a mapping's dirty pages that lie
200  * within the byte offsets <start, end> inclusive.
201  *
202  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
203  * opposed to a regular memory cleansing writeback.  The difference between
204  * these two operations is that if a dirty page/buffer is encountered, it must
205  * be waited upon, and not just skipped over.
206  */
207 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
208                                 loff_t end, int sync_mode)
209 {
210         int ret;
211         struct writeback_control wbc = {
212                 .sync_mode = sync_mode,
213                 .nr_to_write = LONG_MAX,
214                 .range_start = start,
215                 .range_end = end,
216         };
217
218         if (!mapping_cap_writeback_dirty(mapping))
219                 return 0;
220
221         ret = do_writepages(mapping, &wbc);
222         return ret;
223 }
224
225 static inline int __filemap_fdatawrite(struct address_space *mapping,
226         int sync_mode)
227 {
228         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
229 }
230
231 int filemap_fdatawrite(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
234 }
235 EXPORT_SYMBOL(filemap_fdatawrite);
236
237 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
238                                 loff_t end)
239 {
240         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
241 }
242 EXPORT_SYMBOL(filemap_fdatawrite_range);
243
244 /**
245  * filemap_flush - mostly a non-blocking flush
246  * @mapping:    target address_space
247  *
248  * This is a mostly non-blocking flush.  Not suitable for data-integrity
249  * purposes - I/O may not be started against all dirty pages.
250  */
251 int filemap_flush(struct address_space *mapping)
252 {
253         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
254 }
255 EXPORT_SYMBOL(filemap_flush);
256
257 /**
258  * wait_on_page_writeback_range - wait for writeback to complete
259  * @mapping:    target address_space
260  * @start:      beginning page index
261  * @end:        ending page index
262  *
263  * Wait for writeback to complete against pages indexed by start->end
264  * inclusive
265  */
266 int wait_on_page_writeback_range(struct address_space *mapping,
267                                 pgoff_t start, pgoff_t end)
268 {
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272         pgoff_t index;
273
274         if (end < start)
275                 return 0;
276
277         pagevec_init(&pvec, 0);
278         index = start;
279         while ((index <= end) &&
280                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
281                         PAGECACHE_TAG_WRITEBACK,
282                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
283                 unsigned i;
284
285                 for (i = 0; i < nr_pages; i++) {
286                         struct page *page = pvec.pages[i];
287
288                         /* until radix tree lookup accepts end_index */
289                         if (page->index > end)
290                                 continue;
291
292                         wait_on_page_writeback(page);
293                         if (PageError(page))
294                                 ret = -EIO;
295                 }
296                 pagevec_release(&pvec);
297                 cond_resched();
298         }
299
300         /* Check for outstanding write errors */
301         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
302                 ret = -ENOSPC;
303         if (test_and_clear_bit(AS_EIO, &mapping->flags))
304                 ret = -EIO;
305
306         return ret;
307 }
308
309 /**
310  * sync_page_range - write and wait on all pages in the passed range
311  * @inode:      target inode
312  * @mapping:    target address_space
313  * @pos:        beginning offset in pages to write
314  * @count:      number of bytes to write
315  *
316  * Write and wait upon all the pages in the passed range.  This is a "data
317  * integrity" operation.  It waits upon in-flight writeout before starting and
318  * waiting upon new writeout.  If there was an IO error, return it.
319  *
320  * We need to re-take i_mutex during the generic_osync_inode list walk because
321  * it is otherwise livelockable.
322  */
323 int sync_page_range(struct inode *inode, struct address_space *mapping,
324                         loff_t pos, loff_t count)
325 {
326         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
327         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
328         int ret;
329
330         if (!mapping_cap_writeback_dirty(mapping) || !count)
331                 return 0;
332         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
333         if (ret == 0) {
334                 mutex_lock(&inode->i_mutex);
335                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
336                 mutex_unlock(&inode->i_mutex);
337         }
338         if (ret == 0)
339                 ret = wait_on_page_writeback_range(mapping, start, end);
340         return ret;
341 }
342 EXPORT_SYMBOL(sync_page_range);
343
344 /**
345  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
346  * @inode:      target inode
347  * @mapping:    target address_space
348  * @pos:        beginning offset in pages to write
349  * @count:      number of bytes to write
350  *
351  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
352  * as it forces O_SYNC writers to different parts of the same file
353  * to be serialised right until io completion.
354  */
355 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
356                            loff_t pos, loff_t count)
357 {
358         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
359         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
360         int ret;
361
362         if (!mapping_cap_writeback_dirty(mapping) || !count)
363                 return 0;
364         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
365         if (ret == 0)
366                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
367         if (ret == 0)
368                 ret = wait_on_page_writeback_range(mapping, start, end);
369         return ret;
370 }
371 EXPORT_SYMBOL(sync_page_range_nolock);
372
373 /**
374  * filemap_fdatawait - wait for all under-writeback pages to complete
375  * @mapping: address space structure to wait for
376  *
377  * Walk the list of under-writeback pages of the given address space
378  * and wait for all of them.
379  */
380 int filemap_fdatawait(struct address_space *mapping)
381 {
382         loff_t i_size = i_size_read(mapping->host);
383
384         if (i_size == 0)
385                 return 0;
386
387         return wait_on_page_writeback_range(mapping, 0,
388                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
389 }
390 EXPORT_SYMBOL(filemap_fdatawait);
391
392 int filemap_write_and_wait(struct address_space *mapping)
393 {
394         int err = 0;
395
396         if (mapping->nrpages) {
397                 err = filemap_fdatawrite(mapping);
398                 /*
399                  * Even if the above returned error, the pages may be
400                  * written partially (e.g. -ENOSPC), so we wait for it.
401                  * But the -EIO is special case, it may indicate the worst
402                  * thing (e.g. bug) happened, so we avoid waiting for it.
403                  */
404                 if (err != -EIO) {
405                         int err2 = filemap_fdatawait(mapping);
406                         if (!err)
407                                 err = err2;
408                 }
409         }
410         return err;
411 }
412 EXPORT_SYMBOL(filemap_write_and_wait);
413
414 /**
415  * filemap_write_and_wait_range - write out & wait on a file range
416  * @mapping:    the address_space for the pages
417  * @lstart:     offset in bytes where the range starts
418  * @lend:       offset in bytes where the range ends (inclusive)
419  *
420  * Write out and wait upon file offsets lstart->lend, inclusive.
421  *
422  * Note that `lend' is inclusive (describes the last byte to be written) so
423  * that this function can be used to write to the very end-of-file (end = -1).
424  */
425 int filemap_write_and_wait_range(struct address_space *mapping,
426                                  loff_t lstart, loff_t lend)
427 {
428         int err = 0;
429
430         if (mapping->nrpages) {
431                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
432                                                  WB_SYNC_ALL);
433                 /* See comment of filemap_write_and_wait() */
434                 if (err != -EIO) {
435                         int err2 = wait_on_page_writeback_range(mapping,
436                                                 lstart >> PAGE_CACHE_SHIFT,
437                                                 lend >> PAGE_CACHE_SHIFT);
438                         if (!err)
439                                 err = err2;
440                 }
441         }
442         return err;
443 }
444 EXPORT_SYMBOL(filemap_write_and_wait_range);
445
446 /**
447  * add_to_page_cache_locked - add a locked page to the pagecache
448  * @page:       page to add
449  * @mapping:    the page's address_space
450  * @offset:     page index
451  * @gfp_mask:   page allocation mode
452  *
453  * This function is used to add a page to the pagecache. It must be locked.
454  * This function does not add the page to the LRU.  The caller must do that.
455  */
456 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
457                 pgoff_t offset, gfp_t gfp_mask)
458 {
459         int error;
460
461         VM_BUG_ON(!PageLocked(page));
462
463         error = mem_cgroup_cache_charge(page, current->mm,
464                                         gfp_mask & GFP_RECLAIM_MASK);
465         if (error)
466                 goto out;
467
468         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
469         if (error == 0) {
470                 page_cache_get(page);
471                 page->mapping = mapping;
472                 page->index = offset;
473
474                 spin_lock_irq(&mapping->tree_lock);
475                 error = radix_tree_insert(&mapping->page_tree, offset, page);
476                 if (likely(!error)) {
477                         mapping->nrpages++;
478                         __inc_zone_page_state(page, NR_FILE_PAGES);
479                         spin_unlock_irq(&mapping->tree_lock);
480                 } else {
481                         page->mapping = NULL;
482                         spin_unlock_irq(&mapping->tree_lock);
483                         mem_cgroup_uncharge_cache_page(page);
484                         page_cache_release(page);
485                 }
486                 radix_tree_preload_end();
487         } else
488                 mem_cgroup_uncharge_cache_page(page);
489 out:
490         return error;
491 }
492 EXPORT_SYMBOL(add_to_page_cache_locked);
493
494 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
495                                 pgoff_t offset, gfp_t gfp_mask)
496 {
497         int ret;
498
499         /*
500          * Splice_read and readahead add shmem/tmpfs pages into the page cache
501          * before shmem_readpage has a chance to mark them as SwapBacked: they
502          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
503          * (called in add_to_page_cache) needs to know where they're going too.
504          */
505         if (mapping_cap_swap_backed(mapping))
506                 SetPageSwapBacked(page);
507
508         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
509         if (ret == 0) {
510                 if (page_is_file_cache(page))
511                         lru_cache_add_file(page);
512                 else
513                         lru_cache_add_active_anon(page);
514         }
515         return ret;
516 }
517 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
518
519 #ifdef CONFIG_NUMA
520 struct page *__page_cache_alloc(gfp_t gfp)
521 {
522         if (cpuset_do_page_mem_spread()) {
523                 int n = cpuset_mem_spread_node();
524                 return alloc_pages_node(n, gfp, 0);
525         }
526         return alloc_pages(gfp, 0);
527 }
528 EXPORT_SYMBOL(__page_cache_alloc);
529 #endif
530
531 static int __sleep_on_page_lock(void *word)
532 {
533         io_schedule();
534         return 0;
535 }
536
537 /*
538  * In order to wait for pages to become available there must be
539  * waitqueues associated with pages. By using a hash table of
540  * waitqueues where the bucket discipline is to maintain all
541  * waiters on the same queue and wake all when any of the pages
542  * become available, and for the woken contexts to check to be
543  * sure the appropriate page became available, this saves space
544  * at a cost of "thundering herd" phenomena during rare hash
545  * collisions.
546  */
547 static wait_queue_head_t *page_waitqueue(struct page *page)
548 {
549         const struct zone *zone = page_zone(page);
550
551         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
552 }
553
554 static inline void wake_up_page(struct page *page, int bit)
555 {
556         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
557 }
558
559 void wait_on_page_bit(struct page *page, int bit_nr)
560 {
561         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
562
563         if (test_bit(bit_nr, &page->flags))
564                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
565                                                         TASK_UNINTERRUPTIBLE);
566 }
567 EXPORT_SYMBOL(wait_on_page_bit);
568
569 /**
570  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
571  * @page: Page defining the wait queue of interest
572  * @waiter: Waiter to add to the queue
573  *
574  * Add an arbitrary @waiter to the wait queue for the nominated @page.
575  */
576 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
577 {
578         wait_queue_head_t *q = page_waitqueue(page);
579         unsigned long flags;
580
581         spin_lock_irqsave(&q->lock, flags);
582         __add_wait_queue(q, waiter);
583         spin_unlock_irqrestore(&q->lock, flags);
584 }
585 EXPORT_SYMBOL_GPL(add_page_wait_queue);
586
587 /**
588  * unlock_page - unlock a locked page
589  * @page: the page
590  *
591  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
592  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
593  * mechananism between PageLocked pages and PageWriteback pages is shared.
594  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
595  *
596  * The mb is necessary to enforce ordering between the clear_bit and the read
597  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
598  */
599 void unlock_page(struct page *page)
600 {
601         VM_BUG_ON(!PageLocked(page));
602         clear_bit_unlock(PG_locked, &page->flags);
603         smp_mb__after_clear_bit();
604         wake_up_page(page, PG_locked);
605 }
606 EXPORT_SYMBOL(unlock_page);
607
608 /**
609  * end_page_writeback - end writeback against a page
610  * @page: the page
611  */
612 void end_page_writeback(struct page *page)
613 {
614         if (TestClearPageReclaim(page))
615                 rotate_reclaimable_page(page);
616
617         if (!test_clear_page_writeback(page))
618                 BUG();
619
620         smp_mb__after_clear_bit();
621         wake_up_page(page, PG_writeback);
622 }
623 EXPORT_SYMBOL(end_page_writeback);
624
625 /**
626  * __lock_page - get a lock on the page, assuming we need to sleep to get it
627  * @page: the page to lock
628  *
629  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
630  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
631  * chances are that on the second loop, the block layer's plug list is empty,
632  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
633  */
634 void __lock_page(struct page *page)
635 {
636         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
637
638         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
639                                                         TASK_UNINTERRUPTIBLE);
640 }
641 EXPORT_SYMBOL(__lock_page);
642
643 int __lock_page_killable(struct page *page)
644 {
645         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
646
647         return __wait_on_bit_lock(page_waitqueue(page), &wait,
648                                         sync_page_killable, TASK_KILLABLE);
649 }
650 EXPORT_SYMBOL_GPL(__lock_page_killable);
651
652 /**
653  * __lock_page_nosync - get a lock on the page, without calling sync_page()
654  * @page: the page to lock
655  *
656  * Variant of lock_page that does not require the caller to hold a reference
657  * on the page's mapping.
658  */
659 void __lock_page_nosync(struct page *page)
660 {
661         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
662         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
663                                                         TASK_UNINTERRUPTIBLE);
664 }
665
666 /**
667  * find_get_page - find and get a page reference
668  * @mapping: the address_space to search
669  * @offset: the page index
670  *
671  * Is there a pagecache struct page at the given (mapping, offset) tuple?
672  * If yes, increment its refcount and return it; if no, return NULL.
673  */
674 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
675 {
676         void **pagep;
677         struct page *page;
678
679         rcu_read_lock();
680 repeat:
681         page = NULL;
682         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
683         if (pagep) {
684                 page = radix_tree_deref_slot(pagep);
685                 if (unlikely(!page || page == RADIX_TREE_RETRY))
686                         goto repeat;
687
688                 if (!page_cache_get_speculative(page))
689                         goto repeat;
690
691                 /*
692                  * Has the page moved?
693                  * This is part of the lockless pagecache protocol. See
694                  * include/linux/pagemap.h for details.
695                  */
696                 if (unlikely(page != *pagep)) {
697                         page_cache_release(page);
698                         goto repeat;
699                 }
700         }
701         rcu_read_unlock();
702
703         return page;
704 }
705 EXPORT_SYMBOL(find_get_page);
706
707 /**
708  * find_lock_page - locate, pin and lock a pagecache page
709  * @mapping: the address_space to search
710  * @offset: the page index
711  *
712  * Locates the desired pagecache page, locks it, increments its reference
713  * count and returns its address.
714  *
715  * Returns zero if the page was not present. find_lock_page() may sleep.
716  */
717 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
718 {
719         struct page *page;
720
721 repeat:
722         page = find_get_page(mapping, offset);
723         if (page) {
724                 lock_page(page);
725                 /* Has the page been truncated? */
726                 if (unlikely(page->mapping != mapping)) {
727                         unlock_page(page);
728                         page_cache_release(page);
729                         goto repeat;
730                 }
731                 VM_BUG_ON(page->index != offset);
732         }
733         return page;
734 }
735 EXPORT_SYMBOL(find_lock_page);
736
737 /**
738  * find_or_create_page - locate or add a pagecache page
739  * @mapping: the page's address_space
740  * @index: the page's index into the mapping
741  * @gfp_mask: page allocation mode
742  *
743  * Locates a page in the pagecache.  If the page is not present, a new page
744  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
745  * LRU list.  The returned page is locked and has its reference count
746  * incremented.
747  *
748  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
749  * allocation!
750  *
751  * find_or_create_page() returns the desired page's address, or zero on
752  * memory exhaustion.
753  */
754 struct page *find_or_create_page(struct address_space *mapping,
755                 pgoff_t index, gfp_t gfp_mask)
756 {
757         struct page *page;
758         int err;
759 repeat:
760         page = find_lock_page(mapping, index);
761         if (!page) {
762                 page = __page_cache_alloc(gfp_mask);
763                 if (!page)
764                         return NULL;
765                 /*
766                  * We want a regular kernel memory (not highmem or DMA etc)
767                  * allocation for the radix tree nodes, but we need to honour
768                  * the context-specific requirements the caller has asked for.
769                  * GFP_RECLAIM_MASK collects those requirements.
770                  */
771                 err = add_to_page_cache_lru(page, mapping, index,
772                         (gfp_mask & GFP_RECLAIM_MASK));
773                 if (unlikely(err)) {
774                         page_cache_release(page);
775                         page = NULL;
776                         if (err == -EEXIST)
777                                 goto repeat;
778                 }
779         }
780         return page;
781 }
782 EXPORT_SYMBOL(find_or_create_page);
783
784 /**
785  * find_get_pages - gang pagecache lookup
786  * @mapping:    The address_space to search
787  * @start:      The starting page index
788  * @nr_pages:   The maximum number of pages
789  * @pages:      Where the resulting pages are placed
790  *
791  * find_get_pages() will search for and return a group of up to
792  * @nr_pages pages in the mapping.  The pages are placed at @pages.
793  * find_get_pages() takes a reference against the returned pages.
794  *
795  * The search returns a group of mapping-contiguous pages with ascending
796  * indexes.  There may be holes in the indices due to not-present pages.
797  *
798  * find_get_pages() returns the number of pages which were found.
799  */
800 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
801                             unsigned int nr_pages, struct page **pages)
802 {
803         unsigned int i;
804         unsigned int ret;
805         unsigned int nr_found;
806
807         rcu_read_lock();
808 restart:
809         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
810                                 (void ***)pages, start, nr_pages);
811         ret = 0;
812         for (i = 0; i < nr_found; i++) {
813                 struct page *page;
814 repeat:
815                 page = radix_tree_deref_slot((void **)pages[i]);
816                 if (unlikely(!page))
817                         continue;
818                 /*
819                  * this can only trigger if nr_found == 1, making livelock
820                  * a non issue.
821                  */
822                 if (unlikely(page == RADIX_TREE_RETRY))
823                         goto restart;
824
825                 if (!page_cache_get_speculative(page))
826                         goto repeat;
827
828                 /* Has the page moved? */
829                 if (unlikely(page != *((void **)pages[i]))) {
830                         page_cache_release(page);
831                         goto repeat;
832                 }
833
834                 pages[ret] = page;
835                 ret++;
836         }
837         rcu_read_unlock();
838         return ret;
839 }
840
841 /**
842  * find_get_pages_contig - gang contiguous pagecache lookup
843  * @mapping:    The address_space to search
844  * @index:      The starting page index
845  * @nr_pages:   The maximum number of pages
846  * @pages:      Where the resulting pages are placed
847  *
848  * find_get_pages_contig() works exactly like find_get_pages(), except
849  * that the returned number of pages are guaranteed to be contiguous.
850  *
851  * find_get_pages_contig() returns the number of pages which were found.
852  */
853 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
854                                unsigned int nr_pages, struct page **pages)
855 {
856         unsigned int i;
857         unsigned int ret;
858         unsigned int nr_found;
859
860         rcu_read_lock();
861 restart:
862         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
863                                 (void ***)pages, index, nr_pages);
864         ret = 0;
865         for (i = 0; i < nr_found; i++) {
866                 struct page *page;
867 repeat:
868                 page = radix_tree_deref_slot((void **)pages[i]);
869                 if (unlikely(!page))
870                         continue;
871                 /*
872                  * this can only trigger if nr_found == 1, making livelock
873                  * a non issue.
874                  */
875                 if (unlikely(page == RADIX_TREE_RETRY))
876                         goto restart;
877
878                 if (page->mapping == NULL || page->index != index)
879                         break;
880
881                 if (!page_cache_get_speculative(page))
882                         goto repeat;
883
884                 /* Has the page moved? */
885                 if (unlikely(page != *((void **)pages[i]))) {
886                         page_cache_release(page);
887                         goto repeat;
888                 }
889
890                 pages[ret] = page;
891                 ret++;
892                 index++;
893         }
894         rcu_read_unlock();
895         return ret;
896 }
897 EXPORT_SYMBOL(find_get_pages_contig);
898
899 /**
900  * find_get_pages_tag - find and return pages that match @tag
901  * @mapping:    the address_space to search
902  * @index:      the starting page index
903  * @tag:        the tag index
904  * @nr_pages:   the maximum number of pages
905  * @pages:      where the resulting pages are placed
906  *
907  * Like find_get_pages, except we only return pages which are tagged with
908  * @tag.   We update @index to index the next page for the traversal.
909  */
910 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
911                         int tag, unsigned int nr_pages, struct page **pages)
912 {
913         unsigned int i;
914         unsigned int ret;
915         unsigned int nr_found;
916
917         rcu_read_lock();
918 restart:
919         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
920                                 (void ***)pages, *index, nr_pages, tag);
921         ret = 0;
922         for (i = 0; i < nr_found; i++) {
923                 struct page *page;
924 repeat:
925                 page = radix_tree_deref_slot((void **)pages[i]);
926                 if (unlikely(!page))
927                         continue;
928                 /*
929                  * this can only trigger if nr_found == 1, making livelock
930                  * a non issue.
931                  */
932                 if (unlikely(page == RADIX_TREE_RETRY))
933                         goto restart;
934
935                 if (!page_cache_get_speculative(page))
936                         goto repeat;
937
938                 /* Has the page moved? */
939                 if (unlikely(page != *((void **)pages[i]))) {
940                         page_cache_release(page);
941                         goto repeat;
942                 }
943
944                 pages[ret] = page;
945                 ret++;
946         }
947         rcu_read_unlock();
948
949         if (ret)
950                 *index = pages[ret - 1]->index + 1;
951
952         return ret;
953 }
954 EXPORT_SYMBOL(find_get_pages_tag);
955
956 /**
957  * grab_cache_page_nowait - returns locked page at given index in given cache
958  * @mapping: target address_space
959  * @index: the page index
960  *
961  * Same as grab_cache_page(), but do not wait if the page is unavailable.
962  * This is intended for speculative data generators, where the data can
963  * be regenerated if the page couldn't be grabbed.  This routine should
964  * be safe to call while holding the lock for another page.
965  *
966  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
967  * and deadlock against the caller's locked page.
968  */
969 struct page *
970 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
971 {
972         struct page *page = find_get_page(mapping, index);
973
974         if (page) {
975                 if (trylock_page(page))
976                         return page;
977                 page_cache_release(page);
978                 return NULL;
979         }
980         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
981         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
982                 page_cache_release(page);
983                 page = NULL;
984         }
985         return page;
986 }
987 EXPORT_SYMBOL(grab_cache_page_nowait);
988
989 /*
990  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
991  * a _large_ part of the i/o request. Imagine the worst scenario:
992  *
993  *      ---R__________________________________________B__________
994  *         ^ reading here                             ^ bad block(assume 4k)
995  *
996  * read(R) => miss => readahead(R...B) => media error => frustrating retries
997  * => failing the whole request => read(R) => read(R+1) =>
998  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
999  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1000  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1001  *
1002  * It is going insane. Fix it by quickly scaling down the readahead size.
1003  */
1004 static void shrink_readahead_size_eio(struct file *filp,
1005                                         struct file_ra_state *ra)
1006 {
1007         if (!ra->ra_pages)
1008                 return;
1009
1010         ra->ra_pages /= 4;
1011 }
1012
1013 /**
1014  * do_generic_file_read - generic file read routine
1015  * @filp:       the file to read
1016  * @ppos:       current file position
1017  * @desc:       read_descriptor
1018  * @actor:      read method
1019  *
1020  * This is a generic file read routine, and uses the
1021  * mapping->a_ops->readpage() function for the actual low-level stuff.
1022  *
1023  * This is really ugly. But the goto's actually try to clarify some
1024  * of the logic when it comes to error handling etc.
1025  */
1026 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1027                 read_descriptor_t *desc, read_actor_t actor)
1028 {
1029         struct address_space *mapping = filp->f_mapping;
1030         struct inode *inode = mapping->host;
1031         struct file_ra_state *ra = &filp->f_ra;
1032         pgoff_t index;
1033         pgoff_t last_index;
1034         pgoff_t prev_index;
1035         unsigned long offset;      /* offset into pagecache page */
1036         unsigned int prev_offset;
1037         int error;
1038
1039         index = *ppos >> PAGE_CACHE_SHIFT;
1040         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1041         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1042         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1043         offset = *ppos & ~PAGE_CACHE_MASK;
1044
1045         for (;;) {
1046                 struct page *page;
1047                 pgoff_t end_index;
1048                 loff_t isize;
1049                 unsigned long nr, ret;
1050
1051                 cond_resched();
1052 find_page:
1053                 page = find_get_page(mapping, index);
1054                 if (!page) {
1055                         page_cache_sync_readahead(mapping,
1056                                         ra, filp,
1057                                         index, last_index - index);
1058                         page = find_get_page(mapping, index);
1059                         if (unlikely(page == NULL))
1060                                 goto no_cached_page;
1061                 }
1062                 if (PageReadahead(page)) {
1063                         page_cache_async_readahead(mapping,
1064                                         ra, filp, page,
1065                                         index, last_index - index);
1066                 }
1067                 if (!PageUptodate(page)) {
1068                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1069                                         !mapping->a_ops->is_partially_uptodate)
1070                                 goto page_not_up_to_date;
1071                         if (!trylock_page(page))
1072                                 goto page_not_up_to_date;
1073                         if (!mapping->a_ops->is_partially_uptodate(page,
1074                                                                 desc, offset))
1075                                 goto page_not_up_to_date_locked;
1076                         unlock_page(page);
1077                 }
1078 page_ok:
1079                 /*
1080                  * i_size must be checked after we know the page is Uptodate.
1081                  *
1082                  * Checking i_size after the check allows us to calculate
1083                  * the correct value for "nr", which means the zero-filled
1084                  * part of the page is not copied back to userspace (unless
1085                  * another truncate extends the file - this is desired though).
1086                  */
1087
1088                 isize = i_size_read(inode);
1089                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1090                 if (unlikely(!isize || index > end_index)) {
1091                         page_cache_release(page);
1092                         goto out;
1093                 }
1094
1095                 /* nr is the maximum number of bytes to copy from this page */
1096                 nr = PAGE_CACHE_SIZE;
1097                 if (index == end_index) {
1098                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1099                         if (nr <= offset) {
1100                                 page_cache_release(page);
1101                                 goto out;
1102                         }
1103                 }
1104                 nr = nr - offset;
1105
1106                 /* If users can be writing to this page using arbitrary
1107                  * virtual addresses, take care about potential aliasing
1108                  * before reading the page on the kernel side.
1109                  */
1110                 if (mapping_writably_mapped(mapping))
1111                         flush_dcache_page(page);
1112
1113                 /*
1114                  * When a sequential read accesses a page several times,
1115                  * only mark it as accessed the first time.
1116                  */
1117                 if (prev_index != index || offset != prev_offset)
1118                         mark_page_accessed(page);
1119                 prev_index = index;
1120
1121                 /*
1122                  * Ok, we have the page, and it's up-to-date, so
1123                  * now we can copy it to user space...
1124                  *
1125                  * The actor routine returns how many bytes were actually used..
1126                  * NOTE! This may not be the same as how much of a user buffer
1127                  * we filled up (we may be padding etc), so we can only update
1128                  * "pos" here (the actor routine has to update the user buffer
1129                  * pointers and the remaining count).
1130                  */
1131                 ret = actor(desc, page, offset, nr);
1132                 offset += ret;
1133                 index += offset >> PAGE_CACHE_SHIFT;
1134                 offset &= ~PAGE_CACHE_MASK;
1135                 prev_offset = offset;
1136
1137                 page_cache_release(page);
1138                 if (ret == nr && desc->count)
1139                         continue;
1140                 goto out;
1141
1142 page_not_up_to_date:
1143                 /* Get exclusive access to the page ... */
1144                 error = lock_page_killable(page);
1145                 if (unlikely(error))
1146                         goto readpage_error;
1147
1148 page_not_up_to_date_locked:
1149                 /* Did it get truncated before we got the lock? */
1150                 if (!page->mapping) {
1151                         unlock_page(page);
1152                         page_cache_release(page);
1153                         continue;
1154                 }
1155
1156                 /* Did somebody else fill it already? */
1157                 if (PageUptodate(page)) {
1158                         unlock_page(page);
1159                         goto page_ok;
1160                 }
1161
1162 readpage:
1163                 /* Start the actual read. The read will unlock the page. */
1164                 error = mapping->a_ops->readpage(filp, page);
1165
1166                 if (unlikely(error)) {
1167                         if (error == AOP_TRUNCATED_PAGE) {
1168                                 page_cache_release(page);
1169                                 goto find_page;
1170                         }
1171                         goto readpage_error;
1172                 }
1173
1174                 if (!PageUptodate(page)) {
1175                         error = lock_page_killable(page);
1176                         if (unlikely(error))
1177                                 goto readpage_error;
1178                         if (!PageUptodate(page)) {
1179                                 if (page->mapping == NULL) {
1180                                         /*
1181                                          * invalidate_inode_pages got it
1182                                          */
1183                                         unlock_page(page);
1184                                         page_cache_release(page);
1185                                         goto find_page;
1186                                 }
1187                                 unlock_page(page);
1188                                 shrink_readahead_size_eio(filp, ra);
1189                                 error = -EIO;
1190                                 goto readpage_error;
1191                         }
1192                         unlock_page(page);
1193                 }
1194
1195                 goto page_ok;
1196
1197 readpage_error:
1198                 /* UHHUH! A synchronous read error occurred. Report it */
1199                 desc->error = error;
1200                 page_cache_release(page);
1201                 goto out;
1202
1203 no_cached_page:
1204                 /*
1205                  * Ok, it wasn't cached, so we need to create a new
1206                  * page..
1207                  */
1208                 page = page_cache_alloc_cold(mapping);
1209                 if (!page) {
1210                         desc->error = -ENOMEM;
1211                         goto out;
1212                 }
1213                 error = add_to_page_cache_lru(page, mapping,
1214                                                 index, GFP_KERNEL);
1215                 if (error) {
1216                         page_cache_release(page);
1217                         if (error == -EEXIST)
1218                                 goto find_page;
1219                         desc->error = error;
1220                         goto out;
1221                 }
1222                 goto readpage;
1223         }
1224
1225 out:
1226         ra->prev_pos = prev_index;
1227         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1228         ra->prev_pos |= prev_offset;
1229
1230         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1231         file_accessed(filp);
1232 }
1233
1234 int file_read_actor(read_descriptor_t *desc, struct page *page,
1235                         unsigned long offset, unsigned long size)
1236 {
1237         char *kaddr;
1238         unsigned long left, count = desc->count;
1239
1240         if (size > count)
1241                 size = count;
1242
1243         /*
1244          * Faults on the destination of a read are common, so do it before
1245          * taking the kmap.
1246          */
1247         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1248                 kaddr = kmap_atomic(page, KM_USER0);
1249                 left = __copy_to_user_inatomic(desc->arg.buf,
1250                                                 kaddr + offset, size);
1251                 kunmap_atomic(kaddr, KM_USER0);
1252                 if (left == 0)
1253                         goto success;
1254         }
1255
1256         /* Do it the slow way */
1257         kaddr = kmap(page);
1258         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1259         kunmap(page);
1260
1261         if (left) {
1262                 size -= left;
1263                 desc->error = -EFAULT;
1264         }
1265 success:
1266         desc->count = count - size;
1267         desc->written += size;
1268         desc->arg.buf += size;
1269         return size;
1270 }
1271
1272 /*
1273  * Performs necessary checks before doing a write
1274  * @iov:        io vector request
1275  * @nr_segs:    number of segments in the iovec
1276  * @count:      number of bytes to write
1277  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1278  *
1279  * Adjust number of segments and amount of bytes to write (nr_segs should be
1280  * properly initialized first). Returns appropriate error code that caller
1281  * should return or zero in case that write should be allowed.
1282  */
1283 int generic_segment_checks(const struct iovec *iov,
1284                         unsigned long *nr_segs, size_t *count, int access_flags)
1285 {
1286         unsigned long   seg;
1287         size_t cnt = 0;
1288         for (seg = 0; seg < *nr_segs; seg++) {
1289                 const struct iovec *iv = &iov[seg];
1290
1291                 /*
1292                  * If any segment has a negative length, or the cumulative
1293                  * length ever wraps negative then return -EINVAL.
1294                  */
1295                 cnt += iv->iov_len;
1296                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1297                         return -EINVAL;
1298                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1299                         continue;
1300                 if (seg == 0)
1301                         return -EFAULT;
1302                 *nr_segs = seg;
1303                 cnt -= iv->iov_len;     /* This segment is no good */
1304                 break;
1305         }
1306         *count = cnt;
1307         return 0;
1308 }
1309 EXPORT_SYMBOL(generic_segment_checks);
1310
1311 /**
1312  * generic_file_aio_read - generic filesystem read routine
1313  * @iocb:       kernel I/O control block
1314  * @iov:        io vector request
1315  * @nr_segs:    number of segments in the iovec
1316  * @pos:        current file position
1317  *
1318  * This is the "read()" routine for all filesystems
1319  * that can use the page cache directly.
1320  */
1321 ssize_t
1322 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1323                 unsigned long nr_segs, loff_t pos)
1324 {
1325         struct file *filp = iocb->ki_filp;
1326         ssize_t retval;
1327         unsigned long seg;
1328         size_t count;
1329         loff_t *ppos = &iocb->ki_pos;
1330
1331         count = 0;
1332         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1333         if (retval)
1334                 return retval;
1335
1336         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1337         if (filp->f_flags & O_DIRECT) {
1338                 loff_t size;
1339                 struct address_space *mapping;
1340                 struct inode *inode;
1341
1342                 mapping = filp->f_mapping;
1343                 inode = mapping->host;
1344                 if (!count)
1345                         goto out; /* skip atime */
1346                 size = i_size_read(inode);
1347                 if (pos < size) {
1348                         retval = filemap_write_and_wait_range(mapping, pos,
1349                                         pos + iov_length(iov, nr_segs) - 1);
1350                         if (!retval) {
1351                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1352                                                         iov, pos, nr_segs);
1353                         }
1354                         if (retval > 0)
1355                                 *ppos = pos + retval;
1356                         if (retval) {
1357                                 file_accessed(filp);
1358                                 goto out;
1359                         }
1360                 }
1361         }
1362
1363         for (seg = 0; seg < nr_segs; seg++) {
1364                 read_descriptor_t desc;
1365
1366                 desc.written = 0;
1367                 desc.arg.buf = iov[seg].iov_base;
1368                 desc.count = iov[seg].iov_len;
1369                 if (desc.count == 0)
1370                         continue;
1371                 desc.error = 0;
1372                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1373                 retval += desc.written;
1374                 if (desc.error) {
1375                         retval = retval ?: desc.error;
1376                         break;
1377                 }
1378                 if (desc.count > 0)
1379                         break;
1380         }
1381 out:
1382         return retval;
1383 }
1384 EXPORT_SYMBOL(generic_file_aio_read);
1385
1386 static ssize_t
1387 do_readahead(struct address_space *mapping, struct file *filp,
1388              pgoff_t index, unsigned long nr)
1389 {
1390         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1391                 return -EINVAL;
1392
1393         force_page_cache_readahead(mapping, filp, index, nr);
1394         return 0;
1395 }
1396
1397 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1398 {
1399         ssize_t ret;
1400         struct file *file;
1401
1402         ret = -EBADF;
1403         file = fget(fd);
1404         if (file) {
1405                 if (file->f_mode & FMODE_READ) {
1406                         struct address_space *mapping = file->f_mapping;
1407                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1408                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1409                         unsigned long len = end - start + 1;
1410                         ret = do_readahead(mapping, file, start, len);
1411                 }
1412                 fput(file);
1413         }
1414         return ret;
1415 }
1416 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1417 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1418 {
1419         return SYSC_readahead((int) fd, offset, (size_t) count);
1420 }
1421 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1422 #endif
1423
1424 #ifdef CONFIG_MMU
1425 /**
1426  * page_cache_read - adds requested page to the page cache if not already there
1427  * @file:       file to read
1428  * @offset:     page index
1429  *
1430  * This adds the requested page to the page cache if it isn't already there,
1431  * and schedules an I/O to read in its contents from disk.
1432  */
1433 static int page_cache_read(struct file *file, pgoff_t offset)
1434 {
1435         struct address_space *mapping = file->f_mapping;
1436         struct page *page; 
1437         int ret;
1438
1439         do {
1440                 page = page_cache_alloc_cold(mapping);
1441                 if (!page)
1442                         return -ENOMEM;
1443
1444                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1445                 if (ret == 0)
1446                         ret = mapping->a_ops->readpage(file, page);
1447                 else if (ret == -EEXIST)
1448                         ret = 0; /* losing race to add is OK */
1449
1450                 page_cache_release(page);
1451
1452         } while (ret == AOP_TRUNCATED_PAGE);
1453                 
1454         return ret;
1455 }
1456
1457 #define MMAP_LOTSAMISS  (100)
1458
1459 /*
1460  * Synchronous readahead happens when we don't even find
1461  * a page in the page cache at all.
1462  */
1463 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1464                                    struct file_ra_state *ra,
1465                                    struct file *file,
1466                                    pgoff_t offset)
1467 {
1468         unsigned long ra_pages;
1469         struct address_space *mapping = file->f_mapping;
1470
1471         /* If we don't want any read-ahead, don't bother */
1472         if (VM_RandomReadHint(vma))
1473                 return;
1474
1475         if (VM_SequentialReadHint(vma)) {
1476                 page_cache_sync_readahead(mapping, ra, file, offset, 1);
1477                 return;
1478         }
1479
1480         if (ra->mmap_miss < INT_MAX)
1481                 ra->mmap_miss++;
1482
1483         /*
1484          * Do we miss much more than hit in this file? If so,
1485          * stop bothering with read-ahead. It will only hurt.
1486          */
1487         if (ra->mmap_miss > MMAP_LOTSAMISS)
1488                 return;
1489
1490         ra_pages = max_sane_readahead(ra->ra_pages);
1491         if (ra_pages) {
1492                 pgoff_t start = 0;
1493
1494                 if (offset > ra_pages / 2)
1495                         start = offset - ra_pages / 2;
1496                 do_page_cache_readahead(mapping, file, start, ra_pages);
1497         }
1498 }
1499
1500 /*
1501  * Asynchronous readahead happens when we find the page and PG_readahead,
1502  * so we want to possibly extend the readahead further..
1503  */
1504 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1505                                     struct file_ra_state *ra,
1506                                     struct file *file,
1507                                     struct page *page,
1508                                     pgoff_t offset)
1509 {
1510         struct address_space *mapping = file->f_mapping;
1511
1512         /* If we don't want any read-ahead, don't bother */
1513         if (VM_RandomReadHint(vma))
1514                 return;
1515         if (ra->mmap_miss > 0)
1516                 ra->mmap_miss--;
1517         if (PageReadahead(page))
1518                 page_cache_async_readahead(mapping, ra, file, page, offset, 1);
1519 }
1520
1521 /**
1522  * filemap_fault - read in file data for page fault handling
1523  * @vma:        vma in which the fault was taken
1524  * @vmf:        struct vm_fault containing details of the fault
1525  *
1526  * filemap_fault() is invoked via the vma operations vector for a
1527  * mapped memory region to read in file data during a page fault.
1528  *
1529  * The goto's are kind of ugly, but this streamlines the normal case of having
1530  * it in the page cache, and handles the special cases reasonably without
1531  * having a lot of duplicated code.
1532  */
1533 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1534 {
1535         int error;
1536         struct file *file = vma->vm_file;
1537         struct address_space *mapping = file->f_mapping;
1538         struct file_ra_state *ra = &file->f_ra;
1539         struct inode *inode = mapping->host;
1540         pgoff_t offset = vmf->pgoff;
1541         struct page *page;
1542         pgoff_t size;
1543         int ret = 0;
1544
1545         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1546         if (offset >= size)
1547                 return VM_FAULT_SIGBUS;
1548
1549         /*
1550          * Do we have something in the page cache already?
1551          */
1552         page = find_get_page(mapping, offset);
1553         if (likely(page)) {
1554                 /*
1555                  * We found the page, so try async readahead before
1556                  * waiting for the lock.
1557                  */
1558                 do_async_mmap_readahead(vma, ra, file, page, offset);
1559                 lock_page(page);
1560
1561                 /* Did it get truncated? */
1562                 if (unlikely(page->mapping != mapping)) {
1563                         unlock_page(page);
1564                         put_page(page);
1565                         goto no_cached_page;
1566                 }
1567         } else {
1568                 /* No page in the page cache at all */
1569                 do_sync_mmap_readahead(vma, ra, file, offset);
1570                 count_vm_event(PGMAJFAULT);
1571                 ret = VM_FAULT_MAJOR;
1572 retry_find:
1573                 page = find_lock_page(mapping, offset);
1574                 if (!page)
1575                         goto no_cached_page;
1576         }
1577
1578         /*
1579          * We have a locked page in the page cache, now we need to check
1580          * that it's up-to-date. If not, it is going to be due to an error.
1581          */
1582         if (unlikely(!PageUptodate(page)))
1583                 goto page_not_uptodate;
1584
1585         /*
1586          * Found the page and have a reference on it.
1587          * We must recheck i_size under page lock.
1588          */
1589         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1590         if (unlikely(offset >= size)) {
1591                 unlock_page(page);
1592                 page_cache_release(page);
1593                 return VM_FAULT_SIGBUS;
1594         }
1595
1596         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1597         vmf->page = page;
1598         return ret | VM_FAULT_LOCKED;
1599
1600 no_cached_page:
1601         /*
1602          * We're only likely to ever get here if MADV_RANDOM is in
1603          * effect.
1604          */
1605         error = page_cache_read(file, offset);
1606
1607         /*
1608          * The page we want has now been added to the page cache.
1609          * In the unlikely event that someone removed it in the
1610          * meantime, we'll just come back here and read it again.
1611          */
1612         if (error >= 0)
1613                 goto retry_find;
1614
1615         /*
1616          * An error return from page_cache_read can result if the
1617          * system is low on memory, or a problem occurs while trying
1618          * to schedule I/O.
1619          */
1620         if (error == -ENOMEM)
1621                 return VM_FAULT_OOM;
1622         return VM_FAULT_SIGBUS;
1623
1624 page_not_uptodate:
1625         /*
1626          * Umm, take care of errors if the page isn't up-to-date.
1627          * Try to re-read it _once_. We do this synchronously,
1628          * because there really aren't any performance issues here
1629          * and we need to check for errors.
1630          */
1631         ClearPageError(page);
1632         error = mapping->a_ops->readpage(file, page);
1633         if (!error) {
1634                 wait_on_page_locked(page);
1635                 if (!PageUptodate(page))
1636                         error = -EIO;
1637         }
1638         page_cache_release(page);
1639
1640         if (!error || error == AOP_TRUNCATED_PAGE)
1641                 goto retry_find;
1642
1643         /* Things didn't work out. Return zero to tell the mm layer so. */
1644         shrink_readahead_size_eio(file, ra);
1645         return VM_FAULT_SIGBUS;
1646 }
1647 EXPORT_SYMBOL(filemap_fault);
1648
1649 struct vm_operations_struct generic_file_vm_ops = {
1650         .fault          = filemap_fault,
1651 };
1652
1653 /* This is used for a general mmap of a disk file */
1654
1655 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1656 {
1657         struct address_space *mapping = file->f_mapping;
1658
1659         if (!mapping->a_ops->readpage)
1660                 return -ENOEXEC;
1661         file_accessed(file);
1662         vma->vm_ops = &generic_file_vm_ops;
1663         vma->vm_flags |= VM_CAN_NONLINEAR;
1664         return 0;
1665 }
1666
1667 /*
1668  * This is for filesystems which do not implement ->writepage.
1669  */
1670 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1671 {
1672         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1673                 return -EINVAL;
1674         return generic_file_mmap(file, vma);
1675 }
1676 #else
1677 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1678 {
1679         return -ENOSYS;
1680 }
1681 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1682 {
1683         return -ENOSYS;
1684 }
1685 #endif /* CONFIG_MMU */
1686
1687 EXPORT_SYMBOL(generic_file_mmap);
1688 EXPORT_SYMBOL(generic_file_readonly_mmap);
1689
1690 static struct page *__read_cache_page(struct address_space *mapping,
1691                                 pgoff_t index,
1692                                 int (*filler)(void *,struct page*),
1693                                 void *data)
1694 {
1695         struct page *page;
1696         int err;
1697 repeat:
1698         page = find_get_page(mapping, index);
1699         if (!page) {
1700                 page = page_cache_alloc_cold(mapping);
1701                 if (!page)
1702                         return ERR_PTR(-ENOMEM);
1703                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1704                 if (unlikely(err)) {
1705                         page_cache_release(page);
1706                         if (err == -EEXIST)
1707                                 goto repeat;
1708                         /* Presumably ENOMEM for radix tree node */
1709                         return ERR_PTR(err);
1710                 }
1711                 err = filler(data, page);
1712                 if (err < 0) {
1713                         page_cache_release(page);
1714                         page = ERR_PTR(err);
1715                 }
1716         }
1717         return page;
1718 }
1719
1720 /**
1721  * read_cache_page_async - read into page cache, fill it if needed
1722  * @mapping:    the page's address_space
1723  * @index:      the page index
1724  * @filler:     function to perform the read
1725  * @data:       destination for read data
1726  *
1727  * Same as read_cache_page, but don't wait for page to become unlocked
1728  * after submitting it to the filler.
1729  *
1730  * Read into the page cache. If a page already exists, and PageUptodate() is
1731  * not set, try to fill the page but don't wait for it to become unlocked.
1732  *
1733  * If the page does not get brought uptodate, return -EIO.
1734  */
1735 struct page *read_cache_page_async(struct address_space *mapping,
1736                                 pgoff_t index,
1737                                 int (*filler)(void *,struct page*),
1738                                 void *data)
1739 {
1740         struct page *page;
1741         int err;
1742
1743 retry:
1744         page = __read_cache_page(mapping, index, filler, data);
1745         if (IS_ERR(page))
1746                 return page;
1747         if (PageUptodate(page))
1748                 goto out;
1749
1750         lock_page(page);
1751         if (!page->mapping) {
1752                 unlock_page(page);
1753                 page_cache_release(page);
1754                 goto retry;
1755         }
1756         if (PageUptodate(page)) {
1757                 unlock_page(page);
1758                 goto out;
1759         }
1760         err = filler(data, page);
1761         if (err < 0) {
1762                 page_cache_release(page);
1763                 return ERR_PTR(err);
1764         }
1765 out:
1766         mark_page_accessed(page);
1767         return page;
1768 }
1769 EXPORT_SYMBOL(read_cache_page_async);
1770
1771 /**
1772  * read_cache_page - read into page cache, fill it if needed
1773  * @mapping:    the page's address_space
1774  * @index:      the page index
1775  * @filler:     function to perform the read
1776  * @data:       destination for read data
1777  *
1778  * Read into the page cache. If a page already exists, and PageUptodate() is
1779  * not set, try to fill the page then wait for it to become unlocked.
1780  *
1781  * If the page does not get brought uptodate, return -EIO.
1782  */
1783 struct page *read_cache_page(struct address_space *mapping,
1784                                 pgoff_t index,
1785                                 int (*filler)(void *,struct page*),
1786                                 void *data)
1787 {
1788         struct page *page;
1789
1790         page = read_cache_page_async(mapping, index, filler, data);
1791         if (IS_ERR(page))
1792                 goto out;
1793         wait_on_page_locked(page);
1794         if (!PageUptodate(page)) {
1795                 page_cache_release(page);
1796                 page = ERR_PTR(-EIO);
1797         }
1798  out:
1799         return page;
1800 }
1801 EXPORT_SYMBOL(read_cache_page);
1802
1803 /*
1804  * The logic we want is
1805  *
1806  *      if suid or (sgid and xgrp)
1807  *              remove privs
1808  */
1809 int should_remove_suid(struct dentry *dentry)
1810 {
1811         mode_t mode = dentry->d_inode->i_mode;
1812         int kill = 0;
1813
1814         /* suid always must be killed */
1815         if (unlikely(mode & S_ISUID))
1816                 kill = ATTR_KILL_SUID;
1817
1818         /*
1819          * sgid without any exec bits is just a mandatory locking mark; leave
1820          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1821          */
1822         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1823                 kill |= ATTR_KILL_SGID;
1824
1825         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1826                 return kill;
1827
1828         return 0;
1829 }
1830 EXPORT_SYMBOL(should_remove_suid);
1831
1832 static int __remove_suid(struct dentry *dentry, int kill)
1833 {
1834         struct iattr newattrs;
1835
1836         newattrs.ia_valid = ATTR_FORCE | kill;
1837         return notify_change(dentry, &newattrs);
1838 }
1839
1840 int file_remove_suid(struct file *file)
1841 {
1842         struct dentry *dentry = file->f_path.dentry;
1843         int killsuid = should_remove_suid(dentry);
1844         int killpriv = security_inode_need_killpriv(dentry);
1845         int error = 0;
1846
1847         if (killpriv < 0)
1848                 return killpriv;
1849         if (killpriv)
1850                 error = security_inode_killpriv(dentry);
1851         if (!error && killsuid)
1852                 error = __remove_suid(dentry, killsuid);
1853
1854         return error;
1855 }
1856 EXPORT_SYMBOL(file_remove_suid);
1857
1858 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1859                         const struct iovec *iov, size_t base, size_t bytes)
1860 {
1861         size_t copied = 0, left = 0;
1862
1863         while (bytes) {
1864                 char __user *buf = iov->iov_base + base;
1865                 int copy = min(bytes, iov->iov_len - base);
1866
1867                 base = 0;
1868                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1869                 copied += copy;
1870                 bytes -= copy;
1871                 vaddr += copy;
1872                 iov++;
1873
1874                 if (unlikely(left))
1875                         break;
1876         }
1877         return copied - left;
1878 }
1879
1880 /*
1881  * Copy as much as we can into the page and return the number of bytes which
1882  * were sucessfully copied.  If a fault is encountered then return the number of
1883  * bytes which were copied.
1884  */
1885 size_t iov_iter_copy_from_user_atomic(struct page *page,
1886                 struct iov_iter *i, unsigned long offset, size_t bytes)
1887 {
1888         char *kaddr;
1889         size_t copied;
1890
1891         BUG_ON(!in_atomic());
1892         kaddr = kmap_atomic(page, KM_USER0);
1893         if (likely(i->nr_segs == 1)) {
1894                 int left;
1895                 char __user *buf = i->iov->iov_base + i->iov_offset;
1896                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1897                 copied = bytes - left;
1898         } else {
1899                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1900                                                 i->iov, i->iov_offset, bytes);
1901         }
1902         kunmap_atomic(kaddr, KM_USER0);
1903
1904         return copied;
1905 }
1906 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1907
1908 /*
1909  * This has the same sideeffects and return value as
1910  * iov_iter_copy_from_user_atomic().
1911  * The difference is that it attempts to resolve faults.
1912  * Page must not be locked.
1913  */
1914 size_t iov_iter_copy_from_user(struct page *page,
1915                 struct iov_iter *i, unsigned long offset, size_t bytes)
1916 {
1917         char *kaddr;
1918         size_t copied;
1919
1920         kaddr = kmap(page);
1921         if (likely(i->nr_segs == 1)) {
1922                 int left;
1923                 char __user *buf = i->iov->iov_base + i->iov_offset;
1924                 left = __copy_from_user(kaddr + offset, buf, bytes);
1925                 copied = bytes - left;
1926         } else {
1927                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1928                                                 i->iov, i->iov_offset, bytes);
1929         }
1930         kunmap(page);
1931         return copied;
1932 }
1933 EXPORT_SYMBOL(iov_iter_copy_from_user);
1934
1935 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1936 {
1937         BUG_ON(i->count < bytes);
1938
1939         if (likely(i->nr_segs == 1)) {
1940                 i->iov_offset += bytes;
1941                 i->count -= bytes;
1942         } else {
1943                 const struct iovec *iov = i->iov;
1944                 size_t base = i->iov_offset;
1945
1946                 /*
1947                  * The !iov->iov_len check ensures we skip over unlikely
1948                  * zero-length segments (without overruning the iovec).
1949                  */
1950                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1951                         int copy;
1952
1953                         copy = min(bytes, iov->iov_len - base);
1954                         BUG_ON(!i->count || i->count < copy);
1955                         i->count -= copy;
1956                         bytes -= copy;
1957                         base += copy;
1958                         if (iov->iov_len == base) {
1959                                 iov++;
1960                                 base = 0;
1961                         }
1962                 }
1963                 i->iov = iov;
1964                 i->iov_offset = base;
1965         }
1966 }
1967 EXPORT_SYMBOL(iov_iter_advance);
1968
1969 /*
1970  * Fault in the first iovec of the given iov_iter, to a maximum length
1971  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1972  * accessed (ie. because it is an invalid address).
1973  *
1974  * writev-intensive code may want this to prefault several iovecs -- that
1975  * would be possible (callers must not rely on the fact that _only_ the
1976  * first iovec will be faulted with the current implementation).
1977  */
1978 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1979 {
1980         char __user *buf = i->iov->iov_base + i->iov_offset;
1981         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1982         return fault_in_pages_readable(buf, bytes);
1983 }
1984 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1985
1986 /*
1987  * Return the count of just the current iov_iter segment.
1988  */
1989 size_t iov_iter_single_seg_count(struct iov_iter *i)
1990 {
1991         const struct iovec *iov = i->iov;
1992         if (i->nr_segs == 1)
1993                 return i->count;
1994         else
1995                 return min(i->count, iov->iov_len - i->iov_offset);
1996 }
1997 EXPORT_SYMBOL(iov_iter_single_seg_count);
1998
1999 /*
2000  * Performs necessary checks before doing a write
2001  *
2002  * Can adjust writing position or amount of bytes to write.
2003  * Returns appropriate error code that caller should return or
2004  * zero in case that write should be allowed.
2005  */
2006 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2007 {
2008         struct inode *inode = file->f_mapping->host;
2009         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2010
2011         if (unlikely(*pos < 0))
2012                 return -EINVAL;
2013
2014         if (!isblk) {
2015                 /* FIXME: this is for backwards compatibility with 2.4 */
2016                 if (file->f_flags & O_APPEND)
2017                         *pos = i_size_read(inode);
2018
2019                 if (limit != RLIM_INFINITY) {
2020                         if (*pos >= limit) {
2021                                 send_sig(SIGXFSZ, current, 0);
2022                                 return -EFBIG;
2023                         }
2024                         if (*count > limit - (typeof(limit))*pos) {
2025                                 *count = limit - (typeof(limit))*pos;
2026                         }
2027                 }
2028         }
2029
2030         /*
2031          * LFS rule
2032          */
2033         if (unlikely(*pos + *count > MAX_NON_LFS &&
2034                                 !(file->f_flags & O_LARGEFILE))) {
2035                 if (*pos >= MAX_NON_LFS) {
2036                         return -EFBIG;
2037                 }
2038                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2039                         *count = MAX_NON_LFS - (unsigned long)*pos;
2040                 }
2041         }
2042
2043         /*
2044          * Are we about to exceed the fs block limit ?
2045          *
2046          * If we have written data it becomes a short write.  If we have
2047          * exceeded without writing data we send a signal and return EFBIG.
2048          * Linus frestrict idea will clean these up nicely..
2049          */
2050         if (likely(!isblk)) {
2051                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2052                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2053                                 return -EFBIG;
2054                         }
2055                         /* zero-length writes at ->s_maxbytes are OK */
2056                 }
2057
2058                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2059                         *count = inode->i_sb->s_maxbytes - *pos;
2060         } else {
2061 #ifdef CONFIG_BLOCK
2062                 loff_t isize;
2063                 if (bdev_read_only(I_BDEV(inode)))
2064                         return -EPERM;
2065                 isize = i_size_read(inode);
2066                 if (*pos >= isize) {
2067                         if (*count || *pos > isize)
2068                                 return -ENOSPC;
2069                 }
2070
2071                 if (*pos + *count > isize)
2072                         *count = isize - *pos;
2073 #else
2074                 return -EPERM;
2075 #endif
2076         }
2077         return 0;
2078 }
2079 EXPORT_SYMBOL(generic_write_checks);
2080
2081 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2082                                 loff_t pos, unsigned len, unsigned flags,
2083                                 struct page **pagep, void **fsdata)
2084 {
2085         const struct address_space_operations *aops = mapping->a_ops;
2086
2087         return aops->write_begin(file, mapping, pos, len, flags,
2088                                                         pagep, fsdata);
2089 }
2090 EXPORT_SYMBOL(pagecache_write_begin);
2091
2092 int pagecache_write_end(struct file *file, struct address_space *mapping,
2093                                 loff_t pos, unsigned len, unsigned copied,
2094                                 struct page *page, void *fsdata)
2095 {
2096         const struct address_space_operations *aops = mapping->a_ops;
2097
2098         mark_page_accessed(page);
2099         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2100 }
2101 EXPORT_SYMBOL(pagecache_write_end);
2102
2103 ssize_t
2104 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2105                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2106                 size_t count, size_t ocount)
2107 {
2108         struct file     *file = iocb->ki_filp;
2109         struct address_space *mapping = file->f_mapping;
2110         struct inode    *inode = mapping->host;
2111         ssize_t         written;
2112         size_t          write_len;
2113         pgoff_t         end;
2114
2115         if (count != ocount)
2116                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2117
2118         write_len = iov_length(iov, *nr_segs);
2119         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2120
2121         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2122         if (written)
2123                 goto out;
2124
2125         /*
2126          * After a write we want buffered reads to be sure to go to disk to get
2127          * the new data.  We invalidate clean cached page from the region we're
2128          * about to write.  We do this *before* the write so that we can return
2129          * without clobbering -EIOCBQUEUED from ->direct_IO().
2130          */
2131         if (mapping->nrpages) {
2132                 written = invalidate_inode_pages2_range(mapping,
2133                                         pos >> PAGE_CACHE_SHIFT, end);
2134                 /*
2135                  * If a page can not be invalidated, return 0 to fall back
2136                  * to buffered write.
2137                  */
2138                 if (written) {
2139                         if (written == -EBUSY)
2140                                 return 0;
2141                         goto out;
2142                 }
2143         }
2144
2145         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2146
2147         /*
2148          * Finally, try again to invalidate clean pages which might have been
2149          * cached by non-direct readahead, or faulted in by get_user_pages()
2150          * if the source of the write was an mmap'ed region of the file
2151          * we're writing.  Either one is a pretty crazy thing to do,
2152          * so we don't support it 100%.  If this invalidation
2153          * fails, tough, the write still worked...
2154          */
2155         if (mapping->nrpages) {
2156                 invalidate_inode_pages2_range(mapping,
2157                                               pos >> PAGE_CACHE_SHIFT, end);
2158         }
2159
2160         if (written > 0) {
2161                 loff_t end = pos + written;
2162                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2163                         i_size_write(inode,  end);
2164                         mark_inode_dirty(inode);
2165                 }
2166                 *ppos = end;
2167         }
2168
2169         /*
2170          * Sync the fs metadata but not the minor inode changes and
2171          * of course not the data as we did direct DMA for the IO.
2172          * i_mutex is held, which protects generic_osync_inode() from
2173          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2174          */
2175 out:
2176         if ((written >= 0 || written == -EIOCBQUEUED) &&
2177             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2178                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2179                 if (err < 0)
2180                         written = err;
2181         }
2182         return written;
2183 }
2184 EXPORT_SYMBOL(generic_file_direct_write);
2185
2186 /*
2187  * Find or create a page at the given pagecache position. Return the locked
2188  * page. This function is specifically for buffered writes.
2189  */
2190 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2191                                         pgoff_t index, unsigned flags)
2192 {
2193         int status;
2194         struct page *page;
2195         gfp_t gfp_notmask = 0;
2196         if (flags & AOP_FLAG_NOFS)
2197                 gfp_notmask = __GFP_FS;
2198 repeat:
2199         page = find_lock_page(mapping, index);
2200         if (likely(page))
2201                 return page;
2202
2203         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2204         if (!page)
2205                 return NULL;
2206         status = add_to_page_cache_lru(page, mapping, index,
2207                                                 GFP_KERNEL & ~gfp_notmask);
2208         if (unlikely(status)) {
2209                 page_cache_release(page);
2210                 if (status == -EEXIST)
2211                         goto repeat;
2212                 return NULL;
2213         }
2214         return page;
2215 }
2216 EXPORT_SYMBOL(grab_cache_page_write_begin);
2217
2218 static ssize_t generic_perform_write(struct file *file,
2219                                 struct iov_iter *i, loff_t pos)
2220 {
2221         struct address_space *mapping = file->f_mapping;
2222         const struct address_space_operations *a_ops = mapping->a_ops;
2223         long status = 0;
2224         ssize_t written = 0;
2225         unsigned int flags = 0;
2226
2227         /*
2228          * Copies from kernel address space cannot fail (NFSD is a big user).
2229          */
2230         if (segment_eq(get_fs(), KERNEL_DS))
2231                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2232
2233         do {
2234                 struct page *page;
2235                 pgoff_t index;          /* Pagecache index for current page */
2236                 unsigned long offset;   /* Offset into pagecache page */
2237                 unsigned long bytes;    /* Bytes to write to page */
2238                 size_t copied;          /* Bytes copied from user */
2239                 void *fsdata;
2240
2241                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2242                 index = pos >> PAGE_CACHE_SHIFT;
2243                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2244                                                 iov_iter_count(i));
2245
2246 again:
2247
2248                 /*
2249                  * Bring in the user page that we will copy from _first_.
2250                  * Otherwise there's a nasty deadlock on copying from the
2251                  * same page as we're writing to, without it being marked
2252                  * up-to-date.
2253                  *
2254                  * Not only is this an optimisation, but it is also required
2255                  * to check that the address is actually valid, when atomic
2256                  * usercopies are used, below.
2257                  */
2258                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2259                         status = -EFAULT;
2260                         break;
2261                 }
2262
2263                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2264                                                 &page, &fsdata);
2265                 if (unlikely(status))
2266                         break;
2267
2268                 pagefault_disable();
2269                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2270                 pagefault_enable();
2271                 flush_dcache_page(page);
2272
2273                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2274                                                 page, fsdata);
2275                 if (unlikely(status < 0))
2276                         break;
2277                 copied = status;
2278
2279                 cond_resched();
2280
2281                 iov_iter_advance(i, copied);
2282                 if (unlikely(copied == 0)) {
2283                         /*
2284                          * If we were unable to copy any data at all, we must
2285                          * fall back to a single segment length write.
2286                          *
2287                          * If we didn't fallback here, we could livelock
2288                          * because not all segments in the iov can be copied at
2289                          * once without a pagefault.
2290                          */
2291                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2292                                                 iov_iter_single_seg_count(i));
2293                         goto again;
2294                 }
2295                 pos += copied;
2296                 written += copied;
2297
2298                 balance_dirty_pages_ratelimited(mapping);
2299
2300         } while (iov_iter_count(i));
2301
2302         return written ? written : status;
2303 }
2304
2305 ssize_t
2306 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2307                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2308                 size_t count, ssize_t written)
2309 {
2310         struct file *file = iocb->ki_filp;
2311         struct address_space *mapping = file->f_mapping;
2312         const struct address_space_operations *a_ops = mapping->a_ops;
2313         struct inode *inode = mapping->host;
2314         ssize_t status;
2315         struct iov_iter i;
2316
2317         iov_iter_init(&i, iov, nr_segs, count, written);
2318         status = generic_perform_write(file, &i, pos);
2319
2320         if (likely(status >= 0)) {
2321                 written += status;
2322                 *ppos = pos + status;
2323
2324                 /*
2325                  * For now, when the user asks for O_SYNC, we'll actually give
2326                  * O_DSYNC
2327                  */
2328                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2329                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2330                                 status = generic_osync_inode(inode, mapping,
2331                                                 OSYNC_METADATA|OSYNC_DATA);
2332                 }
2333         }
2334         
2335         /*
2336          * If we get here for O_DIRECT writes then we must have fallen through
2337          * to buffered writes (block instantiation inside i_size).  So we sync
2338          * the file data here, to try to honour O_DIRECT expectations.
2339          */
2340         if (unlikely(file->f_flags & O_DIRECT) && written)
2341                 status = filemap_write_and_wait_range(mapping,
2342                                         pos, pos + written - 1);
2343
2344         return written ? written : status;
2345 }
2346 EXPORT_SYMBOL(generic_file_buffered_write);
2347
2348 static ssize_t
2349 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2350                                 unsigned long nr_segs, loff_t *ppos)
2351 {
2352         struct file *file = iocb->ki_filp;
2353         struct address_space * mapping = file->f_mapping;
2354         size_t ocount;          /* original count */
2355         size_t count;           /* after file limit checks */
2356         struct inode    *inode = mapping->host;
2357         loff_t          pos;
2358         ssize_t         written;
2359         ssize_t         err;
2360
2361         ocount = 0;
2362         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2363         if (err)
2364                 return err;
2365
2366         count = ocount;
2367         pos = *ppos;
2368
2369         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2370
2371         /* We can write back this queue in page reclaim */
2372         current->backing_dev_info = mapping->backing_dev_info;
2373         written = 0;
2374
2375         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2376         if (err)
2377                 goto out;
2378
2379         if (count == 0)
2380                 goto out;
2381
2382         err = file_remove_suid(file);
2383         if (err)
2384                 goto out;
2385
2386         file_update_time(file);
2387
2388         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2389         if (unlikely(file->f_flags & O_DIRECT)) {
2390                 loff_t endbyte;
2391                 ssize_t written_buffered;
2392
2393                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2394                                                         ppos, count, ocount);
2395                 if (written < 0 || written == count)
2396                         goto out;
2397                 /*
2398                  * direct-io write to a hole: fall through to buffered I/O
2399                  * for completing the rest of the request.
2400                  */
2401                 pos += written;
2402                 count -= written;
2403                 written_buffered = generic_file_buffered_write(iocb, iov,
2404                                                 nr_segs, pos, ppos, count,
2405                                                 written);
2406                 /*
2407                  * If generic_file_buffered_write() retuned a synchronous error
2408                  * then we want to return the number of bytes which were
2409                  * direct-written, or the error code if that was zero.  Note
2410                  * that this differs from normal direct-io semantics, which
2411                  * will return -EFOO even if some bytes were written.
2412                  */
2413                 if (written_buffered < 0) {
2414                         err = written_buffered;
2415                         goto out;
2416                 }
2417
2418                 /*
2419                  * We need to ensure that the page cache pages are written to
2420                  * disk and invalidated to preserve the expected O_DIRECT
2421                  * semantics.
2422                  */
2423                 endbyte = pos + written_buffered - written - 1;
2424                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2425                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2426                                             SYNC_FILE_RANGE_WRITE|
2427                                             SYNC_FILE_RANGE_WAIT_AFTER);
2428                 if (err == 0) {
2429                         written = written_buffered;
2430                         invalidate_mapping_pages(mapping,
2431                                                  pos >> PAGE_CACHE_SHIFT,
2432                                                  endbyte >> PAGE_CACHE_SHIFT);
2433                 } else {
2434                         /*
2435                          * We don't know how much we wrote, so just return
2436                          * the number of bytes which were direct-written
2437                          */
2438                 }
2439         } else {
2440                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2441                                 pos, ppos, count, written);
2442         }
2443 out:
2444         current->backing_dev_info = NULL;
2445         return written ? written : err;
2446 }
2447
2448 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2449                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2450 {
2451         struct file *file = iocb->ki_filp;
2452         struct address_space *mapping = file->f_mapping;
2453         struct inode *inode = mapping->host;
2454         ssize_t ret;
2455
2456         BUG_ON(iocb->ki_pos != pos);
2457
2458         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2459                         &iocb->ki_pos);
2460
2461         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2462                 ssize_t err;
2463
2464                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2465                 if (err < 0)
2466                         ret = err;
2467         }
2468         return ret;
2469 }
2470 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2471
2472 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2473                 unsigned long nr_segs, loff_t pos)
2474 {
2475         struct file *file = iocb->ki_filp;
2476         struct address_space *mapping = file->f_mapping;
2477         struct inode *inode = mapping->host;
2478         ssize_t ret;
2479
2480         BUG_ON(iocb->ki_pos != pos);
2481
2482         mutex_lock(&inode->i_mutex);
2483         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2484                         &iocb->ki_pos);
2485         mutex_unlock(&inode->i_mutex);
2486
2487         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2488                 ssize_t err;
2489
2490                 err = sync_page_range(inode, mapping, pos, ret);
2491                 if (err < 0)
2492                         ret = err;
2493         }
2494         return ret;
2495 }
2496 EXPORT_SYMBOL(generic_file_aio_write);
2497
2498 /**
2499  * try_to_release_page() - release old fs-specific metadata on a page
2500  *
2501  * @page: the page which the kernel is trying to free
2502  * @gfp_mask: memory allocation flags (and I/O mode)
2503  *
2504  * The address_space is to try to release any data against the page
2505  * (presumably at page->private).  If the release was successful, return `1'.
2506  * Otherwise return zero.
2507  *
2508  * This may also be called if PG_fscache is set on a page, indicating that the
2509  * page is known to the local caching routines.
2510  *
2511  * The @gfp_mask argument specifies whether I/O may be performed to release
2512  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2513  *
2514  */
2515 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2516 {
2517         struct address_space * const mapping = page->mapping;
2518
2519         BUG_ON(!PageLocked(page));
2520         if (PageWriteback(page))
2521                 return 0;
2522
2523         if (mapping && mapping->a_ops->releasepage)
2524                 return mapping->a_ops->releasepage(page, gfp_mask);
2525         return try_to_free_buffers(page);
2526 }
2527
2528 EXPORT_SYMBOL(try_to_release_page);