kill generic_file_direct_IO()
[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 "internal.h"
37
38 /*
39  * FIXME: remove all knowledge of the buffer layer from the core VM
40  */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (vmtruncate)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  */
108
109 /*
110  * Remove a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
113  */
114 void __remove_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         mem_cgroup_uncharge_page(page);
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         write_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         write_unlock_irq(&mapping->tree_lock);
147 }
148
149 static int sync_page(void *word)
150 {
151         struct address_space *mapping;
152         struct page *page;
153
154         page = container_of((unsigned long *)word, struct page, flags);
155
156         /*
157          * page_mapping() is being called without PG_locked held.
158          * Some knowledge of the state and use of the page is used to
159          * reduce the requirements down to a memory barrier.
160          * The danger here is of a stale page_mapping() return value
161          * indicating a struct address_space different from the one it's
162          * associated with when it is associated with one.
163          * After smp_mb(), it's either the correct page_mapping() for
164          * the page, or an old page_mapping() and the page's own
165          * page_mapping() has gone NULL.
166          * The ->sync_page() address_space operation must tolerate
167          * page_mapping() going NULL. By an amazing coincidence,
168          * this comes about because none of the users of the page
169          * in the ->sync_page() methods make essential use of the
170          * page_mapping(), merely passing the page down to the backing
171          * device's unplug functions when it's non-NULL, which in turn
172          * ignore it for all cases but swap, where only page_private(page) is
173          * of interest. When page_mapping() does go NULL, the entire
174          * call stack gracefully ignores the page and returns.
175          * -- wli
176          */
177         smp_mb();
178         mapping = page_mapping(page);
179         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
180                 mapping->a_ops->sync_page(page);
181         io_schedule();
182         return 0;
183 }
184
185 static int sync_page_killable(void *word)
186 {
187         sync_page(word);
188         return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 /**
192  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
193  * @mapping:    address space structure to write
194  * @start:      offset in bytes where the range starts
195  * @end:        offset in bytes where the range ends (inclusive)
196  * @sync_mode:  enable synchronous operation
197  *
198  * Start writeback against all of a mapping's dirty pages that lie
199  * within the byte offsets <start, end> inclusive.
200  *
201  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
202  * opposed to a regular memory cleansing writeback.  The difference between
203  * these two operations is that if a dirty page/buffer is encountered, it must
204  * be waited upon, and not just skipped over.
205  */
206 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
207                                 loff_t end, int sync_mode)
208 {
209         int ret;
210         struct writeback_control wbc = {
211                 .sync_mode = sync_mode,
212                 .nr_to_write = mapping->nrpages * 2,
213                 .range_start = start,
214                 .range_end = end,
215         };
216
217         if (!mapping_cap_writeback_dirty(mapping))
218                 return 0;
219
220         ret = do_writepages(mapping, &wbc);
221         return ret;
222 }
223
224 static inline int __filemap_fdatawrite(struct address_space *mapping,
225         int sync_mode)
226 {
227         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
228 }
229
230 int filemap_fdatawrite(struct address_space *mapping)
231 {
232         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
233 }
234 EXPORT_SYMBOL(filemap_fdatawrite);
235
236 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
237                                 loff_t end)
238 {
239         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite_range);
242
243 /**
244  * filemap_flush - mostly a non-blocking flush
245  * @mapping:    target address_space
246  *
247  * This is a mostly non-blocking flush.  Not suitable for data-integrity
248  * purposes - I/O may not be started against all dirty pages.
249  */
250 int filemap_flush(struct address_space *mapping)
251 {
252         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
253 }
254 EXPORT_SYMBOL(filemap_flush);
255
256 /**
257  * wait_on_page_writeback_range - wait for writeback to complete
258  * @mapping:    target address_space
259  * @start:      beginning page index
260  * @end:        ending page index
261  *
262  * Wait for writeback to complete against pages indexed by start->end
263  * inclusive
264  */
265 int wait_on_page_writeback_range(struct address_space *mapping,
266                                 pgoff_t start, pgoff_t end)
267 {
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271         pgoff_t index;
272
273         if (end < start)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         index = start;
278         while ((index <= end) &&
279                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
280                         PAGECACHE_TAG_WRITEBACK,
281                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
282                 unsigned i;
283
284                 for (i = 0; i < nr_pages; i++) {
285                         struct page *page = pvec.pages[i];
286
287                         /* until radix tree lookup accepts end_index */
288                         if (page->index > end)
289                                 continue;
290
291                         wait_on_page_writeback(page);
292                         if (PageError(page))
293                                 ret = -EIO;
294                 }
295                 pagevec_release(&pvec);
296                 cond_resched();
297         }
298
299         /* Check for outstanding write errors */
300         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
301                 ret = -ENOSPC;
302         if (test_and_clear_bit(AS_EIO, &mapping->flags))
303                 ret = -EIO;
304
305         return ret;
306 }
307
308 /**
309  * sync_page_range - write and wait on all pages in the passed range
310  * @inode:      target inode
311  * @mapping:    target address_space
312  * @pos:        beginning offset in pages to write
313  * @count:      number of bytes to write
314  *
315  * Write and wait upon all the pages in the passed range.  This is a "data
316  * integrity" operation.  It waits upon in-flight writeout before starting and
317  * waiting upon new writeout.  If there was an IO error, return it.
318  *
319  * We need to re-take i_mutex during the generic_osync_inode list walk because
320  * it is otherwise livelockable.
321  */
322 int sync_page_range(struct inode *inode, struct address_space *mapping,
323                         loff_t pos, loff_t count)
324 {
325         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
326         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
327         int ret;
328
329         if (!mapping_cap_writeback_dirty(mapping) || !count)
330                 return 0;
331         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
332         if (ret == 0) {
333                 mutex_lock(&inode->i_mutex);
334                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
335                 mutex_unlock(&inode->i_mutex);
336         }
337         if (ret == 0)
338                 ret = wait_on_page_writeback_range(mapping, start, end);
339         return ret;
340 }
341 EXPORT_SYMBOL(sync_page_range);
342
343 /**
344  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
345  * @inode:      target inode
346  * @mapping:    target address_space
347  * @pos:        beginning offset in pages to write
348  * @count:      number of bytes to write
349  *
350  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
351  * as it forces O_SYNC writers to different parts of the same file
352  * to be serialised right until io completion.
353  */
354 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
355                            loff_t pos, loff_t count)
356 {
357         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
358         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
359         int ret;
360
361         if (!mapping_cap_writeback_dirty(mapping) || !count)
362                 return 0;
363         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
364         if (ret == 0)
365                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
366         if (ret == 0)
367                 ret = wait_on_page_writeback_range(mapping, start, end);
368         return ret;
369 }
370 EXPORT_SYMBOL(sync_page_range_nolock);
371
372 /**
373  * filemap_fdatawait - wait for all under-writeback pages to complete
374  * @mapping: address space structure to wait for
375  *
376  * Walk the list of under-writeback pages of the given address space
377  * and wait for all of them.
378  */
379 int filemap_fdatawait(struct address_space *mapping)
380 {
381         loff_t i_size = i_size_read(mapping->host);
382
383         if (i_size == 0)
384                 return 0;
385
386         return wait_on_page_writeback_range(mapping, 0,
387                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
388 }
389 EXPORT_SYMBOL(filemap_fdatawait);
390
391 int filemap_write_and_wait(struct address_space *mapping)
392 {
393         int err = 0;
394
395         if (mapping->nrpages) {
396                 err = filemap_fdatawrite(mapping);
397                 /*
398                  * Even if the above returned error, the pages may be
399                  * written partially (e.g. -ENOSPC), so we wait for it.
400                  * But the -EIO is special case, it may indicate the worst
401                  * thing (e.g. bug) happened, so we avoid waiting for it.
402                  */
403                 if (err != -EIO) {
404                         int err2 = filemap_fdatawait(mapping);
405                         if (!err)
406                                 err = err2;
407                 }
408         }
409         return err;
410 }
411 EXPORT_SYMBOL(filemap_write_and_wait);
412
413 /**
414  * filemap_write_and_wait_range - write out & wait on a file range
415  * @mapping:    the address_space for the pages
416  * @lstart:     offset in bytes where the range starts
417  * @lend:       offset in bytes where the range ends (inclusive)
418  *
419  * Write out and wait upon file offsets lstart->lend, inclusive.
420  *
421  * Note that `lend' is inclusive (describes the last byte to be written) so
422  * that this function can be used to write to the very end-of-file (end = -1).
423  */
424 int filemap_write_and_wait_range(struct address_space *mapping,
425                                  loff_t lstart, loff_t lend)
426 {
427         int err = 0;
428
429         if (mapping->nrpages) {
430                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
431                                                  WB_SYNC_ALL);
432                 /* See comment of filemap_write_and_wait() */
433                 if (err != -EIO) {
434                         int err2 = wait_on_page_writeback_range(mapping,
435                                                 lstart >> PAGE_CACHE_SHIFT,
436                                                 lend >> PAGE_CACHE_SHIFT);
437                         if (!err)
438                                 err = err2;
439                 }
440         }
441         return err;
442 }
443
444 /**
445  * add_to_page_cache - add newly allocated pagecache pages
446  * @page:       page to add
447  * @mapping:    the page's address_space
448  * @offset:     page index
449  * @gfp_mask:   page allocation mode
450  *
451  * This function is used to add newly allocated pagecache pages;
452  * the page is new, so we can just run SetPageLocked() against it.
453  * The other page state flags were set by rmqueue().
454  *
455  * This function does not add the page to the LRU.  The caller must do that.
456  */
457 int add_to_page_cache(struct page *page, struct address_space *mapping,
458                 pgoff_t offset, gfp_t gfp_mask)
459 {
460         int error = mem_cgroup_cache_charge(page, current->mm,
461                                         gfp_mask & ~__GFP_HIGHMEM);
462         if (error)
463                 goto out;
464
465         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
466         if (error == 0) {
467                 write_lock_irq(&mapping->tree_lock);
468                 error = radix_tree_insert(&mapping->page_tree, offset, page);
469                 if (!error) {
470                         page_cache_get(page);
471                         SetPageLocked(page);
472                         page->mapping = mapping;
473                         page->index = offset;
474                         mapping->nrpages++;
475                         __inc_zone_page_state(page, NR_FILE_PAGES);
476                 } else
477                         mem_cgroup_uncharge_page(page);
478
479                 write_unlock_irq(&mapping->tree_lock);
480                 radix_tree_preload_end();
481         } else
482                 mem_cgroup_uncharge_page(page);
483 out:
484         return error;
485 }
486 EXPORT_SYMBOL(add_to_page_cache);
487
488 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
489                                 pgoff_t offset, gfp_t gfp_mask)
490 {
491         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
492         if (ret == 0)
493                 lru_cache_add(page);
494         return ret;
495 }
496
497 #ifdef CONFIG_NUMA
498 struct page *__page_cache_alloc(gfp_t gfp)
499 {
500         if (cpuset_do_page_mem_spread()) {
501                 int n = cpuset_mem_spread_node();
502                 return alloc_pages_node(n, gfp, 0);
503         }
504         return alloc_pages(gfp, 0);
505 }
506 EXPORT_SYMBOL(__page_cache_alloc);
507 #endif
508
509 static int __sleep_on_page_lock(void *word)
510 {
511         io_schedule();
512         return 0;
513 }
514
515 /*
516  * In order to wait for pages to become available there must be
517  * waitqueues associated with pages. By using a hash table of
518  * waitqueues where the bucket discipline is to maintain all
519  * waiters on the same queue and wake all when any of the pages
520  * become available, and for the woken contexts to check to be
521  * sure the appropriate page became available, this saves space
522  * at a cost of "thundering herd" phenomena during rare hash
523  * collisions.
524  */
525 static wait_queue_head_t *page_waitqueue(struct page *page)
526 {
527         const struct zone *zone = page_zone(page);
528
529         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
530 }
531
532 static inline void wake_up_page(struct page *page, int bit)
533 {
534         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
535 }
536
537 void wait_on_page_bit(struct page *page, int bit_nr)
538 {
539         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
540
541         if (test_bit(bit_nr, &page->flags))
542                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
543                                                         TASK_UNINTERRUPTIBLE);
544 }
545 EXPORT_SYMBOL(wait_on_page_bit);
546
547 /**
548  * unlock_page - unlock a locked page
549  * @page: the page
550  *
551  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
552  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
553  * mechananism between PageLocked pages and PageWriteback pages is shared.
554  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
555  *
556  * The first mb is necessary to safely close the critical section opened by the
557  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
558  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
559  * parallel wait_on_page_locked()).
560  */
561 void unlock_page(struct page *page)
562 {
563         smp_mb__before_clear_bit();
564         if (!TestClearPageLocked(page))
565                 BUG();
566         smp_mb__after_clear_bit(); 
567         wake_up_page(page, PG_locked);
568 }
569 EXPORT_SYMBOL(unlock_page);
570
571 /**
572  * end_page_writeback - end writeback against a page
573  * @page: the page
574  */
575 void end_page_writeback(struct page *page)
576 {
577         if (TestClearPageReclaim(page))
578                 rotate_reclaimable_page(page);
579
580         if (!test_clear_page_writeback(page))
581                 BUG();
582
583         smp_mb__after_clear_bit();
584         wake_up_page(page, PG_writeback);
585 }
586 EXPORT_SYMBOL(end_page_writeback);
587
588 /**
589  * __lock_page - get a lock on the page, assuming we need to sleep to get it
590  * @page: the page to lock
591  *
592  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
593  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
594  * chances are that on the second loop, the block layer's plug list is empty,
595  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
596  */
597 void __lock_page(struct page *page)
598 {
599         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
600
601         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
602                                                         TASK_UNINTERRUPTIBLE);
603 }
604 EXPORT_SYMBOL(__lock_page);
605
606 int __lock_page_killable(struct page *page)
607 {
608         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
609
610         return __wait_on_bit_lock(page_waitqueue(page), &wait,
611                                         sync_page_killable, TASK_KILLABLE);
612 }
613
614 /**
615  * __lock_page_nosync - get a lock on the page, without calling sync_page()
616  * @page: the page to lock
617  *
618  * Variant of lock_page that does not require the caller to hold a reference
619  * on the page's mapping.
620  */
621 void __lock_page_nosync(struct page *page)
622 {
623         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
624         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
625                                                         TASK_UNINTERRUPTIBLE);
626 }
627
628 /**
629  * find_get_page - find and get a page reference
630  * @mapping: the address_space to search
631  * @offset: the page index
632  *
633  * Is there a pagecache struct page at the given (mapping, offset) tuple?
634  * If yes, increment its refcount and return it; if no, return NULL.
635  */
636 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
637 {
638         struct page *page;
639
640         read_lock_irq(&mapping->tree_lock);
641         page = radix_tree_lookup(&mapping->page_tree, offset);
642         if (page)
643                 page_cache_get(page);
644         read_unlock_irq(&mapping->tree_lock);
645         return page;
646 }
647 EXPORT_SYMBOL(find_get_page);
648
649 /**
650  * find_lock_page - locate, pin and lock a pagecache page
651  * @mapping: the address_space to search
652  * @offset: the page index
653  *
654  * Locates the desired pagecache page, locks it, increments its reference
655  * count and returns its address.
656  *
657  * Returns zero if the page was not present. find_lock_page() may sleep.
658  */
659 struct page *find_lock_page(struct address_space *mapping,
660                                 pgoff_t offset)
661 {
662         struct page *page;
663
664 repeat:
665         read_lock_irq(&mapping->tree_lock);
666         page = radix_tree_lookup(&mapping->page_tree, offset);
667         if (page) {
668                 page_cache_get(page);
669                 if (TestSetPageLocked(page)) {
670                         read_unlock_irq(&mapping->tree_lock);
671                         __lock_page(page);
672
673                         /* Has the page been truncated while we slept? */
674                         if (unlikely(page->mapping != mapping)) {
675                                 unlock_page(page);
676                                 page_cache_release(page);
677                                 goto repeat;
678                         }
679                         VM_BUG_ON(page->index != offset);
680                         goto out;
681                 }
682         }
683         read_unlock_irq(&mapping->tree_lock);
684 out:
685         return page;
686 }
687 EXPORT_SYMBOL(find_lock_page);
688
689 /**
690  * find_or_create_page - locate or add a pagecache page
691  * @mapping: the page's address_space
692  * @index: the page's index into the mapping
693  * @gfp_mask: page allocation mode
694  *
695  * Locates a page in the pagecache.  If the page is not present, a new page
696  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
697  * LRU list.  The returned page is locked and has its reference count
698  * incremented.
699  *
700  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
701  * allocation!
702  *
703  * find_or_create_page() returns the desired page's address, or zero on
704  * memory exhaustion.
705  */
706 struct page *find_or_create_page(struct address_space *mapping,
707                 pgoff_t index, gfp_t gfp_mask)
708 {
709         struct page *page;
710         int err;
711 repeat:
712         page = find_lock_page(mapping, index);
713         if (!page) {
714                 page = __page_cache_alloc(gfp_mask);
715                 if (!page)
716                         return NULL;
717                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
718                 if (unlikely(err)) {
719                         page_cache_release(page);
720                         page = NULL;
721                         if (err == -EEXIST)
722                                 goto repeat;
723                 }
724         }
725         return page;
726 }
727 EXPORT_SYMBOL(find_or_create_page);
728
729 /**
730  * find_get_pages - gang pagecache lookup
731  * @mapping:    The address_space to search
732  * @start:      The starting page index
733  * @nr_pages:   The maximum number of pages
734  * @pages:      Where the resulting pages are placed
735  *
736  * find_get_pages() will search for and return a group of up to
737  * @nr_pages pages in the mapping.  The pages are placed at @pages.
738  * find_get_pages() takes a reference against the returned pages.
739  *
740  * The search returns a group of mapping-contiguous pages with ascending
741  * indexes.  There may be holes in the indices due to not-present pages.
742  *
743  * find_get_pages() returns the number of pages which were found.
744  */
745 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
746                             unsigned int nr_pages, struct page **pages)
747 {
748         unsigned int i;
749         unsigned int ret;
750
751         read_lock_irq(&mapping->tree_lock);
752         ret = radix_tree_gang_lookup(&mapping->page_tree,
753                                 (void **)pages, start, nr_pages);
754         for (i = 0; i < ret; i++)
755                 page_cache_get(pages[i]);
756         read_unlock_irq(&mapping->tree_lock);
757         return ret;
758 }
759
760 /**
761  * find_get_pages_contig - gang contiguous pagecache lookup
762  * @mapping:    The address_space to search
763  * @index:      The starting page index
764  * @nr_pages:   The maximum number of pages
765  * @pages:      Where the resulting pages are placed
766  *
767  * find_get_pages_contig() works exactly like find_get_pages(), except
768  * that the returned number of pages are guaranteed to be contiguous.
769  *
770  * find_get_pages_contig() returns the number of pages which were found.
771  */
772 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
773                                unsigned int nr_pages, struct page **pages)
774 {
775         unsigned int i;
776         unsigned int ret;
777
778         read_lock_irq(&mapping->tree_lock);
779         ret = radix_tree_gang_lookup(&mapping->page_tree,
780                                 (void **)pages, index, nr_pages);
781         for (i = 0; i < ret; i++) {
782                 if (pages[i]->mapping == NULL || pages[i]->index != index)
783                         break;
784
785                 page_cache_get(pages[i]);
786                 index++;
787         }
788         read_unlock_irq(&mapping->tree_lock);
789         return i;
790 }
791 EXPORT_SYMBOL(find_get_pages_contig);
792
793 /**
794  * find_get_pages_tag - find and return pages that match @tag
795  * @mapping:    the address_space to search
796  * @index:      the starting page index
797  * @tag:        the tag index
798  * @nr_pages:   the maximum number of pages
799  * @pages:      where the resulting pages are placed
800  *
801  * Like find_get_pages, except we only return pages which are tagged with
802  * @tag.   We update @index to index the next page for the traversal.
803  */
804 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
805                         int tag, unsigned int nr_pages, struct page **pages)
806 {
807         unsigned int i;
808         unsigned int ret;
809
810         read_lock_irq(&mapping->tree_lock);
811         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
812                                 (void **)pages, *index, nr_pages, tag);
813         for (i = 0; i < ret; i++)
814                 page_cache_get(pages[i]);
815         if (ret)
816                 *index = pages[ret - 1]->index + 1;
817         read_unlock_irq(&mapping->tree_lock);
818         return ret;
819 }
820 EXPORT_SYMBOL(find_get_pages_tag);
821
822 /**
823  * grab_cache_page_nowait - returns locked page at given index in given cache
824  * @mapping: target address_space
825  * @index: the page index
826  *
827  * Same as grab_cache_page(), but do not wait if the page is unavailable.
828  * This is intended for speculative data generators, where the data can
829  * be regenerated if the page couldn't be grabbed.  This routine should
830  * be safe to call while holding the lock for another page.
831  *
832  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
833  * and deadlock against the caller's locked page.
834  */
835 struct page *
836 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
837 {
838         struct page *page = find_get_page(mapping, index);
839
840         if (page) {
841                 if (!TestSetPageLocked(page))
842                         return page;
843                 page_cache_release(page);
844                 return NULL;
845         }
846         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
847         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
848                 page_cache_release(page);
849                 page = NULL;
850         }
851         return page;
852 }
853 EXPORT_SYMBOL(grab_cache_page_nowait);
854
855 /*
856  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
857  * a _large_ part of the i/o request. Imagine the worst scenario:
858  *
859  *      ---R__________________________________________B__________
860  *         ^ reading here                             ^ bad block(assume 4k)
861  *
862  * read(R) => miss => readahead(R...B) => media error => frustrating retries
863  * => failing the whole request => read(R) => read(R+1) =>
864  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
865  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
866  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
867  *
868  * It is going insane. Fix it by quickly scaling down the readahead size.
869  */
870 static void shrink_readahead_size_eio(struct file *filp,
871                                         struct file_ra_state *ra)
872 {
873         if (!ra->ra_pages)
874                 return;
875
876         ra->ra_pages /= 4;
877 }
878
879 /**
880  * do_generic_file_read - generic file read routine
881  * @filp:       the file to read
882  * @ppos:       current file position
883  * @desc:       read_descriptor
884  * @actor:      read method
885  *
886  * This is a generic file read routine, and uses the
887  * mapping->a_ops->readpage() function for the actual low-level stuff.
888  *
889  * This is really ugly. But the goto's actually try to clarify some
890  * of the logic when it comes to error handling etc.
891  */
892 static void do_generic_file_read(struct file *filp, loff_t *ppos,
893                 read_descriptor_t *desc, read_actor_t actor)
894 {
895         struct address_space *mapping = filp->f_mapping;
896         struct inode *inode = mapping->host;
897         struct file_ra_state *ra = &filp->f_ra;
898         pgoff_t index;
899         pgoff_t last_index;
900         pgoff_t prev_index;
901         unsigned long offset;      /* offset into pagecache page */
902         unsigned int prev_offset;
903         int error;
904
905         index = *ppos >> PAGE_CACHE_SHIFT;
906         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
907         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
908         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
909         offset = *ppos & ~PAGE_CACHE_MASK;
910
911         for (;;) {
912                 struct page *page;
913                 pgoff_t end_index;
914                 loff_t isize;
915                 unsigned long nr, ret;
916
917                 cond_resched();
918 find_page:
919                 page = find_get_page(mapping, index);
920                 if (!page) {
921                         page_cache_sync_readahead(mapping,
922                                         ra, filp,
923                                         index, last_index - index);
924                         page = find_get_page(mapping, index);
925                         if (unlikely(page == NULL))
926                                 goto no_cached_page;
927                 }
928                 if (PageReadahead(page)) {
929                         page_cache_async_readahead(mapping,
930                                         ra, filp, page,
931                                         index, last_index - index);
932                 }
933                 if (!PageUptodate(page))
934                         goto page_not_up_to_date;
935 page_ok:
936                 /*
937                  * i_size must be checked after we know the page is Uptodate.
938                  *
939                  * Checking i_size after the check allows us to calculate
940                  * the correct value for "nr", which means the zero-filled
941                  * part of the page is not copied back to userspace (unless
942                  * another truncate extends the file - this is desired though).
943                  */
944
945                 isize = i_size_read(inode);
946                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
947                 if (unlikely(!isize || index > end_index)) {
948                         page_cache_release(page);
949                         goto out;
950                 }
951
952                 /* nr is the maximum number of bytes to copy from this page */
953                 nr = PAGE_CACHE_SIZE;
954                 if (index == end_index) {
955                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
956                         if (nr <= offset) {
957                                 page_cache_release(page);
958                                 goto out;
959                         }
960                 }
961                 nr = nr - offset;
962
963                 /* If users can be writing to this page using arbitrary
964                  * virtual addresses, take care about potential aliasing
965                  * before reading the page on the kernel side.
966                  */
967                 if (mapping_writably_mapped(mapping))
968                         flush_dcache_page(page);
969
970                 /*
971                  * When a sequential read accesses a page several times,
972                  * only mark it as accessed the first time.
973                  */
974                 if (prev_index != index || offset != prev_offset)
975                         mark_page_accessed(page);
976                 prev_index = index;
977
978                 /*
979                  * Ok, we have the page, and it's up-to-date, so
980                  * now we can copy it to user space...
981                  *
982                  * The actor routine returns how many bytes were actually used..
983                  * NOTE! This may not be the same as how much of a user buffer
984                  * we filled up (we may be padding etc), so we can only update
985                  * "pos" here (the actor routine has to update the user buffer
986                  * pointers and the remaining count).
987                  */
988                 ret = actor(desc, page, offset, nr);
989                 offset += ret;
990                 index += offset >> PAGE_CACHE_SHIFT;
991                 offset &= ~PAGE_CACHE_MASK;
992                 prev_offset = offset;
993
994                 page_cache_release(page);
995                 if (ret == nr && desc->count)
996                         continue;
997                 goto out;
998
999 page_not_up_to_date:
1000                 /* Get exclusive access to the page ... */
1001                 if (lock_page_killable(page))
1002                         goto readpage_eio;
1003
1004                 /* Did it get truncated before we got the lock? */
1005                 if (!page->mapping) {
1006                         unlock_page(page);
1007                         page_cache_release(page);
1008                         continue;
1009                 }
1010
1011                 /* Did somebody else fill it already? */
1012                 if (PageUptodate(page)) {
1013                         unlock_page(page);
1014                         goto page_ok;
1015                 }
1016
1017 readpage:
1018                 /* Start the actual read. The read will unlock the page. */
1019                 error = mapping->a_ops->readpage(filp, page);
1020
1021                 if (unlikely(error)) {
1022                         if (error == AOP_TRUNCATED_PAGE) {
1023                                 page_cache_release(page);
1024                                 goto find_page;
1025                         }
1026                         goto readpage_error;
1027                 }
1028
1029                 if (!PageUptodate(page)) {
1030                         if (lock_page_killable(page))
1031                                 goto readpage_eio;
1032                         if (!PageUptodate(page)) {
1033                                 if (page->mapping == NULL) {
1034                                         /*
1035                                          * invalidate_inode_pages got it
1036                                          */
1037                                         unlock_page(page);
1038                                         page_cache_release(page);
1039                                         goto find_page;
1040                                 }
1041                                 unlock_page(page);
1042                                 shrink_readahead_size_eio(filp, ra);
1043                                 goto readpage_eio;
1044                         }
1045                         unlock_page(page);
1046                 }
1047
1048                 goto page_ok;
1049
1050 readpage_eio:
1051                 error = -EIO;
1052 readpage_error:
1053                 /* UHHUH! A synchronous read error occurred. Report it */
1054                 desc->error = error;
1055                 page_cache_release(page);
1056                 goto out;
1057
1058 no_cached_page:
1059                 /*
1060                  * Ok, it wasn't cached, so we need to create a new
1061                  * page..
1062                  */
1063                 page = page_cache_alloc_cold(mapping);
1064                 if (!page) {
1065                         desc->error = -ENOMEM;
1066                         goto out;
1067                 }
1068                 error = add_to_page_cache_lru(page, mapping,
1069                                                 index, GFP_KERNEL);
1070                 if (error) {
1071                         page_cache_release(page);
1072                         if (error == -EEXIST)
1073                                 goto find_page;
1074                         desc->error = error;
1075                         goto out;
1076                 }
1077                 goto readpage;
1078         }
1079
1080 out:
1081         ra->prev_pos = prev_index;
1082         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1083         ra->prev_pos |= prev_offset;
1084
1085         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1086         if (filp)
1087                 file_accessed(filp);
1088 }
1089
1090 int file_read_actor(read_descriptor_t *desc, struct page *page,
1091                         unsigned long offset, unsigned long size)
1092 {
1093         char *kaddr;
1094         unsigned long left, count = desc->count;
1095
1096         if (size > count)
1097                 size = count;
1098
1099         /*
1100          * Faults on the destination of a read are common, so do it before
1101          * taking the kmap.
1102          */
1103         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1104                 kaddr = kmap_atomic(page, KM_USER0);
1105                 left = __copy_to_user_inatomic(desc->arg.buf,
1106                                                 kaddr + offset, size);
1107                 kunmap_atomic(kaddr, KM_USER0);
1108                 if (left == 0)
1109                         goto success;
1110         }
1111
1112         /* Do it the slow way */
1113         kaddr = kmap(page);
1114         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1115         kunmap(page);
1116
1117         if (left) {
1118                 size -= left;
1119                 desc->error = -EFAULT;
1120         }
1121 success:
1122         desc->count = count - size;
1123         desc->written += size;
1124         desc->arg.buf += size;
1125         return size;
1126 }
1127
1128 /*
1129  * Performs necessary checks before doing a write
1130  * @iov:        io vector request
1131  * @nr_segs:    number of segments in the iovec
1132  * @count:      number of bytes to write
1133  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1134  *
1135  * Adjust number of segments and amount of bytes to write (nr_segs should be
1136  * properly initialized first). Returns appropriate error code that caller
1137  * should return or zero in case that write should be allowed.
1138  */
1139 int generic_segment_checks(const struct iovec *iov,
1140                         unsigned long *nr_segs, size_t *count, int access_flags)
1141 {
1142         unsigned long   seg;
1143         size_t cnt = 0;
1144         for (seg = 0; seg < *nr_segs; seg++) {
1145                 const struct iovec *iv = &iov[seg];
1146
1147                 /*
1148                  * If any segment has a negative length, or the cumulative
1149                  * length ever wraps negative then return -EINVAL.
1150                  */
1151                 cnt += iv->iov_len;
1152                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1153                         return -EINVAL;
1154                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1155                         continue;
1156                 if (seg == 0)
1157                         return -EFAULT;
1158                 *nr_segs = seg;
1159                 cnt -= iv->iov_len;     /* This segment is no good */
1160                 break;
1161         }
1162         *count = cnt;
1163         return 0;
1164 }
1165 EXPORT_SYMBOL(generic_segment_checks);
1166
1167 /**
1168  * generic_file_aio_read - generic filesystem read routine
1169  * @iocb:       kernel I/O control block
1170  * @iov:        io vector request
1171  * @nr_segs:    number of segments in the iovec
1172  * @pos:        current file position
1173  *
1174  * This is the "read()" routine for all filesystems
1175  * that can use the page cache directly.
1176  */
1177 ssize_t
1178 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1179                 unsigned long nr_segs, loff_t pos)
1180 {
1181         struct file *filp = iocb->ki_filp;
1182         ssize_t retval;
1183         unsigned long seg;
1184         size_t count;
1185         loff_t *ppos = &iocb->ki_pos;
1186
1187         count = 0;
1188         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1189         if (retval)
1190                 return retval;
1191
1192         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1193         if (filp->f_flags & O_DIRECT) {
1194                 loff_t size;
1195                 struct address_space *mapping;
1196                 struct inode *inode;
1197
1198                 mapping = filp->f_mapping;
1199                 inode = mapping->host;
1200                 retval = 0;
1201                 if (!count)
1202                         goto out; /* skip atime */
1203                 size = i_size_read(inode);
1204                 if (pos < size) {
1205                         retval = filemap_write_and_wait(mapping);
1206                         if (!retval) {
1207                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1208                                                         iov, pos, nr_segs);
1209                         }
1210                         if (retval > 0)
1211                                 *ppos = pos + retval;
1212                 }
1213                 if (likely(retval != 0)) {
1214                         file_accessed(filp);
1215                         goto out;
1216                 }
1217         }
1218
1219         retval = 0;
1220         if (count) {
1221                 for (seg = 0; seg < nr_segs; seg++) {
1222                         read_descriptor_t desc;
1223
1224                         desc.written = 0;
1225                         desc.arg.buf = iov[seg].iov_base;
1226                         desc.count = iov[seg].iov_len;
1227                         if (desc.count == 0)
1228                                 continue;
1229                         desc.error = 0;
1230                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1231                         retval += desc.written;
1232                         if (desc.error) {
1233                                 retval = retval ?: desc.error;
1234                                 break;
1235                         }
1236                         if (desc.count > 0)
1237                                 break;
1238                 }
1239         }
1240 out:
1241         return retval;
1242 }
1243 EXPORT_SYMBOL(generic_file_aio_read);
1244
1245 static ssize_t
1246 do_readahead(struct address_space *mapping, struct file *filp,
1247              pgoff_t index, unsigned long nr)
1248 {
1249         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1250                 return -EINVAL;
1251
1252         force_page_cache_readahead(mapping, filp, index,
1253                                         max_sane_readahead(nr));
1254         return 0;
1255 }
1256
1257 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1258 {
1259         ssize_t ret;
1260         struct file *file;
1261
1262         ret = -EBADF;
1263         file = fget(fd);
1264         if (file) {
1265                 if (file->f_mode & FMODE_READ) {
1266                         struct address_space *mapping = file->f_mapping;
1267                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1268                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1269                         unsigned long len = end - start + 1;
1270                         ret = do_readahead(mapping, file, start, len);
1271                 }
1272                 fput(file);
1273         }
1274         return ret;
1275 }
1276
1277 #ifdef CONFIG_MMU
1278 /**
1279  * page_cache_read - adds requested page to the page cache if not already there
1280  * @file:       file to read
1281  * @offset:     page index
1282  *
1283  * This adds the requested page to the page cache if it isn't already there,
1284  * and schedules an I/O to read in its contents from disk.
1285  */
1286 static int page_cache_read(struct file *file, pgoff_t offset)
1287 {
1288         struct address_space *mapping = file->f_mapping;
1289         struct page *page; 
1290         int ret;
1291
1292         do {
1293                 page = page_cache_alloc_cold(mapping);
1294                 if (!page)
1295                         return -ENOMEM;
1296
1297                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1298                 if (ret == 0)
1299                         ret = mapping->a_ops->readpage(file, page);
1300                 else if (ret == -EEXIST)
1301                         ret = 0; /* losing race to add is OK */
1302
1303                 page_cache_release(page);
1304
1305         } while (ret == AOP_TRUNCATED_PAGE);
1306                 
1307         return ret;
1308 }
1309
1310 #define MMAP_LOTSAMISS  (100)
1311
1312 /**
1313  * filemap_fault - read in file data for page fault handling
1314  * @vma:        vma in which the fault was taken
1315  * @vmf:        struct vm_fault containing details of the fault
1316  *
1317  * filemap_fault() is invoked via the vma operations vector for a
1318  * mapped memory region to read in file data during a page fault.
1319  *
1320  * The goto's are kind of ugly, but this streamlines the normal case of having
1321  * it in the page cache, and handles the special cases reasonably without
1322  * having a lot of duplicated code.
1323  */
1324 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1325 {
1326         int error;
1327         struct file *file = vma->vm_file;
1328         struct address_space *mapping = file->f_mapping;
1329         struct file_ra_state *ra = &file->f_ra;
1330         struct inode *inode = mapping->host;
1331         struct page *page;
1332         pgoff_t size;
1333         int did_readaround = 0;
1334         int ret = 0;
1335
1336         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1337         if (vmf->pgoff >= size)
1338                 return VM_FAULT_SIGBUS;
1339
1340         /* If we don't want any read-ahead, don't bother */
1341         if (VM_RandomReadHint(vma))
1342                 goto no_cached_page;
1343
1344         /*
1345          * Do we have something in the page cache already?
1346          */
1347 retry_find:
1348         page = find_lock_page(mapping, vmf->pgoff);
1349         /*
1350          * For sequential accesses, we use the generic readahead logic.
1351          */
1352         if (VM_SequentialReadHint(vma)) {
1353                 if (!page) {
1354                         page_cache_sync_readahead(mapping, ra, file,
1355                                                            vmf->pgoff, 1);
1356                         page = find_lock_page(mapping, vmf->pgoff);
1357                         if (!page)
1358                                 goto no_cached_page;
1359                 }
1360                 if (PageReadahead(page)) {
1361                         page_cache_async_readahead(mapping, ra, file, page,
1362                                                            vmf->pgoff, 1);
1363                 }
1364         }
1365
1366         if (!page) {
1367                 unsigned long ra_pages;
1368
1369                 ra->mmap_miss++;
1370
1371                 /*
1372                  * Do we miss much more than hit in this file? If so,
1373                  * stop bothering with read-ahead. It will only hurt.
1374                  */
1375                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1376                         goto no_cached_page;
1377
1378                 /*
1379                  * To keep the pgmajfault counter straight, we need to
1380                  * check did_readaround, as this is an inner loop.
1381                  */
1382                 if (!did_readaround) {
1383                         ret = VM_FAULT_MAJOR;
1384                         count_vm_event(PGMAJFAULT);
1385                 }
1386                 did_readaround = 1;
1387                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1388                 if (ra_pages) {
1389                         pgoff_t start = 0;
1390
1391                         if (vmf->pgoff > ra_pages / 2)
1392                                 start = vmf->pgoff - ra_pages / 2;
1393                         do_page_cache_readahead(mapping, file, start, ra_pages);
1394                 }
1395                 page = find_lock_page(mapping, vmf->pgoff);
1396                 if (!page)
1397                         goto no_cached_page;
1398         }
1399
1400         if (!did_readaround)
1401                 ra->mmap_miss--;
1402
1403         /*
1404          * We have a locked page in the page cache, now we need to check
1405          * that it's up-to-date. If not, it is going to be due to an error.
1406          */
1407         if (unlikely(!PageUptodate(page)))
1408                 goto page_not_uptodate;
1409
1410         /* Must recheck i_size under page lock */
1411         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1412         if (unlikely(vmf->pgoff >= size)) {
1413                 unlock_page(page);
1414                 page_cache_release(page);
1415                 return VM_FAULT_SIGBUS;
1416         }
1417
1418         /*
1419          * Found the page and have a reference on it.
1420          */
1421         mark_page_accessed(page);
1422         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1423         vmf->page = page;
1424         return ret | VM_FAULT_LOCKED;
1425
1426 no_cached_page:
1427         /*
1428          * We're only likely to ever get here if MADV_RANDOM is in
1429          * effect.
1430          */
1431         error = page_cache_read(file, vmf->pgoff);
1432
1433         /*
1434          * The page we want has now been added to the page cache.
1435          * In the unlikely event that someone removed it in the
1436          * meantime, we'll just come back here and read it again.
1437          */
1438         if (error >= 0)
1439                 goto retry_find;
1440
1441         /*
1442          * An error return from page_cache_read can result if the
1443          * system is low on memory, or a problem occurs while trying
1444          * to schedule I/O.
1445          */
1446         if (error == -ENOMEM)
1447                 return VM_FAULT_OOM;
1448         return VM_FAULT_SIGBUS;
1449
1450 page_not_uptodate:
1451         /* IO error path */
1452         if (!did_readaround) {
1453                 ret = VM_FAULT_MAJOR;
1454                 count_vm_event(PGMAJFAULT);
1455         }
1456
1457         /*
1458          * Umm, take care of errors if the page isn't up-to-date.
1459          * Try to re-read it _once_. We do this synchronously,
1460          * because there really aren't any performance issues here
1461          * and we need to check for errors.
1462          */
1463         ClearPageError(page);
1464         error = mapping->a_ops->readpage(file, page);
1465         if (!error) {
1466                 wait_on_page_locked(page);
1467                 if (!PageUptodate(page))
1468                         error = -EIO;
1469         }
1470         page_cache_release(page);
1471
1472         if (!error || error == AOP_TRUNCATED_PAGE)
1473                 goto retry_find;
1474
1475         /* Things didn't work out. Return zero to tell the mm layer so. */
1476         shrink_readahead_size_eio(file, ra);
1477         return VM_FAULT_SIGBUS;
1478 }
1479 EXPORT_SYMBOL(filemap_fault);
1480
1481 struct vm_operations_struct generic_file_vm_ops = {
1482         .fault          = filemap_fault,
1483 };
1484
1485 /* This is used for a general mmap of a disk file */
1486
1487 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1488 {
1489         struct address_space *mapping = file->f_mapping;
1490
1491         if (!mapping->a_ops->readpage)
1492                 return -ENOEXEC;
1493         file_accessed(file);
1494         vma->vm_ops = &generic_file_vm_ops;
1495         vma->vm_flags |= VM_CAN_NONLINEAR;
1496         return 0;
1497 }
1498
1499 /*
1500  * This is for filesystems which do not implement ->writepage.
1501  */
1502 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1503 {
1504         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1505                 return -EINVAL;
1506         return generic_file_mmap(file, vma);
1507 }
1508 #else
1509 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1510 {
1511         return -ENOSYS;
1512 }
1513 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1514 {
1515         return -ENOSYS;
1516 }
1517 #endif /* CONFIG_MMU */
1518
1519 EXPORT_SYMBOL(generic_file_mmap);
1520 EXPORT_SYMBOL(generic_file_readonly_mmap);
1521
1522 static struct page *__read_cache_page(struct address_space *mapping,
1523                                 pgoff_t index,
1524                                 int (*filler)(void *,struct page*),
1525                                 void *data)
1526 {
1527         struct page *page;
1528         int err;
1529 repeat:
1530         page = find_get_page(mapping, index);
1531         if (!page) {
1532                 page = page_cache_alloc_cold(mapping);
1533                 if (!page)
1534                         return ERR_PTR(-ENOMEM);
1535                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1536                 if (unlikely(err)) {
1537                         page_cache_release(page);
1538                         if (err == -EEXIST)
1539                                 goto repeat;
1540                         /* Presumably ENOMEM for radix tree node */
1541                         return ERR_PTR(err);
1542                 }
1543                 err = filler(data, page);
1544                 if (err < 0) {
1545                         page_cache_release(page);
1546                         page = ERR_PTR(err);
1547                 }
1548         }
1549         return page;
1550 }
1551
1552 /**
1553  * read_cache_page_async - read into page cache, fill it if needed
1554  * @mapping:    the page's address_space
1555  * @index:      the page index
1556  * @filler:     function to perform the read
1557  * @data:       destination for read data
1558  *
1559  * Same as read_cache_page, but don't wait for page to become unlocked
1560  * after submitting it to the filler.
1561  *
1562  * Read into the page cache. If a page already exists, and PageUptodate() is
1563  * not set, try to fill the page but don't wait for it to become unlocked.
1564  *
1565  * If the page does not get brought uptodate, return -EIO.
1566  */
1567 struct page *read_cache_page_async(struct address_space *mapping,
1568                                 pgoff_t index,
1569                                 int (*filler)(void *,struct page*),
1570                                 void *data)
1571 {
1572         struct page *page;
1573         int err;
1574
1575 retry:
1576         page = __read_cache_page(mapping, index, filler, data);
1577         if (IS_ERR(page))
1578                 return page;
1579         if (PageUptodate(page))
1580                 goto out;
1581
1582         lock_page(page);
1583         if (!page->mapping) {
1584                 unlock_page(page);
1585                 page_cache_release(page);
1586                 goto retry;
1587         }
1588         if (PageUptodate(page)) {
1589                 unlock_page(page);
1590                 goto out;
1591         }
1592         err = filler(data, page);
1593         if (err < 0) {
1594                 page_cache_release(page);
1595                 return ERR_PTR(err);
1596         }
1597 out:
1598         mark_page_accessed(page);
1599         return page;
1600 }
1601 EXPORT_SYMBOL(read_cache_page_async);
1602
1603 /**
1604  * read_cache_page - read into page cache, fill it if needed
1605  * @mapping:    the page's address_space
1606  * @index:      the page index
1607  * @filler:     function to perform the read
1608  * @data:       destination for read data
1609  *
1610  * Read into the page cache. If a page already exists, and PageUptodate() is
1611  * not set, try to fill the page then wait for it to become unlocked.
1612  *
1613  * If the page does not get brought uptodate, return -EIO.
1614  */
1615 struct page *read_cache_page(struct address_space *mapping,
1616                                 pgoff_t index,
1617                                 int (*filler)(void *,struct page*),
1618                                 void *data)
1619 {
1620         struct page *page;
1621
1622         page = read_cache_page_async(mapping, index, filler, data);
1623         if (IS_ERR(page))
1624                 goto out;
1625         wait_on_page_locked(page);
1626         if (!PageUptodate(page)) {
1627                 page_cache_release(page);
1628                 page = ERR_PTR(-EIO);
1629         }
1630  out:
1631         return page;
1632 }
1633 EXPORT_SYMBOL(read_cache_page);
1634
1635 /*
1636  * The logic we want is
1637  *
1638  *      if suid or (sgid and xgrp)
1639  *              remove privs
1640  */
1641 int should_remove_suid(struct dentry *dentry)
1642 {
1643         mode_t mode = dentry->d_inode->i_mode;
1644         int kill = 0;
1645
1646         /* suid always must be killed */
1647         if (unlikely(mode & S_ISUID))
1648                 kill = ATTR_KILL_SUID;
1649
1650         /*
1651          * sgid without any exec bits is just a mandatory locking mark; leave
1652          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1653          */
1654         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1655                 kill |= ATTR_KILL_SGID;
1656
1657         if (unlikely(kill && !capable(CAP_FSETID)))
1658                 return kill;
1659
1660         return 0;
1661 }
1662 EXPORT_SYMBOL(should_remove_suid);
1663
1664 static int __remove_suid(struct dentry *dentry, int kill)
1665 {
1666         struct iattr newattrs;
1667
1668         newattrs.ia_valid = ATTR_FORCE | kill;
1669         return notify_change(dentry, &newattrs);
1670 }
1671
1672 int remove_suid(struct dentry *dentry)
1673 {
1674         int killsuid = should_remove_suid(dentry);
1675         int killpriv = security_inode_need_killpriv(dentry);
1676         int error = 0;
1677
1678         if (killpriv < 0)
1679                 return killpriv;
1680         if (killpriv)
1681                 error = security_inode_killpriv(dentry);
1682         if (!error && killsuid)
1683                 error = __remove_suid(dentry, killsuid);
1684
1685         return error;
1686 }
1687 EXPORT_SYMBOL(remove_suid);
1688
1689 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1690                         const struct iovec *iov, size_t base, size_t bytes)
1691 {
1692         size_t copied = 0, left = 0;
1693
1694         while (bytes) {
1695                 char __user *buf = iov->iov_base + base;
1696                 int copy = min(bytes, iov->iov_len - base);
1697
1698                 base = 0;
1699                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1700                 copied += copy;
1701                 bytes -= copy;
1702                 vaddr += copy;
1703                 iov++;
1704
1705                 if (unlikely(left))
1706                         break;
1707         }
1708         return copied - left;
1709 }
1710
1711 /*
1712  * Copy as much as we can into the page and return the number of bytes which
1713  * were sucessfully copied.  If a fault is encountered then return the number of
1714  * bytes which were copied.
1715  */
1716 size_t iov_iter_copy_from_user_atomic(struct page *page,
1717                 struct iov_iter *i, unsigned long offset, size_t bytes)
1718 {
1719         char *kaddr;
1720         size_t copied;
1721
1722         BUG_ON(!in_atomic());
1723         kaddr = kmap_atomic(page, KM_USER0);
1724         if (likely(i->nr_segs == 1)) {
1725                 int left;
1726                 char __user *buf = i->iov->iov_base + i->iov_offset;
1727                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1728                                                         buf, bytes);
1729                 copied = bytes - left;
1730         } else {
1731                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1732                                                 i->iov, i->iov_offset, bytes);
1733         }
1734         kunmap_atomic(kaddr, KM_USER0);
1735
1736         return copied;
1737 }
1738 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1739
1740 /*
1741  * This has the same sideeffects and return value as
1742  * iov_iter_copy_from_user_atomic().
1743  * The difference is that it attempts to resolve faults.
1744  * Page must not be locked.
1745  */
1746 size_t iov_iter_copy_from_user(struct page *page,
1747                 struct iov_iter *i, unsigned long offset, size_t bytes)
1748 {
1749         char *kaddr;
1750         size_t copied;
1751
1752         kaddr = kmap(page);
1753         if (likely(i->nr_segs == 1)) {
1754                 int left;
1755                 char __user *buf = i->iov->iov_base + i->iov_offset;
1756                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1757                 copied = bytes - left;
1758         } else {
1759                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1760                                                 i->iov, i->iov_offset, bytes);
1761         }
1762         kunmap(page);
1763         return copied;
1764 }
1765 EXPORT_SYMBOL(iov_iter_copy_from_user);
1766
1767 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1768 {
1769         BUG_ON(i->count < bytes);
1770
1771         if (likely(i->nr_segs == 1)) {
1772                 i->iov_offset += bytes;
1773                 i->count -= bytes;
1774         } else {
1775                 const struct iovec *iov = i->iov;
1776                 size_t base = i->iov_offset;
1777
1778                 /*
1779                  * The !iov->iov_len check ensures we skip over unlikely
1780                  * zero-length segments (without overruning the iovec).
1781                  */
1782                 while (bytes || unlikely(!iov->iov_len && i->count)) {
1783                         int copy;
1784
1785                         copy = min(bytes, iov->iov_len - base);
1786                         BUG_ON(!i->count || i->count < copy);
1787                         i->count -= copy;
1788                         bytes -= copy;
1789                         base += copy;
1790                         if (iov->iov_len == base) {
1791                                 iov++;
1792                                 base = 0;
1793                         }
1794                 }
1795                 i->iov = iov;
1796                 i->iov_offset = base;
1797         }
1798 }
1799 EXPORT_SYMBOL(iov_iter_advance);
1800
1801 /*
1802  * Fault in the first iovec of the given iov_iter, to a maximum length
1803  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1804  * accessed (ie. because it is an invalid address).
1805  *
1806  * writev-intensive code may want this to prefault several iovecs -- that
1807  * would be possible (callers must not rely on the fact that _only_ the
1808  * first iovec will be faulted with the current implementation).
1809  */
1810 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1811 {
1812         char __user *buf = i->iov->iov_base + i->iov_offset;
1813         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1814         return fault_in_pages_readable(buf, bytes);
1815 }
1816 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1817
1818 /*
1819  * Return the count of just the current iov_iter segment.
1820  */
1821 size_t iov_iter_single_seg_count(struct iov_iter *i)
1822 {
1823         const struct iovec *iov = i->iov;
1824         if (i->nr_segs == 1)
1825                 return i->count;
1826         else
1827                 return min(i->count, iov->iov_len - i->iov_offset);
1828 }
1829 EXPORT_SYMBOL(iov_iter_single_seg_count);
1830
1831 /*
1832  * Performs necessary checks before doing a write
1833  *
1834  * Can adjust writing position or amount of bytes to write.
1835  * Returns appropriate error code that caller should return or
1836  * zero in case that write should be allowed.
1837  */
1838 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1839 {
1840         struct inode *inode = file->f_mapping->host;
1841         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1842
1843         if (unlikely(*pos < 0))
1844                 return -EINVAL;
1845
1846         if (!isblk) {
1847                 /* FIXME: this is for backwards compatibility with 2.4 */
1848                 if (file->f_flags & O_APPEND)
1849                         *pos = i_size_read(inode);
1850
1851                 if (limit != RLIM_INFINITY) {
1852                         if (*pos >= limit) {
1853                                 send_sig(SIGXFSZ, current, 0);
1854                                 return -EFBIG;
1855                         }
1856                         if (*count > limit - (typeof(limit))*pos) {
1857                                 *count = limit - (typeof(limit))*pos;
1858                         }
1859                 }
1860         }
1861
1862         /*
1863          * LFS rule
1864          */
1865         if (unlikely(*pos + *count > MAX_NON_LFS &&
1866                                 !(file->f_flags & O_LARGEFILE))) {
1867                 if (*pos >= MAX_NON_LFS) {
1868                         return -EFBIG;
1869                 }
1870                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1871                         *count = MAX_NON_LFS - (unsigned long)*pos;
1872                 }
1873         }
1874
1875         /*
1876          * Are we about to exceed the fs block limit ?
1877          *
1878          * If we have written data it becomes a short write.  If we have
1879          * exceeded without writing data we send a signal and return EFBIG.
1880          * Linus frestrict idea will clean these up nicely..
1881          */
1882         if (likely(!isblk)) {
1883                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1884                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1885                                 return -EFBIG;
1886                         }
1887                         /* zero-length writes at ->s_maxbytes are OK */
1888                 }
1889
1890                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1891                         *count = inode->i_sb->s_maxbytes - *pos;
1892         } else {
1893 #ifdef CONFIG_BLOCK
1894                 loff_t isize;
1895                 if (bdev_read_only(I_BDEV(inode)))
1896                         return -EPERM;
1897                 isize = i_size_read(inode);
1898                 if (*pos >= isize) {
1899                         if (*count || *pos > isize)
1900                                 return -ENOSPC;
1901                 }
1902
1903                 if (*pos + *count > isize)
1904                         *count = isize - *pos;
1905 #else
1906                 return -EPERM;
1907 #endif
1908         }
1909         return 0;
1910 }
1911 EXPORT_SYMBOL(generic_write_checks);
1912
1913 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1914                                 loff_t pos, unsigned len, unsigned flags,
1915                                 struct page **pagep, void **fsdata)
1916 {
1917         const struct address_space_operations *aops = mapping->a_ops;
1918
1919         if (aops->write_begin) {
1920                 return aops->write_begin(file, mapping, pos, len, flags,
1921                                                         pagep, fsdata);
1922         } else {
1923                 int ret;
1924                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1925                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1926                 struct inode *inode = mapping->host;
1927                 struct page *page;
1928 again:
1929                 page = __grab_cache_page(mapping, index);
1930                 *pagep = page;
1931                 if (!page)
1932                         return -ENOMEM;
1933
1934                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1935                         /*
1936                          * There is no way to resolve a short write situation
1937                          * for a !Uptodate page (except by double copying in
1938                          * the caller done by generic_perform_write_2copy).
1939                          *
1940                          * Instead, we have to bring it uptodate here.
1941                          */
1942                         ret = aops->readpage(file, page);
1943                         page_cache_release(page);
1944                         if (ret) {
1945                                 if (ret == AOP_TRUNCATED_PAGE)
1946                                         goto again;
1947                                 return ret;
1948                         }
1949                         goto again;
1950                 }
1951
1952                 ret = aops->prepare_write(file, page, offset, offset+len);
1953                 if (ret) {
1954                         unlock_page(page);
1955                         page_cache_release(page);
1956                         if (pos + len > inode->i_size)
1957                                 vmtruncate(inode, inode->i_size);
1958                 }
1959                 return ret;
1960         }
1961 }
1962 EXPORT_SYMBOL(pagecache_write_begin);
1963
1964 int pagecache_write_end(struct file *file, struct address_space *mapping,
1965                                 loff_t pos, unsigned len, unsigned copied,
1966                                 struct page *page, void *fsdata)
1967 {
1968         const struct address_space_operations *aops = mapping->a_ops;
1969         int ret;
1970
1971         if (aops->write_end) {
1972                 mark_page_accessed(page);
1973                 ret = aops->write_end(file, mapping, pos, len, copied,
1974                                                         page, fsdata);
1975         } else {
1976                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1977                 struct inode *inode = mapping->host;
1978
1979                 flush_dcache_page(page);
1980                 ret = aops->commit_write(file, page, offset, offset+len);
1981                 unlock_page(page);
1982                 mark_page_accessed(page);
1983                 page_cache_release(page);
1984
1985                 if (ret < 0) {
1986                         if (pos + len > inode->i_size)
1987                                 vmtruncate(inode, inode->i_size);
1988                 } else if (ret > 0)
1989                         ret = min_t(size_t, copied, ret);
1990                 else
1991                         ret = copied;
1992         }
1993
1994         return ret;
1995 }
1996 EXPORT_SYMBOL(pagecache_write_end);
1997
1998 ssize_t
1999 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2000                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2001                 size_t count, size_t ocount)
2002 {
2003         struct file     *file = iocb->ki_filp;
2004         struct address_space *mapping = file->f_mapping;
2005         struct inode    *inode = mapping->host;
2006         ssize_t         written;
2007         size_t          write_len;
2008         pgoff_t         end;
2009
2010         if (count != ocount)
2011                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2012
2013         /*
2014          * Unmap all mmappings of the file up-front.
2015          *
2016          * This will cause any pte dirty bits to be propagated into the
2017          * pageframes for the subsequent filemap_write_and_wait().
2018          */
2019         write_len = iov_length(iov, *nr_segs);
2020         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2021         if (mapping_mapped(mapping))
2022                 unmap_mapping_range(mapping, pos, write_len, 0);
2023
2024         written = filemap_write_and_wait(mapping);
2025         if (written)
2026                 goto out;
2027
2028         /*
2029          * After a write we want buffered reads to be sure to go to disk to get
2030          * the new data.  We invalidate clean cached page from the region we're
2031          * about to write.  We do this *before* the write so that we can return
2032          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2033          */
2034         if (mapping->nrpages) {
2035                 written = invalidate_inode_pages2_range(mapping,
2036                                         pos >> PAGE_CACHE_SHIFT, end);
2037                 if (written)
2038                         goto out;
2039         }
2040
2041         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2042
2043         /*
2044          * Finally, try again to invalidate clean pages which might have been
2045          * cached by non-direct readahead, or faulted in by get_user_pages()
2046          * if the source of the write was an mmap'ed region of the file
2047          * we're writing.  Either one is a pretty crazy thing to do,
2048          * so we don't support it 100%.  If this invalidation
2049          * fails, tough, the write still worked...
2050          */
2051         if (mapping->nrpages) {
2052                 invalidate_inode_pages2_range(mapping,
2053                                               pos >> PAGE_CACHE_SHIFT, end);
2054         }
2055
2056         if (written > 0) {
2057                 loff_t end = pos + written;
2058                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2059                         i_size_write(inode,  end);
2060                         mark_inode_dirty(inode);
2061                 }
2062                 *ppos = end;
2063         }
2064
2065         /*
2066          * Sync the fs metadata but not the minor inode changes and
2067          * of course not the data as we did direct DMA for the IO.
2068          * i_mutex is held, which protects generic_osync_inode() from
2069          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2070          */
2071 out:
2072         if ((written >= 0 || written == -EIOCBQUEUED) &&
2073             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2074                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2075                 if (err < 0)
2076                         written = err;
2077         }
2078         return written;
2079 }
2080 EXPORT_SYMBOL(generic_file_direct_write);
2081
2082 /*
2083  * Find or create a page at the given pagecache position. Return the locked
2084  * page. This function is specifically for buffered writes.
2085  */
2086 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2087 {
2088         int status;
2089         struct page *page;
2090 repeat:
2091         page = find_lock_page(mapping, index);
2092         if (likely(page))
2093                 return page;
2094
2095         page = page_cache_alloc(mapping);
2096         if (!page)
2097                 return NULL;
2098         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2099         if (unlikely(status)) {
2100                 page_cache_release(page);
2101                 if (status == -EEXIST)
2102                         goto repeat;
2103                 return NULL;
2104         }
2105         return page;
2106 }
2107 EXPORT_SYMBOL(__grab_cache_page);
2108
2109 static ssize_t generic_perform_write_2copy(struct file *file,
2110                                 struct iov_iter *i, loff_t pos)
2111 {
2112         struct address_space *mapping = file->f_mapping;
2113         const struct address_space_operations *a_ops = mapping->a_ops;
2114         struct inode *inode = mapping->host;
2115         long status = 0;
2116         ssize_t written = 0;
2117
2118         do {
2119                 struct page *src_page;
2120                 struct page *page;
2121                 pgoff_t index;          /* Pagecache index for current page */
2122                 unsigned long offset;   /* Offset into pagecache page */
2123                 unsigned long bytes;    /* Bytes to write to page */
2124                 size_t copied;          /* Bytes copied from user */
2125
2126                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2127                 index = pos >> PAGE_CACHE_SHIFT;
2128                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2129                                                 iov_iter_count(i));
2130
2131                 /*
2132                  * a non-NULL src_page indicates that we're doing the
2133                  * copy via get_user_pages and kmap.
2134                  */
2135                 src_page = NULL;
2136
2137                 /*
2138                  * Bring in the user page that we will copy from _first_.
2139                  * Otherwise there's a nasty deadlock on copying from the
2140                  * same page as we're writing to, without it being marked
2141                  * up-to-date.
2142                  *
2143                  * Not only is this an optimisation, but it is also required
2144                  * to check that the address is actually valid, when atomic
2145                  * usercopies are used, below.
2146                  */
2147                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2148                         status = -EFAULT;
2149                         break;
2150                 }
2151
2152                 page = __grab_cache_page(mapping, index);
2153                 if (!page) {
2154                         status = -ENOMEM;
2155                         break;
2156                 }
2157
2158                 /*
2159                  * non-uptodate pages cannot cope with short copies, and we
2160                  * cannot take a pagefault with the destination page locked.
2161                  * So pin the source page to copy it.
2162                  */
2163                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2164                         unlock_page(page);
2165
2166                         src_page = alloc_page(GFP_KERNEL);
2167                         if (!src_page) {
2168                                 page_cache_release(page);
2169                                 status = -ENOMEM;
2170                                 break;
2171                         }
2172
2173                         /*
2174                          * Cannot get_user_pages with a page locked for the
2175                          * same reason as we can't take a page fault with a
2176                          * page locked (as explained below).
2177                          */
2178                         copied = iov_iter_copy_from_user(src_page, i,
2179                                                                 offset, bytes);
2180                         if (unlikely(copied == 0)) {
2181                                 status = -EFAULT;
2182                                 page_cache_release(page);
2183                                 page_cache_release(src_page);
2184                                 break;
2185                         }
2186                         bytes = copied;
2187
2188                         lock_page(page);
2189                         /*
2190                          * Can't handle the page going uptodate here, because
2191                          * that means we would use non-atomic usercopies, which
2192                          * zero out the tail of the page, which can cause
2193                          * zeroes to become transiently visible. We could just
2194                          * use a non-zeroing copy, but the APIs aren't too
2195                          * consistent.
2196                          */
2197                         if (unlikely(!page->mapping || PageUptodate(page))) {
2198                                 unlock_page(page);
2199                                 page_cache_release(page);
2200                                 page_cache_release(src_page);
2201                                 continue;
2202                         }
2203                 }
2204
2205                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2206                 if (unlikely(status))
2207                         goto fs_write_aop_error;
2208
2209                 if (!src_page) {
2210                         /*
2211                          * Must not enter the pagefault handler here, because
2212                          * we hold the page lock, so we might recursively
2213                          * deadlock on the same lock, or get an ABBA deadlock
2214                          * against a different lock, or against the mmap_sem
2215                          * (which nests outside the page lock).  So increment
2216                          * preempt count, and use _atomic usercopies.
2217                          *
2218                          * The page is uptodate so we are OK to encounter a
2219                          * short copy: if unmodified parts of the page are
2220                          * marked dirty and written out to disk, it doesn't
2221                          * really matter.
2222                          */
2223                         pagefault_disable();
2224                         copied = iov_iter_copy_from_user_atomic(page, i,
2225                                                                 offset, bytes);
2226                         pagefault_enable();
2227                 } else {
2228                         void *src, *dst;
2229                         src = kmap_atomic(src_page, KM_USER0);
2230                         dst = kmap_atomic(page, KM_USER1);
2231                         memcpy(dst + offset, src + offset, bytes);
2232                         kunmap_atomic(dst, KM_USER1);
2233                         kunmap_atomic(src, KM_USER0);
2234                         copied = bytes;
2235                 }
2236                 flush_dcache_page(page);
2237
2238                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2239                 if (unlikely(status < 0))
2240                         goto fs_write_aop_error;
2241                 if (unlikely(status > 0)) /* filesystem did partial write */
2242                         copied = min_t(size_t, copied, status);
2243
2244                 unlock_page(page);
2245                 mark_page_accessed(page);
2246                 page_cache_release(page);
2247                 if (src_page)
2248                         page_cache_release(src_page);
2249
2250                 iov_iter_advance(i, copied);
2251                 pos += copied;
2252                 written += copied;
2253
2254                 balance_dirty_pages_ratelimited(mapping);
2255                 cond_resched();
2256                 continue;
2257
2258 fs_write_aop_error:
2259                 unlock_page(page);
2260                 page_cache_release(page);
2261                 if (src_page)
2262                         page_cache_release(src_page);
2263
2264                 /*
2265                  * prepare_write() may have instantiated a few blocks
2266                  * outside i_size.  Trim these off again. Don't need
2267                  * i_size_read because we hold i_mutex.
2268                  */
2269                 if (pos + bytes > inode->i_size)
2270                         vmtruncate(inode, inode->i_size);
2271                 break;
2272         } while (iov_iter_count(i));
2273
2274         return written ? written : status;
2275 }
2276
2277 static ssize_t generic_perform_write(struct file *file,
2278                                 struct iov_iter *i, loff_t pos)
2279 {
2280         struct address_space *mapping = file->f_mapping;
2281         const struct address_space_operations *a_ops = mapping->a_ops;
2282         long status = 0;
2283         ssize_t written = 0;
2284         unsigned int flags = 0;
2285
2286         /*
2287          * Copies from kernel address space cannot fail (NFSD is a big user).
2288          */
2289         if (segment_eq(get_fs(), KERNEL_DS))
2290                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2291
2292         do {
2293                 struct page *page;
2294                 pgoff_t index;          /* Pagecache index for current page */
2295                 unsigned long offset;   /* Offset into pagecache page */
2296                 unsigned long bytes;    /* Bytes to write to page */
2297                 size_t copied;          /* Bytes copied from user */
2298                 void *fsdata;
2299
2300                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2301                 index = pos >> PAGE_CACHE_SHIFT;
2302                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2303                                                 iov_iter_count(i));
2304
2305 again:
2306
2307                 /*
2308                  * Bring in the user page that we will copy from _first_.
2309                  * Otherwise there's a nasty deadlock on copying from the
2310                  * same page as we're writing to, without it being marked
2311                  * up-to-date.
2312                  *
2313                  * Not only is this an optimisation, but it is also required
2314                  * to check that the address is actually valid, when atomic
2315                  * usercopies are used, below.
2316                  */
2317                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2318                         status = -EFAULT;
2319                         break;
2320                 }
2321
2322                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2323                                                 &page, &fsdata);
2324                 if (unlikely(status))
2325                         break;
2326
2327                 pagefault_disable();
2328                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2329                 pagefault_enable();
2330                 flush_dcache_page(page);
2331
2332                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2333                                                 page, fsdata);
2334                 if (unlikely(status < 0))
2335                         break;
2336                 copied = status;
2337
2338                 cond_resched();
2339
2340                 iov_iter_advance(i, copied);
2341                 if (unlikely(copied == 0)) {
2342                         /*
2343                          * If we were unable to copy any data at all, we must
2344                          * fall back to a single segment length write.
2345                          *
2346                          * If we didn't fallback here, we could livelock
2347                          * because not all segments in the iov can be copied at
2348                          * once without a pagefault.
2349                          */
2350                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2351                                                 iov_iter_single_seg_count(i));
2352                         goto again;
2353                 }
2354                 pos += copied;
2355                 written += copied;
2356
2357                 balance_dirty_pages_ratelimited(mapping);
2358
2359         } while (iov_iter_count(i));
2360
2361         return written ? written : status;
2362 }
2363
2364 ssize_t
2365 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2366                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2367                 size_t count, ssize_t written)
2368 {
2369         struct file *file = iocb->ki_filp;
2370         struct address_space *mapping = file->f_mapping;
2371         const struct address_space_operations *a_ops = mapping->a_ops;
2372         struct inode *inode = mapping->host;
2373         ssize_t status;
2374         struct iov_iter i;
2375
2376         iov_iter_init(&i, iov, nr_segs, count, written);
2377         if (a_ops->write_begin)
2378                 status = generic_perform_write(file, &i, pos);
2379         else
2380                 status = generic_perform_write_2copy(file, &i, pos);
2381
2382         if (likely(status >= 0)) {
2383                 written += status;
2384                 *ppos = pos + status;
2385
2386                 /*
2387                  * For now, when the user asks for O_SYNC, we'll actually give
2388                  * O_DSYNC
2389                  */
2390                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2391                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2392                                 status = generic_osync_inode(inode, mapping,
2393                                                 OSYNC_METADATA|OSYNC_DATA);
2394                 }
2395         }
2396         
2397         /*
2398          * If we get here for O_DIRECT writes then we must have fallen through
2399          * to buffered writes (block instantiation inside i_size).  So we sync
2400          * the file data here, to try to honour O_DIRECT expectations.
2401          */
2402         if (unlikely(file->f_flags & O_DIRECT) && written)
2403                 status = filemap_write_and_wait(mapping);
2404
2405         return written ? written : status;
2406 }
2407 EXPORT_SYMBOL(generic_file_buffered_write);
2408
2409 static ssize_t
2410 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2411                                 unsigned long nr_segs, loff_t *ppos)
2412 {
2413         struct file *file = iocb->ki_filp;
2414         struct address_space * mapping = file->f_mapping;
2415         size_t ocount;          /* original count */
2416         size_t count;           /* after file limit checks */
2417         struct inode    *inode = mapping->host;
2418         loff_t          pos;
2419         ssize_t         written;
2420         ssize_t         err;
2421
2422         ocount = 0;
2423         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2424         if (err)
2425                 return err;
2426
2427         count = ocount;
2428         pos = *ppos;
2429
2430         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2431
2432         /* We can write back this queue in page reclaim */
2433         current->backing_dev_info = mapping->backing_dev_info;
2434         written = 0;
2435
2436         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2437         if (err)
2438                 goto out;
2439
2440         if (count == 0)
2441                 goto out;
2442
2443         err = remove_suid(file->f_path.dentry);
2444         if (err)
2445                 goto out;
2446
2447         file_update_time(file);
2448
2449         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2450         if (unlikely(file->f_flags & O_DIRECT)) {
2451                 loff_t endbyte;
2452                 ssize_t written_buffered;
2453
2454                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2455                                                         ppos, count, ocount);
2456                 if (written < 0 || written == count)
2457                         goto out;
2458                 /*
2459                  * direct-io write to a hole: fall through to buffered I/O
2460                  * for completing the rest of the request.
2461                  */
2462                 pos += written;
2463                 count -= written;
2464                 written_buffered = generic_file_buffered_write(iocb, iov,
2465                                                 nr_segs, pos, ppos, count,
2466                                                 written);
2467                 /*
2468                  * If generic_file_buffered_write() retuned a synchronous error
2469                  * then we want to return the number of bytes which were
2470                  * direct-written, or the error code if that was zero.  Note
2471                  * that this differs from normal direct-io semantics, which
2472                  * will return -EFOO even if some bytes were written.
2473                  */
2474                 if (written_buffered < 0) {
2475                         err = written_buffered;
2476                         goto out;
2477                 }
2478
2479                 /*
2480                  * We need to ensure that the page cache pages are written to
2481                  * disk and invalidated to preserve the expected O_DIRECT
2482                  * semantics.
2483                  */
2484                 endbyte = pos + written_buffered - written - 1;
2485                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2486                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2487                                             SYNC_FILE_RANGE_WRITE|
2488                                             SYNC_FILE_RANGE_WAIT_AFTER);
2489                 if (err == 0) {
2490                         written = written_buffered;
2491                         invalidate_mapping_pages(mapping,
2492                                                  pos >> PAGE_CACHE_SHIFT,
2493                                                  endbyte >> PAGE_CACHE_SHIFT);
2494                 } else {
2495                         /*
2496                          * We don't know how much we wrote, so just return
2497                          * the number of bytes which were direct-written
2498                          */
2499                 }
2500         } else {
2501                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2502                                 pos, ppos, count, written);
2503         }
2504 out:
2505         current->backing_dev_info = NULL;
2506         return written ? written : err;
2507 }
2508
2509 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2510                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2511 {
2512         struct file *file = iocb->ki_filp;
2513         struct address_space *mapping = file->f_mapping;
2514         struct inode *inode = mapping->host;
2515         ssize_t ret;
2516
2517         BUG_ON(iocb->ki_pos != pos);
2518
2519         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2520                         &iocb->ki_pos);
2521
2522         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2523                 ssize_t err;
2524
2525                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2526                 if (err < 0)
2527                         ret = err;
2528         }
2529         return ret;
2530 }
2531 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2532
2533 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2534                 unsigned long nr_segs, loff_t pos)
2535 {
2536         struct file *file = iocb->ki_filp;
2537         struct address_space *mapping = file->f_mapping;
2538         struct inode *inode = mapping->host;
2539         ssize_t ret;
2540
2541         BUG_ON(iocb->ki_pos != pos);
2542
2543         mutex_lock(&inode->i_mutex);
2544         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2545                         &iocb->ki_pos);
2546         mutex_unlock(&inode->i_mutex);
2547
2548         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2549                 ssize_t err;
2550
2551                 err = sync_page_range(inode, mapping, pos, ret);
2552                 if (err < 0)
2553                         ret = err;
2554         }
2555         return ret;
2556 }
2557 EXPORT_SYMBOL(generic_file_aio_write);
2558
2559 /**
2560  * try_to_release_page() - release old fs-specific metadata on a page
2561  *
2562  * @page: the page which the kernel is trying to free
2563  * @gfp_mask: memory allocation flags (and I/O mode)
2564  *
2565  * The address_space is to try to release any data against the page
2566  * (presumably at page->private).  If the release was successful, return `1'.
2567  * Otherwise return zero.
2568  *
2569  * The @gfp_mask argument specifies whether I/O may be performed to release
2570  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2571  *
2572  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2573  */
2574 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2575 {
2576         struct address_space * const mapping = page->mapping;
2577
2578         BUG_ON(!PageLocked(page));
2579         if (PageWriteback(page))
2580                 return 0;
2581
2582         if (mapping && mapping->a_ops->releasepage)
2583                 return mapping->a_ops->releasepage(page, gfp_mask);
2584         return try_to_free_buffers(page);
2585 }
2586
2587 EXPORT_SYMBOL(try_to_release_page);