x86,mm: make pagefault killable
[pandora-kernel.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  inode_wb_list_lock
84  *    sb_lock                   (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    inode_wb_list_lock        (page_remove_rmap->set_page_dirty)
102  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
103  *    inode_wb_list_lock        (zap_pte_range->set_page_dirty)
104  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
105  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
106  *
107  *  (code doesn't rely on that order, so you could switch it around)
108  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
109  *    ->i_mmap_lock
110  */
111
112 /*
113  * Delete a page from the page cache and free it. Caller has to make
114  * sure the page is locked and that nobody else uses it - or that usage
115  * is safe.  The caller must hold the mapping's tree_lock.
116  */
117 void __delete_from_page_cache(struct page *page)
118 {
119         struct address_space *mapping = page->mapping;
120
121         radix_tree_delete(&mapping->page_tree, page->index);
122         page->mapping = NULL;
123         mapping->nrpages--;
124         __dec_zone_page_state(page, NR_FILE_PAGES);
125         if (PageSwapBacked(page))
126                 __dec_zone_page_state(page, NR_SHMEM);
127         BUG_ON(page_mapped(page));
128
129         /*
130          * Some filesystems seem to re-dirty the page even after
131          * the VM has canceled the dirty bit (eg ext3 journaling).
132          *
133          * Fix it up by doing a final dirty accounting check after
134          * having removed the page entirely.
135          */
136         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
137                 dec_zone_page_state(page, NR_FILE_DIRTY);
138                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
139         }
140 }
141
142 /**
143  * delete_from_page_cache - delete page from page cache
144  * @page: the page which the kernel is trying to remove from page cache
145  *
146  * This must be called only on pages that have been verified to be in the page
147  * cache and locked.  It will never put the page into the free list, the caller
148  * has a reference on the page.
149  */
150 void delete_from_page_cache(struct page *page)
151 {
152         struct address_space *mapping = page->mapping;
153         void (*freepage)(struct page *);
154
155         BUG_ON(!PageLocked(page));
156
157         freepage = mapping->a_ops->freepage;
158         spin_lock_irq(&mapping->tree_lock);
159         __delete_from_page_cache(page);
160         spin_unlock_irq(&mapping->tree_lock);
161         mem_cgroup_uncharge_cache_page(page);
162
163         if (freepage)
164                 freepage(page);
165         page_cache_release(page);
166 }
167 EXPORT_SYMBOL(delete_from_page_cache);
168
169 static int sleep_on_page(void *word)
170 {
171         io_schedule();
172         return 0;
173 }
174
175 static int sleep_on_page_killable(void *word)
176 {
177         sleep_on_page(word);
178         return fatal_signal_pending(current) ? -EINTR : 0;
179 }
180
181 /**
182  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
183  * @mapping:    address space structure to write
184  * @start:      offset in bytes where the range starts
185  * @end:        offset in bytes where the range ends (inclusive)
186  * @sync_mode:  enable synchronous operation
187  *
188  * Start writeback against all of a mapping's dirty pages that lie
189  * within the byte offsets <start, end> inclusive.
190  *
191  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
192  * opposed to a regular memory cleansing writeback.  The difference between
193  * these two operations is that if a dirty page/buffer is encountered, it must
194  * be waited upon, and not just skipped over.
195  */
196 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
197                                 loff_t end, int sync_mode)
198 {
199         int ret;
200         struct writeback_control wbc = {
201                 .sync_mode = sync_mode,
202                 .nr_to_write = LONG_MAX,
203                 .range_start = start,
204                 .range_end = end,
205         };
206
207         if (!mapping_cap_writeback_dirty(mapping))
208                 return 0;
209
210         ret = do_writepages(mapping, &wbc);
211         return ret;
212 }
213
214 static inline int __filemap_fdatawrite(struct address_space *mapping,
215         int sync_mode)
216 {
217         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
218 }
219
220 int filemap_fdatawrite(struct address_space *mapping)
221 {
222         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
223 }
224 EXPORT_SYMBOL(filemap_fdatawrite);
225
226 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
227                                 loff_t end)
228 {
229         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
230 }
231 EXPORT_SYMBOL(filemap_fdatawrite_range);
232
233 /**
234  * filemap_flush - mostly a non-blocking flush
235  * @mapping:    target address_space
236  *
237  * This is a mostly non-blocking flush.  Not suitable for data-integrity
238  * purposes - I/O may not be started against all dirty pages.
239  */
240 int filemap_flush(struct address_space *mapping)
241 {
242         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
243 }
244 EXPORT_SYMBOL(filemap_flush);
245
246 /**
247  * filemap_fdatawait_range - wait for writeback to complete
248  * @mapping:            address space structure to wait for
249  * @start_byte:         offset in bytes where the range starts
250  * @end_byte:           offset in bytes where the range ends (inclusive)
251  *
252  * Walk the list of under-writeback pages of the given address space
253  * in the given range and wait for all of them.
254  */
255 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
256                             loff_t end_byte)
257 {
258         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
259         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
260         struct pagevec pvec;
261         int nr_pages;
262         int ret = 0;
263
264         if (end_byte < start_byte)
265                 return 0;
266
267         pagevec_init(&pvec, 0);
268         while ((index <= end) &&
269                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
270                         PAGECACHE_TAG_WRITEBACK,
271                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
272                 unsigned i;
273
274                 for (i = 0; i < nr_pages; i++) {
275                         struct page *page = pvec.pages[i];
276
277                         /* until radix tree lookup accepts end_index */
278                         if (page->index > end)
279                                 continue;
280
281                         wait_on_page_writeback(page);
282                         if (TestClearPageError(page))
283                                 ret = -EIO;
284                 }
285                 pagevec_release(&pvec);
286                 cond_resched();
287         }
288
289         /* Check for outstanding write errors */
290         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
291                 ret = -ENOSPC;
292         if (test_and_clear_bit(AS_EIO, &mapping->flags))
293                 ret = -EIO;
294
295         return ret;
296 }
297 EXPORT_SYMBOL(filemap_fdatawait_range);
298
299 /**
300  * filemap_fdatawait - wait for all under-writeback pages to complete
301  * @mapping: address space structure to wait for
302  *
303  * Walk the list of under-writeback pages of the given address space
304  * and wait for all of them.
305  */
306 int filemap_fdatawait(struct address_space *mapping)
307 {
308         loff_t i_size = i_size_read(mapping->host);
309
310         if (i_size == 0)
311                 return 0;
312
313         return filemap_fdatawait_range(mapping, 0, i_size - 1);
314 }
315 EXPORT_SYMBOL(filemap_fdatawait);
316
317 int filemap_write_and_wait(struct address_space *mapping)
318 {
319         int err = 0;
320
321         if (mapping->nrpages) {
322                 err = filemap_fdatawrite(mapping);
323                 /*
324                  * Even if the above returned error, the pages may be
325                  * written partially (e.g. -ENOSPC), so we wait for it.
326                  * But the -EIO is special case, it may indicate the worst
327                  * thing (e.g. bug) happened, so we avoid waiting for it.
328                  */
329                 if (err != -EIO) {
330                         int err2 = filemap_fdatawait(mapping);
331                         if (!err)
332                                 err = err2;
333                 }
334         }
335         return err;
336 }
337 EXPORT_SYMBOL(filemap_write_and_wait);
338
339 /**
340  * filemap_write_and_wait_range - write out & wait on a file range
341  * @mapping:    the address_space for the pages
342  * @lstart:     offset in bytes where the range starts
343  * @lend:       offset in bytes where the range ends (inclusive)
344  *
345  * Write out and wait upon file offsets lstart->lend, inclusive.
346  *
347  * Note that `lend' is inclusive (describes the last byte to be written) so
348  * that this function can be used to write to the very end-of-file (end = -1).
349  */
350 int filemap_write_and_wait_range(struct address_space *mapping,
351                                  loff_t lstart, loff_t lend)
352 {
353         int err = 0;
354
355         if (mapping->nrpages) {
356                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
357                                                  WB_SYNC_ALL);
358                 /* See comment of filemap_write_and_wait() */
359                 if (err != -EIO) {
360                         int err2 = filemap_fdatawait_range(mapping,
361                                                 lstart, lend);
362                         if (!err)
363                                 err = err2;
364                 }
365         }
366         return err;
367 }
368 EXPORT_SYMBOL(filemap_write_and_wait_range);
369
370 /**
371  * replace_page_cache_page - replace a pagecache page with a new one
372  * @old:        page to be replaced
373  * @new:        page to replace with
374  * @gfp_mask:   allocation mode
375  *
376  * This function replaces a page in the pagecache with a new one.  On
377  * success it acquires the pagecache reference for the new page and
378  * drops it for the old page.  Both the old and new pages must be
379  * locked.  This function does not add the new page to the LRU, the
380  * caller must do that.
381  *
382  * The remove + add is atomic.  The only way this function can fail is
383  * memory allocation failure.
384  */
385 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
386 {
387         int error;
388         struct mem_cgroup *memcg = NULL;
389
390         VM_BUG_ON(!PageLocked(old));
391         VM_BUG_ON(!PageLocked(new));
392         VM_BUG_ON(new->mapping);
393
394         /*
395          * This is not page migration, but prepare_migration and
396          * end_migration does enough work for charge replacement.
397          *
398          * In the longer term we probably want a specialized function
399          * for moving the charge from old to new in a more efficient
400          * manner.
401          */
402         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
403         if (error)
404                 return error;
405
406         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
407         if (!error) {
408                 struct address_space *mapping = old->mapping;
409                 void (*freepage)(struct page *);
410
411                 pgoff_t offset = old->index;
412                 freepage = mapping->a_ops->freepage;
413
414                 page_cache_get(new);
415                 new->mapping = mapping;
416                 new->index = offset;
417
418                 spin_lock_irq(&mapping->tree_lock);
419                 __delete_from_page_cache(old);
420                 error = radix_tree_insert(&mapping->page_tree, offset, new);
421                 BUG_ON(error);
422                 mapping->nrpages++;
423                 __inc_zone_page_state(new, NR_FILE_PAGES);
424                 if (PageSwapBacked(new))
425                         __inc_zone_page_state(new, NR_SHMEM);
426                 spin_unlock_irq(&mapping->tree_lock);
427                 radix_tree_preload_end();
428                 if (freepage)
429                         freepage(old);
430                 page_cache_release(old);
431                 mem_cgroup_end_migration(memcg, old, new, true);
432         } else {
433                 mem_cgroup_end_migration(memcg, old, new, false);
434         }
435
436         return error;
437 }
438 EXPORT_SYMBOL_GPL(replace_page_cache_page);
439
440 /**
441  * add_to_page_cache_locked - add a locked page to the pagecache
442  * @page:       page to add
443  * @mapping:    the page's address_space
444  * @offset:     page index
445  * @gfp_mask:   page allocation mode
446  *
447  * This function is used to add a page to the pagecache. It must be locked.
448  * This function does not add the page to the LRU.  The caller must do that.
449  */
450 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
451                 pgoff_t offset, gfp_t gfp_mask)
452 {
453         int error;
454
455         VM_BUG_ON(!PageLocked(page));
456
457         error = mem_cgroup_cache_charge(page, current->mm,
458                                         gfp_mask & GFP_RECLAIM_MASK);
459         if (error)
460                 goto out;
461
462         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
463         if (error == 0) {
464                 page_cache_get(page);
465                 page->mapping = mapping;
466                 page->index = offset;
467
468                 spin_lock_irq(&mapping->tree_lock);
469                 error = radix_tree_insert(&mapping->page_tree, offset, page);
470                 if (likely(!error)) {
471                         mapping->nrpages++;
472                         __inc_zone_page_state(page, NR_FILE_PAGES);
473                         if (PageSwapBacked(page))
474                                 __inc_zone_page_state(page, NR_SHMEM);
475                         spin_unlock_irq(&mapping->tree_lock);
476                 } else {
477                         page->mapping = NULL;
478                         spin_unlock_irq(&mapping->tree_lock);
479                         mem_cgroup_uncharge_cache_page(page);
480                         page_cache_release(page);
481                 }
482                 radix_tree_preload_end();
483         } else
484                 mem_cgroup_uncharge_cache_page(page);
485 out:
486         return error;
487 }
488 EXPORT_SYMBOL(add_to_page_cache_locked);
489
490 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
491                                 pgoff_t offset, gfp_t gfp_mask)
492 {
493         int ret;
494
495         /*
496          * Splice_read and readahead add shmem/tmpfs pages into the page cache
497          * before shmem_readpage has a chance to mark them as SwapBacked: they
498          * need to go on the anon lru below, and mem_cgroup_cache_charge
499          * (called in add_to_page_cache) needs to know where they're going too.
500          */
501         if (mapping_cap_swap_backed(mapping))
502                 SetPageSwapBacked(page);
503
504         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
505         if (ret == 0) {
506                 if (page_is_file_cache(page))
507                         lru_cache_add_file(page);
508                 else
509                         lru_cache_add_anon(page);
510         }
511         return ret;
512 }
513 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
514
515 #ifdef CONFIG_NUMA
516 struct page *__page_cache_alloc(gfp_t gfp)
517 {
518         int n;
519         struct page *page;
520
521         if (cpuset_do_page_mem_spread()) {
522                 get_mems_allowed();
523                 n = cpuset_mem_spread_node();
524                 page = alloc_pages_exact_node(n, gfp, 0);
525                 put_mems_allowed();
526                 return page;
527         }
528         return alloc_pages(gfp, 0);
529 }
530 EXPORT_SYMBOL(__page_cache_alloc);
531 #endif
532
533 /*
534  * In order to wait for pages to become available there must be
535  * waitqueues associated with pages. By using a hash table of
536  * waitqueues where the bucket discipline is to maintain all
537  * waiters on the same queue and wake all when any of the pages
538  * become available, and for the woken contexts to check to be
539  * sure the appropriate page became available, this saves space
540  * at a cost of "thundering herd" phenomena during rare hash
541  * collisions.
542  */
543 static wait_queue_head_t *page_waitqueue(struct page *page)
544 {
545         const struct zone *zone = page_zone(page);
546
547         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
548 }
549
550 static inline void wake_up_page(struct page *page, int bit)
551 {
552         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
553 }
554
555 void wait_on_page_bit(struct page *page, int bit_nr)
556 {
557         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
558
559         if (test_bit(bit_nr, &page->flags))
560                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
561                                                         TASK_UNINTERRUPTIBLE);
562 }
563 EXPORT_SYMBOL(wait_on_page_bit);
564
565 int wait_on_page_bit_killable(struct page *page, int bit_nr)
566 {
567         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
568
569         if (!test_bit(bit_nr, &page->flags))
570                 return 0;
571
572         return __wait_on_bit(page_waitqueue(page), &wait,
573                              sleep_on_page_killable, TASK_KILLABLE);
574 }
575
576 /**
577  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
578  * @page: Page defining the wait queue of interest
579  * @waiter: Waiter to add to the queue
580  *
581  * Add an arbitrary @waiter to the wait queue for the nominated @page.
582  */
583 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
584 {
585         wait_queue_head_t *q = page_waitqueue(page);
586         unsigned long flags;
587
588         spin_lock_irqsave(&q->lock, flags);
589         __add_wait_queue(q, waiter);
590         spin_unlock_irqrestore(&q->lock, flags);
591 }
592 EXPORT_SYMBOL_GPL(add_page_wait_queue);
593
594 /**
595  * unlock_page - unlock a locked page
596  * @page: the page
597  *
598  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
599  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
600  * mechananism between PageLocked pages and PageWriteback pages is shared.
601  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
602  *
603  * The mb is necessary to enforce ordering between the clear_bit and the read
604  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
605  */
606 void unlock_page(struct page *page)
607 {
608         VM_BUG_ON(!PageLocked(page));
609         clear_bit_unlock(PG_locked, &page->flags);
610         smp_mb__after_clear_bit();
611         wake_up_page(page, PG_locked);
612 }
613 EXPORT_SYMBOL(unlock_page);
614
615 /**
616  * end_page_writeback - end writeback against a page
617  * @page: the page
618  */
619 void end_page_writeback(struct page *page)
620 {
621         if (TestClearPageReclaim(page))
622                 rotate_reclaimable_page(page);
623
624         if (!test_clear_page_writeback(page))
625                 BUG();
626
627         smp_mb__after_clear_bit();
628         wake_up_page(page, PG_writeback);
629 }
630 EXPORT_SYMBOL(end_page_writeback);
631
632 /**
633  * __lock_page - get a lock on the page, assuming we need to sleep to get it
634  * @page: the page to lock
635  */
636 void __lock_page(struct page *page)
637 {
638         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
639
640         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
641                                                         TASK_UNINTERRUPTIBLE);
642 }
643 EXPORT_SYMBOL(__lock_page);
644
645 int __lock_page_killable(struct page *page)
646 {
647         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
648
649         return __wait_on_bit_lock(page_waitqueue(page), &wait,
650                                         sleep_on_page_killable, TASK_KILLABLE);
651 }
652 EXPORT_SYMBOL_GPL(__lock_page_killable);
653
654 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
655                          unsigned int flags)
656 {
657         if (flags & FAULT_FLAG_ALLOW_RETRY) {
658                 /*
659                  * CAUTION! In this case, mmap_sem is not released
660                  * even though return 0.
661                  */
662                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
663                         return 0;
664
665                 up_read(&mm->mmap_sem);
666                 if (flags & FAULT_FLAG_KILLABLE)
667                         wait_on_page_locked_killable(page);
668                 else
669                         wait_on_page_locked(page);
670                 return 0;
671         } else {
672                 if (flags & FAULT_FLAG_KILLABLE) {
673                         int ret;
674
675                         ret = __lock_page_killable(page);
676                         if (ret) {
677                                 up_read(&mm->mmap_sem);
678                                 return 0;
679                         }
680                 } else
681                         __lock_page(page);
682                 return 1;
683         }
684 }
685
686 /**
687  * find_get_page - find and get a page reference
688  * @mapping: the address_space to search
689  * @offset: the page index
690  *
691  * Is there a pagecache struct page at the given (mapping, offset) tuple?
692  * If yes, increment its refcount and return it; if no, return NULL.
693  */
694 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
695 {
696         void **pagep;
697         struct page *page;
698
699         rcu_read_lock();
700 repeat:
701         page = NULL;
702         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
703         if (pagep) {
704                 page = radix_tree_deref_slot(pagep);
705                 if (unlikely(!page))
706                         goto out;
707                 if (radix_tree_deref_retry(page))
708                         goto repeat;
709
710                 if (!page_cache_get_speculative(page))
711                         goto repeat;
712
713                 /*
714                  * Has the page moved?
715                  * This is part of the lockless pagecache protocol. See
716                  * include/linux/pagemap.h for details.
717                  */
718                 if (unlikely(page != *pagep)) {
719                         page_cache_release(page);
720                         goto repeat;
721                 }
722         }
723 out:
724         rcu_read_unlock();
725
726         return page;
727 }
728 EXPORT_SYMBOL(find_get_page);
729
730 /**
731  * find_lock_page - locate, pin and lock a pagecache page
732  * @mapping: the address_space to search
733  * @offset: the page index
734  *
735  * Locates the desired pagecache page, locks it, increments its reference
736  * count and returns its address.
737  *
738  * Returns zero if the page was not present. find_lock_page() may sleep.
739  */
740 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
741 {
742         struct page *page;
743
744 repeat:
745         page = find_get_page(mapping, offset);
746         if (page) {
747                 lock_page(page);
748                 /* Has the page been truncated? */
749                 if (unlikely(page->mapping != mapping)) {
750                         unlock_page(page);
751                         page_cache_release(page);
752                         goto repeat;
753                 }
754                 VM_BUG_ON(page->index != offset);
755         }
756         return page;
757 }
758 EXPORT_SYMBOL(find_lock_page);
759
760 /**
761  * find_or_create_page - locate or add a pagecache page
762  * @mapping: the page's address_space
763  * @index: the page's index into the mapping
764  * @gfp_mask: page allocation mode
765  *
766  * Locates a page in the pagecache.  If the page is not present, a new page
767  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
768  * LRU list.  The returned page is locked and has its reference count
769  * incremented.
770  *
771  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
772  * allocation!
773  *
774  * find_or_create_page() returns the desired page's address, or zero on
775  * memory exhaustion.
776  */
777 struct page *find_or_create_page(struct address_space *mapping,
778                 pgoff_t index, gfp_t gfp_mask)
779 {
780         struct page *page;
781         int err;
782 repeat:
783         page = find_lock_page(mapping, index);
784         if (!page) {
785                 page = __page_cache_alloc(gfp_mask);
786                 if (!page)
787                         return NULL;
788                 /*
789                  * We want a regular kernel memory (not highmem or DMA etc)
790                  * allocation for the radix tree nodes, but we need to honour
791                  * the context-specific requirements the caller has asked for.
792                  * GFP_RECLAIM_MASK collects those requirements.
793                  */
794                 err = add_to_page_cache_lru(page, mapping, index,
795                         (gfp_mask & GFP_RECLAIM_MASK));
796                 if (unlikely(err)) {
797                         page_cache_release(page);
798                         page = NULL;
799                         if (err == -EEXIST)
800                                 goto repeat;
801                 }
802         }
803         return page;
804 }
805 EXPORT_SYMBOL(find_or_create_page);
806
807 /**
808  * find_get_pages - gang pagecache lookup
809  * @mapping:    The address_space to search
810  * @start:      The starting page index
811  * @nr_pages:   The maximum number of pages
812  * @pages:      Where the resulting pages are placed
813  *
814  * find_get_pages() will search for and return a group of up to
815  * @nr_pages pages in the mapping.  The pages are placed at @pages.
816  * find_get_pages() takes a reference against the returned pages.
817  *
818  * The search returns a group of mapping-contiguous pages with ascending
819  * indexes.  There may be holes in the indices due to not-present pages.
820  *
821  * find_get_pages() returns the number of pages which were found.
822  */
823 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
824                             unsigned int nr_pages, struct page **pages)
825 {
826         unsigned int i;
827         unsigned int ret;
828         unsigned int nr_found;
829
830         rcu_read_lock();
831 restart:
832         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
833                                 (void ***)pages, start, nr_pages);
834         ret = 0;
835         for (i = 0; i < nr_found; i++) {
836                 struct page *page;
837 repeat:
838                 page = radix_tree_deref_slot((void **)pages[i]);
839                 if (unlikely(!page))
840                         continue;
841
842                 /*
843                  * This can only trigger when the entry at index 0 moves out
844                  * of or back to the root: none yet gotten, safe to restart.
845                  */
846                 if (radix_tree_deref_retry(page)) {
847                         WARN_ON(start | i);
848                         goto restart;
849                 }
850
851                 if (!page_cache_get_speculative(page))
852                         goto repeat;
853
854                 /* Has the page moved? */
855                 if (unlikely(page != *((void **)pages[i]))) {
856                         page_cache_release(page);
857                         goto repeat;
858                 }
859
860                 pages[ret] = page;
861                 ret++;
862         }
863
864         /*
865          * If all entries were removed before we could secure them,
866          * try again, because callers stop trying once 0 is returned.
867          */
868         if (unlikely(!ret && nr_found))
869                 goto restart;
870         rcu_read_unlock();
871         return ret;
872 }
873
874 /**
875  * find_get_pages_contig - gang contiguous pagecache lookup
876  * @mapping:    The address_space to search
877  * @index:      The starting page index
878  * @nr_pages:   The maximum number of pages
879  * @pages:      Where the resulting pages are placed
880  *
881  * find_get_pages_contig() works exactly like find_get_pages(), except
882  * that the returned number of pages are guaranteed to be contiguous.
883  *
884  * find_get_pages_contig() returns the number of pages which were found.
885  */
886 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
887                                unsigned int nr_pages, struct page **pages)
888 {
889         unsigned int i;
890         unsigned int ret;
891         unsigned int nr_found;
892
893         rcu_read_lock();
894 restart:
895         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
896                                 (void ***)pages, index, nr_pages);
897         ret = 0;
898         for (i = 0; i < nr_found; i++) {
899                 struct page *page;
900 repeat:
901                 page = radix_tree_deref_slot((void **)pages[i]);
902                 if (unlikely(!page))
903                         continue;
904
905                 /*
906                  * This can only trigger when the entry at index 0 moves out
907                  * of or back to the root: none yet gotten, safe to restart.
908                  */
909                 if (radix_tree_deref_retry(page))
910                         goto restart;
911
912                 if (!page_cache_get_speculative(page))
913                         goto repeat;
914
915                 /* Has the page moved? */
916                 if (unlikely(page != *((void **)pages[i]))) {
917                         page_cache_release(page);
918                         goto repeat;
919                 }
920
921                 /*
922                  * must check mapping and index after taking the ref.
923                  * otherwise we can get both false positives and false
924                  * negatives, which is just confusing to the caller.
925                  */
926                 if (page->mapping == NULL || page->index != index) {
927                         page_cache_release(page);
928                         break;
929                 }
930
931                 pages[ret] = page;
932                 ret++;
933                 index++;
934         }
935         rcu_read_unlock();
936         return ret;
937 }
938 EXPORT_SYMBOL(find_get_pages_contig);
939
940 /**
941  * find_get_pages_tag - find and return pages that match @tag
942  * @mapping:    the address_space to search
943  * @index:      the starting page index
944  * @tag:        the tag index
945  * @nr_pages:   the maximum number of pages
946  * @pages:      where the resulting pages are placed
947  *
948  * Like find_get_pages, except we only return pages which are tagged with
949  * @tag.   We update @index to index the next page for the traversal.
950  */
951 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
952                         int tag, unsigned int nr_pages, struct page **pages)
953 {
954         unsigned int i;
955         unsigned int ret;
956         unsigned int nr_found;
957
958         rcu_read_lock();
959 restart:
960         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
961                                 (void ***)pages, *index, nr_pages, tag);
962         ret = 0;
963         for (i = 0; i < nr_found; i++) {
964                 struct page *page;
965 repeat:
966                 page = radix_tree_deref_slot((void **)pages[i]);
967                 if (unlikely(!page))
968                         continue;
969
970                 /*
971                  * This can only trigger when the entry at index 0 moves out
972                  * of or back to the root: none yet gotten, safe to restart.
973                  */
974                 if (radix_tree_deref_retry(page))
975                         goto restart;
976
977                 if (!page_cache_get_speculative(page))
978                         goto repeat;
979
980                 /* Has the page moved? */
981                 if (unlikely(page != *((void **)pages[i]))) {
982                         page_cache_release(page);
983                         goto repeat;
984                 }
985
986                 pages[ret] = page;
987                 ret++;
988         }
989
990         /*
991          * If all entries were removed before we could secure them,
992          * try again, because callers stop trying once 0 is returned.
993          */
994         if (unlikely(!ret && nr_found))
995                 goto restart;
996         rcu_read_unlock();
997
998         if (ret)
999                 *index = pages[ret - 1]->index + 1;
1000
1001         return ret;
1002 }
1003 EXPORT_SYMBOL(find_get_pages_tag);
1004
1005 /**
1006  * grab_cache_page_nowait - returns locked page at given index in given cache
1007  * @mapping: target address_space
1008  * @index: the page index
1009  *
1010  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1011  * This is intended for speculative data generators, where the data can
1012  * be regenerated if the page couldn't be grabbed.  This routine should
1013  * be safe to call while holding the lock for another page.
1014  *
1015  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1016  * and deadlock against the caller's locked page.
1017  */
1018 struct page *
1019 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1020 {
1021         struct page *page = find_get_page(mapping, index);
1022
1023         if (page) {
1024                 if (trylock_page(page))
1025                         return page;
1026                 page_cache_release(page);
1027                 return NULL;
1028         }
1029         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1030         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1031                 page_cache_release(page);
1032                 page = NULL;
1033         }
1034         return page;
1035 }
1036 EXPORT_SYMBOL(grab_cache_page_nowait);
1037
1038 /*
1039  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1040  * a _large_ part of the i/o request. Imagine the worst scenario:
1041  *
1042  *      ---R__________________________________________B__________
1043  *         ^ reading here                             ^ bad block(assume 4k)
1044  *
1045  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1046  * => failing the whole request => read(R) => read(R+1) =>
1047  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1048  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1049  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1050  *
1051  * It is going insane. Fix it by quickly scaling down the readahead size.
1052  */
1053 static void shrink_readahead_size_eio(struct file *filp,
1054                                         struct file_ra_state *ra)
1055 {
1056         ra->ra_pages /= 4;
1057 }
1058
1059 /**
1060  * do_generic_file_read - generic file read routine
1061  * @filp:       the file to read
1062  * @ppos:       current file position
1063  * @desc:       read_descriptor
1064  * @actor:      read method
1065  *
1066  * This is a generic file read routine, and uses the
1067  * mapping->a_ops->readpage() function for the actual low-level stuff.
1068  *
1069  * This is really ugly. But the goto's actually try to clarify some
1070  * of the logic when it comes to error handling etc.
1071  */
1072 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1073                 read_descriptor_t *desc, read_actor_t actor)
1074 {
1075         struct address_space *mapping = filp->f_mapping;
1076         struct inode *inode = mapping->host;
1077         struct file_ra_state *ra = &filp->f_ra;
1078         pgoff_t index;
1079         pgoff_t last_index;
1080         pgoff_t prev_index;
1081         unsigned long offset;      /* offset into pagecache page */
1082         unsigned int prev_offset;
1083         int error;
1084
1085         index = *ppos >> PAGE_CACHE_SHIFT;
1086         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1087         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1088         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1089         offset = *ppos & ~PAGE_CACHE_MASK;
1090
1091         for (;;) {
1092                 struct page *page;
1093                 pgoff_t end_index;
1094                 loff_t isize;
1095                 unsigned long nr, ret;
1096
1097                 cond_resched();
1098 find_page:
1099                 page = find_get_page(mapping, index);
1100                 if (!page) {
1101                         page_cache_sync_readahead(mapping,
1102                                         ra, filp,
1103                                         index, last_index - index);
1104                         page = find_get_page(mapping, index);
1105                         if (unlikely(page == NULL))
1106                                 goto no_cached_page;
1107                 }
1108                 if (PageReadahead(page)) {
1109                         page_cache_async_readahead(mapping,
1110                                         ra, filp, page,
1111                                         index, last_index - index);
1112                 }
1113                 if (!PageUptodate(page)) {
1114                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1115                                         !mapping->a_ops->is_partially_uptodate)
1116                                 goto page_not_up_to_date;
1117                         if (!trylock_page(page))
1118                                 goto page_not_up_to_date;
1119                         /* Did it get truncated before we got the lock? */
1120                         if (!page->mapping)
1121                                 goto page_not_up_to_date_locked;
1122                         if (!mapping->a_ops->is_partially_uptodate(page,
1123                                                                 desc, offset))
1124                                 goto page_not_up_to_date_locked;
1125                         unlock_page(page);
1126                 }
1127 page_ok:
1128                 /*
1129                  * i_size must be checked after we know the page is Uptodate.
1130                  *
1131                  * Checking i_size after the check allows us to calculate
1132                  * the correct value for "nr", which means the zero-filled
1133                  * part of the page is not copied back to userspace (unless
1134                  * another truncate extends the file - this is desired though).
1135                  */
1136
1137                 isize = i_size_read(inode);
1138                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1139                 if (unlikely(!isize || index > end_index)) {
1140                         page_cache_release(page);
1141                         goto out;
1142                 }
1143
1144                 /* nr is the maximum number of bytes to copy from this page */
1145                 nr = PAGE_CACHE_SIZE;
1146                 if (index == end_index) {
1147                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1148                         if (nr <= offset) {
1149                                 page_cache_release(page);
1150                                 goto out;
1151                         }
1152                 }
1153                 nr = nr - offset;
1154
1155                 /* If users can be writing to this page using arbitrary
1156                  * virtual addresses, take care about potential aliasing
1157                  * before reading the page on the kernel side.
1158                  */
1159                 if (mapping_writably_mapped(mapping))
1160                         flush_dcache_page(page);
1161
1162                 /*
1163                  * When a sequential read accesses a page several times,
1164                  * only mark it as accessed the first time.
1165                  */
1166                 if (prev_index != index || offset != prev_offset)
1167                         mark_page_accessed(page);
1168                 prev_index = index;
1169
1170                 /*
1171                  * Ok, we have the page, and it's up-to-date, so
1172                  * now we can copy it to user space...
1173                  *
1174                  * The actor routine returns how many bytes were actually used..
1175                  * NOTE! This may not be the same as how much of a user buffer
1176                  * we filled up (we may be padding etc), so we can only update
1177                  * "pos" here (the actor routine has to update the user buffer
1178                  * pointers and the remaining count).
1179                  */
1180                 ret = actor(desc, page, offset, nr);
1181                 offset += ret;
1182                 index += offset >> PAGE_CACHE_SHIFT;
1183                 offset &= ~PAGE_CACHE_MASK;
1184                 prev_offset = offset;
1185
1186                 page_cache_release(page);
1187                 if (ret == nr && desc->count)
1188                         continue;
1189                 goto out;
1190
1191 page_not_up_to_date:
1192                 /* Get exclusive access to the page ... */
1193                 error = lock_page_killable(page);
1194                 if (unlikely(error))
1195                         goto readpage_error;
1196
1197 page_not_up_to_date_locked:
1198                 /* Did it get truncated before we got the lock? */
1199                 if (!page->mapping) {
1200                         unlock_page(page);
1201                         page_cache_release(page);
1202                         continue;
1203                 }
1204
1205                 /* Did somebody else fill it already? */
1206                 if (PageUptodate(page)) {
1207                         unlock_page(page);
1208                         goto page_ok;
1209                 }
1210
1211 readpage:
1212                 /*
1213                  * A previous I/O error may have been due to temporary
1214                  * failures, eg. multipath errors.
1215                  * PG_error will be set again if readpage fails.
1216                  */
1217                 ClearPageError(page);
1218                 /* Start the actual read. The read will unlock the page. */
1219                 error = mapping->a_ops->readpage(filp, page);
1220
1221                 if (unlikely(error)) {
1222                         if (error == AOP_TRUNCATED_PAGE) {
1223                                 page_cache_release(page);
1224                                 goto find_page;
1225                         }
1226                         goto readpage_error;
1227                 }
1228
1229                 if (!PageUptodate(page)) {
1230                         error = lock_page_killable(page);
1231                         if (unlikely(error))
1232                                 goto readpage_error;
1233                         if (!PageUptodate(page)) {
1234                                 if (page->mapping == NULL) {
1235                                         /*
1236                                          * invalidate_mapping_pages got it
1237                                          */
1238                                         unlock_page(page);
1239                                         page_cache_release(page);
1240                                         goto find_page;
1241                                 }
1242                                 unlock_page(page);
1243                                 shrink_readahead_size_eio(filp, ra);
1244                                 error = -EIO;
1245                                 goto readpage_error;
1246                         }
1247                         unlock_page(page);
1248                 }
1249
1250                 goto page_ok;
1251
1252 readpage_error:
1253                 /* UHHUH! A synchronous read error occurred. Report it */
1254                 desc->error = error;
1255                 page_cache_release(page);
1256                 goto out;
1257
1258 no_cached_page:
1259                 /*
1260                  * Ok, it wasn't cached, so we need to create a new
1261                  * page..
1262                  */
1263                 page = page_cache_alloc_cold(mapping);
1264                 if (!page) {
1265                         desc->error = -ENOMEM;
1266                         goto out;
1267                 }
1268                 error = add_to_page_cache_lru(page, mapping,
1269                                                 index, GFP_KERNEL);
1270                 if (error) {
1271                         page_cache_release(page);
1272                         if (error == -EEXIST)
1273                                 goto find_page;
1274                         desc->error = error;
1275                         goto out;
1276                 }
1277                 goto readpage;
1278         }
1279
1280 out:
1281         ra->prev_pos = prev_index;
1282         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1283         ra->prev_pos |= prev_offset;
1284
1285         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1286         file_accessed(filp);
1287 }
1288
1289 int file_read_actor(read_descriptor_t *desc, struct page *page,
1290                         unsigned long offset, unsigned long size)
1291 {
1292         char *kaddr;
1293         unsigned long left, count = desc->count;
1294
1295         if (size > count)
1296                 size = count;
1297
1298         /*
1299          * Faults on the destination of a read are common, so do it before
1300          * taking the kmap.
1301          */
1302         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1303                 kaddr = kmap_atomic(page, KM_USER0);
1304                 left = __copy_to_user_inatomic(desc->arg.buf,
1305                                                 kaddr + offset, size);
1306                 kunmap_atomic(kaddr, KM_USER0);
1307                 if (left == 0)
1308                         goto success;
1309         }
1310
1311         /* Do it the slow way */
1312         kaddr = kmap(page);
1313         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1314         kunmap(page);
1315
1316         if (left) {
1317                 size -= left;
1318                 desc->error = -EFAULT;
1319         }
1320 success:
1321         desc->count = count - size;
1322         desc->written += size;
1323         desc->arg.buf += size;
1324         return size;
1325 }
1326
1327 /*
1328  * Performs necessary checks before doing a write
1329  * @iov:        io vector request
1330  * @nr_segs:    number of segments in the iovec
1331  * @count:      number of bytes to write
1332  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1333  *
1334  * Adjust number of segments and amount of bytes to write (nr_segs should be
1335  * properly initialized first). Returns appropriate error code that caller
1336  * should return or zero in case that write should be allowed.
1337  */
1338 int generic_segment_checks(const struct iovec *iov,
1339                         unsigned long *nr_segs, size_t *count, int access_flags)
1340 {
1341         unsigned long   seg;
1342         size_t cnt = 0;
1343         for (seg = 0; seg < *nr_segs; seg++) {
1344                 const struct iovec *iv = &iov[seg];
1345
1346                 /*
1347                  * If any segment has a negative length, or the cumulative
1348                  * length ever wraps negative then return -EINVAL.
1349                  */
1350                 cnt += iv->iov_len;
1351                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1352                         return -EINVAL;
1353                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1354                         continue;
1355                 if (seg == 0)
1356                         return -EFAULT;
1357                 *nr_segs = seg;
1358                 cnt -= iv->iov_len;     /* This segment is no good */
1359                 break;
1360         }
1361         *count = cnt;
1362         return 0;
1363 }
1364 EXPORT_SYMBOL(generic_segment_checks);
1365
1366 /**
1367  * generic_file_aio_read - generic filesystem read routine
1368  * @iocb:       kernel I/O control block
1369  * @iov:        io vector request
1370  * @nr_segs:    number of segments in the iovec
1371  * @pos:        current file position
1372  *
1373  * This is the "read()" routine for all filesystems
1374  * that can use the page cache directly.
1375  */
1376 ssize_t
1377 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1378                 unsigned long nr_segs, loff_t pos)
1379 {
1380         struct file *filp = iocb->ki_filp;
1381         ssize_t retval;
1382         unsigned long seg = 0;
1383         size_t count;
1384         loff_t *ppos = &iocb->ki_pos;
1385         struct blk_plug plug;
1386
1387         count = 0;
1388         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1389         if (retval)
1390                 return retval;
1391
1392         blk_start_plug(&plug);
1393
1394         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1395         if (filp->f_flags & O_DIRECT) {
1396                 loff_t size;
1397                 struct address_space *mapping;
1398                 struct inode *inode;
1399
1400                 mapping = filp->f_mapping;
1401                 inode = mapping->host;
1402                 if (!count)
1403                         goto out; /* skip atime */
1404                 size = i_size_read(inode);
1405                 if (pos < size) {
1406                         retval = filemap_write_and_wait_range(mapping, pos,
1407                                         pos + iov_length(iov, nr_segs) - 1);
1408                         if (!retval) {
1409                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1410                                                         iov, pos, nr_segs);
1411                         }
1412                         if (retval > 0) {
1413                                 *ppos = pos + retval;
1414                                 count -= retval;
1415                         }
1416
1417                         /*
1418                          * Btrfs can have a short DIO read if we encounter
1419                          * compressed extents, so if there was an error, or if
1420                          * we've already read everything we wanted to, or if
1421                          * there was a short read because we hit EOF, go ahead
1422                          * and return.  Otherwise fallthrough to buffered io for
1423                          * the rest of the read.
1424                          */
1425                         if (retval < 0 || !count || *ppos >= size) {
1426                                 file_accessed(filp);
1427                                 goto out;
1428                         }
1429                 }
1430         }
1431
1432         count = retval;
1433         for (seg = 0; seg < nr_segs; seg++) {
1434                 read_descriptor_t desc;
1435                 loff_t offset = 0;
1436
1437                 /*
1438                  * If we did a short DIO read we need to skip the section of the
1439                  * iov that we've already read data into.
1440                  */
1441                 if (count) {
1442                         if (count > iov[seg].iov_len) {
1443                                 count -= iov[seg].iov_len;
1444                                 continue;
1445                         }
1446                         offset = count;
1447                         count = 0;
1448                 }
1449
1450                 desc.written = 0;
1451                 desc.arg.buf = iov[seg].iov_base + offset;
1452                 desc.count = iov[seg].iov_len - offset;
1453                 if (desc.count == 0)
1454                         continue;
1455                 desc.error = 0;
1456                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1457                 retval += desc.written;
1458                 if (desc.error) {
1459                         retval = retval ?: desc.error;
1460                         break;
1461                 }
1462                 if (desc.count > 0)
1463                         break;
1464         }
1465 out:
1466         blk_finish_plug(&plug);
1467         return retval;
1468 }
1469 EXPORT_SYMBOL(generic_file_aio_read);
1470
1471 static ssize_t
1472 do_readahead(struct address_space *mapping, struct file *filp,
1473              pgoff_t index, unsigned long nr)
1474 {
1475         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1476                 return -EINVAL;
1477
1478         force_page_cache_readahead(mapping, filp, index, nr);
1479         return 0;
1480 }
1481
1482 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1483 {
1484         ssize_t ret;
1485         struct file *file;
1486
1487         ret = -EBADF;
1488         file = fget(fd);
1489         if (file) {
1490                 if (file->f_mode & FMODE_READ) {
1491                         struct address_space *mapping = file->f_mapping;
1492                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1493                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1494                         unsigned long len = end - start + 1;
1495                         ret = do_readahead(mapping, file, start, len);
1496                 }
1497                 fput(file);
1498         }
1499         return ret;
1500 }
1501 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1502 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1503 {
1504         return SYSC_readahead((int) fd, offset, (size_t) count);
1505 }
1506 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1507 #endif
1508
1509 #ifdef CONFIG_MMU
1510 /**
1511  * page_cache_read - adds requested page to the page cache if not already there
1512  * @file:       file to read
1513  * @offset:     page index
1514  *
1515  * This adds the requested page to the page cache if it isn't already there,
1516  * and schedules an I/O to read in its contents from disk.
1517  */
1518 static int page_cache_read(struct file *file, pgoff_t offset)
1519 {
1520         struct address_space *mapping = file->f_mapping;
1521         struct page *page; 
1522         int ret;
1523
1524         do {
1525                 page = page_cache_alloc_cold(mapping);
1526                 if (!page)
1527                         return -ENOMEM;
1528
1529                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1530                 if (ret == 0)
1531                         ret = mapping->a_ops->readpage(file, page);
1532                 else if (ret == -EEXIST)
1533                         ret = 0; /* losing race to add is OK */
1534
1535                 page_cache_release(page);
1536
1537         } while (ret == AOP_TRUNCATED_PAGE);
1538                 
1539         return ret;
1540 }
1541
1542 #define MMAP_LOTSAMISS  (100)
1543
1544 /*
1545  * Synchronous readahead happens when we don't even find
1546  * a page in the page cache at all.
1547  */
1548 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1549                                    struct file_ra_state *ra,
1550                                    struct file *file,
1551                                    pgoff_t offset)
1552 {
1553         unsigned long ra_pages;
1554         struct address_space *mapping = file->f_mapping;
1555
1556         /* If we don't want any read-ahead, don't bother */
1557         if (VM_RandomReadHint(vma))
1558                 return;
1559
1560         if (VM_SequentialReadHint(vma) ||
1561                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1562                 page_cache_sync_readahead(mapping, ra, file, offset,
1563                                           ra->ra_pages);
1564                 return;
1565         }
1566
1567         if (ra->mmap_miss < INT_MAX)
1568                 ra->mmap_miss++;
1569
1570         /*
1571          * Do we miss much more than hit in this file? If so,
1572          * stop bothering with read-ahead. It will only hurt.
1573          */
1574         if (ra->mmap_miss > MMAP_LOTSAMISS)
1575                 return;
1576
1577         /*
1578          * mmap read-around
1579          */
1580         ra_pages = max_sane_readahead(ra->ra_pages);
1581         if (ra_pages) {
1582                 ra->start = max_t(long, 0, offset - ra_pages/2);
1583                 ra->size = ra_pages;
1584                 ra->async_size = 0;
1585                 ra_submit(ra, mapping, file);
1586         }
1587 }
1588
1589 /*
1590  * Asynchronous readahead happens when we find the page and PG_readahead,
1591  * so we want to possibly extend the readahead further..
1592  */
1593 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1594                                     struct file_ra_state *ra,
1595                                     struct file *file,
1596                                     struct page *page,
1597                                     pgoff_t offset)
1598 {
1599         struct address_space *mapping = file->f_mapping;
1600
1601         /* If we don't want any read-ahead, don't bother */
1602         if (VM_RandomReadHint(vma))
1603                 return;
1604         if (ra->mmap_miss > 0)
1605                 ra->mmap_miss--;
1606         if (PageReadahead(page))
1607                 page_cache_async_readahead(mapping, ra, file,
1608                                            page, offset, ra->ra_pages);
1609 }
1610
1611 /**
1612  * filemap_fault - read in file data for page fault handling
1613  * @vma:        vma in which the fault was taken
1614  * @vmf:        struct vm_fault containing details of the fault
1615  *
1616  * filemap_fault() is invoked via the vma operations vector for a
1617  * mapped memory region to read in file data during a page fault.
1618  *
1619  * The goto's are kind of ugly, but this streamlines the normal case of having
1620  * it in the page cache, and handles the special cases reasonably without
1621  * having a lot of duplicated code.
1622  */
1623 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1624 {
1625         int error;
1626         struct file *file = vma->vm_file;
1627         struct address_space *mapping = file->f_mapping;
1628         struct file_ra_state *ra = &file->f_ra;
1629         struct inode *inode = mapping->host;
1630         pgoff_t offset = vmf->pgoff;
1631         struct page *page;
1632         pgoff_t size;
1633         int ret = 0;
1634
1635         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1636         if (offset >= size)
1637                 return VM_FAULT_SIGBUS;
1638
1639         /*
1640          * Do we have something in the page cache already?
1641          */
1642         page = find_get_page(mapping, offset);
1643         if (likely(page)) {
1644                 /*
1645                  * We found the page, so try async readahead before
1646                  * waiting for the lock.
1647                  */
1648                 do_async_mmap_readahead(vma, ra, file, page, offset);
1649         } else {
1650                 /* No page in the page cache at all */
1651                 do_sync_mmap_readahead(vma, ra, file, offset);
1652                 count_vm_event(PGMAJFAULT);
1653                 ret = VM_FAULT_MAJOR;
1654 retry_find:
1655                 page = find_get_page(mapping, offset);
1656                 if (!page)
1657                         goto no_cached_page;
1658         }
1659
1660         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1661                 page_cache_release(page);
1662                 return ret | VM_FAULT_RETRY;
1663         }
1664
1665         /* Did it get truncated? */
1666         if (unlikely(page->mapping != mapping)) {
1667                 unlock_page(page);
1668                 put_page(page);
1669                 goto retry_find;
1670         }
1671         VM_BUG_ON(page->index != offset);
1672
1673         /*
1674          * We have a locked page in the page cache, now we need to check
1675          * that it's up-to-date. If not, it is going to be due to an error.
1676          */
1677         if (unlikely(!PageUptodate(page)))
1678                 goto page_not_uptodate;
1679
1680         /*
1681          * Found the page and have a reference on it.
1682          * We must recheck i_size under page lock.
1683          */
1684         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1685         if (unlikely(offset >= size)) {
1686                 unlock_page(page);
1687                 page_cache_release(page);
1688                 return VM_FAULT_SIGBUS;
1689         }
1690
1691         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1692         vmf->page = page;
1693         return ret | VM_FAULT_LOCKED;
1694
1695 no_cached_page:
1696         /*
1697          * We're only likely to ever get here if MADV_RANDOM is in
1698          * effect.
1699          */
1700         error = page_cache_read(file, offset);
1701
1702         /*
1703          * The page we want has now been added to the page cache.
1704          * In the unlikely event that someone removed it in the
1705          * meantime, we'll just come back here and read it again.
1706          */
1707         if (error >= 0)
1708                 goto retry_find;
1709
1710         /*
1711          * An error return from page_cache_read can result if the
1712          * system is low on memory, or a problem occurs while trying
1713          * to schedule I/O.
1714          */
1715         if (error == -ENOMEM)
1716                 return VM_FAULT_OOM;
1717         return VM_FAULT_SIGBUS;
1718
1719 page_not_uptodate:
1720         /*
1721          * Umm, take care of errors if the page isn't up-to-date.
1722          * Try to re-read it _once_. We do this synchronously,
1723          * because there really aren't any performance issues here
1724          * and we need to check for errors.
1725          */
1726         ClearPageError(page);
1727         error = mapping->a_ops->readpage(file, page);
1728         if (!error) {
1729                 wait_on_page_locked(page);
1730                 if (!PageUptodate(page))
1731                         error = -EIO;
1732         }
1733         page_cache_release(page);
1734
1735         if (!error || error == AOP_TRUNCATED_PAGE)
1736                 goto retry_find;
1737
1738         /* Things didn't work out. Return zero to tell the mm layer so. */
1739         shrink_readahead_size_eio(file, ra);
1740         return VM_FAULT_SIGBUS;
1741 }
1742 EXPORT_SYMBOL(filemap_fault);
1743
1744 const struct vm_operations_struct generic_file_vm_ops = {
1745         .fault          = filemap_fault,
1746 };
1747
1748 /* This is used for a general mmap of a disk file */
1749
1750 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1751 {
1752         struct address_space *mapping = file->f_mapping;
1753
1754         if (!mapping->a_ops->readpage)
1755                 return -ENOEXEC;
1756         file_accessed(file);
1757         vma->vm_ops = &generic_file_vm_ops;
1758         vma->vm_flags |= VM_CAN_NONLINEAR;
1759         return 0;
1760 }
1761
1762 /*
1763  * This is for filesystems which do not implement ->writepage.
1764  */
1765 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1766 {
1767         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1768                 return -EINVAL;
1769         return generic_file_mmap(file, vma);
1770 }
1771 #else
1772 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1773 {
1774         return -ENOSYS;
1775 }
1776 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1777 {
1778         return -ENOSYS;
1779 }
1780 #endif /* CONFIG_MMU */
1781
1782 EXPORT_SYMBOL(generic_file_mmap);
1783 EXPORT_SYMBOL(generic_file_readonly_mmap);
1784
1785 static struct page *__read_cache_page(struct address_space *mapping,
1786                                 pgoff_t index,
1787                                 int (*filler)(void *,struct page*),
1788                                 void *data,
1789                                 gfp_t gfp)
1790 {
1791         struct page *page;
1792         int err;
1793 repeat:
1794         page = find_get_page(mapping, index);
1795         if (!page) {
1796                 page = __page_cache_alloc(gfp | __GFP_COLD);
1797                 if (!page)
1798                         return ERR_PTR(-ENOMEM);
1799                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1800                 if (unlikely(err)) {
1801                         page_cache_release(page);
1802                         if (err == -EEXIST)
1803                                 goto repeat;
1804                         /* Presumably ENOMEM for radix tree node */
1805                         return ERR_PTR(err);
1806                 }
1807                 err = filler(data, page);
1808                 if (err < 0) {
1809                         page_cache_release(page);
1810                         page = ERR_PTR(err);
1811                 }
1812         }
1813         return page;
1814 }
1815
1816 static struct page *do_read_cache_page(struct address_space *mapping,
1817                                 pgoff_t index,
1818                                 int (*filler)(void *,struct page*),
1819                                 void *data,
1820                                 gfp_t gfp)
1821
1822 {
1823         struct page *page;
1824         int err;
1825
1826 retry:
1827         page = __read_cache_page(mapping, index, filler, data, gfp);
1828         if (IS_ERR(page))
1829                 return page;
1830         if (PageUptodate(page))
1831                 goto out;
1832
1833         lock_page(page);
1834         if (!page->mapping) {
1835                 unlock_page(page);
1836                 page_cache_release(page);
1837                 goto retry;
1838         }
1839         if (PageUptodate(page)) {
1840                 unlock_page(page);
1841                 goto out;
1842         }
1843         err = filler(data, page);
1844         if (err < 0) {
1845                 page_cache_release(page);
1846                 return ERR_PTR(err);
1847         }
1848 out:
1849         mark_page_accessed(page);
1850         return page;
1851 }
1852
1853 /**
1854  * read_cache_page_async - read into page cache, fill it if needed
1855  * @mapping:    the page's address_space
1856  * @index:      the page index
1857  * @filler:     function to perform the read
1858  * @data:       destination for read data
1859  *
1860  * Same as read_cache_page, but don't wait for page to become unlocked
1861  * after submitting it to the filler.
1862  *
1863  * Read into the page cache. If a page already exists, and PageUptodate() is
1864  * not set, try to fill the page but don't wait for it to become unlocked.
1865  *
1866  * If the page does not get brought uptodate, return -EIO.
1867  */
1868 struct page *read_cache_page_async(struct address_space *mapping,
1869                                 pgoff_t index,
1870                                 int (*filler)(void *,struct page*),
1871                                 void *data)
1872 {
1873         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1874 }
1875 EXPORT_SYMBOL(read_cache_page_async);
1876
1877 static struct page *wait_on_page_read(struct page *page)
1878 {
1879         if (!IS_ERR(page)) {
1880                 wait_on_page_locked(page);
1881                 if (!PageUptodate(page)) {
1882                         page_cache_release(page);
1883                         page = ERR_PTR(-EIO);
1884                 }
1885         }
1886         return page;
1887 }
1888
1889 /**
1890  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1891  * @mapping:    the page's address_space
1892  * @index:      the page index
1893  * @gfp:        the page allocator flags to use if allocating
1894  *
1895  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1896  * any new page allocations done using the specified allocation flags. Note
1897  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1898  * expect to do this atomically or anything like that - but you can pass in
1899  * other page requirements.
1900  *
1901  * If the page does not get brought uptodate, return -EIO.
1902  */
1903 struct page *read_cache_page_gfp(struct address_space *mapping,
1904                                 pgoff_t index,
1905                                 gfp_t gfp)
1906 {
1907         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1908
1909         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1910 }
1911 EXPORT_SYMBOL(read_cache_page_gfp);
1912
1913 /**
1914  * read_cache_page - read into page cache, fill it if needed
1915  * @mapping:    the page's address_space
1916  * @index:      the page index
1917  * @filler:     function to perform the read
1918  * @data:       destination for read data
1919  *
1920  * Read into the page cache. If a page already exists, and PageUptodate() is
1921  * not set, try to fill the page then wait for it to become unlocked.
1922  *
1923  * If the page does not get brought uptodate, return -EIO.
1924  */
1925 struct page *read_cache_page(struct address_space *mapping,
1926                                 pgoff_t index,
1927                                 int (*filler)(void *,struct page*),
1928                                 void *data)
1929 {
1930         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1931 }
1932 EXPORT_SYMBOL(read_cache_page);
1933
1934 /*
1935  * The logic we want is
1936  *
1937  *      if suid or (sgid and xgrp)
1938  *              remove privs
1939  */
1940 int should_remove_suid(struct dentry *dentry)
1941 {
1942         mode_t mode = dentry->d_inode->i_mode;
1943         int kill = 0;
1944
1945         /* suid always must be killed */
1946         if (unlikely(mode & S_ISUID))
1947                 kill = ATTR_KILL_SUID;
1948
1949         /*
1950          * sgid without any exec bits is just a mandatory locking mark; leave
1951          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1952          */
1953         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1954                 kill |= ATTR_KILL_SGID;
1955
1956         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1957                 return kill;
1958
1959         return 0;
1960 }
1961 EXPORT_SYMBOL(should_remove_suid);
1962
1963 static int __remove_suid(struct dentry *dentry, int kill)
1964 {
1965         struct iattr newattrs;
1966
1967         newattrs.ia_valid = ATTR_FORCE | kill;
1968         return notify_change(dentry, &newattrs);
1969 }
1970
1971 int file_remove_suid(struct file *file)
1972 {
1973         struct dentry *dentry = file->f_path.dentry;
1974         int killsuid = should_remove_suid(dentry);
1975         int killpriv = security_inode_need_killpriv(dentry);
1976         int error = 0;
1977
1978         if (killpriv < 0)
1979                 return killpriv;
1980         if (killpriv)
1981                 error = security_inode_killpriv(dentry);
1982         if (!error && killsuid)
1983                 error = __remove_suid(dentry, killsuid);
1984
1985         return error;
1986 }
1987 EXPORT_SYMBOL(file_remove_suid);
1988
1989 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1990                         const struct iovec *iov, size_t base, size_t bytes)
1991 {
1992         size_t copied = 0, left = 0;
1993
1994         while (bytes) {
1995                 char __user *buf = iov->iov_base + base;
1996                 int copy = min(bytes, iov->iov_len - base);
1997
1998                 base = 0;
1999                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2000                 copied += copy;
2001                 bytes -= copy;
2002                 vaddr += copy;
2003                 iov++;
2004
2005                 if (unlikely(left))
2006                         break;
2007         }
2008         return copied - left;
2009 }
2010
2011 /*
2012  * Copy as much as we can into the page and return the number of bytes which
2013  * were successfully copied.  If a fault is encountered then return the number of
2014  * bytes which were copied.
2015  */
2016 size_t iov_iter_copy_from_user_atomic(struct page *page,
2017                 struct iov_iter *i, unsigned long offset, size_t bytes)
2018 {
2019         char *kaddr;
2020         size_t copied;
2021
2022         BUG_ON(!in_atomic());
2023         kaddr = kmap_atomic(page, KM_USER0);
2024         if (likely(i->nr_segs == 1)) {
2025                 int left;
2026                 char __user *buf = i->iov->iov_base + i->iov_offset;
2027                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2028                 copied = bytes - left;
2029         } else {
2030                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2031                                                 i->iov, i->iov_offset, bytes);
2032         }
2033         kunmap_atomic(kaddr, KM_USER0);
2034
2035         return copied;
2036 }
2037 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2038
2039 /*
2040  * This has the same sideeffects and return value as
2041  * iov_iter_copy_from_user_atomic().
2042  * The difference is that it attempts to resolve faults.
2043  * Page must not be locked.
2044  */
2045 size_t iov_iter_copy_from_user(struct page *page,
2046                 struct iov_iter *i, unsigned long offset, size_t bytes)
2047 {
2048         char *kaddr;
2049         size_t copied;
2050
2051         kaddr = kmap(page);
2052         if (likely(i->nr_segs == 1)) {
2053                 int left;
2054                 char __user *buf = i->iov->iov_base + i->iov_offset;
2055                 left = __copy_from_user(kaddr + offset, buf, bytes);
2056                 copied = bytes - left;
2057         } else {
2058                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2059                                                 i->iov, i->iov_offset, bytes);
2060         }
2061         kunmap(page);
2062         return copied;
2063 }
2064 EXPORT_SYMBOL(iov_iter_copy_from_user);
2065
2066 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2067 {
2068         BUG_ON(i->count < bytes);
2069
2070         if (likely(i->nr_segs == 1)) {
2071                 i->iov_offset += bytes;
2072                 i->count -= bytes;
2073         } else {
2074                 const struct iovec *iov = i->iov;
2075                 size_t base = i->iov_offset;
2076
2077                 /*
2078                  * The !iov->iov_len check ensures we skip over unlikely
2079                  * zero-length segments (without overruning the iovec).
2080                  */
2081                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2082                         int copy;
2083
2084                         copy = min(bytes, iov->iov_len - base);
2085                         BUG_ON(!i->count || i->count < copy);
2086                         i->count -= copy;
2087                         bytes -= copy;
2088                         base += copy;
2089                         if (iov->iov_len == base) {
2090                                 iov++;
2091                                 base = 0;
2092                         }
2093                 }
2094                 i->iov = iov;
2095                 i->iov_offset = base;
2096         }
2097 }
2098 EXPORT_SYMBOL(iov_iter_advance);
2099
2100 /*
2101  * Fault in the first iovec of the given iov_iter, to a maximum length
2102  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2103  * accessed (ie. because it is an invalid address).
2104  *
2105  * writev-intensive code may want this to prefault several iovecs -- that
2106  * would be possible (callers must not rely on the fact that _only_ the
2107  * first iovec will be faulted with the current implementation).
2108  */
2109 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2110 {
2111         char __user *buf = i->iov->iov_base + i->iov_offset;
2112         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2113         return fault_in_pages_readable(buf, bytes);
2114 }
2115 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2116
2117 /*
2118  * Return the count of just the current iov_iter segment.
2119  */
2120 size_t iov_iter_single_seg_count(struct iov_iter *i)
2121 {
2122         const struct iovec *iov = i->iov;
2123         if (i->nr_segs == 1)
2124                 return i->count;
2125         else
2126                 return min(i->count, iov->iov_len - i->iov_offset);
2127 }
2128 EXPORT_SYMBOL(iov_iter_single_seg_count);
2129
2130 /*
2131  * Performs necessary checks before doing a write
2132  *
2133  * Can adjust writing position or amount of bytes to write.
2134  * Returns appropriate error code that caller should return or
2135  * zero in case that write should be allowed.
2136  */
2137 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2138 {
2139         struct inode *inode = file->f_mapping->host;
2140         unsigned long limit = rlimit(RLIMIT_FSIZE);
2141
2142         if (unlikely(*pos < 0))
2143                 return -EINVAL;
2144
2145         if (!isblk) {
2146                 /* FIXME: this is for backwards compatibility with 2.4 */
2147                 if (file->f_flags & O_APPEND)
2148                         *pos = i_size_read(inode);
2149
2150                 if (limit != RLIM_INFINITY) {
2151                         if (*pos >= limit) {
2152                                 send_sig(SIGXFSZ, current, 0);
2153                                 return -EFBIG;
2154                         }
2155                         if (*count > limit - (typeof(limit))*pos) {
2156                                 *count = limit - (typeof(limit))*pos;
2157                         }
2158                 }
2159         }
2160
2161         /*
2162          * LFS rule
2163          */
2164         if (unlikely(*pos + *count > MAX_NON_LFS &&
2165                                 !(file->f_flags & O_LARGEFILE))) {
2166                 if (*pos >= MAX_NON_LFS) {
2167                         return -EFBIG;
2168                 }
2169                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2170                         *count = MAX_NON_LFS - (unsigned long)*pos;
2171                 }
2172         }
2173
2174         /*
2175          * Are we about to exceed the fs block limit ?
2176          *
2177          * If we have written data it becomes a short write.  If we have
2178          * exceeded without writing data we send a signal and return EFBIG.
2179          * Linus frestrict idea will clean these up nicely..
2180          */
2181         if (likely(!isblk)) {
2182                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2183                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2184                                 return -EFBIG;
2185                         }
2186                         /* zero-length writes at ->s_maxbytes are OK */
2187                 }
2188
2189                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2190                         *count = inode->i_sb->s_maxbytes - *pos;
2191         } else {
2192 #ifdef CONFIG_BLOCK
2193                 loff_t isize;
2194                 if (bdev_read_only(I_BDEV(inode)))
2195                         return -EPERM;
2196                 isize = i_size_read(inode);
2197                 if (*pos >= isize) {
2198                         if (*count || *pos > isize)
2199                                 return -ENOSPC;
2200                 }
2201
2202                 if (*pos + *count > isize)
2203                         *count = isize - *pos;
2204 #else
2205                 return -EPERM;
2206 #endif
2207         }
2208         return 0;
2209 }
2210 EXPORT_SYMBOL(generic_write_checks);
2211
2212 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2213                                 loff_t pos, unsigned len, unsigned flags,
2214                                 struct page **pagep, void **fsdata)
2215 {
2216         const struct address_space_operations *aops = mapping->a_ops;
2217
2218         return aops->write_begin(file, mapping, pos, len, flags,
2219                                                         pagep, fsdata);
2220 }
2221 EXPORT_SYMBOL(pagecache_write_begin);
2222
2223 int pagecache_write_end(struct file *file, struct address_space *mapping,
2224                                 loff_t pos, unsigned len, unsigned copied,
2225                                 struct page *page, void *fsdata)
2226 {
2227         const struct address_space_operations *aops = mapping->a_ops;
2228
2229         mark_page_accessed(page);
2230         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2231 }
2232 EXPORT_SYMBOL(pagecache_write_end);
2233
2234 ssize_t
2235 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2236                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2237                 size_t count, size_t ocount)
2238 {
2239         struct file     *file = iocb->ki_filp;
2240         struct address_space *mapping = file->f_mapping;
2241         struct inode    *inode = mapping->host;
2242         ssize_t         written;
2243         size_t          write_len;
2244         pgoff_t         end;
2245
2246         if (count != ocount)
2247                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2248
2249         write_len = iov_length(iov, *nr_segs);
2250         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2251
2252         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2253         if (written)
2254                 goto out;
2255
2256         /*
2257          * After a write we want buffered reads to be sure to go to disk to get
2258          * the new data.  We invalidate clean cached page from the region we're
2259          * about to write.  We do this *before* the write so that we can return
2260          * without clobbering -EIOCBQUEUED from ->direct_IO().
2261          */
2262         if (mapping->nrpages) {
2263                 written = invalidate_inode_pages2_range(mapping,
2264                                         pos >> PAGE_CACHE_SHIFT, end);
2265                 /*
2266                  * If a page can not be invalidated, return 0 to fall back
2267                  * to buffered write.
2268                  */
2269                 if (written) {
2270                         if (written == -EBUSY)
2271                                 return 0;
2272                         goto out;
2273                 }
2274         }
2275
2276         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2277
2278         /*
2279          * Finally, try again to invalidate clean pages which might have been
2280          * cached by non-direct readahead, or faulted in by get_user_pages()
2281          * if the source of the write was an mmap'ed region of the file
2282          * we're writing.  Either one is a pretty crazy thing to do,
2283          * so we don't support it 100%.  If this invalidation
2284          * fails, tough, the write still worked...
2285          */
2286         if (mapping->nrpages) {
2287                 invalidate_inode_pages2_range(mapping,
2288                                               pos >> PAGE_CACHE_SHIFT, end);
2289         }
2290
2291         if (written > 0) {
2292                 pos += written;
2293                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2294                         i_size_write(inode, pos);
2295                         mark_inode_dirty(inode);
2296                 }
2297                 *ppos = pos;
2298         }
2299 out:
2300         return written;
2301 }
2302 EXPORT_SYMBOL(generic_file_direct_write);
2303
2304 /*
2305  * Find or create a page at the given pagecache position. Return the locked
2306  * page. This function is specifically for buffered writes.
2307  */
2308 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2309                                         pgoff_t index, unsigned flags)
2310 {
2311         int status;
2312         struct page *page;
2313         gfp_t gfp_notmask = 0;
2314         if (flags & AOP_FLAG_NOFS)
2315                 gfp_notmask = __GFP_FS;
2316 repeat:
2317         page = find_lock_page(mapping, index);
2318         if (page)
2319                 return page;
2320
2321         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2322         if (!page)
2323                 return NULL;
2324         status = add_to_page_cache_lru(page, mapping, index,
2325                                                 GFP_KERNEL & ~gfp_notmask);
2326         if (unlikely(status)) {
2327                 page_cache_release(page);
2328                 if (status == -EEXIST)
2329                         goto repeat;
2330                 return NULL;
2331         }
2332         return page;
2333 }
2334 EXPORT_SYMBOL(grab_cache_page_write_begin);
2335
2336 static ssize_t generic_perform_write(struct file *file,
2337                                 struct iov_iter *i, loff_t pos)
2338 {
2339         struct address_space *mapping = file->f_mapping;
2340         const struct address_space_operations *a_ops = mapping->a_ops;
2341         long status = 0;
2342         ssize_t written = 0;
2343         unsigned int flags = 0;
2344
2345         /*
2346          * Copies from kernel address space cannot fail (NFSD is a big user).
2347          */
2348         if (segment_eq(get_fs(), KERNEL_DS))
2349                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2350
2351         do {
2352                 struct page *page;
2353                 unsigned long offset;   /* Offset into pagecache page */
2354                 unsigned long bytes;    /* Bytes to write to page */
2355                 size_t copied;          /* Bytes copied from user */
2356                 void *fsdata;
2357
2358                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2359                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2360                                                 iov_iter_count(i));
2361
2362 again:
2363
2364                 /*
2365                  * Bring in the user page that we will copy from _first_.
2366                  * Otherwise there's a nasty deadlock on copying from the
2367                  * same page as we're writing to, without it being marked
2368                  * up-to-date.
2369                  *
2370                  * Not only is this an optimisation, but it is also required
2371                  * to check that the address is actually valid, when atomic
2372                  * usercopies are used, below.
2373                  */
2374                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2375                         status = -EFAULT;
2376                         break;
2377                 }
2378
2379                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2380                                                 &page, &fsdata);
2381                 if (unlikely(status))
2382                         break;
2383
2384                 if (mapping_writably_mapped(mapping))
2385                         flush_dcache_page(page);
2386
2387                 pagefault_disable();
2388                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2389                 pagefault_enable();
2390                 flush_dcache_page(page);
2391
2392                 mark_page_accessed(page);
2393                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2394                                                 page, fsdata);
2395                 if (unlikely(status < 0))
2396                         break;
2397                 copied = status;
2398
2399                 cond_resched();
2400
2401                 iov_iter_advance(i, copied);
2402                 if (unlikely(copied == 0)) {
2403                         /*
2404                          * If we were unable to copy any data at all, we must
2405                          * fall back to a single segment length write.
2406                          *
2407                          * If we didn't fallback here, we could livelock
2408                          * because not all segments in the iov can be copied at
2409                          * once without a pagefault.
2410                          */
2411                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2412                                                 iov_iter_single_seg_count(i));
2413                         goto again;
2414                 }
2415                 pos += copied;
2416                 written += copied;
2417
2418                 balance_dirty_pages_ratelimited(mapping);
2419
2420         } while (iov_iter_count(i));
2421
2422         return written ? written : status;
2423 }
2424
2425 ssize_t
2426 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2427                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2428                 size_t count, ssize_t written)
2429 {
2430         struct file *file = iocb->ki_filp;
2431         ssize_t status;
2432         struct iov_iter i;
2433
2434         iov_iter_init(&i, iov, nr_segs, count, written);
2435         status = generic_perform_write(file, &i, pos);
2436
2437         if (likely(status >= 0)) {
2438                 written += status;
2439                 *ppos = pos + status;
2440         }
2441         
2442         return written ? written : status;
2443 }
2444 EXPORT_SYMBOL(generic_file_buffered_write);
2445
2446 /**
2447  * __generic_file_aio_write - write data to a file
2448  * @iocb:       IO state structure (file, offset, etc.)
2449  * @iov:        vector with data to write
2450  * @nr_segs:    number of segments in the vector
2451  * @ppos:       position where to write
2452  *
2453  * This function does all the work needed for actually writing data to a
2454  * file. It does all basic checks, removes SUID from the file, updates
2455  * modification times and calls proper subroutines depending on whether we
2456  * do direct IO or a standard buffered write.
2457  *
2458  * It expects i_mutex to be grabbed unless we work on a block device or similar
2459  * object which does not need locking at all.
2460  *
2461  * This function does *not* take care of syncing data in case of O_SYNC write.
2462  * A caller has to handle it. This is mainly due to the fact that we want to
2463  * avoid syncing under i_mutex.
2464  */
2465 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2466                                  unsigned long nr_segs, loff_t *ppos)
2467 {
2468         struct file *file = iocb->ki_filp;
2469         struct address_space * mapping = file->f_mapping;
2470         size_t ocount;          /* original count */
2471         size_t count;           /* after file limit checks */
2472         struct inode    *inode = mapping->host;
2473         loff_t          pos;
2474         ssize_t         written;
2475         ssize_t         err;
2476
2477         ocount = 0;
2478         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2479         if (err)
2480                 return err;
2481
2482         count = ocount;
2483         pos = *ppos;
2484
2485         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2486
2487         /* We can write back this queue in page reclaim */
2488         current->backing_dev_info = mapping->backing_dev_info;
2489         written = 0;
2490
2491         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2492         if (err)
2493                 goto out;
2494
2495         if (count == 0)
2496                 goto out;
2497
2498         err = file_remove_suid(file);
2499         if (err)
2500                 goto out;
2501
2502         file_update_time(file);
2503
2504         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2505         if (unlikely(file->f_flags & O_DIRECT)) {
2506                 loff_t endbyte;
2507                 ssize_t written_buffered;
2508
2509                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2510                                                         ppos, count, ocount);
2511                 if (written < 0 || written == count)
2512                         goto out;
2513                 /*
2514                  * direct-io write to a hole: fall through to buffered I/O
2515                  * for completing the rest of the request.
2516                  */
2517                 pos += written;
2518                 count -= written;
2519                 written_buffered = generic_file_buffered_write(iocb, iov,
2520                                                 nr_segs, pos, ppos, count,
2521                                                 written);
2522                 /*
2523                  * If generic_file_buffered_write() retuned a synchronous error
2524                  * then we want to return the number of bytes which were
2525                  * direct-written, or the error code if that was zero.  Note
2526                  * that this differs from normal direct-io semantics, which
2527                  * will return -EFOO even if some bytes were written.
2528                  */
2529                 if (written_buffered < 0) {
2530                         err = written_buffered;
2531                         goto out;
2532                 }
2533
2534                 /*
2535                  * We need to ensure that the page cache pages are written to
2536                  * disk and invalidated to preserve the expected O_DIRECT
2537                  * semantics.
2538                  */
2539                 endbyte = pos + written_buffered - written - 1;
2540                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2541                 if (err == 0) {
2542                         written = written_buffered;
2543                         invalidate_mapping_pages(mapping,
2544                                                  pos >> PAGE_CACHE_SHIFT,
2545                                                  endbyte >> PAGE_CACHE_SHIFT);
2546                 } else {
2547                         /*
2548                          * We don't know how much we wrote, so just return
2549                          * the number of bytes which were direct-written
2550                          */
2551                 }
2552         } else {
2553                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2554                                 pos, ppos, count, written);
2555         }
2556 out:
2557         current->backing_dev_info = NULL;
2558         return written ? written : err;
2559 }
2560 EXPORT_SYMBOL(__generic_file_aio_write);
2561
2562 /**
2563  * generic_file_aio_write - write data to a file
2564  * @iocb:       IO state structure
2565  * @iov:        vector with data to write
2566  * @nr_segs:    number of segments in the vector
2567  * @pos:        position in file where to write
2568  *
2569  * This is a wrapper around __generic_file_aio_write() to be used by most
2570  * filesystems. It takes care of syncing the file in case of O_SYNC file
2571  * and acquires i_mutex as needed.
2572  */
2573 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2574                 unsigned long nr_segs, loff_t pos)
2575 {
2576         struct file *file = iocb->ki_filp;
2577         struct inode *inode = file->f_mapping->host;
2578         struct blk_plug plug;
2579         ssize_t ret;
2580
2581         BUG_ON(iocb->ki_pos != pos);
2582
2583         mutex_lock(&inode->i_mutex);
2584         blk_start_plug(&plug);
2585         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2586         mutex_unlock(&inode->i_mutex);
2587
2588         if (ret > 0 || ret == -EIOCBQUEUED) {
2589                 ssize_t err;
2590
2591                 err = generic_write_sync(file, pos, ret);
2592                 if (err < 0 && ret > 0)
2593                         ret = err;
2594         }
2595         blk_finish_plug(&plug);
2596         return ret;
2597 }
2598 EXPORT_SYMBOL(generic_file_aio_write);
2599
2600 /**
2601  * try_to_release_page() - release old fs-specific metadata on a page
2602  *
2603  * @page: the page which the kernel is trying to free
2604  * @gfp_mask: memory allocation flags (and I/O mode)
2605  *
2606  * The address_space is to try to release any data against the page
2607  * (presumably at page->private).  If the release was successful, return `1'.
2608  * Otherwise return zero.
2609  *
2610  * This may also be called if PG_fscache is set on a page, indicating that the
2611  * page is known to the local caching routines.
2612  *
2613  * The @gfp_mask argument specifies whether I/O may be performed to release
2614  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2615  *
2616  */
2617 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2618 {
2619         struct address_space * const mapping = page->mapping;
2620
2621         BUG_ON(!PageLocked(page));
2622         if (PageWriteback(page))
2623                 return 0;
2624
2625         if (mapping && mapping->a_ops->releasepage)
2626                 return mapping->a_ops->releasepage(page, gfp_mask);
2627         return try_to_free_buffers(page);
2628 }
2629
2630 EXPORT_SYMBOL(try_to_release_page);