mm: a few small updates for radix-swap
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/cleancache.h>
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_mutex              (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_mutex
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  bdi->wb.list_lock
81  *    sb_lock                   (fs/fs-writeback.c)
82  *    ->mapping->tree_lock      (__sync_single_inode)
83  *
84  *  ->i_mmap_mutex
85  *    ->anon_vma.lock           (vma_adjust)
86  *
87  *  ->anon_vma.lock
88  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
89  *
90  *  ->page_table_lock or pte_lock
91  *    ->swap_lock               (try_to_unmap_one)
92  *    ->private_lock            (try_to_unmap_one)
93  *    ->tree_lock               (try_to_unmap_one)
94  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
95  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
96  *    ->private_lock            (page_remove_rmap->set_page_dirty)
97  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
98  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
99  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
100  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
101  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
102  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
103  *
104  *  (code doesn't rely on that order, so you could switch it around)
105  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
106  *    ->i_mmap_mutex
107  */
108
109 /*
110  * Delete a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold the mapping's tree_lock.
113  */
114 void __delete_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         /*
119          * if we're uptodate, flush out into the cleancache, otherwise
120          * invalidate any existing cleancache entries.  We can't leave
121          * stale data around in the cleancache once our page is gone
122          */
123         if (PageUptodate(page) && PageMappedToDisk(page))
124                 cleancache_put_page(page);
125         else
126                 cleancache_flush_page(mapping, page);
127
128         radix_tree_delete(&mapping->page_tree, page->index);
129         page->mapping = NULL;
130         /* Leave page->index set: truncation lookup relies upon it */
131         mapping->nrpages--;
132         __dec_zone_page_state(page, NR_FILE_PAGES);
133         if (PageSwapBacked(page))
134                 __dec_zone_page_state(page, NR_SHMEM);
135         BUG_ON(page_mapped(page));
136
137         /*
138          * Some filesystems seem to re-dirty the page even after
139          * the VM has canceled the dirty bit (eg ext3 journaling).
140          *
141          * Fix it up by doing a final dirty accounting check after
142          * having removed the page entirely.
143          */
144         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
145                 dec_zone_page_state(page, NR_FILE_DIRTY);
146                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
147         }
148 }
149
150 /**
151  * delete_from_page_cache - delete page from page cache
152  * @page: the page which the kernel is trying to remove from page cache
153  *
154  * This must be called only on pages that have been verified to be in the page
155  * cache and locked.  It will never put the page into the free list, the caller
156  * has a reference on the page.
157  */
158 void delete_from_page_cache(struct page *page)
159 {
160         struct address_space *mapping = page->mapping;
161         void (*freepage)(struct page *);
162
163         BUG_ON(!PageLocked(page));
164
165         freepage = mapping->a_ops->freepage;
166         spin_lock_irq(&mapping->tree_lock);
167         __delete_from_page_cache(page);
168         spin_unlock_irq(&mapping->tree_lock);
169         mem_cgroup_uncharge_cache_page(page);
170
171         if (freepage)
172                 freepage(page);
173         page_cache_release(page);
174 }
175 EXPORT_SYMBOL(delete_from_page_cache);
176
177 static int sleep_on_page(void *word)
178 {
179         io_schedule();
180         return 0;
181 }
182
183 static int sleep_on_page_killable(void *word)
184 {
185         sleep_on_page(word);
186         return fatal_signal_pending(current) ? -EINTR : 0;
187 }
188
189 /**
190  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
191  * @mapping:    address space structure to write
192  * @start:      offset in bytes where the range starts
193  * @end:        offset in bytes where the range ends (inclusive)
194  * @sync_mode:  enable synchronous operation
195  *
196  * Start writeback against all of a mapping's dirty pages that lie
197  * within the byte offsets <start, end> inclusive.
198  *
199  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
200  * opposed to a regular memory cleansing writeback.  The difference between
201  * these two operations is that if a dirty page/buffer is encountered, it must
202  * be waited upon, and not just skipped over.
203  */
204 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
205                                 loff_t end, int sync_mode)
206 {
207         int ret;
208         struct writeback_control wbc = {
209                 .sync_mode = sync_mode,
210                 .nr_to_write = LONG_MAX,
211                 .range_start = start,
212                 .range_end = end,
213         };
214
215         if (!mapping_cap_writeback_dirty(mapping))
216                 return 0;
217
218         ret = do_writepages(mapping, &wbc);
219         return ret;
220 }
221
222 static inline int __filemap_fdatawrite(struct address_space *mapping,
223         int sync_mode)
224 {
225         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
226 }
227
228 int filemap_fdatawrite(struct address_space *mapping)
229 {
230         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
231 }
232 EXPORT_SYMBOL(filemap_fdatawrite);
233
234 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
235                                 loff_t end)
236 {
237         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
238 }
239 EXPORT_SYMBOL(filemap_fdatawrite_range);
240
241 /**
242  * filemap_flush - mostly a non-blocking flush
243  * @mapping:    target address_space
244  *
245  * This is a mostly non-blocking flush.  Not suitable for data-integrity
246  * purposes - I/O may not be started against all dirty pages.
247  */
248 int filemap_flush(struct address_space *mapping)
249 {
250         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
251 }
252 EXPORT_SYMBOL(filemap_flush);
253
254 /**
255  * filemap_fdatawait_range - wait for writeback to complete
256  * @mapping:            address space structure to wait for
257  * @start_byte:         offset in bytes where the range starts
258  * @end_byte:           offset in bytes where the range ends (inclusive)
259  *
260  * Walk the list of under-writeback pages of the given address space
261  * in the given range and wait for all of them.
262  */
263 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
264                             loff_t end_byte)
265 {
266         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
267         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271
272         if (end_byte < start_byte)
273                 return 0;
274
275         pagevec_init(&pvec, 0);
276         while ((index <= end) &&
277                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
278                         PAGECACHE_TAG_WRITEBACK,
279                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
280                 unsigned i;
281
282                 for (i = 0; i < nr_pages; i++) {
283                         struct page *page = pvec.pages[i];
284
285                         /* until radix tree lookup accepts end_index */
286                         if (page->index > end)
287                                 continue;
288
289                         wait_on_page_writeback(page);
290                         if (TestClearPageError(page))
291                                 ret = -EIO;
292                 }
293                 pagevec_release(&pvec);
294                 cond_resched();
295         }
296
297         /* Check for outstanding write errors */
298         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
299                 ret = -ENOSPC;
300         if (test_and_clear_bit(AS_EIO, &mapping->flags))
301                 ret = -EIO;
302
303         return ret;
304 }
305 EXPORT_SYMBOL(filemap_fdatawait_range);
306
307 /**
308  * filemap_fdatawait - wait for all under-writeback pages to complete
309  * @mapping: address space structure to wait for
310  *
311  * Walk the list of under-writeback pages of the given address space
312  * and wait for all of them.
313  */
314 int filemap_fdatawait(struct address_space *mapping)
315 {
316         loff_t i_size = i_size_read(mapping->host);
317
318         if (i_size == 0)
319                 return 0;
320
321         return filemap_fdatawait_range(mapping, 0, i_size - 1);
322 }
323 EXPORT_SYMBOL(filemap_fdatawait);
324
325 int filemap_write_and_wait(struct address_space *mapping)
326 {
327         int err = 0;
328
329         if (mapping->nrpages) {
330                 err = filemap_fdatawrite(mapping);
331                 /*
332                  * Even if the above returned error, the pages may be
333                  * written partially (e.g. -ENOSPC), so we wait for it.
334                  * But the -EIO is special case, it may indicate the worst
335                  * thing (e.g. bug) happened, so we avoid waiting for it.
336                  */
337                 if (err != -EIO) {
338                         int err2 = filemap_fdatawait(mapping);
339                         if (!err)
340                                 err = err2;
341                 }
342         }
343         return err;
344 }
345 EXPORT_SYMBOL(filemap_write_and_wait);
346
347 /**
348  * filemap_write_and_wait_range - write out & wait on a file range
349  * @mapping:    the address_space for the pages
350  * @lstart:     offset in bytes where the range starts
351  * @lend:       offset in bytes where the range ends (inclusive)
352  *
353  * Write out and wait upon file offsets lstart->lend, inclusive.
354  *
355  * Note that `lend' is inclusive (describes the last byte to be written) so
356  * that this function can be used to write to the very end-of-file (end = -1).
357  */
358 int filemap_write_and_wait_range(struct address_space *mapping,
359                                  loff_t lstart, loff_t lend)
360 {
361         int err = 0;
362
363         if (mapping->nrpages) {
364                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
365                                                  WB_SYNC_ALL);
366                 /* See comment of filemap_write_and_wait() */
367                 if (err != -EIO) {
368                         int err2 = filemap_fdatawait_range(mapping,
369                                                 lstart, lend);
370                         if (!err)
371                                 err = err2;
372                 }
373         }
374         return err;
375 }
376 EXPORT_SYMBOL(filemap_write_and_wait_range);
377
378 /**
379  * replace_page_cache_page - replace a pagecache page with a new one
380  * @old:        page to be replaced
381  * @new:        page to replace with
382  * @gfp_mask:   allocation mode
383  *
384  * This function replaces a page in the pagecache with a new one.  On
385  * success it acquires the pagecache reference for the new page and
386  * drops it for the old page.  Both the old and new pages must be
387  * locked.  This function does not add the new page to the LRU, the
388  * caller must do that.
389  *
390  * The remove + add is atomic.  The only way this function can fail is
391  * memory allocation failure.
392  */
393 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
394 {
395         int error;
396         struct mem_cgroup *memcg = NULL;
397
398         VM_BUG_ON(!PageLocked(old));
399         VM_BUG_ON(!PageLocked(new));
400         VM_BUG_ON(new->mapping);
401
402         /*
403          * This is not page migration, but prepare_migration and
404          * end_migration does enough work for charge replacement.
405          *
406          * In the longer term we probably want a specialized function
407          * for moving the charge from old to new in a more efficient
408          * manner.
409          */
410         error = mem_cgroup_prepare_migration(old, new, &memcg, gfp_mask);
411         if (error)
412                 return error;
413
414         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
415         if (!error) {
416                 struct address_space *mapping = old->mapping;
417                 void (*freepage)(struct page *);
418
419                 pgoff_t offset = old->index;
420                 freepage = mapping->a_ops->freepage;
421
422                 page_cache_get(new);
423                 new->mapping = mapping;
424                 new->index = offset;
425
426                 spin_lock_irq(&mapping->tree_lock);
427                 __delete_from_page_cache(old);
428                 error = radix_tree_insert(&mapping->page_tree, offset, new);
429                 BUG_ON(error);
430                 mapping->nrpages++;
431                 __inc_zone_page_state(new, NR_FILE_PAGES);
432                 if (PageSwapBacked(new))
433                         __inc_zone_page_state(new, NR_SHMEM);
434                 spin_unlock_irq(&mapping->tree_lock);
435                 radix_tree_preload_end();
436                 if (freepage)
437                         freepage(old);
438                 page_cache_release(old);
439                 mem_cgroup_end_migration(memcg, old, new, true);
440         } else {
441                 mem_cgroup_end_migration(memcg, old, new, false);
442         }
443
444         return error;
445 }
446 EXPORT_SYMBOL_GPL(replace_page_cache_page);
447
448 /**
449  * add_to_page_cache_locked - add a locked page to the pagecache
450  * @page:       page to add
451  * @mapping:    the page's address_space
452  * @offset:     page index
453  * @gfp_mask:   page allocation mode
454  *
455  * This function is used to add a page to the pagecache. It must be locked.
456  * This function does not add the page to the LRU.  The caller must do that.
457  */
458 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
459                 pgoff_t offset, gfp_t gfp_mask)
460 {
461         int error;
462
463         VM_BUG_ON(!PageLocked(page));
464         VM_BUG_ON(PageSwapBacked(page));
465
466         error = mem_cgroup_cache_charge(page, current->mm,
467                                         gfp_mask & GFP_RECLAIM_MASK);
468         if (error)
469                 goto out;
470
471         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
472         if (error == 0) {
473                 page_cache_get(page);
474                 page->mapping = mapping;
475                 page->index = offset;
476
477                 spin_lock_irq(&mapping->tree_lock);
478                 error = radix_tree_insert(&mapping->page_tree, offset, page);
479                 if (likely(!error)) {
480                         mapping->nrpages++;
481                         __inc_zone_page_state(page, NR_FILE_PAGES);
482                         spin_unlock_irq(&mapping->tree_lock);
483                 } else {
484                         page->mapping = NULL;
485                         /* Leave page->index set: truncation relies upon it */
486                         spin_unlock_irq(&mapping->tree_lock);
487                         mem_cgroup_uncharge_cache_page(page);
488                         page_cache_release(page);
489                 }
490                 radix_tree_preload_end();
491         } else
492                 mem_cgroup_uncharge_cache_page(page);
493 out:
494         return error;
495 }
496 EXPORT_SYMBOL(add_to_page_cache_locked);
497
498 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
499                                 pgoff_t offset, gfp_t gfp_mask)
500 {
501         int ret;
502
503         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
504         if (ret == 0)
505                 lru_cache_add_file(page);
506         return ret;
507 }
508 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
509
510 #ifdef CONFIG_NUMA
511 struct page *__page_cache_alloc(gfp_t gfp)
512 {
513         int n;
514         struct page *page;
515
516         if (cpuset_do_page_mem_spread()) {
517                 get_mems_allowed();
518                 n = cpuset_mem_spread_node();
519                 page = alloc_pages_exact_node(n, gfp, 0);
520                 put_mems_allowed();
521                 return page;
522         }
523         return alloc_pages(gfp, 0);
524 }
525 EXPORT_SYMBOL(__page_cache_alloc);
526 #endif
527
528 /*
529  * In order to wait for pages to become available there must be
530  * waitqueues associated with pages. By using a hash table of
531  * waitqueues where the bucket discipline is to maintain all
532  * waiters on the same queue and wake all when any of the pages
533  * become available, and for the woken contexts to check to be
534  * sure the appropriate page became available, this saves space
535  * at a cost of "thundering herd" phenomena during rare hash
536  * collisions.
537  */
538 static wait_queue_head_t *page_waitqueue(struct page *page)
539 {
540         const struct zone *zone = page_zone(page);
541
542         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
543 }
544
545 static inline void wake_up_page(struct page *page, int bit)
546 {
547         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
548 }
549
550 void wait_on_page_bit(struct page *page, int bit_nr)
551 {
552         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
553
554         if (test_bit(bit_nr, &page->flags))
555                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
556                                                         TASK_UNINTERRUPTIBLE);
557 }
558 EXPORT_SYMBOL(wait_on_page_bit);
559
560 int wait_on_page_bit_killable(struct page *page, int bit_nr)
561 {
562         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
563
564         if (!test_bit(bit_nr, &page->flags))
565                 return 0;
566
567         return __wait_on_bit(page_waitqueue(page), &wait,
568                              sleep_on_page_killable, TASK_KILLABLE);
569 }
570
571 /**
572  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
573  * @page: Page defining the wait queue of interest
574  * @waiter: Waiter to add to the queue
575  *
576  * Add an arbitrary @waiter to the wait queue for the nominated @page.
577  */
578 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
579 {
580         wait_queue_head_t *q = page_waitqueue(page);
581         unsigned long flags;
582
583         spin_lock_irqsave(&q->lock, flags);
584         __add_wait_queue(q, waiter);
585         spin_unlock_irqrestore(&q->lock, flags);
586 }
587 EXPORT_SYMBOL_GPL(add_page_wait_queue);
588
589 /**
590  * unlock_page - unlock a locked page
591  * @page: the page
592  *
593  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
594  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
595  * mechananism between PageLocked pages and PageWriteback pages is shared.
596  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
597  *
598  * The mb is necessary to enforce ordering between the clear_bit and the read
599  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
600  */
601 void unlock_page(struct page *page)
602 {
603         VM_BUG_ON(!PageLocked(page));
604         clear_bit_unlock(PG_locked, &page->flags);
605         smp_mb__after_clear_bit();
606         wake_up_page(page, PG_locked);
607 }
608 EXPORT_SYMBOL(unlock_page);
609
610 /**
611  * end_page_writeback - end writeback against a page
612  * @page: the page
613  */
614 void end_page_writeback(struct page *page)
615 {
616         if (TestClearPageReclaim(page))
617                 rotate_reclaimable_page(page);
618
619         if (!test_clear_page_writeback(page))
620                 BUG();
621
622         smp_mb__after_clear_bit();
623         wake_up_page(page, PG_writeback);
624 }
625 EXPORT_SYMBOL(end_page_writeback);
626
627 /**
628  * __lock_page - get a lock on the page, assuming we need to sleep to get it
629  * @page: the page to lock
630  */
631 void __lock_page(struct page *page)
632 {
633         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
634
635         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
636                                                         TASK_UNINTERRUPTIBLE);
637 }
638 EXPORT_SYMBOL(__lock_page);
639
640 int __lock_page_killable(struct page *page)
641 {
642         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
643
644         return __wait_on_bit_lock(page_waitqueue(page), &wait,
645                                         sleep_on_page_killable, TASK_KILLABLE);
646 }
647 EXPORT_SYMBOL_GPL(__lock_page_killable);
648
649 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
650                          unsigned int flags)
651 {
652         if (flags & FAULT_FLAG_ALLOW_RETRY) {
653                 /*
654                  * CAUTION! In this case, mmap_sem is not released
655                  * even though return 0.
656                  */
657                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
658                         return 0;
659
660                 up_read(&mm->mmap_sem);
661                 if (flags & FAULT_FLAG_KILLABLE)
662                         wait_on_page_locked_killable(page);
663                 else
664                         wait_on_page_locked(page);
665                 return 0;
666         } else {
667                 if (flags & FAULT_FLAG_KILLABLE) {
668                         int ret;
669
670                         ret = __lock_page_killable(page);
671                         if (ret) {
672                                 up_read(&mm->mmap_sem);
673                                 return 0;
674                         }
675                 } else
676                         __lock_page(page);
677                 return 1;
678         }
679 }
680
681 /**
682  * find_get_page - find and get a page reference
683  * @mapping: the address_space to search
684  * @offset: the page index
685  *
686  * Is there a pagecache struct page at the given (mapping, offset) tuple?
687  * If yes, increment its refcount and return it; if no, return NULL.
688  */
689 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
690 {
691         void **pagep;
692         struct page *page;
693
694         rcu_read_lock();
695 repeat:
696         page = NULL;
697         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
698         if (pagep) {
699                 page = radix_tree_deref_slot(pagep);
700                 if (unlikely(!page))
701                         goto out;
702                 if (radix_tree_exception(page)) {
703                         if (radix_tree_exceptional_entry(page))
704                                 goto out;
705                         /* radix_tree_deref_retry(page) */
706                         goto repeat;
707                 }
708                 if (!page_cache_get_speculative(page))
709                         goto repeat;
710
711                 /*
712                  * Has the page moved?
713                  * This is part of the lockless pagecache protocol. See
714                  * include/linux/pagemap.h for details.
715                  */
716                 if (unlikely(page != *pagep)) {
717                         page_cache_release(page);
718                         goto repeat;
719                 }
720         }
721 out:
722         rcu_read_unlock();
723
724         return page;
725 }
726 EXPORT_SYMBOL(find_get_page);
727
728 /**
729  * find_lock_page - locate, pin and lock a pagecache page
730  * @mapping: the address_space to search
731  * @offset: the page index
732  *
733  * Locates the desired pagecache page, locks it, increments its reference
734  * count and returns its address.
735  *
736  * Returns zero if the page was not present. find_lock_page() may sleep.
737  */
738 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
739 {
740         struct page *page;
741
742 repeat:
743         page = find_get_page(mapping, offset);
744         if (page && !radix_tree_exception(page)) {
745                 lock_page(page);
746                 /* Has the page been truncated? */
747                 if (unlikely(page->mapping != mapping)) {
748                         unlock_page(page);
749                         page_cache_release(page);
750                         goto repeat;
751                 }
752                 VM_BUG_ON(page->index != offset);
753         }
754         return page;
755 }
756 EXPORT_SYMBOL(find_lock_page);
757
758 /**
759  * find_or_create_page - locate or add a pagecache page
760  * @mapping: the page's address_space
761  * @index: the page's index into the mapping
762  * @gfp_mask: page allocation mode
763  *
764  * Locates a page in the pagecache.  If the page is not present, a new page
765  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
766  * LRU list.  The returned page is locked and has its reference count
767  * incremented.
768  *
769  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
770  * allocation!
771  *
772  * find_or_create_page() returns the desired page's address, or zero on
773  * memory exhaustion.
774  */
775 struct page *find_or_create_page(struct address_space *mapping,
776                 pgoff_t index, gfp_t gfp_mask)
777 {
778         struct page *page;
779         int err;
780 repeat:
781         page = find_lock_page(mapping, index);
782         if (!page) {
783                 page = __page_cache_alloc(gfp_mask);
784                 if (!page)
785                         return NULL;
786                 /*
787                  * We want a regular kernel memory (not highmem or DMA etc)
788                  * allocation for the radix tree nodes, but we need to honour
789                  * the context-specific requirements the caller has asked for.
790                  * GFP_RECLAIM_MASK collects those requirements.
791                  */
792                 err = add_to_page_cache_lru(page, mapping, index,
793                         (gfp_mask & GFP_RECLAIM_MASK));
794                 if (unlikely(err)) {
795                         page_cache_release(page);
796                         page = NULL;
797                         if (err == -EEXIST)
798                                 goto repeat;
799                 }
800         }
801         return page;
802 }
803 EXPORT_SYMBOL(find_or_create_page);
804
805 /**
806  * find_get_pages - gang pagecache lookup
807  * @mapping:    The address_space to search
808  * @start:      The starting page index
809  * @nr_pages:   The maximum number of pages
810  * @pages:      Where the resulting pages are placed
811  *
812  * find_get_pages() will search for and return a group of up to
813  * @nr_pages pages in the mapping.  The pages are placed at @pages.
814  * find_get_pages() takes a reference against the returned pages.
815  *
816  * The search returns a group of mapping-contiguous pages with ascending
817  * indexes.  There may be holes in the indices due to not-present pages.
818  *
819  * find_get_pages() returns the number of pages which were found.
820  */
821 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
822                             unsigned int nr_pages, struct page **pages)
823 {
824         unsigned int i;
825         unsigned int ret;
826         unsigned int nr_found;
827
828         rcu_read_lock();
829 restart:
830         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
831                                 (void ***)pages, NULL, start, nr_pages);
832         ret = 0;
833         for (i = 0; i < nr_found; i++) {
834                 struct page *page;
835 repeat:
836                 page = radix_tree_deref_slot((void **)pages[i]);
837                 if (unlikely(!page))
838                         continue;
839
840                 if (radix_tree_exception(page)) {
841                         if (radix_tree_exceptional_entry(page))
842                                 continue;
843                         /*
844                          * radix_tree_deref_retry(page):
845                          * can only trigger when entry at index 0 moves out of
846                          * or back to root: none yet gotten, safe to restart.
847                          */
848                         WARN_ON(start | i);
849                         goto restart;
850                 }
851
852                 if (!page_cache_get_speculative(page))
853                         goto repeat;
854
855                 /* Has the page moved? */
856                 if (unlikely(page != *((void **)pages[i]))) {
857                         page_cache_release(page);
858                         goto repeat;
859                 }
860
861                 pages[ret] = page;
862                 ret++;
863         }
864
865         /*
866          * If all entries were removed before we could secure them,
867          * try again, because callers stop trying once 0 is returned.
868          */
869         if (unlikely(!ret && nr_found))
870                 goto restart;
871         rcu_read_unlock();
872         return ret;
873 }
874
875 /**
876  * find_get_pages_contig - gang contiguous pagecache lookup
877  * @mapping:    The address_space to search
878  * @index:      The starting page index
879  * @nr_pages:   The maximum number of pages
880  * @pages:      Where the resulting pages are placed
881  *
882  * find_get_pages_contig() works exactly like find_get_pages(), except
883  * that the returned number of pages are guaranteed to be contiguous.
884  *
885  * find_get_pages_contig() returns the number of pages which were found.
886  */
887 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
888                                unsigned int nr_pages, struct page **pages)
889 {
890         unsigned int i;
891         unsigned int ret;
892         unsigned int nr_found;
893
894         rcu_read_lock();
895 restart:
896         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
897                                 (void ***)pages, NULL, index, nr_pages);
898         ret = 0;
899         for (i = 0; i < nr_found; i++) {
900                 struct page *page;
901 repeat:
902                 page = radix_tree_deref_slot((void **)pages[i]);
903                 if (unlikely(!page))
904                         continue;
905
906                 if (radix_tree_exception(page)) {
907                         if (radix_tree_exceptional_entry(page))
908                                 break;
909                         /*
910                          * radix_tree_deref_retry(page):
911                          * can only trigger when entry at index 0 moves out of
912                          * or back to root: none yet gotten, safe to restart.
913                          */
914                         goto restart;
915                 }
916
917                 if (!page_cache_get_speculative(page))
918                         goto repeat;
919
920                 /* Has the page moved? */
921                 if (unlikely(page != *((void **)pages[i]))) {
922                         page_cache_release(page);
923                         goto repeat;
924                 }
925
926                 /*
927                  * must check mapping and index after taking the ref.
928                  * otherwise we can get both false positives and false
929                  * negatives, which is just confusing to the caller.
930                  */
931                 if (page->mapping == NULL || page->index != index) {
932                         page_cache_release(page);
933                         break;
934                 }
935
936                 pages[ret] = page;
937                 ret++;
938                 index++;
939         }
940         rcu_read_unlock();
941         return ret;
942 }
943 EXPORT_SYMBOL(find_get_pages_contig);
944
945 /**
946  * find_get_pages_tag - find and return pages that match @tag
947  * @mapping:    the address_space to search
948  * @index:      the starting page index
949  * @tag:        the tag index
950  * @nr_pages:   the maximum number of pages
951  * @pages:      where the resulting pages are placed
952  *
953  * Like find_get_pages, except we only return pages which are tagged with
954  * @tag.   We update @index to index the next page for the traversal.
955  */
956 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
957                         int tag, unsigned int nr_pages, struct page **pages)
958 {
959         unsigned int i;
960         unsigned int ret;
961         unsigned int nr_found;
962
963         rcu_read_lock();
964 restart:
965         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
966                                 (void ***)pages, *index, nr_pages, tag);
967         ret = 0;
968         for (i = 0; i < nr_found; i++) {
969                 struct page *page;
970 repeat:
971                 page = radix_tree_deref_slot((void **)pages[i]);
972                 if (unlikely(!page))
973                         continue;
974
975                 if (radix_tree_exception(page)) {
976                         BUG_ON(radix_tree_exceptional_entry(page));
977                         /*
978                          * radix_tree_deref_retry(page):
979                          * can only trigger when entry at index 0 moves out of
980                          * or back to root: none yet gotten, safe to restart.
981                          */
982                         goto restart;
983                 }
984
985                 if (!page_cache_get_speculative(page))
986                         goto repeat;
987
988                 /* Has the page moved? */
989                 if (unlikely(page != *((void **)pages[i]))) {
990                         page_cache_release(page);
991                         goto repeat;
992                 }
993
994                 pages[ret] = page;
995                 ret++;
996         }
997
998         /*
999          * If all entries were removed before we could secure them,
1000          * try again, because callers stop trying once 0 is returned.
1001          */
1002         if (unlikely(!ret && nr_found))
1003                 goto restart;
1004         rcu_read_unlock();
1005
1006         if (ret)
1007                 *index = pages[ret - 1]->index + 1;
1008
1009         return ret;
1010 }
1011 EXPORT_SYMBOL(find_get_pages_tag);
1012
1013 /**
1014  * grab_cache_page_nowait - returns locked page at given index in given cache
1015  * @mapping: target address_space
1016  * @index: the page index
1017  *
1018  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1019  * This is intended for speculative data generators, where the data can
1020  * be regenerated if the page couldn't be grabbed.  This routine should
1021  * be safe to call while holding the lock for another page.
1022  *
1023  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1024  * and deadlock against the caller's locked page.
1025  */
1026 struct page *
1027 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1028 {
1029         struct page *page = find_get_page(mapping, index);
1030
1031         if (page) {
1032                 if (trylock_page(page))
1033                         return page;
1034                 page_cache_release(page);
1035                 return NULL;
1036         }
1037         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1038         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1039                 page_cache_release(page);
1040                 page = NULL;
1041         }
1042         return page;
1043 }
1044 EXPORT_SYMBOL(grab_cache_page_nowait);
1045
1046 /*
1047  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1048  * a _large_ part of the i/o request. Imagine the worst scenario:
1049  *
1050  *      ---R__________________________________________B__________
1051  *         ^ reading here                             ^ bad block(assume 4k)
1052  *
1053  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1054  * => failing the whole request => read(R) => read(R+1) =>
1055  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1056  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1057  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1058  *
1059  * It is going insane. Fix it by quickly scaling down the readahead size.
1060  */
1061 static void shrink_readahead_size_eio(struct file *filp,
1062                                         struct file_ra_state *ra)
1063 {
1064         ra->ra_pages /= 4;
1065 }
1066
1067 /**
1068  * do_generic_file_read - generic file read routine
1069  * @filp:       the file to read
1070  * @ppos:       current file position
1071  * @desc:       read_descriptor
1072  * @actor:      read method
1073  *
1074  * This is a generic file read routine, and uses the
1075  * mapping->a_ops->readpage() function for the actual low-level stuff.
1076  *
1077  * This is really ugly. But the goto's actually try to clarify some
1078  * of the logic when it comes to error handling etc.
1079  */
1080 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1081                 read_descriptor_t *desc, read_actor_t actor)
1082 {
1083         struct address_space *mapping = filp->f_mapping;
1084         struct inode *inode = mapping->host;
1085         struct file_ra_state *ra = &filp->f_ra;
1086         pgoff_t index;
1087         pgoff_t last_index;
1088         pgoff_t prev_index;
1089         unsigned long offset;      /* offset into pagecache page */
1090         unsigned int prev_offset;
1091         int error;
1092
1093         index = *ppos >> PAGE_CACHE_SHIFT;
1094         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1095         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1096         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1097         offset = *ppos & ~PAGE_CACHE_MASK;
1098
1099         for (;;) {
1100                 struct page *page;
1101                 pgoff_t end_index;
1102                 loff_t isize;
1103                 unsigned long nr, ret;
1104
1105                 cond_resched();
1106 find_page:
1107                 page = find_get_page(mapping, index);
1108                 if (!page) {
1109                         page_cache_sync_readahead(mapping,
1110                                         ra, filp,
1111                                         index, last_index - index);
1112                         page = find_get_page(mapping, index);
1113                         if (unlikely(page == NULL))
1114                                 goto no_cached_page;
1115                 }
1116                 if (PageReadahead(page)) {
1117                         page_cache_async_readahead(mapping,
1118                                         ra, filp, page,
1119                                         index, last_index - index);
1120                 }
1121                 if (!PageUptodate(page)) {
1122                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1123                                         !mapping->a_ops->is_partially_uptodate)
1124                                 goto page_not_up_to_date;
1125                         if (!trylock_page(page))
1126                                 goto page_not_up_to_date;
1127                         /* Did it get truncated before we got the lock? */
1128                         if (!page->mapping)
1129                                 goto page_not_up_to_date_locked;
1130                         if (!mapping->a_ops->is_partially_uptodate(page,
1131                                                                 desc, offset))
1132                                 goto page_not_up_to_date_locked;
1133                         unlock_page(page);
1134                 }
1135 page_ok:
1136                 /*
1137                  * i_size must be checked after we know the page is Uptodate.
1138                  *
1139                  * Checking i_size after the check allows us to calculate
1140                  * the correct value for "nr", which means the zero-filled
1141                  * part of the page is not copied back to userspace (unless
1142                  * another truncate extends the file - this is desired though).
1143                  */
1144
1145                 isize = i_size_read(inode);
1146                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1147                 if (unlikely(!isize || index > end_index)) {
1148                         page_cache_release(page);
1149                         goto out;
1150                 }
1151
1152                 /* nr is the maximum number of bytes to copy from this page */
1153                 nr = PAGE_CACHE_SIZE;
1154                 if (index == end_index) {
1155                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1156                         if (nr <= offset) {
1157                                 page_cache_release(page);
1158                                 goto out;
1159                         }
1160                 }
1161                 nr = nr - offset;
1162
1163                 /* If users can be writing to this page using arbitrary
1164                  * virtual addresses, take care about potential aliasing
1165                  * before reading the page on the kernel side.
1166                  */
1167                 if (mapping_writably_mapped(mapping))
1168                         flush_dcache_page(page);
1169
1170                 /*
1171                  * When a sequential read accesses a page several times,
1172                  * only mark it as accessed the first time.
1173                  */
1174                 if (prev_index != index || offset != prev_offset)
1175                         mark_page_accessed(page);
1176                 prev_index = index;
1177
1178                 /*
1179                  * Ok, we have the page, and it's up-to-date, so
1180                  * now we can copy it to user space...
1181                  *
1182                  * The actor routine returns how many bytes were actually used..
1183                  * NOTE! This may not be the same as how much of a user buffer
1184                  * we filled up (we may be padding etc), so we can only update
1185                  * "pos" here (the actor routine has to update the user buffer
1186                  * pointers and the remaining count).
1187                  */
1188                 ret = actor(desc, page, offset, nr);
1189                 offset += ret;
1190                 index += offset >> PAGE_CACHE_SHIFT;
1191                 offset &= ~PAGE_CACHE_MASK;
1192                 prev_offset = offset;
1193
1194                 page_cache_release(page);
1195                 if (ret == nr && desc->count)
1196                         continue;
1197                 goto out;
1198
1199 page_not_up_to_date:
1200                 /* Get exclusive access to the page ... */
1201                 error = lock_page_killable(page);
1202                 if (unlikely(error))
1203                         goto readpage_error;
1204
1205 page_not_up_to_date_locked:
1206                 /* Did it get truncated before we got the lock? */
1207                 if (!page->mapping) {
1208                         unlock_page(page);
1209                         page_cache_release(page);
1210                         continue;
1211                 }
1212
1213                 /* Did somebody else fill it already? */
1214                 if (PageUptodate(page)) {
1215                         unlock_page(page);
1216                         goto page_ok;
1217                 }
1218
1219 readpage:
1220                 /*
1221                  * A previous I/O error may have been due to temporary
1222                  * failures, eg. multipath errors.
1223                  * PG_error will be set again if readpage fails.
1224                  */
1225                 ClearPageError(page);
1226                 /* Start the actual read. The read will unlock the page. */
1227                 error = mapping->a_ops->readpage(filp, page);
1228
1229                 if (unlikely(error)) {
1230                         if (error == AOP_TRUNCATED_PAGE) {
1231                                 page_cache_release(page);
1232                                 goto find_page;
1233                         }
1234                         goto readpage_error;
1235                 }
1236
1237                 if (!PageUptodate(page)) {
1238                         error = lock_page_killable(page);
1239                         if (unlikely(error))
1240                                 goto readpage_error;
1241                         if (!PageUptodate(page)) {
1242                                 if (page->mapping == NULL) {
1243                                         /*
1244                                          * invalidate_mapping_pages got it
1245                                          */
1246                                         unlock_page(page);
1247                                         page_cache_release(page);
1248                                         goto find_page;
1249                                 }
1250                                 unlock_page(page);
1251                                 shrink_readahead_size_eio(filp, ra);
1252                                 error = -EIO;
1253                                 goto readpage_error;
1254                         }
1255                         unlock_page(page);
1256                 }
1257
1258                 goto page_ok;
1259
1260 readpage_error:
1261                 /* UHHUH! A synchronous read error occurred. Report it */
1262                 desc->error = error;
1263                 page_cache_release(page);
1264                 goto out;
1265
1266 no_cached_page:
1267                 /*
1268                  * Ok, it wasn't cached, so we need to create a new
1269                  * page..
1270                  */
1271                 page = page_cache_alloc_cold(mapping);
1272                 if (!page) {
1273                         desc->error = -ENOMEM;
1274                         goto out;
1275                 }
1276                 error = add_to_page_cache_lru(page, mapping,
1277                                                 index, GFP_KERNEL);
1278                 if (error) {
1279                         page_cache_release(page);
1280                         if (error == -EEXIST)
1281                                 goto find_page;
1282                         desc->error = error;
1283                         goto out;
1284                 }
1285                 goto readpage;
1286         }
1287
1288 out:
1289         ra->prev_pos = prev_index;
1290         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1291         ra->prev_pos |= prev_offset;
1292
1293         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1294         file_accessed(filp);
1295 }
1296
1297 int file_read_actor(read_descriptor_t *desc, struct page *page,
1298                         unsigned long offset, unsigned long size)
1299 {
1300         char *kaddr;
1301         unsigned long left, count = desc->count;
1302
1303         if (size > count)
1304                 size = count;
1305
1306         /*
1307          * Faults on the destination of a read are common, so do it before
1308          * taking the kmap.
1309          */
1310         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1311                 kaddr = kmap_atomic(page, KM_USER0);
1312                 left = __copy_to_user_inatomic(desc->arg.buf,
1313                                                 kaddr + offset, size);
1314                 kunmap_atomic(kaddr, KM_USER0);
1315                 if (left == 0)
1316                         goto success;
1317         }
1318
1319         /* Do it the slow way */
1320         kaddr = kmap(page);
1321         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1322         kunmap(page);
1323
1324         if (left) {
1325                 size -= left;
1326                 desc->error = -EFAULT;
1327         }
1328 success:
1329         desc->count = count - size;
1330         desc->written += size;
1331         desc->arg.buf += size;
1332         return size;
1333 }
1334
1335 /*
1336  * Performs necessary checks before doing a write
1337  * @iov:        io vector request
1338  * @nr_segs:    number of segments in the iovec
1339  * @count:      number of bytes to write
1340  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1341  *
1342  * Adjust number of segments and amount of bytes to write (nr_segs should be
1343  * properly initialized first). Returns appropriate error code that caller
1344  * should return or zero in case that write should be allowed.
1345  */
1346 int generic_segment_checks(const struct iovec *iov,
1347                         unsigned long *nr_segs, size_t *count, int access_flags)
1348 {
1349         unsigned long   seg;
1350         size_t cnt = 0;
1351         for (seg = 0; seg < *nr_segs; seg++) {
1352                 const struct iovec *iv = &iov[seg];
1353
1354                 /*
1355                  * If any segment has a negative length, or the cumulative
1356                  * length ever wraps negative then return -EINVAL.
1357                  */
1358                 cnt += iv->iov_len;
1359                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1360                         return -EINVAL;
1361                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1362                         continue;
1363                 if (seg == 0)
1364                         return -EFAULT;
1365                 *nr_segs = seg;
1366                 cnt -= iv->iov_len;     /* This segment is no good */
1367                 break;
1368         }
1369         *count = cnt;
1370         return 0;
1371 }
1372 EXPORT_SYMBOL(generic_segment_checks);
1373
1374 /**
1375  * generic_file_aio_read - generic filesystem read routine
1376  * @iocb:       kernel I/O control block
1377  * @iov:        io vector request
1378  * @nr_segs:    number of segments in the iovec
1379  * @pos:        current file position
1380  *
1381  * This is the "read()" routine for all filesystems
1382  * that can use the page cache directly.
1383  */
1384 ssize_t
1385 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1386                 unsigned long nr_segs, loff_t pos)
1387 {
1388         struct file *filp = iocb->ki_filp;
1389         ssize_t retval;
1390         unsigned long seg = 0;
1391         size_t count;
1392         loff_t *ppos = &iocb->ki_pos;
1393         struct blk_plug plug;
1394
1395         count = 0;
1396         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1397         if (retval)
1398                 return retval;
1399
1400         blk_start_plug(&plug);
1401
1402         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1403         if (filp->f_flags & O_DIRECT) {
1404                 loff_t size;
1405                 struct address_space *mapping;
1406                 struct inode *inode;
1407
1408                 mapping = filp->f_mapping;
1409                 inode = mapping->host;
1410                 if (!count)
1411                         goto out; /* skip atime */
1412                 size = i_size_read(inode);
1413                 if (pos < size) {
1414                         retval = filemap_write_and_wait_range(mapping, pos,
1415                                         pos + iov_length(iov, nr_segs) - 1);
1416                         if (!retval) {
1417                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1418                                                         iov, pos, nr_segs);
1419                         }
1420                         if (retval > 0) {
1421                                 *ppos = pos + retval;
1422                                 count -= retval;
1423                         }
1424
1425                         /*
1426                          * Btrfs can have a short DIO read if we encounter
1427                          * compressed extents, so if there was an error, or if
1428                          * we've already read everything we wanted to, or if
1429                          * there was a short read because we hit EOF, go ahead
1430                          * and return.  Otherwise fallthrough to buffered io for
1431                          * the rest of the read.
1432                          */
1433                         if (retval < 0 || !count || *ppos >= size) {
1434                                 file_accessed(filp);
1435                                 goto out;
1436                         }
1437                 }
1438         }
1439
1440         count = retval;
1441         for (seg = 0; seg < nr_segs; seg++) {
1442                 read_descriptor_t desc;
1443                 loff_t offset = 0;
1444
1445                 /*
1446                  * If we did a short DIO read we need to skip the section of the
1447                  * iov that we've already read data into.
1448                  */
1449                 if (count) {
1450                         if (count > iov[seg].iov_len) {
1451                                 count -= iov[seg].iov_len;
1452                                 continue;
1453                         }
1454                         offset = count;
1455                         count = 0;
1456                 }
1457
1458                 desc.written = 0;
1459                 desc.arg.buf = iov[seg].iov_base + offset;
1460                 desc.count = iov[seg].iov_len - offset;
1461                 if (desc.count == 0)
1462                         continue;
1463                 desc.error = 0;
1464                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1465                 retval += desc.written;
1466                 if (desc.error) {
1467                         retval = retval ?: desc.error;
1468                         break;
1469                 }
1470                 if (desc.count > 0)
1471                         break;
1472         }
1473 out:
1474         blk_finish_plug(&plug);
1475         return retval;
1476 }
1477 EXPORT_SYMBOL(generic_file_aio_read);
1478
1479 static ssize_t
1480 do_readahead(struct address_space *mapping, struct file *filp,
1481              pgoff_t index, unsigned long nr)
1482 {
1483         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1484                 return -EINVAL;
1485
1486         force_page_cache_readahead(mapping, filp, index, nr);
1487         return 0;
1488 }
1489
1490 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1491 {
1492         ssize_t ret;
1493         struct file *file;
1494
1495         ret = -EBADF;
1496         file = fget(fd);
1497         if (file) {
1498                 if (file->f_mode & FMODE_READ) {
1499                         struct address_space *mapping = file->f_mapping;
1500                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1501                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1502                         unsigned long len = end - start + 1;
1503                         ret = do_readahead(mapping, file, start, len);
1504                 }
1505                 fput(file);
1506         }
1507         return ret;
1508 }
1509 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1510 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1511 {
1512         return SYSC_readahead((int) fd, offset, (size_t) count);
1513 }
1514 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1515 #endif
1516
1517 #ifdef CONFIG_MMU
1518 /**
1519  * page_cache_read - adds requested page to the page cache if not already there
1520  * @file:       file to read
1521  * @offset:     page index
1522  *
1523  * This adds the requested page to the page cache if it isn't already there,
1524  * and schedules an I/O to read in its contents from disk.
1525  */
1526 static int page_cache_read(struct file *file, pgoff_t offset)
1527 {
1528         struct address_space *mapping = file->f_mapping;
1529         struct page *page; 
1530         int ret;
1531
1532         do {
1533                 page = page_cache_alloc_cold(mapping);
1534                 if (!page)
1535                         return -ENOMEM;
1536
1537                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1538                 if (ret == 0)
1539                         ret = mapping->a_ops->readpage(file, page);
1540                 else if (ret == -EEXIST)
1541                         ret = 0; /* losing race to add is OK */
1542
1543                 page_cache_release(page);
1544
1545         } while (ret == AOP_TRUNCATED_PAGE);
1546                 
1547         return ret;
1548 }
1549
1550 #define MMAP_LOTSAMISS  (100)
1551
1552 /*
1553  * Synchronous readahead happens when we don't even find
1554  * a page in the page cache at all.
1555  */
1556 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1557                                    struct file_ra_state *ra,
1558                                    struct file *file,
1559                                    pgoff_t offset)
1560 {
1561         unsigned long ra_pages;
1562         struct address_space *mapping = file->f_mapping;
1563
1564         /* If we don't want any read-ahead, don't bother */
1565         if (VM_RandomReadHint(vma))
1566                 return;
1567         if (!ra->ra_pages)
1568                 return;
1569
1570         if (VM_SequentialReadHint(vma)) {
1571                 page_cache_sync_readahead(mapping, ra, file, offset,
1572                                           ra->ra_pages);
1573                 return;
1574         }
1575
1576         /* Avoid banging the cache line if not needed */
1577         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1578                 ra->mmap_miss++;
1579
1580         /*
1581          * Do we miss much more than hit in this file? If so,
1582          * stop bothering with read-ahead. It will only hurt.
1583          */
1584         if (ra->mmap_miss > MMAP_LOTSAMISS)
1585                 return;
1586
1587         /*
1588          * mmap read-around
1589          */
1590         ra_pages = max_sane_readahead(ra->ra_pages);
1591         ra->start = max_t(long, 0, offset - ra_pages / 2);
1592         ra->size = ra_pages;
1593         ra->async_size = ra_pages / 4;
1594         ra_submit(ra, mapping, file);
1595 }
1596
1597 /*
1598  * Asynchronous readahead happens when we find the page and PG_readahead,
1599  * so we want to possibly extend the readahead further..
1600  */
1601 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1602                                     struct file_ra_state *ra,
1603                                     struct file *file,
1604                                     struct page *page,
1605                                     pgoff_t offset)
1606 {
1607         struct address_space *mapping = file->f_mapping;
1608
1609         /* If we don't want any read-ahead, don't bother */
1610         if (VM_RandomReadHint(vma))
1611                 return;
1612         if (ra->mmap_miss > 0)
1613                 ra->mmap_miss--;
1614         if (PageReadahead(page))
1615                 page_cache_async_readahead(mapping, ra, file,
1616                                            page, offset, ra->ra_pages);
1617 }
1618
1619 /**
1620  * filemap_fault - read in file data for page fault handling
1621  * @vma:        vma in which the fault was taken
1622  * @vmf:        struct vm_fault containing details of the fault
1623  *
1624  * filemap_fault() is invoked via the vma operations vector for a
1625  * mapped memory region to read in file data during a page fault.
1626  *
1627  * The goto's are kind of ugly, but this streamlines the normal case of having
1628  * it in the page cache, and handles the special cases reasonably without
1629  * having a lot of duplicated code.
1630  */
1631 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1632 {
1633         int error;
1634         struct file *file = vma->vm_file;
1635         struct address_space *mapping = file->f_mapping;
1636         struct file_ra_state *ra = &file->f_ra;
1637         struct inode *inode = mapping->host;
1638         pgoff_t offset = vmf->pgoff;
1639         struct page *page;
1640         pgoff_t size;
1641         int ret = 0;
1642
1643         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1644         if (offset >= size)
1645                 return VM_FAULT_SIGBUS;
1646
1647         /*
1648          * Do we have something in the page cache already?
1649          */
1650         page = find_get_page(mapping, offset);
1651         if (likely(page)) {
1652                 /*
1653                  * We found the page, so try async readahead before
1654                  * waiting for the lock.
1655                  */
1656                 do_async_mmap_readahead(vma, ra, file, page, offset);
1657         } else {
1658                 /* No page in the page cache at all */
1659                 do_sync_mmap_readahead(vma, ra, file, offset);
1660                 count_vm_event(PGMAJFAULT);
1661                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1662                 ret = VM_FAULT_MAJOR;
1663 retry_find:
1664                 page = find_get_page(mapping, offset);
1665                 if (!page)
1666                         goto no_cached_page;
1667         }
1668
1669         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1670                 page_cache_release(page);
1671                 return ret | VM_FAULT_RETRY;
1672         }
1673
1674         /* Did it get truncated? */
1675         if (unlikely(page->mapping != mapping)) {
1676                 unlock_page(page);
1677                 put_page(page);
1678                 goto retry_find;
1679         }
1680         VM_BUG_ON(page->index != offset);
1681
1682         /*
1683          * We have a locked page in the page cache, now we need to check
1684          * that it's up-to-date. If not, it is going to be due to an error.
1685          */
1686         if (unlikely(!PageUptodate(page)))
1687                 goto page_not_uptodate;
1688
1689         /*
1690          * Found the page and have a reference on it.
1691          * We must recheck i_size under page lock.
1692          */
1693         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1694         if (unlikely(offset >= size)) {
1695                 unlock_page(page);
1696                 page_cache_release(page);
1697                 return VM_FAULT_SIGBUS;
1698         }
1699
1700         vmf->page = page;
1701         return ret | VM_FAULT_LOCKED;
1702
1703 no_cached_page:
1704         /*
1705          * We're only likely to ever get here if MADV_RANDOM is in
1706          * effect.
1707          */
1708         error = page_cache_read(file, offset);
1709
1710         /*
1711          * The page we want has now been added to the page cache.
1712          * In the unlikely event that someone removed it in the
1713          * meantime, we'll just come back here and read it again.
1714          */
1715         if (error >= 0)
1716                 goto retry_find;
1717
1718         /*
1719          * An error return from page_cache_read can result if the
1720          * system is low on memory, or a problem occurs while trying
1721          * to schedule I/O.
1722          */
1723         if (error == -ENOMEM)
1724                 return VM_FAULT_OOM;
1725         return VM_FAULT_SIGBUS;
1726
1727 page_not_uptodate:
1728         /*
1729          * Umm, take care of errors if the page isn't up-to-date.
1730          * Try to re-read it _once_. We do this synchronously,
1731          * because there really aren't any performance issues here
1732          * and we need to check for errors.
1733          */
1734         ClearPageError(page);
1735         error = mapping->a_ops->readpage(file, page);
1736         if (!error) {
1737                 wait_on_page_locked(page);
1738                 if (!PageUptodate(page))
1739                         error = -EIO;
1740         }
1741         page_cache_release(page);
1742
1743         if (!error || error == AOP_TRUNCATED_PAGE)
1744                 goto retry_find;
1745
1746         /* Things didn't work out. Return zero to tell the mm layer so. */
1747         shrink_readahead_size_eio(file, ra);
1748         return VM_FAULT_SIGBUS;
1749 }
1750 EXPORT_SYMBOL(filemap_fault);
1751
1752 const struct vm_operations_struct generic_file_vm_ops = {
1753         .fault          = filemap_fault,
1754 };
1755
1756 /* This is used for a general mmap of a disk file */
1757
1758 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1759 {
1760         struct address_space *mapping = file->f_mapping;
1761
1762         if (!mapping->a_ops->readpage)
1763                 return -ENOEXEC;
1764         file_accessed(file);
1765         vma->vm_ops = &generic_file_vm_ops;
1766         vma->vm_flags |= VM_CAN_NONLINEAR;
1767         return 0;
1768 }
1769
1770 /*
1771  * This is for filesystems which do not implement ->writepage.
1772  */
1773 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1774 {
1775         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1776                 return -EINVAL;
1777         return generic_file_mmap(file, vma);
1778 }
1779 #else
1780 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1781 {
1782         return -ENOSYS;
1783 }
1784 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1785 {
1786         return -ENOSYS;
1787 }
1788 #endif /* CONFIG_MMU */
1789
1790 EXPORT_SYMBOL(generic_file_mmap);
1791 EXPORT_SYMBOL(generic_file_readonly_mmap);
1792
1793 static struct page *__read_cache_page(struct address_space *mapping,
1794                                 pgoff_t index,
1795                                 int (*filler)(void *, struct page *),
1796                                 void *data,
1797                                 gfp_t gfp)
1798 {
1799         struct page *page;
1800         int err;
1801 repeat:
1802         page = find_get_page(mapping, index);
1803         if (!page) {
1804                 page = __page_cache_alloc(gfp | __GFP_COLD);
1805                 if (!page)
1806                         return ERR_PTR(-ENOMEM);
1807                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1808                 if (unlikely(err)) {
1809                         page_cache_release(page);
1810                         if (err == -EEXIST)
1811                                 goto repeat;
1812                         /* Presumably ENOMEM for radix tree node */
1813                         return ERR_PTR(err);
1814                 }
1815                 err = filler(data, page);
1816                 if (err < 0) {
1817                         page_cache_release(page);
1818                         page = ERR_PTR(err);
1819                 }
1820         }
1821         return page;
1822 }
1823
1824 static struct page *do_read_cache_page(struct address_space *mapping,
1825                                 pgoff_t index,
1826                                 int (*filler)(void *, struct page *),
1827                                 void *data,
1828                                 gfp_t gfp)
1829
1830 {
1831         struct page *page;
1832         int err;
1833
1834 retry:
1835         page = __read_cache_page(mapping, index, filler, data, gfp);
1836         if (IS_ERR(page))
1837                 return page;
1838         if (PageUptodate(page))
1839                 goto out;
1840
1841         lock_page(page);
1842         if (!page->mapping) {
1843                 unlock_page(page);
1844                 page_cache_release(page);
1845                 goto retry;
1846         }
1847         if (PageUptodate(page)) {
1848                 unlock_page(page);
1849                 goto out;
1850         }
1851         err = filler(data, page);
1852         if (err < 0) {
1853                 page_cache_release(page);
1854                 return ERR_PTR(err);
1855         }
1856 out:
1857         mark_page_accessed(page);
1858         return page;
1859 }
1860
1861 /**
1862  * read_cache_page_async - read into page cache, fill it if needed
1863  * @mapping:    the page's address_space
1864  * @index:      the page index
1865  * @filler:     function to perform the read
1866  * @data:       first arg to filler(data, page) function, often left as NULL
1867  *
1868  * Same as read_cache_page, but don't wait for page to become unlocked
1869  * after submitting it to the filler.
1870  *
1871  * Read into the page cache. If a page already exists, and PageUptodate() is
1872  * not set, try to fill the page but don't wait for it to become unlocked.
1873  *
1874  * If the page does not get brought uptodate, return -EIO.
1875  */
1876 struct page *read_cache_page_async(struct address_space *mapping,
1877                                 pgoff_t index,
1878                                 int (*filler)(void *, struct page *),
1879                                 void *data)
1880 {
1881         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1882 }
1883 EXPORT_SYMBOL(read_cache_page_async);
1884
1885 static struct page *wait_on_page_read(struct page *page)
1886 {
1887         if (!IS_ERR(page)) {
1888                 wait_on_page_locked(page);
1889                 if (!PageUptodate(page)) {
1890                         page_cache_release(page);
1891                         page = ERR_PTR(-EIO);
1892                 }
1893         }
1894         return page;
1895 }
1896
1897 /**
1898  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1899  * @mapping:    the page's address_space
1900  * @index:      the page index
1901  * @gfp:        the page allocator flags to use if allocating
1902  *
1903  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1904  * any new page allocations done using the specified allocation flags. Note
1905  * that the Radix tree operations will still use GFP_KERNEL, so you can't
1906  * expect to do this atomically or anything like that - but you can pass in
1907  * other page requirements.
1908  *
1909  * If the page does not get brought uptodate, return -EIO.
1910  */
1911 struct page *read_cache_page_gfp(struct address_space *mapping,
1912                                 pgoff_t index,
1913                                 gfp_t gfp)
1914 {
1915         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1916
1917         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1918 }
1919 EXPORT_SYMBOL(read_cache_page_gfp);
1920
1921 /**
1922  * read_cache_page - read into page cache, fill it if needed
1923  * @mapping:    the page's address_space
1924  * @index:      the page index
1925  * @filler:     function to perform the read
1926  * @data:       first arg to filler(data, page) function, often left as NULL
1927  *
1928  * Read into the page cache. If a page already exists, and PageUptodate() is
1929  * not set, try to fill the page then wait for it to become unlocked.
1930  *
1931  * If the page does not get brought uptodate, return -EIO.
1932  */
1933 struct page *read_cache_page(struct address_space *mapping,
1934                                 pgoff_t index,
1935                                 int (*filler)(void *, struct page *),
1936                                 void *data)
1937 {
1938         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1939 }
1940 EXPORT_SYMBOL(read_cache_page);
1941
1942 /*
1943  * The logic we want is
1944  *
1945  *      if suid or (sgid and xgrp)
1946  *              remove privs
1947  */
1948 int should_remove_suid(struct dentry *dentry)
1949 {
1950         mode_t mode = dentry->d_inode->i_mode;
1951         int kill = 0;
1952
1953         /* suid always must be killed */
1954         if (unlikely(mode & S_ISUID))
1955                 kill = ATTR_KILL_SUID;
1956
1957         /*
1958          * sgid without any exec bits is just a mandatory locking mark; leave
1959          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1960          */
1961         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1962                 kill |= ATTR_KILL_SGID;
1963
1964         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1965                 return kill;
1966
1967         return 0;
1968 }
1969 EXPORT_SYMBOL(should_remove_suid);
1970
1971 static int __remove_suid(struct dentry *dentry, int kill)
1972 {
1973         struct iattr newattrs;
1974
1975         newattrs.ia_valid = ATTR_FORCE | kill;
1976         return notify_change(dentry, &newattrs);
1977 }
1978
1979 int file_remove_suid(struct file *file)
1980 {
1981         struct dentry *dentry = file->f_path.dentry;
1982         struct inode *inode = dentry->d_inode;
1983         int killsuid;
1984         int killpriv;
1985         int error = 0;
1986
1987         /* Fast path for nothing security related */
1988         if (IS_NOSEC(inode))
1989                 return 0;
1990
1991         killsuid = should_remove_suid(dentry);
1992         killpriv = security_inode_need_killpriv(dentry);
1993
1994         if (killpriv < 0)
1995                 return killpriv;
1996         if (killpriv)
1997                 error = security_inode_killpriv(dentry);
1998         if (!error && killsuid)
1999                 error = __remove_suid(dentry, killsuid);
2000         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2001                 inode->i_flags |= S_NOSEC;
2002
2003         return error;
2004 }
2005 EXPORT_SYMBOL(file_remove_suid);
2006
2007 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2008                         const struct iovec *iov, size_t base, size_t bytes)
2009 {
2010         size_t copied = 0, left = 0;
2011
2012         while (bytes) {
2013                 char __user *buf = iov->iov_base + base;
2014                 int copy = min(bytes, iov->iov_len - base);
2015
2016                 base = 0;
2017                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2018                 copied += copy;
2019                 bytes -= copy;
2020                 vaddr += copy;
2021                 iov++;
2022
2023                 if (unlikely(left))
2024                         break;
2025         }
2026         return copied - left;
2027 }
2028
2029 /*
2030  * Copy as much as we can into the page and return the number of bytes which
2031  * were successfully copied.  If a fault is encountered then return the number of
2032  * bytes which were copied.
2033  */
2034 size_t iov_iter_copy_from_user_atomic(struct page *page,
2035                 struct iov_iter *i, unsigned long offset, size_t bytes)
2036 {
2037         char *kaddr;
2038         size_t copied;
2039
2040         BUG_ON(!in_atomic());
2041         kaddr = kmap_atomic(page, KM_USER0);
2042         if (likely(i->nr_segs == 1)) {
2043                 int left;
2044                 char __user *buf = i->iov->iov_base + i->iov_offset;
2045                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2046                 copied = bytes - left;
2047         } else {
2048                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2049                                                 i->iov, i->iov_offset, bytes);
2050         }
2051         kunmap_atomic(kaddr, KM_USER0);
2052
2053         return copied;
2054 }
2055 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2056
2057 /*
2058  * This has the same sideeffects and return value as
2059  * iov_iter_copy_from_user_atomic().
2060  * The difference is that it attempts to resolve faults.
2061  * Page must not be locked.
2062  */
2063 size_t iov_iter_copy_from_user(struct page *page,
2064                 struct iov_iter *i, unsigned long offset, size_t bytes)
2065 {
2066         char *kaddr;
2067         size_t copied;
2068
2069         kaddr = kmap(page);
2070         if (likely(i->nr_segs == 1)) {
2071                 int left;
2072                 char __user *buf = i->iov->iov_base + i->iov_offset;
2073                 left = __copy_from_user(kaddr + offset, buf, bytes);
2074                 copied = bytes - left;
2075         } else {
2076                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2077                                                 i->iov, i->iov_offset, bytes);
2078         }
2079         kunmap(page);
2080         return copied;
2081 }
2082 EXPORT_SYMBOL(iov_iter_copy_from_user);
2083
2084 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2085 {
2086         BUG_ON(i->count < bytes);
2087
2088         if (likely(i->nr_segs == 1)) {
2089                 i->iov_offset += bytes;
2090                 i->count -= bytes;
2091         } else {
2092                 const struct iovec *iov = i->iov;
2093                 size_t base = i->iov_offset;
2094
2095                 /*
2096                  * The !iov->iov_len check ensures we skip over unlikely
2097                  * zero-length segments (without overruning the iovec).
2098                  */
2099                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2100                         int copy;
2101
2102                         copy = min(bytes, iov->iov_len - base);
2103                         BUG_ON(!i->count || i->count < copy);
2104                         i->count -= copy;
2105                         bytes -= copy;
2106                         base += copy;
2107                         if (iov->iov_len == base) {
2108                                 iov++;
2109                                 base = 0;
2110                         }
2111                 }
2112                 i->iov = iov;
2113                 i->iov_offset = base;
2114         }
2115 }
2116 EXPORT_SYMBOL(iov_iter_advance);
2117
2118 /*
2119  * Fault in the first iovec of the given iov_iter, to a maximum length
2120  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2121  * accessed (ie. because it is an invalid address).
2122  *
2123  * writev-intensive code may want this to prefault several iovecs -- that
2124  * would be possible (callers must not rely on the fact that _only_ the
2125  * first iovec will be faulted with the current implementation).
2126  */
2127 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2128 {
2129         char __user *buf = i->iov->iov_base + i->iov_offset;
2130         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2131         return fault_in_pages_readable(buf, bytes);
2132 }
2133 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2134
2135 /*
2136  * Return the count of just the current iov_iter segment.
2137  */
2138 size_t iov_iter_single_seg_count(struct iov_iter *i)
2139 {
2140         const struct iovec *iov = i->iov;
2141         if (i->nr_segs == 1)
2142                 return i->count;
2143         else
2144                 return min(i->count, iov->iov_len - i->iov_offset);
2145 }
2146 EXPORT_SYMBOL(iov_iter_single_seg_count);
2147
2148 /*
2149  * Performs necessary checks before doing a write
2150  *
2151  * Can adjust writing position or amount of bytes to write.
2152  * Returns appropriate error code that caller should return or
2153  * zero in case that write should be allowed.
2154  */
2155 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2156 {
2157         struct inode *inode = file->f_mapping->host;
2158         unsigned long limit = rlimit(RLIMIT_FSIZE);
2159
2160         if (unlikely(*pos < 0))
2161                 return -EINVAL;
2162
2163         if (!isblk) {
2164                 /* FIXME: this is for backwards compatibility with 2.4 */
2165                 if (file->f_flags & O_APPEND)
2166                         *pos = i_size_read(inode);
2167
2168                 if (limit != RLIM_INFINITY) {
2169                         if (*pos >= limit) {
2170                                 send_sig(SIGXFSZ, current, 0);
2171                                 return -EFBIG;
2172                         }
2173                         if (*count > limit - (typeof(limit))*pos) {
2174                                 *count = limit - (typeof(limit))*pos;
2175                         }
2176                 }
2177         }
2178
2179         /*
2180          * LFS rule
2181          */
2182         if (unlikely(*pos + *count > MAX_NON_LFS &&
2183                                 !(file->f_flags & O_LARGEFILE))) {
2184                 if (*pos >= MAX_NON_LFS) {
2185                         return -EFBIG;
2186                 }
2187                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2188                         *count = MAX_NON_LFS - (unsigned long)*pos;
2189                 }
2190         }
2191
2192         /*
2193          * Are we about to exceed the fs block limit ?
2194          *
2195          * If we have written data it becomes a short write.  If we have
2196          * exceeded without writing data we send a signal and return EFBIG.
2197          * Linus frestrict idea will clean these up nicely..
2198          */
2199         if (likely(!isblk)) {
2200                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2201                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2202                                 return -EFBIG;
2203                         }
2204                         /* zero-length writes at ->s_maxbytes are OK */
2205                 }
2206
2207                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2208                         *count = inode->i_sb->s_maxbytes - *pos;
2209         } else {
2210 #ifdef CONFIG_BLOCK
2211                 loff_t isize;
2212                 if (bdev_read_only(I_BDEV(inode)))
2213                         return -EPERM;
2214                 isize = i_size_read(inode);
2215                 if (*pos >= isize) {
2216                         if (*count || *pos > isize)
2217                                 return -ENOSPC;
2218                 }
2219
2220                 if (*pos + *count > isize)
2221                         *count = isize - *pos;
2222 #else
2223                 return -EPERM;
2224 #endif
2225         }
2226         return 0;
2227 }
2228 EXPORT_SYMBOL(generic_write_checks);
2229
2230 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2231                                 loff_t pos, unsigned len, unsigned flags,
2232                                 struct page **pagep, void **fsdata)
2233 {
2234         const struct address_space_operations *aops = mapping->a_ops;
2235
2236         return aops->write_begin(file, mapping, pos, len, flags,
2237                                                         pagep, fsdata);
2238 }
2239 EXPORT_SYMBOL(pagecache_write_begin);
2240
2241 int pagecache_write_end(struct file *file, struct address_space *mapping,
2242                                 loff_t pos, unsigned len, unsigned copied,
2243                                 struct page *page, void *fsdata)
2244 {
2245         const struct address_space_operations *aops = mapping->a_ops;
2246
2247         mark_page_accessed(page);
2248         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2249 }
2250 EXPORT_SYMBOL(pagecache_write_end);
2251
2252 ssize_t
2253 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2254                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2255                 size_t count, size_t ocount)
2256 {
2257         struct file     *file = iocb->ki_filp;
2258         struct address_space *mapping = file->f_mapping;
2259         struct inode    *inode = mapping->host;
2260         ssize_t         written;
2261         size_t          write_len;
2262         pgoff_t         end;
2263
2264         if (count != ocount)
2265                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2266
2267         write_len = iov_length(iov, *nr_segs);
2268         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2269
2270         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2271         if (written)
2272                 goto out;
2273
2274         /*
2275          * After a write we want buffered reads to be sure to go to disk to get
2276          * the new data.  We invalidate clean cached page from the region we're
2277          * about to write.  We do this *before* the write so that we can return
2278          * without clobbering -EIOCBQUEUED from ->direct_IO().
2279          */
2280         if (mapping->nrpages) {
2281                 written = invalidate_inode_pages2_range(mapping,
2282                                         pos >> PAGE_CACHE_SHIFT, end);
2283                 /*
2284                  * If a page can not be invalidated, return 0 to fall back
2285                  * to buffered write.
2286                  */
2287                 if (written) {
2288                         if (written == -EBUSY)
2289                                 return 0;
2290                         goto out;
2291                 }
2292         }
2293
2294         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2295
2296         /*
2297          * Finally, try again to invalidate clean pages which might have been
2298          * cached by non-direct readahead, or faulted in by get_user_pages()
2299          * if the source of the write was an mmap'ed region of the file
2300          * we're writing.  Either one is a pretty crazy thing to do,
2301          * so we don't support it 100%.  If this invalidation
2302          * fails, tough, the write still worked...
2303          */
2304         if (mapping->nrpages) {
2305                 invalidate_inode_pages2_range(mapping,
2306                                               pos >> PAGE_CACHE_SHIFT, end);
2307         }
2308
2309         if (written > 0) {
2310                 pos += written;
2311                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2312                         i_size_write(inode, pos);
2313                         mark_inode_dirty(inode);
2314                 }
2315                 *ppos = pos;
2316         }
2317 out:
2318         return written;
2319 }
2320 EXPORT_SYMBOL(generic_file_direct_write);
2321
2322 /*
2323  * Find or create a page at the given pagecache position. Return the locked
2324  * page. This function is specifically for buffered writes.
2325  */
2326 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2327                                         pgoff_t index, unsigned flags)
2328 {
2329         int status;
2330         struct page *page;
2331         gfp_t gfp_notmask = 0;
2332         if (flags & AOP_FLAG_NOFS)
2333                 gfp_notmask = __GFP_FS;
2334 repeat:
2335         page = find_lock_page(mapping, index);
2336         if (page)
2337                 goto found;
2338
2339         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2340         if (!page)
2341                 return NULL;
2342         status = add_to_page_cache_lru(page, mapping, index,
2343                                                 GFP_KERNEL & ~gfp_notmask);
2344         if (unlikely(status)) {
2345                 page_cache_release(page);
2346                 if (status == -EEXIST)
2347                         goto repeat;
2348                 return NULL;
2349         }
2350 found:
2351         wait_on_page_writeback(page);
2352         return page;
2353 }
2354 EXPORT_SYMBOL(grab_cache_page_write_begin);
2355
2356 static ssize_t generic_perform_write(struct file *file,
2357                                 struct iov_iter *i, loff_t pos)
2358 {
2359         struct address_space *mapping = file->f_mapping;
2360         const struct address_space_operations *a_ops = mapping->a_ops;
2361         long status = 0;
2362         ssize_t written = 0;
2363         unsigned int flags = 0;
2364
2365         /*
2366          * Copies from kernel address space cannot fail (NFSD is a big user).
2367          */
2368         if (segment_eq(get_fs(), KERNEL_DS))
2369                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2370
2371         do {
2372                 struct page *page;
2373                 unsigned long offset;   /* Offset into pagecache page */
2374                 unsigned long bytes;    /* Bytes to write to page */
2375                 size_t copied;          /* Bytes copied from user */
2376                 void *fsdata;
2377
2378                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2379                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2380                                                 iov_iter_count(i));
2381
2382 again:
2383
2384                 /*
2385                  * Bring in the user page that we will copy from _first_.
2386                  * Otherwise there's a nasty deadlock on copying from the
2387                  * same page as we're writing to, without it being marked
2388                  * up-to-date.
2389                  *
2390                  * Not only is this an optimisation, but it is also required
2391                  * to check that the address is actually valid, when atomic
2392                  * usercopies are used, below.
2393                  */
2394                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2395                         status = -EFAULT;
2396                         break;
2397                 }
2398
2399                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2400                                                 &page, &fsdata);
2401                 if (unlikely(status))
2402                         break;
2403
2404                 if (mapping_writably_mapped(mapping))
2405                         flush_dcache_page(page);
2406
2407                 pagefault_disable();
2408                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2409                 pagefault_enable();
2410                 flush_dcache_page(page);
2411
2412                 mark_page_accessed(page);
2413                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2414                                                 page, fsdata);
2415                 if (unlikely(status < 0))
2416                         break;
2417                 copied = status;
2418
2419                 cond_resched();
2420
2421                 iov_iter_advance(i, copied);
2422                 if (unlikely(copied == 0)) {
2423                         /*
2424                          * If we were unable to copy any data at all, we must
2425                          * fall back to a single segment length write.
2426                          *
2427                          * If we didn't fallback here, we could livelock
2428                          * because not all segments in the iov can be copied at
2429                          * once without a pagefault.
2430                          */
2431                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2432                                                 iov_iter_single_seg_count(i));
2433                         goto again;
2434                 }
2435                 pos += copied;
2436                 written += copied;
2437
2438                 balance_dirty_pages_ratelimited(mapping);
2439
2440         } while (iov_iter_count(i));
2441
2442         return written ? written : status;
2443 }
2444
2445 ssize_t
2446 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2447                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2448                 size_t count, ssize_t written)
2449 {
2450         struct file *file = iocb->ki_filp;
2451         ssize_t status;
2452         struct iov_iter i;
2453
2454         iov_iter_init(&i, iov, nr_segs, count, written);
2455         status = generic_perform_write(file, &i, pos);
2456
2457         if (likely(status >= 0)) {
2458                 written += status;
2459                 *ppos = pos + status;
2460         }
2461         
2462         return written ? written : status;
2463 }
2464 EXPORT_SYMBOL(generic_file_buffered_write);
2465
2466 /**
2467  * __generic_file_aio_write - write data to a file
2468  * @iocb:       IO state structure (file, offset, etc.)
2469  * @iov:        vector with data to write
2470  * @nr_segs:    number of segments in the vector
2471  * @ppos:       position where to write
2472  *
2473  * This function does all the work needed for actually writing data to a
2474  * file. It does all basic checks, removes SUID from the file, updates
2475  * modification times and calls proper subroutines depending on whether we
2476  * do direct IO or a standard buffered write.
2477  *
2478  * It expects i_mutex to be grabbed unless we work on a block device or similar
2479  * object which does not need locking at all.
2480  *
2481  * This function does *not* take care of syncing data in case of O_SYNC write.
2482  * A caller has to handle it. This is mainly due to the fact that we want to
2483  * avoid syncing under i_mutex.
2484  */
2485 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2486                                  unsigned long nr_segs, loff_t *ppos)
2487 {
2488         struct file *file = iocb->ki_filp;
2489         struct address_space * mapping = file->f_mapping;
2490         size_t ocount;          /* original count */
2491         size_t count;           /* after file limit checks */
2492         struct inode    *inode = mapping->host;
2493         loff_t          pos;
2494         ssize_t         written;
2495         ssize_t         err;
2496
2497         ocount = 0;
2498         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2499         if (err)
2500                 return err;
2501
2502         count = ocount;
2503         pos = *ppos;
2504
2505         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2506
2507         /* We can write back this queue in page reclaim */
2508         current->backing_dev_info = mapping->backing_dev_info;
2509         written = 0;
2510
2511         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2512         if (err)
2513                 goto out;
2514
2515         if (count == 0)
2516                 goto out;
2517
2518         err = file_remove_suid(file);
2519         if (err)
2520                 goto out;
2521
2522         file_update_time(file);
2523
2524         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2525         if (unlikely(file->f_flags & O_DIRECT)) {
2526                 loff_t endbyte;
2527                 ssize_t written_buffered;
2528
2529                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2530                                                         ppos, count, ocount);
2531                 if (written < 0 || written == count)
2532                         goto out;
2533                 /*
2534                  * direct-io write to a hole: fall through to buffered I/O
2535                  * for completing the rest of the request.
2536                  */
2537                 pos += written;
2538                 count -= written;
2539                 written_buffered = generic_file_buffered_write(iocb, iov,
2540                                                 nr_segs, pos, ppos, count,
2541                                                 written);
2542                 /*
2543                  * If generic_file_buffered_write() retuned a synchronous error
2544                  * then we want to return the number of bytes which were
2545                  * direct-written, or the error code if that was zero.  Note
2546                  * that this differs from normal direct-io semantics, which
2547                  * will return -EFOO even if some bytes were written.
2548                  */
2549                 if (written_buffered < 0) {
2550                         err = written_buffered;
2551                         goto out;
2552                 }
2553
2554                 /*
2555                  * We need to ensure that the page cache pages are written to
2556                  * disk and invalidated to preserve the expected O_DIRECT
2557                  * semantics.
2558                  */
2559                 endbyte = pos + written_buffered - written - 1;
2560                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2561                 if (err == 0) {
2562                         written = written_buffered;
2563                         invalidate_mapping_pages(mapping,
2564                                                  pos >> PAGE_CACHE_SHIFT,
2565                                                  endbyte >> PAGE_CACHE_SHIFT);
2566                 } else {
2567                         /*
2568                          * We don't know how much we wrote, so just return
2569                          * the number of bytes which were direct-written
2570                          */
2571                 }
2572         } else {
2573                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2574                                 pos, ppos, count, written);
2575         }
2576 out:
2577         current->backing_dev_info = NULL;
2578         return written ? written : err;
2579 }
2580 EXPORT_SYMBOL(__generic_file_aio_write);
2581
2582 /**
2583  * generic_file_aio_write - write data to a file
2584  * @iocb:       IO state structure
2585  * @iov:        vector with data to write
2586  * @nr_segs:    number of segments in the vector
2587  * @pos:        position in file where to write
2588  *
2589  * This is a wrapper around __generic_file_aio_write() to be used by most
2590  * filesystems. It takes care of syncing the file in case of O_SYNC file
2591  * and acquires i_mutex as needed.
2592  */
2593 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2594                 unsigned long nr_segs, loff_t pos)
2595 {
2596         struct file *file = iocb->ki_filp;
2597         struct inode *inode = file->f_mapping->host;
2598         struct blk_plug plug;
2599         ssize_t ret;
2600
2601         BUG_ON(iocb->ki_pos != pos);
2602
2603         mutex_lock(&inode->i_mutex);
2604         blk_start_plug(&plug);
2605         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2606         mutex_unlock(&inode->i_mutex);
2607
2608         if (ret > 0 || ret == -EIOCBQUEUED) {
2609                 ssize_t err;
2610
2611                 err = generic_write_sync(file, pos, ret);
2612                 if (err < 0 && ret > 0)
2613                         ret = err;
2614         }
2615         blk_finish_plug(&plug);
2616         return ret;
2617 }
2618 EXPORT_SYMBOL(generic_file_aio_write);
2619
2620 /**
2621  * try_to_release_page() - release old fs-specific metadata on a page
2622  *
2623  * @page: the page which the kernel is trying to free
2624  * @gfp_mask: memory allocation flags (and I/O mode)
2625  *
2626  * The address_space is to try to release any data against the page
2627  * (presumably at page->private).  If the release was successful, return `1'.
2628  * Otherwise return zero.
2629  *
2630  * This may also be called if PG_fscache is set on a page, indicating that the
2631  * page is known to the local caching routines.
2632  *
2633  * The @gfp_mask argument specifies whether I/O may be performed to release
2634  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2635  *
2636  */
2637 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2638 {
2639         struct address_space * const mapping = page->mapping;
2640
2641         BUG_ON(!PageLocked(page));
2642         if (PageWriteback(page))
2643                 return 0;
2644
2645         if (mapping && mapping->a_ops->releasepage)
2646                 return mapping->a_ops->releasepage(page, gfp_mask);
2647         return try_to_free_buffers(page);
2648 }
2649
2650 EXPORT_SYMBOL(try_to_release_page);