mm: spinlock tree_lock
[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 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_cache_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         spin_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         spin_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_locked - add a locked page to the pagecache
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 a page to the pagecache. It must be locked.
452  * This function does not add the page to the LRU.  The caller must do that.
453  */
454 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
455                 pgoff_t offset, gfp_t gfp_mask)
456 {
457         int error;
458
459         VM_BUG_ON(!PageLocked(page));
460
461         error = mem_cgroup_cache_charge(page, current->mm,
462                                         gfp_mask & ~__GFP_HIGHMEM);
463         if (error)
464                 goto out;
465
466         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
467         if (error == 0) {
468                 page_cache_get(page);
469                 page->mapping = mapping;
470                 page->index = offset;
471
472                 spin_lock_irq(&mapping->tree_lock);
473                 error = radix_tree_insert(&mapping->page_tree, offset, page);
474                 if (likely(!error)) {
475                         mapping->nrpages++;
476                         __inc_zone_page_state(page, NR_FILE_PAGES);
477                 } else {
478                         page->mapping = NULL;
479                         mem_cgroup_uncharge_cache_page(page);
480                         page_cache_release(page);
481                 }
482
483                 spin_unlock_irq(&mapping->tree_lock);
484                 radix_tree_preload_end();
485         } else
486                 mem_cgroup_uncharge_cache_page(page);
487 out:
488         return error;
489 }
490 EXPORT_SYMBOL(add_to_page_cache_locked);
491
492 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
493                                 pgoff_t offset, gfp_t gfp_mask)
494 {
495         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
496         if (ret == 0)
497                 lru_cache_add(page);
498         return ret;
499 }
500
501 #ifdef CONFIG_NUMA
502 struct page *__page_cache_alloc(gfp_t gfp)
503 {
504         if (cpuset_do_page_mem_spread()) {
505                 int n = cpuset_mem_spread_node();
506                 return alloc_pages_node(n, gfp, 0);
507         }
508         return alloc_pages(gfp, 0);
509 }
510 EXPORT_SYMBOL(__page_cache_alloc);
511 #endif
512
513 static int __sleep_on_page_lock(void *word)
514 {
515         io_schedule();
516         return 0;
517 }
518
519 /*
520  * In order to wait for pages to become available there must be
521  * waitqueues associated with pages. By using a hash table of
522  * waitqueues where the bucket discipline is to maintain all
523  * waiters on the same queue and wake all when any of the pages
524  * become available, and for the woken contexts to check to be
525  * sure the appropriate page became available, this saves space
526  * at a cost of "thundering herd" phenomena during rare hash
527  * collisions.
528  */
529 static wait_queue_head_t *page_waitqueue(struct page *page)
530 {
531         const struct zone *zone = page_zone(page);
532
533         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
534 }
535
536 static inline void wake_up_page(struct page *page, int bit)
537 {
538         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
539 }
540
541 void wait_on_page_bit(struct page *page, int bit_nr)
542 {
543         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
544
545         if (test_bit(bit_nr, &page->flags))
546                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
547                                                         TASK_UNINTERRUPTIBLE);
548 }
549 EXPORT_SYMBOL(wait_on_page_bit);
550
551 /**
552  * unlock_page - unlock a locked page
553  * @page: the page
554  *
555  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
556  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
557  * mechananism between PageLocked pages and PageWriteback pages is shared.
558  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
559  *
560  * The first mb is necessary to safely close the critical section opened by the
561  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
562  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
563  * parallel wait_on_page_locked()).
564  */
565 void unlock_page(struct page *page)
566 {
567         smp_mb__before_clear_bit();
568         if (!TestClearPageLocked(page))
569                 BUG();
570         smp_mb__after_clear_bit(); 
571         wake_up_page(page, PG_locked);
572 }
573 EXPORT_SYMBOL(unlock_page);
574
575 /**
576  * end_page_writeback - end writeback against a page
577  * @page: the page
578  */
579 void end_page_writeback(struct page *page)
580 {
581         if (TestClearPageReclaim(page))
582                 rotate_reclaimable_page(page);
583
584         if (!test_clear_page_writeback(page))
585                 BUG();
586
587         smp_mb__after_clear_bit();
588         wake_up_page(page, PG_writeback);
589 }
590 EXPORT_SYMBOL(end_page_writeback);
591
592 /**
593  * __lock_page - get a lock on the page, assuming we need to sleep to get it
594  * @page: the page to lock
595  *
596  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
597  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
598  * chances are that on the second loop, the block layer's plug list is empty,
599  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
600  */
601 void __lock_page(struct page *page)
602 {
603         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
604
605         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
606                                                         TASK_UNINTERRUPTIBLE);
607 }
608 EXPORT_SYMBOL(__lock_page);
609
610 int __lock_page_killable(struct page *page)
611 {
612         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
613
614         return __wait_on_bit_lock(page_waitqueue(page), &wait,
615                                         sync_page_killable, TASK_KILLABLE);
616 }
617
618 /**
619  * __lock_page_nosync - get a lock on the page, without calling sync_page()
620  * @page: the page to lock
621  *
622  * Variant of lock_page that does not require the caller to hold a reference
623  * on the page's mapping.
624  */
625 void __lock_page_nosync(struct page *page)
626 {
627         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
628         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
629                                                         TASK_UNINTERRUPTIBLE);
630 }
631
632 /**
633  * find_get_page - find and get a page reference
634  * @mapping: the address_space to search
635  * @offset: the page index
636  *
637  * Is there a pagecache struct page at the given (mapping, offset) tuple?
638  * If yes, increment its refcount and return it; if no, return NULL.
639  */
640 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
641 {
642         void **pagep;
643         struct page *page;
644
645         rcu_read_lock();
646 repeat:
647         page = NULL;
648         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
649         if (pagep) {
650                 page = radix_tree_deref_slot(pagep);
651                 if (unlikely(!page || page == RADIX_TREE_RETRY))
652                         goto repeat;
653
654                 if (!page_cache_get_speculative(page))
655                         goto repeat;
656
657                 /*
658                  * Has the page moved?
659                  * This is part of the lockless pagecache protocol. See
660                  * include/linux/pagemap.h for details.
661                  */
662                 if (unlikely(page != *pagep)) {
663                         page_cache_release(page);
664                         goto repeat;
665                 }
666         }
667         rcu_read_unlock();
668
669         return page;
670 }
671 EXPORT_SYMBOL(find_get_page);
672
673 /**
674  * find_lock_page - locate, pin and lock a pagecache page
675  * @mapping: the address_space to search
676  * @offset: the page index
677  *
678  * Locates the desired pagecache page, locks it, increments its reference
679  * count and returns its address.
680  *
681  * Returns zero if the page was not present. find_lock_page() may sleep.
682  */
683 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
684 {
685         struct page *page;
686
687 repeat:
688         page = find_get_page(mapping, offset);
689         if (page) {
690                 lock_page(page);
691                 /* Has the page been truncated? */
692                 if (unlikely(page->mapping != mapping)) {
693                         unlock_page(page);
694                         page_cache_release(page);
695                         goto repeat;
696                 }
697                 VM_BUG_ON(page->index != offset);
698         }
699         return page;
700 }
701 EXPORT_SYMBOL(find_lock_page);
702
703 /**
704  * find_or_create_page - locate or add a pagecache page
705  * @mapping: the page's address_space
706  * @index: the page's index into the mapping
707  * @gfp_mask: page allocation mode
708  *
709  * Locates a page in the pagecache.  If the page is not present, a new page
710  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
711  * LRU list.  The returned page is locked and has its reference count
712  * incremented.
713  *
714  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
715  * allocation!
716  *
717  * find_or_create_page() returns the desired page's address, or zero on
718  * memory exhaustion.
719  */
720 struct page *find_or_create_page(struct address_space *mapping,
721                 pgoff_t index, gfp_t gfp_mask)
722 {
723         struct page *page;
724         int err;
725 repeat:
726         page = find_lock_page(mapping, index);
727         if (!page) {
728                 page = __page_cache_alloc(gfp_mask);
729                 if (!page)
730                         return NULL;
731                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
732                 if (unlikely(err)) {
733                         page_cache_release(page);
734                         page = NULL;
735                         if (err == -EEXIST)
736                                 goto repeat;
737                 }
738         }
739         return page;
740 }
741 EXPORT_SYMBOL(find_or_create_page);
742
743 /**
744  * find_get_pages - gang pagecache lookup
745  * @mapping:    The address_space to search
746  * @start:      The starting page index
747  * @nr_pages:   The maximum number of pages
748  * @pages:      Where the resulting pages are placed
749  *
750  * find_get_pages() will search for and return a group of up to
751  * @nr_pages pages in the mapping.  The pages are placed at @pages.
752  * find_get_pages() takes a reference against the returned pages.
753  *
754  * The search returns a group of mapping-contiguous pages with ascending
755  * indexes.  There may be holes in the indices due to not-present pages.
756  *
757  * find_get_pages() returns the number of pages which were found.
758  */
759 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
760                             unsigned int nr_pages, struct page **pages)
761 {
762         unsigned int i;
763         unsigned int ret;
764         unsigned int nr_found;
765
766         rcu_read_lock();
767 restart:
768         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
769                                 (void ***)pages, start, nr_pages);
770         ret = 0;
771         for (i = 0; i < nr_found; i++) {
772                 struct page *page;
773 repeat:
774                 page = radix_tree_deref_slot((void **)pages[i]);
775                 if (unlikely(!page))
776                         continue;
777                 /*
778                  * this can only trigger if nr_found == 1, making livelock
779                  * a non issue.
780                  */
781                 if (unlikely(page == RADIX_TREE_RETRY))
782                         goto restart;
783
784                 if (!page_cache_get_speculative(page))
785                         goto repeat;
786
787                 /* Has the page moved? */
788                 if (unlikely(page != *((void **)pages[i]))) {
789                         page_cache_release(page);
790                         goto repeat;
791                 }
792
793                 pages[ret] = page;
794                 ret++;
795         }
796         rcu_read_unlock();
797         return ret;
798 }
799
800 /**
801  * find_get_pages_contig - gang contiguous pagecache lookup
802  * @mapping:    The address_space to search
803  * @index:      The starting page index
804  * @nr_pages:   The maximum number of pages
805  * @pages:      Where the resulting pages are placed
806  *
807  * find_get_pages_contig() works exactly like find_get_pages(), except
808  * that the returned number of pages are guaranteed to be contiguous.
809  *
810  * find_get_pages_contig() returns the number of pages which were found.
811  */
812 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
813                                unsigned int nr_pages, struct page **pages)
814 {
815         unsigned int i;
816         unsigned int ret;
817         unsigned int nr_found;
818
819         rcu_read_lock();
820 restart:
821         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
822                                 (void ***)pages, index, nr_pages);
823         ret = 0;
824         for (i = 0; i < nr_found; i++) {
825                 struct page *page;
826 repeat:
827                 page = radix_tree_deref_slot((void **)pages[i]);
828                 if (unlikely(!page))
829                         continue;
830                 /*
831                  * this can only trigger if nr_found == 1, making livelock
832                  * a non issue.
833                  */
834                 if (unlikely(page == RADIX_TREE_RETRY))
835                         goto restart;
836
837                 if (page->mapping == NULL || page->index != index)
838                         break;
839
840                 if (!page_cache_get_speculative(page))
841                         goto repeat;
842
843                 /* Has the page moved? */
844                 if (unlikely(page != *((void **)pages[i]))) {
845                         page_cache_release(page);
846                         goto repeat;
847                 }
848
849                 pages[ret] = page;
850                 ret++;
851                 index++;
852         }
853         rcu_read_unlock();
854         return ret;
855 }
856 EXPORT_SYMBOL(find_get_pages_contig);
857
858 /**
859  * find_get_pages_tag - find and return pages that match @tag
860  * @mapping:    the address_space to search
861  * @index:      the starting page index
862  * @tag:        the tag index
863  * @nr_pages:   the maximum number of pages
864  * @pages:      where the resulting pages are placed
865  *
866  * Like find_get_pages, except we only return pages which are tagged with
867  * @tag.   We update @index to index the next page for the traversal.
868  */
869 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
870                         int tag, unsigned int nr_pages, struct page **pages)
871 {
872         unsigned int i;
873         unsigned int ret;
874         unsigned int nr_found;
875
876         rcu_read_lock();
877 restart:
878         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
879                                 (void ***)pages, *index, nr_pages, tag);
880         ret = 0;
881         for (i = 0; i < nr_found; i++) {
882                 struct page *page;
883 repeat:
884                 page = radix_tree_deref_slot((void **)pages[i]);
885                 if (unlikely(!page))
886                         continue;
887                 /*
888                  * this can only trigger if nr_found == 1, making livelock
889                  * a non issue.
890                  */
891                 if (unlikely(page == RADIX_TREE_RETRY))
892                         goto restart;
893
894                 if (!page_cache_get_speculative(page))
895                         goto repeat;
896
897                 /* Has the page moved? */
898                 if (unlikely(page != *((void **)pages[i]))) {
899                         page_cache_release(page);
900                         goto repeat;
901                 }
902
903                 pages[ret] = page;
904                 ret++;
905         }
906         rcu_read_unlock();
907
908         if (ret)
909                 *index = pages[ret - 1]->index + 1;
910
911         return ret;
912 }
913 EXPORT_SYMBOL(find_get_pages_tag);
914
915 /**
916  * grab_cache_page_nowait - returns locked page at given index in given cache
917  * @mapping: target address_space
918  * @index: the page index
919  *
920  * Same as grab_cache_page(), but do not wait if the page is unavailable.
921  * This is intended for speculative data generators, where the data can
922  * be regenerated if the page couldn't be grabbed.  This routine should
923  * be safe to call while holding the lock for another page.
924  *
925  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
926  * and deadlock against the caller's locked page.
927  */
928 struct page *
929 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
930 {
931         struct page *page = find_get_page(mapping, index);
932
933         if (page) {
934                 if (!TestSetPageLocked(page))
935                         return page;
936                 page_cache_release(page);
937                 return NULL;
938         }
939         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
940         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
941                 page_cache_release(page);
942                 page = NULL;
943         }
944         return page;
945 }
946 EXPORT_SYMBOL(grab_cache_page_nowait);
947
948 /*
949  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
950  * a _large_ part of the i/o request. Imagine the worst scenario:
951  *
952  *      ---R__________________________________________B__________
953  *         ^ reading here                             ^ bad block(assume 4k)
954  *
955  * read(R) => miss => readahead(R...B) => media error => frustrating retries
956  * => failing the whole request => read(R) => read(R+1) =>
957  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
958  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
959  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
960  *
961  * It is going insane. Fix it by quickly scaling down the readahead size.
962  */
963 static void shrink_readahead_size_eio(struct file *filp,
964                                         struct file_ra_state *ra)
965 {
966         if (!ra->ra_pages)
967                 return;
968
969         ra->ra_pages /= 4;
970 }
971
972 /**
973  * do_generic_file_read - generic file read routine
974  * @filp:       the file to read
975  * @ppos:       current file position
976  * @desc:       read_descriptor
977  * @actor:      read method
978  *
979  * This is a generic file read routine, and uses the
980  * mapping->a_ops->readpage() function for the actual low-level stuff.
981  *
982  * This is really ugly. But the goto's actually try to clarify some
983  * of the logic when it comes to error handling etc.
984  */
985 static void do_generic_file_read(struct file *filp, loff_t *ppos,
986                 read_descriptor_t *desc, read_actor_t actor)
987 {
988         struct address_space *mapping = filp->f_mapping;
989         struct inode *inode = mapping->host;
990         struct file_ra_state *ra = &filp->f_ra;
991         pgoff_t index;
992         pgoff_t last_index;
993         pgoff_t prev_index;
994         unsigned long offset;      /* offset into pagecache page */
995         unsigned int prev_offset;
996         int error;
997
998         index = *ppos >> PAGE_CACHE_SHIFT;
999         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1000         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1001         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1002         offset = *ppos & ~PAGE_CACHE_MASK;
1003
1004         for (;;) {
1005                 struct page *page;
1006                 pgoff_t end_index;
1007                 loff_t isize;
1008                 unsigned long nr, ret;
1009
1010                 cond_resched();
1011 find_page:
1012                 page = find_get_page(mapping, index);
1013                 if (!page) {
1014                         page_cache_sync_readahead(mapping,
1015                                         ra, filp,
1016                                         index, last_index - index);
1017                         page = find_get_page(mapping, index);
1018                         if (unlikely(page == NULL))
1019                                 goto no_cached_page;
1020                 }
1021                 if (PageReadahead(page)) {
1022                         page_cache_async_readahead(mapping,
1023                                         ra, filp, page,
1024                                         index, last_index - index);
1025                 }
1026                 if (!PageUptodate(page))
1027                         goto page_not_up_to_date;
1028 page_ok:
1029                 /*
1030                  * i_size must be checked after we know the page is Uptodate.
1031                  *
1032                  * Checking i_size after the check allows us to calculate
1033                  * the correct value for "nr", which means the zero-filled
1034                  * part of the page is not copied back to userspace (unless
1035                  * another truncate extends the file - this is desired though).
1036                  */
1037
1038                 isize = i_size_read(inode);
1039                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1040                 if (unlikely(!isize || index > end_index)) {
1041                         page_cache_release(page);
1042                         goto out;
1043                 }
1044
1045                 /* nr is the maximum number of bytes to copy from this page */
1046                 nr = PAGE_CACHE_SIZE;
1047                 if (index == end_index) {
1048                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1049                         if (nr <= offset) {
1050                                 page_cache_release(page);
1051                                 goto out;
1052                         }
1053                 }
1054                 nr = nr - offset;
1055
1056                 /* If users can be writing to this page using arbitrary
1057                  * virtual addresses, take care about potential aliasing
1058                  * before reading the page on the kernel side.
1059                  */
1060                 if (mapping_writably_mapped(mapping))
1061                         flush_dcache_page(page);
1062
1063                 /*
1064                  * When a sequential read accesses a page several times,
1065                  * only mark it as accessed the first time.
1066                  */
1067                 if (prev_index != index || offset != prev_offset)
1068                         mark_page_accessed(page);
1069                 prev_index = index;
1070
1071                 /*
1072                  * Ok, we have the page, and it's up-to-date, so
1073                  * now we can copy it to user space...
1074                  *
1075                  * The actor routine returns how many bytes were actually used..
1076                  * NOTE! This may not be the same as how much of a user buffer
1077                  * we filled up (we may be padding etc), so we can only update
1078                  * "pos" here (the actor routine has to update the user buffer
1079                  * pointers and the remaining count).
1080                  */
1081                 ret = actor(desc, page, offset, nr);
1082                 offset += ret;
1083                 index += offset >> PAGE_CACHE_SHIFT;
1084                 offset &= ~PAGE_CACHE_MASK;
1085                 prev_offset = offset;
1086
1087                 page_cache_release(page);
1088                 if (ret == nr && desc->count)
1089                         continue;
1090                 goto out;
1091
1092 page_not_up_to_date:
1093                 /* Get exclusive access to the page ... */
1094                 if (lock_page_killable(page))
1095                         goto readpage_eio;
1096
1097                 /* Did it get truncated before we got the lock? */
1098                 if (!page->mapping) {
1099                         unlock_page(page);
1100                         page_cache_release(page);
1101                         continue;
1102                 }
1103
1104                 /* Did somebody else fill it already? */
1105                 if (PageUptodate(page)) {
1106                         unlock_page(page);
1107                         goto page_ok;
1108                 }
1109
1110 readpage:
1111                 /* Start the actual read. The read will unlock the page. */
1112                 error = mapping->a_ops->readpage(filp, page);
1113
1114                 if (unlikely(error)) {
1115                         if (error == AOP_TRUNCATED_PAGE) {
1116                                 page_cache_release(page);
1117                                 goto find_page;
1118                         }
1119                         goto readpage_error;
1120                 }
1121
1122                 if (!PageUptodate(page)) {
1123                         if (lock_page_killable(page))
1124                                 goto readpage_eio;
1125                         if (!PageUptodate(page)) {
1126                                 if (page->mapping == NULL) {
1127                                         /*
1128                                          * invalidate_inode_pages got it
1129                                          */
1130                                         unlock_page(page);
1131                                         page_cache_release(page);
1132                                         goto find_page;
1133                                 }
1134                                 unlock_page(page);
1135                                 shrink_readahead_size_eio(filp, ra);
1136                                 goto readpage_eio;
1137                         }
1138                         unlock_page(page);
1139                 }
1140
1141                 goto page_ok;
1142
1143 readpage_eio:
1144                 error = -EIO;
1145 readpage_error:
1146                 /* UHHUH! A synchronous read error occurred. Report it */
1147                 desc->error = error;
1148                 page_cache_release(page);
1149                 goto out;
1150
1151 no_cached_page:
1152                 /*
1153                  * Ok, it wasn't cached, so we need to create a new
1154                  * page..
1155                  */
1156                 page = page_cache_alloc_cold(mapping);
1157                 if (!page) {
1158                         desc->error = -ENOMEM;
1159                         goto out;
1160                 }
1161                 error = add_to_page_cache_lru(page, mapping,
1162                                                 index, GFP_KERNEL);
1163                 if (error) {
1164                         page_cache_release(page);
1165                         if (error == -EEXIST)
1166                                 goto find_page;
1167                         desc->error = error;
1168                         goto out;
1169                 }
1170                 goto readpage;
1171         }
1172
1173 out:
1174         ra->prev_pos = prev_index;
1175         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1176         ra->prev_pos |= prev_offset;
1177
1178         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1179         if (filp)
1180                 file_accessed(filp);
1181 }
1182
1183 int file_read_actor(read_descriptor_t *desc, struct page *page,
1184                         unsigned long offset, unsigned long size)
1185 {
1186         char *kaddr;
1187         unsigned long left, count = desc->count;
1188
1189         if (size > count)
1190                 size = count;
1191
1192         /*
1193          * Faults on the destination of a read are common, so do it before
1194          * taking the kmap.
1195          */
1196         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1197                 kaddr = kmap_atomic(page, KM_USER0);
1198                 left = __copy_to_user_inatomic(desc->arg.buf,
1199                                                 kaddr + offset, size);
1200                 kunmap_atomic(kaddr, KM_USER0);
1201                 if (left == 0)
1202                         goto success;
1203         }
1204
1205         /* Do it the slow way */
1206         kaddr = kmap(page);
1207         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1208         kunmap(page);
1209
1210         if (left) {
1211                 size -= left;
1212                 desc->error = -EFAULT;
1213         }
1214 success:
1215         desc->count = count - size;
1216         desc->written += size;
1217         desc->arg.buf += size;
1218         return size;
1219 }
1220
1221 /*
1222  * Performs necessary checks before doing a write
1223  * @iov:        io vector request
1224  * @nr_segs:    number of segments in the iovec
1225  * @count:      number of bytes to write
1226  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1227  *
1228  * Adjust number of segments and amount of bytes to write (nr_segs should be
1229  * properly initialized first). Returns appropriate error code that caller
1230  * should return or zero in case that write should be allowed.
1231  */
1232 int generic_segment_checks(const struct iovec *iov,
1233                         unsigned long *nr_segs, size_t *count, int access_flags)
1234 {
1235         unsigned long   seg;
1236         size_t cnt = 0;
1237         for (seg = 0; seg < *nr_segs; seg++) {
1238                 const struct iovec *iv = &iov[seg];
1239
1240                 /*
1241                  * If any segment has a negative length, or the cumulative
1242                  * length ever wraps negative then return -EINVAL.
1243                  */
1244                 cnt += iv->iov_len;
1245                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1246                         return -EINVAL;
1247                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1248                         continue;
1249                 if (seg == 0)
1250                         return -EFAULT;
1251                 *nr_segs = seg;
1252                 cnt -= iv->iov_len;     /* This segment is no good */
1253                 break;
1254         }
1255         *count = cnt;
1256         return 0;
1257 }
1258 EXPORT_SYMBOL(generic_segment_checks);
1259
1260 /**
1261  * generic_file_aio_read - generic filesystem read routine
1262  * @iocb:       kernel I/O control block
1263  * @iov:        io vector request
1264  * @nr_segs:    number of segments in the iovec
1265  * @pos:        current file position
1266  *
1267  * This is the "read()" routine for all filesystems
1268  * that can use the page cache directly.
1269  */
1270 ssize_t
1271 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1272                 unsigned long nr_segs, loff_t pos)
1273 {
1274         struct file *filp = iocb->ki_filp;
1275         ssize_t retval;
1276         unsigned long seg;
1277         size_t count;
1278         loff_t *ppos = &iocb->ki_pos;
1279
1280         count = 0;
1281         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1282         if (retval)
1283                 return retval;
1284
1285         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1286         if (filp->f_flags & O_DIRECT) {
1287                 loff_t size;
1288                 struct address_space *mapping;
1289                 struct inode *inode;
1290
1291                 mapping = filp->f_mapping;
1292                 inode = mapping->host;
1293                 if (!count)
1294                         goto out; /* skip atime */
1295                 size = i_size_read(inode);
1296                 if (pos < size) {
1297                         retval = filemap_write_and_wait(mapping);
1298                         if (!retval) {
1299                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1300                                                         iov, pos, nr_segs);
1301                         }
1302                         if (retval > 0)
1303                                 *ppos = pos + retval;
1304                         if (retval) {
1305                                 file_accessed(filp);
1306                                 goto out;
1307                         }
1308                 }
1309         }
1310
1311         for (seg = 0; seg < nr_segs; seg++) {
1312                 read_descriptor_t desc;
1313
1314                 desc.written = 0;
1315                 desc.arg.buf = iov[seg].iov_base;
1316                 desc.count = iov[seg].iov_len;
1317                 if (desc.count == 0)
1318                         continue;
1319                 desc.error = 0;
1320                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1321                 retval += desc.written;
1322                 if (desc.error) {
1323                         retval = retval ?: desc.error;
1324                         break;
1325                 }
1326                 if (desc.count > 0)
1327                         break;
1328         }
1329 out:
1330         return retval;
1331 }
1332 EXPORT_SYMBOL(generic_file_aio_read);
1333
1334 static ssize_t
1335 do_readahead(struct address_space *mapping, struct file *filp,
1336              pgoff_t index, unsigned long nr)
1337 {
1338         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1339                 return -EINVAL;
1340
1341         force_page_cache_readahead(mapping, filp, index,
1342                                         max_sane_readahead(nr));
1343         return 0;
1344 }
1345
1346 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1347 {
1348         ssize_t ret;
1349         struct file *file;
1350
1351         ret = -EBADF;
1352         file = fget(fd);
1353         if (file) {
1354                 if (file->f_mode & FMODE_READ) {
1355                         struct address_space *mapping = file->f_mapping;
1356                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1357                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1358                         unsigned long len = end - start + 1;
1359                         ret = do_readahead(mapping, file, start, len);
1360                 }
1361                 fput(file);
1362         }
1363         return ret;
1364 }
1365
1366 #ifdef CONFIG_MMU
1367 /**
1368  * page_cache_read - adds requested page to the page cache if not already there
1369  * @file:       file to read
1370  * @offset:     page index
1371  *
1372  * This adds the requested page to the page cache if it isn't already there,
1373  * and schedules an I/O to read in its contents from disk.
1374  */
1375 static int page_cache_read(struct file *file, pgoff_t offset)
1376 {
1377         struct address_space *mapping = file->f_mapping;
1378         struct page *page; 
1379         int ret;
1380
1381         do {
1382                 page = page_cache_alloc_cold(mapping);
1383                 if (!page)
1384                         return -ENOMEM;
1385
1386                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1387                 if (ret == 0)
1388                         ret = mapping->a_ops->readpage(file, page);
1389                 else if (ret == -EEXIST)
1390                         ret = 0; /* losing race to add is OK */
1391
1392                 page_cache_release(page);
1393
1394         } while (ret == AOP_TRUNCATED_PAGE);
1395                 
1396         return ret;
1397 }
1398
1399 #define MMAP_LOTSAMISS  (100)
1400
1401 /**
1402  * filemap_fault - read in file data for page fault handling
1403  * @vma:        vma in which the fault was taken
1404  * @vmf:        struct vm_fault containing details of the fault
1405  *
1406  * filemap_fault() is invoked via the vma operations vector for a
1407  * mapped memory region to read in file data during a page fault.
1408  *
1409  * The goto's are kind of ugly, but this streamlines the normal case of having
1410  * it in the page cache, and handles the special cases reasonably without
1411  * having a lot of duplicated code.
1412  */
1413 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1414 {
1415         int error;
1416         struct file *file = vma->vm_file;
1417         struct address_space *mapping = file->f_mapping;
1418         struct file_ra_state *ra = &file->f_ra;
1419         struct inode *inode = mapping->host;
1420         struct page *page;
1421         pgoff_t size;
1422         int did_readaround = 0;
1423         int ret = 0;
1424
1425         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1426         if (vmf->pgoff >= size)
1427                 return VM_FAULT_SIGBUS;
1428
1429         /* If we don't want any read-ahead, don't bother */
1430         if (VM_RandomReadHint(vma))
1431                 goto no_cached_page;
1432
1433         /*
1434          * Do we have something in the page cache already?
1435          */
1436 retry_find:
1437         page = find_lock_page(mapping, vmf->pgoff);
1438         /*
1439          * For sequential accesses, we use the generic readahead logic.
1440          */
1441         if (VM_SequentialReadHint(vma)) {
1442                 if (!page) {
1443                         page_cache_sync_readahead(mapping, ra, file,
1444                                                            vmf->pgoff, 1);
1445                         page = find_lock_page(mapping, vmf->pgoff);
1446                         if (!page)
1447                                 goto no_cached_page;
1448                 }
1449                 if (PageReadahead(page)) {
1450                         page_cache_async_readahead(mapping, ra, file, page,
1451                                                            vmf->pgoff, 1);
1452                 }
1453         }
1454
1455         if (!page) {
1456                 unsigned long ra_pages;
1457
1458                 ra->mmap_miss++;
1459
1460                 /*
1461                  * Do we miss much more than hit in this file? If so,
1462                  * stop bothering with read-ahead. It will only hurt.
1463                  */
1464                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1465                         goto no_cached_page;
1466
1467                 /*
1468                  * To keep the pgmajfault counter straight, we need to
1469                  * check did_readaround, as this is an inner loop.
1470                  */
1471                 if (!did_readaround) {
1472                         ret = VM_FAULT_MAJOR;
1473                         count_vm_event(PGMAJFAULT);
1474                 }
1475                 did_readaround = 1;
1476                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1477                 if (ra_pages) {
1478                         pgoff_t start = 0;
1479
1480                         if (vmf->pgoff > ra_pages / 2)
1481                                 start = vmf->pgoff - ra_pages / 2;
1482                         do_page_cache_readahead(mapping, file, start, ra_pages);
1483                 }
1484                 page = find_lock_page(mapping, vmf->pgoff);
1485                 if (!page)
1486                         goto no_cached_page;
1487         }
1488
1489         if (!did_readaround)
1490                 ra->mmap_miss--;
1491
1492         /*
1493          * We have a locked page in the page cache, now we need to check
1494          * that it's up-to-date. If not, it is going to be due to an error.
1495          */
1496         if (unlikely(!PageUptodate(page)))
1497                 goto page_not_uptodate;
1498
1499         /* Must recheck i_size under page lock */
1500         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1501         if (unlikely(vmf->pgoff >= size)) {
1502                 unlock_page(page);
1503                 page_cache_release(page);
1504                 return VM_FAULT_SIGBUS;
1505         }
1506
1507         /*
1508          * Found the page and have a reference on it.
1509          */
1510         mark_page_accessed(page);
1511         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1512         vmf->page = page;
1513         return ret | VM_FAULT_LOCKED;
1514
1515 no_cached_page:
1516         /*
1517          * We're only likely to ever get here if MADV_RANDOM is in
1518          * effect.
1519          */
1520         error = page_cache_read(file, vmf->pgoff);
1521
1522         /*
1523          * The page we want has now been added to the page cache.
1524          * In the unlikely event that someone removed it in the
1525          * meantime, we'll just come back here and read it again.
1526          */
1527         if (error >= 0)
1528                 goto retry_find;
1529
1530         /*
1531          * An error return from page_cache_read can result if the
1532          * system is low on memory, or a problem occurs while trying
1533          * to schedule I/O.
1534          */
1535         if (error == -ENOMEM)
1536                 return VM_FAULT_OOM;
1537         return VM_FAULT_SIGBUS;
1538
1539 page_not_uptodate:
1540         /* IO error path */
1541         if (!did_readaround) {
1542                 ret = VM_FAULT_MAJOR;
1543                 count_vm_event(PGMAJFAULT);
1544         }
1545
1546         /*
1547          * Umm, take care of errors if the page isn't up-to-date.
1548          * Try to re-read it _once_. We do this synchronously,
1549          * because there really aren't any performance issues here
1550          * and we need to check for errors.
1551          */
1552         ClearPageError(page);
1553         error = mapping->a_ops->readpage(file, page);
1554         if (!error) {
1555                 wait_on_page_locked(page);
1556                 if (!PageUptodate(page))
1557                         error = -EIO;
1558         }
1559         page_cache_release(page);
1560
1561         if (!error || error == AOP_TRUNCATED_PAGE)
1562                 goto retry_find;
1563
1564         /* Things didn't work out. Return zero to tell the mm layer so. */
1565         shrink_readahead_size_eio(file, ra);
1566         return VM_FAULT_SIGBUS;
1567 }
1568 EXPORT_SYMBOL(filemap_fault);
1569
1570 struct vm_operations_struct generic_file_vm_ops = {
1571         .fault          = filemap_fault,
1572 };
1573
1574 /* This is used for a general mmap of a disk file */
1575
1576 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1577 {
1578         struct address_space *mapping = file->f_mapping;
1579
1580         if (!mapping->a_ops->readpage)
1581                 return -ENOEXEC;
1582         file_accessed(file);
1583         vma->vm_ops = &generic_file_vm_ops;
1584         vma->vm_flags |= VM_CAN_NONLINEAR;
1585         return 0;
1586 }
1587
1588 /*
1589  * This is for filesystems which do not implement ->writepage.
1590  */
1591 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1592 {
1593         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1594                 return -EINVAL;
1595         return generic_file_mmap(file, vma);
1596 }
1597 #else
1598 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1599 {
1600         return -ENOSYS;
1601 }
1602 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1603 {
1604         return -ENOSYS;
1605 }
1606 #endif /* CONFIG_MMU */
1607
1608 EXPORT_SYMBOL(generic_file_mmap);
1609 EXPORT_SYMBOL(generic_file_readonly_mmap);
1610
1611 static struct page *__read_cache_page(struct address_space *mapping,
1612                                 pgoff_t index,
1613                                 int (*filler)(void *,struct page*),
1614                                 void *data)
1615 {
1616         struct page *page;
1617         int err;
1618 repeat:
1619         page = find_get_page(mapping, index);
1620         if (!page) {
1621                 page = page_cache_alloc_cold(mapping);
1622                 if (!page)
1623                         return ERR_PTR(-ENOMEM);
1624                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1625                 if (unlikely(err)) {
1626                         page_cache_release(page);
1627                         if (err == -EEXIST)
1628                                 goto repeat;
1629                         /* Presumably ENOMEM for radix tree node */
1630                         return ERR_PTR(err);
1631                 }
1632                 err = filler(data, page);
1633                 if (err < 0) {
1634                         page_cache_release(page);
1635                         page = ERR_PTR(err);
1636                 }
1637         }
1638         return page;
1639 }
1640
1641 /**
1642  * read_cache_page_async - read into page cache, fill it if needed
1643  * @mapping:    the page's address_space
1644  * @index:      the page index
1645  * @filler:     function to perform the read
1646  * @data:       destination for read data
1647  *
1648  * Same as read_cache_page, but don't wait for page to become unlocked
1649  * after submitting it to the filler.
1650  *
1651  * Read into the page cache. If a page already exists, and PageUptodate() is
1652  * not set, try to fill the page but don't wait for it to become unlocked.
1653  *
1654  * If the page does not get brought uptodate, return -EIO.
1655  */
1656 struct page *read_cache_page_async(struct address_space *mapping,
1657                                 pgoff_t index,
1658                                 int (*filler)(void *,struct page*),
1659                                 void *data)
1660 {
1661         struct page *page;
1662         int err;
1663
1664 retry:
1665         page = __read_cache_page(mapping, index, filler, data);
1666         if (IS_ERR(page))
1667                 return page;
1668         if (PageUptodate(page))
1669                 goto out;
1670
1671         lock_page(page);
1672         if (!page->mapping) {
1673                 unlock_page(page);
1674                 page_cache_release(page);
1675                 goto retry;
1676         }
1677         if (PageUptodate(page)) {
1678                 unlock_page(page);
1679                 goto out;
1680         }
1681         err = filler(data, page);
1682         if (err < 0) {
1683                 page_cache_release(page);
1684                 return ERR_PTR(err);
1685         }
1686 out:
1687         mark_page_accessed(page);
1688         return page;
1689 }
1690 EXPORT_SYMBOL(read_cache_page_async);
1691
1692 /**
1693  * read_cache_page - read into page cache, fill it if needed
1694  * @mapping:    the page's address_space
1695  * @index:      the page index
1696  * @filler:     function to perform the read
1697  * @data:       destination for read data
1698  *
1699  * Read into the page cache. If a page already exists, and PageUptodate() is
1700  * not set, try to fill the page then wait for it to become unlocked.
1701  *
1702  * If the page does not get brought uptodate, return -EIO.
1703  */
1704 struct page *read_cache_page(struct address_space *mapping,
1705                                 pgoff_t index,
1706                                 int (*filler)(void *,struct page*),
1707                                 void *data)
1708 {
1709         struct page *page;
1710
1711         page = read_cache_page_async(mapping, index, filler, data);
1712         if (IS_ERR(page))
1713                 goto out;
1714         wait_on_page_locked(page);
1715         if (!PageUptodate(page)) {
1716                 page_cache_release(page);
1717                 page = ERR_PTR(-EIO);
1718         }
1719  out:
1720         return page;
1721 }
1722 EXPORT_SYMBOL(read_cache_page);
1723
1724 /*
1725  * The logic we want is
1726  *
1727  *      if suid or (sgid and xgrp)
1728  *              remove privs
1729  */
1730 int should_remove_suid(struct dentry *dentry)
1731 {
1732         mode_t mode = dentry->d_inode->i_mode;
1733         int kill = 0;
1734
1735         /* suid always must be killed */
1736         if (unlikely(mode & S_ISUID))
1737                 kill = ATTR_KILL_SUID;
1738
1739         /*
1740          * sgid without any exec bits is just a mandatory locking mark; leave
1741          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1742          */
1743         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1744                 kill |= ATTR_KILL_SGID;
1745
1746         if (unlikely(kill && !capable(CAP_FSETID)))
1747                 return kill;
1748
1749         return 0;
1750 }
1751 EXPORT_SYMBOL(should_remove_suid);
1752
1753 static int __remove_suid(struct dentry *dentry, int kill)
1754 {
1755         struct iattr newattrs;
1756
1757         newattrs.ia_valid = ATTR_FORCE | kill;
1758         return notify_change(dentry, &newattrs);
1759 }
1760
1761 int remove_suid(struct dentry *dentry)
1762 {
1763         int killsuid = should_remove_suid(dentry);
1764         int killpriv = security_inode_need_killpriv(dentry);
1765         int error = 0;
1766
1767         if (killpriv < 0)
1768                 return killpriv;
1769         if (killpriv)
1770                 error = security_inode_killpriv(dentry);
1771         if (!error && killsuid)
1772                 error = __remove_suid(dentry, killsuid);
1773
1774         return error;
1775 }
1776 EXPORT_SYMBOL(remove_suid);
1777
1778 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1779                         const struct iovec *iov, size_t base, size_t bytes)
1780 {
1781         size_t copied = 0, left = 0;
1782
1783         while (bytes) {
1784                 char __user *buf = iov->iov_base + base;
1785                 int copy = min(bytes, iov->iov_len - base);
1786
1787                 base = 0;
1788                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1789                 copied += copy;
1790                 bytes -= copy;
1791                 vaddr += copy;
1792                 iov++;
1793
1794                 if (unlikely(left))
1795                         break;
1796         }
1797         return copied - left;
1798 }
1799
1800 /*
1801  * Copy as much as we can into the page and return the number of bytes which
1802  * were sucessfully copied.  If a fault is encountered then return the number of
1803  * bytes which were copied.
1804  */
1805 size_t iov_iter_copy_from_user_atomic(struct page *page,
1806                 struct iov_iter *i, unsigned long offset, size_t bytes)
1807 {
1808         char *kaddr;
1809         size_t copied;
1810
1811         BUG_ON(!in_atomic());
1812         kaddr = kmap_atomic(page, KM_USER0);
1813         if (likely(i->nr_segs == 1)) {
1814                 int left;
1815                 char __user *buf = i->iov->iov_base + i->iov_offset;
1816                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1817                                                         buf, bytes);
1818                 copied = bytes - left;
1819         } else {
1820                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1821                                                 i->iov, i->iov_offset, bytes);
1822         }
1823         kunmap_atomic(kaddr, KM_USER0);
1824
1825         return copied;
1826 }
1827 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1828
1829 /*
1830  * This has the same sideeffects and return value as
1831  * iov_iter_copy_from_user_atomic().
1832  * The difference is that it attempts to resolve faults.
1833  * Page must not be locked.
1834  */
1835 size_t iov_iter_copy_from_user(struct page *page,
1836                 struct iov_iter *i, unsigned long offset, size_t bytes)
1837 {
1838         char *kaddr;
1839         size_t copied;
1840
1841         kaddr = kmap(page);
1842         if (likely(i->nr_segs == 1)) {
1843                 int left;
1844                 char __user *buf = i->iov->iov_base + i->iov_offset;
1845                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1846                 copied = bytes - left;
1847         } else {
1848                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1849                                                 i->iov, i->iov_offset, bytes);
1850         }
1851         kunmap(page);
1852         return copied;
1853 }
1854 EXPORT_SYMBOL(iov_iter_copy_from_user);
1855
1856 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1857 {
1858         BUG_ON(i->count < bytes);
1859
1860         if (likely(i->nr_segs == 1)) {
1861                 i->iov_offset += bytes;
1862                 i->count -= bytes;
1863         } else {
1864                 const struct iovec *iov = i->iov;
1865                 size_t base = i->iov_offset;
1866
1867                 /*
1868                  * The !iov->iov_len check ensures we skip over unlikely
1869                  * zero-length segments (without overruning the iovec).
1870                  */
1871                 while (bytes || unlikely(!iov->iov_len && i->count)) {
1872                         int copy;
1873
1874                         copy = min(bytes, iov->iov_len - base);
1875                         BUG_ON(!i->count || i->count < copy);
1876                         i->count -= copy;
1877                         bytes -= copy;
1878                         base += copy;
1879                         if (iov->iov_len == base) {
1880                                 iov++;
1881                                 base = 0;
1882                         }
1883                 }
1884                 i->iov = iov;
1885                 i->iov_offset = base;
1886         }
1887 }
1888 EXPORT_SYMBOL(iov_iter_advance);
1889
1890 /*
1891  * Fault in the first iovec of the given iov_iter, to a maximum length
1892  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1893  * accessed (ie. because it is an invalid address).
1894  *
1895  * writev-intensive code may want this to prefault several iovecs -- that
1896  * would be possible (callers must not rely on the fact that _only_ the
1897  * first iovec will be faulted with the current implementation).
1898  */
1899 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1900 {
1901         char __user *buf = i->iov->iov_base + i->iov_offset;
1902         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1903         return fault_in_pages_readable(buf, bytes);
1904 }
1905 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1906
1907 /*
1908  * Return the count of just the current iov_iter segment.
1909  */
1910 size_t iov_iter_single_seg_count(struct iov_iter *i)
1911 {
1912         const struct iovec *iov = i->iov;
1913         if (i->nr_segs == 1)
1914                 return i->count;
1915         else
1916                 return min(i->count, iov->iov_len - i->iov_offset);
1917 }
1918 EXPORT_SYMBOL(iov_iter_single_seg_count);
1919
1920 /*
1921  * Performs necessary checks before doing a write
1922  *
1923  * Can adjust writing position or amount of bytes to write.
1924  * Returns appropriate error code that caller should return or
1925  * zero in case that write should be allowed.
1926  */
1927 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1928 {
1929         struct inode *inode = file->f_mapping->host;
1930         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1931
1932         if (unlikely(*pos < 0))
1933                 return -EINVAL;
1934
1935         if (!isblk) {
1936                 /* FIXME: this is for backwards compatibility with 2.4 */
1937                 if (file->f_flags & O_APPEND)
1938                         *pos = i_size_read(inode);
1939
1940                 if (limit != RLIM_INFINITY) {
1941                         if (*pos >= limit) {
1942                                 send_sig(SIGXFSZ, current, 0);
1943                                 return -EFBIG;
1944                         }
1945                         if (*count > limit - (typeof(limit))*pos) {
1946                                 *count = limit - (typeof(limit))*pos;
1947                         }
1948                 }
1949         }
1950
1951         /*
1952          * LFS rule
1953          */
1954         if (unlikely(*pos + *count > MAX_NON_LFS &&
1955                                 !(file->f_flags & O_LARGEFILE))) {
1956                 if (*pos >= MAX_NON_LFS) {
1957                         return -EFBIG;
1958                 }
1959                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1960                         *count = MAX_NON_LFS - (unsigned long)*pos;
1961                 }
1962         }
1963
1964         /*
1965          * Are we about to exceed the fs block limit ?
1966          *
1967          * If we have written data it becomes a short write.  If we have
1968          * exceeded without writing data we send a signal and return EFBIG.
1969          * Linus frestrict idea will clean these up nicely..
1970          */
1971         if (likely(!isblk)) {
1972                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1973                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1974                                 return -EFBIG;
1975                         }
1976                         /* zero-length writes at ->s_maxbytes are OK */
1977                 }
1978
1979                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1980                         *count = inode->i_sb->s_maxbytes - *pos;
1981         } else {
1982 #ifdef CONFIG_BLOCK
1983                 loff_t isize;
1984                 if (bdev_read_only(I_BDEV(inode)))
1985                         return -EPERM;
1986                 isize = i_size_read(inode);
1987                 if (*pos >= isize) {
1988                         if (*count || *pos > isize)
1989                                 return -ENOSPC;
1990                 }
1991
1992                 if (*pos + *count > isize)
1993                         *count = isize - *pos;
1994 #else
1995                 return -EPERM;
1996 #endif
1997         }
1998         return 0;
1999 }
2000 EXPORT_SYMBOL(generic_write_checks);
2001
2002 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2003                                 loff_t pos, unsigned len, unsigned flags,
2004                                 struct page **pagep, void **fsdata)
2005 {
2006         const struct address_space_operations *aops = mapping->a_ops;
2007
2008         if (aops->write_begin) {
2009                 return aops->write_begin(file, mapping, pos, len, flags,
2010                                                         pagep, fsdata);
2011         } else {
2012                 int ret;
2013                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2014                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2015                 struct inode *inode = mapping->host;
2016                 struct page *page;
2017 again:
2018                 page = __grab_cache_page(mapping, index);
2019                 *pagep = page;
2020                 if (!page)
2021                         return -ENOMEM;
2022
2023                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
2024                         /*
2025                          * There is no way to resolve a short write situation
2026                          * for a !Uptodate page (except by double copying in
2027                          * the caller done by generic_perform_write_2copy).
2028                          *
2029                          * Instead, we have to bring it uptodate here.
2030                          */
2031                         ret = aops->readpage(file, page);
2032                         page_cache_release(page);
2033                         if (ret) {
2034                                 if (ret == AOP_TRUNCATED_PAGE)
2035                                         goto again;
2036                                 return ret;
2037                         }
2038                         goto again;
2039                 }
2040
2041                 ret = aops->prepare_write(file, page, offset, offset+len);
2042                 if (ret) {
2043                         unlock_page(page);
2044                         page_cache_release(page);
2045                         if (pos + len > inode->i_size)
2046                                 vmtruncate(inode, inode->i_size);
2047                 }
2048                 return ret;
2049         }
2050 }
2051 EXPORT_SYMBOL(pagecache_write_begin);
2052
2053 int pagecache_write_end(struct file *file, struct address_space *mapping,
2054                                 loff_t pos, unsigned len, unsigned copied,
2055                                 struct page *page, void *fsdata)
2056 {
2057         const struct address_space_operations *aops = mapping->a_ops;
2058         int ret;
2059
2060         if (aops->write_end) {
2061                 mark_page_accessed(page);
2062                 ret = aops->write_end(file, mapping, pos, len, copied,
2063                                                         page, fsdata);
2064         } else {
2065                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2066                 struct inode *inode = mapping->host;
2067
2068                 flush_dcache_page(page);
2069                 ret = aops->commit_write(file, page, offset, offset+len);
2070                 unlock_page(page);
2071                 mark_page_accessed(page);
2072                 page_cache_release(page);
2073
2074                 if (ret < 0) {
2075                         if (pos + len > inode->i_size)
2076                                 vmtruncate(inode, inode->i_size);
2077                 } else if (ret > 0)
2078                         ret = min_t(size_t, copied, ret);
2079                 else
2080                         ret = copied;
2081         }
2082
2083         return ret;
2084 }
2085 EXPORT_SYMBOL(pagecache_write_end);
2086
2087 ssize_t
2088 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2089                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2090                 size_t count, size_t ocount)
2091 {
2092         struct file     *file = iocb->ki_filp;
2093         struct address_space *mapping = file->f_mapping;
2094         struct inode    *inode = mapping->host;
2095         ssize_t         written;
2096         size_t          write_len;
2097         pgoff_t         end;
2098
2099         if (count != ocount)
2100                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2101
2102         /*
2103          * Unmap all mmappings of the file up-front.
2104          *
2105          * This will cause any pte dirty bits to be propagated into the
2106          * pageframes for the subsequent filemap_write_and_wait().
2107          */
2108         write_len = iov_length(iov, *nr_segs);
2109         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2110         if (mapping_mapped(mapping))
2111                 unmap_mapping_range(mapping, pos, write_len, 0);
2112
2113         written = filemap_write_and_wait(mapping);
2114         if (written)
2115                 goto out;
2116
2117         /*
2118          * After a write we want buffered reads to be sure to go to disk to get
2119          * the new data.  We invalidate clean cached page from the region we're
2120          * about to write.  We do this *before* the write so that we can return
2121          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2122          */
2123         if (mapping->nrpages) {
2124                 written = invalidate_inode_pages2_range(mapping,
2125                                         pos >> PAGE_CACHE_SHIFT, end);
2126                 if (written)
2127                         goto out;
2128         }
2129
2130         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2131
2132         /*
2133          * Finally, try again to invalidate clean pages which might have been
2134          * cached by non-direct readahead, or faulted in by get_user_pages()
2135          * if the source of the write was an mmap'ed region of the file
2136          * we're writing.  Either one is a pretty crazy thing to do,
2137          * so we don't support it 100%.  If this invalidation
2138          * fails, tough, the write still worked...
2139          */
2140         if (mapping->nrpages) {
2141                 invalidate_inode_pages2_range(mapping,
2142                                               pos >> PAGE_CACHE_SHIFT, end);
2143         }
2144
2145         if (written > 0) {
2146                 loff_t end = pos + written;
2147                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2148                         i_size_write(inode,  end);
2149                         mark_inode_dirty(inode);
2150                 }
2151                 *ppos = end;
2152         }
2153
2154         /*
2155          * Sync the fs metadata but not the minor inode changes and
2156          * of course not the data as we did direct DMA for the IO.
2157          * i_mutex is held, which protects generic_osync_inode() from
2158          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2159          */
2160 out:
2161         if ((written >= 0 || written == -EIOCBQUEUED) &&
2162             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2163                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2164                 if (err < 0)
2165                         written = err;
2166         }
2167         return written;
2168 }
2169 EXPORT_SYMBOL(generic_file_direct_write);
2170
2171 /*
2172  * Find or create a page at the given pagecache position. Return the locked
2173  * page. This function is specifically for buffered writes.
2174  */
2175 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2176 {
2177         int status;
2178         struct page *page;
2179 repeat:
2180         page = find_lock_page(mapping, index);
2181         if (likely(page))
2182                 return page;
2183
2184         page = page_cache_alloc(mapping);
2185         if (!page)
2186                 return NULL;
2187         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2188         if (unlikely(status)) {
2189                 page_cache_release(page);
2190                 if (status == -EEXIST)
2191                         goto repeat;
2192                 return NULL;
2193         }
2194         return page;
2195 }
2196 EXPORT_SYMBOL(__grab_cache_page);
2197
2198 static ssize_t generic_perform_write_2copy(struct file *file,
2199                                 struct iov_iter *i, loff_t pos)
2200 {
2201         struct address_space *mapping = file->f_mapping;
2202         const struct address_space_operations *a_ops = mapping->a_ops;
2203         struct inode *inode = mapping->host;
2204         long status = 0;
2205         ssize_t written = 0;
2206
2207         do {
2208                 struct page *src_page;
2209                 struct page *page;
2210                 pgoff_t index;          /* Pagecache index for current page */
2211                 unsigned long offset;   /* Offset into pagecache page */
2212                 unsigned long bytes;    /* Bytes to write to page */
2213                 size_t copied;          /* Bytes copied from user */
2214
2215                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2216                 index = pos >> PAGE_CACHE_SHIFT;
2217                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2218                                                 iov_iter_count(i));
2219
2220                 /*
2221                  * a non-NULL src_page indicates that we're doing the
2222                  * copy via get_user_pages and kmap.
2223                  */
2224                 src_page = NULL;
2225
2226                 /*
2227                  * Bring in the user page that we will copy from _first_.
2228                  * Otherwise there's a nasty deadlock on copying from the
2229                  * same page as we're writing to, without it being marked
2230                  * up-to-date.
2231                  *
2232                  * Not only is this an optimisation, but it is also required
2233                  * to check that the address is actually valid, when atomic
2234                  * usercopies are used, below.
2235                  */
2236                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2237                         status = -EFAULT;
2238                         break;
2239                 }
2240
2241                 page = __grab_cache_page(mapping, index);
2242                 if (!page) {
2243                         status = -ENOMEM;
2244                         break;
2245                 }
2246
2247                 /*
2248                  * non-uptodate pages cannot cope with short copies, and we
2249                  * cannot take a pagefault with the destination page locked.
2250                  * So pin the source page to copy it.
2251                  */
2252                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2253                         unlock_page(page);
2254
2255                         src_page = alloc_page(GFP_KERNEL);
2256                         if (!src_page) {
2257                                 page_cache_release(page);
2258                                 status = -ENOMEM;
2259                                 break;
2260                         }
2261
2262                         /*
2263                          * Cannot get_user_pages with a page locked for the
2264                          * same reason as we can't take a page fault with a
2265                          * page locked (as explained below).
2266                          */
2267                         copied = iov_iter_copy_from_user(src_page, i,
2268                                                                 offset, bytes);
2269                         if (unlikely(copied == 0)) {
2270                                 status = -EFAULT;
2271                                 page_cache_release(page);
2272                                 page_cache_release(src_page);
2273                                 break;
2274                         }
2275                         bytes = copied;
2276
2277                         lock_page(page);
2278                         /*
2279                          * Can't handle the page going uptodate here, because
2280                          * that means we would use non-atomic usercopies, which
2281                          * zero out the tail of the page, which can cause
2282                          * zeroes to become transiently visible. We could just
2283                          * use a non-zeroing copy, but the APIs aren't too
2284                          * consistent.
2285                          */
2286                         if (unlikely(!page->mapping || PageUptodate(page))) {
2287                                 unlock_page(page);
2288                                 page_cache_release(page);
2289                                 page_cache_release(src_page);
2290                                 continue;
2291                         }
2292                 }
2293
2294                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2295                 if (unlikely(status))
2296                         goto fs_write_aop_error;
2297
2298                 if (!src_page) {
2299                         /*
2300                          * Must not enter the pagefault handler here, because
2301                          * we hold the page lock, so we might recursively
2302                          * deadlock on the same lock, or get an ABBA deadlock
2303                          * against a different lock, or against the mmap_sem
2304                          * (which nests outside the page lock).  So increment
2305                          * preempt count, and use _atomic usercopies.
2306                          *
2307                          * The page is uptodate so we are OK to encounter a
2308                          * short copy: if unmodified parts of the page are
2309                          * marked dirty and written out to disk, it doesn't
2310                          * really matter.
2311                          */
2312                         pagefault_disable();
2313                         copied = iov_iter_copy_from_user_atomic(page, i,
2314                                                                 offset, bytes);
2315                         pagefault_enable();
2316                 } else {
2317                         void *src, *dst;
2318                         src = kmap_atomic(src_page, KM_USER0);
2319                         dst = kmap_atomic(page, KM_USER1);
2320                         memcpy(dst + offset, src + offset, bytes);
2321                         kunmap_atomic(dst, KM_USER1);
2322                         kunmap_atomic(src, KM_USER0);
2323                         copied = bytes;
2324                 }
2325                 flush_dcache_page(page);
2326
2327                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2328                 if (unlikely(status < 0))
2329                         goto fs_write_aop_error;
2330                 if (unlikely(status > 0)) /* filesystem did partial write */
2331                         copied = min_t(size_t, copied, status);
2332
2333                 unlock_page(page);
2334                 mark_page_accessed(page);
2335                 page_cache_release(page);
2336                 if (src_page)
2337                         page_cache_release(src_page);
2338
2339                 iov_iter_advance(i, copied);
2340                 pos += copied;
2341                 written += copied;
2342
2343                 balance_dirty_pages_ratelimited(mapping);
2344                 cond_resched();
2345                 continue;
2346
2347 fs_write_aop_error:
2348                 unlock_page(page);
2349                 page_cache_release(page);
2350                 if (src_page)
2351                         page_cache_release(src_page);
2352
2353                 /*
2354                  * prepare_write() may have instantiated a few blocks
2355                  * outside i_size.  Trim these off again. Don't need
2356                  * i_size_read because we hold i_mutex.
2357                  */
2358                 if (pos + bytes > inode->i_size)
2359                         vmtruncate(inode, inode->i_size);
2360                 break;
2361         } while (iov_iter_count(i));
2362
2363         return written ? written : status;
2364 }
2365
2366 static ssize_t generic_perform_write(struct file *file,
2367                                 struct iov_iter *i, loff_t pos)
2368 {
2369         struct address_space *mapping = file->f_mapping;
2370         const struct address_space_operations *a_ops = mapping->a_ops;
2371         long status = 0;
2372         ssize_t written = 0;
2373         unsigned int flags = 0;
2374
2375         /*
2376          * Copies from kernel address space cannot fail (NFSD is a big user).
2377          */
2378         if (segment_eq(get_fs(), KERNEL_DS))
2379                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2380
2381         do {
2382                 struct page *page;
2383                 pgoff_t index;          /* Pagecache index for current page */
2384                 unsigned long offset;   /* Offset into pagecache page */
2385                 unsigned long bytes;    /* Bytes to write to page */
2386                 size_t copied;          /* Bytes copied from user */
2387                 void *fsdata;
2388
2389                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2390                 index = pos >> PAGE_CACHE_SHIFT;
2391                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2392                                                 iov_iter_count(i));
2393
2394 again:
2395
2396                 /*
2397                  * Bring in the user page that we will copy from _first_.
2398                  * Otherwise there's a nasty deadlock on copying from the
2399                  * same page as we're writing to, without it being marked
2400                  * up-to-date.
2401                  *
2402                  * Not only is this an optimisation, but it is also required
2403                  * to check that the address is actually valid, when atomic
2404                  * usercopies are used, below.
2405                  */
2406                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2407                         status = -EFAULT;
2408                         break;
2409                 }
2410
2411                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2412                                                 &page, &fsdata);
2413                 if (unlikely(status))
2414                         break;
2415
2416                 pagefault_disable();
2417                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2418                 pagefault_enable();
2419                 flush_dcache_page(page);
2420
2421                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2422                                                 page, fsdata);
2423                 if (unlikely(status < 0))
2424                         break;
2425                 copied = status;
2426
2427                 cond_resched();
2428
2429                 iov_iter_advance(i, copied);
2430                 if (unlikely(copied == 0)) {
2431                         /*
2432                          * If we were unable to copy any data at all, we must
2433                          * fall back to a single segment length write.
2434                          *
2435                          * If we didn't fallback here, we could livelock
2436                          * because not all segments in the iov can be copied at
2437                          * once without a pagefault.
2438                          */
2439                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2440                                                 iov_iter_single_seg_count(i));
2441                         goto again;
2442                 }
2443                 pos += copied;
2444                 written += copied;
2445
2446                 balance_dirty_pages_ratelimited(mapping);
2447
2448         } while (iov_iter_count(i));
2449
2450         return written ? written : status;
2451 }
2452
2453 ssize_t
2454 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2455                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2456                 size_t count, ssize_t written)
2457 {
2458         struct file *file = iocb->ki_filp;
2459         struct address_space *mapping = file->f_mapping;
2460         const struct address_space_operations *a_ops = mapping->a_ops;
2461         struct inode *inode = mapping->host;
2462         ssize_t status;
2463         struct iov_iter i;
2464
2465         iov_iter_init(&i, iov, nr_segs, count, written);
2466         if (a_ops->write_begin)
2467                 status = generic_perform_write(file, &i, pos);
2468         else
2469                 status = generic_perform_write_2copy(file, &i, pos);
2470
2471         if (likely(status >= 0)) {
2472                 written += status;
2473                 *ppos = pos + status;
2474
2475                 /*
2476                  * For now, when the user asks for O_SYNC, we'll actually give
2477                  * O_DSYNC
2478                  */
2479                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2480                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2481                                 status = generic_osync_inode(inode, mapping,
2482                                                 OSYNC_METADATA|OSYNC_DATA);
2483                 }
2484         }
2485         
2486         /*
2487          * If we get here for O_DIRECT writes then we must have fallen through
2488          * to buffered writes (block instantiation inside i_size).  So we sync
2489          * the file data here, to try to honour O_DIRECT expectations.
2490          */
2491         if (unlikely(file->f_flags & O_DIRECT) && written)
2492                 status = filemap_write_and_wait(mapping);
2493
2494         return written ? written : status;
2495 }
2496 EXPORT_SYMBOL(generic_file_buffered_write);
2497
2498 static ssize_t
2499 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2500                                 unsigned long nr_segs, loff_t *ppos)
2501 {
2502         struct file *file = iocb->ki_filp;
2503         struct address_space * mapping = file->f_mapping;
2504         size_t ocount;          /* original count */
2505         size_t count;           /* after file limit checks */
2506         struct inode    *inode = mapping->host;
2507         loff_t          pos;
2508         ssize_t         written;
2509         ssize_t         err;
2510
2511         ocount = 0;
2512         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2513         if (err)
2514                 return err;
2515
2516         count = ocount;
2517         pos = *ppos;
2518
2519         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2520
2521         /* We can write back this queue in page reclaim */
2522         current->backing_dev_info = mapping->backing_dev_info;
2523         written = 0;
2524
2525         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2526         if (err)
2527                 goto out;
2528
2529         if (count == 0)
2530                 goto out;
2531
2532         err = remove_suid(file->f_path.dentry);
2533         if (err)
2534                 goto out;
2535
2536         file_update_time(file);
2537
2538         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2539         if (unlikely(file->f_flags & O_DIRECT)) {
2540                 loff_t endbyte;
2541                 ssize_t written_buffered;
2542
2543                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2544                                                         ppos, count, ocount);
2545                 if (written < 0 || written == count)
2546                         goto out;
2547                 /*
2548                  * direct-io write to a hole: fall through to buffered I/O
2549                  * for completing the rest of the request.
2550                  */
2551                 pos += written;
2552                 count -= written;
2553                 written_buffered = generic_file_buffered_write(iocb, iov,
2554                                                 nr_segs, pos, ppos, count,
2555                                                 written);
2556                 /*
2557                  * If generic_file_buffered_write() retuned a synchronous error
2558                  * then we want to return the number of bytes which were
2559                  * direct-written, or the error code if that was zero.  Note
2560                  * that this differs from normal direct-io semantics, which
2561                  * will return -EFOO even if some bytes were written.
2562                  */
2563                 if (written_buffered < 0) {
2564                         err = written_buffered;
2565                         goto out;
2566                 }
2567
2568                 /*
2569                  * We need to ensure that the page cache pages are written to
2570                  * disk and invalidated to preserve the expected O_DIRECT
2571                  * semantics.
2572                  */
2573                 endbyte = pos + written_buffered - written - 1;
2574                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2575                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2576                                             SYNC_FILE_RANGE_WRITE|
2577                                             SYNC_FILE_RANGE_WAIT_AFTER);
2578                 if (err == 0) {
2579                         written = written_buffered;
2580                         invalidate_mapping_pages(mapping,
2581                                                  pos >> PAGE_CACHE_SHIFT,
2582                                                  endbyte >> PAGE_CACHE_SHIFT);
2583                 } else {
2584                         /*
2585                          * We don't know how much we wrote, so just return
2586                          * the number of bytes which were direct-written
2587                          */
2588                 }
2589         } else {
2590                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2591                                 pos, ppos, count, written);
2592         }
2593 out:
2594         current->backing_dev_info = NULL;
2595         return written ? written : err;
2596 }
2597
2598 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2599                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2600 {
2601         struct file *file = iocb->ki_filp;
2602         struct address_space *mapping = file->f_mapping;
2603         struct inode *inode = mapping->host;
2604         ssize_t ret;
2605
2606         BUG_ON(iocb->ki_pos != pos);
2607
2608         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2609                         &iocb->ki_pos);
2610
2611         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2612                 ssize_t err;
2613
2614                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2615                 if (err < 0)
2616                         ret = err;
2617         }
2618         return ret;
2619 }
2620 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2621
2622 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2623                 unsigned long nr_segs, loff_t pos)
2624 {
2625         struct file *file = iocb->ki_filp;
2626         struct address_space *mapping = file->f_mapping;
2627         struct inode *inode = mapping->host;
2628         ssize_t ret;
2629
2630         BUG_ON(iocb->ki_pos != pos);
2631
2632         mutex_lock(&inode->i_mutex);
2633         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2634                         &iocb->ki_pos);
2635         mutex_unlock(&inode->i_mutex);
2636
2637         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2638                 ssize_t err;
2639
2640                 err = sync_page_range(inode, mapping, pos, ret);
2641                 if (err < 0)
2642                         ret = err;
2643         }
2644         return ret;
2645 }
2646 EXPORT_SYMBOL(generic_file_aio_write);
2647
2648 /**
2649  * try_to_release_page() - release old fs-specific metadata on a page
2650  *
2651  * @page: the page which the kernel is trying to free
2652  * @gfp_mask: memory allocation flags (and I/O mode)
2653  *
2654  * The address_space is to try to release any data against the page
2655  * (presumably at page->private).  If the release was successful, return `1'.
2656  * Otherwise return zero.
2657  *
2658  * The @gfp_mask argument specifies whether I/O may be performed to release
2659  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2660  *
2661  */
2662 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2663 {
2664         struct address_space * const mapping = page->mapping;
2665
2666         BUG_ON(!PageLocked(page));
2667         if (PageWriteback(page))
2668                 return 0;
2669
2670         if (mapping && mapping->a_ops->releasepage)
2671                 return mapping->a_ops->releasepage(page, gfp_mask);
2672         return try_to_free_buffers(page);
2673 }
2674
2675 EXPORT_SYMBOL(try_to_release_page);