76bfb6460f5799e1dcb99b63138ca2b2e8712ee6
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include <linux/cleancache.h>
38 #include "internal.h"
39
40 /*
41  * FIXME: remove all knowledge of the buffer layer from the core VM
42  */
43 #include <linux/buffer_head.h> /* for try_to_free_buffers */
44
45 #include <asm/mman.h>
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_mutex              (truncate_pagecache)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_mutex
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  bdi->wb.list_lock
82  *    sb_lock                   (fs/fs-writeback.c)
83  *    ->mapping->tree_lock      (__sync_single_inode)
84  *
85  *  ->i_mmap_mutex
86  *    ->anon_vma.lock           (vma_adjust)
87  *
88  *  ->anon_vma.lock
89  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
90  *
91  *  ->page_table_lock or pte_lock
92  *    ->swap_lock               (try_to_unmap_one)
93  *    ->private_lock            (try_to_unmap_one)
94  *    ->tree_lock               (try_to_unmap_one)
95  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
96  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
97  *    ->private_lock            (page_remove_rmap->set_page_dirty)
98  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
99  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
100  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
101  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
102  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  (code doesn't rely on that order, so you could switch it around)
106  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
107  *    ->i_mmap_mutex
108  */
109
110 /*
111  * Delete a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold the mapping's tree_lock.
114  */
115 void __delete_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         /*
120          * if we're uptodate, flush out into the cleancache, otherwise
121          * invalidate any existing cleancache entries.  We can't leave
122          * stale data around in the cleancache once our page is gone
123          */
124         if (PageUptodate(page) && PageMappedToDisk(page))
125                 cleancache_put_page(page);
126         else
127                 cleancache_flush_page(mapping, page);
128
129         radix_tree_delete(&mapping->page_tree, page->index);
130         page->mapping = NULL;
131         /* Leave page->index set: truncation lookup relies upon it */
132         mapping->nrpages--;
133         __dec_zone_page_state(page, NR_FILE_PAGES);
134         if (PageSwapBacked(page))
135                 __dec_zone_page_state(page, NR_SHMEM);
136         BUG_ON(page_mapped(page));
137
138         /*
139          * Some filesystems seem to re-dirty the page even after
140          * the VM has canceled the dirty bit (eg ext3 journaling).
141          *
142          * Fix it up by doing a final dirty accounting check after
143          * having removed the page entirely.
144          */
145         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
146                 dec_zone_page_state(page, NR_FILE_DIRTY);
147                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
148         }
149 }
150
151 /**
152  * delete_from_page_cache - delete page from page cache
153  * @page: the page which the kernel is trying to remove from page cache
154  *
155  * This must be called only on pages that have been verified to be in the page
156  * cache and locked.  It will never put the page into the free list, the caller
157  * has a reference on the page.
158  */
159 void delete_from_page_cache(struct page *page)
160 {
161         struct address_space *mapping = page->mapping;
162         void (*freepage)(struct page *);
163
164         BUG_ON(!PageLocked(page));
165
166         freepage = mapping->a_ops->freepage;
167         spin_lock_irq(&mapping->tree_lock);
168         __delete_from_page_cache(page);
169         spin_unlock_irq(&mapping->tree_lock);
170         mem_cgroup_uncharge_cache_page(page);
171
172         if (freepage)
173                 freepage(page);
174         page_cache_release(page);
175 }
176 EXPORT_SYMBOL(delete_from_page_cache);
177
178 static int sleep_on_page(void *word)
179 {
180         io_schedule();
181         return 0;
182 }
183
184 static int sleep_on_page_killable(void *word)
185 {
186         sleep_on_page(word);
187         return fatal_signal_pending(current) ? -EINTR : 0;
188 }
189
190 /**
191  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
192  * @mapping:    address space structure to write
193  * @start:      offset in bytes where the range starts
194  * @end:        offset in bytes where the range ends (inclusive)
195  * @sync_mode:  enable synchronous operation
196  *
197  * Start writeback against all of a mapping's dirty pages that lie
198  * within the byte offsets <start, end> inclusive.
199  *
200  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
201  * opposed to a regular memory cleansing writeback.  The difference between
202  * these two operations is that if a dirty page/buffer is encountered, it must
203  * be waited upon, and not just skipped over.
204  */
205 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
206                                 loff_t end, int sync_mode)
207 {
208         int ret;
209         struct writeback_control wbc = {
210                 .sync_mode = sync_mode,
211                 .nr_to_write = LONG_MAX,
212                 .range_start = start,
213                 .range_end = end,
214         };
215
216         if (!mapping_cap_writeback_dirty(mapping))
217                 return 0;
218
219         ret = do_writepages(mapping, &wbc);
220         return ret;
221 }
222
223 static inline int __filemap_fdatawrite(struct address_space *mapping,
224         int sync_mode)
225 {
226         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
227 }
228
229 int filemap_fdatawrite(struct address_space *mapping)
230 {
231         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
232 }
233 EXPORT_SYMBOL(filemap_fdatawrite);
234
235 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
236                                 loff_t end)
237 {
238         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
239 }
240 EXPORT_SYMBOL(filemap_fdatawrite_range);
241
242 /**
243  * filemap_flush - mostly a non-blocking flush
244  * @mapping:    target address_space
245  *
246  * This is a mostly non-blocking flush.  Not suitable for data-integrity
247  * purposes - I/O may not be started against all dirty pages.
248  */
249 int filemap_flush(struct address_space *mapping)
250 {
251         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
252 }
253 EXPORT_SYMBOL(filemap_flush);
254
255 /**
256  * filemap_fdatawait_range - wait for writeback to complete
257  * @mapping:            address space structure to wait for
258  * @start_byte:         offset in bytes where the range starts
259  * @end_byte:           offset in bytes where the range ends (inclusive)
260  *
261  * Walk the list of under-writeback pages of the given address space
262  * in the given range and wait for all of them.
263  */
264 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
265                             loff_t end_byte)
266 {
267         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
268         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272
273         if (end_byte < start_byte)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         while ((index <= end) &&
278                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
279                         PAGECACHE_TAG_WRITEBACK,
280                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
281                 unsigned i;
282
283                 for (i = 0; i < nr_pages; i++) {
284                         struct page *page = pvec.pages[i];
285
286                         /* until radix tree lookup accepts end_index */
287                         if (page->index > end)
288                                 continue;
289
290                         wait_on_page_writeback(page);
291                         if (TestClearPageError(page))
292                                 ret = -EIO;
293                 }
294                 pagevec_release(&pvec);
295                 cond_resched();
296         }
297
298         /* Check for outstanding write errors */
299         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
300                 ret = -ENOSPC;
301         if (test_and_clear_bit(AS_EIO, &mapping->flags))
302                 ret = -EIO;
303
304         return ret;
305 }
306 EXPORT_SYMBOL(filemap_fdatawait_range);
307
308 /**
309  * filemap_fdatawait - wait for all under-writeback pages to complete
310  * @mapping: address space structure to wait for
311  *
312  * Walk the list of under-writeback pages of the given address space
313  * and wait for all of them.
314  */
315 int filemap_fdatawait(struct address_space *mapping)
316 {
317         loff_t i_size = i_size_read(mapping->host);
318
319         if (i_size == 0)
320                 return 0;
321
322         return filemap_fdatawait_range(mapping, 0, i_size - 1);
323 }
324 EXPORT_SYMBOL(filemap_fdatawait);
325
326 int filemap_write_and_wait(struct address_space *mapping)
327 {
328         int err = 0;
329
330         if (mapping->nrpages) {
331                 err = filemap_fdatawrite(mapping);
332                 /*
333                  * Even if the above returned error, the pages may be
334                  * written partially (e.g. -ENOSPC), so we wait for it.
335                  * But the -EIO is special case, it may indicate the worst
336                  * thing (e.g. bug) happened, so we avoid waiting for it.
337                  */
338                 if (err != -EIO) {
339                         int err2 = filemap_fdatawait(mapping);
340                         if (!err)
341                                 err = err2;
342                 }
343         }
344         return err;
345 }
346 EXPORT_SYMBOL(filemap_write_and_wait);
347
348 /**
349  * filemap_write_and_wait_range - write out & wait on a file range
350  * @mapping:    the address_space for the pages
351  * @lstart:     offset in bytes where the range starts
352  * @lend:       offset in bytes where the range ends (inclusive)
353  *
354  * Write out and wait upon file offsets lstart->lend, inclusive.
355  *
356  * Note that `lend' is inclusive (describes the last byte to be written) so
357  * that this function can be used to write to the very end-of-file (end = -1).
358  */
359 int filemap_write_and_wait_range(struct address_space *mapping,
360                                  loff_t lstart, loff_t lend)
361 {
362         int err = 0;
363
364         if (mapping->nrpages) {
365                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
366                                                  WB_SYNC_ALL);
367                 /* See comment of filemap_write_and_wait() */
368                 if (err != -EIO) {
369                         int err2 = filemap_fdatawait_range(mapping,
370                                                 lstart, lend);
371                         if (!err)
372                                 err = err2;
373                 }
374         }
375         return err;
376 }
377 EXPORT_SYMBOL(filemap_write_and_wait_range);
378
379 /**
380  * replace_page_cache_page - replace a pagecache page with a new one
381  * @old:        page to be replaced
382  * @new:        page to replace with
383  * @gfp_mask:   allocation mode
384  *
385  * This function replaces a page in the pagecache with a new one.  On
386  * success it acquires the pagecache reference for the new page and
387  * drops it for the old page.  Both the old and new pages must be
388  * locked.  This function does not add the new page to the LRU, the
389  * caller must do that.
390  *
391  * The remove + add is atomic.  The only way this function can fail is
392  * memory allocation failure.
393  */
394 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
395 {
396         int error;
397         struct mem_cgroup *memcg = NULL;
398
399         VM_BUG_ON(!PageLocked(old));
400         VM_BUG_ON(!PageLocked(new));
401         VM_BUG_ON(new->mapping);
402
403         /*
404          * This is not page migration, but prepare_migration and
405          * end_migration does enough work for charge replacement.
406          *
407          * In the longer term we probably want a specialized function
408          * for moving the charge from old to new in a more efficient
409          * manner.
410          */
411         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
412         if (error)
413                 return error;
414
415         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
416         if (!error) {
417                 struct address_space *mapping = old->mapping;
418                 void (*freepage)(struct page *);
419
420                 pgoff_t offset = old->index;
421                 freepage = mapping->a_ops->freepage;
422
423                 page_cache_get(new);
424                 new->mapping = mapping;
425                 new->index = offset;
426
427                 spin_lock_irq(&mapping->tree_lock);
428                 __delete_from_page_cache(old);
429                 error = radix_tree_insert(&mapping->page_tree, offset, new);
430                 BUG_ON(error);
431                 mapping->nrpages++;
432                 __inc_zone_page_state(new, NR_FILE_PAGES);
433                 if (PageSwapBacked(new))
434                         __inc_zone_page_state(new, NR_SHMEM);
435                 spin_unlock_irq(&mapping->tree_lock);
436                 radix_tree_preload_end();
437                 if (freepage)
438                         freepage(old);
439                 page_cache_release(old);
440                 mem_cgroup_end_migration(memcg, old, new, true);
441         } else {
442                 mem_cgroup_end_migration(memcg, old, new, false);
443         }
444
445         return error;
446 }
447 EXPORT_SYMBOL_GPL(replace_page_cache_page);
448
449 /**
450  * add_to_page_cache_locked - add a locked page to the pagecache
451  * @page:       page to add
452  * @mapping:    the page's address_space
453  * @offset:     page index
454  * @gfp_mask:   page allocation mode
455  *
456  * This function is used to add a page to the pagecache. It must be locked.
457  * This function does not add the page to the LRU.  The caller must do that.
458  */
459 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
460                 pgoff_t offset, gfp_t gfp_mask)
461 {
462         int error;
463
464         VM_BUG_ON(!PageLocked(page));
465
466         error = mem_cgroup_cache_charge(page, current->mm,
467                                         gfp_mask & GFP_RECLAIM_MASK);
468         if (error)
469                 goto out;
470
471         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
472         if (error == 0) {
473                 page_cache_get(page);
474                 page->mapping = mapping;
475                 page->index = offset;
476
477                 spin_lock_irq(&mapping->tree_lock);
478                 error = radix_tree_insert(&mapping->page_tree, offset, page);
479                 if (likely(!error)) {
480                         mapping->nrpages++;
481                         __inc_zone_page_state(page, NR_FILE_PAGES);
482                         if (PageSwapBacked(page))
483                                 __inc_zone_page_state(page, NR_SHMEM);
484                         spin_unlock_irq(&mapping->tree_lock);
485                 } else {
486                         page->mapping = NULL;
487                         /* Leave page->index set: truncation relies upon it */
488                         spin_unlock_irq(&mapping->tree_lock);
489                         mem_cgroup_uncharge_cache_page(page);
490                         page_cache_release(page);
491                 }
492                 radix_tree_preload_end();
493         } else
494                 mem_cgroup_uncharge_cache_page(page);
495 out:
496         return error;
497 }
498 EXPORT_SYMBOL(add_to_page_cache_locked);
499
500 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
501                                 pgoff_t offset, gfp_t gfp_mask)
502 {
503         int ret;
504
505         /*
506          * Splice_read and readahead add shmem/tmpfs pages into the page cache
507          * before shmem_readpage has a chance to mark them as SwapBacked: they
508          * need to go on the anon lru below, and mem_cgroup_cache_charge
509          * (called in add_to_page_cache) needs to know where they're going too.
510          */
511         if (mapping_cap_swap_backed(mapping))
512                 SetPageSwapBacked(page);
513
514         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
515         if (ret == 0) {
516                 if (page_is_file_cache(page))
517                         lru_cache_add_file(page);
518                 else
519                         lru_cache_add_anon(page);
520         }
521         return ret;
522 }
523 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
524
525 #ifdef CONFIG_NUMA
526 struct page *__page_cache_alloc(gfp_t gfp)
527 {
528         int n;
529         struct page *page;
530
531         if (cpuset_do_page_mem_spread()) {
532                 get_mems_allowed();
533                 n = cpuset_mem_spread_node();
534                 page = alloc_pages_exact_node(n, gfp, 0);
535                 put_mems_allowed();
536                 return page;
537         }
538         return alloc_pages(gfp, 0);
539 }
540 EXPORT_SYMBOL(__page_cache_alloc);
541 #endif
542
543 /*
544  * In order to wait for pages to become available there must be
545  * waitqueues associated with pages. By using a hash table of
546  * waitqueues where the bucket discipline is to maintain all
547  * waiters on the same queue and wake all when any of the pages
548  * become available, and for the woken contexts to check to be
549  * sure the appropriate page became available, this saves space
550  * at a cost of "thundering herd" phenomena during rare hash
551  * collisions.
552  */
553 static wait_queue_head_t *page_waitqueue(struct page *page)
554 {
555         const struct zone *zone = page_zone(page);
556
557         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
558 }
559
560 static inline void wake_up_page(struct page *page, int bit)
561 {
562         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
563 }
564
565 void wait_on_page_bit(struct page *page, int bit_nr)
566 {
567         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
568
569         if (test_bit(bit_nr, &page->flags))
570                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
571                                                         TASK_UNINTERRUPTIBLE);
572 }
573 EXPORT_SYMBOL(wait_on_page_bit);
574
575 int wait_on_page_bit_killable(struct page *page, int bit_nr)
576 {
577         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
578
579         if (!test_bit(bit_nr, &page->flags))
580                 return 0;
581
582         return __wait_on_bit(page_waitqueue(page), &wait,
583                              sleep_on_page_killable, TASK_KILLABLE);
584 }
585
586 /**
587  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
588  * @page: Page defining the wait queue of interest
589  * @waiter: Waiter to add to the queue
590  *
591  * Add an arbitrary @waiter to the wait queue for the nominated @page.
592  */
593 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
594 {
595         wait_queue_head_t *q = page_waitqueue(page);
596         unsigned long flags;
597
598         spin_lock_irqsave(&q->lock, flags);
599         __add_wait_queue(q, waiter);
600         spin_unlock_irqrestore(&q->lock, flags);
601 }
602 EXPORT_SYMBOL_GPL(add_page_wait_queue);
603
604 /**
605  * unlock_page - unlock a locked page
606  * @page: the page
607  *
608  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
609  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
610  * mechananism between PageLocked pages and PageWriteback pages is shared.
611  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
612  *
613  * The mb is necessary to enforce ordering between the clear_bit and the read
614  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
615  */
616 void unlock_page(struct page *page)
617 {
618         VM_BUG_ON(!PageLocked(page));
619         clear_bit_unlock(PG_locked, &page->flags);
620         smp_mb__after_clear_bit();
621         wake_up_page(page, PG_locked);
622 }
623 EXPORT_SYMBOL(unlock_page);
624
625 /**
626  * end_page_writeback - end writeback against a page
627  * @page: the page
628  */
629 void end_page_writeback(struct page *page)
630 {
631         if (TestClearPageReclaim(page))
632                 rotate_reclaimable_page(page);
633
634         if (!test_clear_page_writeback(page))
635                 BUG();
636
637         smp_mb__after_clear_bit();
638         wake_up_page(page, PG_writeback);
639 }
640 EXPORT_SYMBOL(end_page_writeback);
641
642 /**
643  * __lock_page - get a lock on the page, assuming we need to sleep to get it
644  * @page: the page to lock
645  */
646 void __lock_page(struct page *page)
647 {
648         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
649
650         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
651                                                         TASK_UNINTERRUPTIBLE);
652 }
653 EXPORT_SYMBOL(__lock_page);
654
655 int __lock_page_killable(struct page *page)
656 {
657         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
658
659         return __wait_on_bit_lock(page_waitqueue(page), &wait,
660                                         sleep_on_page_killable, TASK_KILLABLE);
661 }
662 EXPORT_SYMBOL_GPL(__lock_page_killable);
663
664 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
665                          unsigned int flags)
666 {
667         if (flags & FAULT_FLAG_ALLOW_RETRY) {
668                 /*
669                  * CAUTION! In this case, mmap_sem is not released
670                  * even though return 0.
671                  */
672                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
673                         return 0;
674
675                 up_read(&mm->mmap_sem);
676                 if (flags & FAULT_FLAG_KILLABLE)
677                         wait_on_page_locked_killable(page);
678                 else
679                         wait_on_page_locked(page);
680                 return 0;
681         } else {
682                 if (flags & FAULT_FLAG_KILLABLE) {
683                         int ret;
684
685                         ret = __lock_page_killable(page);
686                         if (ret) {
687                                 up_read(&mm->mmap_sem);
688                                 return 0;
689                         }
690                 } else
691                         __lock_page(page);
692                 return 1;
693         }
694 }
695
696 /**
697  * find_get_page - find and get a page reference
698  * @mapping: the address_space to search
699  * @offset: the page index
700  *
701  * Is there a pagecache struct page at the given (mapping, offset) tuple?
702  * If yes, increment its refcount and return it; if no, return NULL.
703  */
704 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
705 {
706         void **pagep;
707         struct page *page;
708
709         rcu_read_lock();
710 repeat:
711         page = NULL;
712         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
713         if (pagep) {
714                 page = radix_tree_deref_slot(pagep);
715                 if (unlikely(!page))
716                         goto out;
717                 if (radix_tree_exception(page)) {
718                         if (radix_tree_exceptional_entry(page))
719                                 goto out;
720                         /* radix_tree_deref_retry(page) */
721                         goto repeat;
722                 }
723                 if (!page_cache_get_speculative(page))
724                         goto repeat;
725
726                 /*
727                  * Has the page moved?
728                  * This is part of the lockless pagecache protocol. See
729                  * include/linux/pagemap.h for details.
730                  */
731                 if (unlikely(page != *pagep)) {
732                         page_cache_release(page);
733                         goto repeat;
734                 }
735         }
736 out:
737         rcu_read_unlock();
738
739         return page;
740 }
741 EXPORT_SYMBOL(find_get_page);
742
743 /**
744  * find_lock_page - locate, pin and lock a pagecache page
745  * @mapping: the address_space to search
746  * @offset: the page index
747  *
748  * Locates the desired pagecache page, locks it, increments its reference
749  * count and returns its address.
750  *
751  * Returns zero if the page was not present. find_lock_page() may sleep.
752  */
753 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
754 {
755         struct page *page;
756
757 repeat:
758         page = find_get_page(mapping, offset);
759         if (page && !radix_tree_exception(page)) {
760                 lock_page(page);
761                 /* Has the page been truncated? */
762                 if (unlikely(page->mapping != mapping)) {
763                         unlock_page(page);
764                         page_cache_release(page);
765                         goto repeat;
766                 }
767                 VM_BUG_ON(page->index != offset);
768         }
769         return page;
770 }
771 EXPORT_SYMBOL(find_lock_page);
772
773 /**
774  * find_or_create_page - locate or add a pagecache page
775  * @mapping: the page's address_space
776  * @index: the page's index into the mapping
777  * @gfp_mask: page allocation mode
778  *
779  * Locates a page in the pagecache.  If the page is not present, a new page
780  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
781  * LRU list.  The returned page is locked and has its reference count
782  * incremented.
783  *
784  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
785  * allocation!
786  *
787  * find_or_create_page() returns the desired page's address, or zero on
788  * memory exhaustion.
789  */
790 struct page *find_or_create_page(struct address_space *mapping,
791                 pgoff_t index, gfp_t gfp_mask)
792 {
793         struct page *page;
794         int err;
795 repeat:
796         page = find_lock_page(mapping, index);
797         if (!page) {
798                 page = __page_cache_alloc(gfp_mask);
799                 if (!page)
800                         return NULL;
801                 /*
802                  * We want a regular kernel memory (not highmem or DMA etc)
803                  * allocation for the radix tree nodes, but we need to honour
804                  * the context-specific requirements the caller has asked for.
805                  * GFP_RECLAIM_MASK collects those requirements.
806                  */
807                 err = add_to_page_cache_lru(page, mapping, index,
808                         (gfp_mask & GFP_RECLAIM_MASK));
809                 if (unlikely(err)) {
810                         page_cache_release(page);
811                         page = NULL;
812                         if (err == -EEXIST)
813                                 goto repeat;
814                 }
815         }
816         return page;
817 }
818 EXPORT_SYMBOL(find_or_create_page);
819
820 /**
821  * find_get_pages - gang pagecache lookup
822  * @mapping:    The address_space to search
823  * @start:      The starting page index
824  * @nr_pages:   The maximum number of pages
825  * @pages:      Where the resulting pages are placed
826  *
827  * find_get_pages() will search for and return a group of up to
828  * @nr_pages pages in the mapping.  The pages are placed at @pages.
829  * find_get_pages() takes a reference against the returned pages.
830  *
831  * The search returns a group of mapping-contiguous pages with ascending
832  * indexes.  There may be holes in the indices due to not-present pages.
833  *
834  * find_get_pages() returns the number of pages which were found.
835  */
836 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
837                             unsigned int nr_pages, struct page **pages)
838 {
839         unsigned int i;
840         unsigned int ret;
841         unsigned int nr_found;
842
843         rcu_read_lock();
844 restart:
845         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
846                                 (void ***)pages, NULL, start, nr_pages);
847         ret = 0;
848         for (i = 0; i < nr_found; i++) {
849                 struct page *page;
850 repeat:
851                 page = radix_tree_deref_slot((void **)pages[i]);
852                 if (unlikely(!page))
853                         continue;
854
855                 if (radix_tree_exception(page)) {
856                         if (radix_tree_exceptional_entry(page))
857                                 continue;
858                         /*
859                          * radix_tree_deref_retry(page):
860                          * can only trigger when entry at index 0 moves out of
861                          * or back to root: none yet gotten, safe to restart.
862                          */
863                         WARN_ON(start | i);
864                         goto restart;
865                 }
866
867                 if (!page_cache_get_speculative(page))
868                         goto repeat;
869
870                 /* Has the page moved? */
871                 if (unlikely(page != *((void **)pages[i]))) {
872                         page_cache_release(page);
873                         goto repeat;
874                 }
875
876                 pages[ret] = page;
877                 ret++;
878         }
879
880         /*
881          * If all entries were removed before we could secure them,
882          * try again, because callers stop trying once 0 is returned.
883          */
884         if (unlikely(!ret && nr_found))
885                 goto restart;
886         rcu_read_unlock();
887         return ret;
888 }
889
890 /**
891  * find_get_pages_contig - gang contiguous pagecache lookup
892  * @mapping:    The address_space to search
893  * @index:      The starting page index
894  * @nr_pages:   The maximum number of pages
895  * @pages:      Where the resulting pages are placed
896  *
897  * find_get_pages_contig() works exactly like find_get_pages(), except
898  * that the returned number of pages are guaranteed to be contiguous.
899  *
900  * find_get_pages_contig() returns the number of pages which were found.
901  */
902 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
903                                unsigned int nr_pages, struct page **pages)
904 {
905         unsigned int i;
906         unsigned int ret;
907         unsigned int nr_found;
908
909         rcu_read_lock();
910 restart:
911         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
912                                 (void ***)pages, NULL, index, nr_pages);
913         ret = 0;
914         for (i = 0; i < nr_found; i++) {
915                 struct page *page;
916 repeat:
917                 page = radix_tree_deref_slot((void **)pages[i]);
918                 if (unlikely(!page))
919                         continue;
920
921                 if (radix_tree_exception(page)) {
922                         if (radix_tree_exceptional_entry(page))
923                                 break;
924                         /*
925                          * radix_tree_deref_retry(page):
926                          * can only trigger when entry at index 0 moves out of
927                          * or back to root: none yet gotten, safe to restart.
928                          */
929                         goto restart;
930                 }
931
932                 if (!page_cache_get_speculative(page))
933                         goto repeat;
934
935                 /* Has the page moved? */
936                 if (unlikely(page != *((void **)pages[i]))) {
937                         page_cache_release(page);
938                         goto repeat;
939                 }
940
941                 /*
942                  * must check mapping and index after taking the ref.
943                  * otherwise we can get both false positives and false
944                  * negatives, which is just confusing to the caller.
945                  */
946                 if (page->mapping == NULL || page->index != index) {
947                         page_cache_release(page);
948                         break;
949                 }
950
951                 pages[ret] = page;
952                 ret++;
953                 index++;
954         }
955         rcu_read_unlock();
956         return ret;
957 }
958 EXPORT_SYMBOL(find_get_pages_contig);
959
960 /**
961  * find_get_pages_tag - find and return pages that match @tag
962  * @mapping:    the address_space to search
963  * @index:      the starting page index
964  * @tag:        the tag index
965  * @nr_pages:   the maximum number of pages
966  * @pages:      where the resulting pages are placed
967  *
968  * Like find_get_pages, except we only return pages which are tagged with
969  * @tag.   We update @index to index the next page for the traversal.
970  */
971 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
972                         int tag, unsigned int nr_pages, struct page **pages)
973 {
974         unsigned int i;
975         unsigned int ret;
976         unsigned int nr_found;
977
978         rcu_read_lock();
979 restart:
980         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
981                                 (void ***)pages, *index, nr_pages, tag);
982         ret = 0;
983         for (i = 0; i < nr_found; i++) {
984                 struct page *page;
985 repeat:
986                 page = radix_tree_deref_slot((void **)pages[i]);
987                 if (unlikely(!page))
988                         continue;
989
990                 if (radix_tree_exception(page)) {
991                         BUG_ON(radix_tree_exceptional_entry(page));
992                         /*
993                          * radix_tree_deref_retry(page):
994                          * can only trigger when entry at index 0 moves out of
995                          * or back to root: none yet gotten, safe to restart.
996                          */
997                         goto restart;
998                 }
999
1000                 if (!page_cache_get_speculative(page))
1001                         goto repeat;
1002
1003                 /* Has the page moved? */
1004                 if (unlikely(page != *((void **)pages[i]))) {
1005                         page_cache_release(page);
1006                         goto repeat;
1007                 }
1008
1009                 pages[ret] = page;
1010                 ret++;
1011         }
1012
1013         /*
1014          * If all entries were removed before we could secure them,
1015          * try again, because callers stop trying once 0 is returned.
1016          */
1017         if (unlikely(!ret && nr_found))
1018                 goto restart;
1019         rcu_read_unlock();
1020
1021         if (ret)
1022                 *index = pages[ret - 1]->index + 1;
1023
1024         return ret;
1025 }
1026 EXPORT_SYMBOL(find_get_pages_tag);
1027
1028 /**
1029  * grab_cache_page_nowait - returns locked page at given index in given cache
1030  * @mapping: target address_space
1031  * @index: the page index
1032  *
1033  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1034  * This is intended for speculative data generators, where the data can
1035  * be regenerated if the page couldn't be grabbed.  This routine should
1036  * be safe to call while holding the lock for another page.
1037  *
1038  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1039  * and deadlock against the caller's locked page.
1040  */
1041 struct page *
1042 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1043 {
1044         struct page *page = find_get_page(mapping, index);
1045
1046         if (page) {
1047                 if (trylock_page(page))
1048                         return page;
1049                 page_cache_release(page);
1050                 return NULL;
1051         }
1052         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1053         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1054                 page_cache_release(page);
1055                 page = NULL;
1056         }
1057         return page;
1058 }
1059 EXPORT_SYMBOL(grab_cache_page_nowait);
1060
1061 /*
1062  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1063  * a _large_ part of the i/o request. Imagine the worst scenario:
1064  *
1065  *      ---R__________________________________________B__________
1066  *         ^ reading here                             ^ bad block(assume 4k)
1067  *
1068  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1069  * => failing the whole request => read(R) => read(R+1) =>
1070  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1071  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1072  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1073  *
1074  * It is going insane. Fix it by quickly scaling down the readahead size.
1075  */
1076 static void shrink_readahead_size_eio(struct file *filp,
1077                                         struct file_ra_state *ra)
1078 {
1079         ra->ra_pages /= 4;
1080 }
1081
1082 /**
1083  * do_generic_file_read - generic file read routine
1084  * @filp:       the file to read
1085  * @ppos:       current file position
1086  * @desc:       read_descriptor
1087  * @actor:      read method
1088  *
1089  * This is a generic file read routine, and uses the
1090  * mapping->a_ops->readpage() function for the actual low-level stuff.
1091  *
1092  * This is really ugly. But the goto's actually try to clarify some
1093  * of the logic when it comes to error handling etc.
1094  */
1095 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1096                 read_descriptor_t *desc, read_actor_t actor)
1097 {
1098         struct address_space *mapping = filp->f_mapping;
1099         struct inode *inode = mapping->host;
1100         struct file_ra_state *ra = &filp->f_ra;
1101         pgoff_t index;
1102         pgoff_t last_index;
1103         pgoff_t prev_index;
1104         unsigned long offset;      /* offset into pagecache page */
1105         unsigned int prev_offset;
1106         int error;
1107
1108         index = *ppos >> PAGE_CACHE_SHIFT;
1109         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1110         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1111         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1112         offset = *ppos & ~PAGE_CACHE_MASK;
1113
1114         for (;;) {
1115                 struct page *page;
1116                 pgoff_t end_index;
1117                 loff_t isize;
1118                 unsigned long nr, ret;
1119
1120                 cond_resched();
1121 find_page:
1122                 page = find_get_page(mapping, index);
1123                 if (!page) {
1124                         page_cache_sync_readahead(mapping,
1125                                         ra, filp,
1126                                         index, last_index - index);
1127                         page = find_get_page(mapping, index);
1128                         if (unlikely(page == NULL))
1129                                 goto no_cached_page;
1130                 }
1131                 if (PageReadahead(page)) {
1132                         page_cache_async_readahead(mapping,
1133                                         ra, filp, page,
1134                                         index, last_index - index);
1135                 }
1136                 if (!PageUptodate(page)) {
1137                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1138                                         !mapping->a_ops->is_partially_uptodate)
1139                                 goto page_not_up_to_date;
1140                         if (!trylock_page(page))
1141                                 goto page_not_up_to_date;
1142                         /* Did it get truncated before we got the lock? */
1143                         if (!page->mapping)
1144                                 goto page_not_up_to_date_locked;
1145                         if (!mapping->a_ops->is_partially_uptodate(page,
1146                                                                 desc, offset))
1147                                 goto page_not_up_to_date_locked;
1148                         unlock_page(page);
1149                 }
1150 page_ok:
1151                 /*
1152                  * i_size must be checked after we know the page is Uptodate.
1153                  *
1154                  * Checking i_size after the check allows us to calculate
1155                  * the correct value for "nr", which means the zero-filled
1156                  * part of the page is not copied back to userspace (unless
1157                  * another truncate extends the file - this is desired though).
1158                  */
1159
1160                 isize = i_size_read(inode);
1161                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1162                 if (unlikely(!isize || index > end_index)) {
1163                         page_cache_release(page);
1164                         goto out;
1165                 }
1166
1167                 /* nr is the maximum number of bytes to copy from this page */
1168                 nr = PAGE_CACHE_SIZE;
1169                 if (index == end_index) {
1170                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1171                         if (nr <= offset) {
1172                                 page_cache_release(page);
1173                                 goto out;
1174                         }
1175                 }
1176                 nr = nr - offset;
1177
1178                 /* If users can be writing to this page using arbitrary
1179                  * virtual addresses, take care about potential aliasing
1180                  * before reading the page on the kernel side.
1181                  */
1182                 if (mapping_writably_mapped(mapping))
1183                         flush_dcache_page(page);
1184
1185                 /*
1186                  * When a sequential read accesses a page several times,
1187                  * only mark it as accessed the first time.
1188                  */
1189                 if (prev_index != index || offset != prev_offset)
1190                         mark_page_accessed(page);
1191                 prev_index = index;
1192
1193                 /*
1194                  * Ok, we have the page, and it's up-to-date, so
1195                  * now we can copy it to user space...
1196                  *
1197                  * The actor routine returns how many bytes were actually used..
1198                  * NOTE! This may not be the same as how much of a user buffer
1199                  * we filled up (we may be padding etc), so we can only update
1200                  * "pos" here (the actor routine has to update the user buffer
1201                  * pointers and the remaining count).
1202                  */
1203                 ret = actor(desc, page, offset, nr);
1204                 offset += ret;
1205                 index += offset >> PAGE_CACHE_SHIFT;
1206                 offset &= ~PAGE_CACHE_MASK;
1207                 prev_offset = offset;
1208
1209                 page_cache_release(page);
1210                 if (ret == nr && desc->count)
1211                         continue;
1212                 goto out;
1213
1214 page_not_up_to_date:
1215                 /* Get exclusive access to the page ... */
1216                 error = lock_page_killable(page);
1217                 if (unlikely(error))
1218                         goto readpage_error;
1219
1220 page_not_up_to_date_locked:
1221                 /* Did it get truncated before we got the lock? */
1222                 if (!page->mapping) {
1223                         unlock_page(page);
1224                         page_cache_release(page);
1225                         continue;
1226                 }
1227
1228                 /* Did somebody else fill it already? */
1229                 if (PageUptodate(page)) {
1230                         unlock_page(page);
1231                         goto page_ok;
1232                 }
1233
1234 readpage:
1235                 /*
1236                  * A previous I/O error may have been due to temporary
1237                  * failures, eg. multipath errors.
1238                  * PG_error will be set again if readpage fails.
1239                  */
1240                 ClearPageError(page);
1241                 /* Start the actual read. The read will unlock the page. */
1242                 error = mapping->a_ops->readpage(filp, page);
1243
1244                 if (unlikely(error)) {
1245                         if (error == AOP_TRUNCATED_PAGE) {
1246                                 page_cache_release(page);
1247                                 goto find_page;
1248                         }
1249                         goto readpage_error;
1250                 }
1251
1252                 if (!PageUptodate(page)) {
1253                         error = lock_page_killable(page);
1254                         if (unlikely(error))
1255                                 goto readpage_error;
1256                         if (!PageUptodate(page)) {
1257                                 if (page->mapping == NULL) {
1258                                         /*
1259                                          * invalidate_mapping_pages got it
1260                                          */
1261                                         unlock_page(page);
1262                                         page_cache_release(page);
1263                                         goto find_page;
1264                                 }
1265                                 unlock_page(page);
1266                                 shrink_readahead_size_eio(filp, ra);
1267                                 error = -EIO;
1268                                 goto readpage_error;
1269                         }
1270                         unlock_page(page);
1271                 }
1272
1273                 goto page_ok;
1274
1275 readpage_error:
1276                 /* UHHUH! A synchronous read error occurred. Report it */
1277                 desc->error = error;
1278                 page_cache_release(page);
1279                 goto out;
1280
1281 no_cached_page:
1282                 /*
1283                  * Ok, it wasn't cached, so we need to create a new
1284                  * page..
1285                  */
1286                 page = page_cache_alloc_cold(mapping);
1287                 if (!page) {
1288                         desc->error = -ENOMEM;
1289                         goto out;
1290                 }
1291                 error = add_to_page_cache_lru(page, mapping,
1292                                                 index, GFP_KERNEL);
1293                 if (error) {
1294                         page_cache_release(page);
1295                         if (error == -EEXIST)
1296                                 goto find_page;
1297                         desc->error = error;
1298                         goto out;
1299                 }
1300                 goto readpage;
1301         }
1302
1303 out:
1304         ra->prev_pos = prev_index;
1305         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1306         ra->prev_pos |= prev_offset;
1307
1308         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1309         file_accessed(filp);
1310 }
1311
1312 int file_read_actor(read_descriptor_t *desc, struct page *page,
1313                         unsigned long offset, unsigned long size)
1314 {
1315         char *kaddr;
1316         unsigned long left, count = desc->count;
1317
1318         if (size > count)
1319                 size = count;
1320
1321         /*
1322          * Faults on the destination of a read are common, so do it before
1323          * taking the kmap.
1324          */
1325         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1326                 kaddr = kmap_atomic(page, KM_USER0);
1327                 left = __copy_to_user_inatomic(desc->arg.buf,
1328                                                 kaddr + offset, size);
1329                 kunmap_atomic(kaddr, KM_USER0);
1330                 if (left == 0)
1331                         goto success;
1332         }
1333
1334         /* Do it the slow way */
1335         kaddr = kmap(page);
1336         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1337         kunmap(page);
1338
1339         if (left) {
1340                 size -= left;
1341                 desc->error = -EFAULT;
1342         }
1343 success:
1344         desc->count = count - size;
1345         desc->written += size;
1346         desc->arg.buf += size;
1347         return size;
1348 }
1349
1350 /*
1351  * Performs necessary checks before doing a write
1352  * @iov:        io vector request
1353  * @nr_segs:    number of segments in the iovec
1354  * @count:      number of bytes to write
1355  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1356  *
1357  * Adjust number of segments and amount of bytes to write (nr_segs should be
1358  * properly initialized first). Returns appropriate error code that caller
1359  * should return or zero in case that write should be allowed.
1360  */
1361 int generic_segment_checks(const struct iovec *iov,
1362                         unsigned long *nr_segs, size_t *count, int access_flags)
1363 {
1364         unsigned long   seg;
1365         size_t cnt = 0;
1366         for (seg = 0; seg < *nr_segs; seg++) {
1367                 const struct iovec *iv = &iov[seg];
1368
1369                 /*
1370                  * If any segment has a negative length, or the cumulative
1371                  * length ever wraps negative then return -EINVAL.
1372                  */
1373                 cnt += iv->iov_len;
1374                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1375                         return -EINVAL;
1376                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1377                         continue;
1378                 if (seg == 0)
1379                         return -EFAULT;
1380                 *nr_segs = seg;
1381                 cnt -= iv->iov_len;     /* This segment is no good */
1382                 break;
1383         }
1384         *count = cnt;
1385         return 0;
1386 }
1387 EXPORT_SYMBOL(generic_segment_checks);
1388
1389 /**
1390  * generic_file_aio_read - generic filesystem read routine
1391  * @iocb:       kernel I/O control block
1392  * @iov:        io vector request
1393  * @nr_segs:    number of segments in the iovec
1394  * @pos:        current file position
1395  *
1396  * This is the "read()" routine for all filesystems
1397  * that can use the page cache directly.
1398  */
1399 ssize_t
1400 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1401                 unsigned long nr_segs, loff_t pos)
1402 {
1403         struct file *filp = iocb->ki_filp;
1404         ssize_t retval;
1405         unsigned long seg = 0;
1406         size_t count;
1407         loff_t *ppos = &iocb->ki_pos;
1408         struct blk_plug plug;
1409
1410         count = 0;
1411         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1412         if (retval)
1413                 return retval;
1414
1415         blk_start_plug(&plug);
1416
1417         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1418         if (filp->f_flags & O_DIRECT) {
1419                 loff_t size;
1420                 struct address_space *mapping;
1421                 struct inode *inode;
1422
1423                 mapping = filp->f_mapping;
1424                 inode = mapping->host;
1425                 if (!count)
1426                         goto out; /* skip atime */
1427                 size = i_size_read(inode);
1428                 if (pos < size) {
1429                         retval = filemap_write_and_wait_range(mapping, pos,
1430                                         pos + iov_length(iov, nr_segs) - 1);
1431                         if (!retval) {
1432                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1433                                                         iov, pos, nr_segs);
1434                         }
1435                         if (retval > 0) {
1436                                 *ppos = pos + retval;
1437                                 count -= retval;
1438                         }
1439
1440                         /*
1441                          * Btrfs can have a short DIO read if we encounter
1442                          * compressed extents, so if there was an error, or if
1443                          * we've already read everything we wanted to, or if
1444                          * there was a short read because we hit EOF, go ahead
1445                          * and return.  Otherwise fallthrough to buffered io for
1446                          * the rest of the read.
1447                          */
1448                         if (retval < 0 || !count || *ppos >= size) {
1449                                 file_accessed(filp);
1450                                 goto out;
1451                         }
1452                 }
1453         }
1454
1455         count = retval;
1456         for (seg = 0; seg < nr_segs; seg++) {
1457                 read_descriptor_t desc;
1458                 loff_t offset = 0;
1459
1460                 /*
1461                  * If we did a short DIO read we need to skip the section of the
1462                  * iov that we've already read data into.
1463                  */
1464                 if (count) {
1465                         if (count > iov[seg].iov_len) {
1466                                 count -= iov[seg].iov_len;
1467                                 continue;
1468                         }
1469                         offset = count;
1470                         count = 0;
1471                 }
1472
1473                 desc.written = 0;
1474                 desc.arg.buf = iov[seg].iov_base + offset;
1475                 desc.count = iov[seg].iov_len - offset;
1476                 if (desc.count == 0)
1477                         continue;
1478                 desc.error = 0;
1479                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1480                 retval += desc.written;
1481                 if (desc.error) {
1482                         retval = retval ?: desc.error;
1483                         break;
1484                 }
1485                 if (desc.count > 0)
1486                         break;
1487         }
1488 out:
1489         blk_finish_plug(&plug);
1490         return retval;
1491 }
1492 EXPORT_SYMBOL(generic_file_aio_read);
1493
1494 static ssize_t
1495 do_readahead(struct address_space *mapping, struct file *filp,
1496              pgoff_t index, unsigned long nr)
1497 {
1498         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1499                 return -EINVAL;
1500
1501         force_page_cache_readahead(mapping, filp, index, nr);
1502         return 0;
1503 }
1504
1505 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1506 {
1507         ssize_t ret;
1508         struct file *file;
1509
1510         ret = -EBADF;
1511         file = fget(fd);
1512         if (file) {
1513                 if (file->f_mode & FMODE_READ) {
1514                         struct address_space *mapping = file->f_mapping;
1515                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1516                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1517                         unsigned long len = end - start + 1;
1518                         ret = do_readahead(mapping, file, start, len);
1519                 }
1520                 fput(file);
1521         }
1522         return ret;
1523 }
1524 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1525 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1526 {
1527         return SYSC_readahead((int) fd, offset, (size_t) count);
1528 }
1529 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1530 #endif
1531
1532 #ifdef CONFIG_MMU
1533 /**
1534  * page_cache_read - adds requested page to the page cache if not already there
1535  * @file:       file to read
1536  * @offset:     page index
1537  *
1538  * This adds the requested page to the page cache if it isn't already there,
1539  * and schedules an I/O to read in its contents from disk.
1540  */
1541 static int page_cache_read(struct file *file, pgoff_t offset)
1542 {
1543         struct address_space *mapping = file->f_mapping;
1544         struct page *page; 
1545         int ret;
1546
1547         do {
1548                 page = page_cache_alloc_cold(mapping);
1549                 if (!page)
1550                         return -ENOMEM;
1551
1552                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1553                 if (ret == 0)
1554                         ret = mapping->a_ops->readpage(file, page);
1555                 else if (ret == -EEXIST)
1556                         ret = 0; /* losing race to add is OK */
1557
1558                 page_cache_release(page);
1559
1560         } while (ret == AOP_TRUNCATED_PAGE);
1561                 
1562         return ret;
1563 }
1564
1565 #define MMAP_LOTSAMISS  (100)
1566
1567 /*
1568  * Synchronous readahead happens when we don't even find
1569  * a page in the page cache at all.
1570  */
1571 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1572                                    struct file_ra_state *ra,
1573                                    struct file *file,
1574                                    pgoff_t offset)
1575 {
1576         unsigned long ra_pages;
1577         struct address_space *mapping = file->f_mapping;
1578
1579         /* If we don't want any read-ahead, don't bother */
1580         if (VM_RandomReadHint(vma))
1581                 return;
1582         if (!ra->ra_pages)
1583                 return;
1584
1585         if (VM_SequentialReadHint(vma)) {
1586                 page_cache_sync_readahead(mapping, ra, file, offset,
1587                                           ra->ra_pages);
1588                 return;
1589         }
1590
1591         /* Avoid banging the cache line if not needed */
1592         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1593                 ra->mmap_miss++;
1594
1595         /*
1596          * Do we miss much more than hit in this file? If so,
1597          * stop bothering with read-ahead. It will only hurt.
1598          */
1599         if (ra->mmap_miss > MMAP_LOTSAMISS)
1600                 return;
1601
1602         /*
1603          * mmap read-around
1604          */
1605         ra_pages = max_sane_readahead(ra->ra_pages);
1606         ra->start = max_t(long, 0, offset - ra_pages / 2);
1607         ra->size = ra_pages;
1608         ra->async_size = ra_pages / 4;
1609         ra_submit(ra, mapping, file);
1610 }
1611
1612 /*
1613  * Asynchronous readahead happens when we find the page and PG_readahead,
1614  * so we want to possibly extend the readahead further..
1615  */
1616 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1617                                     struct file_ra_state *ra,
1618                                     struct file *file,
1619                                     struct page *page,
1620                                     pgoff_t offset)
1621 {
1622         struct address_space *mapping = file->f_mapping;
1623
1624         /* If we don't want any read-ahead, don't bother */
1625         if (VM_RandomReadHint(vma))
1626                 return;
1627         if (ra->mmap_miss > 0)
1628                 ra->mmap_miss--;
1629         if (PageReadahead(page))
1630                 page_cache_async_readahead(mapping, ra, file,
1631                                            page, offset, ra->ra_pages);
1632 }
1633
1634 /**
1635  * filemap_fault - read in file data for page fault handling
1636  * @vma:        vma in which the fault was taken
1637  * @vmf:        struct vm_fault containing details of the fault
1638  *
1639  * filemap_fault() is invoked via the vma operations vector for a
1640  * mapped memory region to read in file data during a page fault.
1641  *
1642  * The goto's are kind of ugly, but this streamlines the normal case of having
1643  * it in the page cache, and handles the special cases reasonably without
1644  * having a lot of duplicated code.
1645  */
1646 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1647 {
1648         int error;
1649         struct file *file = vma->vm_file;
1650         struct address_space *mapping = file->f_mapping;
1651         struct file_ra_state *ra = &file->f_ra;
1652         struct inode *inode = mapping->host;
1653         pgoff_t offset = vmf->pgoff;
1654         struct page *page;
1655         pgoff_t size;
1656         int ret = 0;
1657
1658         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1659         if (offset >= size)
1660                 return VM_FAULT_SIGBUS;
1661
1662         /*
1663          * Do we have something in the page cache already?
1664          */
1665         page = find_get_page(mapping, offset);
1666         if (likely(page)) {
1667                 /*
1668                  * We found the page, so try async readahead before
1669                  * waiting for the lock.
1670                  */
1671                 do_async_mmap_readahead(vma, ra, file, page, offset);
1672         } else {
1673                 /* No page in the page cache at all */
1674                 do_sync_mmap_readahead(vma, ra, file, offset);
1675                 count_vm_event(PGMAJFAULT);
1676                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1677                 ret = VM_FAULT_MAJOR;
1678 retry_find:
1679                 page = find_get_page(mapping, offset);
1680                 if (!page)
1681                         goto no_cached_page;
1682         }
1683
1684         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1685                 page_cache_release(page);
1686                 return ret | VM_FAULT_RETRY;
1687         }
1688
1689         /* Did it get truncated? */
1690         if (unlikely(page->mapping != mapping)) {
1691                 unlock_page(page);
1692                 put_page(page);
1693                 goto retry_find;
1694         }
1695         VM_BUG_ON(page->index != offset);
1696
1697         /*
1698          * We have a locked page in the page cache, now we need to check
1699          * that it's up-to-date. If not, it is going to be due to an error.
1700          */
1701         if (unlikely(!PageUptodate(page)))
1702                 goto page_not_uptodate;
1703
1704         /*
1705          * Found the page and have a reference on it.
1706          * We must recheck i_size under page lock.
1707          */
1708         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1709         if (unlikely(offset >= size)) {
1710                 unlock_page(page);
1711                 page_cache_release(page);
1712                 return VM_FAULT_SIGBUS;
1713         }
1714
1715         vmf->page = page;
1716         return ret | VM_FAULT_LOCKED;
1717
1718 no_cached_page:
1719         /*
1720          * We're only likely to ever get here if MADV_RANDOM is in
1721          * effect.
1722          */
1723         error = page_cache_read(file, offset);
1724
1725         /*
1726          * The page we want has now been added to the page cache.
1727          * In the unlikely event that someone removed it in the
1728          * meantime, we'll just come back here and read it again.
1729          */
1730         if (error >= 0)
1731                 goto retry_find;
1732
1733         /*
1734          * An error return from page_cache_read can result if the
1735          * system is low on memory, or a problem occurs while trying
1736          * to schedule I/O.
1737          */
1738         if (error == -ENOMEM)
1739                 return VM_FAULT_OOM;
1740         return VM_FAULT_SIGBUS;
1741
1742 page_not_uptodate:
1743         /*
1744          * Umm, take care of errors if the page isn't up-to-date.
1745          * Try to re-read it _once_. We do this synchronously,
1746          * because there really aren't any performance issues here
1747          * and we need to check for errors.
1748          */
1749         ClearPageError(page);
1750         error = mapping->a_ops->readpage(file, page);
1751         if (!error) {
1752                 wait_on_page_locked(page);
1753                 if (!PageUptodate(page))
1754                         error = -EIO;
1755         }
1756         page_cache_release(page);
1757
1758         if (!error || error == AOP_TRUNCATED_PAGE)
1759                 goto retry_find;
1760
1761         /* Things didn't work out. Return zero to tell the mm layer so. */
1762         shrink_readahead_size_eio(file, ra);
1763         return VM_FAULT_SIGBUS;
1764 }
1765 EXPORT_SYMBOL(filemap_fault);
1766
1767 const struct vm_operations_struct generic_file_vm_ops = {
1768         .fault          = filemap_fault,
1769 };
1770
1771 /* This is used for a general mmap of a disk file */
1772
1773 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1774 {
1775         struct address_space *mapping = file->f_mapping;
1776
1777         if (!mapping->a_ops->readpage)
1778                 return -ENOEXEC;
1779         file_accessed(file);
1780         vma->vm_ops = &generic_file_vm_ops;
1781         vma->vm_flags |= VM_CAN_NONLINEAR;
1782         return 0;
1783 }
1784
1785 /*
1786  * This is for filesystems which do not implement ->writepage.
1787  */
1788 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1789 {
1790         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1791                 return -EINVAL;
1792         return generic_file_mmap(file, vma);
1793 }
1794 #else
1795 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1796 {
1797         return -ENOSYS;
1798 }
1799 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1800 {
1801         return -ENOSYS;
1802 }
1803 #endif /* CONFIG_MMU */
1804
1805 EXPORT_SYMBOL(generic_file_mmap);
1806 EXPORT_SYMBOL(generic_file_readonly_mmap);
1807
1808 static struct page *__read_cache_page(struct address_space *mapping,
1809                                 pgoff_t index,
1810                                 int (*filler)(void *, struct page *),
1811                                 void *data,
1812                                 gfp_t gfp)
1813 {
1814         struct page *page;
1815         int err;
1816 repeat:
1817         page = find_get_page(mapping, index);
1818         if (!page) {
1819                 page = __page_cache_alloc(gfp | __GFP_COLD);
1820                 if (!page)
1821                         return ERR_PTR(-ENOMEM);
1822                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1823                 if (unlikely(err)) {
1824                         page_cache_release(page);
1825                         if (err == -EEXIST)
1826                                 goto repeat;
1827                         /* Presumably ENOMEM for radix tree node */
1828                         return ERR_PTR(err);
1829                 }
1830                 err = filler(data, page);
1831                 if (err < 0) {
1832                         page_cache_release(page);
1833                         page = ERR_PTR(err);
1834                 }
1835         }
1836         return page;
1837 }
1838
1839 static struct page *do_read_cache_page(struct address_space *mapping,
1840                                 pgoff_t index,
1841                                 int (*filler)(void *, struct page *),
1842                                 void *data,
1843                                 gfp_t gfp)
1844
1845 {
1846         struct page *page;
1847         int err;
1848
1849 retry:
1850         page = __read_cache_page(mapping, index, filler, data, gfp);
1851         if (IS_ERR(page))
1852                 return page;
1853         if (PageUptodate(page))
1854                 goto out;
1855
1856         lock_page(page);
1857         if (!page->mapping) {
1858                 unlock_page(page);
1859                 page_cache_release(page);
1860                 goto retry;
1861         }
1862         if (PageUptodate(page)) {
1863                 unlock_page(page);
1864                 goto out;
1865         }
1866         err = filler(data, page);
1867         if (err < 0) {
1868                 page_cache_release(page);
1869                 return ERR_PTR(err);
1870         }
1871 out:
1872         mark_page_accessed(page);
1873         return page;
1874 }
1875
1876 /**
1877  * read_cache_page_async - read into page cache, fill it if needed
1878  * @mapping:    the page's address_space
1879  * @index:      the page index
1880  * @filler:     function to perform the read
1881  * @data:       first arg to filler(data, page) function, often left as NULL
1882  *
1883  * Same as read_cache_page, but don't wait for page to become unlocked
1884  * after submitting it to the filler.
1885  *
1886  * Read into the page cache. If a page already exists, and PageUptodate() is
1887  * not set, try to fill the page but don't wait for it to become unlocked.
1888  *
1889  * If the page does not get brought uptodate, return -EIO.
1890  */
1891 struct page *read_cache_page_async(struct address_space *mapping,
1892                                 pgoff_t index,
1893                                 int (*filler)(void *, struct page *),
1894                                 void *data)
1895 {
1896         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1897 }
1898 EXPORT_SYMBOL(read_cache_page_async);
1899
1900 static struct page *wait_on_page_read(struct page *page)
1901 {
1902         if (!IS_ERR(page)) {
1903                 wait_on_page_locked(page);
1904                 if (!PageUptodate(page)) {
1905                         page_cache_release(page);
1906                         page = ERR_PTR(-EIO);
1907                 }
1908         }
1909         return page;
1910 }
1911
1912 /**
1913  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1914  * @mapping:    the page's address_space
1915  * @index:      the page index
1916  * @gfp:        the page allocator flags to use if allocating
1917  *
1918  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1919  * any new page allocations done using the specified allocation flags. Note
1920  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1921  * expect to do this atomically or anything like that - but you can pass in
1922  * other page requirements.
1923  *
1924  * If the page does not get brought uptodate, return -EIO.
1925  */
1926 struct page *read_cache_page_gfp(struct address_space *mapping,
1927                                 pgoff_t index,
1928                                 gfp_t gfp)
1929 {
1930         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1931
1932         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1933 }
1934 EXPORT_SYMBOL(read_cache_page_gfp);
1935
1936 /**
1937  * read_cache_page - read into page cache, fill it if needed
1938  * @mapping:    the page's address_space
1939  * @index:      the page index
1940  * @filler:     function to perform the read
1941  * @data:       first arg to filler(data, page) function, often left as NULL
1942  *
1943  * Read into the page cache. If a page already exists, and PageUptodate() is
1944  * not set, try to fill the page then wait for it to become unlocked.
1945  *
1946  * If the page does not get brought uptodate, return -EIO.
1947  */
1948 struct page *read_cache_page(struct address_space *mapping,
1949                                 pgoff_t index,
1950                                 int (*filler)(void *, struct page *),
1951                                 void *data)
1952 {
1953         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1954 }
1955 EXPORT_SYMBOL(read_cache_page);
1956
1957 /*
1958  * The logic we want is
1959  *
1960  *      if suid or (sgid and xgrp)
1961  *              remove privs
1962  */
1963 int should_remove_suid(struct dentry *dentry)
1964 {
1965         mode_t mode = dentry->d_inode->i_mode;
1966         int kill = 0;
1967
1968         /* suid always must be killed */
1969         if (unlikely(mode & S_ISUID))
1970                 kill = ATTR_KILL_SUID;
1971
1972         /*
1973          * sgid without any exec bits is just a mandatory locking mark; leave
1974          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1975          */
1976         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1977                 kill |= ATTR_KILL_SGID;
1978
1979         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1980                 return kill;
1981
1982         return 0;
1983 }
1984 EXPORT_SYMBOL(should_remove_suid);
1985
1986 static int __remove_suid(struct dentry *dentry, int kill)
1987 {
1988         struct iattr newattrs;
1989
1990         newattrs.ia_valid = ATTR_FORCE | kill;
1991         return notify_change(dentry, &newattrs);
1992 }
1993
1994 int file_remove_suid(struct file *file)
1995 {
1996         struct dentry *dentry = file->f_path.dentry;
1997         struct inode *inode = dentry->d_inode;
1998         int killsuid;
1999         int killpriv;
2000         int error = 0;
2001
2002         /* Fast path for nothing security related */
2003         if (IS_NOSEC(inode))
2004                 return 0;
2005
2006         killsuid = should_remove_suid(dentry);
2007         killpriv = security_inode_need_killpriv(dentry);
2008
2009         if (killpriv < 0)
2010                 return killpriv;
2011         if (killpriv)
2012                 error = security_inode_killpriv(dentry);
2013         if (!error && killsuid)
2014                 error = __remove_suid(dentry, killsuid);
2015         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2016                 inode->i_flags |= S_NOSEC;
2017
2018         return error;
2019 }
2020 EXPORT_SYMBOL(file_remove_suid);
2021
2022 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2023                         const struct iovec *iov, size_t base, size_t bytes)
2024 {
2025         size_t copied = 0, left = 0;
2026
2027         while (bytes) {
2028                 char __user *buf = iov->iov_base + base;
2029                 int copy = min(bytes, iov->iov_len - base);
2030
2031                 base = 0;
2032                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2033                 copied += copy;
2034                 bytes -= copy;
2035                 vaddr += copy;
2036                 iov++;
2037
2038                 if (unlikely(left))
2039                         break;
2040         }
2041         return copied - left;
2042 }
2043
2044 /*
2045  * Copy as much as we can into the page and return the number of bytes which
2046  * were successfully copied.  If a fault is encountered then return the number of
2047  * bytes which were copied.
2048  */
2049 size_t iov_iter_copy_from_user_atomic(struct page *page,
2050                 struct iov_iter *i, unsigned long offset, size_t bytes)
2051 {
2052         char *kaddr;
2053         size_t copied;
2054
2055         BUG_ON(!in_atomic());
2056         kaddr = kmap_atomic(page, KM_USER0);
2057         if (likely(i->nr_segs == 1)) {
2058                 int left;
2059                 char __user *buf = i->iov->iov_base + i->iov_offset;
2060                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2061                 copied = bytes - left;
2062         } else {
2063                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2064                                                 i->iov, i->iov_offset, bytes);
2065         }
2066         kunmap_atomic(kaddr, KM_USER0);
2067
2068         return copied;
2069 }
2070 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2071
2072 /*
2073  * This has the same sideeffects and return value as
2074  * iov_iter_copy_from_user_atomic().
2075  * The difference is that it attempts to resolve faults.
2076  * Page must not be locked.
2077  */
2078 size_t iov_iter_copy_from_user(struct page *page,
2079                 struct iov_iter *i, unsigned long offset, size_t bytes)
2080 {
2081         char *kaddr;
2082         size_t copied;
2083
2084         kaddr = kmap(page);
2085         if (likely(i->nr_segs == 1)) {
2086                 int left;
2087                 char __user *buf = i->iov->iov_base + i->iov_offset;
2088                 left = __copy_from_user(kaddr + offset, buf, bytes);
2089                 copied = bytes - left;
2090         } else {
2091                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2092                                                 i->iov, i->iov_offset, bytes);
2093         }
2094         kunmap(page);
2095         return copied;
2096 }
2097 EXPORT_SYMBOL(iov_iter_copy_from_user);
2098
2099 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2100 {
2101         BUG_ON(i->count < bytes);
2102
2103         if (likely(i->nr_segs == 1)) {
2104                 i->iov_offset += bytes;
2105                 i->count -= bytes;
2106         } else {
2107                 const struct iovec *iov = i->iov;
2108                 size_t base = i->iov_offset;
2109
2110                 /*
2111                  * The !iov->iov_len check ensures we skip over unlikely
2112                  * zero-length segments (without overruning the iovec).
2113                  */
2114                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2115                         int copy;
2116
2117                         copy = min(bytes, iov->iov_len - base);
2118                         BUG_ON(!i->count || i->count < copy);
2119                         i->count -= copy;
2120                         bytes -= copy;
2121                         base += copy;
2122                         if (iov->iov_len == base) {
2123                                 iov++;
2124                                 base = 0;
2125                         }
2126                 }
2127                 i->iov = iov;
2128                 i->iov_offset = base;
2129         }
2130 }
2131 EXPORT_SYMBOL(iov_iter_advance);
2132
2133 /*
2134  * Fault in the first iovec of the given iov_iter, to a maximum length
2135  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2136  * accessed (ie. because it is an invalid address).
2137  *
2138  * writev-intensive code may want this to prefault several iovecs -- that
2139  * would be possible (callers must not rely on the fact that _only_ the
2140  * first iovec will be faulted with the current implementation).
2141  */
2142 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2143 {
2144         char __user *buf = i->iov->iov_base + i->iov_offset;
2145         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2146         return fault_in_pages_readable(buf, bytes);
2147 }
2148 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2149
2150 /*
2151  * Return the count of just the current iov_iter segment.
2152  */
2153 size_t iov_iter_single_seg_count(struct iov_iter *i)
2154 {
2155         const struct iovec *iov = i->iov;
2156         if (i->nr_segs == 1)
2157                 return i->count;
2158         else
2159                 return min(i->count, iov->iov_len - i->iov_offset);
2160 }
2161 EXPORT_SYMBOL(iov_iter_single_seg_count);
2162
2163 /*
2164  * Performs necessary checks before doing a write
2165  *
2166  * Can adjust writing position or amount of bytes to write.
2167  * Returns appropriate error code that caller should return or
2168  * zero in case that write should be allowed.
2169  */
2170 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2171 {
2172         struct inode *inode = file->f_mapping->host;
2173         unsigned long limit = rlimit(RLIMIT_FSIZE);
2174
2175         if (unlikely(*pos < 0))
2176                 return -EINVAL;
2177
2178         if (!isblk) {
2179                 /* FIXME: this is for backwards compatibility with 2.4 */
2180                 if (file->f_flags & O_APPEND)
2181                         *pos = i_size_read(inode);
2182
2183                 if (limit != RLIM_INFINITY) {
2184                         if (*pos >= limit) {
2185                                 send_sig(SIGXFSZ, current, 0);
2186                                 return -EFBIG;
2187                         }
2188                         if (*count > limit - (typeof(limit))*pos) {
2189                                 *count = limit - (typeof(limit))*pos;
2190                         }
2191                 }
2192         }
2193
2194         /*
2195          * LFS rule
2196          */
2197         if (unlikely(*pos + *count > MAX_NON_LFS &&
2198                                 !(file->f_flags & O_LARGEFILE))) {
2199                 if (*pos >= MAX_NON_LFS) {
2200                         return -EFBIG;
2201                 }
2202                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2203                         *count = MAX_NON_LFS - (unsigned long)*pos;
2204                 }
2205         }
2206
2207         /*
2208          * Are we about to exceed the fs block limit ?
2209          *
2210          * If we have written data it becomes a short write.  If we have
2211          * exceeded without writing data we send a signal and return EFBIG.
2212          * Linus frestrict idea will clean these up nicely..
2213          */
2214         if (likely(!isblk)) {
2215                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2216                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2217                                 return -EFBIG;
2218                         }
2219                         /* zero-length writes at ->s_maxbytes are OK */
2220                 }
2221
2222                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2223                         *count = inode->i_sb->s_maxbytes - *pos;
2224         } else {
2225 #ifdef CONFIG_BLOCK
2226                 loff_t isize;
2227                 if (bdev_read_only(I_BDEV(inode)))
2228                         return -EPERM;
2229                 isize = i_size_read(inode);
2230                 if (*pos >= isize) {
2231                         if (*count || *pos > isize)
2232                                 return -ENOSPC;
2233                 }
2234
2235                 if (*pos + *count > isize)
2236                         *count = isize - *pos;
2237 #else
2238                 return -EPERM;
2239 #endif
2240         }
2241         return 0;
2242 }
2243 EXPORT_SYMBOL(generic_write_checks);
2244
2245 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2246                                 loff_t pos, unsigned len, unsigned flags,
2247                                 struct page **pagep, void **fsdata)
2248 {
2249         const struct address_space_operations *aops = mapping->a_ops;
2250
2251         return aops->write_begin(file, mapping, pos, len, flags,
2252                                                         pagep, fsdata);
2253 }
2254 EXPORT_SYMBOL(pagecache_write_begin);
2255
2256 int pagecache_write_end(struct file *file, struct address_space *mapping,
2257                                 loff_t pos, unsigned len, unsigned copied,
2258                                 struct page *page, void *fsdata)
2259 {
2260         const struct address_space_operations *aops = mapping->a_ops;
2261
2262         mark_page_accessed(page);
2263         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2264 }
2265 EXPORT_SYMBOL(pagecache_write_end);
2266
2267 ssize_t
2268 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2269                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2270                 size_t count, size_t ocount)
2271 {
2272         struct file     *file = iocb->ki_filp;
2273         struct address_space *mapping = file->f_mapping;
2274         struct inode    *inode = mapping->host;
2275         ssize_t         written;
2276         size_t          write_len;
2277         pgoff_t         end;
2278
2279         if (count != ocount)
2280                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2281
2282         write_len = iov_length(iov, *nr_segs);
2283         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2284
2285         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2286         if (written)
2287                 goto out;
2288
2289         /*
2290          * After a write we want buffered reads to be sure to go to disk to get
2291          * the new data.  We invalidate clean cached page from the region we're
2292          * about to write.  We do this *before* the write so that we can return
2293          * without clobbering -EIOCBQUEUED from ->direct_IO().
2294          */
2295         if (mapping->nrpages) {
2296                 written = invalidate_inode_pages2_range(mapping,
2297                                         pos >> PAGE_CACHE_SHIFT, end);
2298                 /*
2299                  * If a page can not be invalidated, return 0 to fall back
2300                  * to buffered write.
2301                  */
2302                 if (written) {
2303                         if (written == -EBUSY)
2304                                 return 0;
2305                         goto out;
2306                 }
2307         }
2308
2309         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2310
2311         /*
2312          * Finally, try again to invalidate clean pages which might have been
2313          * cached by non-direct readahead, or faulted in by get_user_pages()
2314          * if the source of the write was an mmap'ed region of the file
2315          * we're writing.  Either one is a pretty crazy thing to do,
2316          * so we don't support it 100%.  If this invalidation
2317          * fails, tough, the write still worked...
2318          */
2319         if (mapping->nrpages) {
2320                 invalidate_inode_pages2_range(mapping,
2321                                               pos >> PAGE_CACHE_SHIFT, end);
2322         }
2323
2324         if (written > 0) {
2325                 pos += written;
2326                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2327                         i_size_write(inode, pos);
2328                         mark_inode_dirty(inode);
2329                 }
2330                 *ppos = pos;
2331         }
2332 out:
2333         return written;
2334 }
2335 EXPORT_SYMBOL(generic_file_direct_write);
2336
2337 /*
2338  * Find or create a page at the given pagecache position. Return the locked
2339  * page. This function is specifically for buffered writes.
2340  */
2341 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2342                                         pgoff_t index, unsigned flags)
2343 {
2344         int status;
2345         struct page *page;
2346         gfp_t gfp_notmask = 0;
2347         if (flags & AOP_FLAG_NOFS)
2348                 gfp_notmask = __GFP_FS;
2349 repeat:
2350         page = find_lock_page(mapping, index);
2351         if (page)
2352                 goto found;
2353
2354         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2355         if (!page)
2356                 return NULL;
2357         status = add_to_page_cache_lru(page, mapping, index,
2358                                                 GFP_KERNEL & ~gfp_notmask);
2359         if (unlikely(status)) {
2360                 page_cache_release(page);
2361                 if (status == -EEXIST)
2362                         goto repeat;
2363                 return NULL;
2364         }
2365 found:
2366         wait_on_page_writeback(page);
2367         return page;
2368 }
2369 EXPORT_SYMBOL(grab_cache_page_write_begin);
2370
2371 static ssize_t generic_perform_write(struct file *file,
2372                                 struct iov_iter *i, loff_t pos)
2373 {
2374         struct address_space *mapping = file->f_mapping;
2375         const struct address_space_operations *a_ops = mapping->a_ops;
2376         long status = 0;
2377         ssize_t written = 0;
2378         unsigned int flags = 0;
2379
2380         /*
2381          * Copies from kernel address space cannot fail (NFSD is a big user).
2382          */
2383         if (segment_eq(get_fs(), KERNEL_DS))
2384                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2385
2386         do {
2387                 struct page *page;
2388                 unsigned long offset;   /* Offset into pagecache page */
2389                 unsigned long bytes;    /* Bytes to write to page */
2390                 size_t copied;          /* Bytes copied from user */
2391                 void *fsdata;
2392
2393                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2394                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2395                                                 iov_iter_count(i));
2396
2397 again:
2398
2399                 /*
2400                  * Bring in the user page that we will copy from _first_.
2401                  * Otherwise there's a nasty deadlock on copying from the
2402                  * same page as we're writing to, without it being marked
2403                  * up-to-date.
2404                  *
2405                  * Not only is this an optimisation, but it is also required
2406                  * to check that the address is actually valid, when atomic
2407                  * usercopies are used, below.
2408                  */
2409                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2410                         status = -EFAULT;
2411                         break;
2412                 }
2413
2414                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2415                                                 &page, &fsdata);
2416                 if (unlikely(status))
2417                         break;
2418
2419                 if (mapping_writably_mapped(mapping))
2420                         flush_dcache_page(page);
2421
2422                 pagefault_disable();
2423                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2424                 pagefault_enable();
2425                 flush_dcache_page(page);
2426
2427                 mark_page_accessed(page);
2428                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2429                                                 page, fsdata);
2430                 if (unlikely(status < 0))
2431                         break;
2432                 copied = status;
2433
2434                 cond_resched();
2435
2436                 iov_iter_advance(i, copied);
2437                 if (unlikely(copied == 0)) {
2438                         /*
2439                          * If we were unable to copy any data at all, we must
2440                          * fall back to a single segment length write.
2441                          *
2442                          * If we didn't fallback here, we could livelock
2443                          * because not all segments in the iov can be copied at
2444                          * once without a pagefault.
2445                          */
2446                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2447                                                 iov_iter_single_seg_count(i));
2448                         goto again;
2449                 }
2450                 pos += copied;
2451                 written += copied;
2452
2453                 balance_dirty_pages_ratelimited(mapping);
2454
2455         } while (iov_iter_count(i));
2456
2457         return written ? written : status;
2458 }
2459
2460 ssize_t
2461 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2462                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2463                 size_t count, ssize_t written)
2464 {
2465         struct file *file = iocb->ki_filp;
2466         ssize_t status;
2467         struct iov_iter i;
2468
2469         iov_iter_init(&i, iov, nr_segs, count, written);
2470         status = generic_perform_write(file, &i, pos);
2471
2472         if (likely(status >= 0)) {
2473                 written += status;
2474                 *ppos = pos + status;
2475         }
2476         
2477         return written ? written : status;
2478 }
2479 EXPORT_SYMBOL(generic_file_buffered_write);
2480
2481 /**
2482  * __generic_file_aio_write - write data to a file
2483  * @iocb:       IO state structure (file, offset, etc.)
2484  * @iov:        vector with data to write
2485  * @nr_segs:    number of segments in the vector
2486  * @ppos:       position where to write
2487  *
2488  * This function does all the work needed for actually writing data to a
2489  * file. It does all basic checks, removes SUID from the file, updates
2490  * modification times and calls proper subroutines depending on whether we
2491  * do direct IO or a standard buffered write.
2492  *
2493  * It expects i_mutex to be grabbed unless we work on a block device or similar
2494  * object which does not need locking at all.
2495  *
2496  * This function does *not* take care of syncing data in case of O_SYNC write.
2497  * A caller has to handle it. This is mainly due to the fact that we want to
2498  * avoid syncing under i_mutex.
2499  */
2500 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2501                                  unsigned long nr_segs, loff_t *ppos)
2502 {
2503         struct file *file = iocb->ki_filp;
2504         struct address_space * mapping = file->f_mapping;
2505         size_t ocount;          /* original count */
2506         size_t count;           /* after file limit checks */
2507         struct inode    *inode = mapping->host;
2508         loff_t          pos;
2509         ssize_t         written;
2510         ssize_t         err;
2511
2512         ocount = 0;
2513         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2514         if (err)
2515                 return err;
2516
2517         count = ocount;
2518         pos = *ppos;
2519
2520         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2521
2522         /* We can write back this queue in page reclaim */
2523         current->backing_dev_info = mapping->backing_dev_info;
2524         written = 0;
2525
2526         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2527         if (err)
2528                 goto out;
2529
2530         if (count == 0)
2531                 goto out;
2532
2533         err = file_remove_suid(file);
2534         if (err)
2535                 goto out;
2536
2537         file_update_time(file);
2538
2539         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2540         if (unlikely(file->f_flags & O_DIRECT)) {
2541                 loff_t endbyte;
2542                 ssize_t written_buffered;
2543
2544                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2545                                                         ppos, count, ocount);
2546                 if (written < 0 || written == count)
2547                         goto out;
2548                 /*
2549                  * direct-io write to a hole: fall through to buffered I/O
2550                  * for completing the rest of the request.
2551                  */
2552                 pos += written;
2553                 count -= written;
2554                 written_buffered = generic_file_buffered_write(iocb, iov,
2555                                                 nr_segs, pos, ppos, count,
2556                                                 written);
2557                 /*
2558                  * If generic_file_buffered_write() retuned a synchronous error
2559                  * then we want to return the number of bytes which were
2560                  * direct-written, or the error code if that was zero.  Note
2561                  * that this differs from normal direct-io semantics, which
2562                  * will return -EFOO even if some bytes were written.
2563                  */
2564                 if (written_buffered < 0) {
2565                         err = written_buffered;
2566                         goto out;
2567                 }
2568
2569                 /*
2570                  * We need to ensure that the page cache pages are written to
2571                  * disk and invalidated to preserve the expected O_DIRECT
2572                  * semantics.
2573                  */
2574                 endbyte = pos + written_buffered - written - 1;
2575                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2576                 if (err == 0) {
2577                         written = written_buffered;
2578                         invalidate_mapping_pages(mapping,
2579                                                  pos >> PAGE_CACHE_SHIFT,
2580                                                  endbyte >> PAGE_CACHE_SHIFT);
2581                 } else {
2582                         /*
2583                          * We don't know how much we wrote, so just return
2584                          * the number of bytes which were direct-written
2585                          */
2586                 }
2587         } else {
2588                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2589                                 pos, ppos, count, written);
2590         }
2591 out:
2592         current->backing_dev_info = NULL;
2593         return written ? written : err;
2594 }
2595 EXPORT_SYMBOL(__generic_file_aio_write);
2596
2597 /**
2598  * generic_file_aio_write - write data to a file
2599  * @iocb:       IO state structure
2600  * @iov:        vector with data to write
2601  * @nr_segs:    number of segments in the vector
2602  * @pos:        position in file where to write
2603  *
2604  * This is a wrapper around __generic_file_aio_write() to be used by most
2605  * filesystems. It takes care of syncing the file in case of O_SYNC file
2606  * and acquires i_mutex as needed.
2607  */
2608 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2609                 unsigned long nr_segs, loff_t pos)
2610 {
2611         struct file *file = iocb->ki_filp;
2612         struct inode *inode = file->f_mapping->host;
2613         struct blk_plug plug;
2614         ssize_t ret;
2615
2616         BUG_ON(iocb->ki_pos != pos);
2617
2618         mutex_lock(&inode->i_mutex);
2619         blk_start_plug(&plug);
2620         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2621         mutex_unlock(&inode->i_mutex);
2622
2623         if (ret > 0 || ret == -EIOCBQUEUED) {
2624                 ssize_t err;
2625
2626                 err = generic_write_sync(file, pos, ret);
2627                 if (err < 0 && ret > 0)
2628                         ret = err;
2629         }
2630         blk_finish_plug(&plug);
2631         return ret;
2632 }
2633 EXPORT_SYMBOL(generic_file_aio_write);
2634
2635 /**
2636  * try_to_release_page() - release old fs-specific metadata on a page
2637  *
2638  * @page: the page which the kernel is trying to free
2639  * @gfp_mask: memory allocation flags (and I/O mode)
2640  *
2641  * The address_space is to try to release any data against the page
2642  * (presumably at page->private).  If the release was successful, return `1'.
2643  * Otherwise return zero.
2644  *
2645  * This may also be called if PG_fscache is set on a page, indicating that the
2646  * page is known to the local caching routines.
2647  *
2648  * The @gfp_mask argument specifies whether I/O may be performed to release
2649  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2650  *
2651  */
2652 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2653 {
2654         struct address_space * const mapping = page->mapping;
2655
2656         BUG_ON(!PageLocked(page));
2657         if (PageWriteback(page))
2658                 return 0;
2659
2660         if (mapping && mapping->a_ops->releasepage)
2661                 return mapping->a_ops->releasepage(page, gfp_mask);
2662         return try_to_free_buffers(page);
2663 }
2664
2665 EXPORT_SYMBOL(try_to_release_page);