MIPS: JZ4740: Fix GCC 4.6.0 build error.
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "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 /**
566  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
567  * @page: Page defining the wait queue of interest
568  * @waiter: Waiter to add to the queue
569  *
570  * Add an arbitrary @waiter to the wait queue for the nominated @page.
571  */
572 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
573 {
574         wait_queue_head_t *q = page_waitqueue(page);
575         unsigned long flags;
576
577         spin_lock_irqsave(&q->lock, flags);
578         __add_wait_queue(q, waiter);
579         spin_unlock_irqrestore(&q->lock, flags);
580 }
581 EXPORT_SYMBOL_GPL(add_page_wait_queue);
582
583 /**
584  * unlock_page - unlock a locked page
585  * @page: the page
586  *
587  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
588  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
589  * mechananism between PageLocked pages and PageWriteback pages is shared.
590  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
591  *
592  * The mb is necessary to enforce ordering between the clear_bit and the read
593  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
594  */
595 void unlock_page(struct page *page)
596 {
597         VM_BUG_ON(!PageLocked(page));
598         clear_bit_unlock(PG_locked, &page->flags);
599         smp_mb__after_clear_bit();
600         wake_up_page(page, PG_locked);
601 }
602 EXPORT_SYMBOL(unlock_page);
603
604 /**
605  * end_page_writeback - end writeback against a page
606  * @page: the page
607  */
608 void end_page_writeback(struct page *page)
609 {
610         if (TestClearPageReclaim(page))
611                 rotate_reclaimable_page(page);
612
613         if (!test_clear_page_writeback(page))
614                 BUG();
615
616         smp_mb__after_clear_bit();
617         wake_up_page(page, PG_writeback);
618 }
619 EXPORT_SYMBOL(end_page_writeback);
620
621 /**
622  * __lock_page - get a lock on the page, assuming we need to sleep to get it
623  * @page: the page to lock
624  */
625 void __lock_page(struct page *page)
626 {
627         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
628
629         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
630                                                         TASK_UNINTERRUPTIBLE);
631 }
632 EXPORT_SYMBOL(__lock_page);
633
634 int __lock_page_killable(struct page *page)
635 {
636         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
637
638         return __wait_on_bit_lock(page_waitqueue(page), &wait,
639                                         sleep_on_page_killable, TASK_KILLABLE);
640 }
641 EXPORT_SYMBOL_GPL(__lock_page_killable);
642
643 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
644                          unsigned int flags)
645 {
646         if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
647                 __lock_page(page);
648                 return 1;
649         } else {
650                 if (!(flags & FAULT_FLAG_RETRY_NOWAIT)) {
651                         up_read(&mm->mmap_sem);
652                         wait_on_page_locked(page);
653                 }
654                 return 0;
655         }
656 }
657
658 /**
659  * find_get_page - find and get a page reference
660  * @mapping: the address_space to search
661  * @offset: the page index
662  *
663  * Is there a pagecache struct page at the given (mapping, offset) tuple?
664  * If yes, increment its refcount and return it; if no, return NULL.
665  */
666 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
667 {
668         void **pagep;
669         struct page *page;
670
671         rcu_read_lock();
672 repeat:
673         page = NULL;
674         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
675         if (pagep) {
676                 page = radix_tree_deref_slot(pagep);
677                 if (unlikely(!page))
678                         goto out;
679                 if (radix_tree_deref_retry(page))
680                         goto repeat;
681
682                 if (!page_cache_get_speculative(page))
683                         goto repeat;
684
685                 /*
686                  * Has the page moved?
687                  * This is part of the lockless pagecache protocol. See
688                  * include/linux/pagemap.h for details.
689                  */
690                 if (unlikely(page != *pagep)) {
691                         page_cache_release(page);
692                         goto repeat;
693                 }
694         }
695 out:
696         rcu_read_unlock();
697
698         return page;
699 }
700 EXPORT_SYMBOL(find_get_page);
701
702 /**
703  * find_lock_page - locate, pin and lock a pagecache page
704  * @mapping: the address_space to search
705  * @offset: the page index
706  *
707  * Locates the desired pagecache page, locks it, increments its reference
708  * count and returns its address.
709  *
710  * Returns zero if the page was not present. find_lock_page() may sleep.
711  */
712 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
713 {
714         struct page *page;
715
716 repeat:
717         page = find_get_page(mapping, offset);
718         if (page) {
719                 lock_page(page);
720                 /* Has the page been truncated? */
721                 if (unlikely(page->mapping != mapping)) {
722                         unlock_page(page);
723                         page_cache_release(page);
724                         goto repeat;
725                 }
726                 VM_BUG_ON(page->index != offset);
727         }
728         return page;
729 }
730 EXPORT_SYMBOL(find_lock_page);
731
732 /**
733  * find_or_create_page - locate or add a pagecache page
734  * @mapping: the page's address_space
735  * @index: the page's index into the mapping
736  * @gfp_mask: page allocation mode
737  *
738  * Locates a page in the pagecache.  If the page is not present, a new page
739  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
740  * LRU list.  The returned page is locked and has its reference count
741  * incremented.
742  *
743  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
744  * allocation!
745  *
746  * find_or_create_page() returns the desired page's address, or zero on
747  * memory exhaustion.
748  */
749 struct page *find_or_create_page(struct address_space *mapping,
750                 pgoff_t index, gfp_t gfp_mask)
751 {
752         struct page *page;
753         int err;
754 repeat:
755         page = find_lock_page(mapping, index);
756         if (!page) {
757                 page = __page_cache_alloc(gfp_mask);
758                 if (!page)
759                         return NULL;
760                 /*
761                  * We want a regular kernel memory (not highmem or DMA etc)
762                  * allocation for the radix tree nodes, but we need to honour
763                  * the context-specific requirements the caller has asked for.
764                  * GFP_RECLAIM_MASK collects those requirements.
765                  */
766                 err = add_to_page_cache_lru(page, mapping, index,
767                         (gfp_mask & GFP_RECLAIM_MASK));
768                 if (unlikely(err)) {
769                         page_cache_release(page);
770                         page = NULL;
771                         if (err == -EEXIST)
772                                 goto repeat;
773                 }
774         }
775         return page;
776 }
777 EXPORT_SYMBOL(find_or_create_page);
778
779 /**
780  * find_get_pages - gang pagecache lookup
781  * @mapping:    The address_space to search
782  * @start:      The starting page index
783  * @nr_pages:   The maximum number of pages
784  * @pages:      Where the resulting pages are placed
785  *
786  * find_get_pages() will search for and return a group of up to
787  * @nr_pages pages in the mapping.  The pages are placed at @pages.
788  * find_get_pages() takes a reference against the returned pages.
789  *
790  * The search returns a group of mapping-contiguous pages with ascending
791  * indexes.  There may be holes in the indices due to not-present pages.
792  *
793  * find_get_pages() returns the number of pages which were found.
794  */
795 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
796                             unsigned int nr_pages, struct page **pages)
797 {
798         unsigned int i;
799         unsigned int ret;
800         unsigned int nr_found;
801
802         rcu_read_lock();
803 restart:
804         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
805                                 (void ***)pages, start, nr_pages);
806         ret = 0;
807         for (i = 0; i < nr_found; i++) {
808                 struct page *page;
809 repeat:
810                 page = radix_tree_deref_slot((void **)pages[i]);
811                 if (unlikely(!page))
812                         continue;
813
814                 /*
815                  * This can only trigger when the entry at index 0 moves out
816                  * of or back to the root: none yet gotten, safe to restart.
817                  */
818                 if (radix_tree_deref_retry(page)) {
819                         WARN_ON(start | i);
820                         goto restart;
821                 }
822
823                 if (!page_cache_get_speculative(page))
824                         goto repeat;
825
826                 /* Has the page moved? */
827                 if (unlikely(page != *((void **)pages[i]))) {
828                         page_cache_release(page);
829                         goto repeat;
830                 }
831
832                 pages[ret] = page;
833                 ret++;
834         }
835
836         /*
837          * If all entries were removed before we could secure them,
838          * try again, because callers stop trying once 0 is returned.
839          */
840         if (unlikely(!ret && nr_found))
841                 goto restart;
842         rcu_read_unlock();
843         return ret;
844 }
845
846 /**
847  * find_get_pages_contig - gang contiguous pagecache lookup
848  * @mapping:    The address_space to search
849  * @index:      The starting page index
850  * @nr_pages:   The maximum number of pages
851  * @pages:      Where the resulting pages are placed
852  *
853  * find_get_pages_contig() works exactly like find_get_pages(), except
854  * that the returned number of pages are guaranteed to be contiguous.
855  *
856  * find_get_pages_contig() returns the number of pages which were found.
857  */
858 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
859                                unsigned int nr_pages, struct page **pages)
860 {
861         unsigned int i;
862         unsigned int ret;
863         unsigned int nr_found;
864
865         rcu_read_lock();
866 restart:
867         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
868                                 (void ***)pages, index, nr_pages);
869         ret = 0;
870         for (i = 0; i < nr_found; i++) {
871                 struct page *page;
872 repeat:
873                 page = radix_tree_deref_slot((void **)pages[i]);
874                 if (unlikely(!page))
875                         continue;
876
877                 /*
878                  * This can only trigger when the entry at index 0 moves out
879                  * of or back to the root: none yet gotten, safe to restart.
880                  */
881                 if (radix_tree_deref_retry(page))
882                         goto restart;
883
884                 if (!page_cache_get_speculative(page))
885                         goto repeat;
886
887                 /* Has the page moved? */
888                 if (unlikely(page != *((void **)pages[i]))) {
889                         page_cache_release(page);
890                         goto repeat;
891                 }
892
893                 /*
894                  * must check mapping and index after taking the ref.
895                  * otherwise we can get both false positives and false
896                  * negatives, which is just confusing to the caller.
897                  */
898                 if (page->mapping == NULL || page->index != index) {
899                         page_cache_release(page);
900                         break;
901                 }
902
903                 pages[ret] = page;
904                 ret++;
905                 index++;
906         }
907         rcu_read_unlock();
908         return ret;
909 }
910 EXPORT_SYMBOL(find_get_pages_contig);
911
912 /**
913  * find_get_pages_tag - find and return pages that match @tag
914  * @mapping:    the address_space to search
915  * @index:      the starting page index
916  * @tag:        the tag index
917  * @nr_pages:   the maximum number of pages
918  * @pages:      where the resulting pages are placed
919  *
920  * Like find_get_pages, except we only return pages which are tagged with
921  * @tag.   We update @index to index the next page for the traversal.
922  */
923 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
924                         int tag, unsigned int nr_pages, struct page **pages)
925 {
926         unsigned int i;
927         unsigned int ret;
928         unsigned int nr_found;
929
930         rcu_read_lock();
931 restart:
932         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
933                                 (void ***)pages, *index, nr_pages, tag);
934         ret = 0;
935         for (i = 0; i < nr_found; i++) {
936                 struct page *page;
937 repeat:
938                 page = radix_tree_deref_slot((void **)pages[i]);
939                 if (unlikely(!page))
940                         continue;
941
942                 /*
943                  * This can only trigger when the entry at index 0 moves out
944                  * of or back to the root: none yet gotten, safe to restart.
945                  */
946                 if (radix_tree_deref_retry(page))
947                         goto restart;
948
949                 if (!page_cache_get_speculative(page))
950                         goto repeat;
951
952                 /* Has the page moved? */
953                 if (unlikely(page != *((void **)pages[i]))) {
954                         page_cache_release(page);
955                         goto repeat;
956                 }
957
958                 pages[ret] = page;
959                 ret++;
960         }
961
962         /*
963          * If all entries were removed before we could secure them,
964          * try again, because callers stop trying once 0 is returned.
965          */
966         if (unlikely(!ret && nr_found))
967                 goto restart;
968         rcu_read_unlock();
969
970         if (ret)
971                 *index = pages[ret - 1]->index + 1;
972
973         return ret;
974 }
975 EXPORT_SYMBOL(find_get_pages_tag);
976
977 /**
978  * grab_cache_page_nowait - returns locked page at given index in given cache
979  * @mapping: target address_space
980  * @index: the page index
981  *
982  * Same as grab_cache_page(), but do not wait if the page is unavailable.
983  * This is intended for speculative data generators, where the data can
984  * be regenerated if the page couldn't be grabbed.  This routine should
985  * be safe to call while holding the lock for another page.
986  *
987  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
988  * and deadlock against the caller's locked page.
989  */
990 struct page *
991 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
992 {
993         struct page *page = find_get_page(mapping, index);
994
995         if (page) {
996                 if (trylock_page(page))
997                         return page;
998                 page_cache_release(page);
999                 return NULL;
1000         }
1001         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1002         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1003                 page_cache_release(page);
1004                 page = NULL;
1005         }
1006         return page;
1007 }
1008 EXPORT_SYMBOL(grab_cache_page_nowait);
1009
1010 /*
1011  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1012  * a _large_ part of the i/o request. Imagine the worst scenario:
1013  *
1014  *      ---R__________________________________________B__________
1015  *         ^ reading here                             ^ bad block(assume 4k)
1016  *
1017  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1018  * => failing the whole request => read(R) => read(R+1) =>
1019  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1020  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1021  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1022  *
1023  * It is going insane. Fix it by quickly scaling down the readahead size.
1024  */
1025 static void shrink_readahead_size_eio(struct file *filp,
1026                                         struct file_ra_state *ra)
1027 {
1028         ra->ra_pages /= 4;
1029 }
1030
1031 /**
1032  * do_generic_file_read - generic file read routine
1033  * @filp:       the file to read
1034  * @ppos:       current file position
1035  * @desc:       read_descriptor
1036  * @actor:      read method
1037  *
1038  * This is a generic file read routine, and uses the
1039  * mapping->a_ops->readpage() function for the actual low-level stuff.
1040  *
1041  * This is really ugly. But the goto's actually try to clarify some
1042  * of the logic when it comes to error handling etc.
1043  */
1044 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1045                 read_descriptor_t *desc, read_actor_t actor)
1046 {
1047         struct address_space *mapping = filp->f_mapping;
1048         struct inode *inode = mapping->host;
1049         struct file_ra_state *ra = &filp->f_ra;
1050         pgoff_t index;
1051         pgoff_t last_index;
1052         pgoff_t prev_index;
1053         unsigned long offset;      /* offset into pagecache page */
1054         unsigned int prev_offset;
1055         int error;
1056
1057         index = *ppos >> PAGE_CACHE_SHIFT;
1058         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1059         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1060         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1061         offset = *ppos & ~PAGE_CACHE_MASK;
1062
1063         for (;;) {
1064                 struct page *page;
1065                 pgoff_t end_index;
1066                 loff_t isize;
1067                 unsigned long nr, ret;
1068
1069                 cond_resched();
1070 find_page:
1071                 page = find_get_page(mapping, index);
1072                 if (!page) {
1073                         page_cache_sync_readahead(mapping,
1074                                         ra, filp,
1075                                         index, last_index - index);
1076                         page = find_get_page(mapping, index);
1077                         if (unlikely(page == NULL))
1078                                 goto no_cached_page;
1079                 }
1080                 if (PageReadahead(page)) {
1081                         page_cache_async_readahead(mapping,
1082                                         ra, filp, page,
1083                                         index, last_index - index);
1084                 }
1085                 if (!PageUptodate(page)) {
1086                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1087                                         !mapping->a_ops->is_partially_uptodate)
1088                                 goto page_not_up_to_date;
1089                         if (!trylock_page(page))
1090                                 goto page_not_up_to_date;
1091                         /* Did it get truncated before we got the lock? */
1092                         if (!page->mapping)
1093                                 goto page_not_up_to_date_locked;
1094                         if (!mapping->a_ops->is_partially_uptodate(page,
1095                                                                 desc, offset))
1096                                 goto page_not_up_to_date_locked;
1097                         unlock_page(page);
1098                 }
1099 page_ok:
1100                 /*
1101                  * i_size must be checked after we know the page is Uptodate.
1102                  *
1103                  * Checking i_size after the check allows us to calculate
1104                  * the correct value for "nr", which means the zero-filled
1105                  * part of the page is not copied back to userspace (unless
1106                  * another truncate extends the file - this is desired though).
1107                  */
1108
1109                 isize = i_size_read(inode);
1110                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1111                 if (unlikely(!isize || index > end_index)) {
1112                         page_cache_release(page);
1113                         goto out;
1114                 }
1115
1116                 /* nr is the maximum number of bytes to copy from this page */
1117                 nr = PAGE_CACHE_SIZE;
1118                 if (index == end_index) {
1119                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1120                         if (nr <= offset) {
1121                                 page_cache_release(page);
1122                                 goto out;
1123                         }
1124                 }
1125                 nr = nr - offset;
1126
1127                 /* If users can be writing to this page using arbitrary
1128                  * virtual addresses, take care about potential aliasing
1129                  * before reading the page on the kernel side.
1130                  */
1131                 if (mapping_writably_mapped(mapping))
1132                         flush_dcache_page(page);
1133
1134                 /*
1135                  * When a sequential read accesses a page several times,
1136                  * only mark it as accessed the first time.
1137                  */
1138                 if (prev_index != index || offset != prev_offset)
1139                         mark_page_accessed(page);
1140                 prev_index = index;
1141
1142                 /*
1143                  * Ok, we have the page, and it's up-to-date, so
1144                  * now we can copy it to user space...
1145                  *
1146                  * The actor routine returns how many bytes were actually used..
1147                  * NOTE! This may not be the same as how much of a user buffer
1148                  * we filled up (we may be padding etc), so we can only update
1149                  * "pos" here (the actor routine has to update the user buffer
1150                  * pointers and the remaining count).
1151                  */
1152                 ret = actor(desc, page, offset, nr);
1153                 offset += ret;
1154                 index += offset >> PAGE_CACHE_SHIFT;
1155                 offset &= ~PAGE_CACHE_MASK;
1156                 prev_offset = offset;
1157
1158                 page_cache_release(page);
1159                 if (ret == nr && desc->count)
1160                         continue;
1161                 goto out;
1162
1163 page_not_up_to_date:
1164                 /* Get exclusive access to the page ... */
1165                 error = lock_page_killable(page);
1166                 if (unlikely(error))
1167                         goto readpage_error;
1168
1169 page_not_up_to_date_locked:
1170                 /* Did it get truncated before we got the lock? */
1171                 if (!page->mapping) {
1172                         unlock_page(page);
1173                         page_cache_release(page);
1174                         continue;
1175                 }
1176
1177                 /* Did somebody else fill it already? */
1178                 if (PageUptodate(page)) {
1179                         unlock_page(page);
1180                         goto page_ok;
1181                 }
1182
1183 readpage:
1184                 /*
1185                  * A previous I/O error may have been due to temporary
1186                  * failures, eg. multipath errors.
1187                  * PG_error will be set again if readpage fails.
1188                  */
1189                 ClearPageError(page);
1190                 /* Start the actual read. The read will unlock the page. */
1191                 error = mapping->a_ops->readpage(filp, page);
1192
1193                 if (unlikely(error)) {
1194                         if (error == AOP_TRUNCATED_PAGE) {
1195                                 page_cache_release(page);
1196                                 goto find_page;
1197                         }
1198                         goto readpage_error;
1199                 }
1200
1201                 if (!PageUptodate(page)) {
1202                         error = lock_page_killable(page);
1203                         if (unlikely(error))
1204                                 goto readpage_error;
1205                         if (!PageUptodate(page)) {
1206                                 if (page->mapping == NULL) {
1207                                         /*
1208                                          * invalidate_mapping_pages got it
1209                                          */
1210                                         unlock_page(page);
1211                                         page_cache_release(page);
1212                                         goto find_page;
1213                                 }
1214                                 unlock_page(page);
1215                                 shrink_readahead_size_eio(filp, ra);
1216                                 error = -EIO;
1217                                 goto readpage_error;
1218                         }
1219                         unlock_page(page);
1220                 }
1221
1222                 goto page_ok;
1223
1224 readpage_error:
1225                 /* UHHUH! A synchronous read error occurred. Report it */
1226                 desc->error = error;
1227                 page_cache_release(page);
1228                 goto out;
1229
1230 no_cached_page:
1231                 /*
1232                  * Ok, it wasn't cached, so we need to create a new
1233                  * page..
1234                  */
1235                 page = page_cache_alloc_cold(mapping);
1236                 if (!page) {
1237                         desc->error = -ENOMEM;
1238                         goto out;
1239                 }
1240                 error = add_to_page_cache_lru(page, mapping,
1241                                                 index, GFP_KERNEL);
1242                 if (error) {
1243                         page_cache_release(page);
1244                         if (error == -EEXIST)
1245                                 goto find_page;
1246                         desc->error = error;
1247                         goto out;
1248                 }
1249                 goto readpage;
1250         }
1251
1252 out:
1253         ra->prev_pos = prev_index;
1254         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1255         ra->prev_pos |= prev_offset;
1256
1257         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1258         file_accessed(filp);
1259 }
1260
1261 int file_read_actor(read_descriptor_t *desc, struct page *page,
1262                         unsigned long offset, unsigned long size)
1263 {
1264         char *kaddr;
1265         unsigned long left, count = desc->count;
1266
1267         if (size > count)
1268                 size = count;
1269
1270         /*
1271          * Faults on the destination of a read are common, so do it before
1272          * taking the kmap.
1273          */
1274         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1275                 kaddr = kmap_atomic(page, KM_USER0);
1276                 left = __copy_to_user_inatomic(desc->arg.buf,
1277                                                 kaddr + offset, size);
1278                 kunmap_atomic(kaddr, KM_USER0);
1279                 if (left == 0)
1280                         goto success;
1281         }
1282
1283         /* Do it the slow way */
1284         kaddr = kmap(page);
1285         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1286         kunmap(page);
1287
1288         if (left) {
1289                 size -= left;
1290                 desc->error = -EFAULT;
1291         }
1292 success:
1293         desc->count = count - size;
1294         desc->written += size;
1295         desc->arg.buf += size;
1296         return size;
1297 }
1298
1299 /*
1300  * Performs necessary checks before doing a write
1301  * @iov:        io vector request
1302  * @nr_segs:    number of segments in the iovec
1303  * @count:      number of bytes to write
1304  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1305  *
1306  * Adjust number of segments and amount of bytes to write (nr_segs should be
1307  * properly initialized first). Returns appropriate error code that caller
1308  * should return or zero in case that write should be allowed.
1309  */
1310 int generic_segment_checks(const struct iovec *iov,
1311                         unsigned long *nr_segs, size_t *count, int access_flags)
1312 {
1313         unsigned long   seg;
1314         size_t cnt = 0;
1315         for (seg = 0; seg < *nr_segs; seg++) {
1316                 const struct iovec *iv = &iov[seg];
1317
1318                 /*
1319                  * If any segment has a negative length, or the cumulative
1320                  * length ever wraps negative then return -EINVAL.
1321                  */
1322                 cnt += iv->iov_len;
1323                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1324                         return -EINVAL;
1325                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1326                         continue;
1327                 if (seg == 0)
1328                         return -EFAULT;
1329                 *nr_segs = seg;
1330                 cnt -= iv->iov_len;     /* This segment is no good */
1331                 break;
1332         }
1333         *count = cnt;
1334         return 0;
1335 }
1336 EXPORT_SYMBOL(generic_segment_checks);
1337
1338 /**
1339  * generic_file_aio_read - generic filesystem read routine
1340  * @iocb:       kernel I/O control block
1341  * @iov:        io vector request
1342  * @nr_segs:    number of segments in the iovec
1343  * @pos:        current file position
1344  *
1345  * This is the "read()" routine for all filesystems
1346  * that can use the page cache directly.
1347  */
1348 ssize_t
1349 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1350                 unsigned long nr_segs, loff_t pos)
1351 {
1352         struct file *filp = iocb->ki_filp;
1353         ssize_t retval;
1354         unsigned long seg = 0;
1355         size_t count;
1356         loff_t *ppos = &iocb->ki_pos;
1357         struct blk_plug plug;
1358
1359         count = 0;
1360         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1361         if (retval)
1362                 return retval;
1363
1364         blk_start_plug(&plug);
1365
1366         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1367         if (filp->f_flags & O_DIRECT) {
1368                 loff_t size;
1369                 struct address_space *mapping;
1370                 struct inode *inode;
1371
1372                 mapping = filp->f_mapping;
1373                 inode = mapping->host;
1374                 if (!count)
1375                         goto out; /* skip atime */
1376                 size = i_size_read(inode);
1377                 if (pos < size) {
1378                         retval = filemap_write_and_wait_range(mapping, pos,
1379                                         pos + iov_length(iov, nr_segs) - 1);
1380                         if (!retval) {
1381                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1382                                                         iov, pos, nr_segs);
1383                         }
1384                         if (retval > 0) {
1385                                 *ppos = pos + retval;
1386                                 count -= retval;
1387                         }
1388
1389                         /*
1390                          * Btrfs can have a short DIO read if we encounter
1391                          * compressed extents, so if there was an error, or if
1392                          * we've already read everything we wanted to, or if
1393                          * there was a short read because we hit EOF, go ahead
1394                          * and return.  Otherwise fallthrough to buffered io for
1395                          * the rest of the read.
1396                          */
1397                         if (retval < 0 || !count || *ppos >= size) {
1398                                 file_accessed(filp);
1399                                 goto out;
1400                         }
1401                 }
1402         }
1403
1404         count = retval;
1405         for (seg = 0; seg < nr_segs; seg++) {
1406                 read_descriptor_t desc;
1407                 loff_t offset = 0;
1408
1409                 /*
1410                  * If we did a short DIO read we need to skip the section of the
1411                  * iov that we've already read data into.
1412                  */
1413                 if (count) {
1414                         if (count > iov[seg].iov_len) {
1415                                 count -= iov[seg].iov_len;
1416                                 continue;
1417                         }
1418                         offset = count;
1419                         count = 0;
1420                 }
1421
1422                 desc.written = 0;
1423                 desc.arg.buf = iov[seg].iov_base + offset;
1424                 desc.count = iov[seg].iov_len - offset;
1425                 if (desc.count == 0)
1426                         continue;
1427                 desc.error = 0;
1428                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1429                 retval += desc.written;
1430                 if (desc.error) {
1431                         retval = retval ?: desc.error;
1432                         break;
1433                 }
1434                 if (desc.count > 0)
1435                         break;
1436         }
1437 out:
1438         blk_finish_plug(&plug);
1439         return retval;
1440 }
1441 EXPORT_SYMBOL(generic_file_aio_read);
1442
1443 static ssize_t
1444 do_readahead(struct address_space *mapping, struct file *filp,
1445              pgoff_t index, unsigned long nr)
1446 {
1447         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1448                 return -EINVAL;
1449
1450         force_page_cache_readahead(mapping, filp, index, nr);
1451         return 0;
1452 }
1453
1454 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1455 {
1456         ssize_t ret;
1457         struct file *file;
1458
1459         ret = -EBADF;
1460         file = fget(fd);
1461         if (file) {
1462                 if (file->f_mode & FMODE_READ) {
1463                         struct address_space *mapping = file->f_mapping;
1464                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1465                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1466                         unsigned long len = end - start + 1;
1467                         ret = do_readahead(mapping, file, start, len);
1468                 }
1469                 fput(file);
1470         }
1471         return ret;
1472 }
1473 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1474 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1475 {
1476         return SYSC_readahead((int) fd, offset, (size_t) count);
1477 }
1478 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1479 #endif
1480
1481 #ifdef CONFIG_MMU
1482 /**
1483  * page_cache_read - adds requested page to the page cache if not already there
1484  * @file:       file to read
1485  * @offset:     page index
1486  *
1487  * This adds the requested page to the page cache if it isn't already there,
1488  * and schedules an I/O to read in its contents from disk.
1489  */
1490 static int page_cache_read(struct file *file, pgoff_t offset)
1491 {
1492         struct address_space *mapping = file->f_mapping;
1493         struct page *page; 
1494         int ret;
1495
1496         do {
1497                 page = page_cache_alloc_cold(mapping);
1498                 if (!page)
1499                         return -ENOMEM;
1500
1501                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1502                 if (ret == 0)
1503                         ret = mapping->a_ops->readpage(file, page);
1504                 else if (ret == -EEXIST)
1505                         ret = 0; /* losing race to add is OK */
1506
1507                 page_cache_release(page);
1508
1509         } while (ret == AOP_TRUNCATED_PAGE);
1510                 
1511         return ret;
1512 }
1513
1514 #define MMAP_LOTSAMISS  (100)
1515
1516 /*
1517  * Synchronous readahead happens when we don't even find
1518  * a page in the page cache at all.
1519  */
1520 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1521                                    struct file_ra_state *ra,
1522                                    struct file *file,
1523                                    pgoff_t offset)
1524 {
1525         unsigned long ra_pages;
1526         struct address_space *mapping = file->f_mapping;
1527
1528         /* If we don't want any read-ahead, don't bother */
1529         if (VM_RandomReadHint(vma))
1530                 return;
1531
1532         if (VM_SequentialReadHint(vma) ||
1533                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1534                 page_cache_sync_readahead(mapping, ra, file, offset,
1535                                           ra->ra_pages);
1536                 return;
1537         }
1538
1539         if (ra->mmap_miss < INT_MAX)
1540                 ra->mmap_miss++;
1541
1542         /*
1543          * Do we miss much more than hit in this file? If so,
1544          * stop bothering with read-ahead. It will only hurt.
1545          */
1546         if (ra->mmap_miss > MMAP_LOTSAMISS)
1547                 return;
1548
1549         /*
1550          * mmap read-around
1551          */
1552         ra_pages = max_sane_readahead(ra->ra_pages);
1553         if (ra_pages) {
1554                 ra->start = max_t(long, 0, offset - ra_pages/2);
1555                 ra->size = ra_pages;
1556                 ra->async_size = 0;
1557                 ra_submit(ra, mapping, file);
1558         }
1559 }
1560
1561 /*
1562  * Asynchronous readahead happens when we find the page and PG_readahead,
1563  * so we want to possibly extend the readahead further..
1564  */
1565 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1566                                     struct file_ra_state *ra,
1567                                     struct file *file,
1568                                     struct page *page,
1569                                     pgoff_t offset)
1570 {
1571         struct address_space *mapping = file->f_mapping;
1572
1573         /* If we don't want any read-ahead, don't bother */
1574         if (VM_RandomReadHint(vma))
1575                 return;
1576         if (ra->mmap_miss > 0)
1577                 ra->mmap_miss--;
1578         if (PageReadahead(page))
1579                 page_cache_async_readahead(mapping, ra, file,
1580                                            page, offset, ra->ra_pages);
1581 }
1582
1583 /**
1584  * filemap_fault - read in file data for page fault handling
1585  * @vma:        vma in which the fault was taken
1586  * @vmf:        struct vm_fault containing details of the fault
1587  *
1588  * filemap_fault() is invoked via the vma operations vector for a
1589  * mapped memory region to read in file data during a page fault.
1590  *
1591  * The goto's are kind of ugly, but this streamlines the normal case of having
1592  * it in the page cache, and handles the special cases reasonably without
1593  * having a lot of duplicated code.
1594  */
1595 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1596 {
1597         int error;
1598         struct file *file = vma->vm_file;
1599         struct address_space *mapping = file->f_mapping;
1600         struct file_ra_state *ra = &file->f_ra;
1601         struct inode *inode = mapping->host;
1602         pgoff_t offset = vmf->pgoff;
1603         struct page *page;
1604         pgoff_t size;
1605         int ret = 0;
1606
1607         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1608         if (offset >= size)
1609                 return VM_FAULT_SIGBUS;
1610
1611         /*
1612          * Do we have something in the page cache already?
1613          */
1614         page = find_get_page(mapping, offset);
1615         if (likely(page)) {
1616                 /*
1617                  * We found the page, so try async readahead before
1618                  * waiting for the lock.
1619                  */
1620                 do_async_mmap_readahead(vma, ra, file, page, offset);
1621         } else {
1622                 /* No page in the page cache at all */
1623                 do_sync_mmap_readahead(vma, ra, file, offset);
1624                 count_vm_event(PGMAJFAULT);
1625                 ret = VM_FAULT_MAJOR;
1626 retry_find:
1627                 page = find_get_page(mapping, offset);
1628                 if (!page)
1629                         goto no_cached_page;
1630         }
1631
1632         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1633                 page_cache_release(page);
1634                 return ret | VM_FAULT_RETRY;
1635         }
1636
1637         /* Did it get truncated? */
1638         if (unlikely(page->mapping != mapping)) {
1639                 unlock_page(page);
1640                 put_page(page);
1641                 goto retry_find;
1642         }
1643         VM_BUG_ON(page->index != offset);
1644
1645         /*
1646          * We have a locked page in the page cache, now we need to check
1647          * that it's up-to-date. If not, it is going to be due to an error.
1648          */
1649         if (unlikely(!PageUptodate(page)))
1650                 goto page_not_uptodate;
1651
1652         /*
1653          * Found the page and have a reference on it.
1654          * We must recheck i_size under page lock.
1655          */
1656         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1657         if (unlikely(offset >= size)) {
1658                 unlock_page(page);
1659                 page_cache_release(page);
1660                 return VM_FAULT_SIGBUS;
1661         }
1662
1663         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1664         vmf->page = page;
1665         return ret | VM_FAULT_LOCKED;
1666
1667 no_cached_page:
1668         /*
1669          * We're only likely to ever get here if MADV_RANDOM is in
1670          * effect.
1671          */
1672         error = page_cache_read(file, offset);
1673
1674         /*
1675          * The page we want has now been added to the page cache.
1676          * In the unlikely event that someone removed it in the
1677          * meantime, we'll just come back here and read it again.
1678          */
1679         if (error >= 0)
1680                 goto retry_find;
1681
1682         /*
1683          * An error return from page_cache_read can result if the
1684          * system is low on memory, or a problem occurs while trying
1685          * to schedule I/O.
1686          */
1687         if (error == -ENOMEM)
1688                 return VM_FAULT_OOM;
1689         return VM_FAULT_SIGBUS;
1690
1691 page_not_uptodate:
1692         /*
1693          * Umm, take care of errors if the page isn't up-to-date.
1694          * Try to re-read it _once_. We do this synchronously,
1695          * because there really aren't any performance issues here
1696          * and we need to check for errors.
1697          */
1698         ClearPageError(page);
1699         error = mapping->a_ops->readpage(file, page);
1700         if (!error) {
1701                 wait_on_page_locked(page);
1702                 if (!PageUptodate(page))
1703                         error = -EIO;
1704         }
1705         page_cache_release(page);
1706
1707         if (!error || error == AOP_TRUNCATED_PAGE)
1708                 goto retry_find;
1709
1710         /* Things didn't work out. Return zero to tell the mm layer so. */
1711         shrink_readahead_size_eio(file, ra);
1712         return VM_FAULT_SIGBUS;
1713 }
1714 EXPORT_SYMBOL(filemap_fault);
1715
1716 const struct vm_operations_struct generic_file_vm_ops = {
1717         .fault          = filemap_fault,
1718 };
1719
1720 /* This is used for a general mmap of a disk file */
1721
1722 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1723 {
1724         struct address_space *mapping = file->f_mapping;
1725
1726         if (!mapping->a_ops->readpage)
1727                 return -ENOEXEC;
1728         file_accessed(file);
1729         vma->vm_ops = &generic_file_vm_ops;
1730         vma->vm_flags |= VM_CAN_NONLINEAR;
1731         return 0;
1732 }
1733
1734 /*
1735  * This is for filesystems which do not implement ->writepage.
1736  */
1737 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1738 {
1739         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1740                 return -EINVAL;
1741         return generic_file_mmap(file, vma);
1742 }
1743 #else
1744 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1745 {
1746         return -ENOSYS;
1747 }
1748 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1749 {
1750         return -ENOSYS;
1751 }
1752 #endif /* CONFIG_MMU */
1753
1754 EXPORT_SYMBOL(generic_file_mmap);
1755 EXPORT_SYMBOL(generic_file_readonly_mmap);
1756
1757 static struct page *__read_cache_page(struct address_space *mapping,
1758                                 pgoff_t index,
1759                                 int (*filler)(void *,struct page*),
1760                                 void *data,
1761                                 gfp_t gfp)
1762 {
1763         struct page *page;
1764         int err;
1765 repeat:
1766         page = find_get_page(mapping, index);
1767         if (!page) {
1768                 page = __page_cache_alloc(gfp | __GFP_COLD);
1769                 if (!page)
1770                         return ERR_PTR(-ENOMEM);
1771                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1772                 if (unlikely(err)) {
1773                         page_cache_release(page);
1774                         if (err == -EEXIST)
1775                                 goto repeat;
1776                         /* Presumably ENOMEM for radix tree node */
1777                         return ERR_PTR(err);
1778                 }
1779                 err = filler(data, page);
1780                 if (err < 0) {
1781                         page_cache_release(page);
1782                         page = ERR_PTR(err);
1783                 }
1784         }
1785         return page;
1786 }
1787
1788 static struct page *do_read_cache_page(struct address_space *mapping,
1789                                 pgoff_t index,
1790                                 int (*filler)(void *,struct page*),
1791                                 void *data,
1792                                 gfp_t gfp)
1793
1794 {
1795         struct page *page;
1796         int err;
1797
1798 retry:
1799         page = __read_cache_page(mapping, index, filler, data, gfp);
1800         if (IS_ERR(page))
1801                 return page;
1802         if (PageUptodate(page))
1803                 goto out;
1804
1805         lock_page(page);
1806         if (!page->mapping) {
1807                 unlock_page(page);
1808                 page_cache_release(page);
1809                 goto retry;
1810         }
1811         if (PageUptodate(page)) {
1812                 unlock_page(page);
1813                 goto out;
1814         }
1815         err = filler(data, page);
1816         if (err < 0) {
1817                 page_cache_release(page);
1818                 return ERR_PTR(err);
1819         }
1820 out:
1821         mark_page_accessed(page);
1822         return page;
1823 }
1824
1825 /**
1826  * read_cache_page_async - read into page cache, fill it if needed
1827  * @mapping:    the page's address_space
1828  * @index:      the page index
1829  * @filler:     function to perform the read
1830  * @data:       destination for read data
1831  *
1832  * Same as read_cache_page, but don't wait for page to become unlocked
1833  * after submitting it to the filler.
1834  *
1835  * Read into the page cache. If a page already exists, and PageUptodate() is
1836  * not set, try to fill the page but don't wait for it to become unlocked.
1837  *
1838  * If the page does not get brought uptodate, return -EIO.
1839  */
1840 struct page *read_cache_page_async(struct address_space *mapping,
1841                                 pgoff_t index,
1842                                 int (*filler)(void *,struct page*),
1843                                 void *data)
1844 {
1845         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1846 }
1847 EXPORT_SYMBOL(read_cache_page_async);
1848
1849 static struct page *wait_on_page_read(struct page *page)
1850 {
1851         if (!IS_ERR(page)) {
1852                 wait_on_page_locked(page);
1853                 if (!PageUptodate(page)) {
1854                         page_cache_release(page);
1855                         page = ERR_PTR(-EIO);
1856                 }
1857         }
1858         return page;
1859 }
1860
1861 /**
1862  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1863  * @mapping:    the page's address_space
1864  * @index:      the page index
1865  * @gfp:        the page allocator flags to use if allocating
1866  *
1867  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1868  * any new page allocations done using the specified allocation flags. Note
1869  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1870  * expect to do this atomically or anything like that - but you can pass in
1871  * other page requirements.
1872  *
1873  * If the page does not get brought uptodate, return -EIO.
1874  */
1875 struct page *read_cache_page_gfp(struct address_space *mapping,
1876                                 pgoff_t index,
1877                                 gfp_t gfp)
1878 {
1879         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1880
1881         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1882 }
1883 EXPORT_SYMBOL(read_cache_page_gfp);
1884
1885 /**
1886  * read_cache_page - read into page cache, fill it if needed
1887  * @mapping:    the page's address_space
1888  * @index:      the page index
1889  * @filler:     function to perform the read
1890  * @data:       destination for read data
1891  *
1892  * Read into the page cache. If a page already exists, and PageUptodate() is
1893  * not set, try to fill the page then wait for it to become unlocked.
1894  *
1895  * If the page does not get brought uptodate, return -EIO.
1896  */
1897 struct page *read_cache_page(struct address_space *mapping,
1898                                 pgoff_t index,
1899                                 int (*filler)(void *,struct page*),
1900                                 void *data)
1901 {
1902         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1903 }
1904 EXPORT_SYMBOL(read_cache_page);
1905
1906 /*
1907  * The logic we want is
1908  *
1909  *      if suid or (sgid and xgrp)
1910  *              remove privs
1911  */
1912 int should_remove_suid(struct dentry *dentry)
1913 {
1914         mode_t mode = dentry->d_inode->i_mode;
1915         int kill = 0;
1916
1917         /* suid always must be killed */
1918         if (unlikely(mode & S_ISUID))
1919                 kill = ATTR_KILL_SUID;
1920
1921         /*
1922          * sgid without any exec bits is just a mandatory locking mark; leave
1923          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1924          */
1925         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1926                 kill |= ATTR_KILL_SGID;
1927
1928         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1929                 return kill;
1930
1931         return 0;
1932 }
1933 EXPORT_SYMBOL(should_remove_suid);
1934
1935 static int __remove_suid(struct dentry *dentry, int kill)
1936 {
1937         struct iattr newattrs;
1938
1939         newattrs.ia_valid = ATTR_FORCE | kill;
1940         return notify_change(dentry, &newattrs);
1941 }
1942
1943 int file_remove_suid(struct file *file)
1944 {
1945         struct dentry *dentry = file->f_path.dentry;
1946         int killsuid = should_remove_suid(dentry);
1947         int killpriv = security_inode_need_killpriv(dentry);
1948         int error = 0;
1949
1950         if (killpriv < 0)
1951                 return killpriv;
1952         if (killpriv)
1953                 error = security_inode_killpriv(dentry);
1954         if (!error && killsuid)
1955                 error = __remove_suid(dentry, killsuid);
1956
1957         return error;
1958 }
1959 EXPORT_SYMBOL(file_remove_suid);
1960
1961 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1962                         const struct iovec *iov, size_t base, size_t bytes)
1963 {
1964         size_t copied = 0, left = 0;
1965
1966         while (bytes) {
1967                 char __user *buf = iov->iov_base + base;
1968                 int copy = min(bytes, iov->iov_len - base);
1969
1970                 base = 0;
1971                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1972                 copied += copy;
1973                 bytes -= copy;
1974                 vaddr += copy;
1975                 iov++;
1976
1977                 if (unlikely(left))
1978                         break;
1979         }
1980         return copied - left;
1981 }
1982
1983 /*
1984  * Copy as much as we can into the page and return the number of bytes which
1985  * were successfully copied.  If a fault is encountered then return the number of
1986  * bytes which were copied.
1987  */
1988 size_t iov_iter_copy_from_user_atomic(struct page *page,
1989                 struct iov_iter *i, unsigned long offset, size_t bytes)
1990 {
1991         char *kaddr;
1992         size_t copied;
1993
1994         BUG_ON(!in_atomic());
1995         kaddr = kmap_atomic(page, KM_USER0);
1996         if (likely(i->nr_segs == 1)) {
1997                 int left;
1998                 char __user *buf = i->iov->iov_base + i->iov_offset;
1999                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2000                 copied = bytes - left;
2001         } else {
2002                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2003                                                 i->iov, i->iov_offset, bytes);
2004         }
2005         kunmap_atomic(kaddr, KM_USER0);
2006
2007         return copied;
2008 }
2009 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2010
2011 /*
2012  * This has the same sideeffects and return value as
2013  * iov_iter_copy_from_user_atomic().
2014  * The difference is that it attempts to resolve faults.
2015  * Page must not be locked.
2016  */
2017 size_t iov_iter_copy_from_user(struct page *page,
2018                 struct iov_iter *i, unsigned long offset, size_t bytes)
2019 {
2020         char *kaddr;
2021         size_t copied;
2022
2023         kaddr = kmap(page);
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(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(page);
2034         return copied;
2035 }
2036 EXPORT_SYMBOL(iov_iter_copy_from_user);
2037
2038 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2039 {
2040         BUG_ON(i->count < bytes);
2041
2042         if (likely(i->nr_segs == 1)) {
2043                 i->iov_offset += bytes;
2044                 i->count -= bytes;
2045         } else {
2046                 const struct iovec *iov = i->iov;
2047                 size_t base = i->iov_offset;
2048
2049                 /*
2050                  * The !iov->iov_len check ensures we skip over unlikely
2051                  * zero-length segments (without overruning the iovec).
2052                  */
2053                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2054                         int copy;
2055
2056                         copy = min(bytes, iov->iov_len - base);
2057                         BUG_ON(!i->count || i->count < copy);
2058                         i->count -= copy;
2059                         bytes -= copy;
2060                         base += copy;
2061                         if (iov->iov_len == base) {
2062                                 iov++;
2063                                 base = 0;
2064                         }
2065                 }
2066                 i->iov = iov;
2067                 i->iov_offset = base;
2068         }
2069 }
2070 EXPORT_SYMBOL(iov_iter_advance);
2071
2072 /*
2073  * Fault in the first iovec of the given iov_iter, to a maximum length
2074  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2075  * accessed (ie. because it is an invalid address).
2076  *
2077  * writev-intensive code may want this to prefault several iovecs -- that
2078  * would be possible (callers must not rely on the fact that _only_ the
2079  * first iovec will be faulted with the current implementation).
2080  */
2081 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2082 {
2083         char __user *buf = i->iov->iov_base + i->iov_offset;
2084         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2085         return fault_in_pages_readable(buf, bytes);
2086 }
2087 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2088
2089 /*
2090  * Return the count of just the current iov_iter segment.
2091  */
2092 size_t iov_iter_single_seg_count(struct iov_iter *i)
2093 {
2094         const struct iovec *iov = i->iov;
2095         if (i->nr_segs == 1)
2096                 return i->count;
2097         else
2098                 return min(i->count, iov->iov_len - i->iov_offset);
2099 }
2100 EXPORT_SYMBOL(iov_iter_single_seg_count);
2101
2102 /*
2103  * Performs necessary checks before doing a write
2104  *
2105  * Can adjust writing position or amount of bytes to write.
2106  * Returns appropriate error code that caller should return or
2107  * zero in case that write should be allowed.
2108  */
2109 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2110 {
2111         struct inode *inode = file->f_mapping->host;
2112         unsigned long limit = rlimit(RLIMIT_FSIZE);
2113
2114         if (unlikely(*pos < 0))
2115                 return -EINVAL;
2116
2117         if (!isblk) {
2118                 /* FIXME: this is for backwards compatibility with 2.4 */
2119                 if (file->f_flags & O_APPEND)
2120                         *pos = i_size_read(inode);
2121
2122                 if (limit != RLIM_INFINITY) {
2123                         if (*pos >= limit) {
2124                                 send_sig(SIGXFSZ, current, 0);
2125                                 return -EFBIG;
2126                         }
2127                         if (*count > limit - (typeof(limit))*pos) {
2128                                 *count = limit - (typeof(limit))*pos;
2129                         }
2130                 }
2131         }
2132
2133         /*
2134          * LFS rule
2135          */
2136         if (unlikely(*pos + *count > MAX_NON_LFS &&
2137                                 !(file->f_flags & O_LARGEFILE))) {
2138                 if (*pos >= MAX_NON_LFS) {
2139                         return -EFBIG;
2140                 }
2141                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2142                         *count = MAX_NON_LFS - (unsigned long)*pos;
2143                 }
2144         }
2145
2146         /*
2147          * Are we about to exceed the fs block limit ?
2148          *
2149          * If we have written data it becomes a short write.  If we have
2150          * exceeded without writing data we send a signal and return EFBIG.
2151          * Linus frestrict idea will clean these up nicely..
2152          */
2153         if (likely(!isblk)) {
2154                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2155                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2156                                 return -EFBIG;
2157                         }
2158                         /* zero-length writes at ->s_maxbytes are OK */
2159                 }
2160
2161                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2162                         *count = inode->i_sb->s_maxbytes - *pos;
2163         } else {
2164 #ifdef CONFIG_BLOCK
2165                 loff_t isize;
2166                 if (bdev_read_only(I_BDEV(inode)))
2167                         return -EPERM;
2168                 isize = i_size_read(inode);
2169                 if (*pos >= isize) {
2170                         if (*count || *pos > isize)
2171                                 return -ENOSPC;
2172                 }
2173
2174                 if (*pos + *count > isize)
2175                         *count = isize - *pos;
2176 #else
2177                 return -EPERM;
2178 #endif
2179         }
2180         return 0;
2181 }
2182 EXPORT_SYMBOL(generic_write_checks);
2183
2184 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2185                                 loff_t pos, unsigned len, unsigned flags,
2186                                 struct page **pagep, void **fsdata)
2187 {
2188         const struct address_space_operations *aops = mapping->a_ops;
2189
2190         return aops->write_begin(file, mapping, pos, len, flags,
2191                                                         pagep, fsdata);
2192 }
2193 EXPORT_SYMBOL(pagecache_write_begin);
2194
2195 int pagecache_write_end(struct file *file, struct address_space *mapping,
2196                                 loff_t pos, unsigned len, unsigned copied,
2197                                 struct page *page, void *fsdata)
2198 {
2199         const struct address_space_operations *aops = mapping->a_ops;
2200
2201         mark_page_accessed(page);
2202         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2203 }
2204 EXPORT_SYMBOL(pagecache_write_end);
2205
2206 ssize_t
2207 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2208                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2209                 size_t count, size_t ocount)
2210 {
2211         struct file     *file = iocb->ki_filp;
2212         struct address_space *mapping = file->f_mapping;
2213         struct inode    *inode = mapping->host;
2214         ssize_t         written;
2215         size_t          write_len;
2216         pgoff_t         end;
2217
2218         if (count != ocount)
2219                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2220
2221         write_len = iov_length(iov, *nr_segs);
2222         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2223
2224         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2225         if (written)
2226                 goto out;
2227
2228         /*
2229          * After a write we want buffered reads to be sure to go to disk to get
2230          * the new data.  We invalidate clean cached page from the region we're
2231          * about to write.  We do this *before* the write so that we can return
2232          * without clobbering -EIOCBQUEUED from ->direct_IO().
2233          */
2234         if (mapping->nrpages) {
2235                 written = invalidate_inode_pages2_range(mapping,
2236                                         pos >> PAGE_CACHE_SHIFT, end);
2237                 /*
2238                  * If a page can not be invalidated, return 0 to fall back
2239                  * to buffered write.
2240                  */
2241                 if (written) {
2242                         if (written == -EBUSY)
2243                                 return 0;
2244                         goto out;
2245                 }
2246         }
2247
2248         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2249
2250         /*
2251          * Finally, try again to invalidate clean pages which might have been
2252          * cached by non-direct readahead, or faulted in by get_user_pages()
2253          * if the source of the write was an mmap'ed region of the file
2254          * we're writing.  Either one is a pretty crazy thing to do,
2255          * so we don't support it 100%.  If this invalidation
2256          * fails, tough, the write still worked...
2257          */
2258         if (mapping->nrpages) {
2259                 invalidate_inode_pages2_range(mapping,
2260                                               pos >> PAGE_CACHE_SHIFT, end);
2261         }
2262
2263         if (written > 0) {
2264                 pos += written;
2265                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2266                         i_size_write(inode, pos);
2267                         mark_inode_dirty(inode);
2268                 }
2269                 *ppos = pos;
2270         }
2271 out:
2272         return written;
2273 }
2274 EXPORT_SYMBOL(generic_file_direct_write);
2275
2276 /*
2277  * Find or create a page at the given pagecache position. Return the locked
2278  * page. This function is specifically for buffered writes.
2279  */
2280 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2281                                         pgoff_t index, unsigned flags)
2282 {
2283         int status;
2284         struct page *page;
2285         gfp_t gfp_notmask = 0;
2286         if (flags & AOP_FLAG_NOFS)
2287                 gfp_notmask = __GFP_FS;
2288 repeat:
2289         page = find_lock_page(mapping, index);
2290         if (page)
2291                 return page;
2292
2293         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2294         if (!page)
2295                 return NULL;
2296         status = add_to_page_cache_lru(page, mapping, index,
2297                                                 GFP_KERNEL & ~gfp_notmask);
2298         if (unlikely(status)) {
2299                 page_cache_release(page);
2300                 if (status == -EEXIST)
2301                         goto repeat;
2302                 return NULL;
2303         }
2304         return page;
2305 }
2306 EXPORT_SYMBOL(grab_cache_page_write_begin);
2307
2308 static ssize_t generic_perform_write(struct file *file,
2309                                 struct iov_iter *i, loff_t pos)
2310 {
2311         struct address_space *mapping = file->f_mapping;
2312         const struct address_space_operations *a_ops = mapping->a_ops;
2313         long status = 0;
2314         ssize_t written = 0;
2315         unsigned int flags = 0;
2316
2317         /*
2318          * Copies from kernel address space cannot fail (NFSD is a big user).
2319          */
2320         if (segment_eq(get_fs(), KERNEL_DS))
2321                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2322
2323         do {
2324                 struct page *page;
2325                 unsigned long offset;   /* Offset into pagecache page */
2326                 unsigned long bytes;    /* Bytes to write to page */
2327                 size_t copied;          /* Bytes copied from user */
2328                 void *fsdata;
2329
2330                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2331                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2332                                                 iov_iter_count(i));
2333
2334 again:
2335
2336                 /*
2337                  * Bring in the user page that we will copy from _first_.
2338                  * Otherwise there's a nasty deadlock on copying from the
2339                  * same page as we're writing to, without it being marked
2340                  * up-to-date.
2341                  *
2342                  * Not only is this an optimisation, but it is also required
2343                  * to check that the address is actually valid, when atomic
2344                  * usercopies are used, below.
2345                  */
2346                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2347                         status = -EFAULT;
2348                         break;
2349                 }
2350
2351                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2352                                                 &page, &fsdata);
2353                 if (unlikely(status))
2354                         break;
2355
2356                 if (mapping_writably_mapped(mapping))
2357                         flush_dcache_page(page);
2358
2359                 pagefault_disable();
2360                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2361                 pagefault_enable();
2362                 flush_dcache_page(page);
2363
2364                 mark_page_accessed(page);
2365                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2366                                                 page, fsdata);
2367                 if (unlikely(status < 0))
2368                         break;
2369                 copied = status;
2370
2371                 cond_resched();
2372
2373                 iov_iter_advance(i, copied);
2374                 if (unlikely(copied == 0)) {
2375                         /*
2376                          * If we were unable to copy any data at all, we must
2377                          * fall back to a single segment length write.
2378                          *
2379                          * If we didn't fallback here, we could livelock
2380                          * because not all segments in the iov can be copied at
2381                          * once without a pagefault.
2382                          */
2383                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2384                                                 iov_iter_single_seg_count(i));
2385                         goto again;
2386                 }
2387                 pos += copied;
2388                 written += copied;
2389
2390                 balance_dirty_pages_ratelimited(mapping);
2391
2392         } while (iov_iter_count(i));
2393
2394         return written ? written : status;
2395 }
2396
2397 ssize_t
2398 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2399                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2400                 size_t count, ssize_t written)
2401 {
2402         struct file *file = iocb->ki_filp;
2403         ssize_t status;
2404         struct iov_iter i;
2405
2406         iov_iter_init(&i, iov, nr_segs, count, written);
2407         status = generic_perform_write(file, &i, pos);
2408
2409         if (likely(status >= 0)) {
2410                 written += status;
2411                 *ppos = pos + status;
2412         }
2413         
2414         return written ? written : status;
2415 }
2416 EXPORT_SYMBOL(generic_file_buffered_write);
2417
2418 /**
2419  * __generic_file_aio_write - write data to a file
2420  * @iocb:       IO state structure (file, offset, etc.)
2421  * @iov:        vector with data to write
2422  * @nr_segs:    number of segments in the vector
2423  * @ppos:       position where to write
2424  *
2425  * This function does all the work needed for actually writing data to a
2426  * file. It does all basic checks, removes SUID from the file, updates
2427  * modification times and calls proper subroutines depending on whether we
2428  * do direct IO or a standard buffered write.
2429  *
2430  * It expects i_mutex to be grabbed unless we work on a block device or similar
2431  * object which does not need locking at all.
2432  *
2433  * This function does *not* take care of syncing data in case of O_SYNC write.
2434  * A caller has to handle it. This is mainly due to the fact that we want to
2435  * avoid syncing under i_mutex.
2436  */
2437 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2438                                  unsigned long nr_segs, loff_t *ppos)
2439 {
2440         struct file *file = iocb->ki_filp;
2441         struct address_space * mapping = file->f_mapping;
2442         size_t ocount;          /* original count */
2443         size_t count;           /* after file limit checks */
2444         struct inode    *inode = mapping->host;
2445         loff_t          pos;
2446         ssize_t         written;
2447         ssize_t         err;
2448
2449         ocount = 0;
2450         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2451         if (err)
2452                 return err;
2453
2454         count = ocount;
2455         pos = *ppos;
2456
2457         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2458
2459         /* We can write back this queue in page reclaim */
2460         current->backing_dev_info = mapping->backing_dev_info;
2461         written = 0;
2462
2463         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2464         if (err)
2465                 goto out;
2466
2467         if (count == 0)
2468                 goto out;
2469
2470         err = file_remove_suid(file);
2471         if (err)
2472                 goto out;
2473
2474         file_update_time(file);
2475
2476         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2477         if (unlikely(file->f_flags & O_DIRECT)) {
2478                 loff_t endbyte;
2479                 ssize_t written_buffered;
2480
2481                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2482                                                         ppos, count, ocount);
2483                 if (written < 0 || written == count)
2484                         goto out;
2485                 /*
2486                  * direct-io write to a hole: fall through to buffered I/O
2487                  * for completing the rest of the request.
2488                  */
2489                 pos += written;
2490                 count -= written;
2491                 written_buffered = generic_file_buffered_write(iocb, iov,
2492                                                 nr_segs, pos, ppos, count,
2493                                                 written);
2494                 /*
2495                  * If generic_file_buffered_write() retuned a synchronous error
2496                  * then we want to return the number of bytes which were
2497                  * direct-written, or the error code if that was zero.  Note
2498                  * that this differs from normal direct-io semantics, which
2499                  * will return -EFOO even if some bytes were written.
2500                  */
2501                 if (written_buffered < 0) {
2502                         err = written_buffered;
2503                         goto out;
2504                 }
2505
2506                 /*
2507                  * We need to ensure that the page cache pages are written to
2508                  * disk and invalidated to preserve the expected O_DIRECT
2509                  * semantics.
2510                  */
2511                 endbyte = pos + written_buffered - written - 1;
2512                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2513                 if (err == 0) {
2514                         written = written_buffered;
2515                         invalidate_mapping_pages(mapping,
2516                                                  pos >> PAGE_CACHE_SHIFT,
2517                                                  endbyte >> PAGE_CACHE_SHIFT);
2518                 } else {
2519                         /*
2520                          * We don't know how much we wrote, so just return
2521                          * the number of bytes which were direct-written
2522                          */
2523                 }
2524         } else {
2525                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2526                                 pos, ppos, count, written);
2527         }
2528 out:
2529         current->backing_dev_info = NULL;
2530         return written ? written : err;
2531 }
2532 EXPORT_SYMBOL(__generic_file_aio_write);
2533
2534 /**
2535  * generic_file_aio_write - write data to a file
2536  * @iocb:       IO state structure
2537  * @iov:        vector with data to write
2538  * @nr_segs:    number of segments in the vector
2539  * @pos:        position in file where to write
2540  *
2541  * This is a wrapper around __generic_file_aio_write() to be used by most
2542  * filesystems. It takes care of syncing the file in case of O_SYNC file
2543  * and acquires i_mutex as needed.
2544  */
2545 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2546                 unsigned long nr_segs, loff_t pos)
2547 {
2548         struct file *file = iocb->ki_filp;
2549         struct inode *inode = file->f_mapping->host;
2550         struct blk_plug plug;
2551         ssize_t ret;
2552
2553         BUG_ON(iocb->ki_pos != pos);
2554
2555         mutex_lock(&inode->i_mutex);
2556         blk_start_plug(&plug);
2557         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2558         mutex_unlock(&inode->i_mutex);
2559
2560         if (ret > 0 || ret == -EIOCBQUEUED) {
2561                 ssize_t err;
2562
2563                 err = generic_write_sync(file, pos, ret);
2564                 if (err < 0 && ret > 0)
2565                         ret = err;
2566         }
2567         blk_finish_plug(&plug);
2568         return ret;
2569 }
2570 EXPORT_SYMBOL(generic_file_aio_write);
2571
2572 /**
2573  * try_to_release_page() - release old fs-specific metadata on a page
2574  *
2575  * @page: the page which the kernel is trying to free
2576  * @gfp_mask: memory allocation flags (and I/O mode)
2577  *
2578  * The address_space is to try to release any data against the page
2579  * (presumably at page->private).  If the release was successful, return `1'.
2580  * Otherwise return zero.
2581  *
2582  * This may also be called if PG_fscache is set on a page, indicating that the
2583  * page is known to the local caching routines.
2584  *
2585  * The @gfp_mask argument specifies whether I/O may be performed to release
2586  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2587  *
2588  */
2589 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2590 {
2591         struct address_space * const mapping = page->mapping;
2592
2593         BUG_ON(!PageLocked(page));
2594         if (PageWriteback(page))
2595                 return 0;
2596
2597         if (mapping && mapping->a_ops->releasepage)
2598                 return mapping->a_ops->releasepage(page, gfp_mask);
2599         return try_to_free_buffers(page);
2600 }
2601
2602 EXPORT_SYMBOL(try_to_release_page);