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