KVM: x86 emulator: Add cmp al, imm and cmp ax, imm instructions (ocodes 3c, 3d)
[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/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.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 "internal.h"
37
38 /*
39  * FIXME: remove all knowledge of the buffer layer from the core VM
40  */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
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               (vmtruncate)
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
109 /*
110  * Remove a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold the mapping's tree_lock.
113  */
114 void __remove_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         mem_cgroup_uncharge_cache_page(page);
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124
125         /*
126          * Some filesystems seem to re-dirty the page even after
127          * the VM has canceled the dirty bit (eg ext3 journaling).
128          *
129          * Fix it up by doing a final dirty accounting check after
130          * having removed the page entirely.
131          */
132         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
133                 dec_zone_page_state(page, NR_FILE_DIRTY);
134                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
135         }
136 }
137
138 void remove_from_page_cache(struct page *page)
139 {
140         struct address_space *mapping = page->mapping;
141
142         BUG_ON(!PageLocked(page));
143
144         spin_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         spin_unlock_irq(&mapping->tree_lock);
147 }
148
149 static int sync_page(void *word)
150 {
151         struct address_space *mapping;
152         struct page *page;
153
154         page = container_of((unsigned long *)word, struct page, flags);
155
156         /*
157          * page_mapping() is being called without PG_locked held.
158          * Some knowledge of the state and use of the page is used to
159          * reduce the requirements down to a memory barrier.
160          * The danger here is of a stale page_mapping() return value
161          * indicating a struct address_space different from the one it's
162          * associated with when it is associated with one.
163          * After smp_mb(), it's either the correct page_mapping() for
164          * the page, or an old page_mapping() and the page's own
165          * page_mapping() has gone NULL.
166          * The ->sync_page() address_space operation must tolerate
167          * page_mapping() going NULL. By an amazing coincidence,
168          * this comes about because none of the users of the page
169          * in the ->sync_page() methods make essential use of the
170          * page_mapping(), merely passing the page down to the backing
171          * device's unplug functions when it's non-NULL, which in turn
172          * ignore it for all cases but swap, where only page_private(page) is
173          * of interest. When page_mapping() does go NULL, the entire
174          * call stack gracefully ignores the page and returns.
175          * -- wli
176          */
177         smp_mb();
178         mapping = page_mapping(page);
179         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
180                 mapping->a_ops->sync_page(page);
181         io_schedule();
182         return 0;
183 }
184
185 static int sync_page_killable(void *word)
186 {
187         sync_page(word);
188         return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 /**
192  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
193  * @mapping:    address space structure to write
194  * @start:      offset in bytes where the range starts
195  * @end:        offset in bytes where the range ends (inclusive)
196  * @sync_mode:  enable synchronous operation
197  *
198  * Start writeback against all of a mapping's dirty pages that lie
199  * within the byte offsets <start, end> inclusive.
200  *
201  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
202  * opposed to a regular memory cleansing writeback.  The difference between
203  * these two operations is that if a dirty page/buffer is encountered, it must
204  * be waited upon, and not just skipped over.
205  */
206 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
207                                 loff_t end, int sync_mode)
208 {
209         int ret;
210         struct writeback_control wbc = {
211                 .sync_mode = sync_mode,
212                 .nr_to_write = mapping->nrpages * 2,
213                 .range_start = start,
214                 .range_end = end,
215         };
216
217         if (!mapping_cap_writeback_dirty(mapping))
218                 return 0;
219
220         ret = do_writepages(mapping, &wbc);
221         return ret;
222 }
223
224 static inline int __filemap_fdatawrite(struct address_space *mapping,
225         int sync_mode)
226 {
227         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
228 }
229
230 int filemap_fdatawrite(struct address_space *mapping)
231 {
232         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
233 }
234 EXPORT_SYMBOL(filemap_fdatawrite);
235
236 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
237                                 loff_t end)
238 {
239         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite_range);
242
243 /**
244  * filemap_flush - mostly a non-blocking flush
245  * @mapping:    target address_space
246  *
247  * This is a mostly non-blocking flush.  Not suitable for data-integrity
248  * purposes - I/O may not be started against all dirty pages.
249  */
250 int filemap_flush(struct address_space *mapping)
251 {
252         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
253 }
254 EXPORT_SYMBOL(filemap_flush);
255
256 /**
257  * wait_on_page_writeback_range - wait for writeback to complete
258  * @mapping:    target address_space
259  * @start:      beginning page index
260  * @end:        ending page index
261  *
262  * Wait for writeback to complete against pages indexed by start->end
263  * inclusive
264  */
265 int wait_on_page_writeback_range(struct address_space *mapping,
266                                 pgoff_t start, pgoff_t end)
267 {
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271         pgoff_t index;
272
273         if (end < start)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         index = start;
278         while ((index <= end) &&
279                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
280                         PAGECACHE_TAG_WRITEBACK,
281                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
282                 unsigned i;
283
284                 for (i = 0; i < nr_pages; i++) {
285                         struct page *page = pvec.pages[i];
286
287                         /* until radix tree lookup accepts end_index */
288                         if (page->index > end)
289                                 continue;
290
291                         wait_on_page_writeback(page);
292                         if (PageError(page))
293                                 ret = -EIO;
294                 }
295                 pagevec_release(&pvec);
296                 cond_resched();
297         }
298
299         /* Check for outstanding write errors */
300         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
301                 ret = -ENOSPC;
302         if (test_and_clear_bit(AS_EIO, &mapping->flags))
303                 ret = -EIO;
304
305         return ret;
306 }
307
308 /**
309  * sync_page_range - write and wait on all pages in the passed range
310  * @inode:      target inode
311  * @mapping:    target address_space
312  * @pos:        beginning offset in pages to write
313  * @count:      number of bytes to write
314  *
315  * Write and wait upon all the pages in the passed range.  This is a "data
316  * integrity" operation.  It waits upon in-flight writeout before starting and
317  * waiting upon new writeout.  If there was an IO error, return it.
318  *
319  * We need to re-take i_mutex during the generic_osync_inode list walk because
320  * it is otherwise livelockable.
321  */
322 int sync_page_range(struct inode *inode, struct address_space *mapping,
323                         loff_t pos, loff_t count)
324 {
325         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
326         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
327         int ret;
328
329         if (!mapping_cap_writeback_dirty(mapping) || !count)
330                 return 0;
331         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
332         if (ret == 0) {
333                 mutex_lock(&inode->i_mutex);
334                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
335                 mutex_unlock(&inode->i_mutex);
336         }
337         if (ret == 0)
338                 ret = wait_on_page_writeback_range(mapping, start, end);
339         return ret;
340 }
341 EXPORT_SYMBOL(sync_page_range);
342
343 /**
344  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
345  * @inode:      target inode
346  * @mapping:    target address_space
347  * @pos:        beginning offset in pages to write
348  * @count:      number of bytes to write
349  *
350  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
351  * as it forces O_SYNC writers to different parts of the same file
352  * to be serialised right until io completion.
353  */
354 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
355                            loff_t pos, loff_t count)
356 {
357         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
358         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
359         int ret;
360
361         if (!mapping_cap_writeback_dirty(mapping) || !count)
362                 return 0;
363         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
364         if (ret == 0)
365                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
366         if (ret == 0)
367                 ret = wait_on_page_writeback_range(mapping, start, end);
368         return ret;
369 }
370 EXPORT_SYMBOL(sync_page_range_nolock);
371
372 /**
373  * filemap_fdatawait - wait for all under-writeback pages to complete
374  * @mapping: address space structure to wait for
375  *
376  * Walk the list of under-writeback pages of the given address space
377  * and wait for all of them.
378  */
379 int filemap_fdatawait(struct address_space *mapping)
380 {
381         loff_t i_size = i_size_read(mapping->host);
382
383         if (i_size == 0)
384                 return 0;
385
386         return wait_on_page_writeback_range(mapping, 0,
387                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
388 }
389 EXPORT_SYMBOL(filemap_fdatawait);
390
391 int filemap_write_and_wait(struct address_space *mapping)
392 {
393         int err = 0;
394
395         if (mapping->nrpages) {
396                 err = filemap_fdatawrite(mapping);
397                 /*
398                  * Even if the above returned error, the pages may be
399                  * written partially (e.g. -ENOSPC), so we wait for it.
400                  * But the -EIO is special case, it may indicate the worst
401                  * thing (e.g. bug) happened, so we avoid waiting for it.
402                  */
403                 if (err != -EIO) {
404                         int err2 = filemap_fdatawait(mapping);
405                         if (!err)
406                                 err = err2;
407                 }
408         }
409         return err;
410 }
411 EXPORT_SYMBOL(filemap_write_and_wait);
412
413 /**
414  * filemap_write_and_wait_range - write out & wait on a file range
415  * @mapping:    the address_space for the pages
416  * @lstart:     offset in bytes where the range starts
417  * @lend:       offset in bytes where the range ends (inclusive)
418  *
419  * Write out and wait upon file offsets lstart->lend, inclusive.
420  *
421  * Note that `lend' is inclusive (describes the last byte to be written) so
422  * that this function can be used to write to the very end-of-file (end = -1).
423  */
424 int filemap_write_and_wait_range(struct address_space *mapping,
425                                  loff_t lstart, loff_t lend)
426 {
427         int err = 0;
428
429         if (mapping->nrpages) {
430                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
431                                                  WB_SYNC_ALL);
432                 /* See comment of filemap_write_and_wait() */
433                 if (err != -EIO) {
434                         int err2 = wait_on_page_writeback_range(mapping,
435                                                 lstart >> PAGE_CACHE_SHIFT,
436                                                 lend >> PAGE_CACHE_SHIFT);
437                         if (!err)
438                                 err = err2;
439                 }
440         }
441         return err;
442 }
443
444 /**
445  * add_to_page_cache_locked - add a locked page to the pagecache
446  * @page:       page to add
447  * @mapping:    the page's address_space
448  * @offset:     page index
449  * @gfp_mask:   page allocation mode
450  *
451  * This function is used to add a page to the pagecache. It must be locked.
452  * This function does not add the page to the LRU.  The caller must do that.
453  */
454 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
455                 pgoff_t offset, gfp_t gfp_mask)
456 {
457         int error;
458
459         VM_BUG_ON(!PageLocked(page));
460
461         error = mem_cgroup_cache_charge(page, current->mm,
462                                         gfp_mask & ~__GFP_HIGHMEM);
463         if (error)
464                 goto out;
465
466         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
467         if (error == 0) {
468                 page_cache_get(page);
469                 page->mapping = mapping;
470                 page->index = offset;
471
472                 spin_lock_irq(&mapping->tree_lock);
473                 error = radix_tree_insert(&mapping->page_tree, offset, page);
474                 if (likely(!error)) {
475                         mapping->nrpages++;
476                         __inc_zone_page_state(page, NR_FILE_PAGES);
477                 } else {
478                         page->mapping = NULL;
479                         mem_cgroup_uncharge_cache_page(page);
480                         page_cache_release(page);
481                 }
482
483                 spin_unlock_irq(&mapping->tree_lock);
484                 radix_tree_preload_end();
485         } else
486                 mem_cgroup_uncharge_cache_page(page);
487 out:
488         return error;
489 }
490 EXPORT_SYMBOL(add_to_page_cache_locked);
491
492 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
493                                 pgoff_t offset, gfp_t gfp_mask)
494 {
495         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
496         if (ret == 0)
497                 lru_cache_add(page);
498         return ret;
499 }
500
501 #ifdef CONFIG_NUMA
502 struct page *__page_cache_alloc(gfp_t gfp)
503 {
504         if (cpuset_do_page_mem_spread()) {
505                 int n = cpuset_mem_spread_node();
506                 return alloc_pages_node(n, gfp, 0);
507         }
508         return alloc_pages(gfp, 0);
509 }
510 EXPORT_SYMBOL(__page_cache_alloc);
511 #endif
512
513 static int __sleep_on_page_lock(void *word)
514 {
515         io_schedule();
516         return 0;
517 }
518
519 /*
520  * In order to wait for pages to become available there must be
521  * waitqueues associated with pages. By using a hash table of
522  * waitqueues where the bucket discipline is to maintain all
523  * waiters on the same queue and wake all when any of the pages
524  * become available, and for the woken contexts to check to be
525  * sure the appropriate page became available, this saves space
526  * at a cost of "thundering herd" phenomena during rare hash
527  * collisions.
528  */
529 static wait_queue_head_t *page_waitqueue(struct page *page)
530 {
531         const struct zone *zone = page_zone(page);
532
533         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
534 }
535
536 static inline void wake_up_page(struct page *page, int bit)
537 {
538         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
539 }
540
541 void wait_on_page_bit(struct page *page, int bit_nr)
542 {
543         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
544
545         if (test_bit(bit_nr, &page->flags))
546                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
547                                                         TASK_UNINTERRUPTIBLE);
548 }
549 EXPORT_SYMBOL(wait_on_page_bit);
550
551 /**
552  * unlock_page - unlock a locked page
553  * @page: the page
554  *
555  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
556  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
557  * mechananism between PageLocked pages and PageWriteback pages is shared.
558  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
559  *
560  * The first mb is necessary to safely close the critical section opened by the
561  * test_and_set_bit() to lock the page; the second mb is necessary to enforce
562  * ordering between the clear_bit and the read of the waitqueue (to avoid SMP
563  * races with a parallel wait_on_page_locked()).
564  */
565 void unlock_page(struct page *page)
566 {
567         smp_mb__before_clear_bit();
568         if (!test_and_clear_bit(PG_locked, &page->flags))
569                 BUG();
570         smp_mb__after_clear_bit(); 
571         wake_up_page(page, PG_locked);
572 }
573 EXPORT_SYMBOL(unlock_page);
574
575 /**
576  * end_page_writeback - end writeback against a page
577  * @page: the page
578  */
579 void end_page_writeback(struct page *page)
580 {
581         if (TestClearPageReclaim(page))
582                 rotate_reclaimable_page(page);
583
584         if (!test_clear_page_writeback(page))
585                 BUG();
586
587         smp_mb__after_clear_bit();
588         wake_up_page(page, PG_writeback);
589 }
590 EXPORT_SYMBOL(end_page_writeback);
591
592 /**
593  * __lock_page - get a lock on the page, assuming we need to sleep to get it
594  * @page: the page to lock
595  *
596  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
597  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
598  * chances are that on the second loop, the block layer's plug list is empty,
599  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
600  */
601 void __lock_page(struct page *page)
602 {
603         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
604
605         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
606                                                         TASK_UNINTERRUPTIBLE);
607 }
608 EXPORT_SYMBOL(__lock_page);
609
610 int __lock_page_killable(struct page *page)
611 {
612         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
613
614         return __wait_on_bit_lock(page_waitqueue(page), &wait,
615                                         sync_page_killable, TASK_KILLABLE);
616 }
617
618 /**
619  * __lock_page_nosync - get a lock on the page, without calling sync_page()
620  * @page: the page to lock
621  *
622  * Variant of lock_page that does not require the caller to hold a reference
623  * on the page's mapping.
624  */
625 void __lock_page_nosync(struct page *page)
626 {
627         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
628         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
629                                                         TASK_UNINTERRUPTIBLE);
630 }
631
632 /**
633  * find_get_page - find and get a page reference
634  * @mapping: the address_space to search
635  * @offset: the page index
636  *
637  * Is there a pagecache struct page at the given (mapping, offset) tuple?
638  * If yes, increment its refcount and return it; if no, return NULL.
639  */
640 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
641 {
642         void **pagep;
643         struct page *page;
644
645         rcu_read_lock();
646 repeat:
647         page = NULL;
648         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
649         if (pagep) {
650                 page = radix_tree_deref_slot(pagep);
651                 if (unlikely(!page || page == RADIX_TREE_RETRY))
652                         goto repeat;
653
654                 if (!page_cache_get_speculative(page))
655                         goto repeat;
656
657                 /*
658                  * Has the page moved?
659                  * This is part of the lockless pagecache protocol. See
660                  * include/linux/pagemap.h for details.
661                  */
662                 if (unlikely(page != *pagep)) {
663                         page_cache_release(page);
664                         goto repeat;
665                 }
666         }
667         rcu_read_unlock();
668
669         return page;
670 }
671 EXPORT_SYMBOL(find_get_page);
672
673 /**
674  * find_lock_page - locate, pin and lock a pagecache page
675  * @mapping: the address_space to search
676  * @offset: the page index
677  *
678  * Locates the desired pagecache page, locks it, increments its reference
679  * count and returns its address.
680  *
681  * Returns zero if the page was not present. find_lock_page() may sleep.
682  */
683 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
684 {
685         struct page *page;
686
687 repeat:
688         page = find_get_page(mapping, offset);
689         if (page) {
690                 lock_page(page);
691                 /* Has the page been truncated? */
692                 if (unlikely(page->mapping != mapping)) {
693                         unlock_page(page);
694                         page_cache_release(page);
695                         goto repeat;
696                 }
697                 VM_BUG_ON(page->index != offset);
698         }
699         return page;
700 }
701 EXPORT_SYMBOL(find_lock_page);
702
703 /**
704  * find_or_create_page - locate or add a pagecache page
705  * @mapping: the page's address_space
706  * @index: the page's index into the mapping
707  * @gfp_mask: page allocation mode
708  *
709  * Locates a page in the pagecache.  If the page is not present, a new page
710  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
711  * LRU list.  The returned page is locked and has its reference count
712  * incremented.
713  *
714  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
715  * allocation!
716  *
717  * find_or_create_page() returns the desired page's address, or zero on
718  * memory exhaustion.
719  */
720 struct page *find_or_create_page(struct address_space *mapping,
721                 pgoff_t index, gfp_t gfp_mask)
722 {
723         struct page *page;
724         int err;
725 repeat:
726         page = find_lock_page(mapping, index);
727         if (!page) {
728                 page = __page_cache_alloc(gfp_mask);
729                 if (!page)
730                         return NULL;
731                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
732                 if (unlikely(err)) {
733                         page_cache_release(page);
734                         page = NULL;
735                         if (err == -EEXIST)
736                                 goto repeat;
737                 }
738         }
739         return page;
740 }
741 EXPORT_SYMBOL(find_or_create_page);
742
743 /**
744  * find_get_pages - gang pagecache lookup
745  * @mapping:    The address_space to search
746  * @start:      The starting page index
747  * @nr_pages:   The maximum number of pages
748  * @pages:      Where the resulting pages are placed
749  *
750  * find_get_pages() will search for and return a group of up to
751  * @nr_pages pages in the mapping.  The pages are placed at @pages.
752  * find_get_pages() takes a reference against the returned pages.
753  *
754  * The search returns a group of mapping-contiguous pages with ascending
755  * indexes.  There may be holes in the indices due to not-present pages.
756  *
757  * find_get_pages() returns the number of pages which were found.
758  */
759 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
760                             unsigned int nr_pages, struct page **pages)
761 {
762         unsigned int i;
763         unsigned int ret;
764         unsigned int nr_found;
765
766         rcu_read_lock();
767 restart:
768         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
769                                 (void ***)pages, start, nr_pages);
770         ret = 0;
771         for (i = 0; i < nr_found; i++) {
772                 struct page *page;
773 repeat:
774                 page = radix_tree_deref_slot((void **)pages[i]);
775                 if (unlikely(!page))
776                         continue;
777                 /*
778                  * this can only trigger if nr_found == 1, making livelock
779                  * a non issue.
780                  */
781                 if (unlikely(page == RADIX_TREE_RETRY))
782                         goto restart;
783
784                 if (!page_cache_get_speculative(page))
785                         goto repeat;
786
787                 /* Has the page moved? */
788                 if (unlikely(page != *((void **)pages[i]))) {
789                         page_cache_release(page);
790                         goto repeat;
791                 }
792
793                 pages[ret] = page;
794                 ret++;
795         }
796         rcu_read_unlock();
797         return ret;
798 }
799
800 /**
801  * find_get_pages_contig - gang contiguous pagecache lookup
802  * @mapping:    The address_space to search
803  * @index:      The starting page index
804  * @nr_pages:   The maximum number of pages
805  * @pages:      Where the resulting pages are placed
806  *
807  * find_get_pages_contig() works exactly like find_get_pages(), except
808  * that the returned number of pages are guaranteed to be contiguous.
809  *
810  * find_get_pages_contig() returns the number of pages which were found.
811  */
812 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
813                                unsigned int nr_pages, struct page **pages)
814 {
815         unsigned int i;
816         unsigned int ret;
817         unsigned int nr_found;
818
819         rcu_read_lock();
820 restart:
821         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
822                                 (void ***)pages, index, nr_pages);
823         ret = 0;
824         for (i = 0; i < nr_found; i++) {
825                 struct page *page;
826 repeat:
827                 page = radix_tree_deref_slot((void **)pages[i]);
828                 if (unlikely(!page))
829                         continue;
830                 /*
831                  * this can only trigger if nr_found == 1, making livelock
832                  * a non issue.
833                  */
834                 if (unlikely(page == RADIX_TREE_RETRY))
835                         goto restart;
836
837                 if (page->mapping == NULL || page->index != index)
838                         break;
839
840                 if (!page_cache_get_speculative(page))
841                         goto repeat;
842
843                 /* Has the page moved? */
844                 if (unlikely(page != *((void **)pages[i]))) {
845                         page_cache_release(page);
846                         goto repeat;
847                 }
848
849                 pages[ret] = page;
850                 ret++;
851                 index++;
852         }
853         rcu_read_unlock();
854         return ret;
855 }
856 EXPORT_SYMBOL(find_get_pages_contig);
857
858 /**
859  * find_get_pages_tag - find and return pages that match @tag
860  * @mapping:    the address_space to search
861  * @index:      the starting page index
862  * @tag:        the tag index
863  * @nr_pages:   the maximum number of pages
864  * @pages:      where the resulting pages are placed
865  *
866  * Like find_get_pages, except we only return pages which are tagged with
867  * @tag.   We update @index to index the next page for the traversal.
868  */
869 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
870                         int tag, unsigned int nr_pages, struct page **pages)
871 {
872         unsigned int i;
873         unsigned int ret;
874         unsigned int nr_found;
875
876         rcu_read_lock();
877 restart:
878         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
879                                 (void ***)pages, *index, nr_pages, tag);
880         ret = 0;
881         for (i = 0; i < nr_found; i++) {
882                 struct page *page;
883 repeat:
884                 page = radix_tree_deref_slot((void **)pages[i]);
885                 if (unlikely(!page))
886                         continue;
887                 /*
888                  * this can only trigger if nr_found == 1, making livelock
889                  * a non issue.
890                  */
891                 if (unlikely(page == RADIX_TREE_RETRY))
892                         goto restart;
893
894                 if (!page_cache_get_speculative(page))
895                         goto repeat;
896
897                 /* Has the page moved? */
898                 if (unlikely(page != *((void **)pages[i]))) {
899                         page_cache_release(page);
900                         goto repeat;
901                 }
902
903                 pages[ret] = page;
904                 ret++;
905         }
906         rcu_read_unlock();
907
908         if (ret)
909                 *index = pages[ret - 1]->index + 1;
910
911         return ret;
912 }
913 EXPORT_SYMBOL(find_get_pages_tag);
914
915 /**
916  * grab_cache_page_nowait - returns locked page at given index in given cache
917  * @mapping: target address_space
918  * @index: the page index
919  *
920  * Same as grab_cache_page(), but do not wait if the page is unavailable.
921  * This is intended for speculative data generators, where the data can
922  * be regenerated if the page couldn't be grabbed.  This routine should
923  * be safe to call while holding the lock for another page.
924  *
925  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
926  * and deadlock against the caller's locked page.
927  */
928 struct page *
929 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
930 {
931         struct page *page = find_get_page(mapping, index);
932
933         if (page) {
934                 if (trylock_page(page))
935                         return page;
936                 page_cache_release(page);
937                 return NULL;
938         }
939         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
940         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
941                 page_cache_release(page);
942                 page = NULL;
943         }
944         return page;
945 }
946 EXPORT_SYMBOL(grab_cache_page_nowait);
947
948 /*
949  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
950  * a _large_ part of the i/o request. Imagine the worst scenario:
951  *
952  *      ---R__________________________________________B__________
953  *         ^ reading here                             ^ bad block(assume 4k)
954  *
955  * read(R) => miss => readahead(R...B) => media error => frustrating retries
956  * => failing the whole request => read(R) => read(R+1) =>
957  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
958  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
959  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
960  *
961  * It is going insane. Fix it by quickly scaling down the readahead size.
962  */
963 static void shrink_readahead_size_eio(struct file *filp,
964                                         struct file_ra_state *ra)
965 {
966         if (!ra->ra_pages)
967                 return;
968
969         ra->ra_pages /= 4;
970 }
971
972 /**
973  * do_generic_file_read - generic file read routine
974  * @filp:       the file to read
975  * @ppos:       current file position
976  * @desc:       read_descriptor
977  * @actor:      read method
978  *
979  * This is a generic file read routine, and uses the
980  * mapping->a_ops->readpage() function for the actual low-level stuff.
981  *
982  * This is really ugly. But the goto's actually try to clarify some
983  * of the logic when it comes to error handling etc.
984  */
985 static void do_generic_file_read(struct file *filp, loff_t *ppos,
986                 read_descriptor_t *desc, read_actor_t actor)
987 {
988         struct address_space *mapping = filp->f_mapping;
989         struct inode *inode = mapping->host;
990         struct file_ra_state *ra = &filp->f_ra;
991         pgoff_t index;
992         pgoff_t last_index;
993         pgoff_t prev_index;
994         unsigned long offset;      /* offset into pagecache page */
995         unsigned int prev_offset;
996         int error;
997
998         index = *ppos >> PAGE_CACHE_SHIFT;
999         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1000         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1001         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1002         offset = *ppos & ~PAGE_CACHE_MASK;
1003
1004         for (;;) {
1005                 struct page *page;
1006                 pgoff_t end_index;
1007                 loff_t isize;
1008                 unsigned long nr, ret;
1009
1010                 cond_resched();
1011 find_page:
1012                 page = find_get_page(mapping, index);
1013                 if (!page) {
1014                         page_cache_sync_readahead(mapping,
1015                                         ra, filp,
1016                                         index, last_index - index);
1017                         page = find_get_page(mapping, index);
1018                         if (unlikely(page == NULL))
1019                                 goto no_cached_page;
1020                 }
1021                 if (PageReadahead(page)) {
1022                         page_cache_async_readahead(mapping,
1023                                         ra, filp, page,
1024                                         index, last_index - index);
1025                 }
1026                 if (!PageUptodate(page)) {
1027                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1028                                         !mapping->a_ops->is_partially_uptodate)
1029                                 goto page_not_up_to_date;
1030                         if (!trylock_page(page))
1031                                 goto page_not_up_to_date;
1032                         if (!mapping->a_ops->is_partially_uptodate(page,
1033                                                                 desc, offset))
1034                                 goto page_not_up_to_date_locked;
1035                         unlock_page(page);
1036                 }
1037 page_ok:
1038                 /*
1039                  * i_size must be checked after we know the page is Uptodate.
1040                  *
1041                  * Checking i_size after the check allows us to calculate
1042                  * the correct value for "nr", which means the zero-filled
1043                  * part of the page is not copied back to userspace (unless
1044                  * another truncate extends the file - this is desired though).
1045                  */
1046
1047                 isize = i_size_read(inode);
1048                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1049                 if (unlikely(!isize || index > end_index)) {
1050                         page_cache_release(page);
1051                         goto out;
1052                 }
1053
1054                 /* nr is the maximum number of bytes to copy from this page */
1055                 nr = PAGE_CACHE_SIZE;
1056                 if (index == end_index) {
1057                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1058                         if (nr <= offset) {
1059                                 page_cache_release(page);
1060                                 goto out;
1061                         }
1062                 }
1063                 nr = nr - offset;
1064
1065                 /* If users can be writing to this page using arbitrary
1066                  * virtual addresses, take care about potential aliasing
1067                  * before reading the page on the kernel side.
1068                  */
1069                 if (mapping_writably_mapped(mapping))
1070                         flush_dcache_page(page);
1071
1072                 /*
1073                  * When a sequential read accesses a page several times,
1074                  * only mark it as accessed the first time.
1075                  */
1076                 if (prev_index != index || offset != prev_offset)
1077                         mark_page_accessed(page);
1078                 prev_index = index;
1079
1080                 /*
1081                  * Ok, we have the page, and it's up-to-date, so
1082                  * now we can copy it to user space...
1083                  *
1084                  * The actor routine returns how many bytes were actually used..
1085                  * NOTE! This may not be the same as how much of a user buffer
1086                  * we filled up (we may be padding etc), so we can only update
1087                  * "pos" here (the actor routine has to update the user buffer
1088                  * pointers and the remaining count).
1089                  */
1090                 ret = actor(desc, page, offset, nr);
1091                 offset += ret;
1092                 index += offset >> PAGE_CACHE_SHIFT;
1093                 offset &= ~PAGE_CACHE_MASK;
1094                 prev_offset = offset;
1095
1096                 page_cache_release(page);
1097                 if (ret == nr && desc->count)
1098                         continue;
1099                 goto out;
1100
1101 page_not_up_to_date:
1102                 /* Get exclusive access to the page ... */
1103                 if (lock_page_killable(page))
1104                         goto readpage_eio;
1105
1106 page_not_up_to_date_locked:
1107                 /* Did it get truncated before we got the lock? */
1108                 if (!page->mapping) {
1109                         unlock_page(page);
1110                         page_cache_release(page);
1111                         continue;
1112                 }
1113
1114                 /* Did somebody else fill it already? */
1115                 if (PageUptodate(page)) {
1116                         unlock_page(page);
1117                         goto page_ok;
1118                 }
1119
1120 readpage:
1121                 /* Start the actual read. The read will unlock the page. */
1122                 error = mapping->a_ops->readpage(filp, page);
1123
1124                 if (unlikely(error)) {
1125                         if (error == AOP_TRUNCATED_PAGE) {
1126                                 page_cache_release(page);
1127                                 goto find_page;
1128                         }
1129                         goto readpage_error;
1130                 }
1131
1132                 if (!PageUptodate(page)) {
1133                         if (lock_page_killable(page))
1134                                 goto readpage_eio;
1135                         if (!PageUptodate(page)) {
1136                                 if (page->mapping == NULL) {
1137                                         /*
1138                                          * invalidate_inode_pages got it
1139                                          */
1140                                         unlock_page(page);
1141                                         page_cache_release(page);
1142                                         goto find_page;
1143                                 }
1144                                 unlock_page(page);
1145                                 shrink_readahead_size_eio(filp, ra);
1146                                 goto readpage_eio;
1147                         }
1148                         unlock_page(page);
1149                 }
1150
1151                 goto page_ok;
1152
1153 readpage_eio:
1154                 error = -EIO;
1155 readpage_error:
1156                 /* UHHUH! A synchronous read error occurred. Report it */
1157                 desc->error = error;
1158                 page_cache_release(page);
1159                 goto out;
1160
1161 no_cached_page:
1162                 /*
1163                  * Ok, it wasn't cached, so we need to create a new
1164                  * page..
1165                  */
1166                 page = page_cache_alloc_cold(mapping);
1167                 if (!page) {
1168                         desc->error = -ENOMEM;
1169                         goto out;
1170                 }
1171                 error = add_to_page_cache_lru(page, mapping,
1172                                                 index, GFP_KERNEL);
1173                 if (error) {
1174                         page_cache_release(page);
1175                         if (error == -EEXIST)
1176                                 goto find_page;
1177                         desc->error = error;
1178                         goto out;
1179                 }
1180                 goto readpage;
1181         }
1182
1183 out:
1184         ra->prev_pos = prev_index;
1185         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1186         ra->prev_pos |= prev_offset;
1187
1188         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1189         if (filp)
1190                 file_accessed(filp);
1191 }
1192
1193 int file_read_actor(read_descriptor_t *desc, struct page *page,
1194                         unsigned long offset, unsigned long size)
1195 {
1196         char *kaddr;
1197         unsigned long left, count = desc->count;
1198
1199         if (size > count)
1200                 size = count;
1201
1202         /*
1203          * Faults on the destination of a read are common, so do it before
1204          * taking the kmap.
1205          */
1206         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1207                 kaddr = kmap_atomic(page, KM_USER0);
1208                 left = __copy_to_user_inatomic(desc->arg.buf,
1209                                                 kaddr + offset, size);
1210                 kunmap_atomic(kaddr, KM_USER0);
1211                 if (left == 0)
1212                         goto success;
1213         }
1214
1215         /* Do it the slow way */
1216         kaddr = kmap(page);
1217         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1218         kunmap(page);
1219
1220         if (left) {
1221                 size -= left;
1222                 desc->error = -EFAULT;
1223         }
1224 success:
1225         desc->count = count - size;
1226         desc->written += size;
1227         desc->arg.buf += size;
1228         return size;
1229 }
1230
1231 /*
1232  * Performs necessary checks before doing a write
1233  * @iov:        io vector request
1234  * @nr_segs:    number of segments in the iovec
1235  * @count:      number of bytes to write
1236  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1237  *
1238  * Adjust number of segments and amount of bytes to write (nr_segs should be
1239  * properly initialized first). Returns appropriate error code that caller
1240  * should return or zero in case that write should be allowed.
1241  */
1242 int generic_segment_checks(const struct iovec *iov,
1243                         unsigned long *nr_segs, size_t *count, int access_flags)
1244 {
1245         unsigned long   seg;
1246         size_t cnt = 0;
1247         for (seg = 0; seg < *nr_segs; seg++) {
1248                 const struct iovec *iv = &iov[seg];
1249
1250                 /*
1251                  * If any segment has a negative length, or the cumulative
1252                  * length ever wraps negative then return -EINVAL.
1253                  */
1254                 cnt += iv->iov_len;
1255                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1256                         return -EINVAL;
1257                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1258                         continue;
1259                 if (seg == 0)
1260                         return -EFAULT;
1261                 *nr_segs = seg;
1262                 cnt -= iv->iov_len;     /* This segment is no good */
1263                 break;
1264         }
1265         *count = cnt;
1266         return 0;
1267 }
1268 EXPORT_SYMBOL(generic_segment_checks);
1269
1270 /**
1271  * generic_file_aio_read - generic filesystem read routine
1272  * @iocb:       kernel I/O control block
1273  * @iov:        io vector request
1274  * @nr_segs:    number of segments in the iovec
1275  * @pos:        current file position
1276  *
1277  * This is the "read()" routine for all filesystems
1278  * that can use the page cache directly.
1279  */
1280 ssize_t
1281 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1282                 unsigned long nr_segs, loff_t pos)
1283 {
1284         struct file *filp = iocb->ki_filp;
1285         ssize_t retval;
1286         unsigned long seg;
1287         size_t count;
1288         loff_t *ppos = &iocb->ki_pos;
1289
1290         count = 0;
1291         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1292         if (retval)
1293                 return retval;
1294
1295         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1296         if (filp->f_flags & O_DIRECT) {
1297                 loff_t size;
1298                 struct address_space *mapping;
1299                 struct inode *inode;
1300
1301                 mapping = filp->f_mapping;
1302                 inode = mapping->host;
1303                 if (!count)
1304                         goto out; /* skip atime */
1305                 size = i_size_read(inode);
1306                 if (pos < size) {
1307                         retval = filemap_write_and_wait(mapping);
1308                         if (!retval) {
1309                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1310                                                         iov, pos, nr_segs);
1311                         }
1312                         if (retval > 0)
1313                                 *ppos = pos + retval;
1314                         if (retval) {
1315                                 file_accessed(filp);
1316                                 goto out;
1317                         }
1318                 }
1319         }
1320
1321         for (seg = 0; seg < nr_segs; seg++) {
1322                 read_descriptor_t desc;
1323
1324                 desc.written = 0;
1325                 desc.arg.buf = iov[seg].iov_base;
1326                 desc.count = iov[seg].iov_len;
1327                 if (desc.count == 0)
1328                         continue;
1329                 desc.error = 0;
1330                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1331                 retval += desc.written;
1332                 if (desc.error) {
1333                         retval = retval ?: desc.error;
1334                         break;
1335                 }
1336                 if (desc.count > 0)
1337                         break;
1338         }
1339 out:
1340         return retval;
1341 }
1342 EXPORT_SYMBOL(generic_file_aio_read);
1343
1344 static ssize_t
1345 do_readahead(struct address_space *mapping, struct file *filp,
1346              pgoff_t index, unsigned long nr)
1347 {
1348         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1349                 return -EINVAL;
1350
1351         force_page_cache_readahead(mapping, filp, index,
1352                                         max_sane_readahead(nr));
1353         return 0;
1354 }
1355
1356 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1357 {
1358         ssize_t ret;
1359         struct file *file;
1360
1361         ret = -EBADF;
1362         file = fget(fd);
1363         if (file) {
1364                 if (file->f_mode & FMODE_READ) {
1365                         struct address_space *mapping = file->f_mapping;
1366                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1367                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1368                         unsigned long len = end - start + 1;
1369                         ret = do_readahead(mapping, file, start, len);
1370                 }
1371                 fput(file);
1372         }
1373         return ret;
1374 }
1375
1376 #ifdef CONFIG_MMU
1377 /**
1378  * page_cache_read - adds requested page to the page cache if not already there
1379  * @file:       file to read
1380  * @offset:     page index
1381  *
1382  * This adds the requested page to the page cache if it isn't already there,
1383  * and schedules an I/O to read in its contents from disk.
1384  */
1385 static int page_cache_read(struct file *file, pgoff_t offset)
1386 {
1387         struct address_space *mapping = file->f_mapping;
1388         struct page *page; 
1389         int ret;
1390
1391         do {
1392                 page = page_cache_alloc_cold(mapping);
1393                 if (!page)
1394                         return -ENOMEM;
1395
1396                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1397                 if (ret == 0)
1398                         ret = mapping->a_ops->readpage(file, page);
1399                 else if (ret == -EEXIST)
1400                         ret = 0; /* losing race to add is OK */
1401
1402                 page_cache_release(page);
1403
1404         } while (ret == AOP_TRUNCATED_PAGE);
1405                 
1406         return ret;
1407 }
1408
1409 #define MMAP_LOTSAMISS  (100)
1410
1411 /**
1412  * filemap_fault - read in file data for page fault handling
1413  * @vma:        vma in which the fault was taken
1414  * @vmf:        struct vm_fault containing details of the fault
1415  *
1416  * filemap_fault() is invoked via the vma operations vector for a
1417  * mapped memory region to read in file data during a page fault.
1418  *
1419  * The goto's are kind of ugly, but this streamlines the normal case of having
1420  * it in the page cache, and handles the special cases reasonably without
1421  * having a lot of duplicated code.
1422  */
1423 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1424 {
1425         int error;
1426         struct file *file = vma->vm_file;
1427         struct address_space *mapping = file->f_mapping;
1428         struct file_ra_state *ra = &file->f_ra;
1429         struct inode *inode = mapping->host;
1430         struct page *page;
1431         pgoff_t size;
1432         int did_readaround = 0;
1433         int ret = 0;
1434
1435         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1436         if (vmf->pgoff >= size)
1437                 return VM_FAULT_SIGBUS;
1438
1439         /* If we don't want any read-ahead, don't bother */
1440         if (VM_RandomReadHint(vma))
1441                 goto no_cached_page;
1442
1443         /*
1444          * Do we have something in the page cache already?
1445          */
1446 retry_find:
1447         page = find_lock_page(mapping, vmf->pgoff);
1448         /*
1449          * For sequential accesses, we use the generic readahead logic.
1450          */
1451         if (VM_SequentialReadHint(vma)) {
1452                 if (!page) {
1453                         page_cache_sync_readahead(mapping, ra, file,
1454                                                            vmf->pgoff, 1);
1455                         page = find_lock_page(mapping, vmf->pgoff);
1456                         if (!page)
1457                                 goto no_cached_page;
1458                 }
1459                 if (PageReadahead(page)) {
1460                         page_cache_async_readahead(mapping, ra, file, page,
1461                                                            vmf->pgoff, 1);
1462                 }
1463         }
1464
1465         if (!page) {
1466                 unsigned long ra_pages;
1467
1468                 ra->mmap_miss++;
1469
1470                 /*
1471                  * Do we miss much more than hit in this file? If so,
1472                  * stop bothering with read-ahead. It will only hurt.
1473                  */
1474                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1475                         goto no_cached_page;
1476
1477                 /*
1478                  * To keep the pgmajfault counter straight, we need to
1479                  * check did_readaround, as this is an inner loop.
1480                  */
1481                 if (!did_readaround) {
1482                         ret = VM_FAULT_MAJOR;
1483                         count_vm_event(PGMAJFAULT);
1484                 }
1485                 did_readaround = 1;
1486                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1487                 if (ra_pages) {
1488                         pgoff_t start = 0;
1489
1490                         if (vmf->pgoff > ra_pages / 2)
1491                                 start = vmf->pgoff - ra_pages / 2;
1492                         do_page_cache_readahead(mapping, file, start, ra_pages);
1493                 }
1494                 page = find_lock_page(mapping, vmf->pgoff);
1495                 if (!page)
1496                         goto no_cached_page;
1497         }
1498
1499         if (!did_readaround)
1500                 ra->mmap_miss--;
1501
1502         /*
1503          * We have a locked page in the page cache, now we need to check
1504          * that it's up-to-date. If not, it is going to be due to an error.
1505          */
1506         if (unlikely(!PageUptodate(page)))
1507                 goto page_not_uptodate;
1508
1509         /* Must recheck i_size under page lock */
1510         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1511         if (unlikely(vmf->pgoff >= size)) {
1512                 unlock_page(page);
1513                 page_cache_release(page);
1514                 return VM_FAULT_SIGBUS;
1515         }
1516
1517         /*
1518          * Found the page and have a reference on it.
1519          */
1520         mark_page_accessed(page);
1521         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1522         vmf->page = page;
1523         return ret | VM_FAULT_LOCKED;
1524
1525 no_cached_page:
1526         /*
1527          * We're only likely to ever get here if MADV_RANDOM is in
1528          * effect.
1529          */
1530         error = page_cache_read(file, vmf->pgoff);
1531
1532         /*
1533          * The page we want has now been added to the page cache.
1534          * In the unlikely event that someone removed it in the
1535          * meantime, we'll just come back here and read it again.
1536          */
1537         if (error >= 0)
1538                 goto retry_find;
1539
1540         /*
1541          * An error return from page_cache_read can result if the
1542          * system is low on memory, or a problem occurs while trying
1543          * to schedule I/O.
1544          */
1545         if (error == -ENOMEM)
1546                 return VM_FAULT_OOM;
1547         return VM_FAULT_SIGBUS;
1548
1549 page_not_uptodate:
1550         /* IO error path */
1551         if (!did_readaround) {
1552                 ret = VM_FAULT_MAJOR;
1553                 count_vm_event(PGMAJFAULT);
1554         }
1555
1556         /*
1557          * Umm, take care of errors if the page isn't up-to-date.
1558          * Try to re-read it _once_. We do this synchronously,
1559          * because there really aren't any performance issues here
1560          * and we need to check for errors.
1561          */
1562         ClearPageError(page);
1563         error = mapping->a_ops->readpage(file, page);
1564         if (!error) {
1565                 wait_on_page_locked(page);
1566                 if (!PageUptodate(page))
1567                         error = -EIO;
1568         }
1569         page_cache_release(page);
1570
1571         if (!error || error == AOP_TRUNCATED_PAGE)
1572                 goto retry_find;
1573
1574         /* Things didn't work out. Return zero to tell the mm layer so. */
1575         shrink_readahead_size_eio(file, ra);
1576         return VM_FAULT_SIGBUS;
1577 }
1578 EXPORT_SYMBOL(filemap_fault);
1579
1580 struct vm_operations_struct generic_file_vm_ops = {
1581         .fault          = filemap_fault,
1582 };
1583
1584 /* This is used for a general mmap of a disk file */
1585
1586 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1587 {
1588         struct address_space *mapping = file->f_mapping;
1589
1590         if (!mapping->a_ops->readpage)
1591                 return -ENOEXEC;
1592         file_accessed(file);
1593         vma->vm_ops = &generic_file_vm_ops;
1594         vma->vm_flags |= VM_CAN_NONLINEAR;
1595         return 0;
1596 }
1597
1598 /*
1599  * This is for filesystems which do not implement ->writepage.
1600  */
1601 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1602 {
1603         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1604                 return -EINVAL;
1605         return generic_file_mmap(file, vma);
1606 }
1607 #else
1608 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1609 {
1610         return -ENOSYS;
1611 }
1612 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1613 {
1614         return -ENOSYS;
1615 }
1616 #endif /* CONFIG_MMU */
1617
1618 EXPORT_SYMBOL(generic_file_mmap);
1619 EXPORT_SYMBOL(generic_file_readonly_mmap);
1620
1621 static struct page *__read_cache_page(struct address_space *mapping,
1622                                 pgoff_t index,
1623                                 int (*filler)(void *,struct page*),
1624                                 void *data)
1625 {
1626         struct page *page;
1627         int err;
1628 repeat:
1629         page = find_get_page(mapping, index);
1630         if (!page) {
1631                 page = page_cache_alloc_cold(mapping);
1632                 if (!page)
1633                         return ERR_PTR(-ENOMEM);
1634                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1635                 if (unlikely(err)) {
1636                         page_cache_release(page);
1637                         if (err == -EEXIST)
1638                                 goto repeat;
1639                         /* Presumably ENOMEM for radix tree node */
1640                         return ERR_PTR(err);
1641                 }
1642                 err = filler(data, page);
1643                 if (err < 0) {
1644                         page_cache_release(page);
1645                         page = ERR_PTR(err);
1646                 }
1647         }
1648         return page;
1649 }
1650
1651 /**
1652  * read_cache_page_async - read into page cache, fill it if needed
1653  * @mapping:    the page's address_space
1654  * @index:      the page index
1655  * @filler:     function to perform the read
1656  * @data:       destination for read data
1657  *
1658  * Same as read_cache_page, but don't wait for page to become unlocked
1659  * after submitting it to the filler.
1660  *
1661  * Read into the page cache. If a page already exists, and PageUptodate() is
1662  * not set, try to fill the page but don't wait for it to become unlocked.
1663  *
1664  * If the page does not get brought uptodate, return -EIO.
1665  */
1666 struct page *read_cache_page_async(struct address_space *mapping,
1667                                 pgoff_t index,
1668                                 int (*filler)(void *,struct page*),
1669                                 void *data)
1670 {
1671         struct page *page;
1672         int err;
1673
1674 retry:
1675         page = __read_cache_page(mapping, index, filler, data);
1676         if (IS_ERR(page))
1677                 return page;
1678         if (PageUptodate(page))
1679                 goto out;
1680
1681         lock_page(page);
1682         if (!page->mapping) {
1683                 unlock_page(page);
1684                 page_cache_release(page);
1685                 goto retry;
1686         }
1687         if (PageUptodate(page)) {
1688                 unlock_page(page);
1689                 goto out;
1690         }
1691         err = filler(data, page);
1692         if (err < 0) {
1693                 page_cache_release(page);
1694                 return ERR_PTR(err);
1695         }
1696 out:
1697         mark_page_accessed(page);
1698         return page;
1699 }
1700 EXPORT_SYMBOL(read_cache_page_async);
1701
1702 /**
1703  * read_cache_page - read into page cache, fill it if needed
1704  * @mapping:    the page's address_space
1705  * @index:      the page index
1706  * @filler:     function to perform the read
1707  * @data:       destination for read data
1708  *
1709  * Read into the page cache. If a page already exists, and PageUptodate() is
1710  * not set, try to fill the page then wait for it to become unlocked.
1711  *
1712  * If the page does not get brought uptodate, return -EIO.
1713  */
1714 struct page *read_cache_page(struct address_space *mapping,
1715                                 pgoff_t index,
1716                                 int (*filler)(void *,struct page*),
1717                                 void *data)
1718 {
1719         struct page *page;
1720
1721         page = read_cache_page_async(mapping, index, filler, data);
1722         if (IS_ERR(page))
1723                 goto out;
1724         wait_on_page_locked(page);
1725         if (!PageUptodate(page)) {
1726                 page_cache_release(page);
1727                 page = ERR_PTR(-EIO);
1728         }
1729  out:
1730         return page;
1731 }
1732 EXPORT_SYMBOL(read_cache_page);
1733
1734 /*
1735  * The logic we want is
1736  *
1737  *      if suid or (sgid and xgrp)
1738  *              remove privs
1739  */
1740 int should_remove_suid(struct dentry *dentry)
1741 {
1742         mode_t mode = dentry->d_inode->i_mode;
1743         int kill = 0;
1744
1745         /* suid always must be killed */
1746         if (unlikely(mode & S_ISUID))
1747                 kill = ATTR_KILL_SUID;
1748
1749         /*
1750          * sgid without any exec bits is just a mandatory locking mark; leave
1751          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1752          */
1753         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1754                 kill |= ATTR_KILL_SGID;
1755
1756         if (unlikely(kill && !capable(CAP_FSETID)))
1757                 return kill;
1758
1759         return 0;
1760 }
1761 EXPORT_SYMBOL(should_remove_suid);
1762
1763 static int __remove_suid(struct dentry *dentry, int kill)
1764 {
1765         struct iattr newattrs;
1766
1767         newattrs.ia_valid = ATTR_FORCE | kill;
1768         return notify_change(dentry, &newattrs);
1769 }
1770
1771 int file_remove_suid(struct file *file)
1772 {
1773         struct dentry *dentry = file->f_path.dentry;
1774         int killsuid = should_remove_suid(dentry);
1775         int killpriv = security_inode_need_killpriv(dentry);
1776         int error = 0;
1777
1778         if (killpriv < 0)
1779                 return killpriv;
1780         if (killpriv)
1781                 error = security_inode_killpriv(dentry);
1782         if (!error && killsuid)
1783                 error = __remove_suid(dentry, killsuid);
1784
1785         return error;
1786 }
1787 EXPORT_SYMBOL(file_remove_suid);
1788
1789 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1790                         const struct iovec *iov, size_t base, size_t bytes)
1791 {
1792         size_t copied = 0, left = 0;
1793
1794         while (bytes) {
1795                 char __user *buf = iov->iov_base + base;
1796                 int copy = min(bytes, iov->iov_len - base);
1797
1798                 base = 0;
1799                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1800                 copied += copy;
1801                 bytes -= copy;
1802                 vaddr += copy;
1803                 iov++;
1804
1805                 if (unlikely(left))
1806                         break;
1807         }
1808         return copied - left;
1809 }
1810
1811 /*
1812  * Copy as much as we can into the page and return the number of bytes which
1813  * were sucessfully copied.  If a fault is encountered then return the number of
1814  * bytes which were copied.
1815  */
1816 size_t iov_iter_copy_from_user_atomic(struct page *page,
1817                 struct iov_iter *i, unsigned long offset, size_t bytes)
1818 {
1819         char *kaddr;
1820         size_t copied;
1821
1822         BUG_ON(!in_atomic());
1823         kaddr = kmap_atomic(page, KM_USER0);
1824         if (likely(i->nr_segs == 1)) {
1825                 int left;
1826                 char __user *buf = i->iov->iov_base + i->iov_offset;
1827                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1828                                                         buf, bytes);
1829                 copied = bytes - left;
1830         } else {
1831                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1832                                                 i->iov, i->iov_offset, bytes);
1833         }
1834         kunmap_atomic(kaddr, KM_USER0);
1835
1836         return copied;
1837 }
1838 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1839
1840 /*
1841  * This has the same sideeffects and return value as
1842  * iov_iter_copy_from_user_atomic().
1843  * The difference is that it attempts to resolve faults.
1844  * Page must not be locked.
1845  */
1846 size_t iov_iter_copy_from_user(struct page *page,
1847                 struct iov_iter *i, unsigned long offset, size_t bytes)
1848 {
1849         char *kaddr;
1850         size_t copied;
1851
1852         kaddr = kmap(page);
1853         if (likely(i->nr_segs == 1)) {
1854                 int left;
1855                 char __user *buf = i->iov->iov_base + i->iov_offset;
1856                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1857                 copied = bytes - left;
1858         } else {
1859                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1860                                                 i->iov, i->iov_offset, bytes);
1861         }
1862         kunmap(page);
1863         return copied;
1864 }
1865 EXPORT_SYMBOL(iov_iter_copy_from_user);
1866
1867 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1868 {
1869         BUG_ON(i->count < bytes);
1870
1871         if (likely(i->nr_segs == 1)) {
1872                 i->iov_offset += bytes;
1873                 i->count -= bytes;
1874         } else {
1875                 const struct iovec *iov = i->iov;
1876                 size_t base = i->iov_offset;
1877
1878                 /*
1879                  * The !iov->iov_len check ensures we skip over unlikely
1880                  * zero-length segments (without overruning the iovec).
1881                  */
1882                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1883                         int copy;
1884
1885                         copy = min(bytes, iov->iov_len - base);
1886                         BUG_ON(!i->count || i->count < copy);
1887                         i->count -= copy;
1888                         bytes -= copy;
1889                         base += copy;
1890                         if (iov->iov_len == base) {
1891                                 iov++;
1892                                 base = 0;
1893                         }
1894                 }
1895                 i->iov = iov;
1896                 i->iov_offset = base;
1897         }
1898 }
1899 EXPORT_SYMBOL(iov_iter_advance);
1900
1901 /*
1902  * Fault in the first iovec of the given iov_iter, to a maximum length
1903  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1904  * accessed (ie. because it is an invalid address).
1905  *
1906  * writev-intensive code may want this to prefault several iovecs -- that
1907  * would be possible (callers must not rely on the fact that _only_ the
1908  * first iovec will be faulted with the current implementation).
1909  */
1910 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1911 {
1912         char __user *buf = i->iov->iov_base + i->iov_offset;
1913         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1914         return fault_in_pages_readable(buf, bytes);
1915 }
1916 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1917
1918 /*
1919  * Return the count of just the current iov_iter segment.
1920  */
1921 size_t iov_iter_single_seg_count(struct iov_iter *i)
1922 {
1923         const struct iovec *iov = i->iov;
1924         if (i->nr_segs == 1)
1925                 return i->count;
1926         else
1927                 return min(i->count, iov->iov_len - i->iov_offset);
1928 }
1929 EXPORT_SYMBOL(iov_iter_single_seg_count);
1930
1931 /*
1932  * Performs necessary checks before doing a write
1933  *
1934  * Can adjust writing position or amount of bytes to write.
1935  * Returns appropriate error code that caller should return or
1936  * zero in case that write should be allowed.
1937  */
1938 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1939 {
1940         struct inode *inode = file->f_mapping->host;
1941         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1942
1943         if (unlikely(*pos < 0))
1944                 return -EINVAL;
1945
1946         if (!isblk) {
1947                 /* FIXME: this is for backwards compatibility with 2.4 */
1948                 if (file->f_flags & O_APPEND)
1949                         *pos = i_size_read(inode);
1950
1951                 if (limit != RLIM_INFINITY) {
1952                         if (*pos >= limit) {
1953                                 send_sig(SIGXFSZ, current, 0);
1954                                 return -EFBIG;
1955                         }
1956                         if (*count > limit - (typeof(limit))*pos) {
1957                                 *count = limit - (typeof(limit))*pos;
1958                         }
1959                 }
1960         }
1961
1962         /*
1963          * LFS rule
1964          */
1965         if (unlikely(*pos + *count > MAX_NON_LFS &&
1966                                 !(file->f_flags & O_LARGEFILE))) {
1967                 if (*pos >= MAX_NON_LFS) {
1968                         return -EFBIG;
1969                 }
1970                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1971                         *count = MAX_NON_LFS - (unsigned long)*pos;
1972                 }
1973         }
1974
1975         /*
1976          * Are we about to exceed the fs block limit ?
1977          *
1978          * If we have written data it becomes a short write.  If we have
1979          * exceeded without writing data we send a signal and return EFBIG.
1980          * Linus frestrict idea will clean these up nicely..
1981          */
1982         if (likely(!isblk)) {
1983                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1984                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1985                                 return -EFBIG;
1986                         }
1987                         /* zero-length writes at ->s_maxbytes are OK */
1988                 }
1989
1990                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1991                         *count = inode->i_sb->s_maxbytes - *pos;
1992         } else {
1993 #ifdef CONFIG_BLOCK
1994                 loff_t isize;
1995                 if (bdev_read_only(I_BDEV(inode)))
1996                         return -EPERM;
1997                 isize = i_size_read(inode);
1998                 if (*pos >= isize) {
1999                         if (*count || *pos > isize)
2000                                 return -ENOSPC;
2001                 }
2002
2003                 if (*pos + *count > isize)
2004                         *count = isize - *pos;
2005 #else
2006                 return -EPERM;
2007 #endif
2008         }
2009         return 0;
2010 }
2011 EXPORT_SYMBOL(generic_write_checks);
2012
2013 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2014                                 loff_t pos, unsigned len, unsigned flags,
2015                                 struct page **pagep, void **fsdata)
2016 {
2017         const struct address_space_operations *aops = mapping->a_ops;
2018
2019         if (aops->write_begin) {
2020                 return aops->write_begin(file, mapping, pos, len, flags,
2021                                                         pagep, fsdata);
2022         } else {
2023                 int ret;
2024                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2025                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2026                 struct inode *inode = mapping->host;
2027                 struct page *page;
2028 again:
2029                 page = __grab_cache_page(mapping, index);
2030                 *pagep = page;
2031                 if (!page)
2032                         return -ENOMEM;
2033
2034                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
2035                         /*
2036                          * There is no way to resolve a short write situation
2037                          * for a !Uptodate page (except by double copying in
2038                          * the caller done by generic_perform_write_2copy).
2039                          *
2040                          * Instead, we have to bring it uptodate here.
2041                          */
2042                         ret = aops->readpage(file, page);
2043                         page_cache_release(page);
2044                         if (ret) {
2045                                 if (ret == AOP_TRUNCATED_PAGE)
2046                                         goto again;
2047                                 return ret;
2048                         }
2049                         goto again;
2050                 }
2051
2052                 ret = aops->prepare_write(file, page, offset, offset+len);
2053                 if (ret) {
2054                         unlock_page(page);
2055                         page_cache_release(page);
2056                         if (pos + len > inode->i_size)
2057                                 vmtruncate(inode, inode->i_size);
2058                 }
2059                 return ret;
2060         }
2061 }
2062 EXPORT_SYMBOL(pagecache_write_begin);
2063
2064 int pagecache_write_end(struct file *file, struct address_space *mapping,
2065                                 loff_t pos, unsigned len, unsigned copied,
2066                                 struct page *page, void *fsdata)
2067 {
2068         const struct address_space_operations *aops = mapping->a_ops;
2069         int ret;
2070
2071         if (aops->write_end) {
2072                 mark_page_accessed(page);
2073                 ret = aops->write_end(file, mapping, pos, len, copied,
2074                                                         page, fsdata);
2075         } else {
2076                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2077                 struct inode *inode = mapping->host;
2078
2079                 flush_dcache_page(page);
2080                 ret = aops->commit_write(file, page, offset, offset+len);
2081                 unlock_page(page);
2082                 mark_page_accessed(page);
2083                 page_cache_release(page);
2084
2085                 if (ret < 0) {
2086                         if (pos + len > inode->i_size)
2087                                 vmtruncate(inode, inode->i_size);
2088                 } else if (ret > 0)
2089                         ret = min_t(size_t, copied, ret);
2090                 else
2091                         ret = copied;
2092         }
2093
2094         return ret;
2095 }
2096 EXPORT_SYMBOL(pagecache_write_end);
2097
2098 ssize_t
2099 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2100                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2101                 size_t count, size_t ocount)
2102 {
2103         struct file     *file = iocb->ki_filp;
2104         struct address_space *mapping = file->f_mapping;
2105         struct inode    *inode = mapping->host;
2106         ssize_t         written;
2107         size_t          write_len;
2108         pgoff_t         end;
2109
2110         if (count != ocount)
2111                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2112
2113         /*
2114          * Unmap all mmappings of the file up-front.
2115          *
2116          * This will cause any pte dirty bits to be propagated into the
2117          * pageframes for the subsequent filemap_write_and_wait().
2118          */
2119         write_len = iov_length(iov, *nr_segs);
2120         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2121         if (mapping_mapped(mapping))
2122                 unmap_mapping_range(mapping, pos, write_len, 0);
2123
2124         written = filemap_write_and_wait(mapping);
2125         if (written)
2126                 goto out;
2127
2128         /*
2129          * After a write we want buffered reads to be sure to go to disk to get
2130          * the new data.  We invalidate clean cached page from the region we're
2131          * about to write.  We do this *before* the write so that we can return
2132          * without clobbering -EIOCBQUEUED from ->direct_IO().
2133          */
2134         if (mapping->nrpages) {
2135                 written = invalidate_inode_pages2_range(mapping,
2136                                         pos >> PAGE_CACHE_SHIFT, end);
2137                 /*
2138                  * If a page can not be invalidated, return 0 to fall back
2139                  * to buffered write.
2140                  */
2141                 if (written) {
2142                         if (written == -EBUSY)
2143                                 return 0;
2144                         goto out;
2145                 }
2146         }
2147
2148         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2149
2150         /*
2151          * Finally, try again to invalidate clean pages which might have been
2152          * cached by non-direct readahead, or faulted in by get_user_pages()
2153          * if the source of the write was an mmap'ed region of the file
2154          * we're writing.  Either one is a pretty crazy thing to do,
2155          * so we don't support it 100%.  If this invalidation
2156          * fails, tough, the write still worked...
2157          */
2158         if (mapping->nrpages) {
2159                 invalidate_inode_pages2_range(mapping,
2160                                               pos >> PAGE_CACHE_SHIFT, end);
2161         }
2162
2163         if (written > 0) {
2164                 loff_t end = pos + written;
2165                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2166                         i_size_write(inode,  end);
2167                         mark_inode_dirty(inode);
2168                 }
2169                 *ppos = end;
2170         }
2171
2172         /*
2173          * Sync the fs metadata but not the minor inode changes and
2174          * of course not the data as we did direct DMA for the IO.
2175          * i_mutex is held, which protects generic_osync_inode() from
2176          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2177          */
2178 out:
2179         if ((written >= 0 || written == -EIOCBQUEUED) &&
2180             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2181                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2182                 if (err < 0)
2183                         written = err;
2184         }
2185         return written;
2186 }
2187 EXPORT_SYMBOL(generic_file_direct_write);
2188
2189 /*
2190  * Find or create a page at the given pagecache position. Return the locked
2191  * page. This function is specifically for buffered writes.
2192  */
2193 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2194 {
2195         int status;
2196         struct page *page;
2197 repeat:
2198         page = find_lock_page(mapping, index);
2199         if (likely(page))
2200                 return page;
2201
2202         page = page_cache_alloc(mapping);
2203         if (!page)
2204                 return NULL;
2205         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2206         if (unlikely(status)) {
2207                 page_cache_release(page);
2208                 if (status == -EEXIST)
2209                         goto repeat;
2210                 return NULL;
2211         }
2212         return page;
2213 }
2214 EXPORT_SYMBOL(__grab_cache_page);
2215
2216 static ssize_t generic_perform_write_2copy(struct file *file,
2217                                 struct iov_iter *i, loff_t pos)
2218 {
2219         struct address_space *mapping = file->f_mapping;
2220         const struct address_space_operations *a_ops = mapping->a_ops;
2221         struct inode *inode = mapping->host;
2222         long status = 0;
2223         ssize_t written = 0;
2224
2225         do {
2226                 struct page *src_page;
2227                 struct page *page;
2228                 pgoff_t index;          /* Pagecache index for current page */
2229                 unsigned long offset;   /* Offset into pagecache page */
2230                 unsigned long bytes;    /* Bytes to write to page */
2231                 size_t copied;          /* Bytes copied from user */
2232
2233                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2234                 index = pos >> PAGE_CACHE_SHIFT;
2235                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2236                                                 iov_iter_count(i));
2237
2238                 /*
2239                  * a non-NULL src_page indicates that we're doing the
2240                  * copy via get_user_pages and kmap.
2241                  */
2242                 src_page = NULL;
2243
2244                 /*
2245                  * Bring in the user page that we will copy from _first_.
2246                  * Otherwise there's a nasty deadlock on copying from the
2247                  * same page as we're writing to, without it being marked
2248                  * up-to-date.
2249                  *
2250                  * Not only is this an optimisation, but it is also required
2251                  * to check that the address is actually valid, when atomic
2252                  * usercopies are used, below.
2253                  */
2254                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2255                         status = -EFAULT;
2256                         break;
2257                 }
2258
2259                 page = __grab_cache_page(mapping, index);
2260                 if (!page) {
2261                         status = -ENOMEM;
2262                         break;
2263                 }
2264
2265                 /*
2266                  * non-uptodate pages cannot cope with short copies, and we
2267                  * cannot take a pagefault with the destination page locked.
2268                  * So pin the source page to copy it.
2269                  */
2270                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2271                         unlock_page(page);
2272
2273                         src_page = alloc_page(GFP_KERNEL);
2274                         if (!src_page) {
2275                                 page_cache_release(page);
2276                                 status = -ENOMEM;
2277                                 break;
2278                         }
2279
2280                         /*
2281                          * Cannot get_user_pages with a page locked for the
2282                          * same reason as we can't take a page fault with a
2283                          * page locked (as explained below).
2284                          */
2285                         copied = iov_iter_copy_from_user(src_page, i,
2286                                                                 offset, bytes);
2287                         if (unlikely(copied == 0)) {
2288                                 status = -EFAULT;
2289                                 page_cache_release(page);
2290                                 page_cache_release(src_page);
2291                                 break;
2292                         }
2293                         bytes = copied;
2294
2295                         lock_page(page);
2296                         /*
2297                          * Can't handle the page going uptodate here, because
2298                          * that means we would use non-atomic usercopies, which
2299                          * zero out the tail of the page, which can cause
2300                          * zeroes to become transiently visible. We could just
2301                          * use a non-zeroing copy, but the APIs aren't too
2302                          * consistent.
2303                          */
2304                         if (unlikely(!page->mapping || PageUptodate(page))) {
2305                                 unlock_page(page);
2306                                 page_cache_release(page);
2307                                 page_cache_release(src_page);
2308                                 continue;
2309                         }
2310                 }
2311
2312                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2313                 if (unlikely(status))
2314                         goto fs_write_aop_error;
2315
2316                 if (!src_page) {
2317                         /*
2318                          * Must not enter the pagefault handler here, because
2319                          * we hold the page lock, so we might recursively
2320                          * deadlock on the same lock, or get an ABBA deadlock
2321                          * against a different lock, or against the mmap_sem
2322                          * (which nests outside the page lock).  So increment
2323                          * preempt count, and use _atomic usercopies.
2324                          *
2325                          * The page is uptodate so we are OK to encounter a
2326                          * short copy: if unmodified parts of the page are
2327                          * marked dirty and written out to disk, it doesn't
2328                          * really matter.
2329                          */
2330                         pagefault_disable();
2331                         copied = iov_iter_copy_from_user_atomic(page, i,
2332                                                                 offset, bytes);
2333                         pagefault_enable();
2334                 } else {
2335                         void *src, *dst;
2336                         src = kmap_atomic(src_page, KM_USER0);
2337                         dst = kmap_atomic(page, KM_USER1);
2338                         memcpy(dst + offset, src + offset, bytes);
2339                         kunmap_atomic(dst, KM_USER1);
2340                         kunmap_atomic(src, KM_USER0);
2341                         copied = bytes;
2342                 }
2343                 flush_dcache_page(page);
2344
2345                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2346                 if (unlikely(status < 0))
2347                         goto fs_write_aop_error;
2348                 if (unlikely(status > 0)) /* filesystem did partial write */
2349                         copied = min_t(size_t, copied, status);
2350
2351                 unlock_page(page);
2352                 mark_page_accessed(page);
2353                 page_cache_release(page);
2354                 if (src_page)
2355                         page_cache_release(src_page);
2356
2357                 iov_iter_advance(i, copied);
2358                 pos += copied;
2359                 written += copied;
2360
2361                 balance_dirty_pages_ratelimited(mapping);
2362                 cond_resched();
2363                 continue;
2364
2365 fs_write_aop_error:
2366                 unlock_page(page);
2367                 page_cache_release(page);
2368                 if (src_page)
2369                         page_cache_release(src_page);
2370
2371                 /*
2372                  * prepare_write() may have instantiated a few blocks
2373                  * outside i_size.  Trim these off again. Don't need
2374                  * i_size_read because we hold i_mutex.
2375                  */
2376                 if (pos + bytes > inode->i_size)
2377                         vmtruncate(inode, inode->i_size);
2378                 break;
2379         } while (iov_iter_count(i));
2380
2381         return written ? written : status;
2382 }
2383
2384 static ssize_t generic_perform_write(struct file *file,
2385                                 struct iov_iter *i, loff_t pos)
2386 {
2387         struct address_space *mapping = file->f_mapping;
2388         const struct address_space_operations *a_ops = mapping->a_ops;
2389         long status = 0;
2390         ssize_t written = 0;
2391         unsigned int flags = 0;
2392
2393         /*
2394          * Copies from kernel address space cannot fail (NFSD is a big user).
2395          */
2396         if (segment_eq(get_fs(), KERNEL_DS))
2397                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2398
2399         do {
2400                 struct page *page;
2401                 pgoff_t index;          /* Pagecache index for current page */
2402                 unsigned long offset;   /* Offset into pagecache page */
2403                 unsigned long bytes;    /* Bytes to write to page */
2404                 size_t copied;          /* Bytes copied from user */
2405                 void *fsdata;
2406
2407                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2408                 index = pos >> PAGE_CACHE_SHIFT;
2409                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2410                                                 iov_iter_count(i));
2411
2412 again:
2413
2414                 /*
2415                  * Bring in the user page that we will copy from _first_.
2416                  * Otherwise there's a nasty deadlock on copying from the
2417                  * same page as we're writing to, without it being marked
2418                  * up-to-date.
2419                  *
2420                  * Not only is this an optimisation, but it is also required
2421                  * to check that the address is actually valid, when atomic
2422                  * usercopies are used, below.
2423                  */
2424                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2425                         status = -EFAULT;
2426                         break;
2427                 }
2428
2429                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2430                                                 &page, &fsdata);
2431                 if (unlikely(status))
2432                         break;
2433
2434                 pagefault_disable();
2435                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2436                 pagefault_enable();
2437                 flush_dcache_page(page);
2438
2439                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2440                                                 page, fsdata);
2441                 if (unlikely(status < 0))
2442                         break;
2443                 copied = status;
2444
2445                 cond_resched();
2446
2447                 iov_iter_advance(i, copied);
2448                 if (unlikely(copied == 0)) {
2449                         /*
2450                          * If we were unable to copy any data at all, we must
2451                          * fall back to a single segment length write.
2452                          *
2453                          * If we didn't fallback here, we could livelock
2454                          * because not all segments in the iov can be copied at
2455                          * once without a pagefault.
2456                          */
2457                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2458                                                 iov_iter_single_seg_count(i));
2459                         goto again;
2460                 }
2461                 pos += copied;
2462                 written += copied;
2463
2464                 balance_dirty_pages_ratelimited(mapping);
2465
2466         } while (iov_iter_count(i));
2467
2468         return written ? written : status;
2469 }
2470
2471 ssize_t
2472 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2473                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2474                 size_t count, ssize_t written)
2475 {
2476         struct file *file = iocb->ki_filp;
2477         struct address_space *mapping = file->f_mapping;
2478         const struct address_space_operations *a_ops = mapping->a_ops;
2479         struct inode *inode = mapping->host;
2480         ssize_t status;
2481         struct iov_iter i;
2482
2483         iov_iter_init(&i, iov, nr_segs, count, written);
2484         if (a_ops->write_begin)
2485                 status = generic_perform_write(file, &i, pos);
2486         else
2487                 status = generic_perform_write_2copy(file, &i, pos);
2488
2489         if (likely(status >= 0)) {
2490                 written += status;
2491                 *ppos = pos + status;
2492
2493                 /*
2494                  * For now, when the user asks for O_SYNC, we'll actually give
2495                  * O_DSYNC
2496                  */
2497                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2498                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2499                                 status = generic_osync_inode(inode, mapping,
2500                                                 OSYNC_METADATA|OSYNC_DATA);
2501                 }
2502         }
2503         
2504         /*
2505          * If we get here for O_DIRECT writes then we must have fallen through
2506          * to buffered writes (block instantiation inside i_size).  So we sync
2507          * the file data here, to try to honour O_DIRECT expectations.
2508          */
2509         if (unlikely(file->f_flags & O_DIRECT) && written)
2510                 status = filemap_write_and_wait(mapping);
2511
2512         return written ? written : status;
2513 }
2514 EXPORT_SYMBOL(generic_file_buffered_write);
2515
2516 static ssize_t
2517 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2518                                 unsigned long nr_segs, loff_t *ppos)
2519 {
2520         struct file *file = iocb->ki_filp;
2521         struct address_space * mapping = file->f_mapping;
2522         size_t ocount;          /* original count */
2523         size_t count;           /* after file limit checks */
2524         struct inode    *inode = mapping->host;
2525         loff_t          pos;
2526         ssize_t         written;
2527         ssize_t         err;
2528
2529         ocount = 0;
2530         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2531         if (err)
2532                 return err;
2533
2534         count = ocount;
2535         pos = *ppos;
2536
2537         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2538
2539         /* We can write back this queue in page reclaim */
2540         current->backing_dev_info = mapping->backing_dev_info;
2541         written = 0;
2542
2543         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2544         if (err)
2545                 goto out;
2546
2547         if (count == 0)
2548                 goto out;
2549
2550         err = file_remove_suid(file);
2551         if (err)
2552                 goto out;
2553
2554         file_update_time(file);
2555
2556         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2557         if (unlikely(file->f_flags & O_DIRECT)) {
2558                 loff_t endbyte;
2559                 ssize_t written_buffered;
2560
2561                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2562                                                         ppos, count, ocount);
2563                 if (written < 0 || written == count)
2564                         goto out;
2565                 /*
2566                  * direct-io write to a hole: fall through to buffered I/O
2567                  * for completing the rest of the request.
2568                  */
2569                 pos += written;
2570                 count -= written;
2571                 written_buffered = generic_file_buffered_write(iocb, iov,
2572                                                 nr_segs, pos, ppos, count,
2573                                                 written);
2574                 /*
2575                  * If generic_file_buffered_write() retuned a synchronous error
2576                  * then we want to return the number of bytes which were
2577                  * direct-written, or the error code if that was zero.  Note
2578                  * that this differs from normal direct-io semantics, which
2579                  * will return -EFOO even if some bytes were written.
2580                  */
2581                 if (written_buffered < 0) {
2582                         err = written_buffered;
2583                         goto out;
2584                 }
2585
2586                 /*
2587                  * We need to ensure that the page cache pages are written to
2588                  * disk and invalidated to preserve the expected O_DIRECT
2589                  * semantics.
2590                  */
2591                 endbyte = pos + written_buffered - written - 1;
2592                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2593                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2594                                             SYNC_FILE_RANGE_WRITE|
2595                                             SYNC_FILE_RANGE_WAIT_AFTER);
2596                 if (err == 0) {
2597                         written = written_buffered;
2598                         invalidate_mapping_pages(mapping,
2599                                                  pos >> PAGE_CACHE_SHIFT,
2600                                                  endbyte >> PAGE_CACHE_SHIFT);
2601                 } else {
2602                         /*
2603                          * We don't know how much we wrote, so just return
2604                          * the number of bytes which were direct-written
2605                          */
2606                 }
2607         } else {
2608                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2609                                 pos, ppos, count, written);
2610         }
2611 out:
2612         current->backing_dev_info = NULL;
2613         return written ? written : err;
2614 }
2615
2616 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2617                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2618 {
2619         struct file *file = iocb->ki_filp;
2620         struct address_space *mapping = file->f_mapping;
2621         struct inode *inode = mapping->host;
2622         ssize_t ret;
2623
2624         BUG_ON(iocb->ki_pos != pos);
2625
2626         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2627                         &iocb->ki_pos);
2628
2629         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2630                 ssize_t err;
2631
2632                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2633                 if (err < 0)
2634                         ret = err;
2635         }
2636         return ret;
2637 }
2638 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2639
2640 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2641                 unsigned long nr_segs, loff_t pos)
2642 {
2643         struct file *file = iocb->ki_filp;
2644         struct address_space *mapping = file->f_mapping;
2645         struct inode *inode = mapping->host;
2646         ssize_t ret;
2647
2648         BUG_ON(iocb->ki_pos != pos);
2649
2650         mutex_lock(&inode->i_mutex);
2651         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2652                         &iocb->ki_pos);
2653         mutex_unlock(&inode->i_mutex);
2654
2655         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2656                 ssize_t err;
2657
2658                 err = sync_page_range(inode, mapping, pos, ret);
2659                 if (err < 0)
2660                         ret = err;
2661         }
2662         return ret;
2663 }
2664 EXPORT_SYMBOL(generic_file_aio_write);
2665
2666 /**
2667  * try_to_release_page() - release old fs-specific metadata on a page
2668  *
2669  * @page: the page which the kernel is trying to free
2670  * @gfp_mask: memory allocation flags (and I/O mode)
2671  *
2672  * The address_space is to try to release any data against the page
2673  * (presumably at page->private).  If the release was successful, return `1'.
2674  * Otherwise return zero.
2675  *
2676  * The @gfp_mask argument specifies whether I/O may be performed to release
2677  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2678  *
2679  */
2680 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2681 {
2682         struct address_space * const mapping = page->mapping;
2683
2684         BUG_ON(!PageLocked(page));
2685         if (PageWriteback(page))
2686                 return 0;
2687
2688         if (mapping && mapping->a_ops->releasepage)
2689                 return mapping->a_ops->releasepage(page, gfp_mask);
2690         return try_to_free_buffers(page);
2691 }
2692
2693 EXPORT_SYMBOL(try_to_release_page);