mm: sync vmalloc address space page tables in alloc_vm_area()
[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/cleancache.h>
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_mutex              (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_mutex
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  bdi->wb.list_lock
81  *    sb_lock                   (fs/fs-writeback.c)
82  *    ->mapping->tree_lock      (__sync_single_inode)
83  *
84  *  ->i_mmap_mutex
85  *    ->anon_vma.lock           (vma_adjust)
86  *
87  *  ->anon_vma.lock
88  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
89  *
90  *  ->page_table_lock or pte_lock
91  *    ->swap_lock               (try_to_unmap_one)
92  *    ->private_lock            (try_to_unmap_one)
93  *    ->tree_lock               (try_to_unmap_one)
94  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
95  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
96  *    ->private_lock            (page_remove_rmap->set_page_dirty)
97  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
98  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
99  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
100  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
101  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
102  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
103  *
104  *  (code doesn't rely on that order, so you could switch it around)
105  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
106  *    ->i_mmap_mutex
107  */
108
109 /*
110  * Delete a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold the mapping's tree_lock.
113  */
114 void __delete_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         /*
119          * if we're uptodate, flush out into the cleancache, otherwise
120          * invalidate any existing cleancache entries.  We can't leave
121          * stale data around in the cleancache once our page is gone
122          */
123         if (PageUptodate(page) && PageMappedToDisk(page))
124                 cleancache_put_page(page);
125         else
126                 cleancache_flush_page(mapping, page);
127
128         radix_tree_delete(&mapping->page_tree, page->index);
129         page->mapping = NULL;
130         /* Leave page->index set: truncation lookup relies upon it */
131         mapping->nrpages--;
132         __dec_zone_page_state(page, NR_FILE_PAGES);
133         if (PageSwapBacked(page))
134                 __dec_zone_page_state(page, NR_SHMEM);
135         BUG_ON(page_mapped(page));
136
137         /*
138          * Some filesystems seem to re-dirty the page even after
139          * the VM has canceled the dirty bit (eg ext3 journaling).
140          *
141          * Fix it up by doing a final dirty accounting check after
142          * having removed the page entirely.
143          */
144         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
145                 dec_zone_page_state(page, NR_FILE_DIRTY);
146                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
147         }
148 }
149
150 /**
151  * delete_from_page_cache - delete page from page cache
152  * @page: the page which the kernel is trying to remove from page cache
153  *
154  * This must be called only on pages that have been verified to be in the page
155  * cache and locked.  It will never put the page into the free list, the caller
156  * has a reference on the page.
157  */
158 void delete_from_page_cache(struct page *page)
159 {
160         struct address_space *mapping = page->mapping;
161         void (*freepage)(struct page *);
162
163         BUG_ON(!PageLocked(page));
164
165         freepage = mapping->a_ops->freepage;
166         spin_lock_irq(&mapping->tree_lock);
167         __delete_from_page_cache(page);
168         spin_unlock_irq(&mapping->tree_lock);
169         mem_cgroup_uncharge_cache_page(page);
170
171         if (freepage)
172                 freepage(page);
173         page_cache_release(page);
174 }
175 EXPORT_SYMBOL(delete_from_page_cache);
176
177 static int sleep_on_page(void *word)
178 {
179         io_schedule();
180         return 0;
181 }
182
183 static int sleep_on_page_killable(void *word)
184 {
185         sleep_on_page(word);
186         return fatal_signal_pending(current) ? -EINTR : 0;
187 }
188
189 /**
190  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
191  * @mapping:    address space structure to write
192  * @start:      offset in bytes where the range starts
193  * @end:        offset in bytes where the range ends (inclusive)
194  * @sync_mode:  enable synchronous operation
195  *
196  * Start writeback against all of a mapping's dirty pages that lie
197  * within the byte offsets <start, end> inclusive.
198  *
199  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
200  * opposed to a regular memory cleansing writeback.  The difference between
201  * these two operations is that if a dirty page/buffer is encountered, it must
202  * be waited upon, and not just skipped over.
203  */
204 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
205                                 loff_t end, int sync_mode)
206 {
207         int ret;
208         struct writeback_control wbc = {
209                 .sync_mode = sync_mode,
210                 .nr_to_write = LONG_MAX,
211                 .range_start = start,
212                 .range_end = end,
213         };
214
215         if (!mapping_cap_writeback_dirty(mapping))
216                 return 0;
217
218         ret = do_writepages(mapping, &wbc);
219         return ret;
220 }
221
222 static inline int __filemap_fdatawrite(struct address_space *mapping,
223         int sync_mode)
224 {
225         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
226 }
227
228 int filemap_fdatawrite(struct address_space *mapping)
229 {
230         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
231 }
232 EXPORT_SYMBOL(filemap_fdatawrite);
233
234 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
235                                 loff_t end)
236 {
237         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
238 }
239 EXPORT_SYMBOL(filemap_fdatawrite_range);
240
241 /**
242  * filemap_flush - mostly a non-blocking flush
243  * @mapping:    target address_space
244  *
245  * This is a mostly non-blocking flush.  Not suitable for data-integrity
246  * purposes - I/O may not be started against all dirty pages.
247  */
248 int filemap_flush(struct address_space *mapping)
249 {
250         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
251 }
252 EXPORT_SYMBOL(filemap_flush);
253
254 /**
255  * filemap_fdatawait_range - wait for writeback to complete
256  * @mapping:            address space structure to wait for
257  * @start_byte:         offset in bytes where the range starts
258  * @end_byte:           offset in bytes where the range ends (inclusive)
259  *
260  * Walk the list of under-writeback pages of the given address space
261  * in the given range and wait for all of them.
262  */
263 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
264                             loff_t end_byte)
265 {
266         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
267         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271
272         if (end_byte < start_byte)
273                 return 0;
274
275         pagevec_init(&pvec, 0);
276         while ((index <= end) &&
277                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
278                         PAGECACHE_TAG_WRITEBACK,
279                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
280                 unsigned i;
281
282                 for (i = 0; i < nr_pages; i++) {
283                         struct page *page = pvec.pages[i];
284
285                         /* until radix tree lookup accepts end_index */
286                         if (page->index > end)
287                                 continue;
288
289                         wait_on_page_writeback(page);
290                         if (TestClearPageError(page))
291                                 ret = -EIO;
292                 }
293                 pagevec_release(&pvec);
294                 cond_resched();
295         }
296
297         /* Check for outstanding write errors */
298         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
299                 ret = -ENOSPC;
300         if (test_and_clear_bit(AS_EIO, &mapping->flags))
301                 ret = -EIO;
302
303         return ret;
304 }
305 EXPORT_SYMBOL(filemap_fdatawait_range);
306
307 /**
308  * filemap_fdatawait - wait for all under-writeback pages to complete
309  * @mapping: address space structure to wait for
310  *
311  * Walk the list of under-writeback pages of the given address space
312  * and wait for all of them.
313  */
314 int filemap_fdatawait(struct address_space *mapping)
315 {
316         loff_t i_size = i_size_read(mapping->host);
317
318         if (i_size == 0)
319                 return 0;
320
321         return filemap_fdatawait_range(mapping, 0, i_size - 1);
322 }
323 EXPORT_SYMBOL(filemap_fdatawait);
324
325 int filemap_write_and_wait(struct address_space *mapping)
326 {
327         int err = 0;
328
329         if (mapping->nrpages) {
330                 err = filemap_fdatawrite(mapping);
331                 /*
332                  * Even if the above returned error, the pages may be
333                  * written partially (e.g. -ENOSPC), so we wait for it.
334                  * But the -EIO is special case, it may indicate the worst
335                  * thing (e.g. bug) happened, so we avoid waiting for it.
336                  */
337                 if (err != -EIO) {
338                         int err2 = filemap_fdatawait(mapping);
339                         if (!err)
340                                 err = err2;
341                 }
342         }
343         return err;
344 }
345 EXPORT_SYMBOL(filemap_write_and_wait);
346
347 /**
348  * filemap_write_and_wait_range - write out & wait on a file range
349  * @mapping:    the address_space for the pages
350  * @lstart:     offset in bytes where the range starts
351  * @lend:       offset in bytes where the range ends (inclusive)
352  *
353  * Write out and wait upon file offsets lstart->lend, inclusive.
354  *
355  * Note that `lend' is inclusive (describes the last byte to be written) so
356  * that this function can be used to write to the very end-of-file (end = -1).
357  */
358 int filemap_write_and_wait_range(struct address_space *mapping,
359                                  loff_t lstart, loff_t lend)
360 {
361         int err = 0;
362
363         if (mapping->nrpages) {
364                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
365                                                  WB_SYNC_ALL);
366                 /* See comment of filemap_write_and_wait() */
367                 if (err != -EIO) {
368                         int err2 = filemap_fdatawait_range(mapping,
369                                                 lstart, lend);
370                         if (!err)
371                                 err = err2;
372                 }
373         }
374         return err;
375 }
376 EXPORT_SYMBOL(filemap_write_and_wait_range);
377
378 /**
379  * replace_page_cache_page - replace a pagecache page with a new one
380  * @old:        page to be replaced
381  * @new:        page to replace with
382  * @gfp_mask:   allocation mode
383  *
384  * This function replaces a page in the pagecache with a new one.  On
385  * success it acquires the pagecache reference for the new page and
386  * drops it for the old page.  Both the old and new pages must be
387  * locked.  This function does not add the new page to the LRU, the
388  * caller must do that.
389  *
390  * The remove + add is atomic.  The only way this function can fail is
391  * memory allocation failure.
392  */
393 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
394 {
395         int error;
396         struct mem_cgroup *memcg = NULL;
397
398         VM_BUG_ON(!PageLocked(old));
399         VM_BUG_ON(!PageLocked(new));
400         VM_BUG_ON(new->mapping);
401
402         /*
403          * This is not page migration, but prepare_migration and
404          * end_migration does enough work for charge replacement.
405          *
406          * In the longer term we probably want a specialized function
407          * for moving the charge from old to new in a more efficient
408          * manner.
409          */
410         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
411         if (error)
412                 return error;
413
414         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
415         if (!error) {
416                 struct address_space *mapping = old->mapping;
417                 void (*freepage)(struct page *);
418
419                 pgoff_t offset = old->index;
420                 freepage = mapping->a_ops->freepage;
421
422                 page_cache_get(new);
423                 new->mapping = mapping;
424                 new->index = offset;
425
426                 spin_lock_irq(&mapping->tree_lock);
427                 __delete_from_page_cache(old);
428                 error = radix_tree_insert(&mapping->page_tree, offset, new);
429                 BUG_ON(error);
430                 mapping->nrpages++;
431                 __inc_zone_page_state(new, NR_FILE_PAGES);
432                 if (PageSwapBacked(new))
433                         __inc_zone_page_state(new, NR_SHMEM);
434                 spin_unlock_irq(&mapping->tree_lock);
435                 radix_tree_preload_end();
436                 if (freepage)
437                         freepage(old);
438                 page_cache_release(old);
439                 mem_cgroup_end_migration(memcg, old, new, true);
440         } else {
441                 mem_cgroup_end_migration(memcg, old, new, false);
442         }
443
444         return error;
445 }
446 EXPORT_SYMBOL_GPL(replace_page_cache_page);
447
448 /**
449  * add_to_page_cache_locked - add a locked page to the pagecache
450  * @page:       page to add
451  * @mapping:    the page's address_space
452  * @offset:     page index
453  * @gfp_mask:   page allocation mode
454  *
455  * This function is used to add a page to the pagecache. It must be locked.
456  * This function does not add the page to the LRU.  The caller must do that.
457  */
458 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
459                 pgoff_t offset, gfp_t gfp_mask)
460 {
461         int error;
462
463         VM_BUG_ON(!PageLocked(page));
464         VM_BUG_ON(PageSwapBacked(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                         spin_unlock_irq(&mapping->tree_lock);
483                 } else {
484                         page->mapping = NULL;
485                         /* Leave page->index set: truncation relies upon it */
486                         spin_unlock_irq(&mapping->tree_lock);
487                         mem_cgroup_uncharge_cache_page(page);
488                         page_cache_release(page);
489                 }
490                 radix_tree_preload_end();
491         } else
492                 mem_cgroup_uncharge_cache_page(page);
493 out:
494         return error;
495 }
496 EXPORT_SYMBOL(add_to_page_cache_locked);
497
498 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
499                                 pgoff_t offset, gfp_t gfp_mask)
500 {
501         int ret;
502
503         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
504         if (ret == 0)
505                 lru_cache_add_file(page);
506         return ret;
507 }
508 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
509
510 #ifdef CONFIG_NUMA
511 struct page *__page_cache_alloc(gfp_t gfp)
512 {
513         int n;
514         struct page *page;
515
516         if (cpuset_do_page_mem_spread()) {
517                 get_mems_allowed();
518                 n = cpuset_mem_spread_node();
519                 page = alloc_pages_exact_node(n, gfp, 0);
520                 put_mems_allowed();
521                 return page;
522         }
523         return alloc_pages(gfp, 0);
524 }
525 EXPORT_SYMBOL(__page_cache_alloc);
526 #endif
527
528 /*
529  * In order to wait for pages to become available there must be
530  * waitqueues associated with pages. By using a hash table of
531  * waitqueues where the bucket discipline is to maintain all
532  * waiters on the same queue and wake all when any of the pages
533  * become available, and for the woken contexts to check to be
534  * sure the appropriate page became available, this saves space
535  * at a cost of "thundering herd" phenomena during rare hash
536  * collisions.
537  */
538 static wait_queue_head_t *page_waitqueue(struct page *page)
539 {
540         const struct zone *zone = page_zone(page);
541
542         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
543 }
544
545 static inline void wake_up_page(struct page *page, int bit)
546 {
547         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
548 }
549
550 void wait_on_page_bit(struct page *page, int bit_nr)
551 {
552         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
553
554         if (test_bit(bit_nr, &page->flags))
555                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
556                                                         TASK_UNINTERRUPTIBLE);
557 }
558 EXPORT_SYMBOL(wait_on_page_bit);
559
560 int wait_on_page_bit_killable(struct page *page, int bit_nr)
561 {
562         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
563
564         if (!test_bit(bit_nr, &page->flags))
565                 return 0;
566
567         return __wait_on_bit(page_waitqueue(page), &wait,
568                              sleep_on_page_killable, TASK_KILLABLE);
569 }
570
571 /**
572  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
573  * @page: Page defining the wait queue of interest
574  * @waiter: Waiter to add to the queue
575  *
576  * Add an arbitrary @waiter to the wait queue for the nominated @page.
577  */
578 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
579 {
580         wait_queue_head_t *q = page_waitqueue(page);
581         unsigned long flags;
582
583         spin_lock_irqsave(&q->lock, flags);
584         __add_wait_queue(q, waiter);
585         spin_unlock_irqrestore(&q->lock, flags);
586 }
587 EXPORT_SYMBOL_GPL(add_page_wait_queue);
588
589 /**
590  * unlock_page - unlock a locked page
591  * @page: the page
592  *
593  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
594  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
595  * mechananism between PageLocked pages and PageWriteback pages is shared.
596  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
597  *
598  * The mb is necessary to enforce ordering between the clear_bit and the read
599  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
600  */
601 void unlock_page(struct page *page)
602 {
603         VM_BUG_ON(!PageLocked(page));
604         clear_bit_unlock(PG_locked, &page->flags);
605         smp_mb__after_clear_bit();
606         wake_up_page(page, PG_locked);
607 }
608 EXPORT_SYMBOL(unlock_page);
609
610 /**
611  * end_page_writeback - end writeback against a page
612  * @page: the page
613  */
614 void end_page_writeback(struct page *page)
615 {
616         if (TestClearPageReclaim(page))
617                 rotate_reclaimable_page(page);
618
619         if (!test_clear_page_writeback(page))
620                 BUG();
621
622         smp_mb__after_clear_bit();
623         wake_up_page(page, PG_writeback);
624 }
625 EXPORT_SYMBOL(end_page_writeback);
626
627 /**
628  * __lock_page - get a lock on the page, assuming we need to sleep to get it
629  * @page: the page to lock
630  */
631 void __lock_page(struct page *page)
632 {
633         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
634
635         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
636                                                         TASK_UNINTERRUPTIBLE);
637 }
638 EXPORT_SYMBOL(__lock_page);
639
640 int __lock_page_killable(struct page *page)
641 {
642         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
643
644         return __wait_on_bit_lock(page_waitqueue(page), &wait,
645                                         sleep_on_page_killable, TASK_KILLABLE);
646 }
647 EXPORT_SYMBOL_GPL(__lock_page_killable);
648
649 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
650                          unsigned int flags)
651 {
652         if (flags & FAULT_FLAG_ALLOW_RETRY) {
653                 /*
654                  * CAUTION! In this case, mmap_sem is not released
655                  * even though return 0.
656                  */
657                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
658                         return 0;
659
660                 up_read(&mm->mmap_sem);
661                 if (flags & FAULT_FLAG_KILLABLE)
662                         wait_on_page_locked_killable(page);
663                 else
664                         wait_on_page_locked(page);
665                 return 0;
666         } else {
667                 if (flags & FAULT_FLAG_KILLABLE) {
668                         int ret;
669
670                         ret = __lock_page_killable(page);
671                         if (ret) {
672                                 up_read(&mm->mmap_sem);
673                                 return 0;
674                         }
675                 } else
676                         __lock_page(page);
677                 return 1;
678         }
679 }
680
681 /**
682  * find_get_page - find and get a page reference
683  * @mapping: the address_space to search
684  * @offset: the page index
685  *
686  * Is there a pagecache struct page at the given (mapping, offset) tuple?
687  * If yes, increment its refcount and return it; if no, return NULL.
688  */
689 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
690 {
691         void **pagep;
692         struct page *page;
693
694         rcu_read_lock();
695 repeat:
696         page = NULL;
697         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
698         if (pagep) {
699                 page = radix_tree_deref_slot(pagep);
700                 if (unlikely(!page))
701                         goto out;
702                 if (radix_tree_exception(page)) {
703                         if (radix_tree_deref_retry(page))
704                                 goto repeat;
705                         /*
706                          * Otherwise, shmem/tmpfs must be storing a swap entry
707                          * here as an exceptional entry: so return it without
708                          * attempting to raise page count.
709                          */
710                         goto out;
711                 }
712                 if (!page_cache_get_speculative(page))
713                         goto repeat;
714
715                 /*
716                  * Has the page moved?
717                  * This is part of the lockless pagecache protocol. See
718                  * include/linux/pagemap.h for details.
719                  */
720                 if (unlikely(page != *pagep)) {
721                         page_cache_release(page);
722                         goto repeat;
723                 }
724         }
725 out:
726         rcu_read_unlock();
727
728         return page;
729 }
730 EXPORT_SYMBOL(find_get_page);
731
732 /**
733  * find_lock_page - locate, pin and lock a pagecache page
734  * @mapping: the address_space to search
735  * @offset: the page index
736  *
737  * Locates the desired pagecache page, locks it, increments its reference
738  * count and returns its address.
739  *
740  * Returns zero if the page was not present. find_lock_page() may sleep.
741  */
742 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
743 {
744         struct page *page;
745
746 repeat:
747         page = find_get_page(mapping, offset);
748         if (page && !radix_tree_exception(page)) {
749                 lock_page(page);
750                 /* Has the page been truncated? */
751                 if (unlikely(page->mapping != mapping)) {
752                         unlock_page(page);
753                         page_cache_release(page);
754                         goto repeat;
755                 }
756                 VM_BUG_ON(page->index != offset);
757         }
758         return page;
759 }
760 EXPORT_SYMBOL(find_lock_page);
761
762 /**
763  * find_or_create_page - locate or add a pagecache page
764  * @mapping: the page's address_space
765  * @index: the page's index into the mapping
766  * @gfp_mask: page allocation mode
767  *
768  * Locates a page in the pagecache.  If the page is not present, a new page
769  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
770  * LRU list.  The returned page is locked and has its reference count
771  * incremented.
772  *
773  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
774  * allocation!
775  *
776  * find_or_create_page() returns the desired page's address, or zero on
777  * memory exhaustion.
778  */
779 struct page *find_or_create_page(struct address_space *mapping,
780                 pgoff_t index, gfp_t gfp_mask)
781 {
782         struct page *page;
783         int err;
784 repeat:
785         page = find_lock_page(mapping, index);
786         if (!page) {
787                 page = __page_cache_alloc(gfp_mask);
788                 if (!page)
789                         return NULL;
790                 /*
791                  * We want a regular kernel memory (not highmem or DMA etc)
792                  * allocation for the radix tree nodes, but we need to honour
793                  * the context-specific requirements the caller has asked for.
794                  * GFP_RECLAIM_MASK collects those requirements.
795                  */
796                 err = add_to_page_cache_lru(page, mapping, index,
797                         (gfp_mask & GFP_RECLAIM_MASK));
798                 if (unlikely(err)) {
799                         page_cache_release(page);
800                         page = NULL;
801                         if (err == -EEXIST)
802                                 goto repeat;
803                 }
804         }
805         return page;
806 }
807 EXPORT_SYMBOL(find_or_create_page);
808
809 /**
810  * find_get_pages - gang pagecache lookup
811  * @mapping:    The address_space to search
812  * @start:      The starting page index
813  * @nr_pages:   The maximum number of pages
814  * @pages:      Where the resulting pages are placed
815  *
816  * find_get_pages() will search for and return a group of up to
817  * @nr_pages pages in the mapping.  The pages are placed at @pages.
818  * find_get_pages() takes a reference against the returned pages.
819  *
820  * The search returns a group of mapping-contiguous pages with ascending
821  * indexes.  There may be holes in the indices due to not-present pages.
822  *
823  * find_get_pages() returns the number of pages which were found.
824  */
825 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
826                             unsigned int nr_pages, struct page **pages)
827 {
828         unsigned int i;
829         unsigned int ret;
830         unsigned int nr_found;
831
832         rcu_read_lock();
833 restart:
834         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
835                                 (void ***)pages, NULL, start, nr_pages);
836         ret = 0;
837         for (i = 0; i < nr_found; i++) {
838                 struct page *page;
839 repeat:
840                 page = radix_tree_deref_slot((void **)pages[i]);
841                 if (unlikely(!page))
842                         continue;
843
844                 if (radix_tree_exception(page)) {
845                         if (radix_tree_deref_retry(page)) {
846                                 /*
847                                  * Transient condition which can only trigger
848                                  * when entry at index 0 moves out of or back
849                                  * to root: none yet gotten, safe to restart.
850                                  */
851                                 WARN_ON(start | i);
852                                 goto restart;
853                         }
854                         /*
855                          * Otherwise, shmem/tmpfs must be storing a swap entry
856                          * here as an exceptional entry: so skip over it -
857                          * we only reach this from invalidate_mapping_pages().
858                          */
859                         continue;
860                 }
861
862                 if (!page_cache_get_speculative(page))
863                         goto repeat;
864
865                 /* Has the page moved? */
866                 if (unlikely(page != *((void **)pages[i]))) {
867                         page_cache_release(page);
868                         goto repeat;
869                 }
870
871                 pages[ret] = page;
872                 ret++;
873         }
874
875         /*
876          * If all entries were removed before we could secure them,
877          * try again, because callers stop trying once 0 is returned.
878          */
879         if (unlikely(!ret && nr_found))
880                 goto restart;
881         rcu_read_unlock();
882         return ret;
883 }
884
885 /**
886  * find_get_pages_contig - gang contiguous pagecache lookup
887  * @mapping:    The address_space to search
888  * @index:      The starting page index
889  * @nr_pages:   The maximum number of pages
890  * @pages:      Where the resulting pages are placed
891  *
892  * find_get_pages_contig() works exactly like find_get_pages(), except
893  * that the returned number of pages are guaranteed to be contiguous.
894  *
895  * find_get_pages_contig() returns the number of pages which were found.
896  */
897 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
898                                unsigned int nr_pages, struct page **pages)
899 {
900         unsigned int i;
901         unsigned int ret;
902         unsigned int nr_found;
903
904         rcu_read_lock();
905 restart:
906         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
907                                 (void ***)pages, NULL, index, nr_pages);
908         ret = 0;
909         for (i = 0; i < nr_found; i++) {
910                 struct page *page;
911 repeat:
912                 page = radix_tree_deref_slot((void **)pages[i]);
913                 if (unlikely(!page))
914                         continue;
915
916                 if (radix_tree_exception(page)) {
917                         if (radix_tree_deref_retry(page)) {
918                                 /*
919                                  * Transient condition which can only trigger
920                                  * when entry at index 0 moves out of or back
921                                  * to root: none yet gotten, safe to restart.
922                                  */
923                                 goto restart;
924                         }
925                         /*
926                          * Otherwise, shmem/tmpfs must be storing a swap entry
927                          * here as an exceptional entry: so stop looking for
928                          * contiguous pages.
929                          */
930                         break;
931                 }
932
933                 if (!page_cache_get_speculative(page))
934                         goto repeat;
935
936                 /* Has the page moved? */
937                 if (unlikely(page != *((void **)pages[i]))) {
938                         page_cache_release(page);
939                         goto repeat;
940                 }
941
942                 /*
943                  * must check mapping and index after taking the ref.
944                  * otherwise we can get both false positives and false
945                  * negatives, which is just confusing to the caller.
946                  */
947                 if (page->mapping == NULL || page->index != index) {
948                         page_cache_release(page);
949                         break;
950                 }
951
952                 pages[ret] = page;
953                 ret++;
954                 index++;
955         }
956         rcu_read_unlock();
957         return ret;
958 }
959 EXPORT_SYMBOL(find_get_pages_contig);
960
961 /**
962  * find_get_pages_tag - find and return pages that match @tag
963  * @mapping:    the address_space to search
964  * @index:      the starting page index
965  * @tag:        the tag index
966  * @nr_pages:   the maximum number of pages
967  * @pages:      where the resulting pages are placed
968  *
969  * Like find_get_pages, except we only return pages which are tagged with
970  * @tag.   We update @index to index the next page for the traversal.
971  */
972 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
973                         int tag, unsigned int nr_pages, struct page **pages)
974 {
975         unsigned int i;
976         unsigned int ret;
977         unsigned int nr_found;
978
979         rcu_read_lock();
980 restart:
981         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
982                                 (void ***)pages, *index, nr_pages, tag);
983         ret = 0;
984         for (i = 0; i < nr_found; i++) {
985                 struct page *page;
986 repeat:
987                 page = radix_tree_deref_slot((void **)pages[i]);
988                 if (unlikely(!page))
989                         continue;
990
991                 if (radix_tree_exception(page)) {
992                         if (radix_tree_deref_retry(page)) {
993                                 /*
994                                  * Transient condition which can only trigger
995                                  * when entry at index 0 moves out of or back
996                                  * to root: none yet gotten, safe to restart.
997                                  */
998                                 goto restart;
999                         }
1000                         /*
1001                          * This function is never used on a shmem/tmpfs
1002                          * mapping, so a swap entry won't be found here.
1003                          */
1004                         BUG();
1005                 }
1006
1007                 if (!page_cache_get_speculative(page))
1008                         goto repeat;
1009
1010                 /* Has the page moved? */
1011                 if (unlikely(page != *((void **)pages[i]))) {
1012                         page_cache_release(page);
1013                         goto repeat;
1014                 }
1015
1016                 pages[ret] = page;
1017                 ret++;
1018         }
1019
1020         /*
1021          * If all entries were removed before we could secure them,
1022          * try again, because callers stop trying once 0 is returned.
1023          */
1024         if (unlikely(!ret && nr_found))
1025                 goto restart;
1026         rcu_read_unlock();
1027
1028         if (ret)
1029                 *index = pages[ret - 1]->index + 1;
1030
1031         return ret;
1032 }
1033 EXPORT_SYMBOL(find_get_pages_tag);
1034
1035 /**
1036  * grab_cache_page_nowait - returns locked page at given index in given cache
1037  * @mapping: target address_space
1038  * @index: the page index
1039  *
1040  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1041  * This is intended for speculative data generators, where the data can
1042  * be regenerated if the page couldn't be grabbed.  This routine should
1043  * be safe to call while holding the lock for another page.
1044  *
1045  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1046  * and deadlock against the caller's locked page.
1047  */
1048 struct page *
1049 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1050 {
1051         struct page *page = find_get_page(mapping, index);
1052
1053         if (page) {
1054                 if (trylock_page(page))
1055                         return page;
1056                 page_cache_release(page);
1057                 return NULL;
1058         }
1059         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1060         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1061                 page_cache_release(page);
1062                 page = NULL;
1063         }
1064         return page;
1065 }
1066 EXPORT_SYMBOL(grab_cache_page_nowait);
1067
1068 /*
1069  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1070  * a _large_ part of the i/o request. Imagine the worst scenario:
1071  *
1072  *      ---R__________________________________________B__________
1073  *         ^ reading here                             ^ bad block(assume 4k)
1074  *
1075  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1076  * => failing the whole request => read(R) => read(R+1) =>
1077  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1078  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1079  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1080  *
1081  * It is going insane. Fix it by quickly scaling down the readahead size.
1082  */
1083 static void shrink_readahead_size_eio(struct file *filp,
1084                                         struct file_ra_state *ra)
1085 {
1086         ra->ra_pages /= 4;
1087 }
1088
1089 /**
1090  * do_generic_file_read - generic file read routine
1091  * @filp:       the file to read
1092  * @ppos:       current file position
1093  * @desc:       read_descriptor
1094  * @actor:      read method
1095  *
1096  * This is a generic file read routine, and uses the
1097  * mapping->a_ops->readpage() function for the actual low-level stuff.
1098  *
1099  * This is really ugly. But the goto's actually try to clarify some
1100  * of the logic when it comes to error handling etc.
1101  */
1102 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1103                 read_descriptor_t *desc, read_actor_t actor)
1104 {
1105         struct address_space *mapping = filp->f_mapping;
1106         struct inode *inode = mapping->host;
1107         struct file_ra_state *ra = &filp->f_ra;
1108         pgoff_t index;
1109         pgoff_t last_index;
1110         pgoff_t prev_index;
1111         unsigned long offset;      /* offset into pagecache page */
1112         unsigned int prev_offset;
1113         int error;
1114
1115         index = *ppos >> PAGE_CACHE_SHIFT;
1116         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1117         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1118         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1119         offset = *ppos & ~PAGE_CACHE_MASK;
1120
1121         for (;;) {
1122                 struct page *page;
1123                 pgoff_t end_index;
1124                 loff_t isize;
1125                 unsigned long nr, ret;
1126
1127                 cond_resched();
1128 find_page:
1129                 page = find_get_page(mapping, index);
1130                 if (!page) {
1131                         page_cache_sync_readahead(mapping,
1132                                         ra, filp,
1133                                         index, last_index - index);
1134                         page = find_get_page(mapping, index);
1135                         if (unlikely(page == NULL))
1136                                 goto no_cached_page;
1137                 }
1138                 if (PageReadahead(page)) {
1139                         page_cache_async_readahead(mapping,
1140                                         ra, filp, page,
1141                                         index, last_index - index);
1142                 }
1143                 if (!PageUptodate(page)) {
1144                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1145                                         !mapping->a_ops->is_partially_uptodate)
1146                                 goto page_not_up_to_date;
1147                         if (!trylock_page(page))
1148                                 goto page_not_up_to_date;
1149                         /* Did it get truncated before we got the lock? */
1150                         if (!page->mapping)
1151                                 goto page_not_up_to_date_locked;
1152                         if (!mapping->a_ops->is_partially_uptodate(page,
1153                                                                 desc, offset))
1154                                 goto page_not_up_to_date_locked;
1155                         unlock_page(page);
1156                 }
1157 page_ok:
1158                 /*
1159                  * i_size must be checked after we know the page is Uptodate.
1160                  *
1161                  * Checking i_size after the check allows us to calculate
1162                  * the correct value for "nr", which means the zero-filled
1163                  * part of the page is not copied back to userspace (unless
1164                  * another truncate extends the file - this is desired though).
1165                  */
1166
1167                 isize = i_size_read(inode);
1168                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1169                 if (unlikely(!isize || index > end_index)) {
1170                         page_cache_release(page);
1171                         goto out;
1172                 }
1173
1174                 /* nr is the maximum number of bytes to copy from this page */
1175                 nr = PAGE_CACHE_SIZE;
1176                 if (index == end_index) {
1177                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1178                         if (nr <= offset) {
1179                                 page_cache_release(page);
1180                                 goto out;
1181                         }
1182                 }
1183                 nr = nr - offset;
1184
1185                 /* If users can be writing to this page using arbitrary
1186                  * virtual addresses, take care about potential aliasing
1187                  * before reading the page on the kernel side.
1188                  */
1189                 if (mapping_writably_mapped(mapping))
1190                         flush_dcache_page(page);
1191
1192                 /*
1193                  * When a sequential read accesses a page several times,
1194                  * only mark it as accessed the first time.
1195                  */
1196                 if (prev_index != index || offset != prev_offset)
1197                         mark_page_accessed(page);
1198                 prev_index = index;
1199
1200                 /*
1201                  * Ok, we have the page, and it's up-to-date, so
1202                  * now we can copy it to user space...
1203                  *
1204                  * The actor routine returns how many bytes were actually used..
1205                  * NOTE! This may not be the same as how much of a user buffer
1206                  * we filled up (we may be padding etc), so we can only update
1207                  * "pos" here (the actor routine has to update the user buffer
1208                  * pointers and the remaining count).
1209                  */
1210                 ret = actor(desc, page, offset, nr);
1211                 offset += ret;
1212                 index += offset >> PAGE_CACHE_SHIFT;
1213                 offset &= ~PAGE_CACHE_MASK;
1214                 prev_offset = offset;
1215
1216                 page_cache_release(page);
1217                 if (ret == nr && desc->count)
1218                         continue;
1219                 goto out;
1220
1221 page_not_up_to_date:
1222                 /* Get exclusive access to the page ... */
1223                 error = lock_page_killable(page);
1224                 if (unlikely(error))
1225                         goto readpage_error;
1226
1227 page_not_up_to_date_locked:
1228                 /* Did it get truncated before we got the lock? */
1229                 if (!page->mapping) {
1230                         unlock_page(page);
1231                         page_cache_release(page);
1232                         continue;
1233                 }
1234
1235                 /* Did somebody else fill it already? */
1236                 if (PageUptodate(page)) {
1237                         unlock_page(page);
1238                         goto page_ok;
1239                 }
1240
1241 readpage:
1242                 /*
1243                  * A previous I/O error may have been due to temporary
1244                  * failures, eg. multipath errors.
1245                  * PG_error will be set again if readpage fails.
1246                  */
1247                 ClearPageError(page);
1248                 /* Start the actual read. The read will unlock the page. */
1249                 error = mapping->a_ops->readpage(filp, page);
1250
1251                 if (unlikely(error)) {
1252                         if (error == AOP_TRUNCATED_PAGE) {
1253                                 page_cache_release(page);
1254                                 goto find_page;
1255                         }
1256                         goto readpage_error;
1257                 }
1258
1259                 if (!PageUptodate(page)) {
1260                         error = lock_page_killable(page);
1261                         if (unlikely(error))
1262                                 goto readpage_error;
1263                         if (!PageUptodate(page)) {
1264                                 if (page->mapping == NULL) {
1265                                         /*
1266                                          * invalidate_mapping_pages got it
1267                                          */
1268                                         unlock_page(page);
1269                                         page_cache_release(page);
1270                                         goto find_page;
1271                                 }
1272                                 unlock_page(page);
1273                                 shrink_readahead_size_eio(filp, ra);
1274                                 error = -EIO;
1275                                 goto readpage_error;
1276                         }
1277                         unlock_page(page);
1278                 }
1279
1280                 goto page_ok;
1281
1282 readpage_error:
1283                 /* UHHUH! A synchronous read error occurred. Report it */
1284                 desc->error = error;
1285                 page_cache_release(page);
1286                 goto out;
1287
1288 no_cached_page:
1289                 /*
1290                  * Ok, it wasn't cached, so we need to create a new
1291                  * page..
1292                  */
1293                 page = page_cache_alloc_cold(mapping);
1294                 if (!page) {
1295                         desc->error = -ENOMEM;
1296                         goto out;
1297                 }
1298                 error = add_to_page_cache_lru(page, mapping,
1299                                                 index, GFP_KERNEL);
1300                 if (error) {
1301                         page_cache_release(page);
1302                         if (error == -EEXIST)
1303                                 goto find_page;
1304                         desc->error = error;
1305                         goto out;
1306                 }
1307                 goto readpage;
1308         }
1309
1310 out:
1311         ra->prev_pos = prev_index;
1312         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1313         ra->prev_pos |= prev_offset;
1314
1315         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1316         file_accessed(filp);
1317 }
1318
1319 int file_read_actor(read_descriptor_t *desc, struct page *page,
1320                         unsigned long offset, unsigned long size)
1321 {
1322         char *kaddr;
1323         unsigned long left, count = desc->count;
1324
1325         if (size > count)
1326                 size = count;
1327
1328         /*
1329          * Faults on the destination of a read are common, so do it before
1330          * taking the kmap.
1331          */
1332         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1333                 kaddr = kmap_atomic(page, KM_USER0);
1334                 left = __copy_to_user_inatomic(desc->arg.buf,
1335                                                 kaddr + offset, size);
1336                 kunmap_atomic(kaddr, KM_USER0);
1337                 if (left == 0)
1338                         goto success;
1339         }
1340
1341         /* Do it the slow way */
1342         kaddr = kmap(page);
1343         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1344         kunmap(page);
1345
1346         if (left) {
1347                 size -= left;
1348                 desc->error = -EFAULT;
1349         }
1350 success:
1351         desc->count = count - size;
1352         desc->written += size;
1353         desc->arg.buf += size;
1354         return size;
1355 }
1356
1357 /*
1358  * Performs necessary checks before doing a write
1359  * @iov:        io vector request
1360  * @nr_segs:    number of segments in the iovec
1361  * @count:      number of bytes to write
1362  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1363  *
1364  * Adjust number of segments and amount of bytes to write (nr_segs should be
1365  * properly initialized first). Returns appropriate error code that caller
1366  * should return or zero in case that write should be allowed.
1367  */
1368 int generic_segment_checks(const struct iovec *iov,
1369                         unsigned long *nr_segs, size_t *count, int access_flags)
1370 {
1371         unsigned long   seg;
1372         size_t cnt = 0;
1373         for (seg = 0; seg < *nr_segs; seg++) {
1374                 const struct iovec *iv = &iov[seg];
1375
1376                 /*
1377                  * If any segment has a negative length, or the cumulative
1378                  * length ever wraps negative then return -EINVAL.
1379                  */
1380                 cnt += iv->iov_len;
1381                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1382                         return -EINVAL;
1383                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1384                         continue;
1385                 if (seg == 0)
1386                         return -EFAULT;
1387                 *nr_segs = seg;
1388                 cnt -= iv->iov_len;     /* This segment is no good */
1389                 break;
1390         }
1391         *count = cnt;
1392         return 0;
1393 }
1394 EXPORT_SYMBOL(generic_segment_checks);
1395
1396 /**
1397  * generic_file_aio_read - generic filesystem read routine
1398  * @iocb:       kernel I/O control block
1399  * @iov:        io vector request
1400  * @nr_segs:    number of segments in the iovec
1401  * @pos:        current file position
1402  *
1403  * This is the "read()" routine for all filesystems
1404  * that can use the page cache directly.
1405  */
1406 ssize_t
1407 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1408                 unsigned long nr_segs, loff_t pos)
1409 {
1410         struct file *filp = iocb->ki_filp;
1411         ssize_t retval;
1412         unsigned long seg = 0;
1413         size_t count;
1414         loff_t *ppos = &iocb->ki_pos;
1415         struct blk_plug plug;
1416
1417         count = 0;
1418         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1419         if (retval)
1420                 return retval;
1421
1422         blk_start_plug(&plug);
1423
1424         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1425         if (filp->f_flags & O_DIRECT) {
1426                 loff_t size;
1427                 struct address_space *mapping;
1428                 struct inode *inode;
1429
1430                 mapping = filp->f_mapping;
1431                 inode = mapping->host;
1432                 if (!count)
1433                         goto out; /* skip atime */
1434                 size = i_size_read(inode);
1435                 if (pos < size) {
1436                         retval = filemap_write_and_wait_range(mapping, pos,
1437                                         pos + iov_length(iov, nr_segs) - 1);
1438                         if (!retval) {
1439                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1440                                                         iov, pos, nr_segs);
1441                         }
1442                         if (retval > 0) {
1443                                 *ppos = pos + retval;
1444                                 count -= retval;
1445                         }
1446
1447                         /*
1448                          * Btrfs can have a short DIO read if we encounter
1449                          * compressed extents, so if there was an error, or if
1450                          * we've already read everything we wanted to, or if
1451                          * there was a short read because we hit EOF, go ahead
1452                          * and return.  Otherwise fallthrough to buffered io for
1453                          * the rest of the read.
1454                          */
1455                         if (retval < 0 || !count || *ppos >= size) {
1456                                 file_accessed(filp);
1457                                 goto out;
1458                         }
1459                 }
1460         }
1461
1462         count = retval;
1463         for (seg = 0; seg < nr_segs; seg++) {
1464                 read_descriptor_t desc;
1465                 loff_t offset = 0;
1466
1467                 /*
1468                  * If we did a short DIO read we need to skip the section of the
1469                  * iov that we've already read data into.
1470                  */
1471                 if (count) {
1472                         if (count > iov[seg].iov_len) {
1473                                 count -= iov[seg].iov_len;
1474                                 continue;
1475                         }
1476                         offset = count;
1477                         count = 0;
1478                 }
1479
1480                 desc.written = 0;
1481                 desc.arg.buf = iov[seg].iov_base + offset;
1482                 desc.count = iov[seg].iov_len - offset;
1483                 if (desc.count == 0)
1484                         continue;
1485                 desc.error = 0;
1486                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1487                 retval += desc.written;
1488                 if (desc.error) {
1489                         retval = retval ?: desc.error;
1490                         break;
1491                 }
1492                 if (desc.count > 0)
1493                         break;
1494         }
1495 out:
1496         blk_finish_plug(&plug);
1497         return retval;
1498 }
1499 EXPORT_SYMBOL(generic_file_aio_read);
1500
1501 static ssize_t
1502 do_readahead(struct address_space *mapping, struct file *filp,
1503              pgoff_t index, unsigned long nr)
1504 {
1505         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1506                 return -EINVAL;
1507
1508         force_page_cache_readahead(mapping, filp, index, nr);
1509         return 0;
1510 }
1511
1512 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1513 {
1514         ssize_t ret;
1515         struct file *file;
1516
1517         ret = -EBADF;
1518         file = fget(fd);
1519         if (file) {
1520                 if (file->f_mode & FMODE_READ) {
1521                         struct address_space *mapping = file->f_mapping;
1522                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1523                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1524                         unsigned long len = end - start + 1;
1525                         ret = do_readahead(mapping, file, start, len);
1526                 }
1527                 fput(file);
1528         }
1529         return ret;
1530 }
1531 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1532 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1533 {
1534         return SYSC_readahead((int) fd, offset, (size_t) count);
1535 }
1536 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1537 #endif
1538
1539 #ifdef CONFIG_MMU
1540 /**
1541  * page_cache_read - adds requested page to the page cache if not already there
1542  * @file:       file to read
1543  * @offset:     page index
1544  *
1545  * This adds the requested page to the page cache if it isn't already there,
1546  * and schedules an I/O to read in its contents from disk.
1547  */
1548 static int page_cache_read(struct file *file, pgoff_t offset)
1549 {
1550         struct address_space *mapping = file->f_mapping;
1551         struct page *page; 
1552         int ret;
1553
1554         do {
1555                 page = page_cache_alloc_cold(mapping);
1556                 if (!page)
1557                         return -ENOMEM;
1558
1559                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1560                 if (ret == 0)
1561                         ret = mapping->a_ops->readpage(file, page);
1562                 else if (ret == -EEXIST)
1563                         ret = 0; /* losing race to add is OK */
1564
1565                 page_cache_release(page);
1566
1567         } while (ret == AOP_TRUNCATED_PAGE);
1568                 
1569         return ret;
1570 }
1571
1572 #define MMAP_LOTSAMISS  (100)
1573
1574 /*
1575  * Synchronous readahead happens when we don't even find
1576  * a page in the page cache at all.
1577  */
1578 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1579                                    struct file_ra_state *ra,
1580                                    struct file *file,
1581                                    pgoff_t offset)
1582 {
1583         unsigned long ra_pages;
1584         struct address_space *mapping = file->f_mapping;
1585
1586         /* If we don't want any read-ahead, don't bother */
1587         if (VM_RandomReadHint(vma))
1588                 return;
1589         if (!ra->ra_pages)
1590                 return;
1591
1592         if (VM_SequentialReadHint(vma)) {
1593                 page_cache_sync_readahead(mapping, ra, file, offset,
1594                                           ra->ra_pages);
1595                 return;
1596         }
1597
1598         /* Avoid banging the cache line if not needed */
1599         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1600                 ra->mmap_miss++;
1601
1602         /*
1603          * Do we miss much more than hit in this file? If so,
1604          * stop bothering with read-ahead. It will only hurt.
1605          */
1606         if (ra->mmap_miss > MMAP_LOTSAMISS)
1607                 return;
1608
1609         /*
1610          * mmap read-around
1611          */
1612         ra_pages = max_sane_readahead(ra->ra_pages);
1613         ra->start = max_t(long, 0, offset - ra_pages / 2);
1614         ra->size = ra_pages;
1615         ra->async_size = ra_pages / 4;
1616         ra_submit(ra, mapping, file);
1617 }
1618
1619 /*
1620  * Asynchronous readahead happens when we find the page and PG_readahead,
1621  * so we want to possibly extend the readahead further..
1622  */
1623 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1624                                     struct file_ra_state *ra,
1625                                     struct file *file,
1626                                     struct page *page,
1627                                     pgoff_t offset)
1628 {
1629         struct address_space *mapping = file->f_mapping;
1630
1631         /* If we don't want any read-ahead, don't bother */
1632         if (VM_RandomReadHint(vma))
1633                 return;
1634         if (ra->mmap_miss > 0)
1635                 ra->mmap_miss--;
1636         if (PageReadahead(page))
1637                 page_cache_async_readahead(mapping, ra, file,
1638                                            page, offset, ra->ra_pages);
1639 }
1640
1641 /**
1642  * filemap_fault - read in file data for page fault handling
1643  * @vma:        vma in which the fault was taken
1644  * @vmf:        struct vm_fault containing details of the fault
1645  *
1646  * filemap_fault() is invoked via the vma operations vector for a
1647  * mapped memory region to read in file data during a page fault.
1648  *
1649  * The goto's are kind of ugly, but this streamlines the normal case of having
1650  * it in the page cache, and handles the special cases reasonably without
1651  * having a lot of duplicated code.
1652  */
1653 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1654 {
1655         int error;
1656         struct file *file = vma->vm_file;
1657         struct address_space *mapping = file->f_mapping;
1658         struct file_ra_state *ra = &file->f_ra;
1659         struct inode *inode = mapping->host;
1660         pgoff_t offset = vmf->pgoff;
1661         struct page *page;
1662         pgoff_t size;
1663         int ret = 0;
1664
1665         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1666         if (offset >= size)
1667                 return VM_FAULT_SIGBUS;
1668
1669         /*
1670          * Do we have something in the page cache already?
1671          */
1672         page = find_get_page(mapping, offset);
1673         if (likely(page)) {
1674                 /*
1675                  * We found the page, so try async readahead before
1676                  * waiting for the lock.
1677                  */
1678                 do_async_mmap_readahead(vma, ra, file, page, offset);
1679         } else {
1680                 /* No page in the page cache at all */
1681                 do_sync_mmap_readahead(vma, ra, file, offset);
1682                 count_vm_event(PGMAJFAULT);
1683                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1684                 ret = VM_FAULT_MAJOR;
1685 retry_find:
1686                 page = find_get_page(mapping, offset);
1687                 if (!page)
1688                         goto no_cached_page;
1689         }
1690
1691         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1692                 page_cache_release(page);
1693                 return ret | VM_FAULT_RETRY;
1694         }
1695
1696         /* Did it get truncated? */
1697         if (unlikely(page->mapping != mapping)) {
1698                 unlock_page(page);
1699                 put_page(page);
1700                 goto retry_find;
1701         }
1702         VM_BUG_ON(page->index != offset);
1703
1704         /*
1705          * We have a locked page in the page cache, now we need to check
1706          * that it's up-to-date. If not, it is going to be due to an error.
1707          */
1708         if (unlikely(!PageUptodate(page)))
1709                 goto page_not_uptodate;
1710
1711         /*
1712          * Found the page and have a reference on it.
1713          * We must recheck i_size under page lock.
1714          */
1715         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1716         if (unlikely(offset >= size)) {
1717                 unlock_page(page);
1718                 page_cache_release(page);
1719                 return VM_FAULT_SIGBUS;
1720         }
1721
1722         vmf->page = page;
1723         return ret | VM_FAULT_LOCKED;
1724
1725 no_cached_page:
1726         /*
1727          * We're only likely to ever get here if MADV_RANDOM is in
1728          * effect.
1729          */
1730         error = page_cache_read(file, offset);
1731
1732         /*
1733          * The page we want has now been added to the page cache.
1734          * In the unlikely event that someone removed it in the
1735          * meantime, we'll just come back here and read it again.
1736          */
1737         if (error >= 0)
1738                 goto retry_find;
1739
1740         /*
1741          * An error return from page_cache_read can result if the
1742          * system is low on memory, or a problem occurs while trying
1743          * to schedule I/O.
1744          */
1745         if (error == -ENOMEM)
1746                 return VM_FAULT_OOM;
1747         return VM_FAULT_SIGBUS;
1748
1749 page_not_uptodate:
1750         /*
1751          * Umm, take care of errors if the page isn't up-to-date.
1752          * Try to re-read it _once_. We do this synchronously,
1753          * because there really aren't any performance issues here
1754          * and we need to check for errors.
1755          */
1756         ClearPageError(page);
1757         error = mapping->a_ops->readpage(file, page);
1758         if (!error) {
1759                 wait_on_page_locked(page);
1760                 if (!PageUptodate(page))
1761                         error = -EIO;
1762         }
1763         page_cache_release(page);
1764
1765         if (!error || error == AOP_TRUNCATED_PAGE)
1766                 goto retry_find;
1767
1768         /* Things didn't work out. Return zero to tell the mm layer so. */
1769         shrink_readahead_size_eio(file, ra);
1770         return VM_FAULT_SIGBUS;
1771 }
1772 EXPORT_SYMBOL(filemap_fault);
1773
1774 const struct vm_operations_struct generic_file_vm_ops = {
1775         .fault          = filemap_fault,
1776 };
1777
1778 /* This is used for a general mmap of a disk file */
1779
1780 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1781 {
1782         struct address_space *mapping = file->f_mapping;
1783
1784         if (!mapping->a_ops->readpage)
1785                 return -ENOEXEC;
1786         file_accessed(file);
1787         vma->vm_ops = &generic_file_vm_ops;
1788         vma->vm_flags |= VM_CAN_NONLINEAR;
1789         return 0;
1790 }
1791
1792 /*
1793  * This is for filesystems which do not implement ->writepage.
1794  */
1795 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1796 {
1797         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1798                 return -EINVAL;
1799         return generic_file_mmap(file, vma);
1800 }
1801 #else
1802 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1803 {
1804         return -ENOSYS;
1805 }
1806 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1807 {
1808         return -ENOSYS;
1809 }
1810 #endif /* CONFIG_MMU */
1811
1812 EXPORT_SYMBOL(generic_file_mmap);
1813 EXPORT_SYMBOL(generic_file_readonly_mmap);
1814
1815 static struct page *__read_cache_page(struct address_space *mapping,
1816                                 pgoff_t index,
1817                                 int (*filler)(void *, struct page *),
1818                                 void *data,
1819                                 gfp_t gfp)
1820 {
1821         struct page *page;
1822         int err;
1823 repeat:
1824         page = find_get_page(mapping, index);
1825         if (!page) {
1826                 page = __page_cache_alloc(gfp | __GFP_COLD);
1827                 if (!page)
1828                         return ERR_PTR(-ENOMEM);
1829                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1830                 if (unlikely(err)) {
1831                         page_cache_release(page);
1832                         if (err == -EEXIST)
1833                                 goto repeat;
1834                         /* Presumably ENOMEM for radix tree node */
1835                         return ERR_PTR(err);
1836                 }
1837                 err = filler(data, page);
1838                 if (err < 0) {
1839                         page_cache_release(page);
1840                         page = ERR_PTR(err);
1841                 }
1842         }
1843         return page;
1844 }
1845
1846 static struct page *do_read_cache_page(struct address_space *mapping,
1847                                 pgoff_t index,
1848                                 int (*filler)(void *, struct page *),
1849                                 void *data,
1850                                 gfp_t gfp)
1851
1852 {
1853         struct page *page;
1854         int err;
1855
1856 retry:
1857         page = __read_cache_page(mapping, index, filler, data, gfp);
1858         if (IS_ERR(page))
1859                 return page;
1860         if (PageUptodate(page))
1861                 goto out;
1862
1863         lock_page(page);
1864         if (!page->mapping) {
1865                 unlock_page(page);
1866                 page_cache_release(page);
1867                 goto retry;
1868         }
1869         if (PageUptodate(page)) {
1870                 unlock_page(page);
1871                 goto out;
1872         }
1873         err = filler(data, page);
1874         if (err < 0) {
1875                 page_cache_release(page);
1876                 return ERR_PTR(err);
1877         }
1878 out:
1879         mark_page_accessed(page);
1880         return page;
1881 }
1882
1883 /**
1884  * read_cache_page_async - read into page cache, fill it if needed
1885  * @mapping:    the page's address_space
1886  * @index:      the page index
1887  * @filler:     function to perform the read
1888  * @data:       first arg to filler(data, page) function, often left as NULL
1889  *
1890  * Same as read_cache_page, but don't wait for page to become unlocked
1891  * after submitting it to the filler.
1892  *
1893  * Read into the page cache. If a page already exists, and PageUptodate() is
1894  * not set, try to fill the page but don't wait for it to become unlocked.
1895  *
1896  * If the page does not get brought uptodate, return -EIO.
1897  */
1898 struct page *read_cache_page_async(struct address_space *mapping,
1899                                 pgoff_t index,
1900                                 int (*filler)(void *, struct page *),
1901                                 void *data)
1902 {
1903         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1904 }
1905 EXPORT_SYMBOL(read_cache_page_async);
1906
1907 static struct page *wait_on_page_read(struct page *page)
1908 {
1909         if (!IS_ERR(page)) {
1910                 wait_on_page_locked(page);
1911                 if (!PageUptodate(page)) {
1912                         page_cache_release(page);
1913                         page = ERR_PTR(-EIO);
1914                 }
1915         }
1916         return page;
1917 }
1918
1919 /**
1920  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1921  * @mapping:    the page's address_space
1922  * @index:      the page index
1923  * @gfp:        the page allocator flags to use if allocating
1924  *
1925  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1926  * any new page allocations done using the specified allocation flags. Note
1927  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1928  * expect to do this atomically or anything like that - but you can pass in
1929  * other page requirements.
1930  *
1931  * If the page does not get brought uptodate, return -EIO.
1932  */
1933 struct page *read_cache_page_gfp(struct address_space *mapping,
1934                                 pgoff_t index,
1935                                 gfp_t gfp)
1936 {
1937         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1938
1939         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1940 }
1941 EXPORT_SYMBOL(read_cache_page_gfp);
1942
1943 /**
1944  * read_cache_page - read into page cache, fill it if needed
1945  * @mapping:    the page's address_space
1946  * @index:      the page index
1947  * @filler:     function to perform the read
1948  * @data:       first arg to filler(data, page) function, often left as NULL
1949  *
1950  * Read into the page cache. If a page already exists, and PageUptodate() is
1951  * not set, try to fill the page then wait for it to become unlocked.
1952  *
1953  * If the page does not get brought uptodate, return -EIO.
1954  */
1955 struct page *read_cache_page(struct address_space *mapping,
1956                                 pgoff_t index,
1957                                 int (*filler)(void *, struct page *),
1958                                 void *data)
1959 {
1960         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1961 }
1962 EXPORT_SYMBOL(read_cache_page);
1963
1964 /*
1965  * The logic we want is
1966  *
1967  *      if suid or (sgid and xgrp)
1968  *              remove privs
1969  */
1970 int should_remove_suid(struct dentry *dentry)
1971 {
1972         mode_t mode = dentry->d_inode->i_mode;
1973         int kill = 0;
1974
1975         /* suid always must be killed */
1976         if (unlikely(mode & S_ISUID))
1977                 kill = ATTR_KILL_SUID;
1978
1979         /*
1980          * sgid without any exec bits is just a mandatory locking mark; leave
1981          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1982          */
1983         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1984                 kill |= ATTR_KILL_SGID;
1985
1986         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1987                 return kill;
1988
1989         return 0;
1990 }
1991 EXPORT_SYMBOL(should_remove_suid);
1992
1993 static int __remove_suid(struct dentry *dentry, int kill)
1994 {
1995         struct iattr newattrs;
1996
1997         newattrs.ia_valid = ATTR_FORCE | kill;
1998         return notify_change(dentry, &newattrs);
1999 }
2000
2001 int file_remove_suid(struct file *file)
2002 {
2003         struct dentry *dentry = file->f_path.dentry;
2004         struct inode *inode = dentry->d_inode;
2005         int killsuid;
2006         int killpriv;
2007         int error = 0;
2008
2009         /* Fast path for nothing security related */
2010         if (IS_NOSEC(inode))
2011                 return 0;
2012
2013         killsuid = should_remove_suid(dentry);
2014         killpriv = security_inode_need_killpriv(dentry);
2015
2016         if (killpriv < 0)
2017                 return killpriv;
2018         if (killpriv)
2019                 error = security_inode_killpriv(dentry);
2020         if (!error && killsuid)
2021                 error = __remove_suid(dentry, killsuid);
2022         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2023                 inode->i_flags |= S_NOSEC;
2024
2025         return error;
2026 }
2027 EXPORT_SYMBOL(file_remove_suid);
2028
2029 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2030                         const struct iovec *iov, size_t base, size_t bytes)
2031 {
2032         size_t copied = 0, left = 0;
2033
2034         while (bytes) {
2035                 char __user *buf = iov->iov_base + base;
2036                 int copy = min(bytes, iov->iov_len - base);
2037
2038                 base = 0;
2039                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2040                 copied += copy;
2041                 bytes -= copy;
2042                 vaddr += copy;
2043                 iov++;
2044
2045                 if (unlikely(left))
2046                         break;
2047         }
2048         return copied - left;
2049 }
2050
2051 /*
2052  * Copy as much as we can into the page and return the number of bytes which
2053  * were successfully copied.  If a fault is encountered then return the number of
2054  * bytes which were copied.
2055  */
2056 size_t iov_iter_copy_from_user_atomic(struct page *page,
2057                 struct iov_iter *i, unsigned long offset, size_t bytes)
2058 {
2059         char *kaddr;
2060         size_t copied;
2061
2062         BUG_ON(!in_atomic());
2063         kaddr = kmap_atomic(page, KM_USER0);
2064         if (likely(i->nr_segs == 1)) {
2065                 int left;
2066                 char __user *buf = i->iov->iov_base + i->iov_offset;
2067                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2068                 copied = bytes - left;
2069         } else {
2070                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2071                                                 i->iov, i->iov_offset, bytes);
2072         }
2073         kunmap_atomic(kaddr, KM_USER0);
2074
2075         return copied;
2076 }
2077 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2078
2079 /*
2080  * This has the same sideeffects and return value as
2081  * iov_iter_copy_from_user_atomic().
2082  * The difference is that it attempts to resolve faults.
2083  * Page must not be locked.
2084  */
2085 size_t iov_iter_copy_from_user(struct page *page,
2086                 struct iov_iter *i, unsigned long offset, size_t bytes)
2087 {
2088         char *kaddr;
2089         size_t copied;
2090
2091         kaddr = kmap(page);
2092         if (likely(i->nr_segs == 1)) {
2093                 int left;
2094                 char __user *buf = i->iov->iov_base + i->iov_offset;
2095                 left = __copy_from_user(kaddr + offset, buf, bytes);
2096                 copied = bytes - left;
2097         } else {
2098                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2099                                                 i->iov, i->iov_offset, bytes);
2100         }
2101         kunmap(page);
2102         return copied;
2103 }
2104 EXPORT_SYMBOL(iov_iter_copy_from_user);
2105
2106 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2107 {
2108         BUG_ON(i->count < bytes);
2109
2110         if (likely(i->nr_segs == 1)) {
2111                 i->iov_offset += bytes;
2112                 i->count -= bytes;
2113         } else {
2114                 const struct iovec *iov = i->iov;
2115                 size_t base = i->iov_offset;
2116
2117                 /*
2118                  * The !iov->iov_len check ensures we skip over unlikely
2119                  * zero-length segments (without overruning the iovec).
2120                  */
2121                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2122                         int copy;
2123
2124                         copy = min(bytes, iov->iov_len - base);
2125                         BUG_ON(!i->count || i->count < copy);
2126                         i->count -= copy;
2127                         bytes -= copy;
2128                         base += copy;
2129                         if (iov->iov_len == base) {
2130                                 iov++;
2131                                 base = 0;
2132                         }
2133                 }
2134                 i->iov = iov;
2135                 i->iov_offset = base;
2136         }
2137 }
2138 EXPORT_SYMBOL(iov_iter_advance);
2139
2140 /*
2141  * Fault in the first iovec of the given iov_iter, to a maximum length
2142  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2143  * accessed (ie. because it is an invalid address).
2144  *
2145  * writev-intensive code may want this to prefault several iovecs -- that
2146  * would be possible (callers must not rely on the fact that _only_ the
2147  * first iovec will be faulted with the current implementation).
2148  */
2149 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2150 {
2151         char __user *buf = i->iov->iov_base + i->iov_offset;
2152         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2153         return fault_in_pages_readable(buf, bytes);
2154 }
2155 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2156
2157 /*
2158  * Return the count of just the current iov_iter segment.
2159  */
2160 size_t iov_iter_single_seg_count(struct iov_iter *i)
2161 {
2162         const struct iovec *iov = i->iov;
2163         if (i->nr_segs == 1)
2164                 return i->count;
2165         else
2166                 return min(i->count, iov->iov_len - i->iov_offset);
2167 }
2168 EXPORT_SYMBOL(iov_iter_single_seg_count);
2169
2170 /*
2171  * Performs necessary checks before doing a write
2172  *
2173  * Can adjust writing position or amount of bytes to write.
2174  * Returns appropriate error code that caller should return or
2175  * zero in case that write should be allowed.
2176  */
2177 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2178 {
2179         struct inode *inode = file->f_mapping->host;
2180         unsigned long limit = rlimit(RLIMIT_FSIZE);
2181
2182         if (unlikely(*pos < 0))
2183                 return -EINVAL;
2184
2185         if (!isblk) {
2186                 /* FIXME: this is for backwards compatibility with 2.4 */
2187                 if (file->f_flags & O_APPEND)
2188                         *pos = i_size_read(inode);
2189
2190                 if (limit != RLIM_INFINITY) {
2191                         if (*pos >= limit) {
2192                                 send_sig(SIGXFSZ, current, 0);
2193                                 return -EFBIG;
2194                         }
2195                         if (*count > limit - (typeof(limit))*pos) {
2196                                 *count = limit - (typeof(limit))*pos;
2197                         }
2198                 }
2199         }
2200
2201         /*
2202          * LFS rule
2203          */
2204         if (unlikely(*pos + *count > MAX_NON_LFS &&
2205                                 !(file->f_flags & O_LARGEFILE))) {
2206                 if (*pos >= MAX_NON_LFS) {
2207                         return -EFBIG;
2208                 }
2209                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2210                         *count = MAX_NON_LFS - (unsigned long)*pos;
2211                 }
2212         }
2213
2214         /*
2215          * Are we about to exceed the fs block limit ?
2216          *
2217          * If we have written data it becomes a short write.  If we have
2218          * exceeded without writing data we send a signal and return EFBIG.
2219          * Linus frestrict idea will clean these up nicely..
2220          */
2221         if (likely(!isblk)) {
2222                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2223                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2224                                 return -EFBIG;
2225                         }
2226                         /* zero-length writes at ->s_maxbytes are OK */
2227                 }
2228
2229                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2230                         *count = inode->i_sb->s_maxbytes - *pos;
2231         } else {
2232 #ifdef CONFIG_BLOCK
2233                 loff_t isize;
2234                 if (bdev_read_only(I_BDEV(inode)))
2235                         return -EPERM;
2236                 isize = i_size_read(inode);
2237                 if (*pos >= isize) {
2238                         if (*count || *pos > isize)
2239                                 return -ENOSPC;
2240                 }
2241
2242                 if (*pos + *count > isize)
2243                         *count = isize - *pos;
2244 #else
2245                 return -EPERM;
2246 #endif
2247         }
2248         return 0;
2249 }
2250 EXPORT_SYMBOL(generic_write_checks);
2251
2252 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2253                                 loff_t pos, unsigned len, unsigned flags,
2254                                 struct page **pagep, void **fsdata)
2255 {
2256         const struct address_space_operations *aops = mapping->a_ops;
2257
2258         return aops->write_begin(file, mapping, pos, len, flags,
2259                                                         pagep, fsdata);
2260 }
2261 EXPORT_SYMBOL(pagecache_write_begin);
2262
2263 int pagecache_write_end(struct file *file, struct address_space *mapping,
2264                                 loff_t pos, unsigned len, unsigned copied,
2265                                 struct page *page, void *fsdata)
2266 {
2267         const struct address_space_operations *aops = mapping->a_ops;
2268
2269         mark_page_accessed(page);
2270         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2271 }
2272 EXPORT_SYMBOL(pagecache_write_end);
2273
2274 ssize_t
2275 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2276                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2277                 size_t count, size_t ocount)
2278 {
2279         struct file     *file = iocb->ki_filp;
2280         struct address_space *mapping = file->f_mapping;
2281         struct inode    *inode = mapping->host;
2282         ssize_t         written;
2283         size_t          write_len;
2284         pgoff_t         end;
2285
2286         if (count != ocount)
2287                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2288
2289         write_len = iov_length(iov, *nr_segs);
2290         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2291
2292         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2293         if (written)
2294                 goto out;
2295
2296         /*
2297          * After a write we want buffered reads to be sure to go to disk to get
2298          * the new data.  We invalidate clean cached page from the region we're
2299          * about to write.  We do this *before* the write so that we can return
2300          * without clobbering -EIOCBQUEUED from ->direct_IO().
2301          */
2302         if (mapping->nrpages) {
2303                 written = invalidate_inode_pages2_range(mapping,
2304                                         pos >> PAGE_CACHE_SHIFT, end);
2305                 /*
2306                  * If a page can not be invalidated, return 0 to fall back
2307                  * to buffered write.
2308                  */
2309                 if (written) {
2310                         if (written == -EBUSY)
2311                                 return 0;
2312                         goto out;
2313                 }
2314         }
2315
2316         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2317
2318         /*
2319          * Finally, try again to invalidate clean pages which might have been
2320          * cached by non-direct readahead, or faulted in by get_user_pages()
2321          * if the source of the write was an mmap'ed region of the file
2322          * we're writing.  Either one is a pretty crazy thing to do,
2323          * so we don't support it 100%.  If this invalidation
2324          * fails, tough, the write still worked...
2325          */
2326         if (mapping->nrpages) {
2327                 invalidate_inode_pages2_range(mapping,
2328                                               pos >> PAGE_CACHE_SHIFT, end);
2329         }
2330
2331         if (written > 0) {
2332                 pos += written;
2333                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2334                         i_size_write(inode, pos);
2335                         mark_inode_dirty(inode);
2336                 }
2337                 *ppos = pos;
2338         }
2339 out:
2340         return written;
2341 }
2342 EXPORT_SYMBOL(generic_file_direct_write);
2343
2344 /*
2345  * Find or create a page at the given pagecache position. Return the locked
2346  * page. This function is specifically for buffered writes.
2347  */
2348 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2349                                         pgoff_t index, unsigned flags)
2350 {
2351         int status;
2352         struct page *page;
2353         gfp_t gfp_notmask = 0;
2354         if (flags & AOP_FLAG_NOFS)
2355                 gfp_notmask = __GFP_FS;
2356 repeat:
2357         page = find_lock_page(mapping, index);
2358         if (page)
2359                 goto found;
2360
2361         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2362         if (!page)
2363                 return NULL;
2364         status = add_to_page_cache_lru(page, mapping, index,
2365                                                 GFP_KERNEL & ~gfp_notmask);
2366         if (unlikely(status)) {
2367                 page_cache_release(page);
2368                 if (status == -EEXIST)
2369                         goto repeat;
2370                 return NULL;
2371         }
2372 found:
2373         wait_on_page_writeback(page);
2374         return page;
2375 }
2376 EXPORT_SYMBOL(grab_cache_page_write_begin);
2377
2378 static ssize_t generic_perform_write(struct file *file,
2379                                 struct iov_iter *i, loff_t pos)
2380 {
2381         struct address_space *mapping = file->f_mapping;
2382         const struct address_space_operations *a_ops = mapping->a_ops;
2383         long status = 0;
2384         ssize_t written = 0;
2385         unsigned int flags = 0;
2386
2387         /*
2388          * Copies from kernel address space cannot fail (NFSD is a big user).
2389          */
2390         if (segment_eq(get_fs(), KERNEL_DS))
2391                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2392
2393         do {
2394                 struct page *page;
2395                 unsigned long offset;   /* Offset into pagecache page */
2396                 unsigned long bytes;    /* Bytes to write to page */
2397                 size_t copied;          /* Bytes copied from user */
2398                 void *fsdata;
2399
2400                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2401                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2402                                                 iov_iter_count(i));
2403
2404 again:
2405
2406                 /*
2407                  * Bring in the user page that we will copy from _first_.
2408                  * Otherwise there's a nasty deadlock on copying from the
2409                  * same page as we're writing to, without it being marked
2410                  * up-to-date.
2411                  *
2412                  * Not only is this an optimisation, but it is also required
2413                  * to check that the address is actually valid, when atomic
2414                  * usercopies are used, below.
2415                  */
2416                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2417                         status = -EFAULT;
2418                         break;
2419                 }
2420
2421                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2422                                                 &page, &fsdata);
2423                 if (unlikely(status))
2424                         break;
2425
2426                 if (mapping_writably_mapped(mapping))
2427                         flush_dcache_page(page);
2428
2429                 pagefault_disable();
2430                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2431                 pagefault_enable();
2432                 flush_dcache_page(page);
2433
2434                 mark_page_accessed(page);
2435                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2436                                                 page, fsdata);
2437                 if (unlikely(status < 0))
2438                         break;
2439                 copied = status;
2440
2441                 cond_resched();
2442
2443                 iov_iter_advance(i, copied);
2444                 if (unlikely(copied == 0)) {
2445                         /*
2446                          * If we were unable to copy any data at all, we must
2447                          * fall back to a single segment length write.
2448                          *
2449                          * If we didn't fallback here, we could livelock
2450                          * because not all segments in the iov can be copied at
2451                          * once without a pagefault.
2452                          */
2453                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2454                                                 iov_iter_single_seg_count(i));
2455                         goto again;
2456                 }
2457                 pos += copied;
2458                 written += copied;
2459
2460                 balance_dirty_pages_ratelimited(mapping);
2461
2462         } while (iov_iter_count(i));
2463
2464         return written ? written : status;
2465 }
2466
2467 ssize_t
2468 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2469                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2470                 size_t count, ssize_t written)
2471 {
2472         struct file *file = iocb->ki_filp;
2473         ssize_t status;
2474         struct iov_iter i;
2475
2476         iov_iter_init(&i, iov, nr_segs, count, written);
2477         status = generic_perform_write(file, &i, pos);
2478
2479         if (likely(status >= 0)) {
2480                 written += status;
2481                 *ppos = pos + status;
2482         }
2483         
2484         return written ? written : status;
2485 }
2486 EXPORT_SYMBOL(generic_file_buffered_write);
2487
2488 /**
2489  * __generic_file_aio_write - write data to a file
2490  * @iocb:       IO state structure (file, offset, etc.)
2491  * @iov:        vector with data to write
2492  * @nr_segs:    number of segments in the vector
2493  * @ppos:       position where to write
2494  *
2495  * This function does all the work needed for actually writing data to a
2496  * file. It does all basic checks, removes SUID from the file, updates
2497  * modification times and calls proper subroutines depending on whether we
2498  * do direct IO or a standard buffered write.
2499  *
2500  * It expects i_mutex to be grabbed unless we work on a block device or similar
2501  * object which does not need locking at all.
2502  *
2503  * This function does *not* take care of syncing data in case of O_SYNC write.
2504  * A caller has to handle it. This is mainly due to the fact that we want to
2505  * avoid syncing under i_mutex.
2506  */
2507 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2508                                  unsigned long nr_segs, loff_t *ppos)
2509 {
2510         struct file *file = iocb->ki_filp;
2511         struct address_space * mapping = file->f_mapping;
2512         size_t ocount;          /* original count */
2513         size_t count;           /* after file limit checks */
2514         struct inode    *inode = mapping->host;
2515         loff_t          pos;
2516         ssize_t         written;
2517         ssize_t         err;
2518
2519         ocount = 0;
2520         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2521         if (err)
2522                 return err;
2523
2524         count = ocount;
2525         pos = *ppos;
2526
2527         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2528
2529         /* We can write back this queue in page reclaim */
2530         current->backing_dev_info = mapping->backing_dev_info;
2531         written = 0;
2532
2533         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2534         if (err)
2535                 goto out;
2536
2537         if (count == 0)
2538                 goto out;
2539
2540         err = file_remove_suid(file);
2541         if (err)
2542                 goto out;
2543
2544         file_update_time(file);
2545
2546         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2547         if (unlikely(file->f_flags & O_DIRECT)) {
2548                 loff_t endbyte;
2549                 ssize_t written_buffered;
2550
2551                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2552                                                         ppos, count, ocount);
2553                 if (written < 0 || written == count)
2554                         goto out;
2555                 /*
2556                  * direct-io write to a hole: fall through to buffered I/O
2557                  * for completing the rest of the request.
2558                  */
2559                 pos += written;
2560                 count -= written;
2561                 written_buffered = generic_file_buffered_write(iocb, iov,
2562                                                 nr_segs, pos, ppos, count,
2563                                                 written);
2564                 /*
2565                  * If generic_file_buffered_write() retuned a synchronous error
2566                  * then we want to return the number of bytes which were
2567                  * direct-written, or the error code if that was zero.  Note
2568                  * that this differs from normal direct-io semantics, which
2569                  * will return -EFOO even if some bytes were written.
2570                  */
2571                 if (written_buffered < 0) {
2572                         err = written_buffered;
2573                         goto out;
2574                 }
2575
2576                 /*
2577                  * We need to ensure that the page cache pages are written to
2578                  * disk and invalidated to preserve the expected O_DIRECT
2579                  * semantics.
2580                  */
2581                 endbyte = pos + written_buffered - written - 1;
2582                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2583                 if (err == 0) {
2584                         written = written_buffered;
2585                         invalidate_mapping_pages(mapping,
2586                                                  pos >> PAGE_CACHE_SHIFT,
2587                                                  endbyte >> PAGE_CACHE_SHIFT);
2588                 } else {
2589                         /*
2590                          * We don't know how much we wrote, so just return
2591                          * the number of bytes which were direct-written
2592                          */
2593                 }
2594         } else {
2595                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2596                                 pos, ppos, count, written);
2597         }
2598 out:
2599         current->backing_dev_info = NULL;
2600         return written ? written : err;
2601 }
2602 EXPORT_SYMBOL(__generic_file_aio_write);
2603
2604 /**
2605  * generic_file_aio_write - write data to a file
2606  * @iocb:       IO state structure
2607  * @iov:        vector with data to write
2608  * @nr_segs:    number of segments in the vector
2609  * @pos:        position in file where to write
2610  *
2611  * This is a wrapper around __generic_file_aio_write() to be used by most
2612  * filesystems. It takes care of syncing the file in case of O_SYNC file
2613  * and acquires i_mutex as needed.
2614  */
2615 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2616                 unsigned long nr_segs, loff_t pos)
2617 {
2618         struct file *file = iocb->ki_filp;
2619         struct inode *inode = file->f_mapping->host;
2620         struct blk_plug plug;
2621         ssize_t ret;
2622
2623         BUG_ON(iocb->ki_pos != pos);
2624
2625         mutex_lock(&inode->i_mutex);
2626         blk_start_plug(&plug);
2627         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2628         mutex_unlock(&inode->i_mutex);
2629
2630         if (ret > 0 || ret == -EIOCBQUEUED) {
2631                 ssize_t err;
2632
2633                 err = generic_write_sync(file, pos, ret);
2634                 if (err < 0 && ret > 0)
2635                         ret = err;
2636         }
2637         blk_finish_plug(&plug);
2638         return ret;
2639 }
2640 EXPORT_SYMBOL(generic_file_aio_write);
2641
2642 /**
2643  * try_to_release_page() - release old fs-specific metadata on a page
2644  *
2645  * @page: the page which the kernel is trying to free
2646  * @gfp_mask: memory allocation flags (and I/O mode)
2647  *
2648  * The address_space is to try to release any data against the page
2649  * (presumably at page->private).  If the release was successful, return `1'.
2650  * Otherwise return zero.
2651  *
2652  * This may also be called if PG_fscache is set on a page, indicating that the
2653  * page is known to the local caching routines.
2654  *
2655  * The @gfp_mask argument specifies whether I/O may be performed to release
2656  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2657  *
2658  */
2659 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2660 {
2661         struct address_space * const mapping = page->mapping;
2662
2663         BUG_ON(!PageLocked(page));
2664         if (PageWriteback(page))
2665                 return 0;
2666
2667         if (mapping && mapping->a_ops->releasepage)
2668                 return mapping->a_ops->releasepage(page, gfp_mask);
2669         return try_to_free_buffers(page);
2670 }
2671
2672 EXPORT_SYMBOL(try_to_release_page);