867d40222ec798ce99a8332b991e8d7c4bdfb8ec
[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 <linux/cleancache.h>
38 #include "internal.h"
39
40 /*
41  * FIXME: remove all knowledge of the buffer layer from the core VM
42  */
43 #include <linux/buffer_head.h> /* for try_to_free_buffers */
44
45 #include <asm/mman.h>
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_mutex              (truncate_pagecache)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_mutex
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  bdi->wb.list_lock
82  *    sb_lock                   (fs/fs-writeback.c)
83  *    ->mapping->tree_lock      (__sync_single_inode)
84  *
85  *  ->i_mmap_mutex
86  *    ->anon_vma.lock           (vma_adjust)
87  *
88  *  ->anon_vma.lock
89  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
90  *
91  *  ->page_table_lock or pte_lock
92  *    ->swap_lock               (try_to_unmap_one)
93  *    ->private_lock            (try_to_unmap_one)
94  *    ->tree_lock               (try_to_unmap_one)
95  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
96  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
97  *    ->private_lock            (page_remove_rmap->set_page_dirty)
98  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
99  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
100  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
101  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
102  *    ->inode->i_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_mutex
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         /*
120          * if we're uptodate, flush out into the cleancache, otherwise
121          * invalidate any existing cleancache entries.  We can't leave
122          * stale data around in the cleancache once our page is gone
123          */
124         if (PageUptodate(page) && PageMappedToDisk(page))
125                 cleancache_put_page(page);
126         else
127                 cleancache_flush_page(mapping, page);
128
129         radix_tree_delete(&mapping->page_tree, page->index);
130         page->mapping = NULL;
131         /* Leave page->index set: truncation lookup relies upon it */
132         mapping->nrpages--;
133         __dec_zone_page_state(page, NR_FILE_PAGES);
134         if (PageSwapBacked(page))
135                 __dec_zone_page_state(page, NR_SHMEM);
136         BUG_ON(page_mapped(page));
137
138         /*
139          * Some filesystems seem to re-dirty the page even after
140          * the VM has canceled the dirty bit (eg ext3 journaling).
141          *
142          * Fix it up by doing a final dirty accounting check after
143          * having removed the page entirely.
144          */
145         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
146                 dec_zone_page_state(page, NR_FILE_DIRTY);
147                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
148         }
149 }
150
151 /**
152  * delete_from_page_cache - delete page from page cache
153  * @page: the page which the kernel is trying to remove from page cache
154  *
155  * This must be called only on pages that have been verified to be in the page
156  * cache and locked.  It will never put the page into the free list, the caller
157  * has a reference on the page.
158  */
159 void delete_from_page_cache(struct page *page)
160 {
161         struct address_space *mapping = page->mapping;
162         void (*freepage)(struct page *);
163
164         BUG_ON(!PageLocked(page));
165
166         freepage = mapping->a_ops->freepage;
167         spin_lock_irq(&mapping->tree_lock);
168         __delete_from_page_cache(page);
169         spin_unlock_irq(&mapping->tree_lock);
170         mem_cgroup_uncharge_cache_page(page);
171
172         if (freepage)
173                 freepage(page);
174         page_cache_release(page);
175 }
176 EXPORT_SYMBOL(delete_from_page_cache);
177
178 static int sleep_on_page(void *word)
179 {
180         io_schedule();
181         return 0;
182 }
183
184 static int sleep_on_page_killable(void *word)
185 {
186         sleep_on_page(word);
187         return fatal_signal_pending(current) ? -EINTR : 0;
188 }
189
190 /**
191  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
192  * @mapping:    address space structure to write
193  * @start:      offset in bytes where the range starts
194  * @end:        offset in bytes where the range ends (inclusive)
195  * @sync_mode:  enable synchronous operation
196  *
197  * Start writeback against all of a mapping's dirty pages that lie
198  * within the byte offsets <start, end> inclusive.
199  *
200  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
201  * opposed to a regular memory cleansing writeback.  The difference between
202  * these two operations is that if a dirty page/buffer is encountered, it must
203  * be waited upon, and not just skipped over.
204  */
205 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
206                                 loff_t end, int sync_mode)
207 {
208         int ret;
209         struct writeback_control wbc = {
210                 .sync_mode = sync_mode,
211                 .nr_to_write = LONG_MAX,
212                 .range_start = start,
213                 .range_end = end,
214         };
215
216         if (!mapping_cap_writeback_dirty(mapping))
217                 return 0;
218
219         ret = do_writepages(mapping, &wbc);
220         return ret;
221 }
222
223 static inline int __filemap_fdatawrite(struct address_space *mapping,
224         int sync_mode)
225 {
226         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
227 }
228
229 int filemap_fdatawrite(struct address_space *mapping)
230 {
231         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
232 }
233 EXPORT_SYMBOL(filemap_fdatawrite);
234
235 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
236                                 loff_t end)
237 {
238         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
239 }
240 EXPORT_SYMBOL(filemap_fdatawrite_range);
241
242 /**
243  * filemap_flush - mostly a non-blocking flush
244  * @mapping:    target address_space
245  *
246  * This is a mostly non-blocking flush.  Not suitable for data-integrity
247  * purposes - I/O may not be started against all dirty pages.
248  */
249 int filemap_flush(struct address_space *mapping)
250 {
251         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
252 }
253 EXPORT_SYMBOL(filemap_flush);
254
255 /**
256  * filemap_fdatawait_range - wait for writeback to complete
257  * @mapping:            address space structure to wait for
258  * @start_byte:         offset in bytes where the range starts
259  * @end_byte:           offset in bytes where the range ends (inclusive)
260  *
261  * Walk the list of under-writeback pages of the given address space
262  * in the given range and wait for all of them.
263  */
264 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
265                             loff_t end_byte)
266 {
267         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
268         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272
273         if (end_byte < start_byte)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         while ((index <= end) &&
278                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
279                         PAGECACHE_TAG_WRITEBACK,
280                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
281                 unsigned i;
282
283                 for (i = 0; i < nr_pages; i++) {
284                         struct page *page = pvec.pages[i];
285
286                         /* until radix tree lookup accepts end_index */
287                         if (page->index > end)
288                                 continue;
289
290                         wait_on_page_writeback(page);
291                         if (TestClearPageError(page))
292                                 ret = -EIO;
293                 }
294                 pagevec_release(&pvec);
295                 cond_resched();
296         }
297
298         /* Check for outstanding write errors */
299         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
300                 ret = -ENOSPC;
301         if (test_and_clear_bit(AS_EIO, &mapping->flags))
302                 ret = -EIO;
303
304         return ret;
305 }
306 EXPORT_SYMBOL(filemap_fdatawait_range);
307
308 /**
309  * filemap_fdatawait - wait for all under-writeback pages to complete
310  * @mapping: address space structure to wait for
311  *
312  * Walk the list of under-writeback pages of the given address space
313  * and wait for all of them.
314  */
315 int filemap_fdatawait(struct address_space *mapping)
316 {
317         loff_t i_size = i_size_read(mapping->host);
318
319         if (i_size == 0)
320                 return 0;
321
322         return filemap_fdatawait_range(mapping, 0, i_size - 1);
323 }
324 EXPORT_SYMBOL(filemap_fdatawait);
325
326 int filemap_write_and_wait(struct address_space *mapping)
327 {
328         int err = 0;
329
330         if (mapping->nrpages) {
331                 err = filemap_fdatawrite(mapping);
332                 /*
333                  * Even if the above returned error, the pages may be
334                  * written partially (e.g. -ENOSPC), so we wait for it.
335                  * But the -EIO is special case, it may indicate the worst
336                  * thing (e.g. bug) happened, so we avoid waiting for it.
337                  */
338                 if (err != -EIO) {
339                         int err2 = filemap_fdatawait(mapping);
340                         if (!err)
341                                 err = err2;
342                 }
343         }
344         return err;
345 }
346 EXPORT_SYMBOL(filemap_write_and_wait);
347
348 /**
349  * filemap_write_and_wait_range - write out & wait on a file range
350  * @mapping:    the address_space for the pages
351  * @lstart:     offset in bytes where the range starts
352  * @lend:       offset in bytes where the range ends (inclusive)
353  *
354  * Write out and wait upon file offsets lstart->lend, inclusive.
355  *
356  * Note that `lend' is inclusive (describes the last byte to be written) so
357  * that this function can be used to write to the very end-of-file (end = -1).
358  */
359 int filemap_write_and_wait_range(struct address_space *mapping,
360                                  loff_t lstart, loff_t lend)
361 {
362         int err = 0;
363
364         if (mapping->nrpages) {
365                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
366                                                  WB_SYNC_ALL);
367                 /* See comment of filemap_write_and_wait() */
368                 if (err != -EIO) {
369                         int err2 = filemap_fdatawait_range(mapping,
370                                                 lstart, lend);
371                         if (!err)
372                                 err = err2;
373                 }
374         }
375         return err;
376 }
377 EXPORT_SYMBOL(filemap_write_and_wait_range);
378
379 /**
380  * replace_page_cache_page - replace a pagecache page with a new one
381  * @old:        page to be replaced
382  * @new:        page to replace with
383  * @gfp_mask:   allocation mode
384  *
385  * This function replaces a page in the pagecache with a new one.  On
386  * success it acquires the pagecache reference for the new page and
387  * drops it for the old page.  Both the old and new pages must be
388  * locked.  This function does not add the new page to the LRU, the
389  * caller must do that.
390  *
391  * The remove + add is atomic.  The only way this function can fail is
392  * memory allocation failure.
393  */
394 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
395 {
396         int error;
397         struct mem_cgroup *memcg = NULL;
398
399         VM_BUG_ON(!PageLocked(old));
400         VM_BUG_ON(!PageLocked(new));
401         VM_BUG_ON(new->mapping);
402
403         /*
404          * This is not page migration, but prepare_migration and
405          * end_migration does enough work for charge replacement.
406          *
407          * In the longer term we probably want a specialized function
408          * for moving the charge from old to new in a more efficient
409          * manner.
410          */
411         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
412         if (error)
413                 return error;
414
415         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
416         if (!error) {
417                 struct address_space *mapping = old->mapping;
418                 void (*freepage)(struct page *);
419
420                 pgoff_t offset = old->index;
421                 freepage = mapping->a_ops->freepage;
422
423                 page_cache_get(new);
424                 new->mapping = mapping;
425                 new->index = offset;
426
427                 spin_lock_irq(&mapping->tree_lock);
428                 __delete_from_page_cache(old);
429                 error = radix_tree_insert(&mapping->page_tree, offset, new);
430                 BUG_ON(error);
431                 mapping->nrpages++;
432                 __inc_zone_page_state(new, NR_FILE_PAGES);
433                 if (PageSwapBacked(new))
434                         __inc_zone_page_state(new, NR_SHMEM);
435                 spin_unlock_irq(&mapping->tree_lock);
436                 radix_tree_preload_end();
437                 if (freepage)
438                         freepage(old);
439                 page_cache_release(old);
440                 mem_cgroup_end_migration(memcg, old, new, true);
441         } else {
442                 mem_cgroup_end_migration(memcg, old, new, false);
443         }
444
445         return error;
446 }
447 EXPORT_SYMBOL_GPL(replace_page_cache_page);
448
449 /**
450  * add_to_page_cache_locked - add a locked page to the pagecache
451  * @page:       page to add
452  * @mapping:    the page's address_space
453  * @offset:     page index
454  * @gfp_mask:   page allocation mode
455  *
456  * This function is used to add a page to the pagecache. It must be locked.
457  * This function does not add the page to the LRU.  The caller must do that.
458  */
459 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
460                 pgoff_t offset, gfp_t gfp_mask)
461 {
462         int error;
463
464         VM_BUG_ON(!PageLocked(page));
465
466         error = mem_cgroup_cache_charge(page, current->mm,
467                                         gfp_mask & GFP_RECLAIM_MASK);
468         if (error)
469                 goto out;
470
471         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
472         if (error == 0) {
473                 page_cache_get(page);
474                 page->mapping = mapping;
475                 page->index = offset;
476
477                 spin_lock_irq(&mapping->tree_lock);
478                 error = radix_tree_insert(&mapping->page_tree, offset, page);
479                 if (likely(!error)) {
480                         mapping->nrpages++;
481                         __inc_zone_page_state(page, NR_FILE_PAGES);
482                         if (PageSwapBacked(page))
483                                 __inc_zone_page_state(page, NR_SHMEM);
484                         spin_unlock_irq(&mapping->tree_lock);
485                 } else {
486                         page->mapping = NULL;
487                         /* Leave page->index set: truncation relies upon it */
488                         spin_unlock_irq(&mapping->tree_lock);
489                         mem_cgroup_uncharge_cache_page(page);
490                         page_cache_release(page);
491                 }
492                 radix_tree_preload_end();
493         } else
494                 mem_cgroup_uncharge_cache_page(page);
495 out:
496         return error;
497 }
498 EXPORT_SYMBOL(add_to_page_cache_locked);
499
500 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
501                                 pgoff_t offset, gfp_t gfp_mask)
502 {
503         int ret;
504
505         /*
506          * Splice_read and readahead add shmem/tmpfs pages into the page cache
507          * before shmem_readpage has a chance to mark them as SwapBacked: they
508          * need to go on the anon lru below, and mem_cgroup_cache_charge
509          * (called in add_to_page_cache) needs to know where they're going too.
510          */
511         if (mapping_cap_swap_backed(mapping))
512                 SetPageSwapBacked(page);
513
514         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
515         if (ret == 0) {
516                 if (page_is_file_cache(page))
517                         lru_cache_add_file(page);
518                 else
519                         lru_cache_add_anon(page);
520         }
521         return ret;
522 }
523 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
524
525 #ifdef CONFIG_NUMA
526 struct page *__page_cache_alloc(gfp_t gfp)
527 {
528         int n;
529         struct page *page;
530
531         if (cpuset_do_page_mem_spread()) {
532                 get_mems_allowed();
533                 n = cpuset_mem_spread_node();
534                 page = alloc_pages_exact_node(n, gfp, 0);
535                 put_mems_allowed();
536                 return page;
537         }
538         return alloc_pages(gfp, 0);
539 }
540 EXPORT_SYMBOL(__page_cache_alloc);
541 #endif
542
543 /*
544  * In order to wait for pages to become available there must be
545  * waitqueues associated with pages. By using a hash table of
546  * waitqueues where the bucket discipline is to maintain all
547  * waiters on the same queue and wake all when any of the pages
548  * become available, and for the woken contexts to check to be
549  * sure the appropriate page became available, this saves space
550  * at a cost of "thundering herd" phenomena during rare hash
551  * collisions.
552  */
553 static wait_queue_head_t *page_waitqueue(struct page *page)
554 {
555         const struct zone *zone = page_zone(page);
556
557         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
558 }
559
560 static inline void wake_up_page(struct page *page, int bit)
561 {
562         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
563 }
564
565 void wait_on_page_bit(struct page *page, int bit_nr)
566 {
567         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
568
569         if (test_bit(bit_nr, &page->flags))
570                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
571                                                         TASK_UNINTERRUPTIBLE);
572 }
573 EXPORT_SYMBOL(wait_on_page_bit);
574
575 int wait_on_page_bit_killable(struct page *page, int bit_nr)
576 {
577         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
578
579         if (!test_bit(bit_nr, &page->flags))
580                 return 0;
581
582         return __wait_on_bit(page_waitqueue(page), &wait,
583                              sleep_on_page_killable, TASK_KILLABLE);
584 }
585
586 /**
587  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
588  * @page: Page defining the wait queue of interest
589  * @waiter: Waiter to add to the queue
590  *
591  * Add an arbitrary @waiter to the wait queue for the nominated @page.
592  */
593 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
594 {
595         wait_queue_head_t *q = page_waitqueue(page);
596         unsigned long flags;
597
598         spin_lock_irqsave(&q->lock, flags);
599         __add_wait_queue(q, waiter);
600         spin_unlock_irqrestore(&q->lock, flags);
601 }
602 EXPORT_SYMBOL_GPL(add_page_wait_queue);
603
604 /**
605  * unlock_page - unlock a locked page
606  * @page: the page
607  *
608  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
609  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
610  * mechananism between PageLocked pages and PageWriteback pages is shared.
611  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
612  *
613  * The mb is necessary to enforce ordering between the clear_bit and the read
614  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
615  */
616 void unlock_page(struct page *page)
617 {
618         VM_BUG_ON(!PageLocked(page));
619         clear_bit_unlock(PG_locked, &page->flags);
620         smp_mb__after_clear_bit();
621         wake_up_page(page, PG_locked);
622 }
623 EXPORT_SYMBOL(unlock_page);
624
625 /**
626  * end_page_writeback - end writeback against a page
627  * @page: the page
628  */
629 void end_page_writeback(struct page *page)
630 {
631         if (TestClearPageReclaim(page))
632                 rotate_reclaimable_page(page);
633
634         if (!test_clear_page_writeback(page))
635                 BUG();
636
637         smp_mb__after_clear_bit();
638         wake_up_page(page, PG_writeback);
639 }
640 EXPORT_SYMBOL(end_page_writeback);
641
642 /**
643  * __lock_page - get a lock on the page, assuming we need to sleep to get it
644  * @page: the page to lock
645  */
646 void __lock_page(struct page *page)
647 {
648         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
649
650         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
651                                                         TASK_UNINTERRUPTIBLE);
652 }
653 EXPORT_SYMBOL(__lock_page);
654
655 int __lock_page_killable(struct page *page)
656 {
657         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
658
659         return __wait_on_bit_lock(page_waitqueue(page), &wait,
660                                         sleep_on_page_killable, TASK_KILLABLE);
661 }
662 EXPORT_SYMBOL_GPL(__lock_page_killable);
663
664 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
665                          unsigned int flags)
666 {
667         if (flags & FAULT_FLAG_ALLOW_RETRY) {
668                 /*
669                  * CAUTION! In this case, mmap_sem is not released
670                  * even though return 0.
671                  */
672                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
673                         return 0;
674
675                 up_read(&mm->mmap_sem);
676                 if (flags & FAULT_FLAG_KILLABLE)
677                         wait_on_page_locked_killable(page);
678                 else
679                         wait_on_page_locked(page);
680                 return 0;
681         } else {
682                 if (flags & FAULT_FLAG_KILLABLE) {
683                         int ret;
684
685                         ret = __lock_page_killable(page);
686                         if (ret) {
687                                 up_read(&mm->mmap_sem);
688                                 return 0;
689                         }
690                 } else
691                         __lock_page(page);
692                 return 1;
693         }
694 }
695
696 /**
697  * find_get_page - find and get a page reference
698  * @mapping: the address_space to search
699  * @offset: the page index
700  *
701  * Is there a pagecache struct page at the given (mapping, offset) tuple?
702  * If yes, increment its refcount and return it; if no, return NULL.
703  */
704 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
705 {
706         void **pagep;
707         struct page *page;
708
709         rcu_read_lock();
710 repeat:
711         page = NULL;
712         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
713         if (pagep) {
714                 page = radix_tree_deref_slot(pagep);
715                 if (unlikely(!page))
716                         goto out;
717                 if (radix_tree_deref_retry(page))
718                         goto repeat;
719
720                 if (!page_cache_get_speculative(page))
721                         goto repeat;
722
723                 /*
724                  * Has the page moved?
725                  * This is part of the lockless pagecache protocol. See
726                  * include/linux/pagemap.h for details.
727                  */
728                 if (unlikely(page != *pagep)) {
729                         page_cache_release(page);
730                         goto repeat;
731                 }
732         }
733 out:
734         rcu_read_unlock();
735
736         return page;
737 }
738 EXPORT_SYMBOL(find_get_page);
739
740 /**
741  * find_lock_page - locate, pin and lock a pagecache page
742  * @mapping: the address_space to search
743  * @offset: the page index
744  *
745  * Locates the desired pagecache page, locks it, increments its reference
746  * count and returns its address.
747  *
748  * Returns zero if the page was not present. find_lock_page() may sleep.
749  */
750 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
751 {
752         struct page *page;
753
754 repeat:
755         page = find_get_page(mapping, offset);
756         if (page) {
757                 lock_page(page);
758                 /* Has the page been truncated? */
759                 if (unlikely(page->mapping != mapping)) {
760                         unlock_page(page);
761                         page_cache_release(page);
762                         goto repeat;
763                 }
764                 VM_BUG_ON(page->index != offset);
765         }
766         return page;
767 }
768 EXPORT_SYMBOL(find_lock_page);
769
770 /**
771  * find_or_create_page - locate or add a pagecache page
772  * @mapping: the page's address_space
773  * @index: the page's index into the mapping
774  * @gfp_mask: page allocation mode
775  *
776  * Locates a page in the pagecache.  If the page is not present, a new page
777  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
778  * LRU list.  The returned page is locked and has its reference count
779  * incremented.
780  *
781  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
782  * allocation!
783  *
784  * find_or_create_page() returns the desired page's address, or zero on
785  * memory exhaustion.
786  */
787 struct page *find_or_create_page(struct address_space *mapping,
788                 pgoff_t index, gfp_t gfp_mask)
789 {
790         struct page *page;
791         int err;
792 repeat:
793         page = find_lock_page(mapping, index);
794         if (!page) {
795                 page = __page_cache_alloc(gfp_mask);
796                 if (!page)
797                         return NULL;
798                 /*
799                  * We want a regular kernel memory (not highmem or DMA etc)
800                  * allocation for the radix tree nodes, but we need to honour
801                  * the context-specific requirements the caller has asked for.
802                  * GFP_RECLAIM_MASK collects those requirements.
803                  */
804                 err = add_to_page_cache_lru(page, mapping, index,
805                         (gfp_mask & GFP_RECLAIM_MASK));
806                 if (unlikely(err)) {
807                         page_cache_release(page);
808                         page = NULL;
809                         if (err == -EEXIST)
810                                 goto repeat;
811                 }
812         }
813         return page;
814 }
815 EXPORT_SYMBOL(find_or_create_page);
816
817 /**
818  * find_get_pages - gang pagecache lookup
819  * @mapping:    The address_space to search
820  * @start:      The starting page index
821  * @nr_pages:   The maximum number of pages
822  * @pages:      Where the resulting pages are placed
823  *
824  * find_get_pages() will search for and return a group of up to
825  * @nr_pages pages in the mapping.  The pages are placed at @pages.
826  * find_get_pages() takes a reference against the returned pages.
827  *
828  * The search returns a group of mapping-contiguous pages with ascending
829  * indexes.  There may be holes in the indices due to not-present pages.
830  *
831  * find_get_pages() returns the number of pages which were found.
832  */
833 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
834                             unsigned int nr_pages, struct page **pages)
835 {
836         unsigned int i;
837         unsigned int ret;
838         unsigned int nr_found;
839
840         rcu_read_lock();
841 restart:
842         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
843                                 (void ***)pages, start, nr_pages);
844         ret = 0;
845         for (i = 0; i < nr_found; i++) {
846                 struct page *page;
847 repeat:
848                 page = radix_tree_deref_slot((void **)pages[i]);
849                 if (unlikely(!page))
850                         continue;
851
852                 /*
853                  * This can only trigger when the entry at index 0 moves out
854                  * of or back to the root: none yet gotten, safe to restart.
855                  */
856                 if (radix_tree_deref_retry(page)) {
857                         WARN_ON(start | i);
858                         goto restart;
859                 }
860
861                 if (!page_cache_get_speculative(page))
862                         goto repeat;
863
864                 /* Has the page moved? */
865                 if (unlikely(page != *((void **)pages[i]))) {
866                         page_cache_release(page);
867                         goto repeat;
868                 }
869
870                 pages[ret] = page;
871                 ret++;
872         }
873
874         /*
875          * If all entries were removed before we could secure them,
876          * try again, because callers stop trying once 0 is returned.
877          */
878         if (unlikely(!ret && nr_found))
879                 goto restart;
880         rcu_read_unlock();
881         return ret;
882 }
883
884 /**
885  * find_get_pages_contig - gang contiguous pagecache lookup
886  * @mapping:    The address_space to search
887  * @index:      The starting page index
888  * @nr_pages:   The maximum number of pages
889  * @pages:      Where the resulting pages are placed
890  *
891  * find_get_pages_contig() works exactly like find_get_pages(), except
892  * that the returned number of pages are guaranteed to be contiguous.
893  *
894  * find_get_pages_contig() returns the number of pages which were found.
895  */
896 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
897                                unsigned int nr_pages, struct page **pages)
898 {
899         unsigned int i;
900         unsigned int ret;
901         unsigned int nr_found;
902
903         rcu_read_lock();
904 restart:
905         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
906                                 (void ***)pages, index, nr_pages);
907         ret = 0;
908         for (i = 0; i < nr_found; i++) {
909                 struct page *page;
910 repeat:
911                 page = radix_tree_deref_slot((void **)pages[i]);
912                 if (unlikely(!page))
913                         continue;
914
915                 /*
916                  * This can only trigger when the entry at index 0 moves out
917                  * of or back to the root: none yet gotten, safe to restart.
918                  */
919                 if (radix_tree_deref_retry(page))
920                         goto restart;
921
922                 if (!page_cache_get_speculative(page))
923                         goto repeat;
924
925                 /* Has the page moved? */
926                 if (unlikely(page != *((void **)pages[i]))) {
927                         page_cache_release(page);
928                         goto repeat;
929                 }
930
931                 /*
932                  * must check mapping and index after taking the ref.
933                  * otherwise we can get both false positives and false
934                  * negatives, which is just confusing to the caller.
935                  */
936                 if (page->mapping == NULL || page->index != index) {
937                         page_cache_release(page);
938                         break;
939                 }
940
941                 pages[ret] = page;
942                 ret++;
943                 index++;
944         }
945         rcu_read_unlock();
946         return ret;
947 }
948 EXPORT_SYMBOL(find_get_pages_contig);
949
950 /**
951  * find_get_pages_tag - find and return pages that match @tag
952  * @mapping:    the address_space to search
953  * @index:      the starting page index
954  * @tag:        the tag index
955  * @nr_pages:   the maximum number of pages
956  * @pages:      where the resulting pages are placed
957  *
958  * Like find_get_pages, except we only return pages which are tagged with
959  * @tag.   We update @index to index the next page for the traversal.
960  */
961 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
962                         int tag, unsigned int nr_pages, struct page **pages)
963 {
964         unsigned int i;
965         unsigned int ret;
966         unsigned int nr_found;
967
968         rcu_read_lock();
969 restart:
970         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
971                                 (void ***)pages, *index, nr_pages, tag);
972         ret = 0;
973         for (i = 0; i < nr_found; i++) {
974                 struct page *page;
975 repeat:
976                 page = radix_tree_deref_slot((void **)pages[i]);
977                 if (unlikely(!page))
978                         continue;
979
980                 /*
981                  * This can only trigger when the entry at index 0 moves out
982                  * of or back to the root: none yet gotten, safe to restart.
983                  */
984                 if (radix_tree_deref_retry(page))
985                         goto restart;
986
987                 if (!page_cache_get_speculative(page))
988                         goto repeat;
989
990                 /* Has the page moved? */
991                 if (unlikely(page != *((void **)pages[i]))) {
992                         page_cache_release(page);
993                         goto repeat;
994                 }
995
996                 pages[ret] = page;
997                 ret++;
998         }
999
1000         /*
1001          * If all entries were removed before we could secure them,
1002          * try again, because callers stop trying once 0 is returned.
1003          */
1004         if (unlikely(!ret && nr_found))
1005                 goto restart;
1006         rcu_read_unlock();
1007
1008         if (ret)
1009                 *index = pages[ret - 1]->index + 1;
1010
1011         return ret;
1012 }
1013 EXPORT_SYMBOL(find_get_pages_tag);
1014
1015 /**
1016  * grab_cache_page_nowait - returns locked page at given index in given cache
1017  * @mapping: target address_space
1018  * @index: the page index
1019  *
1020  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1021  * This is intended for speculative data generators, where the data can
1022  * be regenerated if the page couldn't be grabbed.  This routine should
1023  * be safe to call while holding the lock for another page.
1024  *
1025  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1026  * and deadlock against the caller's locked page.
1027  */
1028 struct page *
1029 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1030 {
1031         struct page *page = find_get_page(mapping, index);
1032
1033         if (page) {
1034                 if (trylock_page(page))
1035                         return page;
1036                 page_cache_release(page);
1037                 return NULL;
1038         }
1039         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1040         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1041                 page_cache_release(page);
1042                 page = NULL;
1043         }
1044         return page;
1045 }
1046 EXPORT_SYMBOL(grab_cache_page_nowait);
1047
1048 /*
1049  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1050  * a _large_ part of the i/o request. Imagine the worst scenario:
1051  *
1052  *      ---R__________________________________________B__________
1053  *         ^ reading here                             ^ bad block(assume 4k)
1054  *
1055  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1056  * => failing the whole request => read(R) => read(R+1) =>
1057  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1058  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1059  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1060  *
1061  * It is going insane. Fix it by quickly scaling down the readahead size.
1062  */
1063 static void shrink_readahead_size_eio(struct file *filp,
1064                                         struct file_ra_state *ra)
1065 {
1066         ra->ra_pages /= 4;
1067 }
1068
1069 /**
1070  * do_generic_file_read - generic file read routine
1071  * @filp:       the file to read
1072  * @ppos:       current file position
1073  * @desc:       read_descriptor
1074  * @actor:      read method
1075  *
1076  * This is a generic file read routine, and uses the
1077  * mapping->a_ops->readpage() function for the actual low-level stuff.
1078  *
1079  * This is really ugly. But the goto's actually try to clarify some
1080  * of the logic when it comes to error handling etc.
1081  */
1082 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1083                 read_descriptor_t *desc, read_actor_t actor)
1084 {
1085         struct address_space *mapping = filp->f_mapping;
1086         struct inode *inode = mapping->host;
1087         struct file_ra_state *ra = &filp->f_ra;
1088         pgoff_t index;
1089         pgoff_t last_index;
1090         pgoff_t prev_index;
1091         unsigned long offset;      /* offset into pagecache page */
1092         unsigned int prev_offset;
1093         int error;
1094
1095         index = *ppos >> PAGE_CACHE_SHIFT;
1096         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1097         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1098         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1099         offset = *ppos & ~PAGE_CACHE_MASK;
1100
1101         for (;;) {
1102                 struct page *page;
1103                 pgoff_t end_index;
1104                 loff_t isize;
1105                 unsigned long nr, ret;
1106
1107                 cond_resched();
1108 find_page:
1109                 page = find_get_page(mapping, index);
1110                 if (!page) {
1111                         page_cache_sync_readahead(mapping,
1112                                         ra, filp,
1113                                         index, last_index - index);
1114                         page = find_get_page(mapping, index);
1115                         if (unlikely(page == NULL))
1116                                 goto no_cached_page;
1117                 }
1118                 if (PageReadahead(page)) {
1119                         page_cache_async_readahead(mapping,
1120                                         ra, filp, page,
1121                                         index, last_index - index);
1122                 }
1123                 if (!PageUptodate(page)) {
1124                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1125                                         !mapping->a_ops->is_partially_uptodate)
1126                                 goto page_not_up_to_date;
1127                         if (!trylock_page(page))
1128                                 goto page_not_up_to_date;
1129                         /* Did it get truncated before we got the lock? */
1130                         if (!page->mapping)
1131                                 goto page_not_up_to_date_locked;
1132                         if (!mapping->a_ops->is_partially_uptodate(page,
1133                                                                 desc, offset))
1134                                 goto page_not_up_to_date_locked;
1135                         unlock_page(page);
1136                 }
1137 page_ok:
1138                 /*
1139                  * i_size must be checked after we know the page is Uptodate.
1140                  *
1141                  * Checking i_size after the check allows us to calculate
1142                  * the correct value for "nr", which means the zero-filled
1143                  * part of the page is not copied back to userspace (unless
1144                  * another truncate extends the file - this is desired though).
1145                  */
1146
1147                 isize = i_size_read(inode);
1148                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1149                 if (unlikely(!isize || index > end_index)) {
1150                         page_cache_release(page);
1151                         goto out;
1152                 }
1153
1154                 /* nr is the maximum number of bytes to copy from this page */
1155                 nr = PAGE_CACHE_SIZE;
1156                 if (index == end_index) {
1157                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1158                         if (nr <= offset) {
1159                                 page_cache_release(page);
1160                                 goto out;
1161                         }
1162                 }
1163                 nr = nr - offset;
1164
1165                 /* If users can be writing to this page using arbitrary
1166                  * virtual addresses, take care about potential aliasing
1167                  * before reading the page on the kernel side.
1168                  */
1169                 if (mapping_writably_mapped(mapping))
1170                         flush_dcache_page(page);
1171
1172                 /*
1173                  * When a sequential read accesses a page several times,
1174                  * only mark it as accessed the first time.
1175                  */
1176                 if (prev_index != index || offset != prev_offset)
1177                         mark_page_accessed(page);
1178                 prev_index = index;
1179
1180                 /*
1181                  * Ok, we have the page, and it's up-to-date, so
1182                  * now we can copy it to user space...
1183                  *
1184                  * The actor routine returns how many bytes were actually used..
1185                  * NOTE! This may not be the same as how much of a user buffer
1186                  * we filled up (we may be padding etc), so we can only update
1187                  * "pos" here (the actor routine has to update the user buffer
1188                  * pointers and the remaining count).
1189                  */
1190                 ret = actor(desc, page, offset, nr);
1191                 offset += ret;
1192                 index += offset >> PAGE_CACHE_SHIFT;
1193                 offset &= ~PAGE_CACHE_MASK;
1194                 prev_offset = offset;
1195
1196                 page_cache_release(page);
1197                 if (ret == nr && desc->count)
1198                         continue;
1199                 goto out;
1200
1201 page_not_up_to_date:
1202                 /* Get exclusive access to the page ... */
1203                 error = lock_page_killable(page);
1204                 if (unlikely(error))
1205                         goto readpage_error;
1206
1207 page_not_up_to_date_locked:
1208                 /* Did it get truncated before we got the lock? */
1209                 if (!page->mapping) {
1210                         unlock_page(page);
1211                         page_cache_release(page);
1212                         continue;
1213                 }
1214
1215                 /* Did somebody else fill it already? */
1216                 if (PageUptodate(page)) {
1217                         unlock_page(page);
1218                         goto page_ok;
1219                 }
1220
1221 readpage:
1222                 /*
1223                  * A previous I/O error may have been due to temporary
1224                  * failures, eg. multipath errors.
1225                  * PG_error will be set again if readpage fails.
1226                  */
1227                 ClearPageError(page);
1228                 /* Start the actual read. The read will unlock the page. */
1229                 error = mapping->a_ops->readpage(filp, page);
1230
1231                 if (unlikely(error)) {
1232                         if (error == AOP_TRUNCATED_PAGE) {
1233                                 page_cache_release(page);
1234                                 goto find_page;
1235                         }
1236                         goto readpage_error;
1237                 }
1238
1239                 if (!PageUptodate(page)) {
1240                         error = lock_page_killable(page);
1241                         if (unlikely(error))
1242                                 goto readpage_error;
1243                         if (!PageUptodate(page)) {
1244                                 if (page->mapping == NULL) {
1245                                         /*
1246                                          * invalidate_mapping_pages got it
1247                                          */
1248                                         unlock_page(page);
1249                                         page_cache_release(page);
1250                                         goto find_page;
1251                                 }
1252                                 unlock_page(page);
1253                                 shrink_readahead_size_eio(filp, ra);
1254                                 error = -EIO;
1255                                 goto readpage_error;
1256                         }
1257                         unlock_page(page);
1258                 }
1259
1260                 goto page_ok;
1261
1262 readpage_error:
1263                 /* UHHUH! A synchronous read error occurred. Report it */
1264                 desc->error = error;
1265                 page_cache_release(page);
1266                 goto out;
1267
1268 no_cached_page:
1269                 /*
1270                  * Ok, it wasn't cached, so we need to create a new
1271                  * page..
1272                  */
1273                 page = page_cache_alloc_cold(mapping);
1274                 if (!page) {
1275                         desc->error = -ENOMEM;
1276                         goto out;
1277                 }
1278                 error = add_to_page_cache_lru(page, mapping,
1279                                                 index, GFP_KERNEL);
1280                 if (error) {
1281                         page_cache_release(page);
1282                         if (error == -EEXIST)
1283                                 goto find_page;
1284                         desc->error = error;
1285                         goto out;
1286                 }
1287                 goto readpage;
1288         }
1289
1290 out:
1291         ra->prev_pos = prev_index;
1292         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1293         ra->prev_pos |= prev_offset;
1294
1295         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1296         file_accessed(filp);
1297 }
1298
1299 int file_read_actor(read_descriptor_t *desc, struct page *page,
1300                         unsigned long offset, unsigned long size)
1301 {
1302         char *kaddr;
1303         unsigned long left, count = desc->count;
1304
1305         if (size > count)
1306                 size = count;
1307
1308         /*
1309          * Faults on the destination of a read are common, so do it before
1310          * taking the kmap.
1311          */
1312         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1313                 kaddr = kmap_atomic(page, KM_USER0);
1314                 left = __copy_to_user_inatomic(desc->arg.buf,
1315                                                 kaddr + offset, size);
1316                 kunmap_atomic(kaddr, KM_USER0);
1317                 if (left == 0)
1318                         goto success;
1319         }
1320
1321         /* Do it the slow way */
1322         kaddr = kmap(page);
1323         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1324         kunmap(page);
1325
1326         if (left) {
1327                 size -= left;
1328                 desc->error = -EFAULT;
1329         }
1330 success:
1331         desc->count = count - size;
1332         desc->written += size;
1333         desc->arg.buf += size;
1334         return size;
1335 }
1336
1337 /*
1338  * Performs necessary checks before doing a write
1339  * @iov:        io vector request
1340  * @nr_segs:    number of segments in the iovec
1341  * @count:      number of bytes to write
1342  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1343  *
1344  * Adjust number of segments and amount of bytes to write (nr_segs should be
1345  * properly initialized first). Returns appropriate error code that caller
1346  * should return or zero in case that write should be allowed.
1347  */
1348 int generic_segment_checks(const struct iovec *iov,
1349                         unsigned long *nr_segs, size_t *count, int access_flags)
1350 {
1351         unsigned long   seg;
1352         size_t cnt = 0;
1353         for (seg = 0; seg < *nr_segs; seg++) {
1354                 const struct iovec *iv = &iov[seg];
1355
1356                 /*
1357                  * If any segment has a negative length, or the cumulative
1358                  * length ever wraps negative then return -EINVAL.
1359                  */
1360                 cnt += iv->iov_len;
1361                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1362                         return -EINVAL;
1363                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1364                         continue;
1365                 if (seg == 0)
1366                         return -EFAULT;
1367                 *nr_segs = seg;
1368                 cnt -= iv->iov_len;     /* This segment is no good */
1369                 break;
1370         }
1371         *count = cnt;
1372         return 0;
1373 }
1374 EXPORT_SYMBOL(generic_segment_checks);
1375
1376 /**
1377  * generic_file_aio_read - generic filesystem read routine
1378  * @iocb:       kernel I/O control block
1379  * @iov:        io vector request
1380  * @nr_segs:    number of segments in the iovec
1381  * @pos:        current file position
1382  *
1383  * This is the "read()" routine for all filesystems
1384  * that can use the page cache directly.
1385  */
1386 ssize_t
1387 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1388                 unsigned long nr_segs, loff_t pos)
1389 {
1390         struct file *filp = iocb->ki_filp;
1391         ssize_t retval;
1392         unsigned long seg = 0;
1393         size_t count;
1394         loff_t *ppos = &iocb->ki_pos;
1395         struct blk_plug plug;
1396
1397         count = 0;
1398         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1399         if (retval)
1400                 return retval;
1401
1402         blk_start_plug(&plug);
1403
1404         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1405         if (filp->f_flags & O_DIRECT) {
1406                 loff_t size;
1407                 struct address_space *mapping;
1408                 struct inode *inode;
1409
1410                 mapping = filp->f_mapping;
1411                 inode = mapping->host;
1412                 if (!count)
1413                         goto out; /* skip atime */
1414                 size = i_size_read(inode);
1415                 if (pos < size) {
1416                         retval = filemap_write_and_wait_range(mapping, pos,
1417                                         pos + iov_length(iov, nr_segs) - 1);
1418                         if (!retval) {
1419                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1420                                                         iov, pos, nr_segs);
1421                         }
1422                         if (retval > 0) {
1423                                 *ppos = pos + retval;
1424                                 count -= retval;
1425                         }
1426
1427                         /*
1428                          * Btrfs can have a short DIO read if we encounter
1429                          * compressed extents, so if there was an error, or if
1430                          * we've already read everything we wanted to, or if
1431                          * there was a short read because we hit EOF, go ahead
1432                          * and return.  Otherwise fallthrough to buffered io for
1433                          * the rest of the read.
1434                          */
1435                         if (retval < 0 || !count || *ppos >= size) {
1436                                 file_accessed(filp);
1437                                 goto out;
1438                         }
1439                 }
1440         }
1441
1442         count = retval;
1443         for (seg = 0; seg < nr_segs; seg++) {
1444                 read_descriptor_t desc;
1445                 loff_t offset = 0;
1446
1447                 /*
1448                  * If we did a short DIO read we need to skip the section of the
1449                  * iov that we've already read data into.
1450                  */
1451                 if (count) {
1452                         if (count > iov[seg].iov_len) {
1453                                 count -= iov[seg].iov_len;
1454                                 continue;
1455                         }
1456                         offset = count;
1457                         count = 0;
1458                 }
1459
1460                 desc.written = 0;
1461                 desc.arg.buf = iov[seg].iov_base + offset;
1462                 desc.count = iov[seg].iov_len - offset;
1463                 if (desc.count == 0)
1464                         continue;
1465                 desc.error = 0;
1466                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1467                 retval += desc.written;
1468                 if (desc.error) {
1469                         retval = retval ?: desc.error;
1470                         break;
1471                 }
1472                 if (desc.count > 0)
1473                         break;
1474         }
1475 out:
1476         blk_finish_plug(&plug);
1477         return retval;
1478 }
1479 EXPORT_SYMBOL(generic_file_aio_read);
1480
1481 static ssize_t
1482 do_readahead(struct address_space *mapping, struct file *filp,
1483              pgoff_t index, unsigned long nr)
1484 {
1485         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1486                 return -EINVAL;
1487
1488         force_page_cache_readahead(mapping, filp, index, nr);
1489         return 0;
1490 }
1491
1492 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1493 {
1494         ssize_t ret;
1495         struct file *file;
1496
1497         ret = -EBADF;
1498         file = fget(fd);
1499         if (file) {
1500                 if (file->f_mode & FMODE_READ) {
1501                         struct address_space *mapping = file->f_mapping;
1502                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1503                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1504                         unsigned long len = end - start + 1;
1505                         ret = do_readahead(mapping, file, start, len);
1506                 }
1507                 fput(file);
1508         }
1509         return ret;
1510 }
1511 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1512 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1513 {
1514         return SYSC_readahead((int) fd, offset, (size_t) count);
1515 }
1516 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1517 #endif
1518
1519 #ifdef CONFIG_MMU
1520 /**
1521  * page_cache_read - adds requested page to the page cache if not already there
1522  * @file:       file to read
1523  * @offset:     page index
1524  *
1525  * This adds the requested page to the page cache if it isn't already there,
1526  * and schedules an I/O to read in its contents from disk.
1527  */
1528 static int page_cache_read(struct file *file, pgoff_t offset)
1529 {
1530         struct address_space *mapping = file->f_mapping;
1531         struct page *page; 
1532         int ret;
1533
1534         do {
1535                 page = page_cache_alloc_cold(mapping);
1536                 if (!page)
1537                         return -ENOMEM;
1538
1539                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1540                 if (ret == 0)
1541                         ret = mapping->a_ops->readpage(file, page);
1542                 else if (ret == -EEXIST)
1543                         ret = 0; /* losing race to add is OK */
1544
1545                 page_cache_release(page);
1546
1547         } while (ret == AOP_TRUNCATED_PAGE);
1548                 
1549         return ret;
1550 }
1551
1552 #define MMAP_LOTSAMISS  (100)
1553
1554 /*
1555  * Synchronous readahead happens when we don't even find
1556  * a page in the page cache at all.
1557  */
1558 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1559                                    struct file_ra_state *ra,
1560                                    struct file *file,
1561                                    pgoff_t offset)
1562 {
1563         unsigned long ra_pages;
1564         struct address_space *mapping = file->f_mapping;
1565
1566         /* If we don't want any read-ahead, don't bother */
1567         if (VM_RandomReadHint(vma))
1568                 return;
1569         if (!ra->ra_pages)
1570                 return;
1571
1572         if (VM_SequentialReadHint(vma)) {
1573                 page_cache_sync_readahead(mapping, ra, file, offset,
1574                                           ra->ra_pages);
1575                 return;
1576         }
1577
1578         /* Avoid banging the cache line if not needed */
1579         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1580                 ra->mmap_miss++;
1581
1582         /*
1583          * Do we miss much more than hit in this file? If so,
1584          * stop bothering with read-ahead. It will only hurt.
1585          */
1586         if (ra->mmap_miss > MMAP_LOTSAMISS)
1587                 return;
1588
1589         /*
1590          * mmap read-around
1591          */
1592         ra_pages = max_sane_readahead(ra->ra_pages);
1593         ra->start = max_t(long, 0, offset - ra_pages / 2);
1594         ra->size = ra_pages;
1595         ra->async_size = ra_pages / 4;
1596         ra_submit(ra, mapping, file);
1597 }
1598
1599 /*
1600  * Asynchronous readahead happens when we find the page and PG_readahead,
1601  * so we want to possibly extend the readahead further..
1602  */
1603 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1604                                     struct file_ra_state *ra,
1605                                     struct file *file,
1606                                     struct page *page,
1607                                     pgoff_t offset)
1608 {
1609         struct address_space *mapping = file->f_mapping;
1610
1611         /* If we don't want any read-ahead, don't bother */
1612         if (VM_RandomReadHint(vma))
1613                 return;
1614         if (ra->mmap_miss > 0)
1615                 ra->mmap_miss--;
1616         if (PageReadahead(page))
1617                 page_cache_async_readahead(mapping, ra, file,
1618                                            page, offset, ra->ra_pages);
1619 }
1620
1621 /**
1622  * filemap_fault - read in file data for page fault handling
1623  * @vma:        vma in which the fault was taken
1624  * @vmf:        struct vm_fault containing details of the fault
1625  *
1626  * filemap_fault() is invoked via the vma operations vector for a
1627  * mapped memory region to read in file data during a page fault.
1628  *
1629  * The goto's are kind of ugly, but this streamlines the normal case of having
1630  * it in the page cache, and handles the special cases reasonably without
1631  * having a lot of duplicated code.
1632  */
1633 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1634 {
1635         int error;
1636         struct file *file = vma->vm_file;
1637         struct address_space *mapping = file->f_mapping;
1638         struct file_ra_state *ra = &file->f_ra;
1639         struct inode *inode = mapping->host;
1640         pgoff_t offset = vmf->pgoff;
1641         struct page *page;
1642         pgoff_t size;
1643         int ret = 0;
1644
1645         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1646         if (offset >= size)
1647                 return VM_FAULT_SIGBUS;
1648
1649         /*
1650          * Do we have something in the page cache already?
1651          */
1652         page = find_get_page(mapping, offset);
1653         if (likely(page)) {
1654                 /*
1655                  * We found the page, so try async readahead before
1656                  * waiting for the lock.
1657                  */
1658                 do_async_mmap_readahead(vma, ra, file, page, offset);
1659         } else {
1660                 /* No page in the page cache at all */
1661                 do_sync_mmap_readahead(vma, ra, file, offset);
1662                 count_vm_event(PGMAJFAULT);
1663                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1664                 ret = VM_FAULT_MAJOR;
1665 retry_find:
1666                 page = find_get_page(mapping, offset);
1667                 if (!page)
1668                         goto no_cached_page;
1669         }
1670
1671         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1672                 page_cache_release(page);
1673                 return ret | VM_FAULT_RETRY;
1674         }
1675
1676         /* Did it get truncated? */
1677         if (unlikely(page->mapping != mapping)) {
1678                 unlock_page(page);
1679                 put_page(page);
1680                 goto retry_find;
1681         }
1682         VM_BUG_ON(page->index != offset);
1683
1684         /*
1685          * We have a locked page in the page cache, now we need to check
1686          * that it's up-to-date. If not, it is going to be due to an error.
1687          */
1688         if (unlikely(!PageUptodate(page)))
1689                 goto page_not_uptodate;
1690
1691         /*
1692          * Found the page and have a reference on it.
1693          * We must recheck i_size under page lock.
1694          */
1695         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1696         if (unlikely(offset >= size)) {
1697                 unlock_page(page);
1698                 page_cache_release(page);
1699                 return VM_FAULT_SIGBUS;
1700         }
1701
1702         vmf->page = page;
1703         return ret | VM_FAULT_LOCKED;
1704
1705 no_cached_page:
1706         /*
1707          * We're only likely to ever get here if MADV_RANDOM is in
1708          * effect.
1709          */
1710         error = page_cache_read(file, offset);
1711
1712         /*
1713          * The page we want has now been added to the page cache.
1714          * In the unlikely event that someone removed it in the
1715          * meantime, we'll just come back here and read it again.
1716          */
1717         if (error >= 0)
1718                 goto retry_find;
1719
1720         /*
1721          * An error return from page_cache_read can result if the
1722          * system is low on memory, or a problem occurs while trying
1723          * to schedule I/O.
1724          */
1725         if (error == -ENOMEM)
1726                 return VM_FAULT_OOM;
1727         return VM_FAULT_SIGBUS;
1728
1729 page_not_uptodate:
1730         /*
1731          * Umm, take care of errors if the page isn't up-to-date.
1732          * Try to re-read it _once_. We do this synchronously,
1733          * because there really aren't any performance issues here
1734          * and we need to check for errors.
1735          */
1736         ClearPageError(page);
1737         error = mapping->a_ops->readpage(file, page);
1738         if (!error) {
1739                 wait_on_page_locked(page);
1740                 if (!PageUptodate(page))
1741                         error = -EIO;
1742         }
1743         page_cache_release(page);
1744
1745         if (!error || error == AOP_TRUNCATED_PAGE)
1746                 goto retry_find;
1747
1748         /* Things didn't work out. Return zero to tell the mm layer so. */
1749         shrink_readahead_size_eio(file, ra);
1750         return VM_FAULT_SIGBUS;
1751 }
1752 EXPORT_SYMBOL(filemap_fault);
1753
1754 const struct vm_operations_struct generic_file_vm_ops = {
1755         .fault          = filemap_fault,
1756 };
1757
1758 /* This is used for a general mmap of a disk file */
1759
1760 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1761 {
1762         struct address_space *mapping = file->f_mapping;
1763
1764         if (!mapping->a_ops->readpage)
1765                 return -ENOEXEC;
1766         file_accessed(file);
1767         vma->vm_ops = &generic_file_vm_ops;
1768         vma->vm_flags |= VM_CAN_NONLINEAR;
1769         return 0;
1770 }
1771
1772 /*
1773  * This is for filesystems which do not implement ->writepage.
1774  */
1775 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1776 {
1777         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1778                 return -EINVAL;
1779         return generic_file_mmap(file, vma);
1780 }
1781 #else
1782 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1783 {
1784         return -ENOSYS;
1785 }
1786 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1787 {
1788         return -ENOSYS;
1789 }
1790 #endif /* CONFIG_MMU */
1791
1792 EXPORT_SYMBOL(generic_file_mmap);
1793 EXPORT_SYMBOL(generic_file_readonly_mmap);
1794
1795 static struct page *__read_cache_page(struct address_space *mapping,
1796                                 pgoff_t index,
1797                                 int (*filler)(void *, struct page *),
1798                                 void *data,
1799                                 gfp_t gfp)
1800 {
1801         struct page *page;
1802         int err;
1803 repeat:
1804         page = find_get_page(mapping, index);
1805         if (!page) {
1806                 page = __page_cache_alloc(gfp | __GFP_COLD);
1807                 if (!page)
1808                         return ERR_PTR(-ENOMEM);
1809                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1810                 if (unlikely(err)) {
1811                         page_cache_release(page);
1812                         if (err == -EEXIST)
1813                                 goto repeat;
1814                         /* Presumably ENOMEM for radix tree node */
1815                         return ERR_PTR(err);
1816                 }
1817                 err = filler(data, page);
1818                 if (err < 0) {
1819                         page_cache_release(page);
1820                         page = ERR_PTR(err);
1821                 }
1822         }
1823         return page;
1824 }
1825
1826 static struct page *do_read_cache_page(struct address_space *mapping,
1827                                 pgoff_t index,
1828                                 int (*filler)(void *, struct page *),
1829                                 void *data,
1830                                 gfp_t gfp)
1831
1832 {
1833         struct page *page;
1834         int err;
1835
1836 retry:
1837         page = __read_cache_page(mapping, index, filler, data, gfp);
1838         if (IS_ERR(page))
1839                 return page;
1840         if (PageUptodate(page))
1841                 goto out;
1842
1843         lock_page(page);
1844         if (!page->mapping) {
1845                 unlock_page(page);
1846                 page_cache_release(page);
1847                 goto retry;
1848         }
1849         if (PageUptodate(page)) {
1850                 unlock_page(page);
1851                 goto out;
1852         }
1853         err = filler(data, page);
1854         if (err < 0) {
1855                 page_cache_release(page);
1856                 return ERR_PTR(err);
1857         }
1858 out:
1859         mark_page_accessed(page);
1860         return page;
1861 }
1862
1863 /**
1864  * read_cache_page_async - read into page cache, fill it if needed
1865  * @mapping:    the page's address_space
1866  * @index:      the page index
1867  * @filler:     function to perform the read
1868  * @data:       first arg to filler(data, page) function, often left as NULL
1869  *
1870  * Same as read_cache_page, but don't wait for page to become unlocked
1871  * after submitting it to the filler.
1872  *
1873  * Read into the page cache. If a page already exists, and PageUptodate() is
1874  * not set, try to fill the page but don't wait for it to become unlocked.
1875  *
1876  * If the page does not get brought uptodate, return -EIO.
1877  */
1878 struct page *read_cache_page_async(struct address_space *mapping,
1879                                 pgoff_t index,
1880                                 int (*filler)(void *, struct page *),
1881                                 void *data)
1882 {
1883         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1884 }
1885 EXPORT_SYMBOL(read_cache_page_async);
1886
1887 static struct page *wait_on_page_read(struct page *page)
1888 {
1889         if (!IS_ERR(page)) {
1890                 wait_on_page_locked(page);
1891                 if (!PageUptodate(page)) {
1892                         page_cache_release(page);
1893                         page = ERR_PTR(-EIO);
1894                 }
1895         }
1896         return page;
1897 }
1898
1899 /**
1900  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1901  * @mapping:    the page's address_space
1902  * @index:      the page index
1903  * @gfp:        the page allocator flags to use if allocating
1904  *
1905  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1906  * any new page allocations done using the specified allocation flags. Note
1907  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1908  * expect to do this atomically or anything like that - but you can pass in
1909  * other page requirements.
1910  *
1911  * If the page does not get brought uptodate, return -EIO.
1912  */
1913 struct page *read_cache_page_gfp(struct address_space *mapping,
1914                                 pgoff_t index,
1915                                 gfp_t gfp)
1916 {
1917         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1918
1919         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1920 }
1921 EXPORT_SYMBOL(read_cache_page_gfp);
1922
1923 /**
1924  * read_cache_page - read into page cache, fill it if needed
1925  * @mapping:    the page's address_space
1926  * @index:      the page index
1927  * @filler:     function to perform the read
1928  * @data:       first arg to filler(data, page) function, often left as NULL
1929  *
1930  * Read into the page cache. If a page already exists, and PageUptodate() is
1931  * not set, try to fill the page then wait for it to become unlocked.
1932  *
1933  * If the page does not get brought uptodate, return -EIO.
1934  */
1935 struct page *read_cache_page(struct address_space *mapping,
1936                                 pgoff_t index,
1937                                 int (*filler)(void *, struct page *),
1938                                 void *data)
1939 {
1940         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1941 }
1942 EXPORT_SYMBOL(read_cache_page);
1943
1944 /*
1945  * The logic we want is
1946  *
1947  *      if suid or (sgid and xgrp)
1948  *              remove privs
1949  */
1950 int should_remove_suid(struct dentry *dentry)
1951 {
1952         mode_t mode = dentry->d_inode->i_mode;
1953         int kill = 0;
1954
1955         /* suid always must be killed */
1956         if (unlikely(mode & S_ISUID))
1957                 kill = ATTR_KILL_SUID;
1958
1959         /*
1960          * sgid without any exec bits is just a mandatory locking mark; leave
1961          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1962          */
1963         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1964                 kill |= ATTR_KILL_SGID;
1965
1966         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1967                 return kill;
1968
1969         return 0;
1970 }
1971 EXPORT_SYMBOL(should_remove_suid);
1972
1973 static int __remove_suid(struct dentry *dentry, int kill)
1974 {
1975         struct iattr newattrs;
1976
1977         newattrs.ia_valid = ATTR_FORCE | kill;
1978         return notify_change(dentry, &newattrs);
1979 }
1980
1981 int file_remove_suid(struct file *file)
1982 {
1983         struct dentry *dentry = file->f_path.dentry;
1984         struct inode *inode = dentry->d_inode;
1985         int killsuid;
1986         int killpriv;
1987         int error = 0;
1988
1989         /* Fast path for nothing security related */
1990         if (IS_NOSEC(inode))
1991                 return 0;
1992
1993         killsuid = should_remove_suid(dentry);
1994         killpriv = security_inode_need_killpriv(dentry);
1995
1996         if (killpriv < 0)
1997                 return killpriv;
1998         if (killpriv)
1999                 error = security_inode_killpriv(dentry);
2000         if (!error && killsuid)
2001                 error = __remove_suid(dentry, killsuid);
2002         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2003                 inode->i_flags |= S_NOSEC;
2004
2005         return error;
2006 }
2007 EXPORT_SYMBOL(file_remove_suid);
2008
2009 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2010                         const struct iovec *iov, size_t base, size_t bytes)
2011 {
2012         size_t copied = 0, left = 0;
2013
2014         while (bytes) {
2015                 char __user *buf = iov->iov_base + base;
2016                 int copy = min(bytes, iov->iov_len - base);
2017
2018                 base = 0;
2019                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2020                 copied += copy;
2021                 bytes -= copy;
2022                 vaddr += copy;
2023                 iov++;
2024
2025                 if (unlikely(left))
2026                         break;
2027         }
2028         return copied - left;
2029 }
2030
2031 /*
2032  * Copy as much as we can into the page and return the number of bytes which
2033  * were successfully copied.  If a fault is encountered then return the number of
2034  * bytes which were copied.
2035  */
2036 size_t iov_iter_copy_from_user_atomic(struct page *page,
2037                 struct iov_iter *i, unsigned long offset, size_t bytes)
2038 {
2039         char *kaddr;
2040         size_t copied;
2041
2042         BUG_ON(!in_atomic());
2043         kaddr = kmap_atomic(page, KM_USER0);
2044         if (likely(i->nr_segs == 1)) {
2045                 int left;
2046                 char __user *buf = i->iov->iov_base + i->iov_offset;
2047                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2048                 copied = bytes - left;
2049         } else {
2050                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2051                                                 i->iov, i->iov_offset, bytes);
2052         }
2053         kunmap_atomic(kaddr, KM_USER0);
2054
2055         return copied;
2056 }
2057 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2058
2059 /*
2060  * This has the same sideeffects and return value as
2061  * iov_iter_copy_from_user_atomic().
2062  * The difference is that it attempts to resolve faults.
2063  * Page must not be locked.
2064  */
2065 size_t iov_iter_copy_from_user(struct page *page,
2066                 struct iov_iter *i, unsigned long offset, size_t bytes)
2067 {
2068         char *kaddr;
2069         size_t copied;
2070
2071         kaddr = kmap(page);
2072         if (likely(i->nr_segs == 1)) {
2073                 int left;
2074                 char __user *buf = i->iov->iov_base + i->iov_offset;
2075                 left = __copy_from_user(kaddr + offset, buf, bytes);
2076                 copied = bytes - left;
2077         } else {
2078                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2079                                                 i->iov, i->iov_offset, bytes);
2080         }
2081         kunmap(page);
2082         return copied;
2083 }
2084 EXPORT_SYMBOL(iov_iter_copy_from_user);
2085
2086 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2087 {
2088         BUG_ON(i->count < bytes);
2089
2090         if (likely(i->nr_segs == 1)) {
2091                 i->iov_offset += bytes;
2092                 i->count -= bytes;
2093         } else {
2094                 const struct iovec *iov = i->iov;
2095                 size_t base = i->iov_offset;
2096
2097                 /*
2098                  * The !iov->iov_len check ensures we skip over unlikely
2099                  * zero-length segments (without overruning the iovec).
2100                  */
2101                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2102                         int copy;
2103
2104                         copy = min(bytes, iov->iov_len - base);
2105                         BUG_ON(!i->count || i->count < copy);
2106                         i->count -= copy;
2107                         bytes -= copy;
2108                         base += copy;
2109                         if (iov->iov_len == base) {
2110                                 iov++;
2111                                 base = 0;
2112                         }
2113                 }
2114                 i->iov = iov;
2115                 i->iov_offset = base;
2116         }
2117 }
2118 EXPORT_SYMBOL(iov_iter_advance);
2119
2120 /*
2121  * Fault in the first iovec of the given iov_iter, to a maximum length
2122  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2123  * accessed (ie. because it is an invalid address).
2124  *
2125  * writev-intensive code may want this to prefault several iovecs -- that
2126  * would be possible (callers must not rely on the fact that _only_ the
2127  * first iovec will be faulted with the current implementation).
2128  */
2129 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2130 {
2131         char __user *buf = i->iov->iov_base + i->iov_offset;
2132         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2133         return fault_in_pages_readable(buf, bytes);
2134 }
2135 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2136
2137 /*
2138  * Return the count of just the current iov_iter segment.
2139  */
2140 size_t iov_iter_single_seg_count(struct iov_iter *i)
2141 {
2142         const struct iovec *iov = i->iov;
2143         if (i->nr_segs == 1)
2144                 return i->count;
2145         else
2146                 return min(i->count, iov->iov_len - i->iov_offset);
2147 }
2148 EXPORT_SYMBOL(iov_iter_single_seg_count);
2149
2150 /*
2151  * Performs necessary checks before doing a write
2152  *
2153  * Can adjust writing position or amount of bytes to write.
2154  * Returns appropriate error code that caller should return or
2155  * zero in case that write should be allowed.
2156  */
2157 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2158 {
2159         struct inode *inode = file->f_mapping->host;
2160         unsigned long limit = rlimit(RLIMIT_FSIZE);
2161
2162         if (unlikely(*pos < 0))
2163                 return -EINVAL;
2164
2165         if (!isblk) {
2166                 /* FIXME: this is for backwards compatibility with 2.4 */
2167                 if (file->f_flags & O_APPEND)
2168                         *pos = i_size_read(inode);
2169
2170                 if (limit != RLIM_INFINITY) {
2171                         if (*pos >= limit) {
2172                                 send_sig(SIGXFSZ, current, 0);
2173                                 return -EFBIG;
2174                         }
2175                         if (*count > limit - (typeof(limit))*pos) {
2176                                 *count = limit - (typeof(limit))*pos;
2177                         }
2178                 }
2179         }
2180
2181         /*
2182          * LFS rule
2183          */
2184         if (unlikely(*pos + *count > MAX_NON_LFS &&
2185                                 !(file->f_flags & O_LARGEFILE))) {
2186                 if (*pos >= MAX_NON_LFS) {
2187                         return -EFBIG;
2188                 }
2189                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2190                         *count = MAX_NON_LFS - (unsigned long)*pos;
2191                 }
2192         }
2193
2194         /*
2195          * Are we about to exceed the fs block limit ?
2196          *
2197          * If we have written data it becomes a short write.  If we have
2198          * exceeded without writing data we send a signal and return EFBIG.
2199          * Linus frestrict idea will clean these up nicely..
2200          */
2201         if (likely(!isblk)) {
2202                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2203                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2204                                 return -EFBIG;
2205                         }
2206                         /* zero-length writes at ->s_maxbytes are OK */
2207                 }
2208
2209                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2210                         *count = inode->i_sb->s_maxbytes - *pos;
2211         } else {
2212 #ifdef CONFIG_BLOCK
2213                 loff_t isize;
2214                 if (bdev_read_only(I_BDEV(inode)))
2215                         return -EPERM;
2216                 isize = i_size_read(inode);
2217                 if (*pos >= isize) {
2218                         if (*count || *pos > isize)
2219                                 return -ENOSPC;
2220                 }
2221
2222                 if (*pos + *count > isize)
2223                         *count = isize - *pos;
2224 #else
2225                 return -EPERM;
2226 #endif
2227         }
2228         return 0;
2229 }
2230 EXPORT_SYMBOL(generic_write_checks);
2231
2232 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2233                                 loff_t pos, unsigned len, unsigned flags,
2234                                 struct page **pagep, void **fsdata)
2235 {
2236         const struct address_space_operations *aops = mapping->a_ops;
2237
2238         return aops->write_begin(file, mapping, pos, len, flags,
2239                                                         pagep, fsdata);
2240 }
2241 EXPORT_SYMBOL(pagecache_write_begin);
2242
2243 int pagecache_write_end(struct file *file, struct address_space *mapping,
2244                                 loff_t pos, unsigned len, unsigned copied,
2245                                 struct page *page, void *fsdata)
2246 {
2247         const struct address_space_operations *aops = mapping->a_ops;
2248
2249         mark_page_accessed(page);
2250         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2251 }
2252 EXPORT_SYMBOL(pagecache_write_end);
2253
2254 ssize_t
2255 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2256                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2257                 size_t count, size_t ocount)
2258 {
2259         struct file     *file = iocb->ki_filp;
2260         struct address_space *mapping = file->f_mapping;
2261         struct inode    *inode = mapping->host;
2262         ssize_t         written;
2263         size_t          write_len;
2264         pgoff_t         end;
2265
2266         if (count != ocount)
2267                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2268
2269         write_len = iov_length(iov, *nr_segs);
2270         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2271
2272         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2273         if (written)
2274                 goto out;
2275
2276         /*
2277          * After a write we want buffered reads to be sure to go to disk to get
2278          * the new data.  We invalidate clean cached page from the region we're
2279          * about to write.  We do this *before* the write so that we can return
2280          * without clobbering -EIOCBQUEUED from ->direct_IO().
2281          */
2282         if (mapping->nrpages) {
2283                 written = invalidate_inode_pages2_range(mapping,
2284                                         pos >> PAGE_CACHE_SHIFT, end);
2285                 /*
2286                  * If a page can not be invalidated, return 0 to fall back
2287                  * to buffered write.
2288                  */
2289                 if (written) {
2290                         if (written == -EBUSY)
2291                                 return 0;
2292                         goto out;
2293                 }
2294         }
2295
2296         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2297
2298         /*
2299          * Finally, try again to invalidate clean pages which might have been
2300          * cached by non-direct readahead, or faulted in by get_user_pages()
2301          * if the source of the write was an mmap'ed region of the file
2302          * we're writing.  Either one is a pretty crazy thing to do,
2303          * so we don't support it 100%.  If this invalidation
2304          * fails, tough, the write still worked...
2305          */
2306         if (mapping->nrpages) {
2307                 invalidate_inode_pages2_range(mapping,
2308                                               pos >> PAGE_CACHE_SHIFT, end);
2309         }
2310
2311         if (written > 0) {
2312                 pos += written;
2313                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2314                         i_size_write(inode, pos);
2315                         mark_inode_dirty(inode);
2316                 }
2317                 *ppos = pos;
2318         }
2319 out:
2320         return written;
2321 }
2322 EXPORT_SYMBOL(generic_file_direct_write);
2323
2324 /*
2325  * Find or create a page at the given pagecache position. Return the locked
2326  * page. This function is specifically for buffered writes.
2327  */
2328 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2329                                         pgoff_t index, unsigned flags)
2330 {
2331         int status;
2332         struct page *page;
2333         gfp_t gfp_notmask = 0;
2334         if (flags & AOP_FLAG_NOFS)
2335                 gfp_notmask = __GFP_FS;
2336 repeat:
2337         page = find_lock_page(mapping, index);
2338         if (page)
2339                 goto found;
2340
2341         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2342         if (!page)
2343                 return NULL;
2344         status = add_to_page_cache_lru(page, mapping, index,
2345                                                 GFP_KERNEL & ~gfp_notmask);
2346         if (unlikely(status)) {
2347                 page_cache_release(page);
2348                 if (status == -EEXIST)
2349                         goto repeat;
2350                 return NULL;
2351         }
2352 found:
2353         wait_on_page_writeback(page);
2354         return page;
2355 }
2356 EXPORT_SYMBOL(grab_cache_page_write_begin);
2357
2358 static ssize_t generic_perform_write(struct file *file,
2359                                 struct iov_iter *i, loff_t pos)
2360 {
2361         struct address_space *mapping = file->f_mapping;
2362         const struct address_space_operations *a_ops = mapping->a_ops;
2363         long status = 0;
2364         ssize_t written = 0;
2365         unsigned int flags = 0;
2366
2367         /*
2368          * Copies from kernel address space cannot fail (NFSD is a big user).
2369          */
2370         if (segment_eq(get_fs(), KERNEL_DS))
2371                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2372
2373         do {
2374                 struct page *page;
2375                 unsigned long offset;   /* Offset into pagecache page */
2376                 unsigned long bytes;    /* Bytes to write to page */
2377                 size_t copied;          /* Bytes copied from user */
2378                 void *fsdata;
2379
2380                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2381                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2382                                                 iov_iter_count(i));
2383
2384 again:
2385
2386                 /*
2387                  * Bring in the user page that we will copy from _first_.
2388                  * Otherwise there's a nasty deadlock on copying from the
2389                  * same page as we're writing to, without it being marked
2390                  * up-to-date.
2391                  *
2392                  * Not only is this an optimisation, but it is also required
2393                  * to check that the address is actually valid, when atomic
2394                  * usercopies are used, below.
2395                  */
2396                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2397                         status = -EFAULT;
2398                         break;
2399                 }
2400
2401                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2402                                                 &page, &fsdata);
2403                 if (unlikely(status))
2404                         break;
2405
2406                 if (mapping_writably_mapped(mapping))
2407                         flush_dcache_page(page);
2408
2409                 pagefault_disable();
2410                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2411                 pagefault_enable();
2412                 flush_dcache_page(page);
2413
2414                 mark_page_accessed(page);
2415                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2416                                                 page, fsdata);
2417                 if (unlikely(status < 0))
2418                         break;
2419                 copied = status;
2420
2421                 cond_resched();
2422
2423                 iov_iter_advance(i, copied);
2424                 if (unlikely(copied == 0)) {
2425                         /*
2426                          * If we were unable to copy any data at all, we must
2427                          * fall back to a single segment length write.
2428                          *
2429                          * If we didn't fallback here, we could livelock
2430                          * because not all segments in the iov can be copied at
2431                          * once without a pagefault.
2432                          */
2433                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2434                                                 iov_iter_single_seg_count(i));
2435                         goto again;
2436                 }
2437                 pos += copied;
2438                 written += copied;
2439
2440                 balance_dirty_pages_ratelimited(mapping);
2441
2442         } while (iov_iter_count(i));
2443
2444         return written ? written : status;
2445 }
2446
2447 ssize_t
2448 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2449                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2450                 size_t count, ssize_t written)
2451 {
2452         struct file *file = iocb->ki_filp;
2453         ssize_t status;
2454         struct iov_iter i;
2455
2456         iov_iter_init(&i, iov, nr_segs, count, written);
2457         status = generic_perform_write(file, &i, pos);
2458
2459         if (likely(status >= 0)) {
2460                 written += status;
2461                 *ppos = pos + status;
2462         }
2463         
2464         return written ? written : status;
2465 }
2466 EXPORT_SYMBOL(generic_file_buffered_write);
2467
2468 /**
2469  * __generic_file_aio_write - write data to a file
2470  * @iocb:       IO state structure (file, offset, etc.)
2471  * @iov:        vector with data to write
2472  * @nr_segs:    number of segments in the vector
2473  * @ppos:       position where to write
2474  *
2475  * This function does all the work needed for actually writing data to a
2476  * file. It does all basic checks, removes SUID from the file, updates
2477  * modification times and calls proper subroutines depending on whether we
2478  * do direct IO or a standard buffered write.
2479  *
2480  * It expects i_mutex to be grabbed unless we work on a block device or similar
2481  * object which does not need locking at all.
2482  *
2483  * This function does *not* take care of syncing data in case of O_SYNC write.
2484  * A caller has to handle it. This is mainly due to the fact that we want to
2485  * avoid syncing under i_mutex.
2486  */
2487 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2488                                  unsigned long nr_segs, loff_t *ppos)
2489 {
2490         struct file *file = iocb->ki_filp;
2491         struct address_space * mapping = file->f_mapping;
2492         size_t ocount;          /* original count */
2493         size_t count;           /* after file limit checks */
2494         struct inode    *inode = mapping->host;
2495         loff_t          pos;
2496         ssize_t         written;
2497         ssize_t         err;
2498
2499         ocount = 0;
2500         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2501         if (err)
2502                 return err;
2503
2504         count = ocount;
2505         pos = *ppos;
2506
2507         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2508
2509         /* We can write back this queue in page reclaim */
2510         current->backing_dev_info = mapping->backing_dev_info;
2511         written = 0;
2512
2513         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2514         if (err)
2515                 goto out;
2516
2517         if (count == 0)
2518                 goto out;
2519
2520         err = file_remove_suid(file);
2521         if (err)
2522                 goto out;
2523
2524         file_update_time(file);
2525
2526         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2527         if (unlikely(file->f_flags & O_DIRECT)) {
2528                 loff_t endbyte;
2529                 ssize_t written_buffered;
2530
2531                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2532                                                         ppos, count, ocount);
2533                 if (written < 0 || written == count)
2534                         goto out;
2535                 /*
2536                  * direct-io write to a hole: fall through to buffered I/O
2537                  * for completing the rest of the request.
2538                  */
2539                 pos += written;
2540                 count -= written;
2541                 written_buffered = generic_file_buffered_write(iocb, iov,
2542                                                 nr_segs, pos, ppos, count,
2543                                                 written);
2544                 /*
2545                  * If generic_file_buffered_write() retuned a synchronous error
2546                  * then we want to return the number of bytes which were
2547                  * direct-written, or the error code if that was zero.  Note
2548                  * that this differs from normal direct-io semantics, which
2549                  * will return -EFOO even if some bytes were written.
2550                  */
2551                 if (written_buffered < 0) {
2552                         err = written_buffered;
2553                         goto out;
2554                 }
2555
2556                 /*
2557                  * We need to ensure that the page cache pages are written to
2558                  * disk and invalidated to preserve the expected O_DIRECT
2559                  * semantics.
2560                  */
2561                 endbyte = pos + written_buffered - written - 1;
2562                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2563                 if (err == 0) {
2564                         written = written_buffered;
2565                         invalidate_mapping_pages(mapping,
2566                                                  pos >> PAGE_CACHE_SHIFT,
2567                                                  endbyte >> PAGE_CACHE_SHIFT);
2568                 } else {
2569                         /*
2570                          * We don't know how much we wrote, so just return
2571                          * the number of bytes which were direct-written
2572                          */
2573                 }
2574         } else {
2575                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2576                                 pos, ppos, count, written);
2577         }
2578 out:
2579         current->backing_dev_info = NULL;
2580         return written ? written : err;
2581 }
2582 EXPORT_SYMBOL(__generic_file_aio_write);
2583
2584 /**
2585  * generic_file_aio_write - write data to a file
2586  * @iocb:       IO state structure
2587  * @iov:        vector with data to write
2588  * @nr_segs:    number of segments in the vector
2589  * @pos:        position in file where to write
2590  *
2591  * This is a wrapper around __generic_file_aio_write() to be used by most
2592  * filesystems. It takes care of syncing the file in case of O_SYNC file
2593  * and acquires i_mutex as needed.
2594  */
2595 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2596                 unsigned long nr_segs, loff_t pos)
2597 {
2598         struct file *file = iocb->ki_filp;
2599         struct inode *inode = file->f_mapping->host;
2600         struct blk_plug plug;
2601         ssize_t ret;
2602
2603         BUG_ON(iocb->ki_pos != pos);
2604
2605         mutex_lock(&inode->i_mutex);
2606         blk_start_plug(&plug);
2607         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2608         mutex_unlock(&inode->i_mutex);
2609
2610         if (ret > 0 || ret == -EIOCBQUEUED) {
2611                 ssize_t err;
2612
2613                 err = generic_write_sync(file, pos, ret);
2614                 if (err < 0 && ret > 0)
2615                         ret = err;
2616         }
2617         blk_finish_plug(&plug);
2618         return ret;
2619 }
2620 EXPORT_SYMBOL(generic_file_aio_write);
2621
2622 /**
2623  * try_to_release_page() - release old fs-specific metadata on a page
2624  *
2625  * @page: the page which the kernel is trying to free
2626  * @gfp_mask: memory allocation flags (and I/O mode)
2627  *
2628  * The address_space is to try to release any data against the page
2629  * (presumably at page->private).  If the release was successful, return `1'.
2630  * Otherwise return zero.
2631  *
2632  * This may also be called if PG_fscache is set on a page, indicating that the
2633  * page is known to the local caching routines.
2634  *
2635  * The @gfp_mask argument specifies whether I/O may be performed to release
2636  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2637  *
2638  */
2639 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2640 {
2641         struct address_space * const mapping = page->mapping;
2642
2643         BUG_ON(!PageLocked(page));
2644         if (PageWriteback(page))
2645                 return 0;
2646
2647         if (mapping && mapping->a_ops->releasepage)
2648                 return mapping->a_ops->releasepage(page, gfp_mask);
2649         return try_to_free_buffers(page);
2650 }
2651
2652 EXPORT_SYMBOL(try_to_release_page);