mm: don't return 0 too early from find_get_pages()
[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/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  (code doesn't rely on that order, so you could switch it around)
106  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
107  *    ->i_mmap_lock
108  */
109
110 /*
111  * Delete a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold the mapping's tree_lock.
114  */
115 void __delete_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         if (PageSwapBacked(page))
124                 __dec_zone_page_state(page, NR_SHMEM);
125         BUG_ON(page_mapped(page));
126
127         /*
128          * Some filesystems seem to re-dirty the page even after
129          * the VM has canceled the dirty bit (eg ext3 journaling).
130          *
131          * Fix it up by doing a final dirty accounting check after
132          * having removed the page entirely.
133          */
134         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
135                 dec_zone_page_state(page, NR_FILE_DIRTY);
136                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
137         }
138 }
139
140 /**
141  * delete_from_page_cache - delete page from page cache
142  * @page: the page which the kernel is trying to remove from page cache
143  *
144  * This must be called only on pages that have been verified to be in the page
145  * cache and locked.  It will never put the page into the free list, the caller
146  * has a reference on the page.
147  */
148 void delete_from_page_cache(struct page *page)
149 {
150         struct address_space *mapping = page->mapping;
151         void (*freepage)(struct page *);
152
153         BUG_ON(!PageLocked(page));
154
155         freepage = mapping->a_ops->freepage;
156         spin_lock_irq(&mapping->tree_lock);
157         __delete_from_page_cache(page);
158         spin_unlock_irq(&mapping->tree_lock);
159         mem_cgroup_uncharge_cache_page(page);
160
161         if (freepage)
162                 freepage(page);
163         page_cache_release(page);
164 }
165 EXPORT_SYMBOL(delete_from_page_cache);
166
167 static int sync_page(void *word)
168 {
169         struct address_space *mapping;
170         struct page *page;
171
172         page = container_of((unsigned long *)word, struct page, flags);
173
174         /*
175          * page_mapping() is being called without PG_locked held.
176          * Some knowledge of the state and use of the page is used to
177          * reduce the requirements down to a memory barrier.
178          * The danger here is of a stale page_mapping() return value
179          * indicating a struct address_space different from the one it's
180          * associated with when it is associated with one.
181          * After smp_mb(), it's either the correct page_mapping() for
182          * the page, or an old page_mapping() and the page's own
183          * page_mapping() has gone NULL.
184          * The ->sync_page() address_space operation must tolerate
185          * page_mapping() going NULL. By an amazing coincidence,
186          * this comes about because none of the users of the page
187          * in the ->sync_page() methods make essential use of the
188          * page_mapping(), merely passing the page down to the backing
189          * device's unplug functions when it's non-NULL, which in turn
190          * ignore it for all cases but swap, where only page_private(page) is
191          * of interest. When page_mapping() does go NULL, the entire
192          * call stack gracefully ignores the page and returns.
193          * -- wli
194          */
195         smp_mb();
196         mapping = page_mapping(page);
197         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
198                 mapping->a_ops->sync_page(page);
199         io_schedule();
200         return 0;
201 }
202
203 static int sync_page_killable(void *word)
204 {
205         sync_page(word);
206         return fatal_signal_pending(current) ? -EINTR : 0;
207 }
208
209 /**
210  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
211  * @mapping:    address space structure to write
212  * @start:      offset in bytes where the range starts
213  * @end:        offset in bytes where the range ends (inclusive)
214  * @sync_mode:  enable synchronous operation
215  *
216  * Start writeback against all of a mapping's dirty pages that lie
217  * within the byte offsets <start, end> inclusive.
218  *
219  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
220  * opposed to a regular memory cleansing writeback.  The difference between
221  * these two operations is that if a dirty page/buffer is encountered, it must
222  * be waited upon, and not just skipped over.
223  */
224 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
225                                 loff_t end, int sync_mode)
226 {
227         int ret;
228         struct writeback_control wbc = {
229                 .sync_mode = sync_mode,
230                 .nr_to_write = LONG_MAX,
231                 .range_start = start,
232                 .range_end = end,
233         };
234
235         if (!mapping_cap_writeback_dirty(mapping))
236                 return 0;
237
238         ret = do_writepages(mapping, &wbc);
239         return ret;
240 }
241
242 static inline int __filemap_fdatawrite(struct address_space *mapping,
243         int sync_mode)
244 {
245         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
246 }
247
248 int filemap_fdatawrite(struct address_space *mapping)
249 {
250         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
251 }
252 EXPORT_SYMBOL(filemap_fdatawrite);
253
254 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
255                                 loff_t end)
256 {
257         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
258 }
259 EXPORT_SYMBOL(filemap_fdatawrite_range);
260
261 /**
262  * filemap_flush - mostly a non-blocking flush
263  * @mapping:    target address_space
264  *
265  * This is a mostly non-blocking flush.  Not suitable for data-integrity
266  * purposes - I/O may not be started against all dirty pages.
267  */
268 int filemap_flush(struct address_space *mapping)
269 {
270         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
271 }
272 EXPORT_SYMBOL(filemap_flush);
273
274 /**
275  * filemap_fdatawait_range - wait for writeback to complete
276  * @mapping:            address space structure to wait for
277  * @start_byte:         offset in bytes where the range starts
278  * @end_byte:           offset in bytes where the range ends (inclusive)
279  *
280  * Walk the list of under-writeback pages of the given address space
281  * in the given range and wait for all of them.
282  */
283 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
284                             loff_t end_byte)
285 {
286         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
287         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
288         struct pagevec pvec;
289         int nr_pages;
290         int ret = 0;
291
292         if (end_byte < start_byte)
293                 return 0;
294
295         pagevec_init(&pvec, 0);
296         while ((index <= end) &&
297                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
298                         PAGECACHE_TAG_WRITEBACK,
299                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
300                 unsigned i;
301
302                 for (i = 0; i < nr_pages; i++) {
303                         struct page *page = pvec.pages[i];
304
305                         /* until radix tree lookup accepts end_index */
306                         if (page->index > end)
307                                 continue;
308
309                         wait_on_page_writeback(page);
310                         if (TestClearPageError(page))
311                                 ret = -EIO;
312                 }
313                 pagevec_release(&pvec);
314                 cond_resched();
315         }
316
317         /* Check for outstanding write errors */
318         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
319                 ret = -ENOSPC;
320         if (test_and_clear_bit(AS_EIO, &mapping->flags))
321                 ret = -EIO;
322
323         return ret;
324 }
325 EXPORT_SYMBOL(filemap_fdatawait_range);
326
327 /**
328  * filemap_fdatawait - wait for all under-writeback pages to complete
329  * @mapping: address space structure to wait for
330  *
331  * Walk the list of under-writeback pages of the given address space
332  * and wait for all of them.
333  */
334 int filemap_fdatawait(struct address_space *mapping)
335 {
336         loff_t i_size = i_size_read(mapping->host);
337
338         if (i_size == 0)
339                 return 0;
340
341         return filemap_fdatawait_range(mapping, 0, i_size - 1);
342 }
343 EXPORT_SYMBOL(filemap_fdatawait);
344
345 int filemap_write_and_wait(struct address_space *mapping)
346 {
347         int err = 0;
348
349         if (mapping->nrpages) {
350                 err = filemap_fdatawrite(mapping);
351                 /*
352                  * Even if the above returned error, the pages may be
353                  * written partially (e.g. -ENOSPC), so we wait for it.
354                  * But the -EIO is special case, it may indicate the worst
355                  * thing (e.g. bug) happened, so we avoid waiting for it.
356                  */
357                 if (err != -EIO) {
358                         int err2 = filemap_fdatawait(mapping);
359                         if (!err)
360                                 err = err2;
361                 }
362         }
363         return err;
364 }
365 EXPORT_SYMBOL(filemap_write_and_wait);
366
367 /**
368  * filemap_write_and_wait_range - write out & wait on a file range
369  * @mapping:    the address_space for the pages
370  * @lstart:     offset in bytes where the range starts
371  * @lend:       offset in bytes where the range ends (inclusive)
372  *
373  * Write out and wait upon file offsets lstart->lend, inclusive.
374  *
375  * Note that `lend' is inclusive (describes the last byte to be written) so
376  * that this function can be used to write to the very end-of-file (end = -1).
377  */
378 int filemap_write_and_wait_range(struct address_space *mapping,
379                                  loff_t lstart, loff_t lend)
380 {
381         int err = 0;
382
383         if (mapping->nrpages) {
384                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
385                                                  WB_SYNC_ALL);
386                 /* See comment of filemap_write_and_wait() */
387                 if (err != -EIO) {
388                         int err2 = filemap_fdatawait_range(mapping,
389                                                 lstart, lend);
390                         if (!err)
391                                 err = err2;
392                 }
393         }
394         return err;
395 }
396 EXPORT_SYMBOL(filemap_write_and_wait_range);
397
398 /**
399  * replace_page_cache_page - replace a pagecache page with a new one
400  * @old:        page to be replaced
401  * @new:        page to replace with
402  * @gfp_mask:   allocation mode
403  *
404  * This function replaces a page in the pagecache with a new one.  On
405  * success it acquires the pagecache reference for the new page and
406  * drops it for the old page.  Both the old and new pages must be
407  * locked.  This function does not add the new page to the LRU, the
408  * caller must do that.
409  *
410  * The remove + add is atomic.  The only way this function can fail is
411  * memory allocation failure.
412  */
413 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
414 {
415         int error;
416         struct mem_cgroup *memcg = NULL;
417
418         VM_BUG_ON(!PageLocked(old));
419         VM_BUG_ON(!PageLocked(new));
420         VM_BUG_ON(new->mapping);
421
422         /*
423          * This is not page migration, but prepare_migration and
424          * end_migration does enough work for charge replacement.
425          *
426          * In the longer term we probably want a specialized function
427          * for moving the charge from old to new in a more efficient
428          * manner.
429          */
430         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
431         if (error)
432                 return error;
433
434         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
435         if (!error) {
436                 struct address_space *mapping = old->mapping;
437                 void (*freepage)(struct page *);
438
439                 pgoff_t offset = old->index;
440                 freepage = mapping->a_ops->freepage;
441
442                 page_cache_get(new);
443                 new->mapping = mapping;
444                 new->index = offset;
445
446                 spin_lock_irq(&mapping->tree_lock);
447                 __delete_from_page_cache(old);
448                 error = radix_tree_insert(&mapping->page_tree, offset, new);
449                 BUG_ON(error);
450                 mapping->nrpages++;
451                 __inc_zone_page_state(new, NR_FILE_PAGES);
452                 if (PageSwapBacked(new))
453                         __inc_zone_page_state(new, NR_SHMEM);
454                 spin_unlock_irq(&mapping->tree_lock);
455                 radix_tree_preload_end();
456                 if (freepage)
457                         freepage(old);
458                 page_cache_release(old);
459                 mem_cgroup_end_migration(memcg, old, new, true);
460         } else {
461                 mem_cgroup_end_migration(memcg, old, new, false);
462         }
463
464         return error;
465 }
466 EXPORT_SYMBOL_GPL(replace_page_cache_page);
467
468 /**
469  * add_to_page_cache_locked - add a locked page to the pagecache
470  * @page:       page to add
471  * @mapping:    the page's address_space
472  * @offset:     page index
473  * @gfp_mask:   page allocation mode
474  *
475  * This function is used to add a page to the pagecache. It must be locked.
476  * This function does not add the page to the LRU.  The caller must do that.
477  */
478 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
479                 pgoff_t offset, gfp_t gfp_mask)
480 {
481         int error;
482
483         VM_BUG_ON(!PageLocked(page));
484
485         error = mem_cgroup_cache_charge(page, current->mm,
486                                         gfp_mask & GFP_RECLAIM_MASK);
487         if (error)
488                 goto out;
489
490         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
491         if (error == 0) {
492                 page_cache_get(page);
493                 page->mapping = mapping;
494                 page->index = offset;
495
496                 spin_lock_irq(&mapping->tree_lock);
497                 error = radix_tree_insert(&mapping->page_tree, offset, page);
498                 if (likely(!error)) {
499                         mapping->nrpages++;
500                         __inc_zone_page_state(page, NR_FILE_PAGES);
501                         if (PageSwapBacked(page))
502                                 __inc_zone_page_state(page, NR_SHMEM);
503                         spin_unlock_irq(&mapping->tree_lock);
504                 } else {
505                         page->mapping = NULL;
506                         spin_unlock_irq(&mapping->tree_lock);
507                         mem_cgroup_uncharge_cache_page(page);
508                         page_cache_release(page);
509                 }
510                 radix_tree_preload_end();
511         } else
512                 mem_cgroup_uncharge_cache_page(page);
513 out:
514         return error;
515 }
516 EXPORT_SYMBOL(add_to_page_cache_locked);
517
518 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
519                                 pgoff_t offset, gfp_t gfp_mask)
520 {
521         int ret;
522
523         /*
524          * Splice_read and readahead add shmem/tmpfs pages into the page cache
525          * before shmem_readpage has a chance to mark them as SwapBacked: they
526          * need to go on the anon lru below, and mem_cgroup_cache_charge
527          * (called in add_to_page_cache) needs to know where they're going too.
528          */
529         if (mapping_cap_swap_backed(mapping))
530                 SetPageSwapBacked(page);
531
532         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
533         if (ret == 0) {
534                 if (page_is_file_cache(page))
535                         lru_cache_add_file(page);
536                 else
537                         lru_cache_add_anon(page);
538         }
539         return ret;
540 }
541 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
542
543 #ifdef CONFIG_NUMA
544 struct page *__page_cache_alloc(gfp_t gfp)
545 {
546         int n;
547         struct page *page;
548
549         if (cpuset_do_page_mem_spread()) {
550                 get_mems_allowed();
551                 n = cpuset_mem_spread_node();
552                 page = alloc_pages_exact_node(n, gfp, 0);
553                 put_mems_allowed();
554                 return page;
555         }
556         return alloc_pages(gfp, 0);
557 }
558 EXPORT_SYMBOL(__page_cache_alloc);
559 #endif
560
561 static int __sleep_on_page_lock(void *word)
562 {
563         io_schedule();
564         return 0;
565 }
566
567 /*
568  * In order to wait for pages to become available there must be
569  * waitqueues associated with pages. By using a hash table of
570  * waitqueues where the bucket discipline is to maintain all
571  * waiters on the same queue and wake all when any of the pages
572  * become available, and for the woken contexts to check to be
573  * sure the appropriate page became available, this saves space
574  * at a cost of "thundering herd" phenomena during rare hash
575  * collisions.
576  */
577 static wait_queue_head_t *page_waitqueue(struct page *page)
578 {
579         const struct zone *zone = page_zone(page);
580
581         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
582 }
583
584 static inline void wake_up_page(struct page *page, int bit)
585 {
586         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
587 }
588
589 void wait_on_page_bit(struct page *page, int bit_nr)
590 {
591         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
592
593         if (test_bit(bit_nr, &page->flags))
594                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
595                                                         TASK_UNINTERRUPTIBLE);
596 }
597 EXPORT_SYMBOL(wait_on_page_bit);
598
599 /**
600  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
601  * @page: Page defining the wait queue of interest
602  * @waiter: Waiter to add to the queue
603  *
604  * Add an arbitrary @waiter to the wait queue for the nominated @page.
605  */
606 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
607 {
608         wait_queue_head_t *q = page_waitqueue(page);
609         unsigned long flags;
610
611         spin_lock_irqsave(&q->lock, flags);
612         __add_wait_queue(q, waiter);
613         spin_unlock_irqrestore(&q->lock, flags);
614 }
615 EXPORT_SYMBOL_GPL(add_page_wait_queue);
616
617 /**
618  * unlock_page - unlock a locked page
619  * @page: the page
620  *
621  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
622  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
623  * mechananism between PageLocked pages and PageWriteback pages is shared.
624  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
625  *
626  * The mb is necessary to enforce ordering between the clear_bit and the read
627  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
628  */
629 void unlock_page(struct page *page)
630 {
631         VM_BUG_ON(!PageLocked(page));
632         clear_bit_unlock(PG_locked, &page->flags);
633         smp_mb__after_clear_bit();
634         wake_up_page(page, PG_locked);
635 }
636 EXPORT_SYMBOL(unlock_page);
637
638 /**
639  * end_page_writeback - end writeback against a page
640  * @page: the page
641  */
642 void end_page_writeback(struct page *page)
643 {
644         if (TestClearPageReclaim(page))
645                 rotate_reclaimable_page(page);
646
647         if (!test_clear_page_writeback(page))
648                 BUG();
649
650         smp_mb__after_clear_bit();
651         wake_up_page(page, PG_writeback);
652 }
653 EXPORT_SYMBOL(end_page_writeback);
654
655 /**
656  * __lock_page - get a lock on the page, assuming we need to sleep to get it
657  * @page: the page to lock
658  *
659  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
660  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
661  * chances are that on the second loop, the block layer's plug list is empty,
662  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
663  */
664 void __lock_page(struct page *page)
665 {
666         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
667
668         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
669                                                         TASK_UNINTERRUPTIBLE);
670 }
671 EXPORT_SYMBOL(__lock_page);
672
673 int __lock_page_killable(struct page *page)
674 {
675         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
676
677         return __wait_on_bit_lock(page_waitqueue(page), &wait,
678                                         sync_page_killable, TASK_KILLABLE);
679 }
680 EXPORT_SYMBOL_GPL(__lock_page_killable);
681
682 /**
683  * __lock_page_nosync - get a lock on the page, without calling sync_page()
684  * @page: the page to lock
685  *
686  * Variant of lock_page that does not require the caller to hold a reference
687  * on the page's mapping.
688  */
689 void __lock_page_nosync(struct page *page)
690 {
691         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
692         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
693                                                         TASK_UNINTERRUPTIBLE);
694 }
695
696 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
697                          unsigned int flags)
698 {
699         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
700                 __lock_page(page);
701                 return 1;
702         } else {
703                 if (!(flags & FAULT_FLAG_RETRY_NOWAIT)) {
704                         up_read(&mm->mmap_sem);
705                         wait_on_page_locked(page);
706                 }
707                 return 0;
708         }
709 }
710
711 /**
712  * find_get_page - find and get a page reference
713  * @mapping: the address_space to search
714  * @offset: the page index
715  *
716  * Is there a pagecache struct page at the given (mapping, offset) tuple?
717  * If yes, increment its refcount and return it; if no, return NULL.
718  */
719 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
720 {
721         void **pagep;
722         struct page *page;
723
724         rcu_read_lock();
725 repeat:
726         page = NULL;
727         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
728         if (pagep) {
729                 page = radix_tree_deref_slot(pagep);
730                 if (unlikely(!page))
731                         goto out;
732                 if (radix_tree_deref_retry(page))
733                         goto repeat;
734
735                 if (!page_cache_get_speculative(page))
736                         goto repeat;
737
738                 /*
739                  * Has the page moved?
740                  * This is part of the lockless pagecache protocol. See
741                  * include/linux/pagemap.h for details.
742                  */
743                 if (unlikely(page != *pagep)) {
744                         page_cache_release(page);
745                         goto repeat;
746                 }
747         }
748 out:
749         rcu_read_unlock();
750
751         return page;
752 }
753 EXPORT_SYMBOL(find_get_page);
754
755 /**
756  * find_lock_page - locate, pin and lock a pagecache page
757  * @mapping: the address_space to search
758  * @offset: the page index
759  *
760  * Locates the desired pagecache page, locks it, increments its reference
761  * count and returns its address.
762  *
763  * Returns zero if the page was not present. find_lock_page() may sleep.
764  */
765 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
766 {
767         struct page *page;
768
769 repeat:
770         page = find_get_page(mapping, offset);
771         if (page) {
772                 lock_page(page);
773                 /* Has the page been truncated? */
774                 if (unlikely(page->mapping != mapping)) {
775                         unlock_page(page);
776                         page_cache_release(page);
777                         goto repeat;
778                 }
779                 VM_BUG_ON(page->index != offset);
780         }
781         return page;
782 }
783 EXPORT_SYMBOL(find_lock_page);
784
785 /**
786  * find_or_create_page - locate or add a pagecache page
787  * @mapping: the page's address_space
788  * @index: the page's index into the mapping
789  * @gfp_mask: page allocation mode
790  *
791  * Locates a page in the pagecache.  If the page is not present, a new page
792  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
793  * LRU list.  The returned page is locked and has its reference count
794  * incremented.
795  *
796  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
797  * allocation!
798  *
799  * find_or_create_page() returns the desired page's address, or zero on
800  * memory exhaustion.
801  */
802 struct page *find_or_create_page(struct address_space *mapping,
803                 pgoff_t index, gfp_t gfp_mask)
804 {
805         struct page *page;
806         int err;
807 repeat:
808         page = find_lock_page(mapping, index);
809         if (!page) {
810                 page = __page_cache_alloc(gfp_mask);
811                 if (!page)
812                         return NULL;
813                 /*
814                  * We want a regular kernel memory (not highmem or DMA etc)
815                  * allocation for the radix tree nodes, but we need to honour
816                  * the context-specific requirements the caller has asked for.
817                  * GFP_RECLAIM_MASK collects those requirements.
818                  */
819                 err = add_to_page_cache_lru(page, mapping, index,
820                         (gfp_mask & GFP_RECLAIM_MASK));
821                 if (unlikely(err)) {
822                         page_cache_release(page);
823                         page = NULL;
824                         if (err == -EEXIST)
825                                 goto repeat;
826                 }
827         }
828         return page;
829 }
830 EXPORT_SYMBOL(find_or_create_page);
831
832 /**
833  * find_get_pages - gang pagecache lookup
834  * @mapping:    The address_space to search
835  * @start:      The starting page index
836  * @nr_pages:   The maximum number of pages
837  * @pages:      Where the resulting pages are placed
838  *
839  * find_get_pages() will search for and return a group of up to
840  * @nr_pages pages in the mapping.  The pages are placed at @pages.
841  * find_get_pages() takes a reference against the returned pages.
842  *
843  * The search returns a group of mapping-contiguous pages with ascending
844  * indexes.  There may be holes in the indices due to not-present pages.
845  *
846  * find_get_pages() returns the number of pages which were found.
847  */
848 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
849                             unsigned int nr_pages, struct page **pages)
850 {
851         unsigned int i;
852         unsigned int ret;
853         unsigned int nr_found;
854
855         rcu_read_lock();
856 restart:
857         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
858                                 (void ***)pages, start, nr_pages);
859         ret = 0;
860         for (i = 0; i < nr_found; i++) {
861                 struct page *page;
862 repeat:
863                 page = radix_tree_deref_slot((void **)pages[i]);
864                 if (unlikely(!page))
865                         continue;
866
867                 /*
868                  * This can only trigger when the entry at index 0 moves out
869                  * of or back to the root: none yet gotten, safe to restart.
870                  */
871                 if (radix_tree_deref_retry(page)) {
872                         WARN_ON(start | i);
873                         goto restart;
874                 }
875
876                 if (!page_cache_get_speculative(page))
877                         goto repeat;
878
879                 /* Has the page moved? */
880                 if (unlikely(page != *((void **)pages[i]))) {
881                         page_cache_release(page);
882                         goto repeat;
883                 }
884
885                 pages[ret] = page;
886                 ret++;
887         }
888
889         /*
890          * If all entries were removed before we could secure them,
891          * try again, because callers stop trying once 0 is returned.
892          */
893         if (unlikely(!ret && nr_found))
894                 goto restart;
895         rcu_read_unlock();
896         return ret;
897 }
898
899 /**
900  * find_get_pages_contig - gang contiguous pagecache lookup
901  * @mapping:    The address_space to search
902  * @index:      The starting page index
903  * @nr_pages:   The maximum number of pages
904  * @pages:      Where the resulting pages are placed
905  *
906  * find_get_pages_contig() works exactly like find_get_pages(), except
907  * that the returned number of pages are guaranteed to be contiguous.
908  *
909  * find_get_pages_contig() returns the number of pages which were found.
910  */
911 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
912                                unsigned int nr_pages, struct page **pages)
913 {
914         unsigned int i;
915         unsigned int ret;
916         unsigned int nr_found;
917
918         rcu_read_lock();
919 restart:
920         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
921                                 (void ***)pages, index, nr_pages);
922         ret = 0;
923         for (i = 0; i < nr_found; i++) {
924                 struct page *page;
925 repeat:
926                 page = radix_tree_deref_slot((void **)pages[i]);
927                 if (unlikely(!page))
928                         continue;
929
930                 /*
931                  * This can only trigger when the entry at index 0 moves out
932                  * of or back to the root: none yet gotten, safe to restart.
933                  */
934                 if (radix_tree_deref_retry(page))
935                         goto restart;
936
937                 if (!page_cache_get_speculative(page))
938                         goto repeat;
939
940                 /* Has the page moved? */
941                 if (unlikely(page != *((void **)pages[i]))) {
942                         page_cache_release(page);
943                         goto repeat;
944                 }
945
946                 /*
947                  * must check mapping and index after taking the ref.
948                  * otherwise we can get both false positives and false
949                  * negatives, which is just confusing to the caller.
950                  */
951                 if (page->mapping == NULL || page->index != index) {
952                         page_cache_release(page);
953                         break;
954                 }
955
956                 pages[ret] = page;
957                 ret++;
958                 index++;
959         }
960         rcu_read_unlock();
961         return ret;
962 }
963 EXPORT_SYMBOL(find_get_pages_contig);
964
965 /**
966  * find_get_pages_tag - find and return pages that match @tag
967  * @mapping:    the address_space to search
968  * @index:      the starting page index
969  * @tag:        the tag index
970  * @nr_pages:   the maximum number of pages
971  * @pages:      where the resulting pages are placed
972  *
973  * Like find_get_pages, except we only return pages which are tagged with
974  * @tag.   We update @index to index the next page for the traversal.
975  */
976 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
977                         int tag, unsigned int nr_pages, struct page **pages)
978 {
979         unsigned int i;
980         unsigned int ret;
981         unsigned int nr_found;
982
983         rcu_read_lock();
984 restart:
985         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
986                                 (void ***)pages, *index, nr_pages, tag);
987         ret = 0;
988         for (i = 0; i < nr_found; i++) {
989                 struct page *page;
990 repeat:
991                 page = radix_tree_deref_slot((void **)pages[i]);
992                 if (unlikely(!page))
993                         continue;
994
995                 /*
996                  * This can only trigger when the entry at index 0 moves out
997                  * of or back to the root: none yet gotten, safe to restart.
998                  */
999                 if (radix_tree_deref_retry(page))
1000                         goto restart;
1001
1002                 if (!page_cache_get_speculative(page))
1003                         goto repeat;
1004
1005                 /* Has the page moved? */
1006                 if (unlikely(page != *((void **)pages[i]))) {
1007                         page_cache_release(page);
1008                         goto repeat;
1009                 }
1010
1011                 pages[ret] = page;
1012                 ret++;
1013         }
1014
1015         /*
1016          * If all entries were removed before we could secure them,
1017          * try again, because callers stop trying once 0 is returned.
1018          */
1019         if (unlikely(!ret && nr_found))
1020                 goto restart;
1021         rcu_read_unlock();
1022
1023         if (ret)
1024                 *index = pages[ret - 1]->index + 1;
1025
1026         return ret;
1027 }
1028 EXPORT_SYMBOL(find_get_pages_tag);
1029
1030 /**
1031  * grab_cache_page_nowait - returns locked page at given index in given cache
1032  * @mapping: target address_space
1033  * @index: the page index
1034  *
1035  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1036  * This is intended for speculative data generators, where the data can
1037  * be regenerated if the page couldn't be grabbed.  This routine should
1038  * be safe to call while holding the lock for another page.
1039  *
1040  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1041  * and deadlock against the caller's locked page.
1042  */
1043 struct page *
1044 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1045 {
1046         struct page *page = find_get_page(mapping, index);
1047
1048         if (page) {
1049                 if (trylock_page(page))
1050                         return page;
1051                 page_cache_release(page);
1052                 return NULL;
1053         }
1054         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1055         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1056                 page_cache_release(page);
1057                 page = NULL;
1058         }
1059         return page;
1060 }
1061 EXPORT_SYMBOL(grab_cache_page_nowait);
1062
1063 /*
1064  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1065  * a _large_ part of the i/o request. Imagine the worst scenario:
1066  *
1067  *      ---R__________________________________________B__________
1068  *         ^ reading here                             ^ bad block(assume 4k)
1069  *
1070  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1071  * => failing the whole request => read(R) => read(R+1) =>
1072  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1073  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1074  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1075  *
1076  * It is going insane. Fix it by quickly scaling down the readahead size.
1077  */
1078 static void shrink_readahead_size_eio(struct file *filp,
1079                                         struct file_ra_state *ra)
1080 {
1081         ra->ra_pages /= 4;
1082 }
1083
1084 /**
1085  * do_generic_file_read - generic file read routine
1086  * @filp:       the file to read
1087  * @ppos:       current file position
1088  * @desc:       read_descriptor
1089  * @actor:      read method
1090  *
1091  * This is a generic file read routine, and uses the
1092  * mapping->a_ops->readpage() function for the actual low-level stuff.
1093  *
1094  * This is really ugly. But the goto's actually try to clarify some
1095  * of the logic when it comes to error handling etc.
1096  */
1097 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1098                 read_descriptor_t *desc, read_actor_t actor)
1099 {
1100         struct address_space *mapping = filp->f_mapping;
1101         struct inode *inode = mapping->host;
1102         struct file_ra_state *ra = &filp->f_ra;
1103         pgoff_t index;
1104         pgoff_t last_index;
1105         pgoff_t prev_index;
1106         unsigned long offset;      /* offset into pagecache page */
1107         unsigned int prev_offset;
1108         int error;
1109
1110         index = *ppos >> PAGE_CACHE_SHIFT;
1111         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1112         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1113         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1114         offset = *ppos & ~PAGE_CACHE_MASK;
1115
1116         for (;;) {
1117                 struct page *page;
1118                 pgoff_t end_index;
1119                 loff_t isize;
1120                 unsigned long nr, ret;
1121
1122                 cond_resched();
1123 find_page:
1124                 page = find_get_page(mapping, index);
1125                 if (!page) {
1126                         page_cache_sync_readahead(mapping,
1127                                         ra, filp,
1128                                         index, last_index - index);
1129                         page = find_get_page(mapping, index);
1130                         if (unlikely(page == NULL))
1131                                 goto no_cached_page;
1132                 }
1133                 if (PageReadahead(page)) {
1134                         page_cache_async_readahead(mapping,
1135                                         ra, filp, page,
1136                                         index, last_index - index);
1137                 }
1138                 if (!PageUptodate(page)) {
1139                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1140                                         !mapping->a_ops->is_partially_uptodate)
1141                                 goto page_not_up_to_date;
1142                         if (!trylock_page(page))
1143                                 goto page_not_up_to_date;
1144                         /* Did it get truncated before we got the lock? */
1145                         if (!page->mapping)
1146                                 goto page_not_up_to_date_locked;
1147                         if (!mapping->a_ops->is_partially_uptodate(page,
1148                                                                 desc, offset))
1149                                 goto page_not_up_to_date_locked;
1150                         unlock_page(page);
1151                 }
1152 page_ok:
1153                 /*
1154                  * i_size must be checked after we know the page is Uptodate.
1155                  *
1156                  * Checking i_size after the check allows us to calculate
1157                  * the correct value for "nr", which means the zero-filled
1158                  * part of the page is not copied back to userspace (unless
1159                  * another truncate extends the file - this is desired though).
1160                  */
1161
1162                 isize = i_size_read(inode);
1163                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1164                 if (unlikely(!isize || index > end_index)) {
1165                         page_cache_release(page);
1166                         goto out;
1167                 }
1168
1169                 /* nr is the maximum number of bytes to copy from this page */
1170                 nr = PAGE_CACHE_SIZE;
1171                 if (index == end_index) {
1172                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1173                         if (nr <= offset) {
1174                                 page_cache_release(page);
1175                                 goto out;
1176                         }
1177                 }
1178                 nr = nr - offset;
1179
1180                 /* If users can be writing to this page using arbitrary
1181                  * virtual addresses, take care about potential aliasing
1182                  * before reading the page on the kernel side.
1183                  */
1184                 if (mapping_writably_mapped(mapping))
1185                         flush_dcache_page(page);
1186
1187                 /*
1188                  * When a sequential read accesses a page several times,
1189                  * only mark it as accessed the first time.
1190                  */
1191                 if (prev_index != index || offset != prev_offset)
1192                         mark_page_accessed(page);
1193                 prev_index = index;
1194
1195                 /*
1196                  * Ok, we have the page, and it's up-to-date, so
1197                  * now we can copy it to user space...
1198                  *
1199                  * The actor routine returns how many bytes were actually used..
1200                  * NOTE! This may not be the same as how much of a user buffer
1201                  * we filled up (we may be padding etc), so we can only update
1202                  * "pos" here (the actor routine has to update the user buffer
1203                  * pointers and the remaining count).
1204                  */
1205                 ret = actor(desc, page, offset, nr);
1206                 offset += ret;
1207                 index += offset >> PAGE_CACHE_SHIFT;
1208                 offset &= ~PAGE_CACHE_MASK;
1209                 prev_offset = offset;
1210
1211                 page_cache_release(page);
1212                 if (ret == nr && desc->count)
1213                         continue;
1214                 goto out;
1215
1216 page_not_up_to_date:
1217                 /* Get exclusive access to the page ... */
1218                 error = lock_page_killable(page);
1219                 if (unlikely(error))
1220                         goto readpage_error;
1221
1222 page_not_up_to_date_locked:
1223                 /* Did it get truncated before we got the lock? */
1224                 if (!page->mapping) {
1225                         unlock_page(page);
1226                         page_cache_release(page);
1227                         continue;
1228                 }
1229
1230                 /* Did somebody else fill it already? */
1231                 if (PageUptodate(page)) {
1232                         unlock_page(page);
1233                         goto page_ok;
1234                 }
1235
1236 readpage:
1237                 /*
1238                  * A previous I/O error may have been due to temporary
1239                  * failures, eg. multipath errors.
1240                  * PG_error will be set again if readpage fails.
1241                  */
1242                 ClearPageError(page);
1243                 /* Start the actual read. The read will unlock the page. */
1244                 error = mapping->a_ops->readpage(filp, page);
1245
1246                 if (unlikely(error)) {
1247                         if (error == AOP_TRUNCATED_PAGE) {
1248                                 page_cache_release(page);
1249                                 goto find_page;
1250                         }
1251                         goto readpage_error;
1252                 }
1253
1254                 if (!PageUptodate(page)) {
1255                         error = lock_page_killable(page);
1256                         if (unlikely(error))
1257                                 goto readpage_error;
1258                         if (!PageUptodate(page)) {
1259                                 if (page->mapping == NULL) {
1260                                         /*
1261                                          * invalidate_mapping_pages got it
1262                                          */
1263                                         unlock_page(page);
1264                                         page_cache_release(page);
1265                                         goto find_page;
1266                                 }
1267                                 unlock_page(page);
1268                                 shrink_readahead_size_eio(filp, ra);
1269                                 error = -EIO;
1270                                 goto readpage_error;
1271                         }
1272                         unlock_page(page);
1273                 }
1274
1275                 goto page_ok;
1276
1277 readpage_error:
1278                 /* UHHUH! A synchronous read error occurred. Report it */
1279                 desc->error = error;
1280                 page_cache_release(page);
1281                 goto out;
1282
1283 no_cached_page:
1284                 /*
1285                  * Ok, it wasn't cached, so we need to create a new
1286                  * page..
1287                  */
1288                 page = page_cache_alloc_cold(mapping);
1289                 if (!page) {
1290                         desc->error = -ENOMEM;
1291                         goto out;
1292                 }
1293                 error = add_to_page_cache_lru(page, mapping,
1294                                                 index, GFP_KERNEL);
1295                 if (error) {
1296                         page_cache_release(page);
1297                         if (error == -EEXIST)
1298                                 goto find_page;
1299                         desc->error = error;
1300                         goto out;
1301                 }
1302                 goto readpage;
1303         }
1304
1305 out:
1306         ra->prev_pos = prev_index;
1307         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1308         ra->prev_pos |= prev_offset;
1309
1310         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1311         file_accessed(filp);
1312 }
1313
1314 int file_read_actor(read_descriptor_t *desc, struct page *page,
1315                         unsigned long offset, unsigned long size)
1316 {
1317         char *kaddr;
1318         unsigned long left, count = desc->count;
1319
1320         if (size > count)
1321                 size = count;
1322
1323         /*
1324          * Faults on the destination of a read are common, so do it before
1325          * taking the kmap.
1326          */
1327         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1328                 kaddr = kmap_atomic(page, KM_USER0);
1329                 left = __copy_to_user_inatomic(desc->arg.buf,
1330                                                 kaddr + offset, size);
1331                 kunmap_atomic(kaddr, KM_USER0);
1332                 if (left == 0)
1333                         goto success;
1334         }
1335
1336         /* Do it the slow way */
1337         kaddr = kmap(page);
1338         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1339         kunmap(page);
1340
1341         if (left) {
1342                 size -= left;
1343                 desc->error = -EFAULT;
1344         }
1345 success:
1346         desc->count = count - size;
1347         desc->written += size;
1348         desc->arg.buf += size;
1349         return size;
1350 }
1351
1352 /*
1353  * Performs necessary checks before doing a write
1354  * @iov:        io vector request
1355  * @nr_segs:    number of segments in the iovec
1356  * @count:      number of bytes to write
1357  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1358  *
1359  * Adjust number of segments and amount of bytes to write (nr_segs should be
1360  * properly initialized first). Returns appropriate error code that caller
1361  * should return or zero in case that write should be allowed.
1362  */
1363 int generic_segment_checks(const struct iovec *iov,
1364                         unsigned long *nr_segs, size_t *count, int access_flags)
1365 {
1366         unsigned long   seg;
1367         size_t cnt = 0;
1368         for (seg = 0; seg < *nr_segs; seg++) {
1369                 const struct iovec *iv = &iov[seg];
1370
1371                 /*
1372                  * If any segment has a negative length, or the cumulative
1373                  * length ever wraps negative then return -EINVAL.
1374                  */
1375                 cnt += iv->iov_len;
1376                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1377                         return -EINVAL;
1378                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1379                         continue;
1380                 if (seg == 0)
1381                         return -EFAULT;
1382                 *nr_segs = seg;
1383                 cnt -= iv->iov_len;     /* This segment is no good */
1384                 break;
1385         }
1386         *count = cnt;
1387         return 0;
1388 }
1389 EXPORT_SYMBOL(generic_segment_checks);
1390
1391 /**
1392  * generic_file_aio_read - generic filesystem read routine
1393  * @iocb:       kernel I/O control block
1394  * @iov:        io vector request
1395  * @nr_segs:    number of segments in the iovec
1396  * @pos:        current file position
1397  *
1398  * This is the "read()" routine for all filesystems
1399  * that can use the page cache directly.
1400  */
1401 ssize_t
1402 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1403                 unsigned long nr_segs, loff_t pos)
1404 {
1405         struct file *filp = iocb->ki_filp;
1406         ssize_t retval;
1407         unsigned long seg = 0;
1408         size_t count;
1409         loff_t *ppos = &iocb->ki_pos;
1410
1411         count = 0;
1412         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1413         if (retval)
1414                 return retval;
1415
1416         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1417         if (filp->f_flags & O_DIRECT) {
1418                 loff_t size;
1419                 struct address_space *mapping;
1420                 struct inode *inode;
1421
1422                 mapping = filp->f_mapping;
1423                 inode = mapping->host;
1424                 if (!count)
1425                         goto out; /* skip atime */
1426                 size = i_size_read(inode);
1427                 if (pos < size) {
1428                         retval = filemap_write_and_wait_range(mapping, pos,
1429                                         pos + iov_length(iov, nr_segs) - 1);
1430                         if (!retval) {
1431                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1432                                                         iov, pos, nr_segs);
1433                         }
1434                         if (retval > 0) {
1435                                 *ppos = pos + retval;
1436                                 count -= retval;
1437                         }
1438
1439                         /*
1440                          * Btrfs can have a short DIO read if we encounter
1441                          * compressed extents, so if there was an error, or if
1442                          * we've already read everything we wanted to, or if
1443                          * there was a short read because we hit EOF, go ahead
1444                          * and return.  Otherwise fallthrough to buffered io for
1445                          * the rest of the read.
1446                          */
1447                         if (retval < 0 || !count || *ppos >= size) {
1448                                 file_accessed(filp);
1449                                 goto out;
1450                         }
1451                 }
1452         }
1453
1454         count = retval;
1455         for (seg = 0; seg < nr_segs; seg++) {
1456                 read_descriptor_t desc;
1457                 loff_t offset = 0;
1458
1459                 /*
1460                  * If we did a short DIO read we need to skip the section of the
1461                  * iov that we've already read data into.
1462                  */
1463                 if (count) {
1464                         if (count > iov[seg].iov_len) {
1465                                 count -= iov[seg].iov_len;
1466                                 continue;
1467                         }
1468                         offset = count;
1469                         count = 0;
1470                 }
1471
1472                 desc.written = 0;
1473                 desc.arg.buf = iov[seg].iov_base + offset;
1474                 desc.count = iov[seg].iov_len - offset;
1475                 if (desc.count == 0)
1476                         continue;
1477                 desc.error = 0;
1478                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1479                 retval += desc.written;
1480                 if (desc.error) {
1481                         retval = retval ?: desc.error;
1482                         break;
1483                 }
1484                 if (desc.count > 0)
1485                         break;
1486         }
1487 out:
1488         return retval;
1489 }
1490 EXPORT_SYMBOL(generic_file_aio_read);
1491
1492 static ssize_t
1493 do_readahead(struct address_space *mapping, struct file *filp,
1494              pgoff_t index, unsigned long nr)
1495 {
1496         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1497                 return -EINVAL;
1498
1499         force_page_cache_readahead(mapping, filp, index, nr);
1500         return 0;
1501 }
1502
1503 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1504 {
1505         ssize_t ret;
1506         struct file *file;
1507
1508         ret = -EBADF;
1509         file = fget(fd);
1510         if (file) {
1511                 if (file->f_mode & FMODE_READ) {
1512                         struct address_space *mapping = file->f_mapping;
1513                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1514                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1515                         unsigned long len = end - start + 1;
1516                         ret = do_readahead(mapping, file, start, len);
1517                 }
1518                 fput(file);
1519         }
1520         return ret;
1521 }
1522 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1523 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1524 {
1525         return SYSC_readahead((int) fd, offset, (size_t) count);
1526 }
1527 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1528 #endif
1529
1530 #ifdef CONFIG_MMU
1531 /**
1532  * page_cache_read - adds requested page to the page cache if not already there
1533  * @file:       file to read
1534  * @offset:     page index
1535  *
1536  * This adds the requested page to the page cache if it isn't already there,
1537  * and schedules an I/O to read in its contents from disk.
1538  */
1539 static int page_cache_read(struct file *file, pgoff_t offset)
1540 {
1541         struct address_space *mapping = file->f_mapping;
1542         struct page *page; 
1543         int ret;
1544
1545         do {
1546                 page = page_cache_alloc_cold(mapping);
1547                 if (!page)
1548                         return -ENOMEM;
1549
1550                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1551                 if (ret == 0)
1552                         ret = mapping->a_ops->readpage(file, page);
1553                 else if (ret == -EEXIST)
1554                         ret = 0; /* losing race to add is OK */
1555
1556                 page_cache_release(page);
1557
1558         } while (ret == AOP_TRUNCATED_PAGE);
1559                 
1560         return ret;
1561 }
1562
1563 #define MMAP_LOTSAMISS  (100)
1564
1565 /*
1566  * Synchronous readahead happens when we don't even find
1567  * a page in the page cache at all.
1568  */
1569 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1570                                    struct file_ra_state *ra,
1571                                    struct file *file,
1572                                    pgoff_t offset)
1573 {
1574         unsigned long ra_pages;
1575         struct address_space *mapping = file->f_mapping;
1576
1577         /* If we don't want any read-ahead, don't bother */
1578         if (VM_RandomReadHint(vma))
1579                 return;
1580
1581         if (VM_SequentialReadHint(vma) ||
1582                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1583                 page_cache_sync_readahead(mapping, ra, file, offset,
1584                                           ra->ra_pages);
1585                 return;
1586         }
1587
1588         if (ra->mmap_miss < INT_MAX)
1589                 ra->mmap_miss++;
1590
1591         /*
1592          * Do we miss much more than hit in this file? If so,
1593          * stop bothering with read-ahead. It will only hurt.
1594          */
1595         if (ra->mmap_miss > MMAP_LOTSAMISS)
1596                 return;
1597
1598         /*
1599          * mmap read-around
1600          */
1601         ra_pages = max_sane_readahead(ra->ra_pages);
1602         if (ra_pages) {
1603                 ra->start = max_t(long, 0, offset - ra_pages/2);
1604                 ra->size = ra_pages;
1605                 ra->async_size = 0;
1606                 ra_submit(ra, mapping, file);
1607         }
1608 }
1609
1610 /*
1611  * Asynchronous readahead happens when we find the page and PG_readahead,
1612  * so we want to possibly extend the readahead further..
1613  */
1614 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1615                                     struct file_ra_state *ra,
1616                                     struct file *file,
1617                                     struct page *page,
1618                                     pgoff_t offset)
1619 {
1620         struct address_space *mapping = file->f_mapping;
1621
1622         /* If we don't want any read-ahead, don't bother */
1623         if (VM_RandomReadHint(vma))
1624                 return;
1625         if (ra->mmap_miss > 0)
1626                 ra->mmap_miss--;
1627         if (PageReadahead(page))
1628                 page_cache_async_readahead(mapping, ra, file,
1629                                            page, offset, ra->ra_pages);
1630 }
1631
1632 /**
1633  * filemap_fault - read in file data for page fault handling
1634  * @vma:        vma in which the fault was taken
1635  * @vmf:        struct vm_fault containing details of the fault
1636  *
1637  * filemap_fault() is invoked via the vma operations vector for a
1638  * mapped memory region to read in file data during a page fault.
1639  *
1640  * The goto's are kind of ugly, but this streamlines the normal case of having
1641  * it in the page cache, and handles the special cases reasonably without
1642  * having a lot of duplicated code.
1643  */
1644 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1645 {
1646         int error;
1647         struct file *file = vma->vm_file;
1648         struct address_space *mapping = file->f_mapping;
1649         struct file_ra_state *ra = &file->f_ra;
1650         struct inode *inode = mapping->host;
1651         pgoff_t offset = vmf->pgoff;
1652         struct page *page;
1653         pgoff_t size;
1654         int ret = 0;
1655
1656         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1657         if (offset >= size)
1658                 return VM_FAULT_SIGBUS;
1659
1660         /*
1661          * Do we have something in the page cache already?
1662          */
1663         page = find_get_page(mapping, offset);
1664         if (likely(page)) {
1665                 /*
1666                  * We found the page, so try async readahead before
1667                  * waiting for the lock.
1668                  */
1669                 do_async_mmap_readahead(vma, ra, file, page, offset);
1670         } else {
1671                 /* No page in the page cache at all */
1672                 do_sync_mmap_readahead(vma, ra, file, offset);
1673                 count_vm_event(PGMAJFAULT);
1674                 ret = VM_FAULT_MAJOR;
1675 retry_find:
1676                 page = find_get_page(mapping, offset);
1677                 if (!page)
1678                         goto no_cached_page;
1679         }
1680
1681         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1682                 page_cache_release(page);
1683                 return ret | VM_FAULT_RETRY;
1684         }
1685
1686         /* Did it get truncated? */
1687         if (unlikely(page->mapping != mapping)) {
1688                 unlock_page(page);
1689                 put_page(page);
1690                 goto retry_find;
1691         }
1692         VM_BUG_ON(page->index != offset);
1693
1694         /*
1695          * We have a locked page in the page cache, now we need to check
1696          * that it's up-to-date. If not, it is going to be due to an error.
1697          */
1698         if (unlikely(!PageUptodate(page)))
1699                 goto page_not_uptodate;
1700
1701         /*
1702          * Found the page and have a reference on it.
1703          * We must recheck i_size under page lock.
1704          */
1705         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1706         if (unlikely(offset >= size)) {
1707                 unlock_page(page);
1708                 page_cache_release(page);
1709                 return VM_FAULT_SIGBUS;
1710         }
1711
1712         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1713         vmf->page = page;
1714         return ret | VM_FAULT_LOCKED;
1715
1716 no_cached_page:
1717         /*
1718          * We're only likely to ever get here if MADV_RANDOM is in
1719          * effect.
1720          */
1721         error = page_cache_read(file, offset);
1722
1723         /*
1724          * The page we want has now been added to the page cache.
1725          * In the unlikely event that someone removed it in the
1726          * meantime, we'll just come back here and read it again.
1727          */
1728         if (error >= 0)
1729                 goto retry_find;
1730
1731         /*
1732          * An error return from page_cache_read can result if the
1733          * system is low on memory, or a problem occurs while trying
1734          * to schedule I/O.
1735          */
1736         if (error == -ENOMEM)
1737                 return VM_FAULT_OOM;
1738         return VM_FAULT_SIGBUS;
1739
1740 page_not_uptodate:
1741         /*
1742          * Umm, take care of errors if the page isn't up-to-date.
1743          * Try to re-read it _once_. We do this synchronously,
1744          * because there really aren't any performance issues here
1745          * and we need to check for errors.
1746          */
1747         ClearPageError(page);
1748         error = mapping->a_ops->readpage(file, page);
1749         if (!error) {
1750                 wait_on_page_locked(page);
1751                 if (!PageUptodate(page))
1752                         error = -EIO;
1753         }
1754         page_cache_release(page);
1755
1756         if (!error || error == AOP_TRUNCATED_PAGE)
1757                 goto retry_find;
1758
1759         /* Things didn't work out. Return zero to tell the mm layer so. */
1760         shrink_readahead_size_eio(file, ra);
1761         return VM_FAULT_SIGBUS;
1762 }
1763 EXPORT_SYMBOL(filemap_fault);
1764
1765 const struct vm_operations_struct generic_file_vm_ops = {
1766         .fault          = filemap_fault,
1767 };
1768
1769 /* This is used for a general mmap of a disk file */
1770
1771 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1772 {
1773         struct address_space *mapping = file->f_mapping;
1774
1775         if (!mapping->a_ops->readpage)
1776                 return -ENOEXEC;
1777         file_accessed(file);
1778         vma->vm_ops = &generic_file_vm_ops;
1779         vma->vm_flags |= VM_CAN_NONLINEAR;
1780         return 0;
1781 }
1782
1783 /*
1784  * This is for filesystems which do not implement ->writepage.
1785  */
1786 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1787 {
1788         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1789                 return -EINVAL;
1790         return generic_file_mmap(file, vma);
1791 }
1792 #else
1793 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1794 {
1795         return -ENOSYS;
1796 }
1797 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1798 {
1799         return -ENOSYS;
1800 }
1801 #endif /* CONFIG_MMU */
1802
1803 EXPORT_SYMBOL(generic_file_mmap);
1804 EXPORT_SYMBOL(generic_file_readonly_mmap);
1805
1806 static struct page *__read_cache_page(struct address_space *mapping,
1807                                 pgoff_t index,
1808                                 int (*filler)(void *,struct page*),
1809                                 void *data,
1810                                 gfp_t gfp)
1811 {
1812         struct page *page;
1813         int err;
1814 repeat:
1815         page = find_get_page(mapping, index);
1816         if (!page) {
1817                 page = __page_cache_alloc(gfp | __GFP_COLD);
1818                 if (!page)
1819                         return ERR_PTR(-ENOMEM);
1820                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1821                 if (unlikely(err)) {
1822                         page_cache_release(page);
1823                         if (err == -EEXIST)
1824                                 goto repeat;
1825                         /* Presumably ENOMEM for radix tree node */
1826                         return ERR_PTR(err);
1827                 }
1828                 err = filler(data, page);
1829                 if (err < 0) {
1830                         page_cache_release(page);
1831                         page = ERR_PTR(err);
1832                 }
1833         }
1834         return page;
1835 }
1836
1837 static struct page *do_read_cache_page(struct address_space *mapping,
1838                                 pgoff_t index,
1839                                 int (*filler)(void *,struct page*),
1840                                 void *data,
1841                                 gfp_t gfp)
1842
1843 {
1844         struct page *page;
1845         int err;
1846
1847 retry:
1848         page = __read_cache_page(mapping, index, filler, data, gfp);
1849         if (IS_ERR(page))
1850                 return page;
1851         if (PageUptodate(page))
1852                 goto out;
1853
1854         lock_page(page);
1855         if (!page->mapping) {
1856                 unlock_page(page);
1857                 page_cache_release(page);
1858                 goto retry;
1859         }
1860         if (PageUptodate(page)) {
1861                 unlock_page(page);
1862                 goto out;
1863         }
1864         err = filler(data, page);
1865         if (err < 0) {
1866                 page_cache_release(page);
1867                 return ERR_PTR(err);
1868         }
1869 out:
1870         mark_page_accessed(page);
1871         return page;
1872 }
1873
1874 /**
1875  * read_cache_page_async - read into page cache, fill it if needed
1876  * @mapping:    the page's address_space
1877  * @index:      the page index
1878  * @filler:     function to perform the read
1879  * @data:       destination for read data
1880  *
1881  * Same as read_cache_page, but don't wait for page to become unlocked
1882  * after submitting it to the filler.
1883  *
1884  * Read into the page cache. If a page already exists, and PageUptodate() is
1885  * not set, try to fill the page but don't wait for it to become unlocked.
1886  *
1887  * If the page does not get brought uptodate, return -EIO.
1888  */
1889 struct page *read_cache_page_async(struct address_space *mapping,
1890                                 pgoff_t index,
1891                                 int (*filler)(void *,struct page*),
1892                                 void *data)
1893 {
1894         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1895 }
1896 EXPORT_SYMBOL(read_cache_page_async);
1897
1898 static struct page *wait_on_page_read(struct page *page)
1899 {
1900         if (!IS_ERR(page)) {
1901                 wait_on_page_locked(page);
1902                 if (!PageUptodate(page)) {
1903                         page_cache_release(page);
1904                         page = ERR_PTR(-EIO);
1905                 }
1906         }
1907         return page;
1908 }
1909
1910 /**
1911  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1912  * @mapping:    the page's address_space
1913  * @index:      the page index
1914  * @gfp:        the page allocator flags to use if allocating
1915  *
1916  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1917  * any new page allocations done using the specified allocation flags. Note
1918  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1919  * expect to do this atomically or anything like that - but you can pass in
1920  * other page requirements.
1921  *
1922  * If the page does not get brought uptodate, return -EIO.
1923  */
1924 struct page *read_cache_page_gfp(struct address_space *mapping,
1925                                 pgoff_t index,
1926                                 gfp_t gfp)
1927 {
1928         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1929
1930         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1931 }
1932 EXPORT_SYMBOL(read_cache_page_gfp);
1933
1934 /**
1935  * read_cache_page - read into page cache, fill it if needed
1936  * @mapping:    the page's address_space
1937  * @index:      the page index
1938  * @filler:     function to perform the read
1939  * @data:       destination for read data
1940  *
1941  * Read into the page cache. If a page already exists, and PageUptodate() is
1942  * not set, try to fill the page then wait for it to become unlocked.
1943  *
1944  * If the page does not get brought uptodate, return -EIO.
1945  */
1946 struct page *read_cache_page(struct address_space *mapping,
1947                                 pgoff_t index,
1948                                 int (*filler)(void *,struct page*),
1949                                 void *data)
1950 {
1951         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1952 }
1953 EXPORT_SYMBOL(read_cache_page);
1954
1955 /*
1956  * The logic we want is
1957  *
1958  *      if suid or (sgid and xgrp)
1959  *              remove privs
1960  */
1961 int should_remove_suid(struct dentry *dentry)
1962 {
1963         mode_t mode = dentry->d_inode->i_mode;
1964         int kill = 0;
1965
1966         /* suid always must be killed */
1967         if (unlikely(mode & S_ISUID))
1968                 kill = ATTR_KILL_SUID;
1969
1970         /*
1971          * sgid without any exec bits is just a mandatory locking mark; leave
1972          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1973          */
1974         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1975                 kill |= ATTR_KILL_SGID;
1976
1977         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1978                 return kill;
1979
1980         return 0;
1981 }
1982 EXPORT_SYMBOL(should_remove_suid);
1983
1984 static int __remove_suid(struct dentry *dentry, int kill)
1985 {
1986         struct iattr newattrs;
1987
1988         newattrs.ia_valid = ATTR_FORCE | kill;
1989         return notify_change(dentry, &newattrs);
1990 }
1991
1992 int file_remove_suid(struct file *file)
1993 {
1994         struct dentry *dentry = file->f_path.dentry;
1995         int killsuid = should_remove_suid(dentry);
1996         int killpriv = security_inode_need_killpriv(dentry);
1997         int error = 0;
1998
1999         if (killpriv < 0)
2000                 return killpriv;
2001         if (killpriv)
2002                 error = security_inode_killpriv(dentry);
2003         if (!error && killsuid)
2004                 error = __remove_suid(dentry, killsuid);
2005
2006         return error;
2007 }
2008 EXPORT_SYMBOL(file_remove_suid);
2009
2010 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2011                         const struct iovec *iov, size_t base, size_t bytes)
2012 {
2013         size_t copied = 0, left = 0;
2014
2015         while (bytes) {
2016                 char __user *buf = iov->iov_base + base;
2017                 int copy = min(bytes, iov->iov_len - base);
2018
2019                 base = 0;
2020                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2021                 copied += copy;
2022                 bytes -= copy;
2023                 vaddr += copy;
2024                 iov++;
2025
2026                 if (unlikely(left))
2027                         break;
2028         }
2029         return copied - left;
2030 }
2031
2032 /*
2033  * Copy as much as we can into the page and return the number of bytes which
2034  * were successfully copied.  If a fault is encountered then return the number of
2035  * bytes which were copied.
2036  */
2037 size_t iov_iter_copy_from_user_atomic(struct page *page,
2038                 struct iov_iter *i, unsigned long offset, size_t bytes)
2039 {
2040         char *kaddr;
2041         size_t copied;
2042
2043         BUG_ON(!in_atomic());
2044         kaddr = kmap_atomic(page, KM_USER0);
2045         if (likely(i->nr_segs == 1)) {
2046                 int left;
2047                 char __user *buf = i->iov->iov_base + i->iov_offset;
2048                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2049                 copied = bytes - left;
2050         } else {
2051                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2052                                                 i->iov, i->iov_offset, bytes);
2053         }
2054         kunmap_atomic(kaddr, KM_USER0);
2055
2056         return copied;
2057 }
2058 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2059
2060 /*
2061  * This has the same sideeffects and return value as
2062  * iov_iter_copy_from_user_atomic().
2063  * The difference is that it attempts to resolve faults.
2064  * Page must not be locked.
2065  */
2066 size_t iov_iter_copy_from_user(struct page *page,
2067                 struct iov_iter *i, unsigned long offset, size_t bytes)
2068 {
2069         char *kaddr;
2070         size_t copied;
2071
2072         kaddr = kmap(page);
2073         if (likely(i->nr_segs == 1)) {
2074                 int left;
2075                 char __user *buf = i->iov->iov_base + i->iov_offset;
2076                 left = __copy_from_user(kaddr + offset, buf, bytes);
2077                 copied = bytes - left;
2078         } else {
2079                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2080                                                 i->iov, i->iov_offset, bytes);
2081         }
2082         kunmap(page);
2083         return copied;
2084 }
2085 EXPORT_SYMBOL(iov_iter_copy_from_user);
2086
2087 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2088 {
2089         BUG_ON(i->count < bytes);
2090
2091         if (likely(i->nr_segs == 1)) {
2092                 i->iov_offset += bytes;
2093                 i->count -= bytes;
2094         } else {
2095                 const struct iovec *iov = i->iov;
2096                 size_t base = i->iov_offset;
2097
2098                 /*
2099                  * The !iov->iov_len check ensures we skip over unlikely
2100                  * zero-length segments (without overruning the iovec).
2101                  */
2102                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2103                         int copy;
2104
2105                         copy = min(bytes, iov->iov_len - base);
2106                         BUG_ON(!i->count || i->count < copy);
2107                         i->count -= copy;
2108                         bytes -= copy;
2109                         base += copy;
2110                         if (iov->iov_len == base) {
2111                                 iov++;
2112                                 base = 0;
2113                         }
2114                 }
2115                 i->iov = iov;
2116                 i->iov_offset = base;
2117         }
2118 }
2119 EXPORT_SYMBOL(iov_iter_advance);
2120
2121 /*
2122  * Fault in the first iovec of the given iov_iter, to a maximum length
2123  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2124  * accessed (ie. because it is an invalid address).
2125  *
2126  * writev-intensive code may want this to prefault several iovecs -- that
2127  * would be possible (callers must not rely on the fact that _only_ the
2128  * first iovec will be faulted with the current implementation).
2129  */
2130 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2131 {
2132         char __user *buf = i->iov->iov_base + i->iov_offset;
2133         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2134         return fault_in_pages_readable(buf, bytes);
2135 }
2136 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2137
2138 /*
2139  * Return the count of just the current iov_iter segment.
2140  */
2141 size_t iov_iter_single_seg_count(struct iov_iter *i)
2142 {
2143         const struct iovec *iov = i->iov;
2144         if (i->nr_segs == 1)
2145                 return i->count;
2146         else
2147                 return min(i->count, iov->iov_len - i->iov_offset);
2148 }
2149 EXPORT_SYMBOL(iov_iter_single_seg_count);
2150
2151 /*
2152  * Performs necessary checks before doing a write
2153  *
2154  * Can adjust writing position or amount of bytes to write.
2155  * Returns appropriate error code that caller should return or
2156  * zero in case that write should be allowed.
2157  */
2158 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2159 {
2160         struct inode *inode = file->f_mapping->host;
2161         unsigned long limit = rlimit(RLIMIT_FSIZE);
2162
2163         if (unlikely(*pos < 0))
2164                 return -EINVAL;
2165
2166         if (!isblk) {
2167                 /* FIXME: this is for backwards compatibility with 2.4 */
2168                 if (file->f_flags & O_APPEND)
2169                         *pos = i_size_read(inode);
2170
2171                 if (limit != RLIM_INFINITY) {
2172                         if (*pos >= limit) {
2173                                 send_sig(SIGXFSZ, current, 0);
2174                                 return -EFBIG;
2175                         }
2176                         if (*count > limit - (typeof(limit))*pos) {
2177                                 *count = limit - (typeof(limit))*pos;
2178                         }
2179                 }
2180         }
2181
2182         /*
2183          * LFS rule
2184          */
2185         if (unlikely(*pos + *count > MAX_NON_LFS &&
2186                                 !(file->f_flags & O_LARGEFILE))) {
2187                 if (*pos >= MAX_NON_LFS) {
2188                         return -EFBIG;
2189                 }
2190                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2191                         *count = MAX_NON_LFS - (unsigned long)*pos;
2192                 }
2193         }
2194
2195         /*
2196          * Are we about to exceed the fs block limit ?
2197          *
2198          * If we have written data it becomes a short write.  If we have
2199          * exceeded without writing data we send a signal and return EFBIG.
2200          * Linus frestrict idea will clean these up nicely..
2201          */
2202         if (likely(!isblk)) {
2203                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2204                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2205                                 return -EFBIG;
2206                         }
2207                         /* zero-length writes at ->s_maxbytes are OK */
2208                 }
2209
2210                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2211                         *count = inode->i_sb->s_maxbytes - *pos;
2212         } else {
2213 #ifdef CONFIG_BLOCK
2214                 loff_t isize;
2215                 if (bdev_read_only(I_BDEV(inode)))
2216                         return -EPERM;
2217                 isize = i_size_read(inode);
2218                 if (*pos >= isize) {
2219                         if (*count || *pos > isize)
2220                                 return -ENOSPC;
2221                 }
2222
2223                 if (*pos + *count > isize)
2224                         *count = isize - *pos;
2225 #else
2226                 return -EPERM;
2227 #endif
2228         }
2229         return 0;
2230 }
2231 EXPORT_SYMBOL(generic_write_checks);
2232
2233 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2234                                 loff_t pos, unsigned len, unsigned flags,
2235                                 struct page **pagep, void **fsdata)
2236 {
2237         const struct address_space_operations *aops = mapping->a_ops;
2238
2239         return aops->write_begin(file, mapping, pos, len, flags,
2240                                                         pagep, fsdata);
2241 }
2242 EXPORT_SYMBOL(pagecache_write_begin);
2243
2244 int pagecache_write_end(struct file *file, struct address_space *mapping,
2245                                 loff_t pos, unsigned len, unsigned copied,
2246                                 struct page *page, void *fsdata)
2247 {
2248         const struct address_space_operations *aops = mapping->a_ops;
2249
2250         mark_page_accessed(page);
2251         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2252 }
2253 EXPORT_SYMBOL(pagecache_write_end);
2254
2255 ssize_t
2256 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2257                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2258                 size_t count, size_t ocount)
2259 {
2260         struct file     *file = iocb->ki_filp;
2261         struct address_space *mapping = file->f_mapping;
2262         struct inode    *inode = mapping->host;
2263         ssize_t         written;
2264         size_t          write_len;
2265         pgoff_t         end;
2266
2267         if (count != ocount)
2268                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2269
2270         write_len = iov_length(iov, *nr_segs);
2271         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2272
2273         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2274         if (written)
2275                 goto out;
2276
2277         /*
2278          * After a write we want buffered reads to be sure to go to disk to get
2279          * the new data.  We invalidate clean cached page from the region we're
2280          * about to write.  We do this *before* the write so that we can return
2281          * without clobbering -EIOCBQUEUED from ->direct_IO().
2282          */
2283         if (mapping->nrpages) {
2284                 written = invalidate_inode_pages2_range(mapping,
2285                                         pos >> PAGE_CACHE_SHIFT, end);
2286                 /*
2287                  * If a page can not be invalidated, return 0 to fall back
2288                  * to buffered write.
2289                  */
2290                 if (written) {
2291                         if (written == -EBUSY)
2292                                 return 0;
2293                         goto out;
2294                 }
2295         }
2296
2297         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2298
2299         /*
2300          * Finally, try again to invalidate clean pages which might have been
2301          * cached by non-direct readahead, or faulted in by get_user_pages()
2302          * if the source of the write was an mmap'ed region of the file
2303          * we're writing.  Either one is a pretty crazy thing to do,
2304          * so we don't support it 100%.  If this invalidation
2305          * fails, tough, the write still worked...
2306          */
2307         if (mapping->nrpages) {
2308                 invalidate_inode_pages2_range(mapping,
2309                                               pos >> PAGE_CACHE_SHIFT, end);
2310         }
2311
2312         if (written > 0) {
2313                 pos += written;
2314                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2315                         i_size_write(inode, pos);
2316                         mark_inode_dirty(inode);
2317                 }
2318                 *ppos = pos;
2319         }
2320 out:
2321         return written;
2322 }
2323 EXPORT_SYMBOL(generic_file_direct_write);
2324
2325 /*
2326  * Find or create a page at the given pagecache position. Return the locked
2327  * page. This function is specifically for buffered writes.
2328  */
2329 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2330                                         pgoff_t index, unsigned flags)
2331 {
2332         int status;
2333         struct page *page;
2334         gfp_t gfp_notmask = 0;
2335         if (flags & AOP_FLAG_NOFS)
2336                 gfp_notmask = __GFP_FS;
2337 repeat:
2338         page = find_lock_page(mapping, index);
2339         if (page)
2340                 return page;
2341
2342         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2343         if (!page)
2344                 return NULL;
2345         status = add_to_page_cache_lru(page, mapping, index,
2346                                                 GFP_KERNEL & ~gfp_notmask);
2347         if (unlikely(status)) {
2348                 page_cache_release(page);
2349                 if (status == -EEXIST)
2350                         goto repeat;
2351                 return NULL;
2352         }
2353         return page;
2354 }
2355 EXPORT_SYMBOL(grab_cache_page_write_begin);
2356
2357 static ssize_t generic_perform_write(struct file *file,
2358                                 struct iov_iter *i, loff_t pos)
2359 {
2360         struct address_space *mapping = file->f_mapping;
2361         const struct address_space_operations *a_ops = mapping->a_ops;
2362         long status = 0;
2363         ssize_t written = 0;
2364         unsigned int flags = 0;
2365
2366         /*
2367          * Copies from kernel address space cannot fail (NFSD is a big user).
2368          */
2369         if (segment_eq(get_fs(), KERNEL_DS))
2370                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2371
2372         do {
2373                 struct page *page;
2374                 unsigned long offset;   /* Offset into pagecache page */
2375                 unsigned long bytes;    /* Bytes to write to page */
2376                 size_t copied;          /* Bytes copied from user */
2377                 void *fsdata;
2378
2379                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2380                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2381                                                 iov_iter_count(i));
2382
2383 again:
2384
2385                 /*
2386                  * Bring in the user page that we will copy from _first_.
2387                  * Otherwise there's a nasty deadlock on copying from the
2388                  * same page as we're writing to, without it being marked
2389                  * up-to-date.
2390                  *
2391                  * Not only is this an optimisation, but it is also required
2392                  * to check that the address is actually valid, when atomic
2393                  * usercopies are used, below.
2394                  */
2395                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2396                         status = -EFAULT;
2397                         break;
2398                 }
2399
2400                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2401                                                 &page, &fsdata);
2402                 if (unlikely(status))
2403                         break;
2404
2405                 if (mapping_writably_mapped(mapping))
2406                         flush_dcache_page(page);
2407
2408                 pagefault_disable();
2409                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2410                 pagefault_enable();
2411                 flush_dcache_page(page);
2412
2413                 mark_page_accessed(page);
2414                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2415                                                 page, fsdata);
2416                 if (unlikely(status < 0))
2417                         break;
2418                 copied = status;
2419
2420                 cond_resched();
2421
2422                 iov_iter_advance(i, copied);
2423                 if (unlikely(copied == 0)) {
2424                         /*
2425                          * If we were unable to copy any data at all, we must
2426                          * fall back to a single segment length write.
2427                          *
2428                          * If we didn't fallback here, we could livelock
2429                          * because not all segments in the iov can be copied at
2430                          * once without a pagefault.
2431                          */
2432                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2433                                                 iov_iter_single_seg_count(i));
2434                         goto again;
2435                 }
2436                 pos += copied;
2437                 written += copied;
2438
2439                 balance_dirty_pages_ratelimited(mapping);
2440
2441         } while (iov_iter_count(i));
2442
2443         return written ? written : status;
2444 }
2445
2446 ssize_t
2447 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2448                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2449                 size_t count, ssize_t written)
2450 {
2451         struct file *file = iocb->ki_filp;
2452         ssize_t status;
2453         struct iov_iter i;
2454
2455         iov_iter_init(&i, iov, nr_segs, count, written);
2456         status = generic_perform_write(file, &i, pos);
2457
2458         if (likely(status >= 0)) {
2459                 written += status;
2460                 *ppos = pos + status;
2461         }
2462         
2463         return written ? written : status;
2464 }
2465 EXPORT_SYMBOL(generic_file_buffered_write);
2466
2467 /**
2468  * __generic_file_aio_write - write data to a file
2469  * @iocb:       IO state structure (file, offset, etc.)
2470  * @iov:        vector with data to write
2471  * @nr_segs:    number of segments in the vector
2472  * @ppos:       position where to write
2473  *
2474  * This function does all the work needed for actually writing data to a
2475  * file. It does all basic checks, removes SUID from the file, updates
2476  * modification times and calls proper subroutines depending on whether we
2477  * do direct IO or a standard buffered write.
2478  *
2479  * It expects i_mutex to be grabbed unless we work on a block device or similar
2480  * object which does not need locking at all.
2481  *
2482  * This function does *not* take care of syncing data in case of O_SYNC write.
2483  * A caller has to handle it. This is mainly due to the fact that we want to
2484  * avoid syncing under i_mutex.
2485  */
2486 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2487                                  unsigned long nr_segs, loff_t *ppos)
2488 {
2489         struct file *file = iocb->ki_filp;
2490         struct address_space * mapping = file->f_mapping;
2491         size_t ocount;          /* original count */
2492         size_t count;           /* after file limit checks */
2493         struct inode    *inode = mapping->host;
2494         loff_t          pos;
2495         ssize_t         written;
2496         ssize_t         err;
2497
2498         ocount = 0;
2499         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2500         if (err)
2501                 return err;
2502
2503         count = ocount;
2504         pos = *ppos;
2505
2506         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2507
2508         /* We can write back this queue in page reclaim */
2509         current->backing_dev_info = mapping->backing_dev_info;
2510         written = 0;
2511
2512         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2513         if (err)
2514                 goto out;
2515
2516         if (count == 0)
2517                 goto out;
2518
2519         err = file_remove_suid(file);
2520         if (err)
2521                 goto out;
2522
2523         file_update_time(file);
2524
2525         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2526         if (unlikely(file->f_flags & O_DIRECT)) {
2527                 loff_t endbyte;
2528                 ssize_t written_buffered;
2529
2530                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2531                                                         ppos, count, ocount);
2532                 if (written < 0 || written == count)
2533                         goto out;
2534                 /*
2535                  * direct-io write to a hole: fall through to buffered I/O
2536                  * for completing the rest of the request.
2537                  */
2538                 pos += written;
2539                 count -= written;
2540                 written_buffered = generic_file_buffered_write(iocb, iov,
2541                                                 nr_segs, pos, ppos, count,
2542                                                 written);
2543                 /*
2544                  * If generic_file_buffered_write() retuned a synchronous error
2545                  * then we want to return the number of bytes which were
2546                  * direct-written, or the error code if that was zero.  Note
2547                  * that this differs from normal direct-io semantics, which
2548                  * will return -EFOO even if some bytes were written.
2549                  */
2550                 if (written_buffered < 0) {
2551                         err = written_buffered;
2552                         goto out;
2553                 }
2554
2555                 /*
2556                  * We need to ensure that the page cache pages are written to
2557                  * disk and invalidated to preserve the expected O_DIRECT
2558                  * semantics.
2559                  */
2560                 endbyte = pos + written_buffered - written - 1;
2561                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2562                 if (err == 0) {
2563                         written = written_buffered;
2564                         invalidate_mapping_pages(mapping,
2565                                                  pos >> PAGE_CACHE_SHIFT,
2566                                                  endbyte >> PAGE_CACHE_SHIFT);
2567                 } else {
2568                         /*
2569                          * We don't know how much we wrote, so just return
2570                          * the number of bytes which were direct-written
2571                          */
2572                 }
2573         } else {
2574                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2575                                 pos, ppos, count, written);
2576         }
2577 out:
2578         current->backing_dev_info = NULL;
2579         return written ? written : err;
2580 }
2581 EXPORT_SYMBOL(__generic_file_aio_write);
2582
2583 /**
2584  * generic_file_aio_write - write data to a file
2585  * @iocb:       IO state structure
2586  * @iov:        vector with data to write
2587  * @nr_segs:    number of segments in the vector
2588  * @pos:        position in file where to write
2589  *
2590  * This is a wrapper around __generic_file_aio_write() to be used by most
2591  * filesystems. It takes care of syncing the file in case of O_SYNC file
2592  * and acquires i_mutex as needed.
2593  */
2594 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2595                 unsigned long nr_segs, loff_t pos)
2596 {
2597         struct file *file = iocb->ki_filp;
2598         struct inode *inode = file->f_mapping->host;
2599         ssize_t ret;
2600
2601         BUG_ON(iocb->ki_pos != pos);
2602
2603         mutex_lock(&inode->i_mutex);
2604         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2605         mutex_unlock(&inode->i_mutex);
2606
2607         if (ret > 0 || ret == -EIOCBQUEUED) {
2608                 ssize_t err;
2609
2610                 err = generic_write_sync(file, pos, ret);
2611                 if (err < 0 && ret > 0)
2612                         ret = err;
2613         }
2614         return ret;
2615 }
2616 EXPORT_SYMBOL(generic_file_aio_write);
2617
2618 /**
2619  * try_to_release_page() - release old fs-specific metadata on a page
2620  *
2621  * @page: the page which the kernel is trying to free
2622  * @gfp_mask: memory allocation flags (and I/O mode)
2623  *
2624  * The address_space is to try to release any data against the page
2625  * (presumably at page->private).  If the release was successful, return `1'.
2626  * Otherwise return zero.
2627  *
2628  * This may also be called if PG_fscache is set on a page, indicating that the
2629  * page is known to the local caching routines.
2630  *
2631  * The @gfp_mask argument specifies whether I/O may be performed to release
2632  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2633  *
2634  */
2635 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2636 {
2637         struct address_space * const mapping = page->mapping;
2638
2639         BUG_ON(!PageLocked(page));
2640         if (PageWriteback(page))
2641                 return 0;
2642
2643         if (mapping && mapping->a_ops->releasepage)
2644                 return mapping->a_ops->releasepage(page, gfp_mask);
2645         return try_to_free_buffers(page);
2646 }
2647
2648 EXPORT_SYMBOL(try_to_release_page);