83efee76a5c0b12c577903346eca32f7ca49be18
[linux-3.10.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/export.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/cpuset.h>
33 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34 #include <linux/memcontrol.h>
35 #include <linux/cleancache.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 try_to_free_buffers */
42
43 #include <asm/mman.h>
44
45 /*
46  * Shared mappings implemented 30.11.1994. It's not fully working yet,
47  * though.
48  *
49  * Shared mappings now work. 15.8.1995  Bruno.
50  *
51  * finished 'unifying' the page and buffer cache and SMP-threaded the
52  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
53  *
54  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
55  */
56
57 /*
58  * Lock ordering:
59  *
60  *  ->i_mmap_mutex              (truncate_pagecache)
61  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
62  *      ->swap_lock             (exclusive_swap_page, others)
63  *        ->mapping->tree_lock
64  *
65  *  ->i_mutex
66  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
67  *
68  *  ->mmap_sem
69  *    ->i_mmap_mutex
70  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
71  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
72  *
73  *  ->mmap_sem
74  *    ->lock_page               (access_process_vm)
75  *
76  *  ->i_mutex                   (generic_file_buffered_write)
77  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
78  *
79  *  bdi->wb.list_lock
80  *    sb_lock                   (fs/fs-writeback.c)
81  *    ->mapping->tree_lock      (__sync_single_inode)
82  *
83  *  ->i_mmap_mutex
84  *    ->anon_vma.lock           (vma_adjust)
85  *
86  *  ->anon_vma.lock
87  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
88  *
89  *  ->page_table_lock or pte_lock
90  *    ->swap_lock               (try_to_unmap_one)
91  *    ->private_lock            (try_to_unmap_one)
92  *    ->tree_lock               (try_to_unmap_one)
93  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
94  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
95  *    ->private_lock            (page_remove_rmap->set_page_dirty)
96  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
97  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
98  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
99  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
100  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
101  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
102  *
103  * ->i_mmap_mutex
104  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
105  */
106
107 /*
108  * Delete a page from the page cache and free it. Caller has to make
109  * sure the page is locked and that nobody else uses it - or that usage
110  * is safe.  The caller must hold the mapping's tree_lock.
111  */
112 void __delete_from_page_cache(struct page *page)
113 {
114         struct address_space *mapping = page->mapping;
115
116         /*
117          * if we're uptodate, flush out into the cleancache, otherwise
118          * invalidate any existing cleancache entries.  We can't leave
119          * stale data around in the cleancache once our page is gone
120          */
121         if (PageUptodate(page) && PageMappedToDisk(page))
122                 cleancache_put_page(page);
123         else
124                 cleancache_invalidate_page(mapping, page);
125
126         radix_tree_delete(&mapping->page_tree, page->index);
127         page->mapping = NULL;
128         /* Leave page->index set: truncation lookup relies upon it */
129         mapping->nrpages--;
130         __dec_zone_page_state(page, NR_FILE_PAGES);
131         if (PageSwapBacked(page))
132                 __dec_zone_page_state(page, NR_SHMEM);
133         BUG_ON(page_mapped(page));
134
135         /*
136          * Some filesystems seem to re-dirty the page even after
137          * the VM has canceled the dirty bit (eg ext3 journaling).
138          *
139          * Fix it up by doing a final dirty accounting check after
140          * having removed the page entirely.
141          */
142         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
143                 dec_zone_page_state(page, NR_FILE_DIRTY);
144                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
145         }
146 }
147
148 /**
149  * delete_from_page_cache - delete page from page cache
150  * @page: the page which the kernel is trying to remove from page cache
151  *
152  * This must be called only on pages that have been verified to be in the page
153  * cache and locked.  It will never put the page into the free list, the caller
154  * has a reference on the page.
155  */
156 void delete_from_page_cache(struct page *page)
157 {
158         struct address_space *mapping = page->mapping;
159         void (*freepage)(struct page *);
160
161         BUG_ON(!PageLocked(page));
162
163         freepage = mapping->a_ops->freepage;
164         spin_lock_irq(&mapping->tree_lock);
165         __delete_from_page_cache(page);
166         spin_unlock_irq(&mapping->tree_lock);
167         mem_cgroup_uncharge_cache_page(page);
168
169         if (freepage)
170                 freepage(page);
171         page_cache_release(page);
172 }
173 EXPORT_SYMBOL(delete_from_page_cache);
174
175 static int sleep_on_page(void *word)
176 {
177         io_schedule();
178         return 0;
179 }
180
181 static int sleep_on_page_killable(void *word)
182 {
183         sleep_on_page(word);
184         return fatal_signal_pending(current) ? -EINTR : 0;
185 }
186
187 /**
188  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
189  * @mapping:    address space structure to write
190  * @start:      offset in bytes where the range starts
191  * @end:        offset in bytes where the range ends (inclusive)
192  * @sync_mode:  enable synchronous operation
193  *
194  * Start writeback against all of a mapping's dirty pages that lie
195  * within the byte offsets <start, end> inclusive.
196  *
197  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
198  * opposed to a regular memory cleansing writeback.  The difference between
199  * these two operations is that if a dirty page/buffer is encountered, it must
200  * be waited upon, and not just skipped over.
201  */
202 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
203                                 loff_t end, int sync_mode)
204 {
205         int ret;
206         struct writeback_control wbc = {
207                 .sync_mode = sync_mode,
208                 .nr_to_write = LONG_MAX,
209                 .range_start = start,
210                 .range_end = end,
211         };
212
213         if (!mapping_cap_writeback_dirty(mapping))
214                 return 0;
215
216         ret = do_writepages(mapping, &wbc);
217         return ret;
218 }
219
220 static inline int __filemap_fdatawrite(struct address_space *mapping,
221         int sync_mode)
222 {
223         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
224 }
225
226 int filemap_fdatawrite(struct address_space *mapping)
227 {
228         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
229 }
230 EXPORT_SYMBOL(filemap_fdatawrite);
231
232 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
233                                 loff_t end)
234 {
235         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
236 }
237 EXPORT_SYMBOL(filemap_fdatawrite_range);
238
239 /**
240  * filemap_flush - mostly a non-blocking flush
241  * @mapping:    target address_space
242  *
243  * This is a mostly non-blocking flush.  Not suitable for data-integrity
244  * purposes - I/O may not be started against all dirty pages.
245  */
246 int filemap_flush(struct address_space *mapping)
247 {
248         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
249 }
250 EXPORT_SYMBOL(filemap_flush);
251
252 /**
253  * filemap_fdatawait_range - wait for writeback to complete
254  * @mapping:            address space structure to wait for
255  * @start_byte:         offset in bytes where the range starts
256  * @end_byte:           offset in bytes where the range ends (inclusive)
257  *
258  * Walk the list of under-writeback pages of the given address space
259  * in the given range and wait for all of them.
260  */
261 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
262                             loff_t end_byte)
263 {
264         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
265         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
266         struct pagevec pvec;
267         int nr_pages;
268         int ret = 0;
269
270         if (end_byte < start_byte)
271                 return 0;
272
273         pagevec_init(&pvec, 0);
274         while ((index <= end) &&
275                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
276                         PAGECACHE_TAG_WRITEBACK,
277                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
278                 unsigned i;
279
280                 for (i = 0; i < nr_pages; i++) {
281                         struct page *page = pvec.pages[i];
282
283                         /* until radix tree lookup accepts end_index */
284                         if (page->index > end)
285                                 continue;
286
287                         wait_on_page_writeback(page);
288                         if (TestClearPageError(page))
289                                 ret = -EIO;
290                 }
291                 pagevec_release(&pvec);
292                 cond_resched();
293         }
294
295         /* Check for outstanding write errors */
296         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
297                 ret = -ENOSPC;
298         if (test_and_clear_bit(AS_EIO, &mapping->flags))
299                 ret = -EIO;
300
301         return ret;
302 }
303 EXPORT_SYMBOL(filemap_fdatawait_range);
304
305 /**
306  * filemap_fdatawait - wait for all under-writeback pages to complete
307  * @mapping: address space structure to wait for
308  *
309  * Walk the list of under-writeback pages of the given address space
310  * and wait for all of them.
311  */
312 int filemap_fdatawait(struct address_space *mapping)
313 {
314         loff_t i_size = i_size_read(mapping->host);
315
316         if (i_size == 0)
317                 return 0;
318
319         return filemap_fdatawait_range(mapping, 0, i_size - 1);
320 }
321 EXPORT_SYMBOL(filemap_fdatawait);
322
323 int filemap_write_and_wait(struct address_space *mapping)
324 {
325         int err = 0;
326
327         if (mapping->nrpages) {
328                 err = filemap_fdatawrite(mapping);
329                 /*
330                  * Even if the above returned error, the pages may be
331                  * written partially (e.g. -ENOSPC), so we wait for it.
332                  * But the -EIO is special case, it may indicate the worst
333                  * thing (e.g. bug) happened, so we avoid waiting for it.
334                  */
335                 if (err != -EIO) {
336                         int err2 = filemap_fdatawait(mapping);
337                         if (!err)
338                                 err = err2;
339                 }
340         }
341         return err;
342 }
343 EXPORT_SYMBOL(filemap_write_and_wait);
344
345 /**
346  * filemap_write_and_wait_range - write out & wait on a file range
347  * @mapping:    the address_space for the pages
348  * @lstart:     offset in bytes where the range starts
349  * @lend:       offset in bytes where the range ends (inclusive)
350  *
351  * Write out and wait upon file offsets lstart->lend, inclusive.
352  *
353  * Note that `lend' is inclusive (describes the last byte to be written) so
354  * that this function can be used to write to the very end-of-file (end = -1).
355  */
356 int filemap_write_and_wait_range(struct address_space *mapping,
357                                  loff_t lstart, loff_t lend)
358 {
359         int err = 0;
360
361         if (mapping->nrpages) {
362                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
363                                                  WB_SYNC_ALL);
364                 /* See comment of filemap_write_and_wait() */
365                 if (err != -EIO) {
366                         int err2 = filemap_fdatawait_range(mapping,
367                                                 lstart, lend);
368                         if (!err)
369                                 err = err2;
370                 }
371         }
372         return err;
373 }
374 EXPORT_SYMBOL(filemap_write_and_wait_range);
375
376 /**
377  * replace_page_cache_page - replace a pagecache page with a new one
378  * @old:        page to be replaced
379  * @new:        page to replace with
380  * @gfp_mask:   allocation mode
381  *
382  * This function replaces a page in the pagecache with a new one.  On
383  * success it acquires the pagecache reference for the new page and
384  * drops it for the old page.  Both the old and new pages must be
385  * locked.  This function does not add the new page to the LRU, the
386  * caller must do that.
387  *
388  * The remove + add is atomic.  The only way this function can fail is
389  * memory allocation failure.
390  */
391 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
392 {
393         int error;
394
395         VM_BUG_ON(!PageLocked(old));
396         VM_BUG_ON(!PageLocked(new));
397         VM_BUG_ON(new->mapping);
398
399         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
400         if (!error) {
401                 struct address_space *mapping = old->mapping;
402                 void (*freepage)(struct page *);
403
404                 pgoff_t offset = old->index;
405                 freepage = mapping->a_ops->freepage;
406
407                 page_cache_get(new);
408                 new->mapping = mapping;
409                 new->index = offset;
410
411                 spin_lock_irq(&mapping->tree_lock);
412                 __delete_from_page_cache(old);
413                 error = radix_tree_insert(&mapping->page_tree, offset, new);
414                 BUG_ON(error);
415                 mapping->nrpages++;
416                 __inc_zone_page_state(new, NR_FILE_PAGES);
417                 if (PageSwapBacked(new))
418                         __inc_zone_page_state(new, NR_SHMEM);
419                 spin_unlock_irq(&mapping->tree_lock);
420                 /* mem_cgroup codes must not be called under tree_lock */
421                 mem_cgroup_replace_page_cache(old, new);
422                 radix_tree_preload_end();
423                 if (freepage)
424                         freepage(old);
425                 page_cache_release(old);
426         }
427
428         return error;
429 }
430 EXPORT_SYMBOL_GPL(replace_page_cache_page);
431
432 /**
433  * add_to_page_cache_locked - add a locked page to the pagecache
434  * @page:       page to add
435  * @mapping:    the page's address_space
436  * @offset:     page index
437  * @gfp_mask:   page allocation mode
438  *
439  * This function is used to add a page to the pagecache. It must be locked.
440  * This function does not add the page to the LRU.  The caller must do that.
441  */
442 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
443                 pgoff_t offset, gfp_t gfp_mask)
444 {
445         int error;
446
447         VM_BUG_ON(!PageLocked(page));
448         VM_BUG_ON(PageSwapBacked(page));
449
450         error = mem_cgroup_cache_charge(page, current->mm,
451                                         gfp_mask & GFP_RECLAIM_MASK);
452         if (error)
453                 goto out;
454
455         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
456         if (error == 0) {
457                 page_cache_get(page);
458                 page->mapping = mapping;
459                 page->index = offset;
460
461                 spin_lock_irq(&mapping->tree_lock);
462                 error = radix_tree_insert(&mapping->page_tree, offset, page);
463                 if (likely(!error)) {
464                         mapping->nrpages++;
465                         __inc_zone_page_state(page, NR_FILE_PAGES);
466                         spin_unlock_irq(&mapping->tree_lock);
467                 } else {
468                         page->mapping = NULL;
469                         /* Leave page->index set: truncation relies upon it */
470                         spin_unlock_irq(&mapping->tree_lock);
471                         mem_cgroup_uncharge_cache_page(page);
472                         page_cache_release(page);
473                 }
474                 radix_tree_preload_end();
475         } else
476                 mem_cgroup_uncharge_cache_page(page);
477 out:
478         return error;
479 }
480 EXPORT_SYMBOL(add_to_page_cache_locked);
481
482 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
483                                 pgoff_t offset, gfp_t gfp_mask)
484 {
485         int ret;
486
487         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
488         if (ret == 0)
489                 lru_cache_add_file(page);
490         return ret;
491 }
492 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
493
494 #ifdef CONFIG_NUMA
495 struct page *__page_cache_alloc(gfp_t gfp)
496 {
497         int n;
498         struct page *page;
499
500         if (cpuset_do_page_mem_spread()) {
501                 unsigned int cpuset_mems_cookie;
502                 do {
503                         cpuset_mems_cookie = get_mems_allowed();
504                         n = cpuset_mem_spread_node();
505                         page = alloc_pages_exact_node(n, gfp, 0);
506                 } while (!put_mems_allowed(cpuset_mems_cookie) && !page);
507
508                 return page;
509         }
510         return alloc_pages(gfp, 0);
511 }
512 EXPORT_SYMBOL(__page_cache_alloc);
513 #endif
514
515 /*
516  * In order to wait for pages to become available there must be
517  * waitqueues associated with pages. By using a hash table of
518  * waitqueues where the bucket discipline is to maintain all
519  * waiters on the same queue and wake all when any of the pages
520  * become available, and for the woken contexts to check to be
521  * sure the appropriate page became available, this saves space
522  * at a cost of "thundering herd" phenomena during rare hash
523  * collisions.
524  */
525 static wait_queue_head_t *page_waitqueue(struct page *page)
526 {
527         const struct zone *zone = page_zone(page);
528
529         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
530 }
531
532 static inline void wake_up_page(struct page *page, int bit)
533 {
534         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
535 }
536
537 void wait_on_page_bit(struct page *page, int bit_nr)
538 {
539         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
540
541         if (test_bit(bit_nr, &page->flags))
542                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
543                                                         TASK_UNINTERRUPTIBLE);
544 }
545 EXPORT_SYMBOL(wait_on_page_bit);
546
547 int wait_on_page_bit_killable(struct page *page, int bit_nr)
548 {
549         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
550
551         if (!test_bit(bit_nr, &page->flags))
552                 return 0;
553
554         return __wait_on_bit(page_waitqueue(page), &wait,
555                              sleep_on_page_killable, TASK_KILLABLE);
556 }
557
558 /**
559  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
560  * @page: Page defining the wait queue of interest
561  * @waiter: Waiter to add to the queue
562  *
563  * Add an arbitrary @waiter to the wait queue for the nominated @page.
564  */
565 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
566 {
567         wait_queue_head_t *q = page_waitqueue(page);
568         unsigned long flags;
569
570         spin_lock_irqsave(&q->lock, flags);
571         __add_wait_queue(q, waiter);
572         spin_unlock_irqrestore(&q->lock, flags);
573 }
574 EXPORT_SYMBOL_GPL(add_page_wait_queue);
575
576 /**
577  * unlock_page - unlock a locked page
578  * @page: the page
579  *
580  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
581  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
582  * mechananism between PageLocked pages and PageWriteback pages is shared.
583  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
584  *
585  * The mb is necessary to enforce ordering between the clear_bit and the read
586  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
587  */
588 void unlock_page(struct page *page)
589 {
590         VM_BUG_ON(!PageLocked(page));
591         clear_bit_unlock(PG_locked, &page->flags);
592         smp_mb__after_clear_bit();
593         wake_up_page(page, PG_locked);
594 }
595 EXPORT_SYMBOL(unlock_page);
596
597 /**
598  * end_page_writeback - end writeback against a page
599  * @page: the page
600  */
601 void end_page_writeback(struct page *page)
602 {
603         if (TestClearPageReclaim(page))
604                 rotate_reclaimable_page(page);
605
606         if (!test_clear_page_writeback(page))
607                 BUG();
608
609         smp_mb__after_clear_bit();
610         wake_up_page(page, PG_writeback);
611 }
612 EXPORT_SYMBOL(end_page_writeback);
613
614 /**
615  * __lock_page - get a lock on the page, assuming we need to sleep to get it
616  * @page: the page to lock
617  */
618 void __lock_page(struct page *page)
619 {
620         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
621
622         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
623                                                         TASK_UNINTERRUPTIBLE);
624 }
625 EXPORT_SYMBOL(__lock_page);
626
627 int __lock_page_killable(struct page *page)
628 {
629         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
630
631         return __wait_on_bit_lock(page_waitqueue(page), &wait,
632                                         sleep_on_page_killable, TASK_KILLABLE);
633 }
634 EXPORT_SYMBOL_GPL(__lock_page_killable);
635
636 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
637                          unsigned int flags)
638 {
639         if (flags & FAULT_FLAG_ALLOW_RETRY) {
640                 /*
641                  * CAUTION! In this case, mmap_sem is not released
642                  * even though return 0.
643                  */
644                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
645                         return 0;
646
647                 up_read(&mm->mmap_sem);
648                 if (flags & FAULT_FLAG_KILLABLE)
649                         wait_on_page_locked_killable(page);
650                 else
651                         wait_on_page_locked(page);
652                 return 0;
653         } else {
654                 if (flags & FAULT_FLAG_KILLABLE) {
655                         int ret;
656
657                         ret = __lock_page_killable(page);
658                         if (ret) {
659                                 up_read(&mm->mmap_sem);
660                                 return 0;
661                         }
662                 } else
663                         __lock_page(page);
664                 return 1;
665         }
666 }
667
668 /**
669  * find_get_page - find and get a page reference
670  * @mapping: the address_space to search
671  * @offset: the page index
672  *
673  * Is there a pagecache struct page at the given (mapping, offset) tuple?
674  * If yes, increment its refcount and return it; if no, return NULL.
675  */
676 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
677 {
678         void **pagep;
679         struct page *page;
680
681         rcu_read_lock();
682 repeat:
683         page = NULL;
684         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
685         if (pagep) {
686                 page = radix_tree_deref_slot(pagep);
687                 if (unlikely(!page))
688                         goto out;
689                 if (radix_tree_exception(page)) {
690                         if (radix_tree_deref_retry(page))
691                                 goto repeat;
692                         /*
693                          * Otherwise, shmem/tmpfs must be storing a swap entry
694                          * here as an exceptional entry: so return it without
695                          * attempting to raise page count.
696                          */
697                         goto out;
698                 }
699                 if (!page_cache_get_speculative(page))
700                         goto repeat;
701
702                 /*
703                  * Has the page moved?
704                  * This is part of the lockless pagecache protocol. See
705                  * include/linux/pagemap.h for details.
706                  */
707                 if (unlikely(page != *pagep)) {
708                         page_cache_release(page);
709                         goto repeat;
710                 }
711         }
712 out:
713         rcu_read_unlock();
714
715         return page;
716 }
717 EXPORT_SYMBOL(find_get_page);
718
719 /**
720  * find_lock_page - locate, pin and lock a pagecache page
721  * @mapping: the address_space to search
722  * @offset: the page index
723  *
724  * Locates the desired pagecache page, locks it, increments its reference
725  * count and returns its address.
726  *
727  * Returns zero if the page was not present. find_lock_page() may sleep.
728  */
729 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
730 {
731         struct page *page;
732
733 repeat:
734         page = find_get_page(mapping, offset);
735         if (page && !radix_tree_exception(page)) {
736                 lock_page(page);
737                 /* Has the page been truncated? */
738                 if (unlikely(page->mapping != mapping)) {
739                         unlock_page(page);
740                         page_cache_release(page);
741                         goto repeat;
742                 }
743                 VM_BUG_ON(page->index != offset);
744         }
745         return page;
746 }
747 EXPORT_SYMBOL(find_lock_page);
748
749 /**
750  * find_or_create_page - locate or add a pagecache page
751  * @mapping: the page's address_space
752  * @index: the page's index into the mapping
753  * @gfp_mask: page allocation mode
754  *
755  * Locates a page in the pagecache.  If the page is not present, a new page
756  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
757  * LRU list.  The returned page is locked and has its reference count
758  * incremented.
759  *
760  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
761  * allocation!
762  *
763  * find_or_create_page() returns the desired page's address, or zero on
764  * memory exhaustion.
765  */
766 struct page *find_or_create_page(struct address_space *mapping,
767                 pgoff_t index, gfp_t gfp_mask)
768 {
769         struct page *page;
770         int err;
771 repeat:
772         page = find_lock_page(mapping, index);
773         if (!page) {
774                 page = __page_cache_alloc(gfp_mask);
775                 if (!page)
776                         return NULL;
777                 /*
778                  * We want a regular kernel memory (not highmem or DMA etc)
779                  * allocation for the radix tree nodes, but we need to honour
780                  * the context-specific requirements the caller has asked for.
781                  * GFP_RECLAIM_MASK collects those requirements.
782                  */
783                 err = add_to_page_cache_lru(page, mapping, index,
784                         (gfp_mask & GFP_RECLAIM_MASK));
785                 if (unlikely(err)) {
786                         page_cache_release(page);
787                         page = NULL;
788                         if (err == -EEXIST)
789                                 goto repeat;
790                 }
791         }
792         return page;
793 }
794 EXPORT_SYMBOL(find_or_create_page);
795
796 /**
797  * find_get_pages - gang pagecache lookup
798  * @mapping:    The address_space to search
799  * @start:      The starting page index
800  * @nr_pages:   The maximum number of pages
801  * @pages:      Where the resulting pages are placed
802  *
803  * find_get_pages() will search for and return a group of up to
804  * @nr_pages pages in the mapping.  The pages are placed at @pages.
805  * find_get_pages() takes a reference against the returned pages.
806  *
807  * The search returns a group of mapping-contiguous pages with ascending
808  * indexes.  There may be holes in the indices due to not-present pages.
809  *
810  * find_get_pages() returns the number of pages which were found.
811  */
812 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
813                             unsigned int nr_pages, struct page **pages)
814 {
815         struct radix_tree_iter iter;
816         void **slot;
817         unsigned ret = 0;
818
819         if (unlikely(!nr_pages))
820                 return 0;
821
822         rcu_read_lock();
823 restart:
824         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
825                 struct page *page;
826 repeat:
827                 page = radix_tree_deref_slot(slot);
828                 if (unlikely(!page))
829                         continue;
830
831                 if (radix_tree_exception(page)) {
832                         if (radix_tree_deref_retry(page)) {
833                                 /*
834                                  * Transient condition which can only trigger
835                                  * when entry at index 0 moves out of or back
836                                  * to root: none yet gotten, safe to restart.
837                                  */
838                                 WARN_ON(iter.index);
839                                 goto restart;
840                         }
841                         /*
842                          * Otherwise, shmem/tmpfs must be storing a swap entry
843                          * here as an exceptional entry: so skip over it -
844                          * we only reach this from invalidate_mapping_pages().
845                          */
846                         continue;
847                 }
848
849                 if (!page_cache_get_speculative(page))
850                         goto repeat;
851
852                 /* Has the page moved? */
853                 if (unlikely(page != *slot)) {
854                         page_cache_release(page);
855                         goto repeat;
856                 }
857
858                 pages[ret] = page;
859                 if (++ret == nr_pages)
860                         break;
861         }
862
863         rcu_read_unlock();
864         return ret;
865 }
866
867 /**
868  * find_get_pages_contig - gang contiguous pagecache lookup
869  * @mapping:    The address_space to search
870  * @index:      The starting page index
871  * @nr_pages:   The maximum number of pages
872  * @pages:      Where the resulting pages are placed
873  *
874  * find_get_pages_contig() works exactly like find_get_pages(), except
875  * that the returned number of pages are guaranteed to be contiguous.
876  *
877  * find_get_pages_contig() returns the number of pages which were found.
878  */
879 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
880                                unsigned int nr_pages, struct page **pages)
881 {
882         struct radix_tree_iter iter;
883         void **slot;
884         unsigned int ret = 0;
885
886         if (unlikely(!nr_pages))
887                 return 0;
888
889         rcu_read_lock();
890 restart:
891         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
892                 struct page *page;
893 repeat:
894                 page = radix_tree_deref_slot(slot);
895                 /* The hole, there no reason to continue */
896                 if (unlikely(!page))
897                         break;
898
899                 if (radix_tree_exception(page)) {
900                         if (radix_tree_deref_retry(page)) {
901                                 /*
902                                  * Transient condition which can only trigger
903                                  * when entry at index 0 moves out of or back
904                                  * to root: none yet gotten, safe to restart.
905                                  */
906                                 goto restart;
907                         }
908                         /*
909                          * Otherwise, shmem/tmpfs must be storing a swap entry
910                          * here as an exceptional entry: so stop looking for
911                          * contiguous pages.
912                          */
913                         break;
914                 }
915
916                 if (!page_cache_get_speculative(page))
917                         goto repeat;
918
919                 /* Has the page moved? */
920                 if (unlikely(page != *slot)) {
921                         page_cache_release(page);
922                         goto repeat;
923                 }
924
925                 /*
926                  * must check mapping and index after taking the ref.
927                  * otherwise we can get both false positives and false
928                  * negatives, which is just confusing to the caller.
929                  */
930                 if (page->mapping == NULL || page->index != iter.index) {
931                         page_cache_release(page);
932                         break;
933                 }
934
935                 pages[ret] = page;
936                 if (++ret == nr_pages)
937                         break;
938         }
939         rcu_read_unlock();
940         return ret;
941 }
942 EXPORT_SYMBOL(find_get_pages_contig);
943
944 /**
945  * find_get_pages_tag - find and return pages that match @tag
946  * @mapping:    the address_space to search
947  * @index:      the starting page index
948  * @tag:        the tag index
949  * @nr_pages:   the maximum number of pages
950  * @pages:      where the resulting pages are placed
951  *
952  * Like find_get_pages, except we only return pages which are tagged with
953  * @tag.   We update @index to index the next page for the traversal.
954  */
955 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
956                         int tag, unsigned int nr_pages, struct page **pages)
957 {
958         struct radix_tree_iter iter;
959         void **slot;
960         unsigned ret = 0;
961
962         if (unlikely(!nr_pages))
963                 return 0;
964
965         rcu_read_lock();
966 restart:
967         radix_tree_for_each_tagged(slot, &mapping->page_tree,
968                                    &iter, *index, tag) {
969                 struct page *page;
970 repeat:
971                 page = radix_tree_deref_slot(slot);
972                 if (unlikely(!page))
973                         continue;
974
975                 if (radix_tree_exception(page)) {
976                         if (radix_tree_deref_retry(page)) {
977                                 /*
978                                  * Transient condition which can only trigger
979                                  * when entry at index 0 moves out of or back
980                                  * to root: none yet gotten, safe to restart.
981                                  */
982                                 goto restart;
983                         }
984                         /*
985                          * This function is never used on a shmem/tmpfs
986                          * mapping, so a swap entry won't be found here.
987                          */
988                         BUG();
989                 }
990
991                 if (!page_cache_get_speculative(page))
992                         goto repeat;
993
994                 /* Has the page moved? */
995                 if (unlikely(page != *slot)) {
996                         page_cache_release(page);
997                         goto repeat;
998                 }
999
1000                 pages[ret] = page;
1001                 if (++ret == nr_pages)
1002                         break;
1003         }
1004
1005         rcu_read_unlock();
1006
1007         if (ret)
1008                 *index = pages[ret - 1]->index + 1;
1009
1010         return ret;
1011 }
1012 EXPORT_SYMBOL(find_get_pages_tag);
1013
1014 /**
1015  * grab_cache_page_nowait - returns locked page at given index in given cache
1016  * @mapping: target address_space
1017  * @index: the page index
1018  *
1019  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1020  * This is intended for speculative data generators, where the data can
1021  * be regenerated if the page couldn't be grabbed.  This routine should
1022  * be safe to call while holding the lock for another page.
1023  *
1024  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1025  * and deadlock against the caller's locked page.
1026  */
1027 struct page *
1028 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1029 {
1030         struct page *page = find_get_page(mapping, index);
1031
1032         if (page) {
1033                 if (trylock_page(page))
1034                         return page;
1035                 page_cache_release(page);
1036                 return NULL;
1037         }
1038         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1039         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1040                 page_cache_release(page);
1041                 page = NULL;
1042         }
1043         return page;
1044 }
1045 EXPORT_SYMBOL(grab_cache_page_nowait);
1046
1047 /*
1048  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1049  * a _large_ part of the i/o request. Imagine the worst scenario:
1050  *
1051  *      ---R__________________________________________B__________
1052  *         ^ reading here                             ^ bad block(assume 4k)
1053  *
1054  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1055  * => failing the whole request => read(R) => read(R+1) =>
1056  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1057  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1058  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1059  *
1060  * It is going insane. Fix it by quickly scaling down the readahead size.
1061  */
1062 static void shrink_readahead_size_eio(struct file *filp,
1063                                         struct file_ra_state *ra)
1064 {
1065         ra->ra_pages /= 4;
1066 }
1067
1068 /**
1069  * do_generic_file_read - generic file read routine
1070  * @filp:       the file to read
1071  * @ppos:       current file position
1072  * @desc:       read_descriptor
1073  * @actor:      read method
1074  *
1075  * This is a generic file read routine, and uses the
1076  * mapping->a_ops->readpage() function for the actual low-level stuff.
1077  *
1078  * This is really ugly. But the goto's actually try to clarify some
1079  * of the logic when it comes to error handling etc.
1080  */
1081 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1082                 read_descriptor_t *desc, read_actor_t actor)
1083 {
1084         struct address_space *mapping = filp->f_mapping;
1085         struct inode *inode = mapping->host;
1086         struct file_ra_state *ra = &filp->f_ra;
1087         pgoff_t index;
1088         pgoff_t last_index;
1089         pgoff_t prev_index;
1090         unsigned long offset;      /* offset into pagecache page */
1091         unsigned int prev_offset;
1092         int error;
1093
1094         index = *ppos >> PAGE_CACHE_SHIFT;
1095         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1096         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1097         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1098         offset = *ppos & ~PAGE_CACHE_MASK;
1099
1100         for (;;) {
1101                 struct page *page;
1102                 pgoff_t end_index;
1103                 loff_t isize;
1104                 unsigned long nr, ret;
1105
1106                 cond_resched();
1107 find_page:
1108                 page = find_get_page(mapping, index);
1109                 if (!page) {
1110                         page_cache_sync_readahead(mapping,
1111                                         ra, filp,
1112                                         index, last_index - index);
1113                         page = find_get_page(mapping, index);
1114                         if (unlikely(page == NULL))
1115                                 goto no_cached_page;
1116                 }
1117                 if (PageReadahead(page)) {
1118                         page_cache_async_readahead(mapping,
1119                                         ra, filp, page,
1120                                         index, last_index - index);
1121                 }
1122                 if (!PageUptodate(page)) {
1123                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1124                                         !mapping->a_ops->is_partially_uptodate)
1125                                 goto page_not_up_to_date;
1126                         if (!trylock_page(page))
1127                                 goto page_not_up_to_date;
1128                         /* Did it get truncated before we got the lock? */
1129                         if (!page->mapping)
1130                                 goto page_not_up_to_date_locked;
1131                         if (!mapping->a_ops->is_partially_uptodate(page,
1132                                                                 desc, offset))
1133                                 goto page_not_up_to_date_locked;
1134                         unlock_page(page);
1135                 }
1136 page_ok:
1137                 /*
1138                  * i_size must be checked after we know the page is Uptodate.
1139                  *
1140                  * Checking i_size after the check allows us to calculate
1141                  * the correct value for "nr", which means the zero-filled
1142                  * part of the page is not copied back to userspace (unless
1143                  * another truncate extends the file - this is desired though).
1144                  */
1145
1146                 isize = i_size_read(inode);
1147                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1148                 if (unlikely(!isize || index > end_index)) {
1149                         page_cache_release(page);
1150                         goto out;
1151                 }
1152
1153                 /* nr is the maximum number of bytes to copy from this page */
1154                 nr = PAGE_CACHE_SIZE;
1155                 if (index == end_index) {
1156                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1157                         if (nr <= offset) {
1158                                 page_cache_release(page);
1159                                 goto out;
1160                         }
1161                 }
1162                 nr = nr - offset;
1163
1164                 /* If users can be writing to this page using arbitrary
1165                  * virtual addresses, take care about potential aliasing
1166                  * before reading the page on the kernel side.
1167                  */
1168                 if (mapping_writably_mapped(mapping))
1169                         flush_dcache_page(page);
1170
1171                 /*
1172                  * When a sequential read accesses a page several times,
1173                  * only mark it as accessed the first time.
1174                  */
1175                 if (prev_index != index || offset != prev_offset)
1176                         mark_page_accessed(page);
1177                 prev_index = index;
1178
1179                 /*
1180                  * Ok, we have the page, and it's up-to-date, so
1181                  * now we can copy it to user space...
1182                  *
1183                  * The actor routine returns how many bytes were actually used..
1184                  * NOTE! This may not be the same as how much of a user buffer
1185                  * we filled up (we may be padding etc), so we can only update
1186                  * "pos" here (the actor routine has to update the user buffer
1187                  * pointers and the remaining count).
1188                  */
1189                 ret = actor(desc, page, offset, nr);
1190                 offset += ret;
1191                 index += offset >> PAGE_CACHE_SHIFT;
1192                 offset &= ~PAGE_CACHE_MASK;
1193                 prev_offset = offset;
1194
1195                 page_cache_release(page);
1196                 if (ret == nr && desc->count)
1197                         continue;
1198                 goto out;
1199
1200 page_not_up_to_date:
1201                 /* Get exclusive access to the page ... */
1202                 error = lock_page_killable(page);
1203                 if (unlikely(error))
1204                         goto readpage_error;
1205
1206 page_not_up_to_date_locked:
1207                 /* Did it get truncated before we got the lock? */
1208                 if (!page->mapping) {
1209                         unlock_page(page);
1210                         page_cache_release(page);
1211                         continue;
1212                 }
1213
1214                 /* Did somebody else fill it already? */
1215                 if (PageUptodate(page)) {
1216                         unlock_page(page);
1217                         goto page_ok;
1218                 }
1219
1220 readpage:
1221                 /*
1222                  * A previous I/O error may have been due to temporary
1223                  * failures, eg. multipath errors.
1224                  * PG_error will be set again if readpage fails.
1225                  */
1226                 ClearPageError(page);
1227                 /* Start the actual read. The read will unlock the page. */
1228                 error = mapping->a_ops->readpage(filp, page);
1229
1230                 if (unlikely(error)) {
1231                         if (error == AOP_TRUNCATED_PAGE) {
1232                                 page_cache_release(page);
1233                                 goto find_page;
1234                         }
1235                         goto readpage_error;
1236                 }
1237
1238                 if (!PageUptodate(page)) {
1239                         error = lock_page_killable(page);
1240                         if (unlikely(error))
1241                                 goto readpage_error;
1242                         if (!PageUptodate(page)) {
1243                                 if (page->mapping == NULL) {
1244                                         /*
1245                                          * invalidate_mapping_pages got it
1246                                          */
1247                                         unlock_page(page);
1248                                         page_cache_release(page);
1249                                         goto find_page;
1250                                 }
1251                                 unlock_page(page);
1252                                 shrink_readahead_size_eio(filp, ra);
1253                                 error = -EIO;
1254                                 goto readpage_error;
1255                         }
1256                         unlock_page(page);
1257                 }
1258
1259                 goto page_ok;
1260
1261 readpage_error:
1262                 /* UHHUH! A synchronous read error occurred. Report it */
1263                 desc->error = error;
1264                 page_cache_release(page);
1265                 goto out;
1266
1267 no_cached_page:
1268                 /*
1269                  * Ok, it wasn't cached, so we need to create a new
1270                  * page..
1271                  */
1272                 page = page_cache_alloc_cold(mapping);
1273                 if (!page) {
1274                         desc->error = -ENOMEM;
1275                         goto out;
1276                 }
1277                 error = add_to_page_cache_lru(page, mapping,
1278                                                 index, GFP_KERNEL);
1279                 if (error) {
1280                         page_cache_release(page);
1281                         if (error == -EEXIST)
1282                                 goto find_page;
1283                         desc->error = error;
1284                         goto out;
1285                 }
1286                 goto readpage;
1287         }
1288
1289 out:
1290         ra->prev_pos = prev_index;
1291         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1292         ra->prev_pos |= prev_offset;
1293
1294         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1295         file_accessed(filp);
1296 }
1297
1298 int file_read_actor(read_descriptor_t *desc, struct page *page,
1299                         unsigned long offset, unsigned long size)
1300 {
1301         char *kaddr;
1302         unsigned long left, count = desc->count;
1303
1304         if (size > count)
1305                 size = count;
1306
1307         /*
1308          * Faults on the destination of a read are common, so do it before
1309          * taking the kmap.
1310          */
1311         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1312                 kaddr = kmap_atomic(page);
1313                 left = __copy_to_user_inatomic(desc->arg.buf,
1314                                                 kaddr + offset, size);
1315                 kunmap_atomic(kaddr);
1316                 if (left == 0)
1317                         goto success;
1318         }
1319
1320         /* Do it the slow way */
1321         kaddr = kmap(page);
1322         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1323         kunmap(page);
1324
1325         if (left) {
1326                 size -= left;
1327                 desc->error = -EFAULT;
1328         }
1329 success:
1330         desc->count = count - size;
1331         desc->written += size;
1332         desc->arg.buf += size;
1333         return size;
1334 }
1335
1336 /*
1337  * Performs necessary checks before doing a write
1338  * @iov:        io vector request
1339  * @nr_segs:    number of segments in the iovec
1340  * @count:      number of bytes to write
1341  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1342  *
1343  * Adjust number of segments and amount of bytes to write (nr_segs should be
1344  * properly initialized first). Returns appropriate error code that caller
1345  * should return or zero in case that write should be allowed.
1346  */
1347 int generic_segment_checks(const struct iovec *iov,
1348                         unsigned long *nr_segs, size_t *count, int access_flags)
1349 {
1350         unsigned long   seg;
1351         size_t cnt = 0;
1352         for (seg = 0; seg < *nr_segs; seg++) {
1353                 const struct iovec *iv = &iov[seg];
1354
1355                 /*
1356                  * If any segment has a negative length, or the cumulative
1357                  * length ever wraps negative then return -EINVAL.
1358                  */
1359                 cnt += iv->iov_len;
1360                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1361                         return -EINVAL;
1362                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1363                         continue;
1364                 if (seg == 0)
1365                         return -EFAULT;
1366                 *nr_segs = seg;
1367                 cnt -= iv->iov_len;     /* This segment is no good */
1368                 break;
1369         }
1370         *count = cnt;
1371         return 0;
1372 }
1373 EXPORT_SYMBOL(generic_segment_checks);
1374
1375 /**
1376  * generic_file_aio_read - generic filesystem read routine
1377  * @iocb:       kernel I/O control block
1378  * @iov:        io vector request
1379  * @nr_segs:    number of segments in the iovec
1380  * @pos:        current file position
1381  *
1382  * This is the "read()" routine for all filesystems
1383  * that can use the page cache directly.
1384  */
1385 ssize_t
1386 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1387                 unsigned long nr_segs, loff_t pos)
1388 {
1389         struct file *filp = iocb->ki_filp;
1390         ssize_t retval;
1391         unsigned long seg = 0;
1392         size_t count;
1393         loff_t *ppos = &iocb->ki_pos;
1394
1395         count = 0;
1396         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1397         if (retval)
1398                 return retval;
1399
1400         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1401         if (filp->f_flags & O_DIRECT) {
1402                 loff_t size;
1403                 struct address_space *mapping;
1404                 struct inode *inode;
1405
1406                 mapping = filp->f_mapping;
1407                 inode = mapping->host;
1408                 if (!count)
1409                         goto out; /* skip atime */
1410                 size = i_size_read(inode);
1411                 if (pos < size) {
1412                         retval = filemap_write_and_wait_range(mapping, pos,
1413                                         pos + iov_length(iov, nr_segs) - 1);
1414                         if (!retval) {
1415                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1416                                                         iov, pos, nr_segs);
1417                         }
1418                         if (retval > 0) {
1419                                 *ppos = pos + retval;
1420                                 count -= retval;
1421                         }
1422
1423                         /*
1424                          * Btrfs can have a short DIO read if we encounter
1425                          * compressed extents, so if there was an error, or if
1426                          * we've already read everything we wanted to, or if
1427                          * there was a short read because we hit EOF, go ahead
1428                          * and return.  Otherwise fallthrough to buffered io for
1429                          * the rest of the read.
1430                          */
1431                         if (retval < 0 || !count || *ppos >= size) {
1432                                 file_accessed(filp);
1433                                 goto out;
1434                         }
1435                 }
1436         }
1437
1438         count = retval;
1439         for (seg = 0; seg < nr_segs; seg++) {
1440                 read_descriptor_t desc;
1441                 loff_t offset = 0;
1442
1443                 /*
1444                  * If we did a short DIO read we need to skip the section of the
1445                  * iov that we've already read data into.
1446                  */
1447                 if (count) {
1448                         if (count > iov[seg].iov_len) {
1449                                 count -= iov[seg].iov_len;
1450                                 continue;
1451                         }
1452                         offset = count;
1453                         count = 0;
1454                 }
1455
1456                 desc.written = 0;
1457                 desc.arg.buf = iov[seg].iov_base + offset;
1458                 desc.count = iov[seg].iov_len - offset;
1459                 if (desc.count == 0)
1460                         continue;
1461                 desc.error = 0;
1462                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1463                 retval += desc.written;
1464                 if (desc.error) {
1465                         retval = retval ?: desc.error;
1466                         break;
1467                 }
1468                 if (desc.count > 0)
1469                         break;
1470         }
1471 out:
1472         return retval;
1473 }
1474 EXPORT_SYMBOL(generic_file_aio_read);
1475
1476 #ifdef CONFIG_MMU
1477 /**
1478  * page_cache_read - adds requested page to the page cache if not already there
1479  * @file:       file to read
1480  * @offset:     page index
1481  *
1482  * This adds the requested page to the page cache if it isn't already there,
1483  * and schedules an I/O to read in its contents from disk.
1484  */
1485 static int page_cache_read(struct file *file, pgoff_t offset)
1486 {
1487         struct address_space *mapping = file->f_mapping;
1488         struct page *page; 
1489         int ret;
1490
1491         do {
1492                 page = page_cache_alloc_cold(mapping);
1493                 if (!page)
1494                         return -ENOMEM;
1495
1496                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1497                 if (ret == 0)
1498                         ret = mapping->a_ops->readpage(file, page);
1499                 else if (ret == -EEXIST)
1500                         ret = 0; /* losing race to add is OK */
1501
1502                 page_cache_release(page);
1503
1504         } while (ret == AOP_TRUNCATED_PAGE);
1505                 
1506         return ret;
1507 }
1508
1509 #define MMAP_LOTSAMISS  (100)
1510
1511 /*
1512  * Synchronous readahead happens when we don't even find
1513  * a page in the page cache at all.
1514  */
1515 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1516                                    struct file_ra_state *ra,
1517                                    struct file *file,
1518                                    pgoff_t offset)
1519 {
1520         unsigned long ra_pages;
1521         struct address_space *mapping = file->f_mapping;
1522
1523         /* If we don't want any read-ahead, don't bother */
1524         if (VM_RandomReadHint(vma))
1525                 return;
1526         if (!ra->ra_pages)
1527                 return;
1528
1529         if (VM_SequentialReadHint(vma)) {
1530                 page_cache_sync_readahead(mapping, ra, file, offset,
1531                                           ra->ra_pages);
1532                 return;
1533         }
1534
1535         /* Avoid banging the cache line if not needed */
1536         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1537                 ra->mmap_miss++;
1538
1539         /*
1540          * Do we miss much more than hit in this file? If so,
1541          * stop bothering with read-ahead. It will only hurt.
1542          */
1543         if (ra->mmap_miss > MMAP_LOTSAMISS)
1544                 return;
1545
1546         /*
1547          * mmap read-around
1548          */
1549         ra_pages = max_sane_readahead(ra->ra_pages);
1550         ra->start = max_t(long, 0, offset - ra_pages / 2);
1551         ra->size = ra_pages;
1552         ra->async_size = ra_pages / 4;
1553         ra_submit(ra, mapping, file);
1554 }
1555
1556 /*
1557  * Asynchronous readahead happens when we find the page and PG_readahead,
1558  * so we want to possibly extend the readahead further..
1559  */
1560 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1561                                     struct file_ra_state *ra,
1562                                     struct file *file,
1563                                     struct page *page,
1564                                     pgoff_t offset)
1565 {
1566         struct address_space *mapping = file->f_mapping;
1567
1568         /* If we don't want any read-ahead, don't bother */
1569         if (VM_RandomReadHint(vma))
1570                 return;
1571         if (ra->mmap_miss > 0)
1572                 ra->mmap_miss--;
1573         if (PageReadahead(page))
1574                 page_cache_async_readahead(mapping, ra, file,
1575                                            page, offset, ra->ra_pages);
1576 }
1577
1578 /**
1579  * filemap_fault - read in file data for page fault handling
1580  * @vma:        vma in which the fault was taken
1581  * @vmf:        struct vm_fault containing details of the fault
1582  *
1583  * filemap_fault() is invoked via the vma operations vector for a
1584  * mapped memory region to read in file data during a page fault.
1585  *
1586  * The goto's are kind of ugly, but this streamlines the normal case of having
1587  * it in the page cache, and handles the special cases reasonably without
1588  * having a lot of duplicated code.
1589  */
1590 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1591 {
1592         int error;
1593         struct file *file = vma->vm_file;
1594         struct address_space *mapping = file->f_mapping;
1595         struct file_ra_state *ra = &file->f_ra;
1596         struct inode *inode = mapping->host;
1597         pgoff_t offset = vmf->pgoff;
1598         struct page *page;
1599         pgoff_t size;
1600         int ret = 0;
1601
1602         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1603         if (offset >= size)
1604                 return VM_FAULT_SIGBUS;
1605
1606         /*
1607          * Do we have something in the page cache already?
1608          */
1609         page = find_get_page(mapping, offset);
1610         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1611                 /*
1612                  * We found the page, so try async readahead before
1613                  * waiting for the lock.
1614                  */
1615                 do_async_mmap_readahead(vma, ra, file, page, offset);
1616         } else if (!page) {
1617                 /* No page in the page cache at all */
1618                 do_sync_mmap_readahead(vma, ra, file, offset);
1619                 count_vm_event(PGMAJFAULT);
1620                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1621                 ret = VM_FAULT_MAJOR;
1622 retry_find:
1623                 page = find_get_page(mapping, offset);
1624                 if (!page)
1625                         goto no_cached_page;
1626         }
1627
1628         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1629                 page_cache_release(page);
1630                 return ret | VM_FAULT_RETRY;
1631         }
1632
1633         /* Did it get truncated? */
1634         if (unlikely(page->mapping != mapping)) {
1635                 unlock_page(page);
1636                 put_page(page);
1637                 goto retry_find;
1638         }
1639         VM_BUG_ON(page->index != offset);
1640
1641         /*
1642          * We have a locked page in the page cache, now we need to check
1643          * that it's up-to-date. If not, it is going to be due to an error.
1644          */
1645         if (unlikely(!PageUptodate(page)))
1646                 goto page_not_uptodate;
1647
1648         /*
1649          * Found the page and have a reference on it.
1650          * We must recheck i_size under page lock.
1651          */
1652         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1653         if (unlikely(offset >= size)) {
1654                 unlock_page(page);
1655                 page_cache_release(page);
1656                 return VM_FAULT_SIGBUS;
1657         }
1658
1659         vmf->page = page;
1660         return ret | VM_FAULT_LOCKED;
1661
1662 no_cached_page:
1663         /*
1664          * We're only likely to ever get here if MADV_RANDOM is in
1665          * effect.
1666          */
1667         error = page_cache_read(file, offset);
1668
1669         /*
1670          * The page we want has now been added to the page cache.
1671          * In the unlikely event that someone removed it in the
1672          * meantime, we'll just come back here and read it again.
1673          */
1674         if (error >= 0)
1675                 goto retry_find;
1676
1677         /*
1678          * An error return from page_cache_read can result if the
1679          * system is low on memory, or a problem occurs while trying
1680          * to schedule I/O.
1681          */
1682         if (error == -ENOMEM)
1683                 return VM_FAULT_OOM;
1684         return VM_FAULT_SIGBUS;
1685
1686 page_not_uptodate:
1687         /*
1688          * Umm, take care of errors if the page isn't up-to-date.
1689          * Try to re-read it _once_. We do this synchronously,
1690          * because there really aren't any performance issues here
1691          * and we need to check for errors.
1692          */
1693         ClearPageError(page);
1694         error = mapping->a_ops->readpage(file, page);
1695         if (!error) {
1696                 wait_on_page_locked(page);
1697                 if (!PageUptodate(page))
1698                         error = -EIO;
1699         }
1700         page_cache_release(page);
1701
1702         if (!error || error == AOP_TRUNCATED_PAGE)
1703                 goto retry_find;
1704
1705         /* Things didn't work out. Return zero to tell the mm layer so. */
1706         shrink_readahead_size_eio(file, ra);
1707         return VM_FAULT_SIGBUS;
1708 }
1709 EXPORT_SYMBOL(filemap_fault);
1710
1711 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1712 {
1713         struct page *page = vmf->page;
1714         struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1715         int ret = VM_FAULT_LOCKED;
1716
1717         sb_start_pagefault(inode->i_sb);
1718         file_update_time(vma->vm_file);
1719         lock_page(page);
1720         if (page->mapping != inode->i_mapping) {
1721                 unlock_page(page);
1722                 ret = VM_FAULT_NOPAGE;
1723                 goto out;
1724         }
1725         /*
1726          * We mark the page dirty already here so that when freeze is in
1727          * progress, we are guaranteed that writeback during freezing will
1728          * see the dirty page and writeprotect it again.
1729          */
1730         set_page_dirty(page);
1731 out:
1732         sb_end_pagefault(inode->i_sb);
1733         return ret;
1734 }
1735 EXPORT_SYMBOL(filemap_page_mkwrite);
1736
1737 const struct vm_operations_struct generic_file_vm_ops = {
1738         .fault          = filemap_fault,
1739         .page_mkwrite   = filemap_page_mkwrite,
1740         .remap_pages    = generic_file_remap_pages,
1741 };
1742
1743 /* This is used for a general mmap of a disk file */
1744
1745 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1746 {
1747         struct address_space *mapping = file->f_mapping;
1748
1749         if (!mapping->a_ops->readpage)
1750                 return -ENOEXEC;
1751         file_accessed(file);
1752         vma->vm_ops = &generic_file_vm_ops;
1753         return 0;
1754 }
1755
1756 /*
1757  * This is for filesystems which do not implement ->writepage.
1758  */
1759 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1760 {
1761         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1762                 return -EINVAL;
1763         return generic_file_mmap(file, vma);
1764 }
1765 #else
1766 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1767 {
1768         return -ENOSYS;
1769 }
1770 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1771 {
1772         return -ENOSYS;
1773 }
1774 #endif /* CONFIG_MMU */
1775
1776 EXPORT_SYMBOL(generic_file_mmap);
1777 EXPORT_SYMBOL(generic_file_readonly_mmap);
1778
1779 static struct page *__read_cache_page(struct address_space *mapping,
1780                                 pgoff_t index,
1781                                 int (*filler)(void *, struct page *),
1782                                 void *data,
1783                                 gfp_t gfp)
1784 {
1785         struct page *page;
1786         int err;
1787 repeat:
1788         page = find_get_page(mapping, index);
1789         if (!page) {
1790                 page = __page_cache_alloc(gfp | __GFP_COLD);
1791                 if (!page)
1792                         return ERR_PTR(-ENOMEM);
1793                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1794                 if (unlikely(err)) {
1795                         page_cache_release(page);
1796                         if (err == -EEXIST)
1797                                 goto repeat;
1798                         /* Presumably ENOMEM for radix tree node */
1799                         return ERR_PTR(err);
1800                 }
1801                 err = filler(data, page);
1802                 if (err < 0) {
1803                         page_cache_release(page);
1804                         page = ERR_PTR(err);
1805                 }
1806         }
1807         return page;
1808 }
1809
1810 static struct page *do_read_cache_page(struct address_space *mapping,
1811                                 pgoff_t index,
1812                                 int (*filler)(void *, struct page *),
1813                                 void *data,
1814                                 gfp_t gfp)
1815
1816 {
1817         struct page *page;
1818         int err;
1819
1820 retry:
1821         page = __read_cache_page(mapping, index, filler, data, gfp);
1822         if (IS_ERR(page))
1823                 return page;
1824         if (PageUptodate(page))
1825                 goto out;
1826
1827         lock_page(page);
1828         if (!page->mapping) {
1829                 unlock_page(page);
1830                 page_cache_release(page);
1831                 goto retry;
1832         }
1833         if (PageUptodate(page)) {
1834                 unlock_page(page);
1835                 goto out;
1836         }
1837         err = filler(data, page);
1838         if (err < 0) {
1839                 page_cache_release(page);
1840                 return ERR_PTR(err);
1841         }
1842 out:
1843         mark_page_accessed(page);
1844         return page;
1845 }
1846
1847 /**
1848  * read_cache_page_async - read into page cache, fill it if needed
1849  * @mapping:    the page's address_space
1850  * @index:      the page index
1851  * @filler:     function to perform the read
1852  * @data:       first arg to filler(data, page) function, often left as NULL
1853  *
1854  * Same as read_cache_page, but don't wait for page to become unlocked
1855  * after submitting it to the filler.
1856  *
1857  * Read into the page cache. If a page already exists, and PageUptodate() is
1858  * not set, try to fill the page but don't wait for it to become unlocked.
1859  *
1860  * If the page does not get brought uptodate, return -EIO.
1861  */
1862 struct page *read_cache_page_async(struct address_space *mapping,
1863                                 pgoff_t index,
1864                                 int (*filler)(void *, struct page *),
1865                                 void *data)
1866 {
1867         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1868 }
1869 EXPORT_SYMBOL(read_cache_page_async);
1870
1871 static struct page *wait_on_page_read(struct page *page)
1872 {
1873         if (!IS_ERR(page)) {
1874                 wait_on_page_locked(page);
1875                 if (!PageUptodate(page)) {
1876                         page_cache_release(page);
1877                         page = ERR_PTR(-EIO);
1878                 }
1879         }
1880         return page;
1881 }
1882
1883 /**
1884  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1885  * @mapping:    the page's address_space
1886  * @index:      the page index
1887  * @gfp:        the page allocator flags to use if allocating
1888  *
1889  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1890  * any new page allocations done using the specified allocation flags.
1891  *
1892  * If the page does not get brought uptodate, return -EIO.
1893  */
1894 struct page *read_cache_page_gfp(struct address_space *mapping,
1895                                 pgoff_t index,
1896                                 gfp_t gfp)
1897 {
1898         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1899
1900         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1901 }
1902 EXPORT_SYMBOL(read_cache_page_gfp);
1903
1904 /**
1905  * read_cache_page - read into page cache, fill it if needed
1906  * @mapping:    the page's address_space
1907  * @index:      the page index
1908  * @filler:     function to perform the read
1909  * @data:       first arg to filler(data, page) function, often left as NULL
1910  *
1911  * Read into the page cache. If a page already exists, and PageUptodate() is
1912  * not set, try to fill the page then wait for it to become unlocked.
1913  *
1914  * If the page does not get brought uptodate, return -EIO.
1915  */
1916 struct page *read_cache_page(struct address_space *mapping,
1917                                 pgoff_t index,
1918                                 int (*filler)(void *, struct page *),
1919                                 void *data)
1920 {
1921         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1922 }
1923 EXPORT_SYMBOL(read_cache_page);
1924
1925 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1926                         const struct iovec *iov, size_t base, size_t bytes)
1927 {
1928         size_t copied = 0, left = 0;
1929
1930         while (bytes) {
1931                 char __user *buf = iov->iov_base + base;
1932                 int copy = min(bytes, iov->iov_len - base);
1933
1934                 base = 0;
1935                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1936                 copied += copy;
1937                 bytes -= copy;
1938                 vaddr += copy;
1939                 iov++;
1940
1941                 if (unlikely(left))
1942                         break;
1943         }
1944         return copied - left;
1945 }
1946
1947 /*
1948  * Copy as much as we can into the page and return the number of bytes which
1949  * were successfully copied.  If a fault is encountered then return the number of
1950  * bytes which were copied.
1951  */
1952 size_t iov_iter_copy_from_user_atomic(struct page *page,
1953                 struct iov_iter *i, unsigned long offset, size_t bytes)
1954 {
1955         char *kaddr;
1956         size_t copied;
1957
1958         BUG_ON(!in_atomic());
1959         kaddr = kmap_atomic(page);
1960         if (likely(i->nr_segs == 1)) {
1961                 int left;
1962                 char __user *buf = i->iov->iov_base + i->iov_offset;
1963                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1964                 copied = bytes - left;
1965         } else {
1966                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1967                                                 i->iov, i->iov_offset, bytes);
1968         }
1969         kunmap_atomic(kaddr);
1970
1971         return copied;
1972 }
1973 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1974
1975 /*
1976  * This has the same sideeffects and return value as
1977  * iov_iter_copy_from_user_atomic().
1978  * The difference is that it attempts to resolve faults.
1979  * Page must not be locked.
1980  */
1981 size_t iov_iter_copy_from_user(struct page *page,
1982                 struct iov_iter *i, unsigned long offset, size_t bytes)
1983 {
1984         char *kaddr;
1985         size_t copied;
1986
1987         kaddr = kmap(page);
1988         if (likely(i->nr_segs == 1)) {
1989                 int left;
1990                 char __user *buf = i->iov->iov_base + i->iov_offset;
1991                 left = __copy_from_user(kaddr + offset, buf, bytes);
1992                 copied = bytes - left;
1993         } else {
1994                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1995                                                 i->iov, i->iov_offset, bytes);
1996         }
1997         kunmap(page);
1998         return copied;
1999 }
2000 EXPORT_SYMBOL(iov_iter_copy_from_user);
2001
2002 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2003 {
2004         BUG_ON(i->count < bytes);
2005
2006         if (likely(i->nr_segs == 1)) {
2007                 i->iov_offset += bytes;
2008                 i->count -= bytes;
2009         } else {
2010                 const struct iovec *iov = i->iov;
2011                 size_t base = i->iov_offset;
2012                 unsigned long nr_segs = i->nr_segs;
2013
2014                 /*
2015                  * The !iov->iov_len check ensures we skip over unlikely
2016                  * zero-length segments (without overruning the iovec).
2017                  */
2018                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2019                         int copy;
2020
2021                         copy = min(bytes, iov->iov_len - base);
2022                         BUG_ON(!i->count || i->count < copy);
2023                         i->count -= copy;
2024                         bytes -= copy;
2025                         base += copy;
2026                         if (iov->iov_len == base) {
2027                                 iov++;
2028                                 nr_segs--;
2029                                 base = 0;
2030                         }
2031                 }
2032                 i->iov = iov;
2033                 i->iov_offset = base;
2034                 i->nr_segs = nr_segs;
2035         }
2036 }
2037 EXPORT_SYMBOL(iov_iter_advance);
2038
2039 /*
2040  * Fault in the first iovec of the given iov_iter, to a maximum length
2041  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2042  * accessed (ie. because it is an invalid address).
2043  *
2044  * writev-intensive code may want this to prefault several iovecs -- that
2045  * would be possible (callers must not rely on the fact that _only_ the
2046  * first iovec will be faulted with the current implementation).
2047  */
2048 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2049 {
2050         char __user *buf = i->iov->iov_base + i->iov_offset;
2051         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2052         return fault_in_pages_readable(buf, bytes);
2053 }
2054 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2055
2056 /*
2057  * Return the count of just the current iov_iter segment.
2058  */
2059 size_t iov_iter_single_seg_count(struct iov_iter *i)
2060 {
2061         const struct iovec *iov = i->iov;
2062         if (i->nr_segs == 1)
2063                 return i->count;
2064         else
2065                 return min(i->count, iov->iov_len - i->iov_offset);
2066 }
2067 EXPORT_SYMBOL(iov_iter_single_seg_count);
2068
2069 /*
2070  * Performs necessary checks before doing a write
2071  *
2072  * Can adjust writing position or amount of bytes to write.
2073  * Returns appropriate error code that caller should return or
2074  * zero in case that write should be allowed.
2075  */
2076 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2077 {
2078         struct inode *inode = file->f_mapping->host;
2079         unsigned long limit = rlimit(RLIMIT_FSIZE);
2080
2081         if (unlikely(*pos < 0))
2082                 return -EINVAL;
2083
2084         if (!isblk) {
2085                 /* FIXME: this is for backwards compatibility with 2.4 */
2086                 if (file->f_flags & O_APPEND)
2087                         *pos = i_size_read(inode);
2088
2089                 if (limit != RLIM_INFINITY) {
2090                         if (*pos >= limit) {
2091                                 send_sig(SIGXFSZ, current, 0);
2092                                 return -EFBIG;
2093                         }
2094                         if (*count > limit - (typeof(limit))*pos) {
2095                                 *count = limit - (typeof(limit))*pos;
2096                         }
2097                 }
2098         }
2099
2100         /*
2101          * LFS rule
2102          */
2103         if (unlikely(*pos + *count > MAX_NON_LFS &&
2104                                 !(file->f_flags & O_LARGEFILE))) {
2105                 if (*pos >= MAX_NON_LFS) {
2106                         return -EFBIG;
2107                 }
2108                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2109                         *count = MAX_NON_LFS - (unsigned long)*pos;
2110                 }
2111         }
2112
2113         /*
2114          * Are we about to exceed the fs block limit ?
2115          *
2116          * If we have written data it becomes a short write.  If we have
2117          * exceeded without writing data we send a signal and return EFBIG.
2118          * Linus frestrict idea will clean these up nicely..
2119          */
2120         if (likely(!isblk)) {
2121                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2122                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2123                                 return -EFBIG;
2124                         }
2125                         /* zero-length writes at ->s_maxbytes are OK */
2126                 }
2127
2128                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2129                         *count = inode->i_sb->s_maxbytes - *pos;
2130         } else {
2131 #ifdef CONFIG_BLOCK
2132                 loff_t isize;
2133                 if (bdev_read_only(I_BDEV(inode)))
2134                         return -EPERM;
2135                 isize = i_size_read(inode);
2136                 if (*pos >= isize) {
2137                         if (*count || *pos > isize)
2138                                 return -ENOSPC;
2139                 }
2140
2141                 if (*pos + *count > isize)
2142                         *count = isize - *pos;
2143 #else
2144                 return -EPERM;
2145 #endif
2146         }
2147         return 0;
2148 }
2149 EXPORT_SYMBOL(generic_write_checks);
2150
2151 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2152                                 loff_t pos, unsigned len, unsigned flags,
2153                                 struct page **pagep, void **fsdata)
2154 {
2155         const struct address_space_operations *aops = mapping->a_ops;
2156
2157         return aops->write_begin(file, mapping, pos, len, flags,
2158                                                         pagep, fsdata);
2159 }
2160 EXPORT_SYMBOL(pagecache_write_begin);
2161
2162 int pagecache_write_end(struct file *file, struct address_space *mapping,
2163                                 loff_t pos, unsigned len, unsigned copied,
2164                                 struct page *page, void *fsdata)
2165 {
2166         const struct address_space_operations *aops = mapping->a_ops;
2167
2168         mark_page_accessed(page);
2169         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2170 }
2171 EXPORT_SYMBOL(pagecache_write_end);
2172
2173 ssize_t
2174 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2175                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2176                 size_t count, size_t ocount)
2177 {
2178         struct file     *file = iocb->ki_filp;
2179         struct address_space *mapping = file->f_mapping;
2180         struct inode    *inode = mapping->host;
2181         ssize_t         written;
2182         size_t          write_len;
2183         pgoff_t         end;
2184
2185         if (count != ocount)
2186                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2187
2188         write_len = iov_length(iov, *nr_segs);
2189         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2190
2191         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2192         if (written)
2193                 goto out;
2194
2195         /*
2196          * After a write we want buffered reads to be sure to go to disk to get
2197          * the new data.  We invalidate clean cached page from the region we're
2198          * about to write.  We do this *before* the write so that we can return
2199          * without clobbering -EIOCBQUEUED from ->direct_IO().
2200          */
2201         if (mapping->nrpages) {
2202                 written = invalidate_inode_pages2_range(mapping,
2203                                         pos >> PAGE_CACHE_SHIFT, end);
2204                 /*
2205                  * If a page can not be invalidated, return 0 to fall back
2206                  * to buffered write.
2207                  */
2208                 if (written) {
2209                         if (written == -EBUSY)
2210                                 return 0;
2211                         goto out;
2212                 }
2213         }
2214
2215         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2216
2217         /*
2218          * Finally, try again to invalidate clean pages which might have been
2219          * cached by non-direct readahead, or faulted in by get_user_pages()
2220          * if the source of the write was an mmap'ed region of the file
2221          * we're writing.  Either one is a pretty crazy thing to do,
2222          * so we don't support it 100%.  If this invalidation
2223          * fails, tough, the write still worked...
2224          */
2225         if (mapping->nrpages) {
2226                 invalidate_inode_pages2_range(mapping,
2227                                               pos >> PAGE_CACHE_SHIFT, end);
2228         }
2229
2230         if (written > 0) {
2231                 pos += written;
2232                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2233                         i_size_write(inode, pos);
2234                         mark_inode_dirty(inode);
2235                 }
2236                 *ppos = pos;
2237         }
2238 out:
2239         return written;
2240 }
2241 EXPORT_SYMBOL(generic_file_direct_write);
2242
2243 /*
2244  * Find or create a page at the given pagecache position. Return the locked
2245  * page. This function is specifically for buffered writes.
2246  */
2247 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2248                                         pgoff_t index, unsigned flags)
2249 {
2250         int status;
2251         gfp_t gfp_mask;
2252         struct page *page;
2253         gfp_t gfp_notmask = 0;
2254
2255         gfp_mask = mapping_gfp_mask(mapping);
2256         if (mapping_cap_account_dirty(mapping))
2257                 gfp_mask |= __GFP_WRITE;
2258         if (flags & AOP_FLAG_NOFS)
2259                 gfp_notmask = __GFP_FS;
2260 repeat:
2261         page = find_lock_page(mapping, index);
2262         if (page)
2263                 goto found;
2264
2265         page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2266         if (!page)
2267                 return NULL;
2268         status = add_to_page_cache_lru(page, mapping, index,
2269                                                 GFP_KERNEL & ~gfp_notmask);
2270         if (unlikely(status)) {
2271                 page_cache_release(page);
2272                 if (status == -EEXIST)
2273                         goto repeat;
2274                 return NULL;
2275         }
2276 found:
2277         wait_on_page_writeback(page);
2278         return page;
2279 }
2280 EXPORT_SYMBOL(grab_cache_page_write_begin);
2281
2282 static ssize_t generic_perform_write(struct file *file,
2283                                 struct iov_iter *i, loff_t pos)
2284 {
2285         struct address_space *mapping = file->f_mapping;
2286         const struct address_space_operations *a_ops = mapping->a_ops;
2287         long status = 0;
2288         ssize_t written = 0;
2289         unsigned int flags = 0;
2290
2291         /*
2292          * Copies from kernel address space cannot fail (NFSD is a big user).
2293          */
2294         if (segment_eq(get_fs(), KERNEL_DS))
2295                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2296
2297         do {
2298                 struct page *page;
2299                 unsigned long offset;   /* Offset into pagecache page */
2300                 unsigned long bytes;    /* Bytes to write to page */
2301                 size_t copied;          /* Bytes copied from user */
2302                 void *fsdata;
2303
2304                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2305                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2306                                                 iov_iter_count(i));
2307
2308 again:
2309                 /*
2310                  * Bring in the user page that we will copy from _first_.
2311                  * Otherwise there's a nasty deadlock on copying from the
2312                  * same page as we're writing to, without it being marked
2313                  * up-to-date.
2314                  *
2315                  * Not only is this an optimisation, but it is also required
2316                  * to check that the address is actually valid, when atomic
2317                  * usercopies are used, below.
2318                  */
2319                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2320                         status = -EFAULT;
2321                         break;
2322                 }
2323
2324                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2325                                                 &page, &fsdata);
2326                 if (unlikely(status))
2327                         break;
2328
2329                 if (mapping_writably_mapped(mapping))
2330                         flush_dcache_page(page);
2331
2332                 pagefault_disable();
2333                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2334                 pagefault_enable();
2335                 flush_dcache_page(page);
2336
2337                 mark_page_accessed(page);
2338                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2339                                                 page, fsdata);
2340                 if (unlikely(status < 0))
2341                         break;
2342                 copied = status;
2343
2344                 cond_resched();
2345
2346                 iov_iter_advance(i, copied);
2347                 if (unlikely(copied == 0)) {
2348                         /*
2349                          * If we were unable to copy any data at all, we must
2350                          * fall back to a single segment length write.
2351                          *
2352                          * If we didn't fallback here, we could livelock
2353                          * because not all segments in the iov can be copied at
2354                          * once without a pagefault.
2355                          */
2356                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2357                                                 iov_iter_single_seg_count(i));
2358                         goto again;
2359                 }
2360                 pos += copied;
2361                 written += copied;
2362
2363                 balance_dirty_pages_ratelimited(mapping);
2364                 if (fatal_signal_pending(current)) {
2365                         status = -EINTR;
2366                         break;
2367                 }
2368         } while (iov_iter_count(i));
2369
2370         return written ? written : status;
2371 }
2372
2373 ssize_t
2374 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2375                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2376                 size_t count, ssize_t written)
2377 {
2378         struct file *file = iocb->ki_filp;
2379         ssize_t status;
2380         struct iov_iter i;
2381
2382         iov_iter_init(&i, iov, nr_segs, count, written);
2383         status = generic_perform_write(file, &i, pos);
2384
2385         if (likely(status >= 0)) {
2386                 written += status;
2387                 *ppos = pos + status;
2388         }
2389         
2390         return written ? written : status;
2391 }
2392 EXPORT_SYMBOL(generic_file_buffered_write);
2393
2394 /**
2395  * __generic_file_aio_write - write data to a file
2396  * @iocb:       IO state structure (file, offset, etc.)
2397  * @iov:        vector with data to write
2398  * @nr_segs:    number of segments in the vector
2399  * @ppos:       position where to write
2400  *
2401  * This function does all the work needed for actually writing data to a
2402  * file. It does all basic checks, removes SUID from the file, updates
2403  * modification times and calls proper subroutines depending on whether we
2404  * do direct IO or a standard buffered write.
2405  *
2406  * It expects i_mutex to be grabbed unless we work on a block device or similar
2407  * object which does not need locking at all.
2408  *
2409  * This function does *not* take care of syncing data in case of O_SYNC write.
2410  * A caller has to handle it. This is mainly due to the fact that we want to
2411  * avoid syncing under i_mutex.
2412  */
2413 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2414                                  unsigned long nr_segs, loff_t *ppos)
2415 {
2416         struct file *file = iocb->ki_filp;
2417         struct address_space * mapping = file->f_mapping;
2418         size_t ocount;          /* original count */
2419         size_t count;           /* after file limit checks */
2420         struct inode    *inode = mapping->host;
2421         loff_t          pos;
2422         ssize_t         written;
2423         ssize_t         err;
2424
2425         ocount = 0;
2426         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2427         if (err)
2428                 return err;
2429
2430         count = ocount;
2431         pos = *ppos;
2432
2433         /* We can write back this queue in page reclaim */
2434         current->backing_dev_info = mapping->backing_dev_info;
2435         written = 0;
2436
2437         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2438         if (err)
2439                 goto out;
2440
2441         if (count == 0)
2442                 goto out;
2443
2444         err = file_remove_suid(file);
2445         if (err)
2446                 goto out;
2447
2448         err = file_update_time(file);
2449         if (err)
2450                 goto out;
2451
2452         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2453         if (unlikely(file->f_flags & O_DIRECT)) {
2454                 loff_t endbyte;
2455                 ssize_t written_buffered;
2456
2457                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2458                                                         ppos, count, ocount);
2459                 if (written < 0 || written == count)
2460                         goto out;
2461                 /*
2462                  * direct-io write to a hole: fall through to buffered I/O
2463                  * for completing the rest of the request.
2464                  */
2465                 pos += written;
2466                 count -= written;
2467                 written_buffered = generic_file_buffered_write(iocb, iov,
2468                                                 nr_segs, pos, ppos, count,
2469                                                 written);
2470                 /*
2471                  * If generic_file_buffered_write() retuned a synchronous error
2472                  * then we want to return the number of bytes which were
2473                  * direct-written, or the error code if that was zero.  Note
2474                  * that this differs from normal direct-io semantics, which
2475                  * will return -EFOO even if some bytes were written.
2476                  */
2477                 if (written_buffered < 0) {
2478                         err = written_buffered;
2479                         goto out;
2480                 }
2481
2482                 /*
2483                  * We need to ensure that the page cache pages are written to
2484                  * disk and invalidated to preserve the expected O_DIRECT
2485                  * semantics.
2486                  */
2487                 endbyte = pos + written_buffered - written - 1;
2488                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2489                 if (err == 0) {
2490                         written = written_buffered;
2491                         invalidate_mapping_pages(mapping,
2492                                                  pos >> PAGE_CACHE_SHIFT,
2493                                                  endbyte >> PAGE_CACHE_SHIFT);
2494                 } else {
2495                         /*
2496                          * We don't know how much we wrote, so just return
2497                          * the number of bytes which were direct-written
2498                          */
2499                 }
2500         } else {
2501                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2502                                 pos, ppos, count, written);
2503         }
2504 out:
2505         current->backing_dev_info = NULL;
2506         return written ? written : err;
2507 }
2508 EXPORT_SYMBOL(__generic_file_aio_write);
2509
2510 /**
2511  * generic_file_aio_write - write data to a file
2512  * @iocb:       IO state structure
2513  * @iov:        vector with data to write
2514  * @nr_segs:    number of segments in the vector
2515  * @pos:        position in file where to write
2516  *
2517  * This is a wrapper around __generic_file_aio_write() to be used by most
2518  * filesystems. It takes care of syncing the file in case of O_SYNC file
2519  * and acquires i_mutex as needed.
2520  */
2521 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2522                 unsigned long nr_segs, loff_t pos)
2523 {
2524         struct file *file = iocb->ki_filp;
2525         struct inode *inode = file->f_mapping->host;
2526         ssize_t ret;
2527
2528         BUG_ON(iocb->ki_pos != pos);
2529
2530         sb_start_write(inode->i_sb);
2531         mutex_lock(&inode->i_mutex);
2532         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2533         mutex_unlock(&inode->i_mutex);
2534
2535         if (ret > 0 || ret == -EIOCBQUEUED) {
2536                 ssize_t err;
2537
2538                 err = generic_write_sync(file, pos, ret);
2539                 if (err < 0 && ret > 0)
2540                         ret = err;
2541         }
2542         sb_end_write(inode->i_sb);
2543         return ret;
2544 }
2545 EXPORT_SYMBOL(generic_file_aio_write);
2546
2547 /**
2548  * try_to_release_page() - release old fs-specific metadata on a page
2549  *
2550  * @page: the page which the kernel is trying to free
2551  * @gfp_mask: memory allocation flags (and I/O mode)
2552  *
2553  * The address_space is to try to release any data against the page
2554  * (presumably at page->private).  If the release was successful, return `1'.
2555  * Otherwise return zero.
2556  *
2557  * This may also be called if PG_fscache is set on a page, indicating that the
2558  * page is known to the local caching routines.
2559  *
2560  * The @gfp_mask argument specifies whether I/O may be performed to release
2561  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2562  *
2563  */
2564 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2565 {
2566         struct address_space * const mapping = page->mapping;
2567
2568         BUG_ON(!PageLocked(page));
2569         if (PageWriteback(page))
2570                 return 0;
2571
2572         if (mapping && mapping->a_ops->releasepage)
2573                 return mapping->a_ops->releasepage(page, gfp_mask);
2574         return try_to_free_buffers(page);
2575 }
2576
2577 EXPORT_SYMBOL(try_to_release_page);