page allocator: do not check NUMA node ID when the caller knows the node is valid
[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_exact_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         ra->ra_pages /= 4;
1008 }
1009
1010 /**
1011  * do_generic_file_read - generic file read routine
1012  * @filp:       the file to read
1013  * @ppos:       current file position
1014  * @desc:       read_descriptor
1015  * @actor:      read method
1016  *
1017  * This is a generic file read routine, and uses the
1018  * mapping->a_ops->readpage() function for the actual low-level stuff.
1019  *
1020  * This is really ugly. But the goto's actually try to clarify some
1021  * of the logic when it comes to error handling etc.
1022  */
1023 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1024                 read_descriptor_t *desc, read_actor_t actor)
1025 {
1026         struct address_space *mapping = filp->f_mapping;
1027         struct inode *inode = mapping->host;
1028         struct file_ra_state *ra = &filp->f_ra;
1029         pgoff_t index;
1030         pgoff_t last_index;
1031         pgoff_t prev_index;
1032         unsigned long offset;      /* offset into pagecache page */
1033         unsigned int prev_offset;
1034         int error;
1035
1036         index = *ppos >> PAGE_CACHE_SHIFT;
1037         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1038         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1039         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1040         offset = *ppos & ~PAGE_CACHE_MASK;
1041
1042         for (;;) {
1043                 struct page *page;
1044                 pgoff_t end_index;
1045                 loff_t isize;
1046                 unsigned long nr, ret;
1047
1048                 cond_resched();
1049 find_page:
1050                 page = find_get_page(mapping, index);
1051                 if (!page) {
1052                         page_cache_sync_readahead(mapping,
1053                                         ra, filp,
1054                                         index, last_index - index);
1055                         page = find_get_page(mapping, index);
1056                         if (unlikely(page == NULL))
1057                                 goto no_cached_page;
1058                 }
1059                 if (PageReadahead(page)) {
1060                         page_cache_async_readahead(mapping,
1061                                         ra, filp, page,
1062                                         index, last_index - index);
1063                 }
1064                 if (!PageUptodate(page)) {
1065                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1066                                         !mapping->a_ops->is_partially_uptodate)
1067                                 goto page_not_up_to_date;
1068                         if (!trylock_page(page))
1069                                 goto page_not_up_to_date;
1070                         if (!mapping->a_ops->is_partially_uptodate(page,
1071                                                                 desc, offset))
1072                                 goto page_not_up_to_date_locked;
1073                         unlock_page(page);
1074                 }
1075 page_ok:
1076                 /*
1077                  * i_size must be checked after we know the page is Uptodate.
1078                  *
1079                  * Checking i_size after the check allows us to calculate
1080                  * the correct value for "nr", which means the zero-filled
1081                  * part of the page is not copied back to userspace (unless
1082                  * another truncate extends the file - this is desired though).
1083                  */
1084
1085                 isize = i_size_read(inode);
1086                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1087                 if (unlikely(!isize || index > end_index)) {
1088                         page_cache_release(page);
1089                         goto out;
1090                 }
1091
1092                 /* nr is the maximum number of bytes to copy from this page */
1093                 nr = PAGE_CACHE_SIZE;
1094                 if (index == end_index) {
1095                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1096                         if (nr <= offset) {
1097                                 page_cache_release(page);
1098                                 goto out;
1099                         }
1100                 }
1101                 nr = nr - offset;
1102
1103                 /* If users can be writing to this page using arbitrary
1104                  * virtual addresses, take care about potential aliasing
1105                  * before reading the page on the kernel side.
1106                  */
1107                 if (mapping_writably_mapped(mapping))
1108                         flush_dcache_page(page);
1109
1110                 /*
1111                  * When a sequential read accesses a page several times,
1112                  * only mark it as accessed the first time.
1113                  */
1114                 if (prev_index != index || offset != prev_offset)
1115                         mark_page_accessed(page);
1116                 prev_index = index;
1117
1118                 /*
1119                  * Ok, we have the page, and it's up-to-date, so
1120                  * now we can copy it to user space...
1121                  *
1122                  * The actor routine returns how many bytes were actually used..
1123                  * NOTE! This may not be the same as how much of a user buffer
1124                  * we filled up (we may be padding etc), so we can only update
1125                  * "pos" here (the actor routine has to update the user buffer
1126                  * pointers and the remaining count).
1127                  */
1128                 ret = actor(desc, page, offset, nr);
1129                 offset += ret;
1130                 index += offset >> PAGE_CACHE_SHIFT;
1131                 offset &= ~PAGE_CACHE_MASK;
1132                 prev_offset = offset;
1133
1134                 page_cache_release(page);
1135                 if (ret == nr && desc->count)
1136                         continue;
1137                 goto out;
1138
1139 page_not_up_to_date:
1140                 /* Get exclusive access to the page ... */
1141                 error = lock_page_killable(page);
1142                 if (unlikely(error))
1143                         goto readpage_error;
1144
1145 page_not_up_to_date_locked:
1146                 /* Did it get truncated before we got the lock? */
1147                 if (!page->mapping) {
1148                         unlock_page(page);
1149                         page_cache_release(page);
1150                         continue;
1151                 }
1152
1153                 /* Did somebody else fill it already? */
1154                 if (PageUptodate(page)) {
1155                         unlock_page(page);
1156                         goto page_ok;
1157                 }
1158
1159 readpage:
1160                 /* Start the actual read. The read will unlock the page. */
1161                 error = mapping->a_ops->readpage(filp, page);
1162
1163                 if (unlikely(error)) {
1164                         if (error == AOP_TRUNCATED_PAGE) {
1165                                 page_cache_release(page);
1166                                 goto find_page;
1167                         }
1168                         goto readpage_error;
1169                 }
1170
1171                 if (!PageUptodate(page)) {
1172                         error = lock_page_killable(page);
1173                         if (unlikely(error))
1174                                 goto readpage_error;
1175                         if (!PageUptodate(page)) {
1176                                 if (page->mapping == NULL) {
1177                                         /*
1178                                          * invalidate_inode_pages got it
1179                                          */
1180                                         unlock_page(page);
1181                                         page_cache_release(page);
1182                                         goto find_page;
1183                                 }
1184                                 unlock_page(page);
1185                                 shrink_readahead_size_eio(filp, ra);
1186                                 error = -EIO;
1187                                 goto readpage_error;
1188                         }
1189                         unlock_page(page);
1190                 }
1191
1192                 goto page_ok;
1193
1194 readpage_error:
1195                 /* UHHUH! A synchronous read error occurred. Report it */
1196                 desc->error = error;
1197                 page_cache_release(page);
1198                 goto out;
1199
1200 no_cached_page:
1201                 /*
1202                  * Ok, it wasn't cached, so we need to create a new
1203                  * page..
1204                  */
1205                 page = page_cache_alloc_cold(mapping);
1206                 if (!page) {
1207                         desc->error = -ENOMEM;
1208                         goto out;
1209                 }
1210                 error = add_to_page_cache_lru(page, mapping,
1211                                                 index, GFP_KERNEL);
1212                 if (error) {
1213                         page_cache_release(page);
1214                         if (error == -EEXIST)
1215                                 goto find_page;
1216                         desc->error = error;
1217                         goto out;
1218                 }
1219                 goto readpage;
1220         }
1221
1222 out:
1223         ra->prev_pos = prev_index;
1224         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1225         ra->prev_pos |= prev_offset;
1226
1227         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1228         file_accessed(filp);
1229 }
1230
1231 int file_read_actor(read_descriptor_t *desc, struct page *page,
1232                         unsigned long offset, unsigned long size)
1233 {
1234         char *kaddr;
1235         unsigned long left, count = desc->count;
1236
1237         if (size > count)
1238                 size = count;
1239
1240         /*
1241          * Faults on the destination of a read are common, so do it before
1242          * taking the kmap.
1243          */
1244         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1245                 kaddr = kmap_atomic(page, KM_USER0);
1246                 left = __copy_to_user_inatomic(desc->arg.buf,
1247                                                 kaddr + offset, size);
1248                 kunmap_atomic(kaddr, KM_USER0);
1249                 if (left == 0)
1250                         goto success;
1251         }
1252
1253         /* Do it the slow way */
1254         kaddr = kmap(page);
1255         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1256         kunmap(page);
1257
1258         if (left) {
1259                 size -= left;
1260                 desc->error = -EFAULT;
1261         }
1262 success:
1263         desc->count = count - size;
1264         desc->written += size;
1265         desc->arg.buf += size;
1266         return size;
1267 }
1268
1269 /*
1270  * Performs necessary checks before doing a write
1271  * @iov:        io vector request
1272  * @nr_segs:    number of segments in the iovec
1273  * @count:      number of bytes to write
1274  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1275  *
1276  * Adjust number of segments and amount of bytes to write (nr_segs should be
1277  * properly initialized first). Returns appropriate error code that caller
1278  * should return or zero in case that write should be allowed.
1279  */
1280 int generic_segment_checks(const struct iovec *iov,
1281                         unsigned long *nr_segs, size_t *count, int access_flags)
1282 {
1283         unsigned long   seg;
1284         size_t cnt = 0;
1285         for (seg = 0; seg < *nr_segs; seg++) {
1286                 const struct iovec *iv = &iov[seg];
1287
1288                 /*
1289                  * If any segment has a negative length, or the cumulative
1290                  * length ever wraps negative then return -EINVAL.
1291                  */
1292                 cnt += iv->iov_len;
1293                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1294                         return -EINVAL;
1295                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1296                         continue;
1297                 if (seg == 0)
1298                         return -EFAULT;
1299                 *nr_segs = seg;
1300                 cnt -= iv->iov_len;     /* This segment is no good */
1301                 break;
1302         }
1303         *count = cnt;
1304         return 0;
1305 }
1306 EXPORT_SYMBOL(generic_segment_checks);
1307
1308 /**
1309  * generic_file_aio_read - generic filesystem read routine
1310  * @iocb:       kernel I/O control block
1311  * @iov:        io vector request
1312  * @nr_segs:    number of segments in the iovec
1313  * @pos:        current file position
1314  *
1315  * This is the "read()" routine for all filesystems
1316  * that can use the page cache directly.
1317  */
1318 ssize_t
1319 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1320                 unsigned long nr_segs, loff_t pos)
1321 {
1322         struct file *filp = iocb->ki_filp;
1323         ssize_t retval;
1324         unsigned long seg;
1325         size_t count;
1326         loff_t *ppos = &iocb->ki_pos;
1327
1328         count = 0;
1329         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1330         if (retval)
1331                 return retval;
1332
1333         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1334         if (filp->f_flags & O_DIRECT) {
1335                 loff_t size;
1336                 struct address_space *mapping;
1337                 struct inode *inode;
1338
1339                 mapping = filp->f_mapping;
1340                 inode = mapping->host;
1341                 if (!count)
1342                         goto out; /* skip atime */
1343                 size = i_size_read(inode);
1344                 if (pos < size) {
1345                         retval = filemap_write_and_wait_range(mapping, pos,
1346                                         pos + iov_length(iov, nr_segs) - 1);
1347                         if (!retval) {
1348                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1349                                                         iov, pos, nr_segs);
1350                         }
1351                         if (retval > 0)
1352                                 *ppos = pos + retval;
1353                         if (retval) {
1354                                 file_accessed(filp);
1355                                 goto out;
1356                         }
1357                 }
1358         }
1359
1360         for (seg = 0; seg < nr_segs; seg++) {
1361                 read_descriptor_t desc;
1362
1363                 desc.written = 0;
1364                 desc.arg.buf = iov[seg].iov_base;
1365                 desc.count = iov[seg].iov_len;
1366                 if (desc.count == 0)
1367                         continue;
1368                 desc.error = 0;
1369                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1370                 retval += desc.written;
1371                 if (desc.error) {
1372                         retval = retval ?: desc.error;
1373                         break;
1374                 }
1375                 if (desc.count > 0)
1376                         break;
1377         }
1378 out:
1379         return retval;
1380 }
1381 EXPORT_SYMBOL(generic_file_aio_read);
1382
1383 static ssize_t
1384 do_readahead(struct address_space *mapping, struct file *filp,
1385              pgoff_t index, unsigned long nr)
1386 {
1387         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1388                 return -EINVAL;
1389
1390         force_page_cache_readahead(mapping, filp, index, nr);
1391         return 0;
1392 }
1393
1394 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1395 {
1396         ssize_t ret;
1397         struct file *file;
1398
1399         ret = -EBADF;
1400         file = fget(fd);
1401         if (file) {
1402                 if (file->f_mode & FMODE_READ) {
1403                         struct address_space *mapping = file->f_mapping;
1404                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1405                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1406                         unsigned long len = end - start + 1;
1407                         ret = do_readahead(mapping, file, start, len);
1408                 }
1409                 fput(file);
1410         }
1411         return ret;
1412 }
1413 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1414 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1415 {
1416         return SYSC_readahead((int) fd, offset, (size_t) count);
1417 }
1418 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1419 #endif
1420
1421 #ifdef CONFIG_MMU
1422 /**
1423  * page_cache_read - adds requested page to the page cache if not already there
1424  * @file:       file to read
1425  * @offset:     page index
1426  *
1427  * This adds the requested page to the page cache if it isn't already there,
1428  * and schedules an I/O to read in its contents from disk.
1429  */
1430 static int page_cache_read(struct file *file, pgoff_t offset)
1431 {
1432         struct address_space *mapping = file->f_mapping;
1433         struct page *page; 
1434         int ret;
1435
1436         do {
1437                 page = page_cache_alloc_cold(mapping);
1438                 if (!page)
1439                         return -ENOMEM;
1440
1441                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1442                 if (ret == 0)
1443                         ret = mapping->a_ops->readpage(file, page);
1444                 else if (ret == -EEXIST)
1445                         ret = 0; /* losing race to add is OK */
1446
1447                 page_cache_release(page);
1448
1449         } while (ret == AOP_TRUNCATED_PAGE);
1450                 
1451         return ret;
1452 }
1453
1454 #define MMAP_LOTSAMISS  (100)
1455
1456 /*
1457  * Synchronous readahead happens when we don't even find
1458  * a page in the page cache at all.
1459  */
1460 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1461                                    struct file_ra_state *ra,
1462                                    struct file *file,
1463                                    pgoff_t offset)
1464 {
1465         unsigned long ra_pages;
1466         struct address_space *mapping = file->f_mapping;
1467
1468         /* If we don't want any read-ahead, don't bother */
1469         if (VM_RandomReadHint(vma))
1470                 return;
1471
1472         if (VM_SequentialReadHint(vma) ||
1473                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1474                 page_cache_sync_readahead(mapping, ra, file, offset,
1475                                           ra->ra_pages);
1476                 return;
1477         }
1478
1479         if (ra->mmap_miss < INT_MAX)
1480                 ra->mmap_miss++;
1481
1482         /*
1483          * Do we miss much more than hit in this file? If so,
1484          * stop bothering with read-ahead. It will only hurt.
1485          */
1486         if (ra->mmap_miss > MMAP_LOTSAMISS)
1487                 return;
1488
1489         /*
1490          * mmap read-around
1491          */
1492         ra_pages = max_sane_readahead(ra->ra_pages);
1493         if (ra_pages) {
1494                 ra->start = max_t(long, 0, offset - ra_pages/2);
1495                 ra->size = ra_pages;
1496                 ra->async_size = 0;
1497                 ra_submit(ra, mapping, file);
1498         }
1499 }
1500
1501 /*
1502  * Asynchronous readahead happens when we find the page and PG_readahead,
1503  * so we want to possibly extend the readahead further..
1504  */
1505 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1506                                     struct file_ra_state *ra,
1507                                     struct file *file,
1508                                     struct page *page,
1509                                     pgoff_t offset)
1510 {
1511         struct address_space *mapping = file->f_mapping;
1512
1513         /* If we don't want any read-ahead, don't bother */
1514         if (VM_RandomReadHint(vma))
1515                 return;
1516         if (ra->mmap_miss > 0)
1517                 ra->mmap_miss--;
1518         if (PageReadahead(page))
1519                 page_cache_async_readahead(mapping, ra, file,
1520                                            page, offset, ra->ra_pages);
1521 }
1522
1523 /**
1524  * filemap_fault - read in file data for page fault handling
1525  * @vma:        vma in which the fault was taken
1526  * @vmf:        struct vm_fault containing details of the fault
1527  *
1528  * filemap_fault() is invoked via the vma operations vector for a
1529  * mapped memory region to read in file data during a page fault.
1530  *
1531  * The goto's are kind of ugly, but this streamlines the normal case of having
1532  * it in the page cache, and handles the special cases reasonably without
1533  * having a lot of duplicated code.
1534  */
1535 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1536 {
1537         int error;
1538         struct file *file = vma->vm_file;
1539         struct address_space *mapping = file->f_mapping;
1540         struct file_ra_state *ra = &file->f_ra;
1541         struct inode *inode = mapping->host;
1542         pgoff_t offset = vmf->pgoff;
1543         struct page *page;
1544         pgoff_t size;
1545         int ret = 0;
1546
1547         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1548         if (offset >= size)
1549                 return VM_FAULT_SIGBUS;
1550
1551         /*
1552          * Do we have something in the page cache already?
1553          */
1554         page = find_get_page(mapping, offset);
1555         if (likely(page)) {
1556                 /*
1557                  * We found the page, so try async readahead before
1558                  * waiting for the lock.
1559                  */
1560                 do_async_mmap_readahead(vma, ra, file, page, offset);
1561                 lock_page(page);
1562
1563                 /* Did it get truncated? */
1564                 if (unlikely(page->mapping != mapping)) {
1565                         unlock_page(page);
1566                         put_page(page);
1567                         goto no_cached_page;
1568                 }
1569         } else {
1570                 /* No page in the page cache at all */
1571                 do_sync_mmap_readahead(vma, ra, file, offset);
1572                 count_vm_event(PGMAJFAULT);
1573                 ret = VM_FAULT_MAJOR;
1574 retry_find:
1575                 page = find_lock_page(mapping, offset);
1576                 if (!page)
1577                         goto no_cached_page;
1578         }
1579
1580         /*
1581          * We have a locked page in the page cache, now we need to check
1582          * that it's up-to-date. If not, it is going to be due to an error.
1583          */
1584         if (unlikely(!PageUptodate(page)))
1585                 goto page_not_uptodate;
1586
1587         /*
1588          * Found the page and have a reference on it.
1589          * We must recheck i_size under page lock.
1590          */
1591         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1592         if (unlikely(offset >= size)) {
1593                 unlock_page(page);
1594                 page_cache_release(page);
1595                 return VM_FAULT_SIGBUS;
1596         }
1597
1598         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1599         vmf->page = page;
1600         return ret | VM_FAULT_LOCKED;
1601
1602 no_cached_page:
1603         /*
1604          * We're only likely to ever get here if MADV_RANDOM is in
1605          * effect.
1606          */
1607         error = page_cache_read(file, offset);
1608
1609         /*
1610          * The page we want has now been added to the page cache.
1611          * In the unlikely event that someone removed it in the
1612          * meantime, we'll just come back here and read it again.
1613          */
1614         if (error >= 0)
1615                 goto retry_find;
1616
1617         /*
1618          * An error return from page_cache_read can result if the
1619          * system is low on memory, or a problem occurs while trying
1620          * to schedule I/O.
1621          */
1622         if (error == -ENOMEM)
1623                 return VM_FAULT_OOM;
1624         return VM_FAULT_SIGBUS;
1625
1626 page_not_uptodate:
1627         /*
1628          * Umm, take care of errors if the page isn't up-to-date.
1629          * Try to re-read it _once_. We do this synchronously,
1630          * because there really aren't any performance issues here
1631          * and we need to check for errors.
1632          */
1633         ClearPageError(page);
1634         error = mapping->a_ops->readpage(file, page);
1635         if (!error) {
1636                 wait_on_page_locked(page);
1637                 if (!PageUptodate(page))
1638                         error = -EIO;
1639         }
1640         page_cache_release(page);
1641
1642         if (!error || error == AOP_TRUNCATED_PAGE)
1643                 goto retry_find;
1644
1645         /* Things didn't work out. Return zero to tell the mm layer so. */
1646         shrink_readahead_size_eio(file, ra);
1647         return VM_FAULT_SIGBUS;
1648 }
1649 EXPORT_SYMBOL(filemap_fault);
1650
1651 struct vm_operations_struct generic_file_vm_ops = {
1652         .fault          = filemap_fault,
1653 };
1654
1655 /* This is used for a general mmap of a disk file */
1656
1657 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1658 {
1659         struct address_space *mapping = file->f_mapping;
1660
1661         if (!mapping->a_ops->readpage)
1662                 return -ENOEXEC;
1663         file_accessed(file);
1664         vma->vm_ops = &generic_file_vm_ops;
1665         vma->vm_flags |= VM_CAN_NONLINEAR;
1666         return 0;
1667 }
1668
1669 /*
1670  * This is for filesystems which do not implement ->writepage.
1671  */
1672 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1673 {
1674         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1675                 return -EINVAL;
1676         return generic_file_mmap(file, vma);
1677 }
1678 #else
1679 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1680 {
1681         return -ENOSYS;
1682 }
1683 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1684 {
1685         return -ENOSYS;
1686 }
1687 #endif /* CONFIG_MMU */
1688
1689 EXPORT_SYMBOL(generic_file_mmap);
1690 EXPORT_SYMBOL(generic_file_readonly_mmap);
1691
1692 static struct page *__read_cache_page(struct address_space *mapping,
1693                                 pgoff_t index,
1694                                 int (*filler)(void *,struct page*),
1695                                 void *data)
1696 {
1697         struct page *page;
1698         int err;
1699 repeat:
1700         page = find_get_page(mapping, index);
1701         if (!page) {
1702                 page = page_cache_alloc_cold(mapping);
1703                 if (!page)
1704                         return ERR_PTR(-ENOMEM);
1705                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1706                 if (unlikely(err)) {
1707                         page_cache_release(page);
1708                         if (err == -EEXIST)
1709                                 goto repeat;
1710                         /* Presumably ENOMEM for radix tree node */
1711                         return ERR_PTR(err);
1712                 }
1713                 err = filler(data, page);
1714                 if (err < 0) {
1715                         page_cache_release(page);
1716                         page = ERR_PTR(err);
1717                 }
1718         }
1719         return page;
1720 }
1721
1722 /**
1723  * read_cache_page_async - read into page cache, fill it if needed
1724  * @mapping:    the page's address_space
1725  * @index:      the page index
1726  * @filler:     function to perform the read
1727  * @data:       destination for read data
1728  *
1729  * Same as read_cache_page, but don't wait for page to become unlocked
1730  * after submitting it to the filler.
1731  *
1732  * Read into the page cache. If a page already exists, and PageUptodate() is
1733  * not set, try to fill the page but don't wait for it to become unlocked.
1734  *
1735  * If the page does not get brought uptodate, return -EIO.
1736  */
1737 struct page *read_cache_page_async(struct address_space *mapping,
1738                                 pgoff_t index,
1739                                 int (*filler)(void *,struct page*),
1740                                 void *data)
1741 {
1742         struct page *page;
1743         int err;
1744
1745 retry:
1746         page = __read_cache_page(mapping, index, filler, data);
1747         if (IS_ERR(page))
1748                 return page;
1749         if (PageUptodate(page))
1750                 goto out;
1751
1752         lock_page(page);
1753         if (!page->mapping) {
1754                 unlock_page(page);
1755                 page_cache_release(page);
1756                 goto retry;
1757         }
1758         if (PageUptodate(page)) {
1759                 unlock_page(page);
1760                 goto out;
1761         }
1762         err = filler(data, page);
1763         if (err < 0) {
1764                 page_cache_release(page);
1765                 return ERR_PTR(err);
1766         }
1767 out:
1768         mark_page_accessed(page);
1769         return page;
1770 }
1771 EXPORT_SYMBOL(read_cache_page_async);
1772
1773 /**
1774  * read_cache_page - read into page cache, fill it if needed
1775  * @mapping:    the page's address_space
1776  * @index:      the page index
1777  * @filler:     function to perform the read
1778  * @data:       destination for read data
1779  *
1780  * Read into the page cache. If a page already exists, and PageUptodate() is
1781  * not set, try to fill the page then wait for it to become unlocked.
1782  *
1783  * If the page does not get brought uptodate, return -EIO.
1784  */
1785 struct page *read_cache_page(struct address_space *mapping,
1786                                 pgoff_t index,
1787                                 int (*filler)(void *,struct page*),
1788                                 void *data)
1789 {
1790         struct page *page;
1791
1792         page = read_cache_page_async(mapping, index, filler, data);
1793         if (IS_ERR(page))
1794                 goto out;
1795         wait_on_page_locked(page);
1796         if (!PageUptodate(page)) {
1797                 page_cache_release(page);
1798                 page = ERR_PTR(-EIO);
1799         }
1800  out:
1801         return page;
1802 }
1803 EXPORT_SYMBOL(read_cache_page);
1804
1805 /*
1806  * The logic we want is
1807  *
1808  *      if suid or (sgid and xgrp)
1809  *              remove privs
1810  */
1811 int should_remove_suid(struct dentry *dentry)
1812 {
1813         mode_t mode = dentry->d_inode->i_mode;
1814         int kill = 0;
1815
1816         /* suid always must be killed */
1817         if (unlikely(mode & S_ISUID))
1818                 kill = ATTR_KILL_SUID;
1819
1820         /*
1821          * sgid without any exec bits is just a mandatory locking mark; leave
1822          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1823          */
1824         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1825                 kill |= ATTR_KILL_SGID;
1826
1827         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1828                 return kill;
1829
1830         return 0;
1831 }
1832 EXPORT_SYMBOL(should_remove_suid);
1833
1834 static int __remove_suid(struct dentry *dentry, int kill)
1835 {
1836         struct iattr newattrs;
1837
1838         newattrs.ia_valid = ATTR_FORCE | kill;
1839         return notify_change(dentry, &newattrs);
1840 }
1841
1842 int file_remove_suid(struct file *file)
1843 {
1844         struct dentry *dentry = file->f_path.dentry;
1845         int killsuid = should_remove_suid(dentry);
1846         int killpriv = security_inode_need_killpriv(dentry);
1847         int error = 0;
1848
1849         if (killpriv < 0)
1850                 return killpriv;
1851         if (killpriv)
1852                 error = security_inode_killpriv(dentry);
1853         if (!error && killsuid)
1854                 error = __remove_suid(dentry, killsuid);
1855
1856         return error;
1857 }
1858 EXPORT_SYMBOL(file_remove_suid);
1859
1860 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1861                         const struct iovec *iov, size_t base, size_t bytes)
1862 {
1863         size_t copied = 0, left = 0;
1864
1865         while (bytes) {
1866                 char __user *buf = iov->iov_base + base;
1867                 int copy = min(bytes, iov->iov_len - base);
1868
1869                 base = 0;
1870                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1871                 copied += copy;
1872                 bytes -= copy;
1873                 vaddr += copy;
1874                 iov++;
1875
1876                 if (unlikely(left))
1877                         break;
1878         }
1879         return copied - left;
1880 }
1881
1882 /*
1883  * Copy as much as we can into the page and return the number of bytes which
1884  * were sucessfully copied.  If a fault is encountered then return the number of
1885  * bytes which were copied.
1886  */
1887 size_t iov_iter_copy_from_user_atomic(struct page *page,
1888                 struct iov_iter *i, unsigned long offset, size_t bytes)
1889 {
1890         char *kaddr;
1891         size_t copied;
1892
1893         BUG_ON(!in_atomic());
1894         kaddr = kmap_atomic(page, KM_USER0);
1895         if (likely(i->nr_segs == 1)) {
1896                 int left;
1897                 char __user *buf = i->iov->iov_base + i->iov_offset;
1898                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1899                 copied = bytes - left;
1900         } else {
1901                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1902                                                 i->iov, i->iov_offset, bytes);
1903         }
1904         kunmap_atomic(kaddr, KM_USER0);
1905
1906         return copied;
1907 }
1908 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1909
1910 /*
1911  * This has the same sideeffects and return value as
1912  * iov_iter_copy_from_user_atomic().
1913  * The difference is that it attempts to resolve faults.
1914  * Page must not be locked.
1915  */
1916 size_t iov_iter_copy_from_user(struct page *page,
1917                 struct iov_iter *i, unsigned long offset, size_t bytes)
1918 {
1919         char *kaddr;
1920         size_t copied;
1921
1922         kaddr = kmap(page);
1923         if (likely(i->nr_segs == 1)) {
1924                 int left;
1925                 char __user *buf = i->iov->iov_base + i->iov_offset;
1926                 left = __copy_from_user(kaddr + offset, buf, bytes);
1927                 copied = bytes - left;
1928         } else {
1929                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1930                                                 i->iov, i->iov_offset, bytes);
1931         }
1932         kunmap(page);
1933         return copied;
1934 }
1935 EXPORT_SYMBOL(iov_iter_copy_from_user);
1936
1937 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1938 {
1939         BUG_ON(i->count < bytes);
1940
1941         if (likely(i->nr_segs == 1)) {
1942                 i->iov_offset += bytes;
1943                 i->count -= bytes;
1944         } else {
1945                 const struct iovec *iov = i->iov;
1946                 size_t base = i->iov_offset;
1947
1948                 /*
1949                  * The !iov->iov_len check ensures we skip over unlikely
1950                  * zero-length segments (without overruning the iovec).
1951                  */
1952                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1953                         int copy;
1954
1955                         copy = min(bytes, iov->iov_len - base);
1956                         BUG_ON(!i->count || i->count < copy);
1957                         i->count -= copy;
1958                         bytes -= copy;
1959                         base += copy;
1960                         if (iov->iov_len == base) {
1961                                 iov++;
1962                                 base = 0;
1963                         }
1964                 }
1965                 i->iov = iov;
1966                 i->iov_offset = base;
1967         }
1968 }
1969 EXPORT_SYMBOL(iov_iter_advance);
1970
1971 /*
1972  * Fault in the first iovec of the given iov_iter, to a maximum length
1973  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1974  * accessed (ie. because it is an invalid address).
1975  *
1976  * writev-intensive code may want this to prefault several iovecs -- that
1977  * would be possible (callers must not rely on the fact that _only_ the
1978  * first iovec will be faulted with the current implementation).
1979  */
1980 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1981 {
1982         char __user *buf = i->iov->iov_base + i->iov_offset;
1983         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1984         return fault_in_pages_readable(buf, bytes);
1985 }
1986 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1987
1988 /*
1989  * Return the count of just the current iov_iter segment.
1990  */
1991 size_t iov_iter_single_seg_count(struct iov_iter *i)
1992 {
1993         const struct iovec *iov = i->iov;
1994         if (i->nr_segs == 1)
1995                 return i->count;
1996         else
1997                 return min(i->count, iov->iov_len - i->iov_offset);
1998 }
1999 EXPORT_SYMBOL(iov_iter_single_seg_count);
2000
2001 /*
2002  * Performs necessary checks before doing a write
2003  *
2004  * Can adjust writing position or amount of bytes to write.
2005  * Returns appropriate error code that caller should return or
2006  * zero in case that write should be allowed.
2007  */
2008 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2009 {
2010         struct inode *inode = file->f_mapping->host;
2011         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2012
2013         if (unlikely(*pos < 0))
2014                 return -EINVAL;
2015
2016         if (!isblk) {
2017                 /* FIXME: this is for backwards compatibility with 2.4 */
2018                 if (file->f_flags & O_APPEND)
2019                         *pos = i_size_read(inode);
2020
2021                 if (limit != RLIM_INFINITY) {
2022                         if (*pos >= limit) {
2023                                 send_sig(SIGXFSZ, current, 0);
2024                                 return -EFBIG;
2025                         }
2026                         if (*count > limit - (typeof(limit))*pos) {
2027                                 *count = limit - (typeof(limit))*pos;
2028                         }
2029                 }
2030         }
2031
2032         /*
2033          * LFS rule
2034          */
2035         if (unlikely(*pos + *count > MAX_NON_LFS &&
2036                                 !(file->f_flags & O_LARGEFILE))) {
2037                 if (*pos >= MAX_NON_LFS) {
2038                         return -EFBIG;
2039                 }
2040                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2041                         *count = MAX_NON_LFS - (unsigned long)*pos;
2042                 }
2043         }
2044
2045         /*
2046          * Are we about to exceed the fs block limit ?
2047          *
2048          * If we have written data it becomes a short write.  If we have
2049          * exceeded without writing data we send a signal and return EFBIG.
2050          * Linus frestrict idea will clean these up nicely..
2051          */
2052         if (likely(!isblk)) {
2053                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2054                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2055                                 return -EFBIG;
2056                         }
2057                         /* zero-length writes at ->s_maxbytes are OK */
2058                 }
2059
2060                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2061                         *count = inode->i_sb->s_maxbytes - *pos;
2062         } else {
2063 #ifdef CONFIG_BLOCK
2064                 loff_t isize;
2065                 if (bdev_read_only(I_BDEV(inode)))
2066                         return -EPERM;
2067                 isize = i_size_read(inode);
2068                 if (*pos >= isize) {
2069                         if (*count || *pos > isize)
2070                                 return -ENOSPC;
2071                 }
2072
2073                 if (*pos + *count > isize)
2074                         *count = isize - *pos;
2075 #else
2076                 return -EPERM;
2077 #endif
2078         }
2079         return 0;
2080 }
2081 EXPORT_SYMBOL(generic_write_checks);
2082
2083 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2084                                 loff_t pos, unsigned len, unsigned flags,
2085                                 struct page **pagep, void **fsdata)
2086 {
2087         const struct address_space_operations *aops = mapping->a_ops;
2088
2089         return aops->write_begin(file, mapping, pos, len, flags,
2090                                                         pagep, fsdata);
2091 }
2092 EXPORT_SYMBOL(pagecache_write_begin);
2093
2094 int pagecache_write_end(struct file *file, struct address_space *mapping,
2095                                 loff_t pos, unsigned len, unsigned copied,
2096                                 struct page *page, void *fsdata)
2097 {
2098         const struct address_space_operations *aops = mapping->a_ops;
2099
2100         mark_page_accessed(page);
2101         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2102 }
2103 EXPORT_SYMBOL(pagecache_write_end);
2104
2105 ssize_t
2106 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2107                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2108                 size_t count, size_t ocount)
2109 {
2110         struct file     *file = iocb->ki_filp;
2111         struct address_space *mapping = file->f_mapping;
2112         struct inode    *inode = mapping->host;
2113         ssize_t         written;
2114         size_t          write_len;
2115         pgoff_t         end;
2116
2117         if (count != ocount)
2118                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2119
2120         write_len = iov_length(iov, *nr_segs);
2121         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2122
2123         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2124         if (written)
2125                 goto out;
2126
2127         /*
2128          * After a write we want buffered reads to be sure to go to disk to get
2129          * the new data.  We invalidate clean cached page from the region we're
2130          * about to write.  We do this *before* the write so that we can return
2131          * without clobbering -EIOCBQUEUED from ->direct_IO().
2132          */
2133         if (mapping->nrpages) {
2134                 written = invalidate_inode_pages2_range(mapping,
2135                                         pos >> PAGE_CACHE_SHIFT, end);
2136                 /*
2137                  * If a page can not be invalidated, return 0 to fall back
2138                  * to buffered write.
2139                  */
2140                 if (written) {
2141                         if (written == -EBUSY)
2142                                 return 0;
2143                         goto out;
2144                 }
2145         }
2146
2147         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2148
2149         /*
2150          * Finally, try again to invalidate clean pages which might have been
2151          * cached by non-direct readahead, or faulted in by get_user_pages()
2152          * if the source of the write was an mmap'ed region of the file
2153          * we're writing.  Either one is a pretty crazy thing to do,
2154          * so we don't support it 100%.  If this invalidation
2155          * fails, tough, the write still worked...
2156          */
2157         if (mapping->nrpages) {
2158                 invalidate_inode_pages2_range(mapping,
2159                                               pos >> PAGE_CACHE_SHIFT, end);
2160         }
2161
2162         if (written > 0) {
2163                 loff_t end = pos + written;
2164                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2165                         i_size_write(inode,  end);
2166                         mark_inode_dirty(inode);
2167                 }
2168                 *ppos = end;
2169         }
2170
2171         /*
2172          * Sync the fs metadata but not the minor inode changes and
2173          * of course not the data as we did direct DMA for the IO.
2174          * i_mutex is held, which protects generic_osync_inode() from
2175          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2176          */
2177 out:
2178         if ((written >= 0 || written == -EIOCBQUEUED) &&
2179             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2180                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2181                 if (err < 0)
2182                         written = err;
2183         }
2184         return written;
2185 }
2186 EXPORT_SYMBOL(generic_file_direct_write);
2187
2188 /*
2189  * Find or create a page at the given pagecache position. Return the locked
2190  * page. This function is specifically for buffered writes.
2191  */
2192 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2193                                         pgoff_t index, unsigned flags)
2194 {
2195         int status;
2196         struct page *page;
2197         gfp_t gfp_notmask = 0;
2198         if (flags & AOP_FLAG_NOFS)
2199                 gfp_notmask = __GFP_FS;
2200 repeat:
2201         page = find_lock_page(mapping, index);
2202         if (likely(page))
2203                 return page;
2204
2205         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2206         if (!page)
2207                 return NULL;
2208         status = add_to_page_cache_lru(page, mapping, index,
2209                                                 GFP_KERNEL & ~gfp_notmask);
2210         if (unlikely(status)) {
2211                 page_cache_release(page);
2212                 if (status == -EEXIST)
2213                         goto repeat;
2214                 return NULL;
2215         }
2216         return page;
2217 }
2218 EXPORT_SYMBOL(grab_cache_page_write_begin);
2219
2220 static ssize_t generic_perform_write(struct file *file,
2221                                 struct iov_iter *i, loff_t pos)
2222 {
2223         struct address_space *mapping = file->f_mapping;
2224         const struct address_space_operations *a_ops = mapping->a_ops;
2225         long status = 0;
2226         ssize_t written = 0;
2227         unsigned int flags = 0;
2228
2229         /*
2230          * Copies from kernel address space cannot fail (NFSD is a big user).
2231          */
2232         if (segment_eq(get_fs(), KERNEL_DS))
2233                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2234
2235         do {
2236                 struct page *page;
2237                 pgoff_t index;          /* Pagecache index for current page */
2238                 unsigned long offset;   /* Offset into pagecache page */
2239                 unsigned long bytes;    /* Bytes to write to page */
2240                 size_t copied;          /* Bytes copied from user */
2241                 void *fsdata;
2242
2243                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2244                 index = pos >> PAGE_CACHE_SHIFT;
2245                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2246                                                 iov_iter_count(i));
2247
2248 again:
2249
2250                 /*
2251                  * Bring in the user page that we will copy from _first_.
2252                  * Otherwise there's a nasty deadlock on copying from the
2253                  * same page as we're writing to, without it being marked
2254                  * up-to-date.
2255                  *
2256                  * Not only is this an optimisation, but it is also required
2257                  * to check that the address is actually valid, when atomic
2258                  * usercopies are used, below.
2259                  */
2260                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2261                         status = -EFAULT;
2262                         break;
2263                 }
2264
2265                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2266                                                 &page, &fsdata);
2267                 if (unlikely(status))
2268                         break;
2269
2270                 pagefault_disable();
2271                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2272                 pagefault_enable();
2273                 flush_dcache_page(page);
2274
2275                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2276                                                 page, fsdata);
2277                 if (unlikely(status < 0))
2278                         break;
2279                 copied = status;
2280
2281                 cond_resched();
2282
2283                 iov_iter_advance(i, copied);
2284                 if (unlikely(copied == 0)) {
2285                         /*
2286                          * If we were unable to copy any data at all, we must
2287                          * fall back to a single segment length write.
2288                          *
2289                          * If we didn't fallback here, we could livelock
2290                          * because not all segments in the iov can be copied at
2291                          * once without a pagefault.
2292                          */
2293                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2294                                                 iov_iter_single_seg_count(i));
2295                         goto again;
2296                 }
2297                 pos += copied;
2298                 written += copied;
2299
2300                 balance_dirty_pages_ratelimited(mapping);
2301
2302         } while (iov_iter_count(i));
2303
2304         return written ? written : status;
2305 }
2306
2307 ssize_t
2308 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2309                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2310                 size_t count, ssize_t written)
2311 {
2312         struct file *file = iocb->ki_filp;
2313         struct address_space *mapping = file->f_mapping;
2314         const struct address_space_operations *a_ops = mapping->a_ops;
2315         struct inode *inode = mapping->host;
2316         ssize_t status;
2317         struct iov_iter i;
2318
2319         iov_iter_init(&i, iov, nr_segs, count, written);
2320         status = generic_perform_write(file, &i, pos);
2321
2322         if (likely(status >= 0)) {
2323                 written += status;
2324                 *ppos = pos + status;
2325
2326                 /*
2327                  * For now, when the user asks for O_SYNC, we'll actually give
2328                  * O_DSYNC
2329                  */
2330                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2331                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2332                                 status = generic_osync_inode(inode, mapping,
2333                                                 OSYNC_METADATA|OSYNC_DATA);
2334                 }
2335         }
2336         
2337         /*
2338          * If we get here for O_DIRECT writes then we must have fallen through
2339          * to buffered writes (block instantiation inside i_size).  So we sync
2340          * the file data here, to try to honour O_DIRECT expectations.
2341          */
2342         if (unlikely(file->f_flags & O_DIRECT) && written)
2343                 status = filemap_write_and_wait_range(mapping,
2344                                         pos, pos + written - 1);
2345
2346         return written ? written : status;
2347 }
2348 EXPORT_SYMBOL(generic_file_buffered_write);
2349
2350 static ssize_t
2351 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2352                                 unsigned long nr_segs, loff_t *ppos)
2353 {
2354         struct file *file = iocb->ki_filp;
2355         struct address_space * mapping = file->f_mapping;
2356         size_t ocount;          /* original count */
2357         size_t count;           /* after file limit checks */
2358         struct inode    *inode = mapping->host;
2359         loff_t          pos;
2360         ssize_t         written;
2361         ssize_t         err;
2362
2363         ocount = 0;
2364         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2365         if (err)
2366                 return err;
2367
2368         count = ocount;
2369         pos = *ppos;
2370
2371         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2372
2373         /* We can write back this queue in page reclaim */
2374         current->backing_dev_info = mapping->backing_dev_info;
2375         written = 0;
2376
2377         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2378         if (err)
2379                 goto out;
2380
2381         if (count == 0)
2382                 goto out;
2383
2384         err = file_remove_suid(file);
2385         if (err)
2386                 goto out;
2387
2388         file_update_time(file);
2389
2390         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2391         if (unlikely(file->f_flags & O_DIRECT)) {
2392                 loff_t endbyte;
2393                 ssize_t written_buffered;
2394
2395                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2396                                                         ppos, count, ocount);
2397                 if (written < 0 || written == count)
2398                         goto out;
2399                 /*
2400                  * direct-io write to a hole: fall through to buffered I/O
2401                  * for completing the rest of the request.
2402                  */
2403                 pos += written;
2404                 count -= written;
2405                 written_buffered = generic_file_buffered_write(iocb, iov,
2406                                                 nr_segs, pos, ppos, count,
2407                                                 written);
2408                 /*
2409                  * If generic_file_buffered_write() retuned a synchronous error
2410                  * then we want to return the number of bytes which were
2411                  * direct-written, or the error code if that was zero.  Note
2412                  * that this differs from normal direct-io semantics, which
2413                  * will return -EFOO even if some bytes were written.
2414                  */
2415                 if (written_buffered < 0) {
2416                         err = written_buffered;
2417                         goto out;
2418                 }
2419
2420                 /*
2421                  * We need to ensure that the page cache pages are written to
2422                  * disk and invalidated to preserve the expected O_DIRECT
2423                  * semantics.
2424                  */
2425                 endbyte = pos + written_buffered - written - 1;
2426                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2427                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2428                                             SYNC_FILE_RANGE_WRITE|
2429                                             SYNC_FILE_RANGE_WAIT_AFTER);
2430                 if (err == 0) {
2431                         written = written_buffered;
2432                         invalidate_mapping_pages(mapping,
2433                                                  pos >> PAGE_CACHE_SHIFT,
2434                                                  endbyte >> PAGE_CACHE_SHIFT);
2435                 } else {
2436                         /*
2437                          * We don't know how much we wrote, so just return
2438                          * the number of bytes which were direct-written
2439                          */
2440                 }
2441         } else {
2442                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2443                                 pos, ppos, count, written);
2444         }
2445 out:
2446         current->backing_dev_info = NULL;
2447         return written ? written : err;
2448 }
2449
2450 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2451                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2452 {
2453         struct file *file = iocb->ki_filp;
2454         struct address_space *mapping = file->f_mapping;
2455         struct inode *inode = mapping->host;
2456         ssize_t ret;
2457
2458         BUG_ON(iocb->ki_pos != pos);
2459
2460         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2461                         &iocb->ki_pos);
2462
2463         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2464                 ssize_t err;
2465
2466                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2467                 if (err < 0)
2468                         ret = err;
2469         }
2470         return ret;
2471 }
2472 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2473
2474 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2475                 unsigned long nr_segs, loff_t pos)
2476 {
2477         struct file *file = iocb->ki_filp;
2478         struct address_space *mapping = file->f_mapping;
2479         struct inode *inode = mapping->host;
2480         ssize_t ret;
2481
2482         BUG_ON(iocb->ki_pos != pos);
2483
2484         mutex_lock(&inode->i_mutex);
2485         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2486                         &iocb->ki_pos);
2487         mutex_unlock(&inode->i_mutex);
2488
2489         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2490                 ssize_t err;
2491
2492                 err = sync_page_range(inode, mapping, pos, ret);
2493                 if (err < 0)
2494                         ret = err;
2495         }
2496         return ret;
2497 }
2498 EXPORT_SYMBOL(generic_file_aio_write);
2499
2500 /**
2501  * try_to_release_page() - release old fs-specific metadata on a page
2502  *
2503  * @page: the page which the kernel is trying to free
2504  * @gfp_mask: memory allocation flags (and I/O mode)
2505  *
2506  * The address_space is to try to release any data against the page
2507  * (presumably at page->private).  If the release was successful, return `1'.
2508  * Otherwise return zero.
2509  *
2510  * This may also be called if PG_fscache is set on a page, indicating that the
2511  * page is known to the local caching routines.
2512  *
2513  * The @gfp_mask argument specifies whether I/O may be performed to release
2514  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2515  *
2516  */
2517 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2518 {
2519         struct address_space * const mapping = page->mapping;
2520
2521         BUG_ON(!PageLocked(page));
2522         if (PageWriteback(page))
2523                 return 0;
2524
2525         if (mapping && mapping->a_ops->releasepage)
2526                 return mapping->a_ops->releasepage(page, gfp_mask);
2527         return try_to_free_buffers(page);
2528 }
2529
2530 EXPORT_SYMBOL(try_to_release_page);