mm: make swapin readahead skip over holes
[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/export.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/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
397         VM_BUG_ON(!PageLocked(old));
398         VM_BUG_ON(!PageLocked(new));
399         VM_BUG_ON(new->mapping);
400
401         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
402         if (!error) {
403                 struct address_space *mapping = old->mapping;
404                 void (*freepage)(struct page *);
405
406                 pgoff_t offset = old->index;
407                 freepage = mapping->a_ops->freepage;
408
409                 page_cache_get(new);
410                 new->mapping = mapping;
411                 new->index = offset;
412
413                 spin_lock_irq(&mapping->tree_lock);
414                 __delete_from_page_cache(old);
415                 error = radix_tree_insert(&mapping->page_tree, offset, new);
416                 BUG_ON(error);
417                 mapping->nrpages++;
418                 __inc_zone_page_state(new, NR_FILE_PAGES);
419                 if (PageSwapBacked(new))
420                         __inc_zone_page_state(new, NR_SHMEM);
421                 spin_unlock_irq(&mapping->tree_lock);
422                 /* mem_cgroup codes must not be called under tree_lock */
423                 mem_cgroup_replace_page_cache(old, new);
424                 radix_tree_preload_end();
425                 if (freepage)
426                         freepage(old);
427                 page_cache_release(old);
428         }
429
430         return error;
431 }
432 EXPORT_SYMBOL_GPL(replace_page_cache_page);
433
434 /**
435  * add_to_page_cache_locked - add a locked page to the pagecache
436  * @page:       page to add
437  * @mapping:    the page's address_space
438  * @offset:     page index
439  * @gfp_mask:   page allocation mode
440  *
441  * This function is used to add a page to the pagecache. It must be locked.
442  * This function does not add the page to the LRU.  The caller must do that.
443  */
444 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
445                 pgoff_t offset, gfp_t gfp_mask)
446 {
447         int error;
448
449         VM_BUG_ON(!PageLocked(page));
450         VM_BUG_ON(PageSwapBacked(page));
451
452         error = mem_cgroup_cache_charge(page, current->mm,
453                                         gfp_mask & GFP_RECLAIM_MASK);
454         if (error)
455                 goto out;
456
457         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
458         if (error == 0) {
459                 page_cache_get(page);
460                 page->mapping = mapping;
461                 page->index = offset;
462
463                 spin_lock_irq(&mapping->tree_lock);
464                 error = radix_tree_insert(&mapping->page_tree, offset, page);
465                 if (likely(!error)) {
466                         mapping->nrpages++;
467                         __inc_zone_page_state(page, NR_FILE_PAGES);
468                         spin_unlock_irq(&mapping->tree_lock);
469                 } else {
470                         page->mapping = NULL;
471                         /* Leave page->index set: truncation relies upon it */
472                         spin_unlock_irq(&mapping->tree_lock);
473                         mem_cgroup_uncharge_cache_page(page);
474                         page_cache_release(page);
475                 }
476                 radix_tree_preload_end();
477         } else
478                 mem_cgroup_uncharge_cache_page(page);
479 out:
480         return error;
481 }
482 EXPORT_SYMBOL(add_to_page_cache_locked);
483
484 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
485                                 pgoff_t offset, gfp_t gfp_mask)
486 {
487         int ret;
488
489         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
490         if (ret == 0)
491                 lru_cache_add_file(page);
492         return ret;
493 }
494 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
495
496 #ifdef CONFIG_NUMA
497 struct page *__page_cache_alloc(gfp_t gfp)
498 {
499         int n;
500         struct page *page;
501
502         if (cpuset_do_page_mem_spread()) {
503                 get_mems_allowed();
504                 n = cpuset_mem_spread_node();
505                 page = alloc_pages_exact_node(n, gfp, 0);
506                 put_mems_allowed();
507                 return page;
508         }
509         return alloc_pages(gfp, 0);
510 }
511 EXPORT_SYMBOL(__page_cache_alloc);
512 #endif
513
514 /*
515  * In order to wait for pages to become available there must be
516  * waitqueues associated with pages. By using a hash table of
517  * waitqueues where the bucket discipline is to maintain all
518  * waiters on the same queue and wake all when any of the pages
519  * become available, and for the woken contexts to check to be
520  * sure the appropriate page became available, this saves space
521  * at a cost of "thundering herd" phenomena during rare hash
522  * collisions.
523  */
524 static wait_queue_head_t *page_waitqueue(struct page *page)
525 {
526         const struct zone *zone = page_zone(page);
527
528         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
529 }
530
531 static inline void wake_up_page(struct page *page, int bit)
532 {
533         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
534 }
535
536 void wait_on_page_bit(struct page *page, int bit_nr)
537 {
538         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
539
540         if (test_bit(bit_nr, &page->flags))
541                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
542                                                         TASK_UNINTERRUPTIBLE);
543 }
544 EXPORT_SYMBOL(wait_on_page_bit);
545
546 int wait_on_page_bit_killable(struct page *page, int bit_nr)
547 {
548         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
549
550         if (!test_bit(bit_nr, &page->flags))
551                 return 0;
552
553         return __wait_on_bit(page_waitqueue(page), &wait,
554                              sleep_on_page_killable, TASK_KILLABLE);
555 }
556
557 /**
558  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
559  * @page: Page defining the wait queue of interest
560  * @waiter: Waiter to add to the queue
561  *
562  * Add an arbitrary @waiter to the wait queue for the nominated @page.
563  */
564 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
565 {
566         wait_queue_head_t *q = page_waitqueue(page);
567         unsigned long flags;
568
569         spin_lock_irqsave(&q->lock, flags);
570         __add_wait_queue(q, waiter);
571         spin_unlock_irqrestore(&q->lock, flags);
572 }
573 EXPORT_SYMBOL_GPL(add_page_wait_queue);
574
575 /**
576  * unlock_page - unlock a locked page
577  * @page: the page
578  *
579  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
580  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
581  * mechananism between PageLocked pages and PageWriteback pages is shared.
582  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
583  *
584  * The mb is necessary to enforce ordering between the clear_bit and the read
585  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
586  */
587 void unlock_page(struct page *page)
588 {
589         VM_BUG_ON(!PageLocked(page));
590         clear_bit_unlock(PG_locked, &page->flags);
591         smp_mb__after_clear_bit();
592         wake_up_page(page, PG_locked);
593 }
594 EXPORT_SYMBOL(unlock_page);
595
596 /**
597  * end_page_writeback - end writeback against a page
598  * @page: the page
599  */
600 void end_page_writeback(struct page *page)
601 {
602         if (TestClearPageReclaim(page))
603                 rotate_reclaimable_page(page);
604
605         if (!test_clear_page_writeback(page))
606                 BUG();
607
608         smp_mb__after_clear_bit();
609         wake_up_page(page, PG_writeback);
610 }
611 EXPORT_SYMBOL(end_page_writeback);
612
613 /**
614  * __lock_page - get a lock on the page, assuming we need to sleep to get it
615  * @page: the page to lock
616  */
617 void __lock_page(struct page *page)
618 {
619         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
620
621         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
622                                                         TASK_UNINTERRUPTIBLE);
623 }
624 EXPORT_SYMBOL(__lock_page);
625
626 int __lock_page_killable(struct page *page)
627 {
628         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
629
630         return __wait_on_bit_lock(page_waitqueue(page), &wait,
631                                         sleep_on_page_killable, TASK_KILLABLE);
632 }
633 EXPORT_SYMBOL_GPL(__lock_page_killable);
634
635 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
636                          unsigned int flags)
637 {
638         if (flags & FAULT_FLAG_ALLOW_RETRY) {
639                 /*
640                  * CAUTION! In this case, mmap_sem is not released
641                  * even though return 0.
642                  */
643                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
644                         return 0;
645
646                 up_read(&mm->mmap_sem);
647                 if (flags & FAULT_FLAG_KILLABLE)
648                         wait_on_page_locked_killable(page);
649                 else
650                         wait_on_page_locked(page);
651                 return 0;
652         } else {
653                 if (flags & FAULT_FLAG_KILLABLE) {
654                         int ret;
655
656                         ret = __lock_page_killable(page);
657                         if (ret) {
658                                 up_read(&mm->mmap_sem);
659                                 return 0;
660                         }
661                 } else
662                         __lock_page(page);
663                 return 1;
664         }
665 }
666
667 /**
668  * find_get_page - find and get a page reference
669  * @mapping: the address_space to search
670  * @offset: the page index
671  *
672  * Is there a pagecache struct page at the given (mapping, offset) tuple?
673  * If yes, increment its refcount and return it; if no, return NULL.
674  */
675 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
676 {
677         void **pagep;
678         struct page *page;
679
680         rcu_read_lock();
681 repeat:
682         page = NULL;
683         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
684         if (pagep) {
685                 page = radix_tree_deref_slot(pagep);
686                 if (unlikely(!page))
687                         goto out;
688                 if (radix_tree_exception(page)) {
689                         if (radix_tree_deref_retry(page))
690                                 goto repeat;
691                         /*
692                          * Otherwise, shmem/tmpfs must be storing a swap entry
693                          * here as an exceptional entry: so return it without
694                          * attempting to raise page count.
695                          */
696                         goto out;
697                 }
698                 if (!page_cache_get_speculative(page))
699                         goto repeat;
700
701                 /*
702                  * Has the page moved?
703                  * This is part of the lockless pagecache protocol. See
704                  * include/linux/pagemap.h for details.
705                  */
706                 if (unlikely(page != *pagep)) {
707                         page_cache_release(page);
708                         goto repeat;
709                 }
710         }
711 out:
712         rcu_read_unlock();
713
714         return page;
715 }
716 EXPORT_SYMBOL(find_get_page);
717
718 /**
719  * find_lock_page - locate, pin and lock a pagecache page
720  * @mapping: the address_space to search
721  * @offset: the page index
722  *
723  * Locates the desired pagecache page, locks it, increments its reference
724  * count and returns its address.
725  *
726  * Returns zero if the page was not present. find_lock_page() may sleep.
727  */
728 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
729 {
730         struct page *page;
731
732 repeat:
733         page = find_get_page(mapping, offset);
734         if (page && !radix_tree_exception(page)) {
735                 lock_page(page);
736                 /* Has the page been truncated? */
737                 if (unlikely(page->mapping != mapping)) {
738                         unlock_page(page);
739                         page_cache_release(page);
740                         goto repeat;
741                 }
742                 VM_BUG_ON(page->index != offset);
743         }
744         return page;
745 }
746 EXPORT_SYMBOL(find_lock_page);
747
748 /**
749  * find_or_create_page - locate or add a pagecache page
750  * @mapping: the page's address_space
751  * @index: the page's index into the mapping
752  * @gfp_mask: page allocation mode
753  *
754  * Locates a page in the pagecache.  If the page is not present, a new page
755  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
756  * LRU list.  The returned page is locked and has its reference count
757  * incremented.
758  *
759  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
760  * allocation!
761  *
762  * find_or_create_page() returns the desired page's address, or zero on
763  * memory exhaustion.
764  */
765 struct page *find_or_create_page(struct address_space *mapping,
766                 pgoff_t index, gfp_t gfp_mask)
767 {
768         struct page *page;
769         int err;
770 repeat:
771         page = find_lock_page(mapping, index);
772         if (!page) {
773                 page = __page_cache_alloc(gfp_mask);
774                 if (!page)
775                         return NULL;
776                 /*
777                  * We want a regular kernel memory (not highmem or DMA etc)
778                  * allocation for the radix tree nodes, but we need to honour
779                  * the context-specific requirements the caller has asked for.
780                  * GFP_RECLAIM_MASK collects those requirements.
781                  */
782                 err = add_to_page_cache_lru(page, mapping, index,
783                         (gfp_mask & GFP_RECLAIM_MASK));
784                 if (unlikely(err)) {
785                         page_cache_release(page);
786                         page = NULL;
787                         if (err == -EEXIST)
788                                 goto repeat;
789                 }
790         }
791         return page;
792 }
793 EXPORT_SYMBOL(find_or_create_page);
794
795 /**
796  * find_get_pages - gang pagecache lookup
797  * @mapping:    The address_space to search
798  * @start:      The starting page index
799  * @nr_pages:   The maximum number of pages
800  * @pages:      Where the resulting pages are placed
801  *
802  * find_get_pages() will search for and return a group of up to
803  * @nr_pages pages in the mapping.  The pages are placed at @pages.
804  * find_get_pages() takes a reference against the returned pages.
805  *
806  * The search returns a group of mapping-contiguous pages with ascending
807  * indexes.  There may be holes in the indices due to not-present pages.
808  *
809  * find_get_pages() returns the number of pages which were found.
810  */
811 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
812                             unsigned int nr_pages, struct page **pages)
813 {
814         unsigned int i;
815         unsigned int ret;
816         unsigned int nr_found, nr_skip;
817
818         rcu_read_lock();
819 restart:
820         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
821                                 (void ***)pages, NULL, start, nr_pages);
822         ret = 0;
823         nr_skip = 0;
824         for (i = 0; i < nr_found; i++) {
825                 struct page *page;
826 repeat:
827                 page = radix_tree_deref_slot((void **)pages[i]);
828                 if (unlikely(!page))
829                         continue;
830
831                 if (radix_tree_exception(page)) {
832                         if (radix_tree_deref_retry(page)) {
833                                 /*
834                                  * Transient condition which can only trigger
835                                  * when entry at index 0 moves out of or back
836                                  * to root: none yet gotten, safe to restart.
837                                  */
838                                 WARN_ON(start | i);
839                                 goto restart;
840                         }
841                         /*
842                          * Otherwise, shmem/tmpfs must be storing a swap entry
843                          * here as an exceptional entry: so skip over it -
844                          * we only reach this from invalidate_mapping_pages().
845                          */
846                         nr_skip++;
847                         continue;
848                 }
849
850                 if (!page_cache_get_speculative(page))
851                         goto repeat;
852
853                 /* Has the page moved? */
854                 if (unlikely(page != *((void **)pages[i]))) {
855                         page_cache_release(page);
856                         goto repeat;
857                 }
858
859                 pages[ret] = page;
860                 ret++;
861         }
862
863         /*
864          * If all entries were removed before we could secure them,
865          * try again, because callers stop trying once 0 is returned.
866          */
867         if (unlikely(!ret && nr_found > nr_skip))
868                 goto restart;
869         rcu_read_unlock();
870         return ret;
871 }
872
873 /**
874  * find_get_pages_contig - gang contiguous pagecache lookup
875  * @mapping:    The address_space to search
876  * @index:      The starting page index
877  * @nr_pages:   The maximum number of pages
878  * @pages:      Where the resulting pages are placed
879  *
880  * find_get_pages_contig() works exactly like find_get_pages(), except
881  * that the returned number of pages are guaranteed to be contiguous.
882  *
883  * find_get_pages_contig() returns the number of pages which were found.
884  */
885 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
886                                unsigned int nr_pages, struct page **pages)
887 {
888         unsigned int i;
889         unsigned int ret;
890         unsigned int nr_found;
891
892         rcu_read_lock();
893 restart:
894         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
895                                 (void ***)pages, NULL, index, nr_pages);
896         ret = 0;
897         for (i = 0; i < nr_found; i++) {
898                 struct page *page;
899 repeat:
900                 page = radix_tree_deref_slot((void **)pages[i]);
901                 if (unlikely(!page))
902                         continue;
903
904                 if (radix_tree_exception(page)) {
905                         if (radix_tree_deref_retry(page)) {
906                                 /*
907                                  * Transient condition which can only trigger
908                                  * when entry at index 0 moves out of or back
909                                  * to root: none yet gotten, safe to restart.
910                                  */
911                                 goto restart;
912                         }
913                         /*
914                          * Otherwise, shmem/tmpfs must be storing a swap entry
915                          * here as an exceptional entry: so stop looking for
916                          * contiguous pages.
917                          */
918                         break;
919                 }
920
921                 if (!page_cache_get_speculative(page))
922                         goto repeat;
923
924                 /* Has the page moved? */
925                 if (unlikely(page != *((void **)pages[i]))) {
926                         page_cache_release(page);
927                         goto repeat;
928                 }
929
930                 /*
931                  * must check mapping and index after taking the ref.
932                  * otherwise we can get both false positives and false
933                  * negatives, which is just confusing to the caller.
934                  */
935                 if (page->mapping == NULL || page->index != index) {
936                         page_cache_release(page);
937                         break;
938                 }
939
940                 pages[ret] = page;
941                 ret++;
942                 index++;
943         }
944         rcu_read_unlock();
945         return ret;
946 }
947 EXPORT_SYMBOL(find_get_pages_contig);
948
949 /**
950  * find_get_pages_tag - find and return pages that match @tag
951  * @mapping:    the address_space to search
952  * @index:      the starting page index
953  * @tag:        the tag index
954  * @nr_pages:   the maximum number of pages
955  * @pages:      where the resulting pages are placed
956  *
957  * Like find_get_pages, except we only return pages which are tagged with
958  * @tag.   We update @index to index the next page for the traversal.
959  */
960 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
961                         int tag, unsigned int nr_pages, struct page **pages)
962 {
963         unsigned int i;
964         unsigned int ret;
965         unsigned int nr_found;
966
967         rcu_read_lock();
968 restart:
969         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
970                                 (void ***)pages, *index, nr_pages, tag);
971         ret = 0;
972         for (i = 0; i < nr_found; i++) {
973                 struct page *page;
974 repeat:
975                 page = radix_tree_deref_slot((void **)pages[i]);
976                 if (unlikely(!page))
977                         continue;
978
979                 if (radix_tree_exception(page)) {
980                         if (radix_tree_deref_retry(page)) {
981                                 /*
982                                  * Transient condition which can only trigger
983                                  * when entry at index 0 moves out of or back
984                                  * to root: none yet gotten, safe to restart.
985                                  */
986                                 goto restart;
987                         }
988                         /*
989                          * This function is never used on a shmem/tmpfs
990                          * mapping, so a swap entry won't be found here.
991                          */
992                         BUG();
993                 }
994
995                 if (!page_cache_get_speculative(page))
996                         goto repeat;
997
998                 /* Has the page moved? */
999                 if (unlikely(page != *((void **)pages[i]))) {
1000                         page_cache_release(page);
1001                         goto repeat;
1002                 }
1003
1004                 pages[ret] = page;
1005                 ret++;
1006         }
1007
1008         /*
1009          * If all entries were removed before we could secure them,
1010          * try again, because callers stop trying once 0 is returned.
1011          */
1012         if (unlikely(!ret && nr_found))
1013                 goto restart;
1014         rcu_read_unlock();
1015
1016         if (ret)
1017                 *index = pages[ret - 1]->index + 1;
1018
1019         return ret;
1020 }
1021 EXPORT_SYMBOL(find_get_pages_tag);
1022
1023 /**
1024  * grab_cache_page_nowait - returns locked page at given index in given cache
1025  * @mapping: target address_space
1026  * @index: the page index
1027  *
1028  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1029  * This is intended for speculative data generators, where the data can
1030  * be regenerated if the page couldn't be grabbed.  This routine should
1031  * be safe to call while holding the lock for another page.
1032  *
1033  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1034  * and deadlock against the caller's locked page.
1035  */
1036 struct page *
1037 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1038 {
1039         struct page *page = find_get_page(mapping, index);
1040
1041         if (page) {
1042                 if (trylock_page(page))
1043                         return page;
1044                 page_cache_release(page);
1045                 return NULL;
1046         }
1047         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1048         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1049                 page_cache_release(page);
1050                 page = NULL;
1051         }
1052         return page;
1053 }
1054 EXPORT_SYMBOL(grab_cache_page_nowait);
1055
1056 /*
1057  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1058  * a _large_ part of the i/o request. Imagine the worst scenario:
1059  *
1060  *      ---R__________________________________________B__________
1061  *         ^ reading here                             ^ bad block(assume 4k)
1062  *
1063  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1064  * => failing the whole request => read(R) => read(R+1) =>
1065  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1066  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1067  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1068  *
1069  * It is going insane. Fix it by quickly scaling down the readahead size.
1070  */
1071 static void shrink_readahead_size_eio(struct file *filp,
1072                                         struct file_ra_state *ra)
1073 {
1074         ra->ra_pages /= 4;
1075 }
1076
1077 /**
1078  * do_generic_file_read - generic file read routine
1079  * @filp:       the file to read
1080  * @ppos:       current file position
1081  * @desc:       read_descriptor
1082  * @actor:      read method
1083  *
1084  * This is a generic file read routine, and uses the
1085  * mapping->a_ops->readpage() function for the actual low-level stuff.
1086  *
1087  * This is really ugly. But the goto's actually try to clarify some
1088  * of the logic when it comes to error handling etc.
1089  */
1090 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1091                 read_descriptor_t *desc, read_actor_t actor)
1092 {
1093         struct address_space *mapping = filp->f_mapping;
1094         struct inode *inode = mapping->host;
1095         struct file_ra_state *ra = &filp->f_ra;
1096         pgoff_t index;
1097         pgoff_t last_index;
1098         pgoff_t prev_index;
1099         unsigned long offset;      /* offset into pagecache page */
1100         unsigned int prev_offset;
1101         int error;
1102
1103         index = *ppos >> PAGE_CACHE_SHIFT;
1104         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1105         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1106         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1107         offset = *ppos & ~PAGE_CACHE_MASK;
1108
1109         for (;;) {
1110                 struct page *page;
1111                 pgoff_t end_index;
1112                 loff_t isize;
1113                 unsigned long nr, ret;
1114
1115                 cond_resched();
1116 find_page:
1117                 page = find_get_page(mapping, index);
1118                 if (!page) {
1119                         page_cache_sync_readahead(mapping,
1120                                         ra, filp,
1121                                         index, last_index - index);
1122                         page = find_get_page(mapping, index);
1123                         if (unlikely(page == NULL))
1124                                 goto no_cached_page;
1125                 }
1126                 if (PageReadahead(page)) {
1127                         page_cache_async_readahead(mapping,
1128                                         ra, filp, page,
1129                                         index, last_index - index);
1130                 }
1131                 if (!PageUptodate(page)) {
1132                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1133                                         !mapping->a_ops->is_partially_uptodate)
1134                                 goto page_not_up_to_date;
1135                         if (!trylock_page(page))
1136                                 goto page_not_up_to_date;
1137                         /* Did it get truncated before we got the lock? */
1138                         if (!page->mapping)
1139                                 goto page_not_up_to_date_locked;
1140                         if (!mapping->a_ops->is_partially_uptodate(page,
1141                                                                 desc, offset))
1142                                 goto page_not_up_to_date_locked;
1143                         unlock_page(page);
1144                 }
1145 page_ok:
1146                 /*
1147                  * i_size must be checked after we know the page is Uptodate.
1148                  *
1149                  * Checking i_size after the check allows us to calculate
1150                  * the correct value for "nr", which means the zero-filled
1151                  * part of the page is not copied back to userspace (unless
1152                  * another truncate extends the file - this is desired though).
1153                  */
1154
1155                 isize = i_size_read(inode);
1156                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1157                 if (unlikely(!isize || index > end_index)) {
1158                         page_cache_release(page);
1159                         goto out;
1160                 }
1161
1162                 /* nr is the maximum number of bytes to copy from this page */
1163                 nr = PAGE_CACHE_SIZE;
1164                 if (index == end_index) {
1165                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1166                         if (nr <= offset) {
1167                                 page_cache_release(page);
1168                                 goto out;
1169                         }
1170                 }
1171                 nr = nr - offset;
1172
1173                 /* If users can be writing to this page using arbitrary
1174                  * virtual addresses, take care about potential aliasing
1175                  * before reading the page on the kernel side.
1176                  */
1177                 if (mapping_writably_mapped(mapping))
1178                         flush_dcache_page(page);
1179
1180                 /*
1181                  * When a sequential read accesses a page several times,
1182                  * only mark it as accessed the first time.
1183                  */
1184                 if (prev_index != index || offset != prev_offset)
1185                         mark_page_accessed(page);
1186                 prev_index = index;
1187
1188                 /*
1189                  * Ok, we have the page, and it's up-to-date, so
1190                  * now we can copy it to user space...
1191                  *
1192                  * The actor routine returns how many bytes were actually used..
1193                  * NOTE! This may not be the same as how much of a user buffer
1194                  * we filled up (we may be padding etc), so we can only update
1195                  * "pos" here (the actor routine has to update the user buffer
1196                  * pointers and the remaining count).
1197                  */
1198                 ret = actor(desc, page, offset, nr);
1199                 offset += ret;
1200                 index += offset >> PAGE_CACHE_SHIFT;
1201                 offset &= ~PAGE_CACHE_MASK;
1202                 prev_offset = offset;
1203
1204                 page_cache_release(page);
1205                 if (ret == nr && desc->count)
1206                         continue;
1207                 goto out;
1208
1209 page_not_up_to_date:
1210                 /* Get exclusive access to the page ... */
1211                 error = lock_page_killable(page);
1212                 if (unlikely(error))
1213                         goto readpage_error;
1214
1215 page_not_up_to_date_locked:
1216                 /* Did it get truncated before we got the lock? */
1217                 if (!page->mapping) {
1218                         unlock_page(page);
1219                         page_cache_release(page);
1220                         continue;
1221                 }
1222
1223                 /* Did somebody else fill it already? */
1224                 if (PageUptodate(page)) {
1225                         unlock_page(page);
1226                         goto page_ok;
1227                 }
1228
1229 readpage:
1230                 /*
1231                  * A previous I/O error may have been due to temporary
1232                  * failures, eg. multipath errors.
1233                  * PG_error will be set again if readpage fails.
1234                  */
1235                 ClearPageError(page);
1236                 /* Start the actual read. The read will unlock the page. */
1237                 error = mapping->a_ops->readpage(filp, page);
1238
1239                 if (unlikely(error)) {
1240                         if (error == AOP_TRUNCATED_PAGE) {
1241                                 page_cache_release(page);
1242                                 goto find_page;
1243                         }
1244                         goto readpage_error;
1245                 }
1246
1247                 if (!PageUptodate(page)) {
1248                         error = lock_page_killable(page);
1249                         if (unlikely(error))
1250                                 goto readpage_error;
1251                         if (!PageUptodate(page)) {
1252                                 if (page->mapping == NULL) {
1253                                         /*
1254                                          * invalidate_mapping_pages got it
1255                                          */
1256                                         unlock_page(page);
1257                                         page_cache_release(page);
1258                                         goto find_page;
1259                                 }
1260                                 unlock_page(page);
1261                                 shrink_readahead_size_eio(filp, ra);
1262                                 error = -EIO;
1263                                 goto readpage_error;
1264                         }
1265                         unlock_page(page);
1266                 }
1267
1268                 goto page_ok;
1269
1270 readpage_error:
1271                 /* UHHUH! A synchronous read error occurred. Report it */
1272                 desc->error = error;
1273                 page_cache_release(page);
1274                 goto out;
1275
1276 no_cached_page:
1277                 /*
1278                  * Ok, it wasn't cached, so we need to create a new
1279                  * page..
1280                  */
1281                 page = page_cache_alloc_cold(mapping);
1282                 if (!page) {
1283                         desc->error = -ENOMEM;
1284                         goto out;
1285                 }
1286                 error = add_to_page_cache_lru(page, mapping,
1287                                                 index, GFP_KERNEL);
1288                 if (error) {
1289                         page_cache_release(page);
1290                         if (error == -EEXIST)
1291                                 goto find_page;
1292                         desc->error = error;
1293                         goto out;
1294                 }
1295                 goto readpage;
1296         }
1297
1298 out:
1299         ra->prev_pos = prev_index;
1300         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1301         ra->prev_pos |= prev_offset;
1302
1303         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1304         file_accessed(filp);
1305 }
1306
1307 int file_read_actor(read_descriptor_t *desc, struct page *page,
1308                         unsigned long offset, unsigned long size)
1309 {
1310         char *kaddr;
1311         unsigned long left, count = desc->count;
1312
1313         if (size > count)
1314                 size = count;
1315
1316         /*
1317          * Faults on the destination of a read are common, so do it before
1318          * taking the kmap.
1319          */
1320         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1321                 kaddr = kmap_atomic(page);
1322                 left = __copy_to_user_inatomic(desc->arg.buf,
1323                                                 kaddr + offset, size);
1324                 kunmap_atomic(kaddr);
1325                 if (left == 0)
1326                         goto success;
1327         }
1328
1329         /* Do it the slow way */
1330         kaddr = kmap(page);
1331         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1332         kunmap(page);
1333
1334         if (left) {
1335                 size -= left;
1336                 desc->error = -EFAULT;
1337         }
1338 success:
1339         desc->count = count - size;
1340         desc->written += size;
1341         desc->arg.buf += size;
1342         return size;
1343 }
1344
1345 /*
1346  * Performs necessary checks before doing a write
1347  * @iov:        io vector request
1348  * @nr_segs:    number of segments in the iovec
1349  * @count:      number of bytes to write
1350  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1351  *
1352  * Adjust number of segments and amount of bytes to write (nr_segs should be
1353  * properly initialized first). Returns appropriate error code that caller
1354  * should return or zero in case that write should be allowed.
1355  */
1356 int generic_segment_checks(const struct iovec *iov,
1357                         unsigned long *nr_segs, size_t *count, int access_flags)
1358 {
1359         unsigned long   seg;
1360         size_t cnt = 0;
1361         for (seg = 0; seg < *nr_segs; seg++) {
1362                 const struct iovec *iv = &iov[seg];
1363
1364                 /*
1365                  * If any segment has a negative length, or the cumulative
1366                  * length ever wraps negative then return -EINVAL.
1367                  */
1368                 cnt += iv->iov_len;
1369                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1370                         return -EINVAL;
1371                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1372                         continue;
1373                 if (seg == 0)
1374                         return -EFAULT;
1375                 *nr_segs = seg;
1376                 cnt -= iv->iov_len;     /* This segment is no good */
1377                 break;
1378         }
1379         *count = cnt;
1380         return 0;
1381 }
1382 EXPORT_SYMBOL(generic_segment_checks);
1383
1384 /**
1385  * generic_file_aio_read - generic filesystem read routine
1386  * @iocb:       kernel I/O control block
1387  * @iov:        io vector request
1388  * @nr_segs:    number of segments in the iovec
1389  * @pos:        current file position
1390  *
1391  * This is the "read()" routine for all filesystems
1392  * that can use the page cache directly.
1393  */
1394 ssize_t
1395 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1396                 unsigned long nr_segs, loff_t pos)
1397 {
1398         struct file *filp = iocb->ki_filp;
1399         ssize_t retval;
1400         unsigned long seg = 0;
1401         size_t count;
1402         loff_t *ppos = &iocb->ki_pos;
1403
1404         count = 0;
1405         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1406         if (retval)
1407                 return retval;
1408
1409         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1410         if (filp->f_flags & O_DIRECT) {
1411                 loff_t size;
1412                 struct address_space *mapping;
1413                 struct inode *inode;
1414
1415                 mapping = filp->f_mapping;
1416                 inode = mapping->host;
1417                 if (!count)
1418                         goto out; /* skip atime */
1419                 size = i_size_read(inode);
1420                 if (pos < size) {
1421                         retval = filemap_write_and_wait_range(mapping, pos,
1422                                         pos + iov_length(iov, nr_segs) - 1);
1423                         if (!retval) {
1424                                 struct blk_plug plug;
1425
1426                                 blk_start_plug(&plug);
1427                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1428                                                         iov, pos, nr_segs);
1429                                 blk_finish_plug(&plug);
1430                         }
1431                         if (retval > 0) {
1432                                 *ppos = pos + retval;
1433                                 count -= retval;
1434                         }
1435
1436                         /*
1437                          * Btrfs can have a short DIO read if we encounter
1438                          * compressed extents, so if there was an error, or if
1439                          * we've already read everything we wanted to, or if
1440                          * there was a short read because we hit EOF, go ahead
1441                          * and return.  Otherwise fallthrough to buffered io for
1442                          * the rest of the read.
1443                          */
1444                         if (retval < 0 || !count || *ppos >= size) {
1445                                 file_accessed(filp);
1446                                 goto out;
1447                         }
1448                 }
1449         }
1450
1451         count = retval;
1452         for (seg = 0; seg < nr_segs; seg++) {
1453                 read_descriptor_t desc;
1454                 loff_t offset = 0;
1455
1456                 /*
1457                  * If we did a short DIO read we need to skip the section of the
1458                  * iov that we've already read data into.
1459                  */
1460                 if (count) {
1461                         if (count > iov[seg].iov_len) {
1462                                 count -= iov[seg].iov_len;
1463                                 continue;
1464                         }
1465                         offset = count;
1466                         count = 0;
1467                 }
1468
1469                 desc.written = 0;
1470                 desc.arg.buf = iov[seg].iov_base + offset;
1471                 desc.count = iov[seg].iov_len - offset;
1472                 if (desc.count == 0)
1473                         continue;
1474                 desc.error = 0;
1475                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1476                 retval += desc.written;
1477                 if (desc.error) {
1478                         retval = retval ?: desc.error;
1479                         break;
1480                 }
1481                 if (desc.count > 0)
1482                         break;
1483         }
1484 out:
1485         return retval;
1486 }
1487 EXPORT_SYMBOL(generic_file_aio_read);
1488
1489 static ssize_t
1490 do_readahead(struct address_space *mapping, struct file *filp,
1491              pgoff_t index, unsigned long nr)
1492 {
1493         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1494                 return -EINVAL;
1495
1496         force_page_cache_readahead(mapping, filp, index, nr);
1497         return 0;
1498 }
1499
1500 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1501 {
1502         ssize_t ret;
1503         struct file *file;
1504
1505         ret = -EBADF;
1506         file = fget(fd);
1507         if (file) {
1508                 if (file->f_mode & FMODE_READ) {
1509                         struct address_space *mapping = file->f_mapping;
1510                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1511                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1512                         unsigned long len = end - start + 1;
1513                         ret = do_readahead(mapping, file, start, len);
1514                 }
1515                 fput(file);
1516         }
1517         return ret;
1518 }
1519 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1520 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1521 {
1522         return SYSC_readahead((int) fd, offset, (size_t) count);
1523 }
1524 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1525 #endif
1526
1527 #ifdef CONFIG_MMU
1528 /**
1529  * page_cache_read - adds requested page to the page cache if not already there
1530  * @file:       file to read
1531  * @offset:     page index
1532  *
1533  * This adds the requested page to the page cache if it isn't already there,
1534  * and schedules an I/O to read in its contents from disk.
1535  */
1536 static int page_cache_read(struct file *file, pgoff_t offset)
1537 {
1538         struct address_space *mapping = file->f_mapping;
1539         struct page *page; 
1540         int ret;
1541
1542         do {
1543                 page = page_cache_alloc_cold(mapping);
1544                 if (!page)
1545                         return -ENOMEM;
1546
1547                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1548                 if (ret == 0)
1549                         ret = mapping->a_ops->readpage(file, page);
1550                 else if (ret == -EEXIST)
1551                         ret = 0; /* losing race to add is OK */
1552
1553                 page_cache_release(page);
1554
1555         } while (ret == AOP_TRUNCATED_PAGE);
1556                 
1557         return ret;
1558 }
1559
1560 #define MMAP_LOTSAMISS  (100)
1561
1562 /*
1563  * Synchronous readahead happens when we don't even find
1564  * a page in the page cache at all.
1565  */
1566 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1567                                    struct file_ra_state *ra,
1568                                    struct file *file,
1569                                    pgoff_t offset)
1570 {
1571         unsigned long ra_pages;
1572         struct address_space *mapping = file->f_mapping;
1573
1574         /* If we don't want any read-ahead, don't bother */
1575         if (VM_RandomReadHint(vma))
1576                 return;
1577         if (!ra->ra_pages)
1578                 return;
1579
1580         if (VM_SequentialReadHint(vma)) {
1581                 page_cache_sync_readahead(mapping, ra, file, offset,
1582                                           ra->ra_pages);
1583                 return;
1584         }
1585
1586         /* Avoid banging the cache line if not needed */
1587         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1588                 ra->mmap_miss++;
1589
1590         /*
1591          * Do we miss much more than hit in this file? If so,
1592          * stop bothering with read-ahead. It will only hurt.
1593          */
1594         if (ra->mmap_miss > MMAP_LOTSAMISS)
1595                 return;
1596
1597         /*
1598          * mmap read-around
1599          */
1600         ra_pages = max_sane_readahead(ra->ra_pages);
1601         ra->start = max_t(long, 0, offset - ra_pages / 2);
1602         ra->size = ra_pages;
1603         ra->async_size = ra_pages / 4;
1604         ra_submit(ra, mapping, file);
1605 }
1606
1607 /*
1608  * Asynchronous readahead happens when we find the page and PG_readahead,
1609  * so we want to possibly extend the readahead further..
1610  */
1611 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1612                                     struct file_ra_state *ra,
1613                                     struct file *file,
1614                                     struct page *page,
1615                                     pgoff_t offset)
1616 {
1617         struct address_space *mapping = file->f_mapping;
1618
1619         /* If we don't want any read-ahead, don't bother */
1620         if (VM_RandomReadHint(vma))
1621                 return;
1622         if (ra->mmap_miss > 0)
1623                 ra->mmap_miss--;
1624         if (PageReadahead(page))
1625                 page_cache_async_readahead(mapping, ra, file,
1626                                            page, offset, ra->ra_pages);
1627 }
1628
1629 /**
1630  * filemap_fault - read in file data for page fault handling
1631  * @vma:        vma in which the fault was taken
1632  * @vmf:        struct vm_fault containing details of the fault
1633  *
1634  * filemap_fault() is invoked via the vma operations vector for a
1635  * mapped memory region to read in file data during a page fault.
1636  *
1637  * The goto's are kind of ugly, but this streamlines the normal case of having
1638  * it in the page cache, and handles the special cases reasonably without
1639  * having a lot of duplicated code.
1640  */
1641 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1642 {
1643         int error;
1644         struct file *file = vma->vm_file;
1645         struct address_space *mapping = file->f_mapping;
1646         struct file_ra_state *ra = &file->f_ra;
1647         struct inode *inode = mapping->host;
1648         pgoff_t offset = vmf->pgoff;
1649         struct page *page;
1650         pgoff_t size;
1651         int ret = 0;
1652
1653         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1654         if (offset >= size)
1655                 return VM_FAULT_SIGBUS;
1656
1657         /*
1658          * Do we have something in the page cache already?
1659          */
1660         page = find_get_page(mapping, offset);
1661         if (likely(page)) {
1662                 /*
1663                  * We found the page, so try async readahead before
1664                  * waiting for the lock.
1665                  */
1666                 do_async_mmap_readahead(vma, ra, file, page, offset);
1667         } else {
1668                 /* No page in the page cache at all */
1669                 do_sync_mmap_readahead(vma, ra, file, offset);
1670                 count_vm_event(PGMAJFAULT);
1671                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1672                 ret = VM_FAULT_MAJOR;
1673 retry_find:
1674                 page = find_get_page(mapping, offset);
1675                 if (!page)
1676                         goto no_cached_page;
1677         }
1678
1679         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1680                 page_cache_release(page);
1681                 return ret | VM_FAULT_RETRY;
1682         }
1683
1684         /* Did it get truncated? */
1685         if (unlikely(page->mapping != mapping)) {
1686                 unlock_page(page);
1687                 put_page(page);
1688                 goto retry_find;
1689         }
1690         VM_BUG_ON(page->index != offset);
1691
1692         /*
1693          * We have a locked page in the page cache, now we need to check
1694          * that it's up-to-date. If not, it is going to be due to an error.
1695          */
1696         if (unlikely(!PageUptodate(page)))
1697                 goto page_not_uptodate;
1698
1699         /*
1700          * Found the page and have a reference on it.
1701          * We must recheck i_size under page lock.
1702          */
1703         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1704         if (unlikely(offset >= size)) {
1705                 unlock_page(page);
1706                 page_cache_release(page);
1707                 return VM_FAULT_SIGBUS;
1708         }
1709
1710         vmf->page = page;
1711         return ret | VM_FAULT_LOCKED;
1712
1713 no_cached_page:
1714         /*
1715          * We're only likely to ever get here if MADV_RANDOM is in
1716          * effect.
1717          */
1718         error = page_cache_read(file, offset);
1719
1720         /*
1721          * The page we want has now been added to the page cache.
1722          * In the unlikely event that someone removed it in the
1723          * meantime, we'll just come back here and read it again.
1724          */
1725         if (error >= 0)
1726                 goto retry_find;
1727
1728         /*
1729          * An error return from page_cache_read can result if the
1730          * system is low on memory, or a problem occurs while trying
1731          * to schedule I/O.
1732          */
1733         if (error == -ENOMEM)
1734                 return VM_FAULT_OOM;
1735         return VM_FAULT_SIGBUS;
1736
1737 page_not_uptodate:
1738         /*
1739          * Umm, take care of errors if the page isn't up-to-date.
1740          * Try to re-read it _once_. We do this synchronously,
1741          * because there really aren't any performance issues here
1742          * and we need to check for errors.
1743          */
1744         ClearPageError(page);
1745         error = mapping->a_ops->readpage(file, page);
1746         if (!error) {
1747                 wait_on_page_locked(page);
1748                 if (!PageUptodate(page))
1749                         error = -EIO;
1750         }
1751         page_cache_release(page);
1752
1753         if (!error || error == AOP_TRUNCATED_PAGE)
1754                 goto retry_find;
1755
1756         /* Things didn't work out. Return zero to tell the mm layer so. */
1757         shrink_readahead_size_eio(file, ra);
1758         return VM_FAULT_SIGBUS;
1759 }
1760 EXPORT_SYMBOL(filemap_fault);
1761
1762 const struct vm_operations_struct generic_file_vm_ops = {
1763         .fault          = filemap_fault,
1764 };
1765
1766 /* This is used for a general mmap of a disk file */
1767
1768 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1769 {
1770         struct address_space *mapping = file->f_mapping;
1771
1772         if (!mapping->a_ops->readpage)
1773                 return -ENOEXEC;
1774         file_accessed(file);
1775         vma->vm_ops = &generic_file_vm_ops;
1776         vma->vm_flags |= VM_CAN_NONLINEAR;
1777         return 0;
1778 }
1779
1780 /*
1781  * This is for filesystems which do not implement ->writepage.
1782  */
1783 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1784 {
1785         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1786                 return -EINVAL;
1787         return generic_file_mmap(file, vma);
1788 }
1789 #else
1790 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1791 {
1792         return -ENOSYS;
1793 }
1794 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1795 {
1796         return -ENOSYS;
1797 }
1798 #endif /* CONFIG_MMU */
1799
1800 EXPORT_SYMBOL(generic_file_mmap);
1801 EXPORT_SYMBOL(generic_file_readonly_mmap);
1802
1803 static struct page *__read_cache_page(struct address_space *mapping,
1804                                 pgoff_t index,
1805                                 int (*filler)(void *, struct page *),
1806                                 void *data,
1807                                 gfp_t gfp)
1808 {
1809         struct page *page;
1810         int err;
1811 repeat:
1812         page = find_get_page(mapping, index);
1813         if (!page) {
1814                 page = __page_cache_alloc(gfp | __GFP_COLD);
1815                 if (!page)
1816                         return ERR_PTR(-ENOMEM);
1817                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1818                 if (unlikely(err)) {
1819                         page_cache_release(page);
1820                         if (err == -EEXIST)
1821                                 goto repeat;
1822                         /* Presumably ENOMEM for radix tree node */
1823                         return ERR_PTR(err);
1824                 }
1825                 err = filler(data, page);
1826                 if (err < 0) {
1827                         page_cache_release(page);
1828                         page = ERR_PTR(err);
1829                 }
1830         }
1831         return page;
1832 }
1833
1834 static struct page *do_read_cache_page(struct address_space *mapping,
1835                                 pgoff_t index,
1836                                 int (*filler)(void *, struct page *),
1837                                 void *data,
1838                                 gfp_t gfp)
1839
1840 {
1841         struct page *page;
1842         int err;
1843
1844 retry:
1845         page = __read_cache_page(mapping, index, filler, data, gfp);
1846         if (IS_ERR(page))
1847                 return page;
1848         if (PageUptodate(page))
1849                 goto out;
1850
1851         lock_page(page);
1852         if (!page->mapping) {
1853                 unlock_page(page);
1854                 page_cache_release(page);
1855                 goto retry;
1856         }
1857         if (PageUptodate(page)) {
1858                 unlock_page(page);
1859                 goto out;
1860         }
1861         err = filler(data, page);
1862         if (err < 0) {
1863                 page_cache_release(page);
1864                 return ERR_PTR(err);
1865         }
1866 out:
1867         mark_page_accessed(page);
1868         return page;
1869 }
1870
1871 /**
1872  * read_cache_page_async - read into page cache, fill it if needed
1873  * @mapping:    the page's address_space
1874  * @index:      the page index
1875  * @filler:     function to perform the read
1876  * @data:       first arg to filler(data, page) function, often left as NULL
1877  *
1878  * Same as read_cache_page, but don't wait for page to become unlocked
1879  * after submitting it to the filler.
1880  *
1881  * Read into the page cache. If a page already exists, and PageUptodate() is
1882  * not set, try to fill the page but don't wait for it to become unlocked.
1883  *
1884  * If the page does not get brought uptodate, return -EIO.
1885  */
1886 struct page *read_cache_page_async(struct address_space *mapping,
1887                                 pgoff_t index,
1888                                 int (*filler)(void *, struct page *),
1889                                 void *data)
1890 {
1891         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1892 }
1893 EXPORT_SYMBOL(read_cache_page_async);
1894
1895 static struct page *wait_on_page_read(struct page *page)
1896 {
1897         if (!IS_ERR(page)) {
1898                 wait_on_page_locked(page);
1899                 if (!PageUptodate(page)) {
1900                         page_cache_release(page);
1901                         page = ERR_PTR(-EIO);
1902                 }
1903         }
1904         return page;
1905 }
1906
1907 /**
1908  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1909  * @mapping:    the page's address_space
1910  * @index:      the page index
1911  * @gfp:        the page allocator flags to use if allocating
1912  *
1913  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1914  * any new page allocations done using the specified allocation flags.
1915  *
1916  * If the page does not get brought uptodate, return -EIO.
1917  */
1918 struct page *read_cache_page_gfp(struct address_space *mapping,
1919                                 pgoff_t index,
1920                                 gfp_t gfp)
1921 {
1922         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1923
1924         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1925 }
1926 EXPORT_SYMBOL(read_cache_page_gfp);
1927
1928 /**
1929  * read_cache_page - read into page cache, fill it if needed
1930  * @mapping:    the page's address_space
1931  * @index:      the page index
1932  * @filler:     function to perform the read
1933  * @data:       first arg to filler(data, page) function, often left as NULL
1934  *
1935  * Read into the page cache. If a page already exists, and PageUptodate() is
1936  * not set, try to fill the page then wait for it to become unlocked.
1937  *
1938  * If the page does not get brought uptodate, return -EIO.
1939  */
1940 struct page *read_cache_page(struct address_space *mapping,
1941                                 pgoff_t index,
1942                                 int (*filler)(void *, struct page *),
1943                                 void *data)
1944 {
1945         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1946 }
1947 EXPORT_SYMBOL(read_cache_page);
1948
1949 /*
1950  * The logic we want is
1951  *
1952  *      if suid or (sgid and xgrp)
1953  *              remove privs
1954  */
1955 int should_remove_suid(struct dentry *dentry)
1956 {
1957         umode_t mode = dentry->d_inode->i_mode;
1958         int kill = 0;
1959
1960         /* suid always must be killed */
1961         if (unlikely(mode & S_ISUID))
1962                 kill = ATTR_KILL_SUID;
1963
1964         /*
1965          * sgid without any exec bits is just a mandatory locking mark; leave
1966          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1967          */
1968         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1969                 kill |= ATTR_KILL_SGID;
1970
1971         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1972                 return kill;
1973
1974         return 0;
1975 }
1976 EXPORT_SYMBOL(should_remove_suid);
1977
1978 static int __remove_suid(struct dentry *dentry, int kill)
1979 {
1980         struct iattr newattrs;
1981
1982         newattrs.ia_valid = ATTR_FORCE | kill;
1983         return notify_change(dentry, &newattrs);
1984 }
1985
1986 int file_remove_suid(struct file *file)
1987 {
1988         struct dentry *dentry = file->f_path.dentry;
1989         struct inode *inode = dentry->d_inode;
1990         int killsuid;
1991         int killpriv;
1992         int error = 0;
1993
1994         /* Fast path for nothing security related */
1995         if (IS_NOSEC(inode))
1996                 return 0;
1997
1998         killsuid = should_remove_suid(dentry);
1999         killpriv = security_inode_need_killpriv(dentry);
2000
2001         if (killpriv < 0)
2002                 return killpriv;
2003         if (killpriv)
2004                 error = security_inode_killpriv(dentry);
2005         if (!error && killsuid)
2006                 error = __remove_suid(dentry, killsuid);
2007         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2008                 inode->i_flags |= S_NOSEC;
2009
2010         return error;
2011 }
2012 EXPORT_SYMBOL(file_remove_suid);
2013
2014 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2015                         const struct iovec *iov, size_t base, size_t bytes)
2016 {
2017         size_t copied = 0, left = 0;
2018
2019         while (bytes) {
2020                 char __user *buf = iov->iov_base + base;
2021                 int copy = min(bytes, iov->iov_len - base);
2022
2023                 base = 0;
2024                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2025                 copied += copy;
2026                 bytes -= copy;
2027                 vaddr += copy;
2028                 iov++;
2029
2030                 if (unlikely(left))
2031                         break;
2032         }
2033         return copied - left;
2034 }
2035
2036 /*
2037  * Copy as much as we can into the page and return the number of bytes which
2038  * were successfully copied.  If a fault is encountered then return the number of
2039  * bytes which were copied.
2040  */
2041 size_t iov_iter_copy_from_user_atomic(struct page *page,
2042                 struct iov_iter *i, unsigned long offset, size_t bytes)
2043 {
2044         char *kaddr;
2045         size_t copied;
2046
2047         BUG_ON(!in_atomic());
2048         kaddr = kmap_atomic(page);
2049         if (likely(i->nr_segs == 1)) {
2050                 int left;
2051                 char __user *buf = i->iov->iov_base + i->iov_offset;
2052                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2053                 copied = bytes - left;
2054         } else {
2055                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2056                                                 i->iov, i->iov_offset, bytes);
2057         }
2058         kunmap_atomic(kaddr);
2059
2060         return copied;
2061 }
2062 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2063
2064 /*
2065  * This has the same sideeffects and return value as
2066  * iov_iter_copy_from_user_atomic().
2067  * The difference is that it attempts to resolve faults.
2068  * Page must not be locked.
2069  */
2070 size_t iov_iter_copy_from_user(struct page *page,
2071                 struct iov_iter *i, unsigned long offset, size_t bytes)
2072 {
2073         char *kaddr;
2074         size_t copied;
2075
2076         kaddr = kmap(page);
2077         if (likely(i->nr_segs == 1)) {
2078                 int left;
2079                 char __user *buf = i->iov->iov_base + i->iov_offset;
2080                 left = __copy_from_user(kaddr + offset, buf, bytes);
2081                 copied = bytes - left;
2082         } else {
2083                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2084                                                 i->iov, i->iov_offset, bytes);
2085         }
2086         kunmap(page);
2087         return copied;
2088 }
2089 EXPORT_SYMBOL(iov_iter_copy_from_user);
2090
2091 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2092 {
2093         BUG_ON(i->count < bytes);
2094
2095         if (likely(i->nr_segs == 1)) {
2096                 i->iov_offset += bytes;
2097                 i->count -= bytes;
2098         } else {
2099                 const struct iovec *iov = i->iov;
2100                 size_t base = i->iov_offset;
2101                 unsigned long nr_segs = i->nr_segs;
2102
2103                 /*
2104                  * The !iov->iov_len check ensures we skip over unlikely
2105                  * zero-length segments (without overruning the iovec).
2106                  */
2107                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2108                         int copy;
2109
2110                         copy = min(bytes, iov->iov_len - base);
2111                         BUG_ON(!i->count || i->count < copy);
2112                         i->count -= copy;
2113                         bytes -= copy;
2114                         base += copy;
2115                         if (iov->iov_len == base) {
2116                                 iov++;
2117                                 nr_segs--;
2118                                 base = 0;
2119                         }
2120                 }
2121                 i->iov = iov;
2122                 i->iov_offset = base;
2123                 i->nr_segs = nr_segs;
2124         }
2125 }
2126 EXPORT_SYMBOL(iov_iter_advance);
2127
2128 /*
2129  * Fault in the first iovec of the given iov_iter, to a maximum length
2130  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2131  * accessed (ie. because it is an invalid address).
2132  *
2133  * writev-intensive code may want this to prefault several iovecs -- that
2134  * would be possible (callers must not rely on the fact that _only_ the
2135  * first iovec will be faulted with the current implementation).
2136  */
2137 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2138 {
2139         char __user *buf = i->iov->iov_base + i->iov_offset;
2140         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2141         return fault_in_pages_readable(buf, bytes);
2142 }
2143 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2144
2145 /*
2146  * Return the count of just the current iov_iter segment.
2147  */
2148 size_t iov_iter_single_seg_count(struct iov_iter *i)
2149 {
2150         const struct iovec *iov = i->iov;
2151         if (i->nr_segs == 1)
2152                 return i->count;
2153         else
2154                 return min(i->count, iov->iov_len - i->iov_offset);
2155 }
2156 EXPORT_SYMBOL(iov_iter_single_seg_count);
2157
2158 /*
2159  * Performs necessary checks before doing a write
2160  *
2161  * Can adjust writing position or amount of bytes to write.
2162  * Returns appropriate error code that caller should return or
2163  * zero in case that write should be allowed.
2164  */
2165 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2166 {
2167         struct inode *inode = file->f_mapping->host;
2168         unsigned long limit = rlimit(RLIMIT_FSIZE);
2169
2170         if (unlikely(*pos < 0))
2171                 return -EINVAL;
2172
2173         if (!isblk) {
2174                 /* FIXME: this is for backwards compatibility with 2.4 */
2175                 if (file->f_flags & O_APPEND)
2176                         *pos = i_size_read(inode);
2177
2178                 if (limit != RLIM_INFINITY) {
2179                         if (*pos >= limit) {
2180                                 send_sig(SIGXFSZ, current, 0);
2181                                 return -EFBIG;
2182                         }
2183                         if (*count > limit - (typeof(limit))*pos) {
2184                                 *count = limit - (typeof(limit))*pos;
2185                         }
2186                 }
2187         }
2188
2189         /*
2190          * LFS rule
2191          */
2192         if (unlikely(*pos + *count > MAX_NON_LFS &&
2193                                 !(file->f_flags & O_LARGEFILE))) {
2194                 if (*pos >= MAX_NON_LFS) {
2195                         return -EFBIG;
2196                 }
2197                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2198                         *count = MAX_NON_LFS - (unsigned long)*pos;
2199                 }
2200         }
2201
2202         /*
2203          * Are we about to exceed the fs block limit ?
2204          *
2205          * If we have written data it becomes a short write.  If we have
2206          * exceeded without writing data we send a signal and return EFBIG.
2207          * Linus frestrict idea will clean these up nicely..
2208          */
2209         if (likely(!isblk)) {
2210                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2211                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2212                                 return -EFBIG;
2213                         }
2214                         /* zero-length writes at ->s_maxbytes are OK */
2215                 }
2216
2217                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2218                         *count = inode->i_sb->s_maxbytes - *pos;
2219         } else {
2220 #ifdef CONFIG_BLOCK
2221                 loff_t isize;
2222                 if (bdev_read_only(I_BDEV(inode)))
2223                         return -EPERM;
2224                 isize = i_size_read(inode);
2225                 if (*pos >= isize) {
2226                         if (*count || *pos > isize)
2227                                 return -ENOSPC;
2228                 }
2229
2230                 if (*pos + *count > isize)
2231                         *count = isize - *pos;
2232 #else
2233                 return -EPERM;
2234 #endif
2235         }
2236         return 0;
2237 }
2238 EXPORT_SYMBOL(generic_write_checks);
2239
2240 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2241                                 loff_t pos, unsigned len, unsigned flags,
2242                                 struct page **pagep, void **fsdata)
2243 {
2244         const struct address_space_operations *aops = mapping->a_ops;
2245
2246         return aops->write_begin(file, mapping, pos, len, flags,
2247                                                         pagep, fsdata);
2248 }
2249 EXPORT_SYMBOL(pagecache_write_begin);
2250
2251 int pagecache_write_end(struct file *file, struct address_space *mapping,
2252                                 loff_t pos, unsigned len, unsigned copied,
2253                                 struct page *page, void *fsdata)
2254 {
2255         const struct address_space_operations *aops = mapping->a_ops;
2256
2257         mark_page_accessed(page);
2258         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2259 }
2260 EXPORT_SYMBOL(pagecache_write_end);
2261
2262 ssize_t
2263 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2264                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2265                 size_t count, size_t ocount)
2266 {
2267         struct file     *file = iocb->ki_filp;
2268         struct address_space *mapping = file->f_mapping;
2269         struct inode    *inode = mapping->host;
2270         ssize_t         written;
2271         size_t          write_len;
2272         pgoff_t         end;
2273
2274         if (count != ocount)
2275                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2276
2277         write_len = iov_length(iov, *nr_segs);
2278         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2279
2280         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2281         if (written)
2282                 goto out;
2283
2284         /*
2285          * After a write we want buffered reads to be sure to go to disk to get
2286          * the new data.  We invalidate clean cached page from the region we're
2287          * about to write.  We do this *before* the write so that we can return
2288          * without clobbering -EIOCBQUEUED from ->direct_IO().
2289          */
2290         if (mapping->nrpages) {
2291                 written = invalidate_inode_pages2_range(mapping,
2292                                         pos >> PAGE_CACHE_SHIFT, end);
2293                 /*
2294                  * If a page can not be invalidated, return 0 to fall back
2295                  * to buffered write.
2296                  */
2297                 if (written) {
2298                         if (written == -EBUSY)
2299                                 return 0;
2300                         goto out;
2301                 }
2302         }
2303
2304         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2305
2306         /*
2307          * Finally, try again to invalidate clean pages which might have been
2308          * cached by non-direct readahead, or faulted in by get_user_pages()
2309          * if the source of the write was an mmap'ed region of the file
2310          * we're writing.  Either one is a pretty crazy thing to do,
2311          * so we don't support it 100%.  If this invalidation
2312          * fails, tough, the write still worked...
2313          */
2314         if (mapping->nrpages) {
2315                 invalidate_inode_pages2_range(mapping,
2316                                               pos >> PAGE_CACHE_SHIFT, end);
2317         }
2318
2319         if (written > 0) {
2320                 pos += written;
2321                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2322                         i_size_write(inode, pos);
2323                         mark_inode_dirty(inode);
2324                 }
2325                 *ppos = pos;
2326         }
2327 out:
2328         return written;
2329 }
2330 EXPORT_SYMBOL(generic_file_direct_write);
2331
2332 /*
2333  * Find or create a page at the given pagecache position. Return the locked
2334  * page. This function is specifically for buffered writes.
2335  */
2336 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2337                                         pgoff_t index, unsigned flags)
2338 {
2339         int status;
2340         gfp_t gfp_mask;
2341         struct page *page;
2342         gfp_t gfp_notmask = 0;
2343
2344         gfp_mask = mapping_gfp_mask(mapping) | __GFP_WRITE;
2345         if (flags & AOP_FLAG_NOFS)
2346                 gfp_notmask = __GFP_FS;
2347 repeat:
2348         page = find_lock_page(mapping, index);
2349         if (page)
2350                 goto found;
2351
2352         page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2353         if (!page)
2354                 return NULL;
2355         status = add_to_page_cache_lru(page, mapping, index,
2356                                                 GFP_KERNEL & ~gfp_notmask);
2357         if (unlikely(status)) {
2358                 page_cache_release(page);
2359                 if (status == -EEXIST)
2360                         goto repeat;
2361                 return NULL;
2362         }
2363 found:
2364         wait_on_page_writeback(page);
2365         return page;
2366 }
2367 EXPORT_SYMBOL(grab_cache_page_write_begin);
2368
2369 static ssize_t generic_perform_write(struct file *file,
2370                                 struct iov_iter *i, loff_t pos)
2371 {
2372         struct address_space *mapping = file->f_mapping;
2373         const struct address_space_operations *a_ops = mapping->a_ops;
2374         long status = 0;
2375         ssize_t written = 0;
2376         unsigned int flags = 0;
2377
2378         /*
2379          * Copies from kernel address space cannot fail (NFSD is a big user).
2380          */
2381         if (segment_eq(get_fs(), KERNEL_DS))
2382                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2383
2384         do {
2385                 struct page *page;
2386                 unsigned long offset;   /* Offset into pagecache page */
2387                 unsigned long bytes;    /* Bytes to write to page */
2388                 size_t copied;          /* Bytes copied from user */
2389                 void *fsdata;
2390
2391                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2392                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2393                                                 iov_iter_count(i));
2394
2395 again:
2396                 /*
2397                  * Bring in the user page that we will copy from _first_.
2398                  * Otherwise there's a nasty deadlock on copying from the
2399                  * same page as we're writing to, without it being marked
2400                  * up-to-date.
2401                  *
2402                  * Not only is this an optimisation, but it is also required
2403                  * to check that the address is actually valid, when atomic
2404                  * usercopies are used, below.
2405                  */
2406                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2407                         status = -EFAULT;
2408                         break;
2409                 }
2410
2411                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2412                                                 &page, &fsdata);
2413                 if (unlikely(status))
2414                         break;
2415
2416                 if (mapping_writably_mapped(mapping))
2417                         flush_dcache_page(page);
2418
2419                 pagefault_disable();
2420                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2421                 pagefault_enable();
2422                 flush_dcache_page(page);
2423
2424                 mark_page_accessed(page);
2425                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2426                                                 page, fsdata);
2427                 if (unlikely(status < 0))
2428                         break;
2429                 copied = status;
2430
2431                 cond_resched();
2432
2433                 iov_iter_advance(i, copied);
2434                 if (unlikely(copied == 0)) {
2435                         /*
2436                          * If we were unable to copy any data at all, we must
2437                          * fall back to a single segment length write.
2438                          *
2439                          * If we didn't fallback here, we could livelock
2440                          * because not all segments in the iov can be copied at
2441                          * once without a pagefault.
2442                          */
2443                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2444                                                 iov_iter_single_seg_count(i));
2445                         goto again;
2446                 }
2447                 pos += copied;
2448                 written += copied;
2449
2450                 balance_dirty_pages_ratelimited(mapping);
2451                 if (fatal_signal_pending(current)) {
2452                         status = -EINTR;
2453                         break;
2454                 }
2455         } while (iov_iter_count(i));
2456
2457         return written ? written : status;
2458 }
2459
2460 ssize_t
2461 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2462                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2463                 size_t count, ssize_t written)
2464 {
2465         struct file *file = iocb->ki_filp;
2466         ssize_t status;
2467         struct iov_iter i;
2468
2469         iov_iter_init(&i, iov, nr_segs, count, written);
2470         status = generic_perform_write(file, &i, pos);
2471
2472         if (likely(status >= 0)) {
2473                 written += status;
2474                 *ppos = pos + status;
2475         }
2476         
2477         return written ? written : status;
2478 }
2479 EXPORT_SYMBOL(generic_file_buffered_write);
2480
2481 /**
2482  * __generic_file_aio_write - write data to a file
2483  * @iocb:       IO state structure (file, offset, etc.)
2484  * @iov:        vector with data to write
2485  * @nr_segs:    number of segments in the vector
2486  * @ppos:       position where to write
2487  *
2488  * This function does all the work needed for actually writing data to a
2489  * file. It does all basic checks, removes SUID from the file, updates
2490  * modification times and calls proper subroutines depending on whether we
2491  * do direct IO or a standard buffered write.
2492  *
2493  * It expects i_mutex to be grabbed unless we work on a block device or similar
2494  * object which does not need locking at all.
2495  *
2496  * This function does *not* take care of syncing data in case of O_SYNC write.
2497  * A caller has to handle it. This is mainly due to the fact that we want to
2498  * avoid syncing under i_mutex.
2499  */
2500 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2501                                  unsigned long nr_segs, loff_t *ppos)
2502 {
2503         struct file *file = iocb->ki_filp;
2504         struct address_space * mapping = file->f_mapping;
2505         size_t ocount;          /* original count */
2506         size_t count;           /* after file limit checks */
2507         struct inode    *inode = mapping->host;
2508         loff_t          pos;
2509         ssize_t         written;
2510         ssize_t         err;
2511
2512         ocount = 0;
2513         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2514         if (err)
2515                 return err;
2516
2517         count = ocount;
2518         pos = *ppos;
2519
2520         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2521
2522         /* We can write back this queue in page reclaim */
2523         current->backing_dev_info = mapping->backing_dev_info;
2524         written = 0;
2525
2526         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2527         if (err)
2528                 goto out;
2529
2530         if (count == 0)
2531                 goto out;
2532
2533         err = file_remove_suid(file);
2534         if (err)
2535                 goto out;
2536
2537         file_update_time(file);
2538
2539         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2540         if (unlikely(file->f_flags & O_DIRECT)) {
2541                 loff_t endbyte;
2542                 ssize_t written_buffered;
2543
2544                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2545                                                         ppos, count, ocount);
2546                 if (written < 0 || written == count)
2547                         goto out;
2548                 /*
2549                  * direct-io write to a hole: fall through to buffered I/O
2550                  * for completing the rest of the request.
2551                  */
2552                 pos += written;
2553                 count -= written;
2554                 written_buffered = generic_file_buffered_write(iocb, iov,
2555                                                 nr_segs, pos, ppos, count,
2556                                                 written);
2557                 /*
2558                  * If generic_file_buffered_write() retuned a synchronous error
2559                  * then we want to return the number of bytes which were
2560                  * direct-written, or the error code if that was zero.  Note
2561                  * that this differs from normal direct-io semantics, which
2562                  * will return -EFOO even if some bytes were written.
2563                  */
2564                 if (written_buffered < 0) {
2565                         err = written_buffered;
2566                         goto out;
2567                 }
2568
2569                 /*
2570                  * We need to ensure that the page cache pages are written to
2571                  * disk and invalidated to preserve the expected O_DIRECT
2572                  * semantics.
2573                  */
2574                 endbyte = pos + written_buffered - written - 1;
2575                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2576                 if (err == 0) {
2577                         written = written_buffered;
2578                         invalidate_mapping_pages(mapping,
2579                                                  pos >> PAGE_CACHE_SHIFT,
2580                                                  endbyte >> PAGE_CACHE_SHIFT);
2581                 } else {
2582                         /*
2583                          * We don't know how much we wrote, so just return
2584                          * the number of bytes which were direct-written
2585                          */
2586                 }
2587         } else {
2588                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2589                                 pos, ppos, count, written);
2590         }
2591 out:
2592         current->backing_dev_info = NULL;
2593         return written ? written : err;
2594 }
2595 EXPORT_SYMBOL(__generic_file_aio_write);
2596
2597 /**
2598  * generic_file_aio_write - write data to a file
2599  * @iocb:       IO state structure
2600  * @iov:        vector with data to write
2601  * @nr_segs:    number of segments in the vector
2602  * @pos:        position in file where to write
2603  *
2604  * This is a wrapper around __generic_file_aio_write() to be used by most
2605  * filesystems. It takes care of syncing the file in case of O_SYNC file
2606  * and acquires i_mutex as needed.
2607  */
2608 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2609                 unsigned long nr_segs, loff_t pos)
2610 {
2611         struct file *file = iocb->ki_filp;
2612         struct inode *inode = file->f_mapping->host;
2613         struct blk_plug plug;
2614         ssize_t ret;
2615
2616         BUG_ON(iocb->ki_pos != pos);
2617
2618         mutex_lock(&inode->i_mutex);
2619         blk_start_plug(&plug);
2620         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2621         mutex_unlock(&inode->i_mutex);
2622
2623         if (ret > 0 || ret == -EIOCBQUEUED) {
2624                 ssize_t err;
2625
2626                 err = generic_write_sync(file, pos, ret);
2627                 if (err < 0 && ret > 0)
2628                         ret = err;
2629         }
2630         blk_finish_plug(&plug);
2631         return ret;
2632 }
2633 EXPORT_SYMBOL(generic_file_aio_write);
2634
2635 /**
2636  * try_to_release_page() - release old fs-specific metadata on a page
2637  *
2638  * @page: the page which the kernel is trying to free
2639  * @gfp_mask: memory allocation flags (and I/O mode)
2640  *
2641  * The address_space is to try to release any data against the page
2642  * (presumably at page->private).  If the release was successful, return `1'.
2643  * Otherwise return zero.
2644  *
2645  * This may also be called if PG_fscache is set on a page, indicating that the
2646  * page is known to the local caching routines.
2647  *
2648  * The @gfp_mask argument specifies whether I/O may be performed to release
2649  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2650  *
2651  */
2652 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2653 {
2654         struct address_space * const mapping = page->mapping;
2655
2656         BUG_ON(!PageLocked(page));
2657         if (PageWriteback(page))
2658                 return 0;
2659
2660         if (mapping && mapping->a_ops->releasepage)
2661                 return mapping->a_ops->releasepage(page, gfp_mask);
2662         return try_to_free_buffers(page);
2663 }
2664
2665 EXPORT_SYMBOL(try_to_release_page);