Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
[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/backing-dev.h>
32 #include <linux/security.h>
33 #include <linux/syscalls.h>
34 #include <linux/cpuset.h>
35 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
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 static ssize_t
46 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
47         loff_t offset, unsigned long nr_segs);
48
49 /*
50  * Shared mappings implemented 30.11.1994. It's not fully working yet,
51  * though.
52  *
53  * Shared mappings now work. 15.8.1995  Bruno.
54  *
55  * finished 'unifying' the page and buffer cache and SMP-threaded the
56  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57  *
58  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59  */
60
61 /*
62  * Lock ordering:
63  *
64  *  ->i_mmap_lock               (vmtruncate)
65  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
66  *      ->swap_lock             (exclusive_swap_page, others)
67  *        ->mapping->tree_lock
68  *          ->zone.lock
69  *
70  *  ->i_mutex
71  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
72  *
73  *  ->mmap_sem
74  *    ->i_mmap_lock
75  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
76  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
77  *
78  *  ->mmap_sem
79  *    ->lock_page               (access_process_vm)
80  *
81  *  ->i_mutex                   (generic_file_buffered_write)
82  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
83  *
84  *  ->i_mutex
85  *    ->i_alloc_sem             (various)
86  *
87  *  ->inode_lock
88  *    ->sb_lock                 (fs/fs-writeback.c)
89  *    ->mapping->tree_lock      (__sync_single_inode)
90  *
91  *  ->i_mmap_lock
92  *    ->anon_vma.lock           (vma_adjust)
93  *
94  *  ->anon_vma.lock
95  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
96  *
97  *  ->page_table_lock or pte_lock
98  *    ->swap_lock               (try_to_unmap_one)
99  *    ->private_lock            (try_to_unmap_one)
100  *    ->tree_lock               (try_to_unmap_one)
101  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
102  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
103  *    ->private_lock            (page_remove_rmap->set_page_dirty)
104  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
106  *    ->inode_lock              (zap_pte_range->set_page_dirty)
107  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
108  *
109  *  ->task->proc_lock
110  *    ->dcache_lock             (proc_pid_lookup)
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         BUG_ON(page_mapped(page));
127
128         /*
129          * Some filesystems seem to re-dirty the page even after
130          * the VM has canceled the dirty bit (eg ext3 journaling).
131          *
132          * Fix it up by doing a final dirty accounting check after
133          * having removed the page entirely.
134          */
135         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
136                 dec_zone_page_state(page, NR_FILE_DIRTY);
137                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
138         }
139 }
140
141 void remove_from_page_cache(struct page *page)
142 {
143         struct address_space *mapping = page->mapping;
144
145         BUG_ON(!PageLocked(page));
146
147         write_lock_irq(&mapping->tree_lock);
148         __remove_from_page_cache(page);
149         write_unlock_irq(&mapping->tree_lock);
150 }
151
152 static int sync_page(void *word)
153 {
154         struct address_space *mapping;
155         struct page *page;
156
157         page = container_of((unsigned long *)word, struct page, flags);
158
159         /*
160          * page_mapping() is being called without PG_locked held.
161          * Some knowledge of the state and use of the page is used to
162          * reduce the requirements down to a memory barrier.
163          * The danger here is of a stale page_mapping() return value
164          * indicating a struct address_space different from the one it's
165          * associated with when it is associated with one.
166          * After smp_mb(), it's either the correct page_mapping() for
167          * the page, or an old page_mapping() and the page's own
168          * page_mapping() has gone NULL.
169          * The ->sync_page() address_space operation must tolerate
170          * page_mapping() going NULL. By an amazing coincidence,
171          * this comes about because none of the users of the page
172          * in the ->sync_page() methods make essential use of the
173          * page_mapping(), merely passing the page down to the backing
174          * device's unplug functions when it's non-NULL, which in turn
175          * ignore it for all cases but swap, where only page_private(page) is
176          * of interest. When page_mapping() does go NULL, the entire
177          * call stack gracefully ignores the page and returns.
178          * -- wli
179          */
180         smp_mb();
181         mapping = page_mapping(page);
182         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
183                 mapping->a_ops->sync_page(page);
184         io_schedule();
185         return 0;
186 }
187
188 /**
189  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
190  * @mapping:    address space structure to write
191  * @start:      offset in bytes where the range starts
192  * @end:        offset in bytes where the range ends (inclusive)
193  * @sync_mode:  enable synchronous operation
194  *
195  * Start writeback against all of a mapping's dirty pages that lie
196  * within the byte offsets <start, end> inclusive.
197  *
198  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
199  * opposed to a regular memory cleansing writeback.  The difference between
200  * these two operations is that if a dirty page/buffer is encountered, it must
201  * be waited upon, and not just skipped over.
202  */
203 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
204                                 loff_t end, int sync_mode)
205 {
206         int ret;
207         struct writeback_control wbc = {
208                 .sync_mode = sync_mode,
209                 .nr_to_write = mapping->nrpages * 2,
210                 .range_start = start,
211                 .range_end = end,
212         };
213
214         if (!mapping_cap_writeback_dirty(mapping))
215                 return 0;
216
217         ret = do_writepages(mapping, &wbc);
218         return ret;
219 }
220
221 static inline int __filemap_fdatawrite(struct address_space *mapping,
222         int sync_mode)
223 {
224         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
225 }
226
227 int filemap_fdatawrite(struct address_space *mapping)
228 {
229         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
230 }
231 EXPORT_SYMBOL(filemap_fdatawrite);
232
233 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
234                                 loff_t end)
235 {
236         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
237 }
238
239 /**
240  * filemap_flush - mostly a non-blocking flush
241  * @mapping:    target address_space
242  *
243  * This is a mostly non-blocking flush.  Not suitable for data-integrity
244  * purposes - I/O may not be started against all dirty pages.
245  */
246 int filemap_flush(struct address_space *mapping)
247 {
248         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
249 }
250 EXPORT_SYMBOL(filemap_flush);
251
252 /**
253  * wait_on_page_writeback_range - wait for writeback to complete
254  * @mapping:    target address_space
255  * @start:      beginning page index
256  * @end:        ending page index
257  *
258  * Wait for writeback to complete against pages indexed by start->end
259  * inclusive
260  */
261 int wait_on_page_writeback_range(struct address_space *mapping,
262                                 pgoff_t start, pgoff_t end)
263 {
264         struct pagevec pvec;
265         int nr_pages;
266         int ret = 0;
267         pgoff_t index;
268
269         if (end < start)
270                 return 0;
271
272         pagevec_init(&pvec, 0);
273         index = start;
274         while ((index <= end) &&
275                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
276                         PAGECACHE_TAG_WRITEBACK,
277                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
278                 unsigned i;
279
280                 for (i = 0; i < nr_pages; i++) {
281                         struct page *page = pvec.pages[i];
282
283                         /* until radix tree lookup accepts end_index */
284                         if (page->index > end)
285                                 continue;
286
287                         wait_on_page_writeback(page);
288                         if (PageError(page))
289                                 ret = -EIO;
290                 }
291                 pagevec_release(&pvec);
292                 cond_resched();
293         }
294
295         /* Check for outstanding write errors */
296         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_and_clear_bit(AS_EIO, &mapping->flags))
299                 ret = -EIO;
300
301         return ret;
302 }
303
304 /**
305  * sync_page_range - write and wait on all pages in the passed range
306  * @inode:      target inode
307  * @mapping:    target address_space
308  * @pos:        beginning offset in pages to write
309  * @count:      number of bytes to write
310  *
311  * Write and wait upon all the pages in the passed range.  This is a "data
312  * integrity" operation.  It waits upon in-flight writeout before starting and
313  * waiting upon new writeout.  If there was an IO error, return it.
314  *
315  * We need to re-take i_mutex during the generic_osync_inode list walk because
316  * it is otherwise livelockable.
317  */
318 int sync_page_range(struct inode *inode, struct address_space *mapping,
319                         loff_t pos, loff_t count)
320 {
321         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
322         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
323         int ret;
324
325         if (!mapping_cap_writeback_dirty(mapping) || !count)
326                 return 0;
327         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
328         if (ret == 0) {
329                 mutex_lock(&inode->i_mutex);
330                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
331                 mutex_unlock(&inode->i_mutex);
332         }
333         if (ret == 0)
334                 ret = wait_on_page_writeback_range(mapping, start, end);
335         return ret;
336 }
337 EXPORT_SYMBOL(sync_page_range);
338
339 /**
340  * sync_page_range_nolock
341  * @inode:      target inode
342  * @mapping:    target address_space
343  * @pos:        beginning offset in pages to write
344  * @count:      number of bytes to write
345  *
346  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
347  * as it forces O_SYNC writers to different parts of the same file
348  * to be serialised right until io completion.
349  */
350 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
351                            loff_t pos, loff_t count)
352 {
353         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
354         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
355         int ret;
356
357         if (!mapping_cap_writeback_dirty(mapping) || !count)
358                 return 0;
359         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
360         if (ret == 0)
361                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
362         if (ret == 0)
363                 ret = wait_on_page_writeback_range(mapping, start, end);
364         return ret;
365 }
366 EXPORT_SYMBOL(sync_page_range_nolock);
367
368 /**
369  * filemap_fdatawait - wait for all under-writeback pages to complete
370  * @mapping: address space structure to wait for
371  *
372  * Walk the list of under-writeback pages of the given address space
373  * and wait for all of them.
374  */
375 int filemap_fdatawait(struct address_space *mapping)
376 {
377         loff_t i_size = i_size_read(mapping->host);
378
379         if (i_size == 0)
380                 return 0;
381
382         return wait_on_page_writeback_range(mapping, 0,
383                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
384 }
385 EXPORT_SYMBOL(filemap_fdatawait);
386
387 int filemap_write_and_wait(struct address_space *mapping)
388 {
389         int err = 0;
390
391         if (mapping->nrpages) {
392                 err = filemap_fdatawrite(mapping);
393                 /*
394                  * Even if the above returned error, the pages may be
395                  * written partially (e.g. -ENOSPC), so we wait for it.
396                  * But the -EIO is special case, it may indicate the worst
397                  * thing (e.g. bug) happened, so we avoid waiting for it.
398                  */
399                 if (err != -EIO) {
400                         int err2 = filemap_fdatawait(mapping);
401                         if (!err)
402                                 err = err2;
403                 }
404         }
405         return err;
406 }
407 EXPORT_SYMBOL(filemap_write_and_wait);
408
409 /**
410  * filemap_write_and_wait_range - write out & wait on a file range
411  * @mapping:    the address_space for the pages
412  * @lstart:     offset in bytes where the range starts
413  * @lend:       offset in bytes where the range ends (inclusive)
414  *
415  * Write out and wait upon file offsets lstart->lend, inclusive.
416  *
417  * Note that `lend' is inclusive (describes the last byte to be written) so
418  * that this function can be used to write to the very end-of-file (end = -1).
419  */
420 int filemap_write_and_wait_range(struct address_space *mapping,
421                                  loff_t lstart, loff_t lend)
422 {
423         int err = 0;
424
425         if (mapping->nrpages) {
426                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
427                                                  WB_SYNC_ALL);
428                 /* See comment of filemap_write_and_wait() */
429                 if (err != -EIO) {
430                         int err2 = wait_on_page_writeback_range(mapping,
431                                                 lstart >> PAGE_CACHE_SHIFT,
432                                                 lend >> PAGE_CACHE_SHIFT);
433                         if (!err)
434                                 err = err2;
435                 }
436         }
437         return err;
438 }
439
440 /**
441  * add_to_page_cache - add newly allocated pagecache pages
442  * @page:       page to add
443  * @mapping:    the page's address_space
444  * @offset:     page index
445  * @gfp_mask:   page allocation mode
446  *
447  * This function is used to add newly allocated pagecache pages;
448  * the page is new, so we can just run SetPageLocked() against it.
449  * The other page state flags were set by rmqueue().
450  *
451  * This function does not add the page to the LRU.  The caller must do that.
452  */
453 int add_to_page_cache(struct page *page, struct address_space *mapping,
454                 pgoff_t offset, gfp_t gfp_mask)
455 {
456         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
457
458         if (error == 0) {
459                 write_lock_irq(&mapping->tree_lock);
460                 error = radix_tree_insert(&mapping->page_tree, offset, page);
461                 if (!error) {
462                         page_cache_get(page);
463                         SetPageLocked(page);
464                         page->mapping = mapping;
465                         page->index = offset;
466                         mapping->nrpages++;
467                         __inc_zone_page_state(page, NR_FILE_PAGES);
468                 }
469                 write_unlock_irq(&mapping->tree_lock);
470                 radix_tree_preload_end();
471         }
472         return error;
473 }
474 EXPORT_SYMBOL(add_to_page_cache);
475
476 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
477                                 pgoff_t offset, gfp_t gfp_mask)
478 {
479         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
480         if (ret == 0)
481                 lru_cache_add(page);
482         return ret;
483 }
484
485 #ifdef CONFIG_NUMA
486 struct page *__page_cache_alloc(gfp_t gfp)
487 {
488         if (cpuset_do_page_mem_spread()) {
489                 int n = cpuset_mem_spread_node();
490                 return alloc_pages_node(n, gfp, 0);
491         }
492         return alloc_pages(gfp, 0);
493 }
494 EXPORT_SYMBOL(__page_cache_alloc);
495 #endif
496
497 static int __sleep_on_page_lock(void *word)
498 {
499         io_schedule();
500         return 0;
501 }
502
503 /*
504  * In order to wait for pages to become available there must be
505  * waitqueues associated with pages. By using a hash table of
506  * waitqueues where the bucket discipline is to maintain all
507  * waiters on the same queue and wake all when any of the pages
508  * become available, and for the woken contexts to check to be
509  * sure the appropriate page became available, this saves space
510  * at a cost of "thundering herd" phenomena during rare hash
511  * collisions.
512  */
513 static wait_queue_head_t *page_waitqueue(struct page *page)
514 {
515         const struct zone *zone = page_zone(page);
516
517         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
518 }
519
520 static inline void wake_up_page(struct page *page, int bit)
521 {
522         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
523 }
524
525 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
526 {
527         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
528
529         if (test_bit(bit_nr, &page->flags))
530                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
531                                                         TASK_UNINTERRUPTIBLE);
532 }
533 EXPORT_SYMBOL(wait_on_page_bit);
534
535 /**
536  * unlock_page - unlock a locked page
537  * @page: the page
538  *
539  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
540  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
541  * mechananism between PageLocked pages and PageWriteback pages is shared.
542  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
543  *
544  * The first mb is necessary to safely close the critical section opened by the
545  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
546  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
547  * parallel wait_on_page_locked()).
548  */
549 void fastcall unlock_page(struct page *page)
550 {
551         smp_mb__before_clear_bit();
552         if (!TestClearPageLocked(page))
553                 BUG();
554         smp_mb__after_clear_bit(); 
555         wake_up_page(page, PG_locked);
556 }
557 EXPORT_SYMBOL(unlock_page);
558
559 /**
560  * end_page_writeback - end writeback against a page
561  * @page: the page
562  */
563 void end_page_writeback(struct page *page)
564 {
565         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
566                 if (!test_clear_page_writeback(page))
567                         BUG();
568         }
569         smp_mb__after_clear_bit();
570         wake_up_page(page, PG_writeback);
571 }
572 EXPORT_SYMBOL(end_page_writeback);
573
574 /**
575  * __lock_page - get a lock on the page, assuming we need to sleep to get it
576  * @page: the page to lock
577  *
578  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
579  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
580  * chances are that on the second loop, the block layer's plug list is empty,
581  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
582  */
583 void fastcall __lock_page(struct page *page)
584 {
585         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
586
587         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
588                                                         TASK_UNINTERRUPTIBLE);
589 }
590 EXPORT_SYMBOL(__lock_page);
591
592 /*
593  * Variant of lock_page that does not require the caller to hold a reference
594  * on the page's mapping.
595  */
596 void fastcall __lock_page_nosync(struct page *page)
597 {
598         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
599         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
600                                                         TASK_UNINTERRUPTIBLE);
601 }
602
603 /**
604  * find_get_page - find and get a page reference
605  * @mapping: the address_space to search
606  * @offset: the page index
607  *
608  * Is there a pagecache struct page at the given (mapping, offset) tuple?
609  * If yes, increment its refcount and return it; if no, return NULL.
610  */
611 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
612 {
613         struct page *page;
614
615         read_lock_irq(&mapping->tree_lock);
616         page = radix_tree_lookup(&mapping->page_tree, offset);
617         if (page)
618                 page_cache_get(page);
619         read_unlock_irq(&mapping->tree_lock);
620         return page;
621 }
622 EXPORT_SYMBOL(find_get_page);
623
624 /**
625  * find_lock_page - locate, pin and lock a pagecache page
626  * @mapping: the address_space to search
627  * @offset: the page index
628  *
629  * Locates the desired pagecache page, locks it, increments its reference
630  * count and returns its address.
631  *
632  * Returns zero if the page was not present. find_lock_page() may sleep.
633  */
634 struct page *find_lock_page(struct address_space *mapping,
635                                 pgoff_t offset)
636 {
637         struct page *page;
638
639 repeat:
640         read_lock_irq(&mapping->tree_lock);
641         page = radix_tree_lookup(&mapping->page_tree, offset);
642         if (page) {
643                 page_cache_get(page);
644                 if (TestSetPageLocked(page)) {
645                         read_unlock_irq(&mapping->tree_lock);
646                         __lock_page(page);
647
648                         /* Has the page been truncated while we slept? */
649                         if (unlikely(page->mapping != mapping)) {
650                                 unlock_page(page);
651                                 page_cache_release(page);
652                                 goto repeat;
653                         }
654                         VM_BUG_ON(page->index != offset);
655                         goto out;
656                 }
657         }
658         read_unlock_irq(&mapping->tree_lock);
659 out:
660         return page;
661 }
662 EXPORT_SYMBOL(find_lock_page);
663
664 /**
665  * find_or_create_page - locate or add a pagecache page
666  * @mapping: the page's address_space
667  * @index: the page's index into the mapping
668  * @gfp_mask: page allocation mode
669  *
670  * Locates a page in the pagecache.  If the page is not present, a new page
671  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
672  * LRU list.  The returned page is locked and has its reference count
673  * incremented.
674  *
675  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
676  * allocation!
677  *
678  * find_or_create_page() returns the desired page's address, or zero on
679  * memory exhaustion.
680  */
681 struct page *find_or_create_page(struct address_space *mapping,
682                 pgoff_t index, gfp_t gfp_mask)
683 {
684         struct page *page;
685         int err;
686 repeat:
687         page = find_lock_page(mapping, index);
688         if (!page) {
689                 page = __page_cache_alloc(gfp_mask);
690                 if (!page)
691                         return NULL;
692                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
693                 if (unlikely(err)) {
694                         page_cache_release(page);
695                         page = NULL;
696                         if (err == -EEXIST)
697                                 goto repeat;
698                 }
699         }
700         return page;
701 }
702 EXPORT_SYMBOL(find_or_create_page);
703
704 /**
705  * find_get_pages - gang pagecache lookup
706  * @mapping:    The address_space to search
707  * @start:      The starting page index
708  * @nr_pages:   The maximum number of pages
709  * @pages:      Where the resulting pages are placed
710  *
711  * find_get_pages() will search for and return a group of up to
712  * @nr_pages pages in the mapping.  The pages are placed at @pages.
713  * find_get_pages() takes a reference against the returned pages.
714  *
715  * The search returns a group of mapping-contiguous pages with ascending
716  * indexes.  There may be holes in the indices due to not-present pages.
717  *
718  * find_get_pages() returns the number of pages which were found.
719  */
720 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
721                             unsigned int nr_pages, struct page **pages)
722 {
723         unsigned int i;
724         unsigned int ret;
725
726         read_lock_irq(&mapping->tree_lock);
727         ret = radix_tree_gang_lookup(&mapping->page_tree,
728                                 (void **)pages, start, nr_pages);
729         for (i = 0; i < ret; i++)
730                 page_cache_get(pages[i]);
731         read_unlock_irq(&mapping->tree_lock);
732         return ret;
733 }
734
735 /**
736  * find_get_pages_contig - gang contiguous pagecache lookup
737  * @mapping:    The address_space to search
738  * @index:      The starting page index
739  * @nr_pages:   The maximum number of pages
740  * @pages:      Where the resulting pages are placed
741  *
742  * find_get_pages_contig() works exactly like find_get_pages(), except
743  * that the returned number of pages are guaranteed to be contiguous.
744  *
745  * find_get_pages_contig() returns the number of pages which were found.
746  */
747 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
748                                unsigned int nr_pages, struct page **pages)
749 {
750         unsigned int i;
751         unsigned int ret;
752
753         read_lock_irq(&mapping->tree_lock);
754         ret = radix_tree_gang_lookup(&mapping->page_tree,
755                                 (void **)pages, index, nr_pages);
756         for (i = 0; i < ret; i++) {
757                 if (pages[i]->mapping == NULL || pages[i]->index != index)
758                         break;
759
760                 page_cache_get(pages[i]);
761                 index++;
762         }
763         read_unlock_irq(&mapping->tree_lock);
764         return i;
765 }
766 EXPORT_SYMBOL(find_get_pages_contig);
767
768 /**
769  * find_get_pages_tag - find and return pages that match @tag
770  * @mapping:    the address_space to search
771  * @index:      the starting page index
772  * @tag:        the tag index
773  * @nr_pages:   the maximum number of pages
774  * @pages:      where the resulting pages are placed
775  *
776  * Like find_get_pages, except we only return pages which are tagged with
777  * @tag.   We update @index to index the next page for the traversal.
778  */
779 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
780                         int tag, unsigned int nr_pages, struct page **pages)
781 {
782         unsigned int i;
783         unsigned int ret;
784
785         read_lock_irq(&mapping->tree_lock);
786         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
787                                 (void **)pages, *index, nr_pages, tag);
788         for (i = 0; i < ret; i++)
789                 page_cache_get(pages[i]);
790         if (ret)
791                 *index = pages[ret - 1]->index + 1;
792         read_unlock_irq(&mapping->tree_lock);
793         return ret;
794 }
795 EXPORT_SYMBOL(find_get_pages_tag);
796
797 /**
798  * grab_cache_page_nowait - returns locked page at given index in given cache
799  * @mapping: target address_space
800  * @index: the page index
801  *
802  * Same as grab_cache_page(), but do not wait if the page is unavailable.
803  * This is intended for speculative data generators, where the data can
804  * be regenerated if the page couldn't be grabbed.  This routine should
805  * be safe to call while holding the lock for another page.
806  *
807  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
808  * and deadlock against the caller's locked page.
809  */
810 struct page *
811 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
812 {
813         struct page *page = find_get_page(mapping, index);
814
815         if (page) {
816                 if (!TestSetPageLocked(page))
817                         return page;
818                 page_cache_release(page);
819                 return NULL;
820         }
821         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
822         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
823                 page_cache_release(page);
824                 page = NULL;
825         }
826         return page;
827 }
828 EXPORT_SYMBOL(grab_cache_page_nowait);
829
830 /*
831  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
832  * a _large_ part of the i/o request. Imagine the worst scenario:
833  *
834  *      ---R__________________________________________B__________
835  *         ^ reading here                             ^ bad block(assume 4k)
836  *
837  * read(R) => miss => readahead(R...B) => media error => frustrating retries
838  * => failing the whole request => read(R) => read(R+1) =>
839  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
840  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
841  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
842  *
843  * It is going insane. Fix it by quickly scaling down the readahead size.
844  */
845 static void shrink_readahead_size_eio(struct file *filp,
846                                         struct file_ra_state *ra)
847 {
848         if (!ra->ra_pages)
849                 return;
850
851         ra->ra_pages /= 4;
852 }
853
854 /**
855  * do_generic_mapping_read - generic file read routine
856  * @mapping:    address_space to be read
857  * @ra:         file's readahead state
858  * @filp:       the file to read
859  * @ppos:       current file position
860  * @desc:       read_descriptor
861  * @actor:      read method
862  *
863  * This is a generic file read routine, and uses the
864  * mapping->a_ops->readpage() function for the actual low-level stuff.
865  *
866  * This is really ugly. But the goto's actually try to clarify some
867  * of the logic when it comes to error handling etc.
868  *
869  * Note the struct file* is only passed for the use of readpage.
870  * It may be NULL.
871  */
872 void do_generic_mapping_read(struct address_space *mapping,
873                              struct file_ra_state *ra,
874                              struct file *filp,
875                              loff_t *ppos,
876                              read_descriptor_t *desc,
877                              read_actor_t actor)
878 {
879         struct inode *inode = mapping->host;
880         pgoff_t index;
881         pgoff_t last_index;
882         pgoff_t prev_index;
883         unsigned long offset;      /* offset into pagecache page */
884         unsigned int prev_offset;
885         int error;
886
887         index = *ppos >> PAGE_CACHE_SHIFT;
888         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
889         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
890         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
891         offset = *ppos & ~PAGE_CACHE_MASK;
892
893         for (;;) {
894                 struct page *page;
895                 pgoff_t end_index;
896                 loff_t isize;
897                 unsigned long nr, ret;
898
899                 cond_resched();
900 find_page:
901                 page = find_get_page(mapping, index);
902                 if (!page) {
903                         page_cache_sync_readahead(mapping,
904                                         ra, filp,
905                                         index, last_index - index);
906                         page = find_get_page(mapping, index);
907                         if (unlikely(page == NULL))
908                                 goto no_cached_page;
909                 }
910                 if (PageReadahead(page)) {
911                         page_cache_async_readahead(mapping,
912                                         ra, filp, page,
913                                         index, last_index - index);
914                 }
915                 if (!PageUptodate(page))
916                         goto page_not_up_to_date;
917 page_ok:
918                 /*
919                  * i_size must be checked after we know the page is Uptodate.
920                  *
921                  * Checking i_size after the check allows us to calculate
922                  * the correct value for "nr", which means the zero-filled
923                  * part of the page is not copied back to userspace (unless
924                  * another truncate extends the file - this is desired though).
925                  */
926
927                 isize = i_size_read(inode);
928                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
929                 if (unlikely(!isize || index > end_index)) {
930                         page_cache_release(page);
931                         goto out;
932                 }
933
934                 /* nr is the maximum number of bytes to copy from this page */
935                 nr = PAGE_CACHE_SIZE;
936                 if (index == end_index) {
937                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
938                         if (nr <= offset) {
939                                 page_cache_release(page);
940                                 goto out;
941                         }
942                 }
943                 nr = nr - offset;
944
945                 /* If users can be writing to this page using arbitrary
946                  * virtual addresses, take care about potential aliasing
947                  * before reading the page on the kernel side.
948                  */
949                 if (mapping_writably_mapped(mapping))
950                         flush_dcache_page(page);
951
952                 /*
953                  * When a sequential read accesses a page several times,
954                  * only mark it as accessed the first time.
955                  */
956                 if (prev_index != index || offset != prev_offset)
957                         mark_page_accessed(page);
958                 prev_index = index;
959
960                 /*
961                  * Ok, we have the page, and it's up-to-date, so
962                  * now we can copy it to user space...
963                  *
964                  * The actor routine returns how many bytes were actually used..
965                  * NOTE! This may not be the same as how much of a user buffer
966                  * we filled up (we may be padding etc), so we can only update
967                  * "pos" here (the actor routine has to update the user buffer
968                  * pointers and the remaining count).
969                  */
970                 ret = actor(desc, page, offset, nr);
971                 offset += ret;
972                 index += offset >> PAGE_CACHE_SHIFT;
973                 offset &= ~PAGE_CACHE_MASK;
974                 prev_offset = offset;
975
976                 page_cache_release(page);
977                 if (ret == nr && desc->count)
978                         continue;
979                 goto out;
980
981 page_not_up_to_date:
982                 /* Get exclusive access to the page ... */
983                 lock_page(page);
984
985                 /* Did it get truncated before we got the lock? */
986                 if (!page->mapping) {
987                         unlock_page(page);
988                         page_cache_release(page);
989                         continue;
990                 }
991
992                 /* Did somebody else fill it already? */
993                 if (PageUptodate(page)) {
994                         unlock_page(page);
995                         goto page_ok;
996                 }
997
998 readpage:
999                 /* Start the actual read. The read will unlock the page. */
1000                 error = mapping->a_ops->readpage(filp, page);
1001
1002                 if (unlikely(error)) {
1003                         if (error == AOP_TRUNCATED_PAGE) {
1004                                 page_cache_release(page);
1005                                 goto find_page;
1006                         }
1007                         goto readpage_error;
1008                 }
1009
1010                 if (!PageUptodate(page)) {
1011                         lock_page(page);
1012                         if (!PageUptodate(page)) {
1013                                 if (page->mapping == NULL) {
1014                                         /*
1015                                          * invalidate_inode_pages got it
1016                                          */
1017                                         unlock_page(page);
1018                                         page_cache_release(page);
1019                                         goto find_page;
1020                                 }
1021                                 unlock_page(page);
1022                                 error = -EIO;
1023                                 shrink_readahead_size_eio(filp, ra);
1024                                 goto readpage_error;
1025                         }
1026                         unlock_page(page);
1027                 }
1028
1029                 goto page_ok;
1030
1031 readpage_error:
1032                 /* UHHUH! A synchronous read error occurred. Report it */
1033                 desc->error = error;
1034                 page_cache_release(page);
1035                 goto out;
1036
1037 no_cached_page:
1038                 /*
1039                  * Ok, it wasn't cached, so we need to create a new
1040                  * page..
1041                  */
1042                 page = page_cache_alloc_cold(mapping);
1043                 if (!page) {
1044                         desc->error = -ENOMEM;
1045                         goto out;
1046                 }
1047                 error = add_to_page_cache_lru(page, mapping,
1048                                                 index, GFP_KERNEL);
1049                 if (error) {
1050                         page_cache_release(page);
1051                         if (error == -EEXIST)
1052                                 goto find_page;
1053                         desc->error = error;
1054                         goto out;
1055                 }
1056                 goto readpage;
1057         }
1058
1059 out:
1060         ra->prev_pos = prev_index;
1061         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1062         ra->prev_pos |= prev_offset;
1063
1064         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1065         if (filp)
1066                 file_accessed(filp);
1067 }
1068 EXPORT_SYMBOL(do_generic_mapping_read);
1069
1070 int file_read_actor(read_descriptor_t *desc, struct page *page,
1071                         unsigned long offset, unsigned long size)
1072 {
1073         char *kaddr;
1074         unsigned long left, count = desc->count;
1075
1076         if (size > count)
1077                 size = count;
1078
1079         /*
1080          * Faults on the destination of a read are common, so do it before
1081          * taking the kmap.
1082          */
1083         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1084                 kaddr = kmap_atomic(page, KM_USER0);
1085                 left = __copy_to_user_inatomic(desc->arg.buf,
1086                                                 kaddr + offset, size);
1087                 kunmap_atomic(kaddr, KM_USER0);
1088                 if (left == 0)
1089                         goto success;
1090         }
1091
1092         /* Do it the slow way */
1093         kaddr = kmap(page);
1094         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1095         kunmap(page);
1096
1097         if (left) {
1098                 size -= left;
1099                 desc->error = -EFAULT;
1100         }
1101 success:
1102         desc->count = count - size;
1103         desc->written += size;
1104         desc->arg.buf += size;
1105         return size;
1106 }
1107
1108 /*
1109  * Performs necessary checks before doing a write
1110  * @iov:        io vector request
1111  * @nr_segs:    number of segments in the iovec
1112  * @count:      number of bytes to write
1113  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1114  *
1115  * Adjust number of segments and amount of bytes to write (nr_segs should be
1116  * properly initialized first). Returns appropriate error code that caller
1117  * should return or zero in case that write should be allowed.
1118  */
1119 int generic_segment_checks(const struct iovec *iov,
1120                         unsigned long *nr_segs, size_t *count, int access_flags)
1121 {
1122         unsigned long   seg;
1123         size_t cnt = 0;
1124         for (seg = 0; seg < *nr_segs; seg++) {
1125                 const struct iovec *iv = &iov[seg];
1126
1127                 /*
1128                  * If any segment has a negative length, or the cumulative
1129                  * length ever wraps negative then return -EINVAL.
1130                  */
1131                 cnt += iv->iov_len;
1132                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1133                         return -EINVAL;
1134                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1135                         continue;
1136                 if (seg == 0)
1137                         return -EFAULT;
1138                 *nr_segs = seg;
1139                 cnt -= iv->iov_len;     /* This segment is no good */
1140                 break;
1141         }
1142         *count = cnt;
1143         return 0;
1144 }
1145 EXPORT_SYMBOL(generic_segment_checks);
1146
1147 /**
1148  * generic_file_aio_read - generic filesystem read routine
1149  * @iocb:       kernel I/O control block
1150  * @iov:        io vector request
1151  * @nr_segs:    number of segments in the iovec
1152  * @pos:        current file position
1153  *
1154  * This is the "read()" routine for all filesystems
1155  * that can use the page cache directly.
1156  */
1157 ssize_t
1158 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1159                 unsigned long nr_segs, loff_t pos)
1160 {
1161         struct file *filp = iocb->ki_filp;
1162         ssize_t retval;
1163         unsigned long seg;
1164         size_t count;
1165         loff_t *ppos = &iocb->ki_pos;
1166
1167         count = 0;
1168         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1169         if (retval)
1170                 return retval;
1171
1172         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1173         if (filp->f_flags & O_DIRECT) {
1174                 loff_t size;
1175                 struct address_space *mapping;
1176                 struct inode *inode;
1177
1178                 mapping = filp->f_mapping;
1179                 inode = mapping->host;
1180                 retval = 0;
1181                 if (!count)
1182                         goto out; /* skip atime */
1183                 size = i_size_read(inode);
1184                 if (pos < size) {
1185                         retval = generic_file_direct_IO(READ, iocb,
1186                                                 iov, pos, nr_segs);
1187                         if (retval > 0)
1188                                 *ppos = pos + retval;
1189                 }
1190                 if (likely(retval != 0)) {
1191                         file_accessed(filp);
1192                         goto out;
1193                 }
1194         }
1195
1196         retval = 0;
1197         if (count) {
1198                 for (seg = 0; seg < nr_segs; seg++) {
1199                         read_descriptor_t desc;
1200
1201                         desc.written = 0;
1202                         desc.arg.buf = iov[seg].iov_base;
1203                         desc.count = iov[seg].iov_len;
1204                         if (desc.count == 0)
1205                                 continue;
1206                         desc.error = 0;
1207                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1208                         retval += desc.written;
1209                         if (desc.error) {
1210                                 retval = retval ?: desc.error;
1211                                 break;
1212                         }
1213                         if (desc.count > 0)
1214                                 break;
1215                 }
1216         }
1217 out:
1218         return retval;
1219 }
1220 EXPORT_SYMBOL(generic_file_aio_read);
1221
1222 static ssize_t
1223 do_readahead(struct address_space *mapping, struct file *filp,
1224              pgoff_t index, unsigned long nr)
1225 {
1226         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1227                 return -EINVAL;
1228
1229         force_page_cache_readahead(mapping, filp, index,
1230                                         max_sane_readahead(nr));
1231         return 0;
1232 }
1233
1234 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1235 {
1236         ssize_t ret;
1237         struct file *file;
1238
1239         ret = -EBADF;
1240         file = fget(fd);
1241         if (file) {
1242                 if (file->f_mode & FMODE_READ) {
1243                         struct address_space *mapping = file->f_mapping;
1244                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1245                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1246                         unsigned long len = end - start + 1;
1247                         ret = do_readahead(mapping, file, start, len);
1248                 }
1249                 fput(file);
1250         }
1251         return ret;
1252 }
1253
1254 #ifdef CONFIG_MMU
1255 /**
1256  * page_cache_read - adds requested page to the page cache if not already there
1257  * @file:       file to read
1258  * @offset:     page index
1259  *
1260  * This adds the requested page to the page cache if it isn't already there,
1261  * and schedules an I/O to read in its contents from disk.
1262  */
1263 static int fastcall page_cache_read(struct file * file, pgoff_t offset)
1264 {
1265         struct address_space *mapping = file->f_mapping;
1266         struct page *page; 
1267         int ret;
1268
1269         do {
1270                 page = page_cache_alloc_cold(mapping);
1271                 if (!page)
1272                         return -ENOMEM;
1273
1274                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1275                 if (ret == 0)
1276                         ret = mapping->a_ops->readpage(file, page);
1277                 else if (ret == -EEXIST)
1278                         ret = 0; /* losing race to add is OK */
1279
1280                 page_cache_release(page);
1281
1282         } while (ret == AOP_TRUNCATED_PAGE);
1283                 
1284         return ret;
1285 }
1286
1287 #define MMAP_LOTSAMISS  (100)
1288
1289 /**
1290  * filemap_fault - read in file data for page fault handling
1291  * @vma:        vma in which the fault was taken
1292  * @vmf:        struct vm_fault containing details of the fault
1293  *
1294  * filemap_fault() is invoked via the vma operations vector for a
1295  * mapped memory region to read in file data during a page fault.
1296  *
1297  * The goto's are kind of ugly, but this streamlines the normal case of having
1298  * it in the page cache, and handles the special cases reasonably without
1299  * having a lot of duplicated code.
1300  */
1301 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1302 {
1303         int error;
1304         struct file *file = vma->vm_file;
1305         struct address_space *mapping = file->f_mapping;
1306         struct file_ra_state *ra = &file->f_ra;
1307         struct inode *inode = mapping->host;
1308         struct page *page;
1309         unsigned long size;
1310         int did_readaround = 0;
1311         int ret = 0;
1312
1313         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1314         if (vmf->pgoff >= size)
1315                 return VM_FAULT_SIGBUS;
1316
1317         /* If we don't want any read-ahead, don't bother */
1318         if (VM_RandomReadHint(vma))
1319                 goto no_cached_page;
1320
1321         /*
1322          * Do we have something in the page cache already?
1323          */
1324 retry_find:
1325         page = find_lock_page(mapping, vmf->pgoff);
1326         /*
1327          * For sequential accesses, we use the generic readahead logic.
1328          */
1329         if (VM_SequentialReadHint(vma)) {
1330                 if (!page) {
1331                         page_cache_sync_readahead(mapping, ra, file,
1332                                                            vmf->pgoff, 1);
1333                         page = find_lock_page(mapping, vmf->pgoff);
1334                         if (!page)
1335                                 goto no_cached_page;
1336                 }
1337                 if (PageReadahead(page)) {
1338                         page_cache_async_readahead(mapping, ra, file, page,
1339                                                            vmf->pgoff, 1);
1340                 }
1341         }
1342
1343         if (!page) {
1344                 unsigned long ra_pages;
1345
1346                 ra->mmap_miss++;
1347
1348                 /*
1349                  * Do we miss much more than hit in this file? If so,
1350                  * stop bothering with read-ahead. It will only hurt.
1351                  */
1352                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1353                         goto no_cached_page;
1354
1355                 /*
1356                  * To keep the pgmajfault counter straight, we need to
1357                  * check did_readaround, as this is an inner loop.
1358                  */
1359                 if (!did_readaround) {
1360                         ret = VM_FAULT_MAJOR;
1361                         count_vm_event(PGMAJFAULT);
1362                 }
1363                 did_readaround = 1;
1364                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1365                 if (ra_pages) {
1366                         pgoff_t start = 0;
1367
1368                         if (vmf->pgoff > ra_pages / 2)
1369                                 start = vmf->pgoff - ra_pages / 2;
1370                         do_page_cache_readahead(mapping, file, start, ra_pages);
1371                 }
1372                 page = find_lock_page(mapping, vmf->pgoff);
1373                 if (!page)
1374                         goto no_cached_page;
1375         }
1376
1377         if (!did_readaround)
1378                 ra->mmap_miss--;
1379
1380         /*
1381          * We have a locked page in the page cache, now we need to check
1382          * that it's up-to-date. If not, it is going to be due to an error.
1383          */
1384         if (unlikely(!PageUptodate(page)))
1385                 goto page_not_uptodate;
1386
1387         /* Must recheck i_size under page lock */
1388         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1389         if (unlikely(vmf->pgoff >= size)) {
1390                 unlock_page(page);
1391                 page_cache_release(page);
1392                 return VM_FAULT_SIGBUS;
1393         }
1394
1395         /*
1396          * Found the page and have a reference on it.
1397          */
1398         mark_page_accessed(page);
1399         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1400         vmf->page = page;
1401         return ret | VM_FAULT_LOCKED;
1402
1403 no_cached_page:
1404         /*
1405          * We're only likely to ever get here if MADV_RANDOM is in
1406          * effect.
1407          */
1408         error = page_cache_read(file, vmf->pgoff);
1409
1410         /*
1411          * The page we want has now been added to the page cache.
1412          * In the unlikely event that someone removed it in the
1413          * meantime, we'll just come back here and read it again.
1414          */
1415         if (error >= 0)
1416                 goto retry_find;
1417
1418         /*
1419          * An error return from page_cache_read can result if the
1420          * system is low on memory, or a problem occurs while trying
1421          * to schedule I/O.
1422          */
1423         if (error == -ENOMEM)
1424                 return VM_FAULT_OOM;
1425         return VM_FAULT_SIGBUS;
1426
1427 page_not_uptodate:
1428         /* IO error path */
1429         if (!did_readaround) {
1430                 ret = VM_FAULT_MAJOR;
1431                 count_vm_event(PGMAJFAULT);
1432         }
1433
1434         /*
1435          * Umm, take care of errors if the page isn't up-to-date.
1436          * Try to re-read it _once_. We do this synchronously,
1437          * because there really aren't any performance issues here
1438          * and we need to check for errors.
1439          */
1440         ClearPageError(page);
1441         error = mapping->a_ops->readpage(file, page);
1442         page_cache_release(page);
1443
1444         if (!error || error == AOP_TRUNCATED_PAGE)
1445                 goto retry_find;
1446
1447         /* Things didn't work out. Return zero to tell the mm layer so. */
1448         shrink_readahead_size_eio(file, ra);
1449         return VM_FAULT_SIGBUS;
1450 }
1451 EXPORT_SYMBOL(filemap_fault);
1452
1453 struct vm_operations_struct generic_file_vm_ops = {
1454         .fault          = filemap_fault,
1455 };
1456
1457 /* This is used for a general mmap of a disk file */
1458
1459 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1460 {
1461         struct address_space *mapping = file->f_mapping;
1462
1463         if (!mapping->a_ops->readpage)
1464                 return -ENOEXEC;
1465         file_accessed(file);
1466         vma->vm_ops = &generic_file_vm_ops;
1467         vma->vm_flags |= VM_CAN_NONLINEAR;
1468         return 0;
1469 }
1470
1471 /*
1472  * This is for filesystems which do not implement ->writepage.
1473  */
1474 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1475 {
1476         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1477                 return -EINVAL;
1478         return generic_file_mmap(file, vma);
1479 }
1480 #else
1481 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1482 {
1483         return -ENOSYS;
1484 }
1485 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1486 {
1487         return -ENOSYS;
1488 }
1489 #endif /* CONFIG_MMU */
1490
1491 EXPORT_SYMBOL(generic_file_mmap);
1492 EXPORT_SYMBOL(generic_file_readonly_mmap);
1493
1494 static struct page *__read_cache_page(struct address_space *mapping,
1495                                 pgoff_t index,
1496                                 int (*filler)(void *,struct page*),
1497                                 void *data)
1498 {
1499         struct page *page;
1500         int err;
1501 repeat:
1502         page = find_get_page(mapping, index);
1503         if (!page) {
1504                 page = page_cache_alloc_cold(mapping);
1505                 if (!page)
1506                         return ERR_PTR(-ENOMEM);
1507                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1508                 if (unlikely(err)) {
1509                         page_cache_release(page);
1510                         if (err == -EEXIST)
1511                                 goto repeat;
1512                         /* Presumably ENOMEM for radix tree node */
1513                         return ERR_PTR(err);
1514                 }
1515                 err = filler(data, page);
1516                 if (err < 0) {
1517                         page_cache_release(page);
1518                         page = ERR_PTR(err);
1519                 }
1520         }
1521         return page;
1522 }
1523
1524 /*
1525  * Same as read_cache_page, but don't wait for page to become unlocked
1526  * after submitting it to the filler.
1527  */
1528 struct page *read_cache_page_async(struct address_space *mapping,
1529                                 pgoff_t index,
1530                                 int (*filler)(void *,struct page*),
1531                                 void *data)
1532 {
1533         struct page *page;
1534         int err;
1535
1536 retry:
1537         page = __read_cache_page(mapping, index, filler, data);
1538         if (IS_ERR(page))
1539                 return page;
1540         if (PageUptodate(page))
1541                 goto out;
1542
1543         lock_page(page);
1544         if (!page->mapping) {
1545                 unlock_page(page);
1546                 page_cache_release(page);
1547                 goto retry;
1548         }
1549         if (PageUptodate(page)) {
1550                 unlock_page(page);
1551                 goto out;
1552         }
1553         err = filler(data, page);
1554         if (err < 0) {
1555                 page_cache_release(page);
1556                 return ERR_PTR(err);
1557         }
1558 out:
1559         mark_page_accessed(page);
1560         return page;
1561 }
1562 EXPORT_SYMBOL(read_cache_page_async);
1563
1564 /**
1565  * read_cache_page - read into page cache, fill it if needed
1566  * @mapping:    the page's address_space
1567  * @index:      the page index
1568  * @filler:     function to perform the read
1569  * @data:       destination for read data
1570  *
1571  * Read into the page cache. If a page already exists, and PageUptodate() is
1572  * not set, try to fill the page then wait for it to become unlocked.
1573  *
1574  * If the page does not get brought uptodate, return -EIO.
1575  */
1576 struct page *read_cache_page(struct address_space *mapping,
1577                                 pgoff_t index,
1578                                 int (*filler)(void *,struct page*),
1579                                 void *data)
1580 {
1581         struct page *page;
1582
1583         page = read_cache_page_async(mapping, index, filler, data);
1584         if (IS_ERR(page))
1585                 goto out;
1586         wait_on_page_locked(page);
1587         if (!PageUptodate(page)) {
1588                 page_cache_release(page);
1589                 page = ERR_PTR(-EIO);
1590         }
1591  out:
1592         return page;
1593 }
1594 EXPORT_SYMBOL(read_cache_page);
1595
1596 /*
1597  * The logic we want is
1598  *
1599  *      if suid or (sgid and xgrp)
1600  *              remove privs
1601  */
1602 int should_remove_suid(struct dentry *dentry)
1603 {
1604         mode_t mode = dentry->d_inode->i_mode;
1605         int kill = 0;
1606
1607         /* suid always must be killed */
1608         if (unlikely(mode & S_ISUID))
1609                 kill = ATTR_KILL_SUID;
1610
1611         /*
1612          * sgid without any exec bits is just a mandatory locking mark; leave
1613          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1614          */
1615         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1616                 kill |= ATTR_KILL_SGID;
1617
1618         if (unlikely(kill && !capable(CAP_FSETID)))
1619                 return kill;
1620
1621         return 0;
1622 }
1623 EXPORT_SYMBOL(should_remove_suid);
1624
1625 int __remove_suid(struct dentry *dentry, int kill)
1626 {
1627         struct iattr newattrs;
1628
1629         newattrs.ia_valid = ATTR_FORCE | kill;
1630         return notify_change(dentry, &newattrs);
1631 }
1632
1633 int remove_suid(struct dentry *dentry)
1634 {
1635         int killsuid = should_remove_suid(dentry);
1636         int killpriv = security_inode_need_killpriv(dentry);
1637         int error = 0;
1638
1639         if (killpriv < 0)
1640                 return killpriv;
1641         if (killpriv)
1642                 error = security_inode_killpriv(dentry);
1643         if (!error && killsuid)
1644                 error = __remove_suid(dentry, killsuid);
1645
1646         return error;
1647 }
1648 EXPORT_SYMBOL(remove_suid);
1649
1650 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1651                         const struct iovec *iov, size_t base, size_t bytes)
1652 {
1653         size_t copied = 0, left = 0;
1654
1655         while (bytes) {
1656                 char __user *buf = iov->iov_base + base;
1657                 int copy = min(bytes, iov->iov_len - base);
1658
1659                 base = 0;
1660                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1661                 copied += copy;
1662                 bytes -= copy;
1663                 vaddr += copy;
1664                 iov++;
1665
1666                 if (unlikely(left))
1667                         break;
1668         }
1669         return copied - left;
1670 }
1671
1672 /*
1673  * Copy as much as we can into the page and return the number of bytes which
1674  * were sucessfully copied.  If a fault is encountered then return the number of
1675  * bytes which were copied.
1676  */
1677 size_t iov_iter_copy_from_user_atomic(struct page *page,
1678                 struct iov_iter *i, unsigned long offset, size_t bytes)
1679 {
1680         char *kaddr;
1681         size_t copied;
1682
1683         BUG_ON(!in_atomic());
1684         kaddr = kmap_atomic(page, KM_USER0);
1685         if (likely(i->nr_segs == 1)) {
1686                 int left;
1687                 char __user *buf = i->iov->iov_base + i->iov_offset;
1688                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1689                                                         buf, bytes);
1690                 copied = bytes - left;
1691         } else {
1692                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1693                                                 i->iov, i->iov_offset, bytes);
1694         }
1695         kunmap_atomic(kaddr, KM_USER0);
1696
1697         return copied;
1698 }
1699 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1700
1701 /*
1702  * This has the same sideeffects and return value as
1703  * iov_iter_copy_from_user_atomic().
1704  * The difference is that it attempts to resolve faults.
1705  * Page must not be locked.
1706  */
1707 size_t iov_iter_copy_from_user(struct page *page,
1708                 struct iov_iter *i, unsigned long offset, size_t bytes)
1709 {
1710         char *kaddr;
1711         size_t copied;
1712
1713         kaddr = kmap(page);
1714         if (likely(i->nr_segs == 1)) {
1715                 int left;
1716                 char __user *buf = i->iov->iov_base + i->iov_offset;
1717                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1718                 copied = bytes - left;
1719         } else {
1720                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1721                                                 i->iov, i->iov_offset, bytes);
1722         }
1723         kunmap(page);
1724         return copied;
1725 }
1726 EXPORT_SYMBOL(iov_iter_copy_from_user);
1727
1728 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1729 {
1730         if (likely(i->nr_segs == 1)) {
1731                 i->iov_offset += bytes;
1732         } else {
1733                 const struct iovec *iov = i->iov;
1734                 size_t base = i->iov_offset;
1735
1736                 while (bytes) {
1737                         int copy = min(bytes, iov->iov_len - base);
1738
1739                         bytes -= copy;
1740                         base += copy;
1741                         if (iov->iov_len == base) {
1742                                 iov++;
1743                                 base = 0;
1744                         }
1745                 }
1746                 i->iov = iov;
1747                 i->iov_offset = base;
1748         }
1749 }
1750
1751 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1752 {
1753         BUG_ON(i->count < bytes);
1754
1755         __iov_iter_advance_iov(i, bytes);
1756         i->count -= bytes;
1757 }
1758 EXPORT_SYMBOL(iov_iter_advance);
1759
1760 /*
1761  * Fault in the first iovec of the given iov_iter, to a maximum length
1762  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1763  * accessed (ie. because it is an invalid address).
1764  *
1765  * writev-intensive code may want this to prefault several iovecs -- that
1766  * would be possible (callers must not rely on the fact that _only_ the
1767  * first iovec will be faulted with the current implementation).
1768  */
1769 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1770 {
1771         char __user *buf = i->iov->iov_base + i->iov_offset;
1772         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1773         return fault_in_pages_readable(buf, bytes);
1774 }
1775 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1776
1777 /*
1778  * Return the count of just the current iov_iter segment.
1779  */
1780 size_t iov_iter_single_seg_count(struct iov_iter *i)
1781 {
1782         const struct iovec *iov = i->iov;
1783         if (i->nr_segs == 1)
1784                 return i->count;
1785         else
1786                 return min(i->count, iov->iov_len - i->iov_offset);
1787 }
1788 EXPORT_SYMBOL(iov_iter_single_seg_count);
1789
1790 /*
1791  * Performs necessary checks before doing a write
1792  *
1793  * Can adjust writing position or amount of bytes to write.
1794  * Returns appropriate error code that caller should return or
1795  * zero in case that write should be allowed.
1796  */
1797 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1798 {
1799         struct inode *inode = file->f_mapping->host;
1800         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1801
1802         if (unlikely(*pos < 0))
1803                 return -EINVAL;
1804
1805         if (!isblk) {
1806                 /* FIXME: this is for backwards compatibility with 2.4 */
1807                 if (file->f_flags & O_APPEND)
1808                         *pos = i_size_read(inode);
1809
1810                 if (limit != RLIM_INFINITY) {
1811                         if (*pos >= limit) {
1812                                 send_sig(SIGXFSZ, current, 0);
1813                                 return -EFBIG;
1814                         }
1815                         if (*count > limit - (typeof(limit))*pos) {
1816                                 *count = limit - (typeof(limit))*pos;
1817                         }
1818                 }
1819         }
1820
1821         /*
1822          * LFS rule
1823          */
1824         if (unlikely(*pos + *count > MAX_NON_LFS &&
1825                                 !(file->f_flags & O_LARGEFILE))) {
1826                 if (*pos >= MAX_NON_LFS) {
1827                         return -EFBIG;
1828                 }
1829                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1830                         *count = MAX_NON_LFS - (unsigned long)*pos;
1831                 }
1832         }
1833
1834         /*
1835          * Are we about to exceed the fs block limit ?
1836          *
1837          * If we have written data it becomes a short write.  If we have
1838          * exceeded without writing data we send a signal and return EFBIG.
1839          * Linus frestrict idea will clean these up nicely..
1840          */
1841         if (likely(!isblk)) {
1842                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1843                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1844                                 return -EFBIG;
1845                         }
1846                         /* zero-length writes at ->s_maxbytes are OK */
1847                 }
1848
1849                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1850                         *count = inode->i_sb->s_maxbytes - *pos;
1851         } else {
1852 #ifdef CONFIG_BLOCK
1853                 loff_t isize;
1854                 if (bdev_read_only(I_BDEV(inode)))
1855                         return -EPERM;
1856                 isize = i_size_read(inode);
1857                 if (*pos >= isize) {
1858                         if (*count || *pos > isize)
1859                                 return -ENOSPC;
1860                 }
1861
1862                 if (*pos + *count > isize)
1863                         *count = isize - *pos;
1864 #else
1865                 return -EPERM;
1866 #endif
1867         }
1868         return 0;
1869 }
1870 EXPORT_SYMBOL(generic_write_checks);
1871
1872 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1873                                 loff_t pos, unsigned len, unsigned flags,
1874                                 struct page **pagep, void **fsdata)
1875 {
1876         const struct address_space_operations *aops = mapping->a_ops;
1877
1878         if (aops->write_begin) {
1879                 return aops->write_begin(file, mapping, pos, len, flags,
1880                                                         pagep, fsdata);
1881         } else {
1882                 int ret;
1883                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1884                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1885                 struct inode *inode = mapping->host;
1886                 struct page *page;
1887 again:
1888                 page = __grab_cache_page(mapping, index);
1889                 *pagep = page;
1890                 if (!page)
1891                         return -ENOMEM;
1892
1893                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1894                         /*
1895                          * There is no way to resolve a short write situation
1896                          * for a !Uptodate page (except by double copying in
1897                          * the caller done by generic_perform_write_2copy).
1898                          *
1899                          * Instead, we have to bring it uptodate here.
1900                          */
1901                         ret = aops->readpage(file, page);
1902                         page_cache_release(page);
1903                         if (ret) {
1904                                 if (ret == AOP_TRUNCATED_PAGE)
1905                                         goto again;
1906                                 return ret;
1907                         }
1908                         goto again;
1909                 }
1910
1911                 ret = aops->prepare_write(file, page, offset, offset+len);
1912                 if (ret) {
1913                         unlock_page(page);
1914                         page_cache_release(page);
1915                         if (pos + len > inode->i_size)
1916                                 vmtruncate(inode, inode->i_size);
1917                 }
1918                 return ret;
1919         }
1920 }
1921 EXPORT_SYMBOL(pagecache_write_begin);
1922
1923 int pagecache_write_end(struct file *file, struct address_space *mapping,
1924                                 loff_t pos, unsigned len, unsigned copied,
1925                                 struct page *page, void *fsdata)
1926 {
1927         const struct address_space_operations *aops = mapping->a_ops;
1928         int ret;
1929
1930         if (aops->write_end) {
1931                 mark_page_accessed(page);
1932                 ret = aops->write_end(file, mapping, pos, len, copied,
1933                                                         page, fsdata);
1934         } else {
1935                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1936                 struct inode *inode = mapping->host;
1937
1938                 flush_dcache_page(page);
1939                 ret = aops->commit_write(file, page, offset, offset+len);
1940                 unlock_page(page);
1941                 mark_page_accessed(page);
1942                 page_cache_release(page);
1943
1944                 if (ret < 0) {
1945                         if (pos + len > inode->i_size)
1946                                 vmtruncate(inode, inode->i_size);
1947                 } else if (ret > 0)
1948                         ret = min_t(size_t, copied, ret);
1949                 else
1950                         ret = copied;
1951         }
1952
1953         return ret;
1954 }
1955 EXPORT_SYMBOL(pagecache_write_end);
1956
1957 ssize_t
1958 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1959                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1960                 size_t count, size_t ocount)
1961 {
1962         struct file     *file = iocb->ki_filp;
1963         struct address_space *mapping = file->f_mapping;
1964         struct inode    *inode = mapping->host;
1965         ssize_t         written;
1966
1967         if (count != ocount)
1968                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1969
1970         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1971         if (written > 0) {
1972                 loff_t end = pos + written;
1973                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1974                         i_size_write(inode,  end);
1975                         mark_inode_dirty(inode);
1976                 }
1977                 *ppos = end;
1978         }
1979
1980         /*
1981          * Sync the fs metadata but not the minor inode changes and
1982          * of course not the data as we did direct DMA for the IO.
1983          * i_mutex is held, which protects generic_osync_inode() from
1984          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1985          */
1986         if ((written >= 0 || written == -EIOCBQUEUED) &&
1987             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1988                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1989                 if (err < 0)
1990                         written = err;
1991         }
1992         return written;
1993 }
1994 EXPORT_SYMBOL(generic_file_direct_write);
1995
1996 /*
1997  * Find or create a page at the given pagecache position. Return the locked
1998  * page. This function is specifically for buffered writes.
1999  */
2000 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2001 {
2002         int status;
2003         struct page *page;
2004 repeat:
2005         page = find_lock_page(mapping, index);
2006         if (likely(page))
2007                 return page;
2008
2009         page = page_cache_alloc(mapping);
2010         if (!page)
2011                 return NULL;
2012         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2013         if (unlikely(status)) {
2014                 page_cache_release(page);
2015                 if (status == -EEXIST)
2016                         goto repeat;
2017                 return NULL;
2018         }
2019         return page;
2020 }
2021 EXPORT_SYMBOL(__grab_cache_page);
2022
2023 static ssize_t generic_perform_write_2copy(struct file *file,
2024                                 struct iov_iter *i, loff_t pos)
2025 {
2026         struct address_space *mapping = file->f_mapping;
2027         const struct address_space_operations *a_ops = mapping->a_ops;
2028         struct inode *inode = mapping->host;
2029         long status = 0;
2030         ssize_t written = 0;
2031
2032         do {
2033                 struct page *src_page;
2034                 struct page *page;
2035                 pgoff_t index;          /* Pagecache index for current page */
2036                 unsigned long offset;   /* Offset into pagecache page */
2037                 unsigned long bytes;    /* Bytes to write to page */
2038                 size_t copied;          /* Bytes copied from user */
2039
2040                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2041                 index = pos >> PAGE_CACHE_SHIFT;
2042                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2043                                                 iov_iter_count(i));
2044
2045                 /*
2046                  * a non-NULL src_page indicates that we're doing the
2047                  * copy via get_user_pages and kmap.
2048                  */
2049                 src_page = NULL;
2050
2051                 /*
2052                  * Bring in the user page that we will copy from _first_.
2053                  * Otherwise there's a nasty deadlock on copying from the
2054                  * same page as we're writing to, without it being marked
2055                  * up-to-date.
2056                  *
2057                  * Not only is this an optimisation, but it is also required
2058                  * to check that the address is actually valid, when atomic
2059                  * usercopies are used, below.
2060                  */
2061                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2062                         status = -EFAULT;
2063                         break;
2064                 }
2065
2066                 page = __grab_cache_page(mapping, index);
2067                 if (!page) {
2068                         status = -ENOMEM;
2069                         break;
2070                 }
2071
2072                 /*
2073                  * non-uptodate pages cannot cope with short copies, and we
2074                  * cannot take a pagefault with the destination page locked.
2075                  * So pin the source page to copy it.
2076                  */
2077                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2078                         unlock_page(page);
2079
2080                         src_page = alloc_page(GFP_KERNEL);
2081                         if (!src_page) {
2082                                 page_cache_release(page);
2083                                 status = -ENOMEM;
2084                                 break;
2085                         }
2086
2087                         /*
2088                          * Cannot get_user_pages with a page locked for the
2089                          * same reason as we can't take a page fault with a
2090                          * page locked (as explained below).
2091                          */
2092                         copied = iov_iter_copy_from_user(src_page, i,
2093                                                                 offset, bytes);
2094                         if (unlikely(copied == 0)) {
2095                                 status = -EFAULT;
2096                                 page_cache_release(page);
2097                                 page_cache_release(src_page);
2098                                 break;
2099                         }
2100                         bytes = copied;
2101
2102                         lock_page(page);
2103                         /*
2104                          * Can't handle the page going uptodate here, because
2105                          * that means we would use non-atomic usercopies, which
2106                          * zero out the tail of the page, which can cause
2107                          * zeroes to become transiently visible. We could just
2108                          * use a non-zeroing copy, but the APIs aren't too
2109                          * consistent.
2110                          */
2111                         if (unlikely(!page->mapping || PageUptodate(page))) {
2112                                 unlock_page(page);
2113                                 page_cache_release(page);
2114                                 page_cache_release(src_page);
2115                                 continue;
2116                         }
2117                 }
2118
2119                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2120                 if (unlikely(status))
2121                         goto fs_write_aop_error;
2122
2123                 if (!src_page) {
2124                         /*
2125                          * Must not enter the pagefault handler here, because
2126                          * we hold the page lock, so we might recursively
2127                          * deadlock on the same lock, or get an ABBA deadlock
2128                          * against a different lock, or against the mmap_sem
2129                          * (which nests outside the page lock).  So increment
2130                          * preempt count, and use _atomic usercopies.
2131                          *
2132                          * The page is uptodate so we are OK to encounter a
2133                          * short copy: if unmodified parts of the page are
2134                          * marked dirty and written out to disk, it doesn't
2135                          * really matter.
2136                          */
2137                         pagefault_disable();
2138                         copied = iov_iter_copy_from_user_atomic(page, i,
2139                                                                 offset, bytes);
2140                         pagefault_enable();
2141                 } else {
2142                         void *src, *dst;
2143                         src = kmap_atomic(src_page, KM_USER0);
2144                         dst = kmap_atomic(page, KM_USER1);
2145                         memcpy(dst + offset, src + offset, bytes);
2146                         kunmap_atomic(dst, KM_USER1);
2147                         kunmap_atomic(src, KM_USER0);
2148                         copied = bytes;
2149                 }
2150                 flush_dcache_page(page);
2151
2152                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2153                 if (unlikely(status < 0))
2154                         goto fs_write_aop_error;
2155                 if (unlikely(status > 0)) /* filesystem did partial write */
2156                         copied = min_t(size_t, copied, status);
2157
2158                 unlock_page(page);
2159                 mark_page_accessed(page);
2160                 page_cache_release(page);
2161                 if (src_page)
2162                         page_cache_release(src_page);
2163
2164                 iov_iter_advance(i, copied);
2165                 pos += copied;
2166                 written += copied;
2167
2168                 balance_dirty_pages_ratelimited(mapping);
2169                 cond_resched();
2170                 continue;
2171
2172 fs_write_aop_error:
2173                 unlock_page(page);
2174                 page_cache_release(page);
2175                 if (src_page)
2176                         page_cache_release(src_page);
2177
2178                 /*
2179                  * prepare_write() may have instantiated a few blocks
2180                  * outside i_size.  Trim these off again. Don't need
2181                  * i_size_read because we hold i_mutex.
2182                  */
2183                 if (pos + bytes > inode->i_size)
2184                         vmtruncate(inode, inode->i_size);
2185                 break;
2186         } while (iov_iter_count(i));
2187
2188         return written ? written : status;
2189 }
2190
2191 static ssize_t generic_perform_write(struct file *file,
2192                                 struct iov_iter *i, loff_t pos)
2193 {
2194         struct address_space *mapping = file->f_mapping;
2195         const struct address_space_operations *a_ops = mapping->a_ops;
2196         long status = 0;
2197         ssize_t written = 0;
2198         unsigned int flags = 0;
2199
2200         /*
2201          * Copies from kernel address space cannot fail (NFSD is a big user).
2202          */
2203         if (segment_eq(get_fs(), KERNEL_DS))
2204                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2205
2206         do {
2207                 struct page *page;
2208                 pgoff_t index;          /* Pagecache index for current page */
2209                 unsigned long offset;   /* Offset into pagecache page */
2210                 unsigned long bytes;    /* Bytes to write to page */
2211                 size_t copied;          /* Bytes copied from user */
2212                 void *fsdata;
2213
2214                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2215                 index = pos >> PAGE_CACHE_SHIFT;
2216                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2217                                                 iov_iter_count(i));
2218
2219 again:
2220
2221                 /*
2222                  * Bring in the user page that we will copy from _first_.
2223                  * Otherwise there's a nasty deadlock on copying from the
2224                  * same page as we're writing to, without it being marked
2225                  * up-to-date.
2226                  *
2227                  * Not only is this an optimisation, but it is also required
2228                  * to check that the address is actually valid, when atomic
2229                  * usercopies are used, below.
2230                  */
2231                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2232                         status = -EFAULT;
2233                         break;
2234                 }
2235
2236                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2237                                                 &page, &fsdata);
2238                 if (unlikely(status))
2239                         break;
2240
2241                 pagefault_disable();
2242                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2243                 pagefault_enable();
2244                 flush_dcache_page(page);
2245
2246                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2247                                                 page, fsdata);
2248                 if (unlikely(status < 0))
2249                         break;
2250                 copied = status;
2251
2252                 cond_resched();
2253
2254                 if (unlikely(copied == 0)) {
2255                         /*
2256                          * If we were unable to copy any data at all, we must
2257                          * fall back to a single segment length write.
2258                          *
2259                          * If we didn't fallback here, we could livelock
2260                          * because not all segments in the iov can be copied at
2261                          * once without a pagefault.
2262                          */
2263                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2264                                                 iov_iter_single_seg_count(i));
2265                         goto again;
2266                 }
2267                 iov_iter_advance(i, copied);
2268                 pos += copied;
2269                 written += copied;
2270
2271                 balance_dirty_pages_ratelimited(mapping);
2272
2273         } while (iov_iter_count(i));
2274
2275         return written ? written : status;
2276 }
2277
2278 ssize_t
2279 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2280                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2281                 size_t count, ssize_t written)
2282 {
2283         struct file *file = iocb->ki_filp;
2284         struct address_space *mapping = file->f_mapping;
2285         const struct address_space_operations *a_ops = mapping->a_ops;
2286         struct inode *inode = mapping->host;
2287         ssize_t status;
2288         struct iov_iter i;
2289
2290         iov_iter_init(&i, iov, nr_segs, count, written);
2291         if (a_ops->write_begin)
2292                 status = generic_perform_write(file, &i, pos);
2293         else
2294                 status = generic_perform_write_2copy(file, &i, pos);
2295
2296         if (likely(status >= 0)) {
2297                 written += status;
2298                 *ppos = pos + status;
2299
2300                 /*
2301                  * For now, when the user asks for O_SYNC, we'll actually give
2302                  * O_DSYNC
2303                  */
2304                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2305                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2306                                 status = generic_osync_inode(inode, mapping,
2307                                                 OSYNC_METADATA|OSYNC_DATA);
2308                 }
2309         }
2310         
2311         /*
2312          * If we get here for O_DIRECT writes then we must have fallen through
2313          * to buffered writes (block instantiation inside i_size).  So we sync
2314          * the file data here, to try to honour O_DIRECT expectations.
2315          */
2316         if (unlikely(file->f_flags & O_DIRECT) && written)
2317                 status = filemap_write_and_wait(mapping);
2318
2319         return written ? written : status;
2320 }
2321 EXPORT_SYMBOL(generic_file_buffered_write);
2322
2323 static ssize_t
2324 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2325                                 unsigned long nr_segs, loff_t *ppos)
2326 {
2327         struct file *file = iocb->ki_filp;
2328         struct address_space * mapping = file->f_mapping;
2329         size_t ocount;          /* original count */
2330         size_t count;           /* after file limit checks */
2331         struct inode    *inode = mapping->host;
2332         loff_t          pos;
2333         ssize_t         written;
2334         ssize_t         err;
2335
2336         ocount = 0;
2337         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2338         if (err)
2339                 return err;
2340
2341         count = ocount;
2342         pos = *ppos;
2343
2344         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2345
2346         /* We can write back this queue in page reclaim */
2347         current->backing_dev_info = mapping->backing_dev_info;
2348         written = 0;
2349
2350         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2351         if (err)
2352                 goto out;
2353
2354         if (count == 0)
2355                 goto out;
2356
2357         err = remove_suid(file->f_path.dentry);
2358         if (err)
2359                 goto out;
2360
2361         file_update_time(file);
2362
2363         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2364         if (unlikely(file->f_flags & O_DIRECT)) {
2365                 loff_t endbyte;
2366                 ssize_t written_buffered;
2367
2368                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2369                                                         ppos, count, ocount);
2370                 if (written < 0 || written == count)
2371                         goto out;
2372                 /*
2373                  * direct-io write to a hole: fall through to buffered I/O
2374                  * for completing the rest of the request.
2375                  */
2376                 pos += written;
2377                 count -= written;
2378                 written_buffered = generic_file_buffered_write(iocb, iov,
2379                                                 nr_segs, pos, ppos, count,
2380                                                 written);
2381                 /*
2382                  * If generic_file_buffered_write() retuned a synchronous error
2383                  * then we want to return the number of bytes which were
2384                  * direct-written, or the error code if that was zero.  Note
2385                  * that this differs from normal direct-io semantics, which
2386                  * will return -EFOO even if some bytes were written.
2387                  */
2388                 if (written_buffered < 0) {
2389                         err = written_buffered;
2390                         goto out;
2391                 }
2392
2393                 /*
2394                  * We need to ensure that the page cache pages are written to
2395                  * disk and invalidated to preserve the expected O_DIRECT
2396                  * semantics.
2397                  */
2398                 endbyte = pos + written_buffered - written - 1;
2399                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2400                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2401                                             SYNC_FILE_RANGE_WRITE|
2402                                             SYNC_FILE_RANGE_WAIT_AFTER);
2403                 if (err == 0) {
2404                         written = written_buffered;
2405                         invalidate_mapping_pages(mapping,
2406                                                  pos >> PAGE_CACHE_SHIFT,
2407                                                  endbyte >> PAGE_CACHE_SHIFT);
2408                 } else {
2409                         /*
2410                          * We don't know how much we wrote, so just return
2411                          * the number of bytes which were direct-written
2412                          */
2413                 }
2414         } else {
2415                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2416                                 pos, ppos, count, written);
2417         }
2418 out:
2419         current->backing_dev_info = NULL;
2420         return written ? written : err;
2421 }
2422
2423 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2424                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2425 {
2426         struct file *file = iocb->ki_filp;
2427         struct address_space *mapping = file->f_mapping;
2428         struct inode *inode = mapping->host;
2429         ssize_t ret;
2430
2431         BUG_ON(iocb->ki_pos != pos);
2432
2433         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2434                         &iocb->ki_pos);
2435
2436         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2437                 ssize_t err;
2438
2439                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2440                 if (err < 0)
2441                         ret = err;
2442         }
2443         return ret;
2444 }
2445 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2446
2447 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2448                 unsigned long nr_segs, loff_t pos)
2449 {
2450         struct file *file = iocb->ki_filp;
2451         struct address_space *mapping = file->f_mapping;
2452         struct inode *inode = mapping->host;
2453         ssize_t ret;
2454
2455         BUG_ON(iocb->ki_pos != pos);
2456
2457         mutex_lock(&inode->i_mutex);
2458         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2459                         &iocb->ki_pos);
2460         mutex_unlock(&inode->i_mutex);
2461
2462         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2463                 ssize_t err;
2464
2465                 err = sync_page_range(inode, mapping, pos, ret);
2466                 if (err < 0)
2467                         ret = err;
2468         }
2469         return ret;
2470 }
2471 EXPORT_SYMBOL(generic_file_aio_write);
2472
2473 /*
2474  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2475  * went wrong during pagecache shootdown.
2476  */
2477 static ssize_t
2478 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2479         loff_t offset, unsigned long nr_segs)
2480 {
2481         struct file *file = iocb->ki_filp;
2482         struct address_space *mapping = file->f_mapping;
2483         ssize_t retval;
2484         size_t write_len;
2485         pgoff_t end = 0; /* silence gcc */
2486
2487         /*
2488          * If it's a write, unmap all mmappings of the file up-front.  This
2489          * will cause any pte dirty bits to be propagated into the pageframes
2490          * for the subsequent filemap_write_and_wait().
2491          */
2492         if (rw == WRITE) {
2493                 write_len = iov_length(iov, nr_segs);
2494                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2495                 if (mapping_mapped(mapping))
2496                         unmap_mapping_range(mapping, offset, write_len, 0);
2497         }
2498
2499         retval = filemap_write_and_wait(mapping);
2500         if (retval)
2501                 goto out;
2502
2503         /*
2504          * After a write we want buffered reads to be sure to go to disk to get
2505          * the new data.  We invalidate clean cached page from the region we're
2506          * about to write.  We do this *before* the write so that we can return
2507          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2508          */
2509         if (rw == WRITE && mapping->nrpages) {
2510                 retval = invalidate_inode_pages2_range(mapping,
2511                                         offset >> PAGE_CACHE_SHIFT, end);
2512                 if (retval)
2513                         goto out;
2514         }
2515
2516         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2517
2518         /*
2519          * Finally, try again to invalidate clean pages which might have been
2520          * cached by non-direct readahead, or faulted in by get_user_pages()
2521          * if the source of the write was an mmap'ed region of the file
2522          * we're writing.  Either one is a pretty crazy thing to do,
2523          * so we don't support it 100%.  If this invalidation
2524          * fails, tough, the write still worked...
2525          */
2526         if (rw == WRITE && mapping->nrpages) {
2527                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2528         }
2529 out:
2530         return retval;
2531 }
2532
2533 /**
2534  * try_to_release_page() - release old fs-specific metadata on a page
2535  *
2536  * @page: the page which the kernel is trying to free
2537  * @gfp_mask: memory allocation flags (and I/O mode)
2538  *
2539  * The address_space is to try to release any data against the page
2540  * (presumably at page->private).  If the release was successful, return `1'.
2541  * Otherwise return zero.
2542  *
2543  * The @gfp_mask argument specifies whether I/O may be performed to release
2544  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2545  *
2546  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2547  */
2548 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2549 {
2550         struct address_space * const mapping = page->mapping;
2551
2552         BUG_ON(!PageLocked(page));
2553         if (PageWriteback(page))
2554                 return 0;
2555
2556         if (mapping && mapping->a_ops->releasepage)
2557                 return mapping->a_ops->releasepage(page, gfp_mask);
2558         return try_to_free_buffers(page);
2559 }
2560
2561 EXPORT_SYMBOL(try_to_release_page);