Cache xattr security drop check for write v2
[linux-3.10.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  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  inode_wb_list_lock
85  *    sb_lock                   (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_mutex
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    inode_wb_list_lock        (page_remove_rmap->set_page_dirty)
103  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
104  *    inode_wb_list_lock        (zap_pte_range->set_page_dirty)
105  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  *  (code doesn't rely on that order, so you could switch it around)
109  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
110  *    ->i_mmap_mutex
111  */
112
113 /*
114  * Delete a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold the mapping's tree_lock.
117  */
118 void __delete_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         /*
123          * if we're uptodate, flush out into the cleancache, otherwise
124          * invalidate any existing cleancache entries.  We can't leave
125          * stale data around in the cleancache once our page is gone
126          */
127         if (PageUptodate(page) && PageMappedToDisk(page))
128                 cleancache_put_page(page);
129         else
130                 cleancache_flush_page(mapping, page);
131
132         radix_tree_delete(&mapping->page_tree, page->index);
133         page->mapping = NULL;
134         mapping->nrpages--;
135         __dec_zone_page_state(page, NR_FILE_PAGES);
136         if (PageSwapBacked(page))
137                 __dec_zone_page_state(page, NR_SHMEM);
138         BUG_ON(page_mapped(page));
139
140         /*
141          * Some filesystems seem to re-dirty the page even after
142          * the VM has canceled the dirty bit (eg ext3 journaling).
143          *
144          * Fix it up by doing a final dirty accounting check after
145          * having removed the page entirely.
146          */
147         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
148                 dec_zone_page_state(page, NR_FILE_DIRTY);
149                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
150         }
151 }
152
153 /**
154  * delete_from_page_cache - delete page from page cache
155  * @page: the page which the kernel is trying to remove from page cache
156  *
157  * This must be called only on pages that have been verified to be in the page
158  * cache and locked.  It will never put the page into the free list, the caller
159  * has a reference on the page.
160  */
161 void delete_from_page_cache(struct page *page)
162 {
163         struct address_space *mapping = page->mapping;
164         void (*freepage)(struct page *);
165
166         BUG_ON(!PageLocked(page));
167
168         freepage = mapping->a_ops->freepage;
169         spin_lock_irq(&mapping->tree_lock);
170         __delete_from_page_cache(page);
171         spin_unlock_irq(&mapping->tree_lock);
172         mem_cgroup_uncharge_cache_page(page);
173
174         if (freepage)
175                 freepage(page);
176         page_cache_release(page);
177 }
178 EXPORT_SYMBOL(delete_from_page_cache);
179
180 static int sleep_on_page(void *word)
181 {
182         io_schedule();
183         return 0;
184 }
185
186 static int sleep_on_page_killable(void *word)
187 {
188         sleep_on_page(word);
189         return fatal_signal_pending(current) ? -EINTR : 0;
190 }
191
192 /**
193  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
194  * @mapping:    address space structure to write
195  * @start:      offset in bytes where the range starts
196  * @end:        offset in bytes where the range ends (inclusive)
197  * @sync_mode:  enable synchronous operation
198  *
199  * Start writeback against all of a mapping's dirty pages that lie
200  * within the byte offsets <start, end> inclusive.
201  *
202  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
203  * opposed to a regular memory cleansing writeback.  The difference between
204  * these two operations is that if a dirty page/buffer is encountered, it must
205  * be waited upon, and not just skipped over.
206  */
207 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
208                                 loff_t end, int sync_mode)
209 {
210         int ret;
211         struct writeback_control wbc = {
212                 .sync_mode = sync_mode,
213                 .nr_to_write = LONG_MAX,
214                 .range_start = start,
215                 .range_end = end,
216         };
217
218         if (!mapping_cap_writeback_dirty(mapping))
219                 return 0;
220
221         ret = do_writepages(mapping, &wbc);
222         return ret;
223 }
224
225 static inline int __filemap_fdatawrite(struct address_space *mapping,
226         int sync_mode)
227 {
228         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
229 }
230
231 int filemap_fdatawrite(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
234 }
235 EXPORT_SYMBOL(filemap_fdatawrite);
236
237 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
238                                 loff_t end)
239 {
240         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
241 }
242 EXPORT_SYMBOL(filemap_fdatawrite_range);
243
244 /**
245  * filemap_flush - mostly a non-blocking flush
246  * @mapping:    target address_space
247  *
248  * This is a mostly non-blocking flush.  Not suitable for data-integrity
249  * purposes - I/O may not be started against all dirty pages.
250  */
251 int filemap_flush(struct address_space *mapping)
252 {
253         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
254 }
255 EXPORT_SYMBOL(filemap_flush);
256
257 /**
258  * filemap_fdatawait_range - wait for writeback to complete
259  * @mapping:            address space structure to wait for
260  * @start_byte:         offset in bytes where the range starts
261  * @end_byte:           offset in bytes where the range ends (inclusive)
262  *
263  * Walk the list of under-writeback pages of the given address space
264  * in the given range and wait for all of them.
265  */
266 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
267                             loff_t end_byte)
268 {
269         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
270         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
271         struct pagevec pvec;
272         int nr_pages;
273         int ret = 0;
274
275         if (end_byte < start_byte)
276                 return 0;
277
278         pagevec_init(&pvec, 0);
279         while ((index <= end) &&
280                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
281                         PAGECACHE_TAG_WRITEBACK,
282                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
283                 unsigned i;
284
285                 for (i = 0; i < nr_pages; i++) {
286                         struct page *page = pvec.pages[i];
287
288                         /* until radix tree lookup accepts end_index */
289                         if (page->index > end)
290                                 continue;
291
292                         wait_on_page_writeback(page);
293                         if (TestClearPageError(page))
294                                 ret = -EIO;
295                 }
296                 pagevec_release(&pvec);
297                 cond_resched();
298         }
299
300         /* Check for outstanding write errors */
301         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
302                 ret = -ENOSPC;
303         if (test_and_clear_bit(AS_EIO, &mapping->flags))
304                 ret = -EIO;
305
306         return ret;
307 }
308 EXPORT_SYMBOL(filemap_fdatawait_range);
309
310 /**
311  * filemap_fdatawait - wait for all under-writeback pages to complete
312  * @mapping: address space structure to wait for
313  *
314  * Walk the list of under-writeback pages of the given address space
315  * and wait for all of them.
316  */
317 int filemap_fdatawait(struct address_space *mapping)
318 {
319         loff_t i_size = i_size_read(mapping->host);
320
321         if (i_size == 0)
322                 return 0;
323
324         return filemap_fdatawait_range(mapping, 0, i_size - 1);
325 }
326 EXPORT_SYMBOL(filemap_fdatawait);
327
328 int filemap_write_and_wait(struct address_space *mapping)
329 {
330         int err = 0;
331
332         if (mapping->nrpages) {
333                 err = filemap_fdatawrite(mapping);
334                 /*
335                  * Even if the above returned error, the pages may be
336                  * written partially (e.g. -ENOSPC), so we wait for it.
337                  * But the -EIO is special case, it may indicate the worst
338                  * thing (e.g. bug) happened, so we avoid waiting for it.
339                  */
340                 if (err != -EIO) {
341                         int err2 = filemap_fdatawait(mapping);
342                         if (!err)
343                                 err = err2;
344                 }
345         }
346         return err;
347 }
348 EXPORT_SYMBOL(filemap_write_and_wait);
349
350 /**
351  * filemap_write_and_wait_range - write out & wait on a file range
352  * @mapping:    the address_space for the pages
353  * @lstart:     offset in bytes where the range starts
354  * @lend:       offset in bytes where the range ends (inclusive)
355  *
356  * Write out and wait upon file offsets lstart->lend, inclusive.
357  *
358  * Note that `lend' is inclusive (describes the last byte to be written) so
359  * that this function can be used to write to the very end-of-file (end = -1).
360  */
361 int filemap_write_and_wait_range(struct address_space *mapping,
362                                  loff_t lstart, loff_t lend)
363 {
364         int err = 0;
365
366         if (mapping->nrpages) {
367                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
368                                                  WB_SYNC_ALL);
369                 /* See comment of filemap_write_and_wait() */
370                 if (err != -EIO) {
371                         int err2 = filemap_fdatawait_range(mapping,
372                                                 lstart, lend);
373                         if (!err)
374                                 err = err2;
375                 }
376         }
377         return err;
378 }
379 EXPORT_SYMBOL(filemap_write_and_wait_range);
380
381 /**
382  * replace_page_cache_page - replace a pagecache page with a new one
383  * @old:        page to be replaced
384  * @new:        page to replace with
385  * @gfp_mask:   allocation mode
386  *
387  * This function replaces a page in the pagecache with a new one.  On
388  * success it acquires the pagecache reference for the new page and
389  * drops it for the old page.  Both the old and new pages must be
390  * locked.  This function does not add the new page to the LRU, the
391  * caller must do that.
392  *
393  * The remove + add is atomic.  The only way this function can fail is
394  * memory allocation failure.
395  */
396 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
397 {
398         int error;
399         struct mem_cgroup *memcg = NULL;
400
401         VM_BUG_ON(!PageLocked(old));
402         VM_BUG_ON(!PageLocked(new));
403         VM_BUG_ON(new->mapping);
404
405         /*
406          * This is not page migration, but prepare_migration and
407          * end_migration does enough work for charge replacement.
408          *
409          * In the longer term we probably want a specialized function
410          * for moving the charge from old to new in a more efficient
411          * manner.
412          */
413         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
414         if (error)
415                 return error;
416
417         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
418         if (!error) {
419                 struct address_space *mapping = old->mapping;
420                 void (*freepage)(struct page *);
421
422                 pgoff_t offset = old->index;
423                 freepage = mapping->a_ops->freepage;
424
425                 page_cache_get(new);
426                 new->mapping = mapping;
427                 new->index = offset;
428
429                 spin_lock_irq(&mapping->tree_lock);
430                 __delete_from_page_cache(old);
431                 error = radix_tree_insert(&mapping->page_tree, offset, new);
432                 BUG_ON(error);
433                 mapping->nrpages++;
434                 __inc_zone_page_state(new, NR_FILE_PAGES);
435                 if (PageSwapBacked(new))
436                         __inc_zone_page_state(new, NR_SHMEM);
437                 spin_unlock_irq(&mapping->tree_lock);
438                 radix_tree_preload_end();
439                 if (freepage)
440                         freepage(old);
441                 page_cache_release(old);
442                 mem_cgroup_end_migration(memcg, old, new, true);
443         } else {
444                 mem_cgroup_end_migration(memcg, old, new, false);
445         }
446
447         return error;
448 }
449 EXPORT_SYMBOL_GPL(replace_page_cache_page);
450
451 /**
452  * add_to_page_cache_locked - add a locked page to the pagecache
453  * @page:       page to add
454  * @mapping:    the page's address_space
455  * @offset:     page index
456  * @gfp_mask:   page allocation mode
457  *
458  * This function is used to add a page to the pagecache. It must be locked.
459  * This function does not add the page to the LRU.  The caller must do that.
460  */
461 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
462                 pgoff_t offset, gfp_t gfp_mask)
463 {
464         int error;
465
466         VM_BUG_ON(!PageLocked(page));
467
468         error = mem_cgroup_cache_charge(page, current->mm,
469                                         gfp_mask & GFP_RECLAIM_MASK);
470         if (error)
471                 goto out;
472
473         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
474         if (error == 0) {
475                 page_cache_get(page);
476                 page->mapping = mapping;
477                 page->index = offset;
478
479                 spin_lock_irq(&mapping->tree_lock);
480                 error = radix_tree_insert(&mapping->page_tree, offset, page);
481                 if (likely(!error)) {
482                         mapping->nrpages++;
483                         __inc_zone_page_state(page, NR_FILE_PAGES);
484                         if (PageSwapBacked(page))
485                                 __inc_zone_page_state(page, NR_SHMEM);
486                         spin_unlock_irq(&mapping->tree_lock);
487                 } else {
488                         page->mapping = NULL;
489                         spin_unlock_irq(&mapping->tree_lock);
490                         mem_cgroup_uncharge_cache_page(page);
491                         page_cache_release(page);
492                 }
493                 radix_tree_preload_end();
494         } else
495                 mem_cgroup_uncharge_cache_page(page);
496 out:
497         return error;
498 }
499 EXPORT_SYMBOL(add_to_page_cache_locked);
500
501 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
502                                 pgoff_t offset, gfp_t gfp_mask)
503 {
504         int ret;
505
506         /*
507          * Splice_read and readahead add shmem/tmpfs pages into the page cache
508          * before shmem_readpage has a chance to mark them as SwapBacked: they
509          * need to go on the anon lru below, and mem_cgroup_cache_charge
510          * (called in add_to_page_cache) needs to know where they're going too.
511          */
512         if (mapping_cap_swap_backed(mapping))
513                 SetPageSwapBacked(page);
514
515         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
516         if (ret == 0) {
517                 if (page_is_file_cache(page))
518                         lru_cache_add_file(page);
519                 else
520                         lru_cache_add_anon(page);
521         }
522         return ret;
523 }
524 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
525
526 #ifdef CONFIG_NUMA
527 struct page *__page_cache_alloc(gfp_t gfp)
528 {
529         int n;
530         struct page *page;
531
532         if (cpuset_do_page_mem_spread()) {
533                 get_mems_allowed();
534                 n = cpuset_mem_spread_node();
535                 page = alloc_pages_exact_node(n, gfp, 0);
536                 put_mems_allowed();
537                 return page;
538         }
539         return alloc_pages(gfp, 0);
540 }
541 EXPORT_SYMBOL(__page_cache_alloc);
542 #endif
543
544 /*
545  * In order to wait for pages to become available there must be
546  * waitqueues associated with pages. By using a hash table of
547  * waitqueues where the bucket discipline is to maintain all
548  * waiters on the same queue and wake all when any of the pages
549  * become available, and for the woken contexts to check to be
550  * sure the appropriate page became available, this saves space
551  * at a cost of "thundering herd" phenomena during rare hash
552  * collisions.
553  */
554 static wait_queue_head_t *page_waitqueue(struct page *page)
555 {
556         const struct zone *zone = page_zone(page);
557
558         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
559 }
560
561 static inline void wake_up_page(struct page *page, int bit)
562 {
563         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
564 }
565
566 void wait_on_page_bit(struct page *page, int bit_nr)
567 {
568         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
569
570         if (test_bit(bit_nr, &page->flags))
571                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
572                                                         TASK_UNINTERRUPTIBLE);
573 }
574 EXPORT_SYMBOL(wait_on_page_bit);
575
576 int wait_on_page_bit_killable(struct page *page, int bit_nr)
577 {
578         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
579
580         if (!test_bit(bit_nr, &page->flags))
581                 return 0;
582
583         return __wait_on_bit(page_waitqueue(page), &wait,
584                              sleep_on_page_killable, TASK_KILLABLE);
585 }
586
587 /**
588  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
589  * @page: Page defining the wait queue of interest
590  * @waiter: Waiter to add to the queue
591  *
592  * Add an arbitrary @waiter to the wait queue for the nominated @page.
593  */
594 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
595 {
596         wait_queue_head_t *q = page_waitqueue(page);
597         unsigned long flags;
598
599         spin_lock_irqsave(&q->lock, flags);
600         __add_wait_queue(q, waiter);
601         spin_unlock_irqrestore(&q->lock, flags);
602 }
603 EXPORT_SYMBOL_GPL(add_page_wait_queue);
604
605 /**
606  * unlock_page - unlock a locked page
607  * @page: the page
608  *
609  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
610  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
611  * mechananism between PageLocked pages and PageWriteback pages is shared.
612  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
613  *
614  * The mb is necessary to enforce ordering between the clear_bit and the read
615  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
616  */
617 void unlock_page(struct page *page)
618 {
619         VM_BUG_ON(!PageLocked(page));
620         clear_bit_unlock(PG_locked, &page->flags);
621         smp_mb__after_clear_bit();
622         wake_up_page(page, PG_locked);
623 }
624 EXPORT_SYMBOL(unlock_page);
625
626 /**
627  * end_page_writeback - end writeback against a page
628  * @page: the page
629  */
630 void end_page_writeback(struct page *page)
631 {
632         if (TestClearPageReclaim(page))
633                 rotate_reclaimable_page(page);
634
635         if (!test_clear_page_writeback(page))
636                 BUG();
637
638         smp_mb__after_clear_bit();
639         wake_up_page(page, PG_writeback);
640 }
641 EXPORT_SYMBOL(end_page_writeback);
642
643 /**
644  * __lock_page - get a lock on the page, assuming we need to sleep to get it
645  * @page: the page to lock
646  */
647 void __lock_page(struct page *page)
648 {
649         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
650
651         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
652                                                         TASK_UNINTERRUPTIBLE);
653 }
654 EXPORT_SYMBOL(__lock_page);
655
656 int __lock_page_killable(struct page *page)
657 {
658         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
659
660         return __wait_on_bit_lock(page_waitqueue(page), &wait,
661                                         sleep_on_page_killable, TASK_KILLABLE);
662 }
663 EXPORT_SYMBOL_GPL(__lock_page_killable);
664
665 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
666                          unsigned int flags)
667 {
668         if (flags & FAULT_FLAG_ALLOW_RETRY) {
669                 /*
670                  * CAUTION! In this case, mmap_sem is not released
671                  * even though return 0.
672                  */
673                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
674                         return 0;
675
676                 up_read(&mm->mmap_sem);
677                 if (flags & FAULT_FLAG_KILLABLE)
678                         wait_on_page_locked_killable(page);
679                 else
680                         wait_on_page_locked(page);
681                 return 0;
682         } else {
683                 if (flags & FAULT_FLAG_KILLABLE) {
684                         int ret;
685
686                         ret = __lock_page_killable(page);
687                         if (ret) {
688                                 up_read(&mm->mmap_sem);
689                                 return 0;
690                         }
691                 } else
692                         __lock_page(page);
693                 return 1;
694         }
695 }
696
697 /**
698  * find_get_page - find and get a page reference
699  * @mapping: the address_space to search
700  * @offset: the page index
701  *
702  * Is there a pagecache struct page at the given (mapping, offset) tuple?
703  * If yes, increment its refcount and return it; if no, return NULL.
704  */
705 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
706 {
707         void **pagep;
708         struct page *page;
709
710         rcu_read_lock();
711 repeat:
712         page = NULL;
713         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
714         if (pagep) {
715                 page = radix_tree_deref_slot(pagep);
716                 if (unlikely(!page))
717                         goto out;
718                 if (radix_tree_deref_retry(page))
719                         goto repeat;
720
721                 if (!page_cache_get_speculative(page))
722                         goto repeat;
723
724                 /*
725                  * Has the page moved?
726                  * This is part of the lockless pagecache protocol. See
727                  * include/linux/pagemap.h for details.
728                  */
729                 if (unlikely(page != *pagep)) {
730                         page_cache_release(page);
731                         goto repeat;
732                 }
733         }
734 out:
735         rcu_read_unlock();
736
737         return page;
738 }
739 EXPORT_SYMBOL(find_get_page);
740
741 /**
742  * find_lock_page - locate, pin and lock a pagecache page
743  * @mapping: the address_space to search
744  * @offset: the page index
745  *
746  * Locates the desired pagecache page, locks it, increments its reference
747  * count and returns its address.
748  *
749  * Returns zero if the page was not present. find_lock_page() may sleep.
750  */
751 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
752 {
753         struct page *page;
754
755 repeat:
756         page = find_get_page(mapping, offset);
757         if (page) {
758                 lock_page(page);
759                 /* Has the page been truncated? */
760                 if (unlikely(page->mapping != mapping)) {
761                         unlock_page(page);
762                         page_cache_release(page);
763                         goto repeat;
764                 }
765                 VM_BUG_ON(page->index != offset);
766         }
767         return page;
768 }
769 EXPORT_SYMBOL(find_lock_page);
770
771 /**
772  * find_or_create_page - locate or add a pagecache page
773  * @mapping: the page's address_space
774  * @index: the page's index into the mapping
775  * @gfp_mask: page allocation mode
776  *
777  * Locates a page in the pagecache.  If the page is not present, a new page
778  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
779  * LRU list.  The returned page is locked and has its reference count
780  * incremented.
781  *
782  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
783  * allocation!
784  *
785  * find_or_create_page() returns the desired page's address, or zero on
786  * memory exhaustion.
787  */
788 struct page *find_or_create_page(struct address_space *mapping,
789                 pgoff_t index, gfp_t gfp_mask)
790 {
791         struct page *page;
792         int err;
793 repeat:
794         page = find_lock_page(mapping, index);
795         if (!page) {
796                 page = __page_cache_alloc(gfp_mask);
797                 if (!page)
798                         return NULL;
799                 /*
800                  * We want a regular kernel memory (not highmem or DMA etc)
801                  * allocation for the radix tree nodes, but we need to honour
802                  * the context-specific requirements the caller has asked for.
803                  * GFP_RECLAIM_MASK collects those requirements.
804                  */
805                 err = add_to_page_cache_lru(page, mapping, index,
806                         (gfp_mask & GFP_RECLAIM_MASK));
807                 if (unlikely(err)) {
808                         page_cache_release(page);
809                         page = NULL;
810                         if (err == -EEXIST)
811                                 goto repeat;
812                 }
813         }
814         return page;
815 }
816 EXPORT_SYMBOL(find_or_create_page);
817
818 /**
819  * find_get_pages - gang pagecache lookup
820  * @mapping:    The address_space to search
821  * @start:      The starting page index
822  * @nr_pages:   The maximum number of pages
823  * @pages:      Where the resulting pages are placed
824  *
825  * find_get_pages() will search for and return a group of up to
826  * @nr_pages pages in the mapping.  The pages are placed at @pages.
827  * find_get_pages() takes a reference against the returned pages.
828  *
829  * The search returns a group of mapping-contiguous pages with ascending
830  * indexes.  There may be holes in the indices due to not-present pages.
831  *
832  * find_get_pages() returns the number of pages which were found.
833  */
834 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
835                             unsigned int nr_pages, struct page **pages)
836 {
837         unsigned int i;
838         unsigned int ret;
839         unsigned int nr_found;
840
841         rcu_read_lock();
842 restart:
843         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
844                                 (void ***)pages, start, nr_pages);
845         ret = 0;
846         for (i = 0; i < nr_found; i++) {
847                 struct page *page;
848 repeat:
849                 page = radix_tree_deref_slot((void **)pages[i]);
850                 if (unlikely(!page))
851                         continue;
852
853                 /*
854                  * This can only trigger when the entry at index 0 moves out
855                  * of or back to the root: none yet gotten, safe to restart.
856                  */
857                 if (radix_tree_deref_retry(page)) {
858                         WARN_ON(start | i);
859                         goto restart;
860                 }
861
862                 if (!page_cache_get_speculative(page))
863                         goto repeat;
864
865                 /* Has the page moved? */
866                 if (unlikely(page != *((void **)pages[i]))) {
867                         page_cache_release(page);
868                         goto repeat;
869                 }
870
871                 pages[ret] = page;
872                 ret++;
873         }
874
875         /*
876          * If all entries were removed before we could secure them,
877          * try again, because callers stop trying once 0 is returned.
878          */
879         if (unlikely(!ret && nr_found))
880                 goto restart;
881         rcu_read_unlock();
882         return ret;
883 }
884
885 /**
886  * find_get_pages_contig - gang contiguous pagecache lookup
887  * @mapping:    The address_space to search
888  * @index:      The starting page index
889  * @nr_pages:   The maximum number of pages
890  * @pages:      Where the resulting pages are placed
891  *
892  * find_get_pages_contig() works exactly like find_get_pages(), except
893  * that the returned number of pages are guaranteed to be contiguous.
894  *
895  * find_get_pages_contig() returns the number of pages which were found.
896  */
897 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
898                                unsigned int nr_pages, struct page **pages)
899 {
900         unsigned int i;
901         unsigned int ret;
902         unsigned int nr_found;
903
904         rcu_read_lock();
905 restart:
906         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
907                                 (void ***)pages, index, nr_pages);
908         ret = 0;
909         for (i = 0; i < nr_found; i++) {
910                 struct page *page;
911 repeat:
912                 page = radix_tree_deref_slot((void **)pages[i]);
913                 if (unlikely(!page))
914                         continue;
915
916                 /*
917                  * This can only trigger when the entry at index 0 moves out
918                  * of or back to the root: none yet gotten, safe to restart.
919                  */
920                 if (radix_tree_deref_retry(page))
921                         goto restart;
922
923                 if (!page_cache_get_speculative(page))
924                         goto repeat;
925
926                 /* Has the page moved? */
927                 if (unlikely(page != *((void **)pages[i]))) {
928                         page_cache_release(page);
929                         goto repeat;
930                 }
931
932                 /*
933                  * must check mapping and index after taking the ref.
934                  * otherwise we can get both false positives and false
935                  * negatives, which is just confusing to the caller.
936                  */
937                 if (page->mapping == NULL || page->index != index) {
938                         page_cache_release(page);
939                         break;
940                 }
941
942                 pages[ret] = page;
943                 ret++;
944                 index++;
945         }
946         rcu_read_unlock();
947         return ret;
948 }
949 EXPORT_SYMBOL(find_get_pages_contig);
950
951 /**
952  * find_get_pages_tag - find and return pages that match @tag
953  * @mapping:    the address_space to search
954  * @index:      the starting page index
955  * @tag:        the tag index
956  * @nr_pages:   the maximum number of pages
957  * @pages:      where the resulting pages are placed
958  *
959  * Like find_get_pages, except we only return pages which are tagged with
960  * @tag.   We update @index to index the next page for the traversal.
961  */
962 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
963                         int tag, unsigned int nr_pages, struct page **pages)
964 {
965         unsigned int i;
966         unsigned int ret;
967         unsigned int nr_found;
968
969         rcu_read_lock();
970 restart:
971         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
972                                 (void ***)pages, *index, nr_pages, tag);
973         ret = 0;
974         for (i = 0; i < nr_found; i++) {
975                 struct page *page;
976 repeat:
977                 page = radix_tree_deref_slot((void **)pages[i]);
978                 if (unlikely(!page))
979                         continue;
980
981                 /*
982                  * This can only trigger when the entry at index 0 moves out
983                  * of or back to the root: none yet gotten, safe to restart.
984                  */
985                 if (radix_tree_deref_retry(page))
986                         goto restart;
987
988                 if (!page_cache_get_speculative(page))
989                         goto repeat;
990
991                 /* Has the page moved? */
992                 if (unlikely(page != *((void **)pages[i]))) {
993                         page_cache_release(page);
994                         goto repeat;
995                 }
996
997                 pages[ret] = page;
998                 ret++;
999         }
1000
1001         /*
1002          * If all entries were removed before we could secure them,
1003          * try again, because callers stop trying once 0 is returned.
1004          */
1005         if (unlikely(!ret && nr_found))
1006                 goto restart;
1007         rcu_read_unlock();
1008
1009         if (ret)
1010                 *index = pages[ret - 1]->index + 1;
1011
1012         return ret;
1013 }
1014 EXPORT_SYMBOL(find_get_pages_tag);
1015
1016 /**
1017  * grab_cache_page_nowait - returns locked page at given index in given cache
1018  * @mapping: target address_space
1019  * @index: the page index
1020  *
1021  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1022  * This is intended for speculative data generators, where the data can
1023  * be regenerated if the page couldn't be grabbed.  This routine should
1024  * be safe to call while holding the lock for another page.
1025  *
1026  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1027  * and deadlock against the caller's locked page.
1028  */
1029 struct page *
1030 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1031 {
1032         struct page *page = find_get_page(mapping, index);
1033
1034         if (page) {
1035                 if (trylock_page(page))
1036                         return page;
1037                 page_cache_release(page);
1038                 return NULL;
1039         }
1040         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1041         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1042                 page_cache_release(page);
1043                 page = NULL;
1044         }
1045         return page;
1046 }
1047 EXPORT_SYMBOL(grab_cache_page_nowait);
1048
1049 /*
1050  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1051  * a _large_ part of the i/o request. Imagine the worst scenario:
1052  *
1053  *      ---R__________________________________________B__________
1054  *         ^ reading here                             ^ bad block(assume 4k)
1055  *
1056  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1057  * => failing the whole request => read(R) => read(R+1) =>
1058  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1059  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1060  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1061  *
1062  * It is going insane. Fix it by quickly scaling down the readahead size.
1063  */
1064 static void shrink_readahead_size_eio(struct file *filp,
1065                                         struct file_ra_state *ra)
1066 {
1067         ra->ra_pages /= 4;
1068 }
1069
1070 /**
1071  * do_generic_file_read - generic file read routine
1072  * @filp:       the file to read
1073  * @ppos:       current file position
1074  * @desc:       read_descriptor
1075  * @actor:      read method
1076  *
1077  * This is a generic file read routine, and uses the
1078  * mapping->a_ops->readpage() function for the actual low-level stuff.
1079  *
1080  * This is really ugly. But the goto's actually try to clarify some
1081  * of the logic when it comes to error handling etc.
1082  */
1083 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1084                 read_descriptor_t *desc, read_actor_t actor)
1085 {
1086         struct address_space *mapping = filp->f_mapping;
1087         struct inode *inode = mapping->host;
1088         struct file_ra_state *ra = &filp->f_ra;
1089         pgoff_t index;
1090         pgoff_t last_index;
1091         pgoff_t prev_index;
1092         unsigned long offset;      /* offset into pagecache page */
1093         unsigned int prev_offset;
1094         int error;
1095
1096         index = *ppos >> PAGE_CACHE_SHIFT;
1097         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1098         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1099         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1100         offset = *ppos & ~PAGE_CACHE_MASK;
1101
1102         for (;;) {
1103                 struct page *page;
1104                 pgoff_t end_index;
1105                 loff_t isize;
1106                 unsigned long nr, ret;
1107
1108                 cond_resched();
1109 find_page:
1110                 page = find_get_page(mapping, index);
1111                 if (!page) {
1112                         page_cache_sync_readahead(mapping,
1113                                         ra, filp,
1114                                         index, last_index - index);
1115                         page = find_get_page(mapping, index);
1116                         if (unlikely(page == NULL))
1117                                 goto no_cached_page;
1118                 }
1119                 if (PageReadahead(page)) {
1120                         page_cache_async_readahead(mapping,
1121                                         ra, filp, page,
1122                                         index, last_index - index);
1123                 }
1124                 if (!PageUptodate(page)) {
1125                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1126                                         !mapping->a_ops->is_partially_uptodate)
1127                                 goto page_not_up_to_date;
1128                         if (!trylock_page(page))
1129                                 goto page_not_up_to_date;
1130                         /* Did it get truncated before we got the lock? */
1131                         if (!page->mapping)
1132                                 goto page_not_up_to_date_locked;
1133                         if (!mapping->a_ops->is_partially_uptodate(page,
1134                                                                 desc, offset))
1135                                 goto page_not_up_to_date_locked;
1136                         unlock_page(page);
1137                 }
1138 page_ok:
1139                 /*
1140                  * i_size must be checked after we know the page is Uptodate.
1141                  *
1142                  * Checking i_size after the check allows us to calculate
1143                  * the correct value for "nr", which means the zero-filled
1144                  * part of the page is not copied back to userspace (unless
1145                  * another truncate extends the file - this is desired though).
1146                  */
1147
1148                 isize = i_size_read(inode);
1149                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1150                 if (unlikely(!isize || index > end_index)) {
1151                         page_cache_release(page);
1152                         goto out;
1153                 }
1154
1155                 /* nr is the maximum number of bytes to copy from this page */
1156                 nr = PAGE_CACHE_SIZE;
1157                 if (index == end_index) {
1158                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1159                         if (nr <= offset) {
1160                                 page_cache_release(page);
1161                                 goto out;
1162                         }
1163                 }
1164                 nr = nr - offset;
1165
1166                 /* If users can be writing to this page using arbitrary
1167                  * virtual addresses, take care about potential aliasing
1168                  * before reading the page on the kernel side.
1169                  */
1170                 if (mapping_writably_mapped(mapping))
1171                         flush_dcache_page(page);
1172
1173                 /*
1174                  * When a sequential read accesses a page several times,
1175                  * only mark it as accessed the first time.
1176                  */
1177                 if (prev_index != index || offset != prev_offset)
1178                         mark_page_accessed(page);
1179                 prev_index = index;
1180
1181                 /*
1182                  * Ok, we have the page, and it's up-to-date, so
1183                  * now we can copy it to user space...
1184                  *
1185                  * The actor routine returns how many bytes were actually used..
1186                  * NOTE! This may not be the same as how much of a user buffer
1187                  * we filled up (we may be padding etc), so we can only update
1188                  * "pos" here (the actor routine has to update the user buffer
1189                  * pointers and the remaining count).
1190                  */
1191                 ret = actor(desc, page, offset, nr);
1192                 offset += ret;
1193                 index += offset >> PAGE_CACHE_SHIFT;
1194                 offset &= ~PAGE_CACHE_MASK;
1195                 prev_offset = offset;
1196
1197                 page_cache_release(page);
1198                 if (ret == nr && desc->count)
1199                         continue;
1200                 goto out;
1201
1202 page_not_up_to_date:
1203                 /* Get exclusive access to the page ... */
1204                 error = lock_page_killable(page);
1205                 if (unlikely(error))
1206                         goto readpage_error;
1207
1208 page_not_up_to_date_locked:
1209                 /* Did it get truncated before we got the lock? */
1210                 if (!page->mapping) {
1211                         unlock_page(page);
1212                         page_cache_release(page);
1213                         continue;
1214                 }
1215
1216                 /* Did somebody else fill it already? */
1217                 if (PageUptodate(page)) {
1218                         unlock_page(page);
1219                         goto page_ok;
1220                 }
1221
1222 readpage:
1223                 /*
1224                  * A previous I/O error may have been due to temporary
1225                  * failures, eg. multipath errors.
1226                  * PG_error will be set again if readpage fails.
1227                  */
1228                 ClearPageError(page);
1229                 /* Start the actual read. The read will unlock the page. */
1230                 error = mapping->a_ops->readpage(filp, page);
1231
1232                 if (unlikely(error)) {
1233                         if (error == AOP_TRUNCATED_PAGE) {
1234                                 page_cache_release(page);
1235                                 goto find_page;
1236                         }
1237                         goto readpage_error;
1238                 }
1239
1240                 if (!PageUptodate(page)) {
1241                         error = lock_page_killable(page);
1242                         if (unlikely(error))
1243                                 goto readpage_error;
1244                         if (!PageUptodate(page)) {
1245                                 if (page->mapping == NULL) {
1246                                         /*
1247                                          * invalidate_mapping_pages got it
1248                                          */
1249                                         unlock_page(page);
1250                                         page_cache_release(page);
1251                                         goto find_page;
1252                                 }
1253                                 unlock_page(page);
1254                                 shrink_readahead_size_eio(filp, ra);
1255                                 error = -EIO;
1256                                 goto readpage_error;
1257                         }
1258                         unlock_page(page);
1259                 }
1260
1261                 goto page_ok;
1262
1263 readpage_error:
1264                 /* UHHUH! A synchronous read error occurred. Report it */
1265                 desc->error = error;
1266                 page_cache_release(page);
1267                 goto out;
1268
1269 no_cached_page:
1270                 /*
1271                  * Ok, it wasn't cached, so we need to create a new
1272                  * page..
1273                  */
1274                 page = page_cache_alloc_cold(mapping);
1275                 if (!page) {
1276                         desc->error = -ENOMEM;
1277                         goto out;
1278                 }
1279                 error = add_to_page_cache_lru(page, mapping,
1280                                                 index, GFP_KERNEL);
1281                 if (error) {
1282                         page_cache_release(page);
1283                         if (error == -EEXIST)
1284                                 goto find_page;
1285                         desc->error = error;
1286                         goto out;
1287                 }
1288                 goto readpage;
1289         }
1290
1291 out:
1292         ra->prev_pos = prev_index;
1293         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1294         ra->prev_pos |= prev_offset;
1295
1296         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1297         file_accessed(filp);
1298 }
1299
1300 int file_read_actor(read_descriptor_t *desc, struct page *page,
1301                         unsigned long offset, unsigned long size)
1302 {
1303         char *kaddr;
1304         unsigned long left, count = desc->count;
1305
1306         if (size > count)
1307                 size = count;
1308
1309         /*
1310          * Faults on the destination of a read are common, so do it before
1311          * taking the kmap.
1312          */
1313         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1314                 kaddr = kmap_atomic(page, KM_USER0);
1315                 left = __copy_to_user_inatomic(desc->arg.buf,
1316                                                 kaddr + offset, size);
1317                 kunmap_atomic(kaddr, KM_USER0);
1318                 if (left == 0)
1319                         goto success;
1320         }
1321
1322         /* Do it the slow way */
1323         kaddr = kmap(page);
1324         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1325         kunmap(page);
1326
1327         if (left) {
1328                 size -= left;
1329                 desc->error = -EFAULT;
1330         }
1331 success:
1332         desc->count = count - size;
1333         desc->written += size;
1334         desc->arg.buf += size;
1335         return size;
1336 }
1337
1338 /*
1339  * Performs necessary checks before doing a write
1340  * @iov:        io vector request
1341  * @nr_segs:    number of segments in the iovec
1342  * @count:      number of bytes to write
1343  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1344  *
1345  * Adjust number of segments and amount of bytes to write (nr_segs should be
1346  * properly initialized first). Returns appropriate error code that caller
1347  * should return or zero in case that write should be allowed.
1348  */
1349 int generic_segment_checks(const struct iovec *iov,
1350                         unsigned long *nr_segs, size_t *count, int access_flags)
1351 {
1352         unsigned long   seg;
1353         size_t cnt = 0;
1354         for (seg = 0; seg < *nr_segs; seg++) {
1355                 const struct iovec *iv = &iov[seg];
1356
1357                 /*
1358                  * If any segment has a negative length, or the cumulative
1359                  * length ever wraps negative then return -EINVAL.
1360                  */
1361                 cnt += iv->iov_len;
1362                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1363                         return -EINVAL;
1364                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1365                         continue;
1366                 if (seg == 0)
1367                         return -EFAULT;
1368                 *nr_segs = seg;
1369                 cnt -= iv->iov_len;     /* This segment is no good */
1370                 break;
1371         }
1372         *count = cnt;
1373         return 0;
1374 }
1375 EXPORT_SYMBOL(generic_segment_checks);
1376
1377 /**
1378  * generic_file_aio_read - generic filesystem read routine
1379  * @iocb:       kernel I/O control block
1380  * @iov:        io vector request
1381  * @nr_segs:    number of segments in the iovec
1382  * @pos:        current file position
1383  *
1384  * This is the "read()" routine for all filesystems
1385  * that can use the page cache directly.
1386  */
1387 ssize_t
1388 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1389                 unsigned long nr_segs, loff_t pos)
1390 {
1391         struct file *filp = iocb->ki_filp;
1392         ssize_t retval;
1393         unsigned long seg = 0;
1394         size_t count;
1395         loff_t *ppos = &iocb->ki_pos;
1396         struct blk_plug plug;
1397
1398         count = 0;
1399         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1400         if (retval)
1401                 return retval;
1402
1403         blk_start_plug(&plug);
1404
1405         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1406         if (filp->f_flags & O_DIRECT) {
1407                 loff_t size;
1408                 struct address_space *mapping;
1409                 struct inode *inode;
1410
1411                 mapping = filp->f_mapping;
1412                 inode = mapping->host;
1413                 if (!count)
1414                         goto out; /* skip atime */
1415                 size = i_size_read(inode);
1416                 if (pos < size) {
1417                         retval = filemap_write_and_wait_range(mapping, pos,
1418                                         pos + iov_length(iov, nr_segs) - 1);
1419                         if (!retval) {
1420                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1421                                                         iov, pos, nr_segs);
1422                         }
1423                         if (retval > 0) {
1424                                 *ppos = pos + retval;
1425                                 count -= retval;
1426                         }
1427
1428                         /*
1429                          * Btrfs can have a short DIO read if we encounter
1430                          * compressed extents, so if there was an error, or if
1431                          * we've already read everything we wanted to, or if
1432                          * there was a short read because we hit EOF, go ahead
1433                          * and return.  Otherwise fallthrough to buffered io for
1434                          * the rest of the read.
1435                          */
1436                         if (retval < 0 || !count || *ppos >= size) {
1437                                 file_accessed(filp);
1438                                 goto out;
1439                         }
1440                 }
1441         }
1442
1443         count = retval;
1444         for (seg = 0; seg < nr_segs; seg++) {
1445                 read_descriptor_t desc;
1446                 loff_t offset = 0;
1447
1448                 /*
1449                  * If we did a short DIO read we need to skip the section of the
1450                  * iov that we've already read data into.
1451                  */
1452                 if (count) {
1453                         if (count > iov[seg].iov_len) {
1454                                 count -= iov[seg].iov_len;
1455                                 continue;
1456                         }
1457                         offset = count;
1458                         count = 0;
1459                 }
1460
1461                 desc.written = 0;
1462                 desc.arg.buf = iov[seg].iov_base + offset;
1463                 desc.count = iov[seg].iov_len - offset;
1464                 if (desc.count == 0)
1465                         continue;
1466                 desc.error = 0;
1467                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1468                 retval += desc.written;
1469                 if (desc.error) {
1470                         retval = retval ?: desc.error;
1471                         break;
1472                 }
1473                 if (desc.count > 0)
1474                         break;
1475         }
1476 out:
1477         blk_finish_plug(&plug);
1478         return retval;
1479 }
1480 EXPORT_SYMBOL(generic_file_aio_read);
1481
1482 static ssize_t
1483 do_readahead(struct address_space *mapping, struct file *filp,
1484              pgoff_t index, unsigned long nr)
1485 {
1486         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1487                 return -EINVAL;
1488
1489         force_page_cache_readahead(mapping, filp, index, nr);
1490         return 0;
1491 }
1492
1493 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1494 {
1495         ssize_t ret;
1496         struct file *file;
1497
1498         ret = -EBADF;
1499         file = fget(fd);
1500         if (file) {
1501                 if (file->f_mode & FMODE_READ) {
1502                         struct address_space *mapping = file->f_mapping;
1503                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1504                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1505                         unsigned long len = end - start + 1;
1506                         ret = do_readahead(mapping, file, start, len);
1507                 }
1508                 fput(file);
1509         }
1510         return ret;
1511 }
1512 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1513 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1514 {
1515         return SYSC_readahead((int) fd, offset, (size_t) count);
1516 }
1517 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1518 #endif
1519
1520 #ifdef CONFIG_MMU
1521 /**
1522  * page_cache_read - adds requested page to the page cache if not already there
1523  * @file:       file to read
1524  * @offset:     page index
1525  *
1526  * This adds the requested page to the page cache if it isn't already there,
1527  * and schedules an I/O to read in its contents from disk.
1528  */
1529 static int page_cache_read(struct file *file, pgoff_t offset)
1530 {
1531         struct address_space *mapping = file->f_mapping;
1532         struct page *page; 
1533         int ret;
1534
1535         do {
1536                 page = page_cache_alloc_cold(mapping);
1537                 if (!page)
1538                         return -ENOMEM;
1539
1540                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1541                 if (ret == 0)
1542                         ret = mapping->a_ops->readpage(file, page);
1543                 else if (ret == -EEXIST)
1544                         ret = 0; /* losing race to add is OK */
1545
1546                 page_cache_release(page);
1547
1548         } while (ret == AOP_TRUNCATED_PAGE);
1549                 
1550         return ret;
1551 }
1552
1553 #define MMAP_LOTSAMISS  (100)
1554
1555 /*
1556  * Synchronous readahead happens when we don't even find
1557  * a page in the page cache at all.
1558  */
1559 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1560                                    struct file_ra_state *ra,
1561                                    struct file *file,
1562                                    pgoff_t offset)
1563 {
1564         unsigned long ra_pages;
1565         struct address_space *mapping = file->f_mapping;
1566
1567         /* If we don't want any read-ahead, don't bother */
1568         if (VM_RandomReadHint(vma))
1569                 return;
1570         if (!ra->ra_pages)
1571                 return;
1572
1573         if (VM_SequentialReadHint(vma)) {
1574                 page_cache_sync_readahead(mapping, ra, file, offset,
1575                                           ra->ra_pages);
1576                 return;
1577         }
1578
1579         /* Avoid banging the cache line if not needed */
1580         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1581                 ra->mmap_miss++;
1582
1583         /*
1584          * Do we miss much more than hit in this file? If so,
1585          * stop bothering with read-ahead. It will only hurt.
1586          */
1587         if (ra->mmap_miss > MMAP_LOTSAMISS)
1588                 return;
1589
1590         /*
1591          * mmap read-around
1592          */
1593         ra_pages = max_sane_readahead(ra->ra_pages);
1594         ra->start = max_t(long, 0, offset - ra_pages / 2);
1595         ra->size = ra_pages;
1596         ra->async_size = ra_pages / 4;
1597         ra_submit(ra, mapping, file);
1598 }
1599
1600 /*
1601  * Asynchronous readahead happens when we find the page and PG_readahead,
1602  * so we want to possibly extend the readahead further..
1603  */
1604 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1605                                     struct file_ra_state *ra,
1606                                     struct file *file,
1607                                     struct page *page,
1608                                     pgoff_t offset)
1609 {
1610         struct address_space *mapping = file->f_mapping;
1611
1612         /* If we don't want any read-ahead, don't bother */
1613         if (VM_RandomReadHint(vma))
1614                 return;
1615         if (ra->mmap_miss > 0)
1616                 ra->mmap_miss--;
1617         if (PageReadahead(page))
1618                 page_cache_async_readahead(mapping, ra, file,
1619                                            page, offset, ra->ra_pages);
1620 }
1621
1622 /**
1623  * filemap_fault - read in file data for page fault handling
1624  * @vma:        vma in which the fault was taken
1625  * @vmf:        struct vm_fault containing details of the fault
1626  *
1627  * filemap_fault() is invoked via the vma operations vector for a
1628  * mapped memory region to read in file data during a page fault.
1629  *
1630  * The goto's are kind of ugly, but this streamlines the normal case of having
1631  * it in the page cache, and handles the special cases reasonably without
1632  * having a lot of duplicated code.
1633  */
1634 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1635 {
1636         int error;
1637         struct file *file = vma->vm_file;
1638         struct address_space *mapping = file->f_mapping;
1639         struct file_ra_state *ra = &file->f_ra;
1640         struct inode *inode = mapping->host;
1641         pgoff_t offset = vmf->pgoff;
1642         struct page *page;
1643         pgoff_t size;
1644         int ret = 0;
1645
1646         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1647         if (offset >= size)
1648                 return VM_FAULT_SIGBUS;
1649
1650         /*
1651          * Do we have something in the page cache already?
1652          */
1653         page = find_get_page(mapping, offset);
1654         if (likely(page)) {
1655                 /*
1656                  * We found the page, so try async readahead before
1657                  * waiting for the lock.
1658                  */
1659                 do_async_mmap_readahead(vma, ra, file, page, offset);
1660         } else {
1661                 /* No page in the page cache at all */
1662                 do_sync_mmap_readahead(vma, ra, file, offset);
1663                 count_vm_event(PGMAJFAULT);
1664                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1665                 ret = VM_FAULT_MAJOR;
1666 retry_find:
1667                 page = find_get_page(mapping, offset);
1668                 if (!page)
1669                         goto no_cached_page;
1670         }
1671
1672         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1673                 page_cache_release(page);
1674                 return ret | VM_FAULT_RETRY;
1675         }
1676
1677         /* Did it get truncated? */
1678         if (unlikely(page->mapping != mapping)) {
1679                 unlock_page(page);
1680                 put_page(page);
1681                 goto retry_find;
1682         }
1683         VM_BUG_ON(page->index != offset);
1684
1685         /*
1686          * We have a locked page in the page cache, now we need to check
1687          * that it's up-to-date. If not, it is going to be due to an error.
1688          */
1689         if (unlikely(!PageUptodate(page)))
1690                 goto page_not_uptodate;
1691
1692         /*
1693          * Found the page and have a reference on it.
1694          * We must recheck i_size under page lock.
1695          */
1696         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1697         if (unlikely(offset >= size)) {
1698                 unlock_page(page);
1699                 page_cache_release(page);
1700                 return VM_FAULT_SIGBUS;
1701         }
1702
1703         vmf->page = page;
1704         return ret | VM_FAULT_LOCKED;
1705
1706 no_cached_page:
1707         /*
1708          * We're only likely to ever get here if MADV_RANDOM is in
1709          * effect.
1710          */
1711         error = page_cache_read(file, offset);
1712
1713         /*
1714          * The page we want has now been added to the page cache.
1715          * In the unlikely event that someone removed it in the
1716          * meantime, we'll just come back here and read it again.
1717          */
1718         if (error >= 0)
1719                 goto retry_find;
1720
1721         /*
1722          * An error return from page_cache_read can result if the
1723          * system is low on memory, or a problem occurs while trying
1724          * to schedule I/O.
1725          */
1726         if (error == -ENOMEM)
1727                 return VM_FAULT_OOM;
1728         return VM_FAULT_SIGBUS;
1729
1730 page_not_uptodate:
1731         /*
1732          * Umm, take care of errors if the page isn't up-to-date.
1733          * Try to re-read it _once_. We do this synchronously,
1734          * because there really aren't any performance issues here
1735          * and we need to check for errors.
1736          */
1737         ClearPageError(page);
1738         error = mapping->a_ops->readpage(file, page);
1739         if (!error) {
1740                 wait_on_page_locked(page);
1741                 if (!PageUptodate(page))
1742                         error = -EIO;
1743         }
1744         page_cache_release(page);
1745
1746         if (!error || error == AOP_TRUNCATED_PAGE)
1747                 goto retry_find;
1748
1749         /* Things didn't work out. Return zero to tell the mm layer so. */
1750         shrink_readahead_size_eio(file, ra);
1751         return VM_FAULT_SIGBUS;
1752 }
1753 EXPORT_SYMBOL(filemap_fault);
1754
1755 const struct vm_operations_struct generic_file_vm_ops = {
1756         .fault          = filemap_fault,
1757 };
1758
1759 /* This is used for a general mmap of a disk file */
1760
1761 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1762 {
1763         struct address_space *mapping = file->f_mapping;
1764
1765         if (!mapping->a_ops->readpage)
1766                 return -ENOEXEC;
1767         file_accessed(file);
1768         vma->vm_ops = &generic_file_vm_ops;
1769         vma->vm_flags |= VM_CAN_NONLINEAR;
1770         return 0;
1771 }
1772
1773 /*
1774  * This is for filesystems which do not implement ->writepage.
1775  */
1776 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1777 {
1778         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1779                 return -EINVAL;
1780         return generic_file_mmap(file, vma);
1781 }
1782 #else
1783 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1784 {
1785         return -ENOSYS;
1786 }
1787 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1788 {
1789         return -ENOSYS;
1790 }
1791 #endif /* CONFIG_MMU */
1792
1793 EXPORT_SYMBOL(generic_file_mmap);
1794 EXPORT_SYMBOL(generic_file_readonly_mmap);
1795
1796 static struct page *__read_cache_page(struct address_space *mapping,
1797                                 pgoff_t index,
1798                                 int (*filler)(void *,struct page*),
1799                                 void *data,
1800                                 gfp_t gfp)
1801 {
1802         struct page *page;
1803         int err;
1804 repeat:
1805         page = find_get_page(mapping, index);
1806         if (!page) {
1807                 page = __page_cache_alloc(gfp | __GFP_COLD);
1808                 if (!page)
1809                         return ERR_PTR(-ENOMEM);
1810                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1811                 if (unlikely(err)) {
1812                         page_cache_release(page);
1813                         if (err == -EEXIST)
1814                                 goto repeat;
1815                         /* Presumably ENOMEM for radix tree node */
1816                         return ERR_PTR(err);
1817                 }
1818                 err = filler(data, page);
1819                 if (err < 0) {
1820                         page_cache_release(page);
1821                         page = ERR_PTR(err);
1822                 }
1823         }
1824         return page;
1825 }
1826
1827 static struct page *do_read_cache_page(struct address_space *mapping,
1828                                 pgoff_t index,
1829                                 int (*filler)(void *,struct page*),
1830                                 void *data,
1831                                 gfp_t gfp)
1832
1833 {
1834         struct page *page;
1835         int err;
1836
1837 retry:
1838         page = __read_cache_page(mapping, index, filler, data, gfp);
1839         if (IS_ERR(page))
1840                 return page;
1841         if (PageUptodate(page))
1842                 goto out;
1843
1844         lock_page(page);
1845         if (!page->mapping) {
1846                 unlock_page(page);
1847                 page_cache_release(page);
1848                 goto retry;
1849         }
1850         if (PageUptodate(page)) {
1851                 unlock_page(page);
1852                 goto out;
1853         }
1854         err = filler(data, page);
1855         if (err < 0) {
1856                 page_cache_release(page);
1857                 return ERR_PTR(err);
1858         }
1859 out:
1860         mark_page_accessed(page);
1861         return page;
1862 }
1863
1864 /**
1865  * read_cache_page_async - read into page cache, fill it if needed
1866  * @mapping:    the page's address_space
1867  * @index:      the page index
1868  * @filler:     function to perform the read
1869  * @data:       destination for read data
1870  *
1871  * Same as read_cache_page, but don't wait for page to become unlocked
1872  * after submitting it to the filler.
1873  *
1874  * Read into the page cache. If a page already exists, and PageUptodate() is
1875  * not set, try to fill the page but don't wait for it to become unlocked.
1876  *
1877  * If the page does not get brought uptodate, return -EIO.
1878  */
1879 struct page *read_cache_page_async(struct address_space *mapping,
1880                                 pgoff_t index,
1881                                 int (*filler)(void *,struct page*),
1882                                 void *data)
1883 {
1884         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1885 }
1886 EXPORT_SYMBOL(read_cache_page_async);
1887
1888 static struct page *wait_on_page_read(struct page *page)
1889 {
1890         if (!IS_ERR(page)) {
1891                 wait_on_page_locked(page);
1892                 if (!PageUptodate(page)) {
1893                         page_cache_release(page);
1894                         page = ERR_PTR(-EIO);
1895                 }
1896         }
1897         return page;
1898 }
1899
1900 /**
1901  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1902  * @mapping:    the page's address_space
1903  * @index:      the page index
1904  * @gfp:        the page allocator flags to use if allocating
1905  *
1906  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1907  * any new page allocations done using the specified allocation flags. Note
1908  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1909  * expect to do this atomically or anything like that - but you can pass in
1910  * other page requirements.
1911  *
1912  * If the page does not get brought uptodate, return -EIO.
1913  */
1914 struct page *read_cache_page_gfp(struct address_space *mapping,
1915                                 pgoff_t index,
1916                                 gfp_t gfp)
1917 {
1918         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1919
1920         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1921 }
1922 EXPORT_SYMBOL(read_cache_page_gfp);
1923
1924 /**
1925  * read_cache_page - read into page cache, fill it if needed
1926  * @mapping:    the page's address_space
1927  * @index:      the page index
1928  * @filler:     function to perform the read
1929  * @data:       destination for read data
1930  *
1931  * Read into the page cache. If a page already exists, and PageUptodate() is
1932  * not set, try to fill the page then wait for it to become unlocked.
1933  *
1934  * If the page does not get brought uptodate, return -EIO.
1935  */
1936 struct page *read_cache_page(struct address_space *mapping,
1937                                 pgoff_t index,
1938                                 int (*filler)(void *,struct page*),
1939                                 void *data)
1940 {
1941         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1942 }
1943 EXPORT_SYMBOL(read_cache_page);
1944
1945 /*
1946  * The logic we want is
1947  *
1948  *      if suid or (sgid and xgrp)
1949  *              remove privs
1950  */
1951 int should_remove_suid(struct dentry *dentry)
1952 {
1953         mode_t mode = dentry->d_inode->i_mode;
1954         int kill = 0;
1955
1956         /* suid always must be killed */
1957         if (unlikely(mode & S_ISUID))
1958                 kill = ATTR_KILL_SUID;
1959
1960         /*
1961          * sgid without any exec bits is just a mandatory locking mark; leave
1962          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1963          */
1964         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1965                 kill |= ATTR_KILL_SGID;
1966
1967         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1968                 return kill;
1969
1970         return 0;
1971 }
1972 EXPORT_SYMBOL(should_remove_suid);
1973
1974 static int __remove_suid(struct dentry *dentry, int kill)
1975 {
1976         struct iattr newattrs;
1977
1978         newattrs.ia_valid = ATTR_FORCE | kill;
1979         return notify_change(dentry, &newattrs);
1980 }
1981
1982 int file_remove_suid(struct file *file)
1983 {
1984         struct dentry *dentry = file->f_path.dentry;
1985         struct inode *inode = dentry->d_inode;
1986         int killsuid;
1987         int killpriv;
1988         int error = 0;
1989
1990         /* Fast path for nothing security related */
1991         if (IS_NOSEC(inode))
1992                 return 0;
1993
1994         killsuid = should_remove_suid(dentry);
1995         killpriv = security_inode_need_killpriv(dentry);
1996
1997         if (killpriv < 0)
1998                 return killpriv;
1999         if (killpriv)
2000                 error = security_inode_killpriv(dentry);
2001         if (!error && killsuid)
2002                 error = __remove_suid(dentry, killsuid);
2003         if (!error)
2004                 inode->i_flags |= S_NOSEC;
2005
2006         return error;
2007 }
2008 EXPORT_SYMBOL(file_remove_suid);
2009
2010 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2011                         const struct iovec *iov, size_t base, size_t bytes)
2012 {
2013         size_t copied = 0, left = 0;
2014
2015         while (bytes) {
2016                 char __user *buf = iov->iov_base + base;
2017                 int copy = min(bytes, iov->iov_len - base);
2018
2019                 base = 0;
2020                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2021                 copied += copy;
2022                 bytes -= copy;
2023                 vaddr += copy;
2024                 iov++;
2025
2026                 if (unlikely(left))
2027                         break;
2028         }
2029         return copied - left;
2030 }
2031
2032 /*
2033  * Copy as much as we can into the page and return the number of bytes which
2034  * were successfully copied.  If a fault is encountered then return the number of
2035  * bytes which were copied.
2036  */
2037 size_t iov_iter_copy_from_user_atomic(struct page *page,
2038                 struct iov_iter *i, unsigned long offset, size_t bytes)
2039 {
2040         char *kaddr;
2041         size_t copied;
2042
2043         BUG_ON(!in_atomic());
2044         kaddr = kmap_atomic(page, KM_USER0);
2045         if (likely(i->nr_segs == 1)) {
2046                 int left;
2047                 char __user *buf = i->iov->iov_base + i->iov_offset;
2048                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2049                 copied = bytes - left;
2050         } else {
2051                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2052                                                 i->iov, i->iov_offset, bytes);
2053         }
2054         kunmap_atomic(kaddr, KM_USER0);
2055
2056         return copied;
2057 }
2058 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2059
2060 /*
2061  * This has the same sideeffects and return value as
2062  * iov_iter_copy_from_user_atomic().
2063  * The difference is that it attempts to resolve faults.
2064  * Page must not be locked.
2065  */
2066 size_t iov_iter_copy_from_user(struct page *page,
2067                 struct iov_iter *i, unsigned long offset, size_t bytes)
2068 {
2069         char *kaddr;
2070         size_t copied;
2071
2072         kaddr = kmap(page);
2073         if (likely(i->nr_segs == 1)) {
2074                 int left;
2075                 char __user *buf = i->iov->iov_base + i->iov_offset;
2076                 left = __copy_from_user(kaddr + offset, buf, bytes);
2077                 copied = bytes - left;
2078         } else {
2079                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2080                                                 i->iov, i->iov_offset, bytes);
2081         }
2082         kunmap(page);
2083         return copied;
2084 }
2085 EXPORT_SYMBOL(iov_iter_copy_from_user);
2086
2087 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2088 {
2089         BUG_ON(i->count < bytes);
2090
2091         if (likely(i->nr_segs == 1)) {
2092                 i->iov_offset += bytes;
2093                 i->count -= bytes;
2094         } else {
2095                 const struct iovec *iov = i->iov;
2096                 size_t base = i->iov_offset;
2097
2098                 /*
2099                  * The !iov->iov_len check ensures we skip over unlikely
2100                  * zero-length segments (without overruning the iovec).
2101                  */
2102                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2103                         int copy;
2104
2105                         copy = min(bytes, iov->iov_len - base);
2106                         BUG_ON(!i->count || i->count < copy);
2107                         i->count -= copy;
2108                         bytes -= copy;
2109                         base += copy;
2110                         if (iov->iov_len == base) {
2111                                 iov++;
2112                                 base = 0;
2113                         }
2114                 }
2115                 i->iov = iov;
2116                 i->iov_offset = base;
2117         }
2118 }
2119 EXPORT_SYMBOL(iov_iter_advance);
2120
2121 /*
2122  * Fault in the first iovec of the given iov_iter, to a maximum length
2123  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2124  * accessed (ie. because it is an invalid address).
2125  *
2126  * writev-intensive code may want this to prefault several iovecs -- that
2127  * would be possible (callers must not rely on the fact that _only_ the
2128  * first iovec will be faulted with the current implementation).
2129  */
2130 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2131 {
2132         char __user *buf = i->iov->iov_base + i->iov_offset;
2133         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2134         return fault_in_pages_readable(buf, bytes);
2135 }
2136 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2137
2138 /*
2139  * Return the count of just the current iov_iter segment.
2140  */
2141 size_t iov_iter_single_seg_count(struct iov_iter *i)
2142 {
2143         const struct iovec *iov = i->iov;
2144         if (i->nr_segs == 1)
2145                 return i->count;
2146         else
2147                 return min(i->count, iov->iov_len - i->iov_offset);
2148 }
2149 EXPORT_SYMBOL(iov_iter_single_seg_count);
2150
2151 /*
2152  * Performs necessary checks before doing a write
2153  *
2154  * Can adjust writing position or amount of bytes to write.
2155  * Returns appropriate error code that caller should return or
2156  * zero in case that write should be allowed.
2157  */
2158 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2159 {
2160         struct inode *inode = file->f_mapping->host;
2161         unsigned long limit = rlimit(RLIMIT_FSIZE);
2162
2163         if (unlikely(*pos < 0))
2164                 return -EINVAL;
2165
2166         if (!isblk) {
2167                 /* FIXME: this is for backwards compatibility with 2.4 */
2168                 if (file->f_flags & O_APPEND)
2169                         *pos = i_size_read(inode);
2170
2171                 if (limit != RLIM_INFINITY) {
2172                         if (*pos >= limit) {
2173                                 send_sig(SIGXFSZ, current, 0);
2174                                 return -EFBIG;
2175                         }
2176                         if (*count > limit - (typeof(limit))*pos) {
2177                                 *count = limit - (typeof(limit))*pos;
2178                         }
2179                 }
2180         }
2181
2182         /*
2183          * LFS rule
2184          */
2185         if (unlikely(*pos + *count > MAX_NON_LFS &&
2186                                 !(file->f_flags & O_LARGEFILE))) {
2187                 if (*pos >= MAX_NON_LFS) {
2188                         return -EFBIG;
2189                 }
2190                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2191                         *count = MAX_NON_LFS - (unsigned long)*pos;
2192                 }
2193         }
2194
2195         /*
2196          * Are we about to exceed the fs block limit ?
2197          *
2198          * If we have written data it becomes a short write.  If we have
2199          * exceeded without writing data we send a signal and return EFBIG.
2200          * Linus frestrict idea will clean these up nicely..
2201          */
2202         if (likely(!isblk)) {
2203                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2204                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2205                                 return -EFBIG;
2206                         }
2207                         /* zero-length writes at ->s_maxbytes are OK */
2208                 }
2209
2210                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2211                         *count = inode->i_sb->s_maxbytes - *pos;
2212         } else {
2213 #ifdef CONFIG_BLOCK
2214                 loff_t isize;
2215                 if (bdev_read_only(I_BDEV(inode)))
2216                         return -EPERM;
2217                 isize = i_size_read(inode);
2218                 if (*pos >= isize) {
2219                         if (*count || *pos > isize)
2220                                 return -ENOSPC;
2221                 }
2222
2223                 if (*pos + *count > isize)
2224                         *count = isize - *pos;
2225 #else
2226                 return -EPERM;
2227 #endif
2228         }
2229         return 0;
2230 }
2231 EXPORT_SYMBOL(generic_write_checks);
2232
2233 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2234                                 loff_t pos, unsigned len, unsigned flags,
2235                                 struct page **pagep, void **fsdata)
2236 {
2237         const struct address_space_operations *aops = mapping->a_ops;
2238
2239         return aops->write_begin(file, mapping, pos, len, flags,
2240                                                         pagep, fsdata);
2241 }
2242 EXPORT_SYMBOL(pagecache_write_begin);
2243
2244 int pagecache_write_end(struct file *file, struct address_space *mapping,
2245                                 loff_t pos, unsigned len, unsigned copied,
2246                                 struct page *page, void *fsdata)
2247 {
2248         const struct address_space_operations *aops = mapping->a_ops;
2249
2250         mark_page_accessed(page);
2251         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2252 }
2253 EXPORT_SYMBOL(pagecache_write_end);
2254
2255 ssize_t
2256 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2257                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2258                 size_t count, size_t ocount)
2259 {
2260         struct file     *file = iocb->ki_filp;
2261         struct address_space *mapping = file->f_mapping;
2262         struct inode    *inode = mapping->host;
2263         ssize_t         written;
2264         size_t          write_len;
2265         pgoff_t         end;
2266
2267         if (count != ocount)
2268                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2269
2270         write_len = iov_length(iov, *nr_segs);
2271         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2272
2273         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2274         if (written)
2275                 goto out;
2276
2277         /*
2278          * After a write we want buffered reads to be sure to go to disk to get
2279          * the new data.  We invalidate clean cached page from the region we're
2280          * about to write.  We do this *before* the write so that we can return
2281          * without clobbering -EIOCBQUEUED from ->direct_IO().
2282          */
2283         if (mapping->nrpages) {
2284                 written = invalidate_inode_pages2_range(mapping,
2285                                         pos >> PAGE_CACHE_SHIFT, end);
2286                 /*
2287                  * If a page can not be invalidated, return 0 to fall back
2288                  * to buffered write.
2289                  */
2290                 if (written) {
2291                         if (written == -EBUSY)
2292                                 return 0;
2293                         goto out;
2294                 }
2295         }
2296
2297         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2298
2299         /*
2300          * Finally, try again to invalidate clean pages which might have been
2301          * cached by non-direct readahead, or faulted in by get_user_pages()
2302          * if the source of the write was an mmap'ed region of the file
2303          * we're writing.  Either one is a pretty crazy thing to do,
2304          * so we don't support it 100%.  If this invalidation
2305          * fails, tough, the write still worked...
2306          */
2307         if (mapping->nrpages) {
2308                 invalidate_inode_pages2_range(mapping,
2309                                               pos >> PAGE_CACHE_SHIFT, end);
2310         }
2311
2312         if (written > 0) {
2313                 pos += written;
2314                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2315                         i_size_write(inode, pos);
2316                         mark_inode_dirty(inode);
2317                 }
2318                 *ppos = pos;
2319         }
2320 out:
2321         return written;
2322 }
2323 EXPORT_SYMBOL(generic_file_direct_write);
2324
2325 /*
2326  * Find or create a page at the given pagecache position. Return the locked
2327  * page. This function is specifically for buffered writes.
2328  */
2329 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2330                                         pgoff_t index, unsigned flags)
2331 {
2332         int status;
2333         struct page *page;
2334         gfp_t gfp_notmask = 0;
2335         if (flags & AOP_FLAG_NOFS)
2336                 gfp_notmask = __GFP_FS;
2337 repeat:
2338         page = find_lock_page(mapping, index);
2339         if (page)
2340                 goto found;
2341
2342         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2343         if (!page)
2344                 return NULL;
2345         status = add_to_page_cache_lru(page, mapping, index,
2346                                                 GFP_KERNEL & ~gfp_notmask);
2347         if (unlikely(status)) {
2348                 page_cache_release(page);
2349                 if (status == -EEXIST)
2350                         goto repeat;
2351                 return NULL;
2352         }
2353 found:
2354         wait_on_page_writeback(page);
2355         return page;
2356 }
2357 EXPORT_SYMBOL(grab_cache_page_write_begin);
2358
2359 static ssize_t generic_perform_write(struct file *file,
2360                                 struct iov_iter *i, loff_t pos)
2361 {
2362         struct address_space *mapping = file->f_mapping;
2363         const struct address_space_operations *a_ops = mapping->a_ops;
2364         long status = 0;
2365         ssize_t written = 0;
2366         unsigned int flags = 0;
2367
2368         /*
2369          * Copies from kernel address space cannot fail (NFSD is a big user).
2370          */
2371         if (segment_eq(get_fs(), KERNEL_DS))
2372                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2373
2374         do {
2375                 struct page *page;
2376                 unsigned long offset;   /* Offset into pagecache page */
2377                 unsigned long bytes;    /* Bytes to write to page */
2378                 size_t copied;          /* Bytes copied from user */
2379                 void *fsdata;
2380
2381                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2382                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2383                                                 iov_iter_count(i));
2384
2385 again:
2386
2387                 /*
2388                  * Bring in the user page that we will copy from _first_.
2389                  * Otherwise there's a nasty deadlock on copying from the
2390                  * same page as we're writing to, without it being marked
2391                  * up-to-date.
2392                  *
2393                  * Not only is this an optimisation, but it is also required
2394                  * to check that the address is actually valid, when atomic
2395                  * usercopies are used, below.
2396                  */
2397                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2398                         status = -EFAULT;
2399                         break;
2400                 }
2401
2402                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2403                                                 &page, &fsdata);
2404                 if (unlikely(status))
2405                         break;
2406
2407                 if (mapping_writably_mapped(mapping))
2408                         flush_dcache_page(page);
2409
2410                 pagefault_disable();
2411                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2412                 pagefault_enable();
2413                 flush_dcache_page(page);
2414
2415                 mark_page_accessed(page);
2416                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2417                                                 page, fsdata);
2418                 if (unlikely(status < 0))
2419                         break;
2420                 copied = status;
2421
2422                 cond_resched();
2423
2424                 iov_iter_advance(i, copied);
2425                 if (unlikely(copied == 0)) {
2426                         /*
2427                          * If we were unable to copy any data at all, we must
2428                          * fall back to a single segment length write.
2429                          *
2430                          * If we didn't fallback here, we could livelock
2431                          * because not all segments in the iov can be copied at
2432                          * once without a pagefault.
2433                          */
2434                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2435                                                 iov_iter_single_seg_count(i));
2436                         goto again;
2437                 }
2438                 pos += copied;
2439                 written += copied;
2440
2441                 balance_dirty_pages_ratelimited(mapping);
2442
2443         } while (iov_iter_count(i));
2444
2445         return written ? written : status;
2446 }
2447
2448 ssize_t
2449 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2450                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2451                 size_t count, ssize_t written)
2452 {
2453         struct file *file = iocb->ki_filp;
2454         ssize_t status;
2455         struct iov_iter i;
2456
2457         iov_iter_init(&i, iov, nr_segs, count, written);
2458         status = generic_perform_write(file, &i, pos);
2459
2460         if (likely(status >= 0)) {
2461                 written += status;
2462                 *ppos = pos + status;
2463         }
2464         
2465         return written ? written : status;
2466 }
2467 EXPORT_SYMBOL(generic_file_buffered_write);
2468
2469 /**
2470  * __generic_file_aio_write - write data to a file
2471  * @iocb:       IO state structure (file, offset, etc.)
2472  * @iov:        vector with data to write
2473  * @nr_segs:    number of segments in the vector
2474  * @ppos:       position where to write
2475  *
2476  * This function does all the work needed for actually writing data to a
2477  * file. It does all basic checks, removes SUID from the file, updates
2478  * modification times and calls proper subroutines depending on whether we
2479  * do direct IO or a standard buffered write.
2480  *
2481  * It expects i_mutex to be grabbed unless we work on a block device or similar
2482  * object which does not need locking at all.
2483  *
2484  * This function does *not* take care of syncing data in case of O_SYNC write.
2485  * A caller has to handle it. This is mainly due to the fact that we want to
2486  * avoid syncing under i_mutex.
2487  */
2488 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2489                                  unsigned long nr_segs, loff_t *ppos)
2490 {
2491         struct file *file = iocb->ki_filp;
2492         struct address_space * mapping = file->f_mapping;
2493         size_t ocount;          /* original count */
2494         size_t count;           /* after file limit checks */
2495         struct inode    *inode = mapping->host;
2496         loff_t          pos;
2497         ssize_t         written;
2498         ssize_t         err;
2499
2500         ocount = 0;
2501         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2502         if (err)
2503                 return err;
2504
2505         count = ocount;
2506         pos = *ppos;
2507
2508         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2509
2510         /* We can write back this queue in page reclaim */
2511         current->backing_dev_info = mapping->backing_dev_info;
2512         written = 0;
2513
2514         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2515         if (err)
2516                 goto out;
2517
2518         if (count == 0)
2519                 goto out;
2520
2521         err = file_remove_suid(file);
2522         if (err)
2523                 goto out;
2524
2525         file_update_time(file);
2526
2527         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2528         if (unlikely(file->f_flags & O_DIRECT)) {
2529                 loff_t endbyte;
2530                 ssize_t written_buffered;
2531
2532                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2533                                                         ppos, count, ocount);
2534                 if (written < 0 || written == count)
2535                         goto out;
2536                 /*
2537                  * direct-io write to a hole: fall through to buffered I/O
2538                  * for completing the rest of the request.
2539                  */
2540                 pos += written;
2541                 count -= written;
2542                 written_buffered = generic_file_buffered_write(iocb, iov,
2543                                                 nr_segs, pos, ppos, count,
2544                                                 written);
2545                 /*
2546                  * If generic_file_buffered_write() retuned a synchronous error
2547                  * then we want to return the number of bytes which were
2548                  * direct-written, or the error code if that was zero.  Note
2549                  * that this differs from normal direct-io semantics, which
2550                  * will return -EFOO even if some bytes were written.
2551                  */
2552                 if (written_buffered < 0) {
2553                         err = written_buffered;
2554                         goto out;
2555                 }
2556
2557                 /*
2558                  * We need to ensure that the page cache pages are written to
2559                  * disk and invalidated to preserve the expected O_DIRECT
2560                  * semantics.
2561                  */
2562                 endbyte = pos + written_buffered - written - 1;
2563                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2564                 if (err == 0) {
2565                         written = written_buffered;
2566                         invalidate_mapping_pages(mapping,
2567                                                  pos >> PAGE_CACHE_SHIFT,
2568                                                  endbyte >> PAGE_CACHE_SHIFT);
2569                 } else {
2570                         /*
2571                          * We don't know how much we wrote, so just return
2572                          * the number of bytes which were direct-written
2573                          */
2574                 }
2575         } else {
2576                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2577                                 pos, ppos, count, written);
2578         }
2579 out:
2580         current->backing_dev_info = NULL;
2581         return written ? written : err;
2582 }
2583 EXPORT_SYMBOL(__generic_file_aio_write);
2584
2585 /**
2586  * generic_file_aio_write - write data to a file
2587  * @iocb:       IO state structure
2588  * @iov:        vector with data to write
2589  * @nr_segs:    number of segments in the vector
2590  * @pos:        position in file where to write
2591  *
2592  * This is a wrapper around __generic_file_aio_write() to be used by most
2593  * filesystems. It takes care of syncing the file in case of O_SYNC file
2594  * and acquires i_mutex as needed.
2595  */
2596 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2597                 unsigned long nr_segs, loff_t pos)
2598 {
2599         struct file *file = iocb->ki_filp;
2600         struct inode *inode = file->f_mapping->host;
2601         struct blk_plug plug;
2602         ssize_t ret;
2603
2604         BUG_ON(iocb->ki_pos != pos);
2605
2606         mutex_lock(&inode->i_mutex);
2607         blk_start_plug(&plug);
2608         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2609         mutex_unlock(&inode->i_mutex);
2610
2611         if (ret > 0 || ret == -EIOCBQUEUED) {
2612                 ssize_t err;
2613
2614                 err = generic_write_sync(file, pos, ret);
2615                 if (err < 0 && ret > 0)
2616                         ret = err;
2617         }
2618         blk_finish_plug(&plug);
2619         return ret;
2620 }
2621 EXPORT_SYMBOL(generic_file_aio_write);
2622
2623 /**
2624  * try_to_release_page() - release old fs-specific metadata on a page
2625  *
2626  * @page: the page which the kernel is trying to free
2627  * @gfp_mask: memory allocation flags (and I/O mode)
2628  *
2629  * The address_space is to try to release any data against the page
2630  * (presumably at page->private).  If the release was successful, return `1'.
2631  * Otherwise return zero.
2632  *
2633  * This may also be called if PG_fscache is set on a page, indicating that the
2634  * page is known to the local caching routines.
2635  *
2636  * The @gfp_mask argument specifies whether I/O may be performed to release
2637  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2638  *
2639  */
2640 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2641 {
2642         struct address_space * const mapping = page->mapping;
2643
2644         BUG_ON(!PageLocked(page));
2645         if (PageWriteback(page))
2646                 return 0;
2647
2648         if (mapping && mapping->a_ops->releasepage)
2649                 return mapping->a_ops->releasepage(page, gfp_mask);
2650         return try_to_free_buffers(page);
2651 }
2652
2653 EXPORT_SYMBOL(try_to_release_page);