ARM: OMAP3: PM: remove access to PRM_VOLTCTRL register
[pandora-kernel.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                 unsigned int cpuset_mems_cookie;
504                 do {
505                         cpuset_mems_cookie = get_mems_allowed();
506                         n = cpuset_mem_spread_node();
507                         page = alloc_pages_exact_node(n, gfp, 0);
508                 } while (!put_mems_allowed(cpuset_mems_cookie) && !page);
509
510                 return page;
511         }
512         return alloc_pages(gfp, 0);
513 }
514 EXPORT_SYMBOL(__page_cache_alloc);
515 #endif
516
517 /*
518  * In order to wait for pages to become available there must be
519  * waitqueues associated with pages. By using a hash table of
520  * waitqueues where the bucket discipline is to maintain all
521  * waiters on the same queue and wake all when any of the pages
522  * become available, and for the woken contexts to check to be
523  * sure the appropriate page became available, this saves space
524  * at a cost of "thundering herd" phenomena during rare hash
525  * collisions.
526  */
527 static wait_queue_head_t *page_waitqueue(struct page *page)
528 {
529         const struct zone *zone = page_zone(page);
530
531         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
532 }
533
534 static inline void wake_up_page(struct page *page, int bit)
535 {
536         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
537 }
538
539 void wait_on_page_bit(struct page *page, int bit_nr)
540 {
541         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
542
543         if (test_bit(bit_nr, &page->flags))
544                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
545                                                         TASK_UNINTERRUPTIBLE);
546 }
547 EXPORT_SYMBOL(wait_on_page_bit);
548
549 int wait_on_page_bit_killable(struct page *page, int bit_nr)
550 {
551         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
552
553         if (!test_bit(bit_nr, &page->flags))
554                 return 0;
555
556         return __wait_on_bit(page_waitqueue(page), &wait,
557                              sleep_on_page_killable, TASK_KILLABLE);
558 }
559
560 /**
561  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
562  * @page: Page defining the wait queue of interest
563  * @waiter: Waiter to add to the queue
564  *
565  * Add an arbitrary @waiter to the wait queue for the nominated @page.
566  */
567 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
568 {
569         wait_queue_head_t *q = page_waitqueue(page);
570         unsigned long flags;
571
572         spin_lock_irqsave(&q->lock, flags);
573         __add_wait_queue(q, waiter);
574         spin_unlock_irqrestore(&q->lock, flags);
575 }
576 EXPORT_SYMBOL_GPL(add_page_wait_queue);
577
578 /**
579  * unlock_page - unlock a locked page
580  * @page: the page
581  *
582  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
583  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
584  * mechananism between PageLocked pages and PageWriteback pages is shared.
585  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
586  *
587  * The mb is necessary to enforce ordering between the clear_bit and the read
588  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
589  */
590 void unlock_page(struct page *page)
591 {
592         VM_BUG_ON(!PageLocked(page));
593         clear_bit_unlock(PG_locked, &page->flags);
594         smp_mb__after_clear_bit();
595         wake_up_page(page, PG_locked);
596 }
597 EXPORT_SYMBOL(unlock_page);
598
599 /**
600  * end_page_writeback - end writeback against a page
601  * @page: the page
602  */
603 void end_page_writeback(struct page *page)
604 {
605         if (TestClearPageReclaim(page))
606                 rotate_reclaimable_page(page);
607
608         if (!test_clear_page_writeback(page))
609                 BUG();
610
611         smp_mb__after_clear_bit();
612         wake_up_page(page, PG_writeback);
613 }
614 EXPORT_SYMBOL(end_page_writeback);
615
616 /**
617  * __lock_page - get a lock on the page, assuming we need to sleep to get it
618  * @page: the page to lock
619  */
620 void __lock_page(struct page *page)
621 {
622         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
623
624         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
625                                                         TASK_UNINTERRUPTIBLE);
626 }
627 EXPORT_SYMBOL(__lock_page);
628
629 int __lock_page_killable(struct page *page)
630 {
631         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
632
633         return __wait_on_bit_lock(page_waitqueue(page), &wait,
634                                         sleep_on_page_killable, TASK_KILLABLE);
635 }
636 EXPORT_SYMBOL_GPL(__lock_page_killable);
637
638 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
639                          unsigned int flags)
640 {
641         if (flags & FAULT_FLAG_ALLOW_RETRY) {
642                 /*
643                  * CAUTION! In this case, mmap_sem is not released
644                  * even though return 0.
645                  */
646                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
647                         return 0;
648
649                 up_read(&mm->mmap_sem);
650                 if (flags & FAULT_FLAG_KILLABLE)
651                         wait_on_page_locked_killable(page);
652                 else
653                         wait_on_page_locked(page);
654                 return 0;
655         } else {
656                 if (flags & FAULT_FLAG_KILLABLE) {
657                         int ret;
658
659                         ret = __lock_page_killable(page);
660                         if (ret) {
661                                 up_read(&mm->mmap_sem);
662                                 return 0;
663                         }
664                 } else
665                         __lock_page(page);
666                 return 1;
667         }
668 }
669
670 /**
671  * find_get_page - find and get a page reference
672  * @mapping: the address_space to search
673  * @offset: the page index
674  *
675  * Is there a pagecache struct page at the given (mapping, offset) tuple?
676  * If yes, increment its refcount and return it; if no, return NULL.
677  */
678 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
679 {
680         void **pagep;
681         struct page *page;
682
683         rcu_read_lock();
684 repeat:
685         page = NULL;
686         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
687         if (pagep) {
688                 page = radix_tree_deref_slot(pagep);
689                 if (unlikely(!page))
690                         goto out;
691                 if (radix_tree_exception(page)) {
692                         if (radix_tree_deref_retry(page))
693                                 goto repeat;
694                         /*
695                          * Otherwise, shmem/tmpfs must be storing a swap entry
696                          * here as an exceptional entry: so return it without
697                          * attempting to raise page count.
698                          */
699                         goto out;
700                 }
701                 if (!page_cache_get_speculative(page))
702                         goto repeat;
703
704                 /*
705                  * Has the page moved?
706                  * This is part of the lockless pagecache protocol. See
707                  * include/linux/pagemap.h for details.
708                  */
709                 if (unlikely(page != *pagep)) {
710                         page_cache_release(page);
711                         goto repeat;
712                 }
713         }
714 out:
715         rcu_read_unlock();
716
717         return page;
718 }
719 EXPORT_SYMBOL(find_get_page);
720
721 /**
722  * find_lock_page - locate, pin and lock a pagecache page
723  * @mapping: the address_space to search
724  * @offset: the page index
725  *
726  * Locates the desired pagecache page, locks it, increments its reference
727  * count and returns its address.
728  *
729  * Returns zero if the page was not present. find_lock_page() may sleep.
730  */
731 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
732 {
733         struct page *page;
734
735 repeat:
736         page = find_get_page(mapping, offset);
737         if (page && !radix_tree_exception(page)) {
738                 lock_page(page);
739                 /* Has the page been truncated? */
740                 if (unlikely(page->mapping != mapping)) {
741                         unlock_page(page);
742                         page_cache_release(page);
743                         goto repeat;
744                 }
745                 VM_BUG_ON(page->index != offset);
746         }
747         return page;
748 }
749 EXPORT_SYMBOL(find_lock_page);
750
751 /**
752  * find_or_create_page - locate or add a pagecache page
753  * @mapping: the page's address_space
754  * @index: the page's index into the mapping
755  * @gfp_mask: page allocation mode
756  *
757  * Locates a page in the pagecache.  If the page is not present, a new page
758  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
759  * LRU list.  The returned page is locked and has its reference count
760  * incremented.
761  *
762  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
763  * allocation!
764  *
765  * find_or_create_page() returns the desired page's address, or zero on
766  * memory exhaustion.
767  */
768 struct page *find_or_create_page(struct address_space *mapping,
769                 pgoff_t index, gfp_t gfp_mask)
770 {
771         struct page *page;
772         int err;
773 repeat:
774         page = find_lock_page(mapping, index);
775         if (!page) {
776                 page = __page_cache_alloc(gfp_mask);
777                 if (!page)
778                         return NULL;
779                 /*
780                  * We want a regular kernel memory (not highmem or DMA etc)
781                  * allocation for the radix tree nodes, but we need to honour
782                  * the context-specific requirements the caller has asked for.
783                  * GFP_RECLAIM_MASK collects those requirements.
784                  */
785                 err = add_to_page_cache_lru(page, mapping, index,
786                         (gfp_mask & GFP_RECLAIM_MASK));
787                 if (unlikely(err)) {
788                         page_cache_release(page);
789                         page = NULL;
790                         if (err == -EEXIST)
791                                 goto repeat;
792                 }
793         }
794         return page;
795 }
796 EXPORT_SYMBOL(find_or_create_page);
797
798 /**
799  * find_get_pages - gang pagecache lookup
800  * @mapping:    The address_space to search
801  * @start:      The starting page index
802  * @nr_pages:   The maximum number of pages
803  * @pages:      Where the resulting pages are placed
804  *
805  * find_get_pages() will search for and return a group of up to
806  * @nr_pages pages in the mapping.  The pages are placed at @pages.
807  * find_get_pages() takes a reference against the returned pages.
808  *
809  * The search returns a group of mapping-contiguous pages with ascending
810  * indexes.  There may be holes in the indices due to not-present pages.
811  *
812  * find_get_pages() returns the number of pages which were found.
813  */
814 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
815                             unsigned int nr_pages, struct page **pages)
816 {
817         unsigned int i;
818         unsigned int ret;
819         unsigned int nr_found, nr_skip;
820
821         rcu_read_lock();
822 restart:
823         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
824                                 (void ***)pages, NULL, start, nr_pages);
825         ret = 0;
826         nr_skip = 0;
827         for (i = 0; i < nr_found; i++) {
828                 struct page *page;
829 repeat:
830                 page = radix_tree_deref_slot((void **)pages[i]);
831                 if (unlikely(!page))
832                         continue;
833
834                 if (radix_tree_exception(page)) {
835                         if (radix_tree_deref_retry(page)) {
836                                 /*
837                                  * Transient condition which can only trigger
838                                  * when entry at index 0 moves out of or back
839                                  * to root: none yet gotten, safe to restart.
840                                  */
841                                 WARN_ON(start | i);
842                                 goto restart;
843                         }
844                         /*
845                          * Otherwise, shmem/tmpfs must be storing a swap entry
846                          * here as an exceptional entry: so skip over it -
847                          * we only reach this from invalidate_mapping_pages().
848                          */
849                         nr_skip++;
850                         continue;
851                 }
852
853                 if (!page_cache_get_speculative(page))
854                         goto repeat;
855
856                 /* Has the page moved? */
857                 if (unlikely(page != *((void **)pages[i]))) {
858                         page_cache_release(page);
859                         goto repeat;
860                 }
861
862                 pages[ret] = page;
863                 ret++;
864         }
865
866         /*
867          * If all entries were removed before we could secure them,
868          * try again, because callers stop trying once 0 is returned.
869          */
870         if (unlikely(!ret && nr_found > nr_skip))
871                 goto restart;
872         rcu_read_unlock();
873         return ret;
874 }
875
876 /**
877  * find_get_pages_contig - gang contiguous pagecache lookup
878  * @mapping:    The address_space to search
879  * @index:      The starting page index
880  * @nr_pages:   The maximum number of pages
881  * @pages:      Where the resulting pages are placed
882  *
883  * find_get_pages_contig() works exactly like find_get_pages(), except
884  * that the returned number of pages are guaranteed to be contiguous.
885  *
886  * find_get_pages_contig() returns the number of pages which were found.
887  */
888 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
889                                unsigned int nr_pages, struct page **pages)
890 {
891         unsigned int i;
892         unsigned int ret;
893         unsigned int nr_found;
894
895         rcu_read_lock();
896 restart:
897         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
898                                 (void ***)pages, NULL, index, nr_pages);
899         ret = 0;
900         for (i = 0; i < nr_found; i++) {
901                 struct page *page;
902 repeat:
903                 page = radix_tree_deref_slot((void **)pages[i]);
904                 if (unlikely(!page))
905                         continue;
906
907                 if (radix_tree_exception(page)) {
908                         if (radix_tree_deref_retry(page)) {
909                                 /*
910                                  * Transient condition which can only trigger
911                                  * when entry at index 0 moves out of or back
912                                  * to root: none yet gotten, safe to restart.
913                                  */
914                                 goto restart;
915                         }
916                         /*
917                          * Otherwise, shmem/tmpfs must be storing a swap entry
918                          * here as an exceptional entry: so stop looking for
919                          * contiguous pages.
920                          */
921                         break;
922                 }
923
924                 if (!page_cache_get_speculative(page))
925                         goto repeat;
926
927                 /* Has the page moved? */
928                 if (unlikely(page != *((void **)pages[i]))) {
929                         page_cache_release(page);
930                         goto repeat;
931                 }
932
933                 /*
934                  * must check mapping and index after taking the ref.
935                  * otherwise we can get both false positives and false
936                  * negatives, which is just confusing to the caller.
937                  */
938                 if (page->mapping == NULL || page->index != index) {
939                         page_cache_release(page);
940                         break;
941                 }
942
943                 pages[ret] = page;
944                 ret++;
945                 index++;
946         }
947         rcu_read_unlock();
948         return ret;
949 }
950 EXPORT_SYMBOL(find_get_pages_contig);
951
952 /**
953  * find_get_pages_tag - find and return pages that match @tag
954  * @mapping:    the address_space to search
955  * @index:      the starting page index
956  * @tag:        the tag index
957  * @nr_pages:   the maximum number of pages
958  * @pages:      where the resulting pages are placed
959  *
960  * Like find_get_pages, except we only return pages which are tagged with
961  * @tag.   We update @index to index the next page for the traversal.
962  */
963 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
964                         int tag, unsigned int nr_pages, struct page **pages)
965 {
966         unsigned int i;
967         unsigned int ret;
968         unsigned int nr_found;
969
970         rcu_read_lock();
971 restart:
972         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
973                                 (void ***)pages, *index, nr_pages, tag);
974         ret = 0;
975         for (i = 0; i < nr_found; i++) {
976                 struct page *page;
977 repeat:
978                 page = radix_tree_deref_slot((void **)pages[i]);
979                 if (unlikely(!page))
980                         continue;
981
982                 if (radix_tree_exception(page)) {
983                         if (radix_tree_deref_retry(page)) {
984                                 /*
985                                  * Transient condition which can only trigger
986                                  * when entry at index 0 moves out of or back
987                                  * to root: none yet gotten, safe to restart.
988                                  */
989                                 goto restart;
990                         }
991                         /*
992                          * This function is never used on a shmem/tmpfs
993                          * mapping, so a swap entry won't be found here.
994                          */
995                         BUG();
996                 }
997
998                 if (!page_cache_get_speculative(page))
999                         goto repeat;
1000
1001                 /* Has the page moved? */
1002                 if (unlikely(page != *((void **)pages[i]))) {
1003                         page_cache_release(page);
1004                         goto repeat;
1005                 }
1006
1007                 pages[ret] = page;
1008                 ret++;
1009         }
1010
1011         /*
1012          * If all entries were removed before we could secure them,
1013          * try again, because callers stop trying once 0 is returned.
1014          */
1015         if (unlikely(!ret && nr_found))
1016                 goto restart;
1017         rcu_read_unlock();
1018
1019         if (ret)
1020                 *index = pages[ret - 1]->index + 1;
1021
1022         return ret;
1023 }
1024 EXPORT_SYMBOL(find_get_pages_tag);
1025
1026 /**
1027  * grab_cache_page_nowait - returns locked page at given index in given cache
1028  * @mapping: target address_space
1029  * @index: the page index
1030  *
1031  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1032  * This is intended for speculative data generators, where the data can
1033  * be regenerated if the page couldn't be grabbed.  This routine should
1034  * be safe to call while holding the lock for another page.
1035  *
1036  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1037  * and deadlock against the caller's locked page.
1038  */
1039 struct page *
1040 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1041 {
1042         struct page *page = find_get_page(mapping, index);
1043
1044         if (page) {
1045                 if (trylock_page(page))
1046                         return page;
1047                 page_cache_release(page);
1048                 return NULL;
1049         }
1050         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1051         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1052                 page_cache_release(page);
1053                 page = NULL;
1054         }
1055         return page;
1056 }
1057 EXPORT_SYMBOL(grab_cache_page_nowait);
1058
1059 /*
1060  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1061  * a _large_ part of the i/o request. Imagine the worst scenario:
1062  *
1063  *      ---R__________________________________________B__________
1064  *         ^ reading here                             ^ bad block(assume 4k)
1065  *
1066  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1067  * => failing the whole request => read(R) => read(R+1) =>
1068  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1069  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1070  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1071  *
1072  * It is going insane. Fix it by quickly scaling down the readahead size.
1073  */
1074 static void shrink_readahead_size_eio(struct file *filp,
1075                                         struct file_ra_state *ra)
1076 {
1077         ra->ra_pages /= 4;
1078 }
1079
1080 /**
1081  * do_generic_file_read - generic file read routine
1082  * @filp:       the file to read
1083  * @ppos:       current file position
1084  * @desc:       read_descriptor
1085  * @actor:      read method
1086  *
1087  * This is a generic file read routine, and uses the
1088  * mapping->a_ops->readpage() function for the actual low-level stuff.
1089  *
1090  * This is really ugly. But the goto's actually try to clarify some
1091  * of the logic when it comes to error handling etc.
1092  */
1093 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1094                 read_descriptor_t *desc, read_actor_t actor)
1095 {
1096         struct address_space *mapping = filp->f_mapping;
1097         struct inode *inode = mapping->host;
1098         struct file_ra_state *ra = &filp->f_ra;
1099         pgoff_t index;
1100         pgoff_t last_index;
1101         pgoff_t prev_index;
1102         unsigned long offset;      /* offset into pagecache page */
1103         unsigned int prev_offset;
1104         int error;
1105
1106         index = *ppos >> PAGE_CACHE_SHIFT;
1107         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1108         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1109         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1110         offset = *ppos & ~PAGE_CACHE_MASK;
1111
1112         for (;;) {
1113                 struct page *page;
1114                 pgoff_t end_index;
1115                 loff_t isize;
1116                 unsigned long nr, ret;
1117
1118                 cond_resched();
1119 find_page:
1120                 page = find_get_page(mapping, index);
1121                 if (!page) {
1122                         page_cache_sync_readahead(mapping,
1123                                         ra, filp,
1124                                         index, last_index - index);
1125                         page = find_get_page(mapping, index);
1126                         if (unlikely(page == NULL))
1127                                 goto no_cached_page;
1128                 }
1129                 if (PageReadahead(page)) {
1130                         page_cache_async_readahead(mapping,
1131                                         ra, filp, page,
1132                                         index, last_index - index);
1133                 }
1134                 if (!PageUptodate(page)) {
1135                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1136                                         !mapping->a_ops->is_partially_uptodate)
1137                                 goto page_not_up_to_date;
1138                         if (!trylock_page(page))
1139                                 goto page_not_up_to_date;
1140                         /* Did it get truncated before we got the lock? */
1141                         if (!page->mapping)
1142                                 goto page_not_up_to_date_locked;
1143                         if (!mapping->a_ops->is_partially_uptodate(page,
1144                                                                 desc, offset))
1145                                 goto page_not_up_to_date_locked;
1146                         unlock_page(page);
1147                 }
1148 page_ok:
1149                 /*
1150                  * i_size must be checked after we know the page is Uptodate.
1151                  *
1152                  * Checking i_size after the check allows us to calculate
1153                  * the correct value for "nr", which means the zero-filled
1154                  * part of the page is not copied back to userspace (unless
1155                  * another truncate extends the file - this is desired though).
1156                  */
1157
1158                 isize = i_size_read(inode);
1159                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1160                 if (unlikely(!isize || index > end_index)) {
1161                         page_cache_release(page);
1162                         goto out;
1163                 }
1164
1165                 /* nr is the maximum number of bytes to copy from this page */
1166                 nr = PAGE_CACHE_SIZE;
1167                 if (index == end_index) {
1168                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1169                         if (nr <= offset) {
1170                                 page_cache_release(page);
1171                                 goto out;
1172                         }
1173                 }
1174                 nr = nr - offset;
1175
1176                 /* If users can be writing to this page using arbitrary
1177                  * virtual addresses, take care about potential aliasing
1178                  * before reading the page on the kernel side.
1179                  */
1180                 if (mapping_writably_mapped(mapping))
1181                         flush_dcache_page(page);
1182
1183                 /*
1184                  * When a sequential read accesses a page several times,
1185                  * only mark it as accessed the first time.
1186                  */
1187                 if (prev_index != index || offset != prev_offset)
1188                         mark_page_accessed(page);
1189                 prev_index = index;
1190
1191                 /*
1192                  * Ok, we have the page, and it's up-to-date, so
1193                  * now we can copy it to user space...
1194                  *
1195                  * The actor routine returns how many bytes were actually used..
1196                  * NOTE! This may not be the same as how much of a user buffer
1197                  * we filled up (we may be padding etc), so we can only update
1198                  * "pos" here (the actor routine has to update the user buffer
1199                  * pointers and the remaining count).
1200                  */
1201                 ret = actor(desc, page, offset, nr);
1202                 offset += ret;
1203                 index += offset >> PAGE_CACHE_SHIFT;
1204                 offset &= ~PAGE_CACHE_MASK;
1205                 prev_offset = offset;
1206
1207                 page_cache_release(page);
1208                 if (ret == nr && desc->count)
1209                         continue;
1210                 goto out;
1211
1212 page_not_up_to_date:
1213                 /* Get exclusive access to the page ... */
1214                 error = lock_page_killable(page);
1215                 if (unlikely(error))
1216                         goto readpage_error;
1217
1218 page_not_up_to_date_locked:
1219                 /* Did it get truncated before we got the lock? */
1220                 if (!page->mapping) {
1221                         unlock_page(page);
1222                         page_cache_release(page);
1223                         continue;
1224                 }
1225
1226                 /* Did somebody else fill it already? */
1227                 if (PageUptodate(page)) {
1228                         unlock_page(page);
1229                         goto page_ok;
1230                 }
1231
1232 readpage:
1233                 /*
1234                  * A previous I/O error may have been due to temporary
1235                  * failures, eg. multipath errors.
1236                  * PG_error will be set again if readpage fails.
1237                  */
1238                 ClearPageError(page);
1239                 /* Start the actual read. The read will unlock the page. */
1240                 error = mapping->a_ops->readpage(filp, page);
1241
1242                 if (unlikely(error)) {
1243                         if (error == AOP_TRUNCATED_PAGE) {
1244                                 page_cache_release(page);
1245                                 goto find_page;
1246                         }
1247                         goto readpage_error;
1248                 }
1249
1250                 if (!PageUptodate(page)) {
1251                         error = lock_page_killable(page);
1252                         if (unlikely(error))
1253                                 goto readpage_error;
1254                         if (!PageUptodate(page)) {
1255                                 if (page->mapping == NULL) {
1256                                         /*
1257                                          * invalidate_mapping_pages got it
1258                                          */
1259                                         unlock_page(page);
1260                                         page_cache_release(page);
1261                                         goto find_page;
1262                                 }
1263                                 unlock_page(page);
1264                                 shrink_readahead_size_eio(filp, ra);
1265                                 error = -EIO;
1266                                 goto readpage_error;
1267                         }
1268                         unlock_page(page);
1269                 }
1270
1271                 goto page_ok;
1272
1273 readpage_error:
1274                 /* UHHUH! A synchronous read error occurred. Report it */
1275                 desc->error = error;
1276                 page_cache_release(page);
1277                 goto out;
1278
1279 no_cached_page:
1280                 /*
1281                  * Ok, it wasn't cached, so we need to create a new
1282                  * page..
1283                  */
1284                 page = page_cache_alloc_cold(mapping);
1285                 if (!page) {
1286                         desc->error = -ENOMEM;
1287                         goto out;
1288                 }
1289                 error = add_to_page_cache_lru(page, mapping,
1290                                                 index, GFP_KERNEL);
1291                 if (error) {
1292                         page_cache_release(page);
1293                         if (error == -EEXIST)
1294                                 goto find_page;
1295                         desc->error = error;
1296                         goto out;
1297                 }
1298                 goto readpage;
1299         }
1300
1301 out:
1302         ra->prev_pos = prev_index;
1303         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1304         ra->prev_pos |= prev_offset;
1305
1306         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1307         file_accessed(filp);
1308 }
1309
1310 int file_read_actor(read_descriptor_t *desc, struct page *page,
1311                         unsigned long offset, unsigned long size)
1312 {
1313         char *kaddr;
1314         unsigned long left, count = desc->count;
1315
1316         if (size > count)
1317                 size = count;
1318
1319         /*
1320          * Faults on the destination of a read are common, so do it before
1321          * taking the kmap.
1322          */
1323         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1324                 kaddr = kmap_atomic(page, KM_USER0);
1325                 left = __copy_to_user_inatomic(desc->arg.buf,
1326                                                 kaddr + offset, size);
1327                 kunmap_atomic(kaddr, KM_USER0);
1328                 if (left == 0)
1329                         goto success;
1330         }
1331
1332         /* Do it the slow way */
1333         kaddr = kmap(page);
1334         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1335         kunmap(page);
1336
1337         if (left) {
1338                 size -= left;
1339                 desc->error = -EFAULT;
1340         }
1341 success:
1342         desc->count = count - size;
1343         desc->written += size;
1344         desc->arg.buf += size;
1345         return size;
1346 }
1347
1348 /*
1349  * Performs necessary checks before doing a write
1350  * @iov:        io vector request
1351  * @nr_segs:    number of segments in the iovec
1352  * @count:      number of bytes to write
1353  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1354  *
1355  * Adjust number of segments and amount of bytes to write (nr_segs should be
1356  * properly initialized first). Returns appropriate error code that caller
1357  * should return or zero in case that write should be allowed.
1358  */
1359 int generic_segment_checks(const struct iovec *iov,
1360                         unsigned long *nr_segs, size_t *count, int access_flags)
1361 {
1362         unsigned long   seg;
1363         size_t cnt = 0;
1364         for (seg = 0; seg < *nr_segs; seg++) {
1365                 const struct iovec *iv = &iov[seg];
1366
1367                 /*
1368                  * If any segment has a negative length, or the cumulative
1369                  * length ever wraps negative then return -EINVAL.
1370                  */
1371                 cnt += iv->iov_len;
1372                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1373                         return -EINVAL;
1374                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1375                         continue;
1376                 if (seg == 0)
1377                         return -EFAULT;
1378                 *nr_segs = seg;
1379                 cnt -= iv->iov_len;     /* This segment is no good */
1380                 break;
1381         }
1382         *count = cnt;
1383         return 0;
1384 }
1385 EXPORT_SYMBOL(generic_segment_checks);
1386
1387 /**
1388  * generic_file_aio_read - generic filesystem read routine
1389  * @iocb:       kernel I/O control block
1390  * @iov:        io vector request
1391  * @nr_segs:    number of segments in the iovec
1392  * @pos:        current file position
1393  *
1394  * This is the "read()" routine for all filesystems
1395  * that can use the page cache directly.
1396  */
1397 ssize_t
1398 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1399                 unsigned long nr_segs, loff_t pos)
1400 {
1401         struct file *filp = iocb->ki_filp;
1402         ssize_t retval;
1403         unsigned long seg = 0;
1404         size_t count;
1405         loff_t *ppos = &iocb->ki_pos;
1406
1407         count = 0;
1408         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1409         if (retval)
1410                 return retval;
1411
1412         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1413         if (filp->f_flags & O_DIRECT) {
1414                 loff_t size;
1415                 struct address_space *mapping;
1416                 struct inode *inode;
1417
1418                 mapping = filp->f_mapping;
1419                 inode = mapping->host;
1420                 if (!count)
1421                         goto out; /* skip atime */
1422                 size = i_size_read(inode);
1423                 if (pos < size) {
1424                         retval = filemap_write_and_wait_range(mapping, pos,
1425                                         pos + iov_length(iov, nr_segs) - 1);
1426                         if (!retval) {
1427                                 struct blk_plug plug;
1428
1429                                 blk_start_plug(&plug);
1430                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1431                                                         iov, pos, nr_segs);
1432                                 blk_finish_plug(&plug);
1433                         }
1434                         if (retval > 0) {
1435                                 *ppos = pos + retval;
1436                                 count -= retval;
1437                         }
1438
1439                         /*
1440                          * Btrfs can have a short DIO read if we encounter
1441                          * compressed extents, so if there was an error, or if
1442                          * we've already read everything we wanted to, or if
1443                          * there was a short read because we hit EOF, go ahead
1444                          * and return.  Otherwise fallthrough to buffered io for
1445                          * the rest of the read.
1446                          */
1447                         if (retval < 0 || !count || *ppos >= size) {
1448                                 file_accessed(filp);
1449                                 goto out;
1450                         }
1451                 }
1452         }
1453
1454         count = retval;
1455         for (seg = 0; seg < nr_segs; seg++) {
1456                 read_descriptor_t desc;
1457                 loff_t offset = 0;
1458
1459                 /*
1460                  * If we did a short DIO read we need to skip the section of the
1461                  * iov that we've already read data into.
1462                  */
1463                 if (count) {
1464                         if (count > iov[seg].iov_len) {
1465                                 count -= iov[seg].iov_len;
1466                                 continue;
1467                         }
1468                         offset = count;
1469                         count = 0;
1470                 }
1471
1472                 desc.written = 0;
1473                 desc.arg.buf = iov[seg].iov_base + offset;
1474                 desc.count = iov[seg].iov_len - offset;
1475                 if (desc.count == 0)
1476                         continue;
1477                 desc.error = 0;
1478                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1479                 retval += desc.written;
1480                 if (desc.error) {
1481                         retval = retval ?: desc.error;
1482                         break;
1483                 }
1484                 if (desc.count > 0)
1485                         break;
1486         }
1487 out:
1488         return retval;
1489 }
1490 EXPORT_SYMBOL(generic_file_aio_read);
1491
1492 static ssize_t
1493 do_readahead(struct address_space *mapping, struct file *filp,
1494              pgoff_t index, unsigned long nr)
1495 {
1496         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1497                 return -EINVAL;
1498
1499         force_page_cache_readahead(mapping, filp, index, nr);
1500         return 0;
1501 }
1502
1503 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1504 {
1505         ssize_t ret;
1506         struct file *file;
1507
1508         ret = -EBADF;
1509         file = fget(fd);
1510         if (file) {
1511                 if (file->f_mode & FMODE_READ) {
1512                         struct address_space *mapping = file->f_mapping;
1513                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1514                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1515                         unsigned long len = end - start + 1;
1516                         ret = do_readahead(mapping, file, start, len);
1517                 }
1518                 fput(file);
1519         }
1520         return ret;
1521 }
1522 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1523 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1524 {
1525         return SYSC_readahead((int) fd, offset, (size_t) count);
1526 }
1527 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1528 #endif
1529
1530 #ifdef CONFIG_MMU
1531 /**
1532  * page_cache_read - adds requested page to the page cache if not already there
1533  * @file:       file to read
1534  * @offset:     page index
1535  *
1536  * This adds the requested page to the page cache if it isn't already there,
1537  * and schedules an I/O to read in its contents from disk.
1538  */
1539 static int page_cache_read(struct file *file, pgoff_t offset)
1540 {
1541         struct address_space *mapping = file->f_mapping;
1542         struct page *page; 
1543         int ret;
1544
1545         do {
1546                 page = page_cache_alloc_cold(mapping);
1547                 if (!page)
1548                         return -ENOMEM;
1549
1550                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1551                 if (ret == 0)
1552                         ret = mapping->a_ops->readpage(file, page);
1553                 else if (ret == -EEXIST)
1554                         ret = 0; /* losing race to add is OK */
1555
1556                 page_cache_release(page);
1557
1558         } while (ret == AOP_TRUNCATED_PAGE);
1559                 
1560         return ret;
1561 }
1562
1563 #define MMAP_LOTSAMISS  (100)
1564
1565 /*
1566  * Synchronous readahead happens when we don't even find
1567  * a page in the page cache at all.
1568  */
1569 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1570                                    struct file_ra_state *ra,
1571                                    struct file *file,
1572                                    pgoff_t offset)
1573 {
1574         unsigned long ra_pages;
1575         struct address_space *mapping = file->f_mapping;
1576
1577         /* If we don't want any read-ahead, don't bother */
1578         if (VM_RandomReadHint(vma))
1579                 return;
1580         if (!ra->ra_pages)
1581                 return;
1582
1583         if (VM_SequentialReadHint(vma)) {
1584                 page_cache_sync_readahead(mapping, ra, file, offset,
1585                                           ra->ra_pages);
1586                 return;
1587         }
1588
1589         /* Avoid banging the cache line if not needed */
1590         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1591                 ra->mmap_miss++;
1592
1593         /*
1594          * Do we miss much more than hit in this file? If so,
1595          * stop bothering with read-ahead. It will only hurt.
1596          */
1597         if (ra->mmap_miss > MMAP_LOTSAMISS)
1598                 return;
1599
1600         /*
1601          * mmap read-around
1602          */
1603         ra_pages = max_sane_readahead(ra->ra_pages);
1604         ra->start = max_t(long, 0, offset - ra_pages / 2);
1605         ra->size = ra_pages;
1606         ra->async_size = ra_pages / 4;
1607         ra_submit(ra, mapping, file);
1608 }
1609
1610 /*
1611  * Asynchronous readahead happens when we find the page and PG_readahead,
1612  * so we want to possibly extend the readahead further..
1613  */
1614 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1615                                     struct file_ra_state *ra,
1616                                     struct file *file,
1617                                     struct page *page,
1618                                     pgoff_t offset)
1619 {
1620         struct address_space *mapping = file->f_mapping;
1621
1622         /* If we don't want any read-ahead, don't bother */
1623         if (VM_RandomReadHint(vma))
1624                 return;
1625         if (ra->mmap_miss > 0)
1626                 ra->mmap_miss--;
1627         if (PageReadahead(page))
1628                 page_cache_async_readahead(mapping, ra, file,
1629                                            page, offset, ra->ra_pages);
1630 }
1631
1632 /**
1633  * filemap_fault - read in file data for page fault handling
1634  * @vma:        vma in which the fault was taken
1635  * @vmf:        struct vm_fault containing details of the fault
1636  *
1637  * filemap_fault() is invoked via the vma operations vector for a
1638  * mapped memory region to read in file data during a page fault.
1639  *
1640  * The goto's are kind of ugly, but this streamlines the normal case of having
1641  * it in the page cache, and handles the special cases reasonably without
1642  * having a lot of duplicated code.
1643  */
1644 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1645 {
1646         int error;
1647         struct file *file = vma->vm_file;
1648         struct address_space *mapping = file->f_mapping;
1649         struct file_ra_state *ra = &file->f_ra;
1650         struct inode *inode = mapping->host;
1651         pgoff_t offset = vmf->pgoff;
1652         struct page *page;
1653         pgoff_t size;
1654         int ret = 0;
1655
1656         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1657         if (offset >= size)
1658                 return VM_FAULT_SIGBUS;
1659
1660         /*
1661          * Do we have something in the page cache already?
1662          */
1663         page = find_get_page(mapping, offset);
1664         if (likely(page)) {
1665                 /*
1666                  * We found the page, so try async readahead before
1667                  * waiting for the lock.
1668                  */
1669                 do_async_mmap_readahead(vma, ra, file, page, offset);
1670         } else {
1671                 /* No page in the page cache at all */
1672                 do_sync_mmap_readahead(vma, ra, file, offset);
1673                 count_vm_event(PGMAJFAULT);
1674                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1675                 ret = VM_FAULT_MAJOR;
1676 retry_find:
1677                 page = find_get_page(mapping, offset);
1678                 if (!page)
1679                         goto no_cached_page;
1680         }
1681
1682         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1683                 page_cache_release(page);
1684                 return ret | VM_FAULT_RETRY;
1685         }
1686
1687         /* Did it get truncated? */
1688         if (unlikely(page->mapping != mapping)) {
1689                 unlock_page(page);
1690                 put_page(page);
1691                 goto retry_find;
1692         }
1693         VM_BUG_ON(page->index != offset);
1694
1695         /*
1696          * We have a locked page in the page cache, now we need to check
1697          * that it's up-to-date. If not, it is going to be due to an error.
1698          */
1699         if (unlikely(!PageUptodate(page)))
1700                 goto page_not_uptodate;
1701
1702         /*
1703          * Found the page and have a reference on it.
1704          * We must recheck i_size under page lock.
1705          */
1706         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1707         if (unlikely(offset >= size)) {
1708                 unlock_page(page);
1709                 page_cache_release(page);
1710                 return VM_FAULT_SIGBUS;
1711         }
1712
1713         vmf->page = page;
1714         return ret | VM_FAULT_LOCKED;
1715
1716 no_cached_page:
1717         /*
1718          * We're only likely to ever get here if MADV_RANDOM is in
1719          * effect.
1720          */
1721         error = page_cache_read(file, offset);
1722
1723         /*
1724          * The page we want has now been added to the page cache.
1725          * In the unlikely event that someone removed it in the
1726          * meantime, we'll just come back here and read it again.
1727          */
1728         if (error >= 0)
1729                 goto retry_find;
1730
1731         /*
1732          * An error return from page_cache_read can result if the
1733          * system is low on memory, or a problem occurs while trying
1734          * to schedule I/O.
1735          */
1736         if (error == -ENOMEM)
1737                 return VM_FAULT_OOM;
1738         return VM_FAULT_SIGBUS;
1739
1740 page_not_uptodate:
1741         /*
1742          * Umm, take care of errors if the page isn't up-to-date.
1743          * Try to re-read it _once_. We do this synchronously,
1744          * because there really aren't any performance issues here
1745          * and we need to check for errors.
1746          */
1747         ClearPageError(page);
1748         error = mapping->a_ops->readpage(file, page);
1749         if (!error) {
1750                 wait_on_page_locked(page);
1751                 if (!PageUptodate(page))
1752                         error = -EIO;
1753         }
1754         page_cache_release(page);
1755
1756         if (!error || error == AOP_TRUNCATED_PAGE)
1757                 goto retry_find;
1758
1759         /* Things didn't work out. Return zero to tell the mm layer so. */
1760         shrink_readahead_size_eio(file, ra);
1761         return VM_FAULT_SIGBUS;
1762 }
1763 EXPORT_SYMBOL(filemap_fault);
1764
1765 const struct vm_operations_struct generic_file_vm_ops = {
1766         .fault          = filemap_fault,
1767 };
1768
1769 /* This is used for a general mmap of a disk file */
1770
1771 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1772 {
1773         struct address_space *mapping = file->f_mapping;
1774
1775         if (!mapping->a_ops->readpage)
1776                 return -ENOEXEC;
1777         file_accessed(file);
1778         vma->vm_ops = &generic_file_vm_ops;
1779         vma->vm_flags |= VM_CAN_NONLINEAR;
1780         return 0;
1781 }
1782
1783 /*
1784  * This is for filesystems which do not implement ->writepage.
1785  */
1786 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1787 {
1788         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1789                 return -EINVAL;
1790         return generic_file_mmap(file, vma);
1791 }
1792 #else
1793 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1794 {
1795         return -ENOSYS;
1796 }
1797 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1798 {
1799         return -ENOSYS;
1800 }
1801 #endif /* CONFIG_MMU */
1802
1803 EXPORT_SYMBOL(generic_file_mmap);
1804 EXPORT_SYMBOL(generic_file_readonly_mmap);
1805
1806 static struct page *__read_cache_page(struct address_space *mapping,
1807                                 pgoff_t index,
1808                                 int (*filler)(void *, struct page *),
1809                                 void *data,
1810                                 gfp_t gfp)
1811 {
1812         struct page *page;
1813         int err;
1814 repeat:
1815         page = find_get_page(mapping, index);
1816         if (!page) {
1817                 page = __page_cache_alloc(gfp | __GFP_COLD);
1818                 if (!page)
1819                         return ERR_PTR(-ENOMEM);
1820                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1821                 if (unlikely(err)) {
1822                         page_cache_release(page);
1823                         if (err == -EEXIST)
1824                                 goto repeat;
1825                         /* Presumably ENOMEM for radix tree node */
1826                         return ERR_PTR(err);
1827                 }
1828                 err = filler(data, page);
1829                 if (err < 0) {
1830                         page_cache_release(page);
1831                         page = ERR_PTR(err);
1832                 }
1833         }
1834         return page;
1835 }
1836
1837 static struct page *do_read_cache_page(struct address_space *mapping,
1838                                 pgoff_t index,
1839                                 int (*filler)(void *, struct page *),
1840                                 void *data,
1841                                 gfp_t gfp)
1842
1843 {
1844         struct page *page;
1845         int err;
1846
1847 retry:
1848         page = __read_cache_page(mapping, index, filler, data, gfp);
1849         if (IS_ERR(page))
1850                 return page;
1851         if (PageUptodate(page))
1852                 goto out;
1853
1854         lock_page(page);
1855         if (!page->mapping) {
1856                 unlock_page(page);
1857                 page_cache_release(page);
1858                 goto retry;
1859         }
1860         if (PageUptodate(page)) {
1861                 unlock_page(page);
1862                 goto out;
1863         }
1864         err = filler(data, page);
1865         if (err < 0) {
1866                 page_cache_release(page);
1867                 return ERR_PTR(err);
1868         }
1869 out:
1870         mark_page_accessed(page);
1871         return page;
1872 }
1873
1874 /**
1875  * read_cache_page_async - read into page cache, fill it if needed
1876  * @mapping:    the page's address_space
1877  * @index:      the page index
1878  * @filler:     function to perform the read
1879  * @data:       first arg to filler(data, page) function, often left as NULL
1880  *
1881  * Same as read_cache_page, but don't wait for page to become unlocked
1882  * after submitting it to the filler.
1883  *
1884  * Read into the page cache. If a page already exists, and PageUptodate() is
1885  * not set, try to fill the page but don't wait for it to become unlocked.
1886  *
1887  * If the page does not get brought uptodate, return -EIO.
1888  */
1889 struct page *read_cache_page_async(struct address_space *mapping,
1890                                 pgoff_t index,
1891                                 int (*filler)(void *, struct page *),
1892                                 void *data)
1893 {
1894         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1895 }
1896 EXPORT_SYMBOL(read_cache_page_async);
1897
1898 static struct page *wait_on_page_read(struct page *page)
1899 {
1900         if (!IS_ERR(page)) {
1901                 wait_on_page_locked(page);
1902                 if (!PageUptodate(page)) {
1903                         page_cache_release(page);
1904                         page = ERR_PTR(-EIO);
1905                 }
1906         }
1907         return page;
1908 }
1909
1910 /**
1911  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1912  * @mapping:    the page's address_space
1913  * @index:      the page index
1914  * @gfp:        the page allocator flags to use if allocating
1915  *
1916  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1917  * any new page allocations done using the specified allocation flags.
1918  *
1919  * If the page does not get brought uptodate, return -EIO.
1920  */
1921 struct page *read_cache_page_gfp(struct address_space *mapping,
1922                                 pgoff_t index,
1923                                 gfp_t gfp)
1924 {
1925         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1926
1927         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1928 }
1929 EXPORT_SYMBOL(read_cache_page_gfp);
1930
1931 /**
1932  * read_cache_page - read into page cache, fill it if needed
1933  * @mapping:    the page's address_space
1934  * @index:      the page index
1935  * @filler:     function to perform the read
1936  * @data:       first arg to filler(data, page) function, often left as NULL
1937  *
1938  * Read into the page cache. If a page already exists, and PageUptodate() is
1939  * not set, try to fill the page then wait for it to become unlocked.
1940  *
1941  * If the page does not get brought uptodate, return -EIO.
1942  */
1943 struct page *read_cache_page(struct address_space *mapping,
1944                                 pgoff_t index,
1945                                 int (*filler)(void *, struct page *),
1946                                 void *data)
1947 {
1948         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1949 }
1950 EXPORT_SYMBOL(read_cache_page);
1951
1952 /*
1953  * The logic we want is
1954  *
1955  *      if suid or (sgid and xgrp)
1956  *              remove privs
1957  */
1958 int should_remove_suid(struct dentry *dentry)
1959 {
1960         mode_t mode = dentry->d_inode->i_mode;
1961         int kill = 0;
1962
1963         /* suid always must be killed */
1964         if (unlikely(mode & S_ISUID))
1965                 kill = ATTR_KILL_SUID;
1966
1967         /*
1968          * sgid without any exec bits is just a mandatory locking mark; leave
1969          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1970          */
1971         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1972                 kill |= ATTR_KILL_SGID;
1973
1974         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1975                 return kill;
1976
1977         return 0;
1978 }
1979 EXPORT_SYMBOL(should_remove_suid);
1980
1981 static int __remove_suid(struct dentry *dentry, int kill)
1982 {
1983         struct iattr newattrs;
1984
1985         newattrs.ia_valid = ATTR_FORCE | kill;
1986         return notify_change(dentry, &newattrs);
1987 }
1988
1989 int file_remove_suid(struct file *file)
1990 {
1991         struct dentry *dentry = file->f_path.dentry;
1992         struct inode *inode = dentry->d_inode;
1993         int killsuid;
1994         int killpriv;
1995         int error = 0;
1996
1997         /* Fast path for nothing security related */
1998         if (IS_NOSEC(inode))
1999                 return 0;
2000
2001         killsuid = should_remove_suid(dentry);
2002         killpriv = security_inode_need_killpriv(dentry);
2003
2004         if (killpriv < 0)
2005                 return killpriv;
2006         if (killpriv)
2007                 error = security_inode_killpriv(dentry);
2008         if (!error && killsuid)
2009                 error = __remove_suid(dentry, killsuid);
2010         if (!error && (inode->i_sb->s_flags & MS_NOSEC))
2011                 inode->i_flags |= S_NOSEC;
2012
2013         return error;
2014 }
2015 EXPORT_SYMBOL(file_remove_suid);
2016
2017 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2018                         const struct iovec *iov, size_t base, size_t bytes)
2019 {
2020         size_t copied = 0, left = 0;
2021
2022         while (bytes) {
2023                 char __user *buf = iov->iov_base + base;
2024                 int copy = min(bytes, iov->iov_len - base);
2025
2026                 base = 0;
2027                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2028                 copied += copy;
2029                 bytes -= copy;
2030                 vaddr += copy;
2031                 iov++;
2032
2033                 if (unlikely(left))
2034                         break;
2035         }
2036         return copied - left;
2037 }
2038
2039 /*
2040  * Copy as much as we can into the page and return the number of bytes which
2041  * were successfully copied.  If a fault is encountered then return the number of
2042  * bytes which were copied.
2043  */
2044 size_t iov_iter_copy_from_user_atomic(struct page *page,
2045                 struct iov_iter *i, unsigned long offset, size_t bytes)
2046 {
2047         char *kaddr;
2048         size_t copied;
2049
2050         BUG_ON(!in_atomic());
2051         kaddr = kmap_atomic(page, KM_USER0);
2052         if (likely(i->nr_segs == 1)) {
2053                 int left;
2054                 char __user *buf = i->iov->iov_base + i->iov_offset;
2055                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2056                 copied = bytes - left;
2057         } else {
2058                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2059                                                 i->iov, i->iov_offset, bytes);
2060         }
2061         kunmap_atomic(kaddr, KM_USER0);
2062
2063         return copied;
2064 }
2065 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2066
2067 /*
2068  * This has the same sideeffects and return value as
2069  * iov_iter_copy_from_user_atomic().
2070  * The difference is that it attempts to resolve faults.
2071  * Page must not be locked.
2072  */
2073 size_t iov_iter_copy_from_user(struct page *page,
2074                 struct iov_iter *i, unsigned long offset, size_t bytes)
2075 {
2076         char *kaddr;
2077         size_t copied;
2078
2079         kaddr = kmap(page);
2080         if (likely(i->nr_segs == 1)) {
2081                 int left;
2082                 char __user *buf = i->iov->iov_base + i->iov_offset;
2083                 left = __copy_from_user(kaddr + offset, buf, bytes);
2084                 copied = bytes - left;
2085         } else {
2086                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2087                                                 i->iov, i->iov_offset, bytes);
2088         }
2089         kunmap(page);
2090         return copied;
2091 }
2092 EXPORT_SYMBOL(iov_iter_copy_from_user);
2093
2094 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2095 {
2096         BUG_ON(i->count < bytes);
2097
2098         if (likely(i->nr_segs == 1)) {
2099                 i->iov_offset += bytes;
2100                 i->count -= bytes;
2101         } else {
2102                 const struct iovec *iov = i->iov;
2103                 size_t base = i->iov_offset;
2104                 unsigned long nr_segs = i->nr_segs;
2105
2106                 /*
2107                  * The !iov->iov_len check ensures we skip over unlikely
2108                  * zero-length segments (without overruning the iovec).
2109                  */
2110                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2111                         int copy;
2112
2113                         copy = min(bytes, iov->iov_len - base);
2114                         BUG_ON(!i->count || i->count < copy);
2115                         i->count -= copy;
2116                         bytes -= copy;
2117                         base += copy;
2118                         if (iov->iov_len == base) {
2119                                 iov++;
2120                                 nr_segs--;
2121                                 base = 0;
2122                         }
2123                 }
2124                 i->iov = iov;
2125                 i->iov_offset = base;
2126                 i->nr_segs = nr_segs;
2127         }
2128 }
2129 EXPORT_SYMBOL(iov_iter_advance);
2130
2131 /*
2132  * Fault in the first iovec of the given iov_iter, to a maximum length
2133  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2134  * accessed (ie. because it is an invalid address).
2135  *
2136  * writev-intensive code may want this to prefault several iovecs -- that
2137  * would be possible (callers must not rely on the fact that _only_ the
2138  * first iovec will be faulted with the current implementation).
2139  */
2140 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2141 {
2142         char __user *buf = i->iov->iov_base + i->iov_offset;
2143         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2144         return fault_in_pages_readable(buf, bytes);
2145 }
2146 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2147
2148 /*
2149  * Return the count of just the current iov_iter segment.
2150  */
2151 size_t iov_iter_single_seg_count(struct iov_iter *i)
2152 {
2153         const struct iovec *iov = i->iov;
2154         if (i->nr_segs == 1)
2155                 return i->count;
2156         else
2157                 return min(i->count, iov->iov_len - i->iov_offset);
2158 }
2159 EXPORT_SYMBOL(iov_iter_single_seg_count);
2160
2161 /*
2162  * Performs necessary checks before doing a write
2163  *
2164  * Can adjust writing position or amount of bytes to write.
2165  * Returns appropriate error code that caller should return or
2166  * zero in case that write should be allowed.
2167  */
2168 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2169 {
2170         struct inode *inode = file->f_mapping->host;
2171         unsigned long limit = rlimit(RLIMIT_FSIZE);
2172
2173         if (unlikely(*pos < 0))
2174                 return -EINVAL;
2175
2176         if (!isblk) {
2177                 /* FIXME: this is for backwards compatibility with 2.4 */
2178                 if (file->f_flags & O_APPEND)
2179                         *pos = i_size_read(inode);
2180
2181                 if (limit != RLIM_INFINITY) {
2182                         if (*pos >= limit) {
2183                                 send_sig(SIGXFSZ, current, 0);
2184                                 return -EFBIG;
2185                         }
2186                         if (*count > limit - (typeof(limit))*pos) {
2187                                 *count = limit - (typeof(limit))*pos;
2188                         }
2189                 }
2190         }
2191
2192         /*
2193          * LFS rule
2194          */
2195         if (unlikely(*pos + *count > MAX_NON_LFS &&
2196                                 !(file->f_flags & O_LARGEFILE))) {
2197                 if (*pos >= MAX_NON_LFS) {
2198                         return -EFBIG;
2199                 }
2200                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2201                         *count = MAX_NON_LFS - (unsigned long)*pos;
2202                 }
2203         }
2204
2205         /*
2206          * Are we about to exceed the fs block limit ?
2207          *
2208          * If we have written data it becomes a short write.  If we have
2209          * exceeded without writing data we send a signal and return EFBIG.
2210          * Linus frestrict idea will clean these up nicely..
2211          */
2212         if (likely(!isblk)) {
2213                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2214                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2215                                 return -EFBIG;
2216                         }
2217                         /* zero-length writes at ->s_maxbytes are OK */
2218                 }
2219
2220                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2221                         *count = inode->i_sb->s_maxbytes - *pos;
2222         } else {
2223 #ifdef CONFIG_BLOCK
2224                 loff_t isize;
2225                 if (bdev_read_only(I_BDEV(inode)))
2226                         return -EPERM;
2227                 isize = i_size_read(inode);
2228                 if (*pos >= isize) {
2229                         if (*count || *pos > isize)
2230                                 return -ENOSPC;
2231                 }
2232
2233                 if (*pos + *count > isize)
2234                         *count = isize - *pos;
2235 #else
2236                 return -EPERM;
2237 #endif
2238         }
2239         return 0;
2240 }
2241 EXPORT_SYMBOL(generic_write_checks);
2242
2243 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2244                                 loff_t pos, unsigned len, unsigned flags,
2245                                 struct page **pagep, void **fsdata)
2246 {
2247         const struct address_space_operations *aops = mapping->a_ops;
2248
2249         return aops->write_begin(file, mapping, pos, len, flags,
2250                                                         pagep, fsdata);
2251 }
2252 EXPORT_SYMBOL(pagecache_write_begin);
2253
2254 int pagecache_write_end(struct file *file, struct address_space *mapping,
2255                                 loff_t pos, unsigned len, unsigned copied,
2256                                 struct page *page, void *fsdata)
2257 {
2258         const struct address_space_operations *aops = mapping->a_ops;
2259
2260         mark_page_accessed(page);
2261         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2262 }
2263 EXPORT_SYMBOL(pagecache_write_end);
2264
2265 ssize_t
2266 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2267                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2268                 size_t count, size_t ocount)
2269 {
2270         struct file     *file = iocb->ki_filp;
2271         struct address_space *mapping = file->f_mapping;
2272         struct inode    *inode = mapping->host;
2273         ssize_t         written;
2274         size_t          write_len;
2275         pgoff_t         end;
2276
2277         if (count != ocount)
2278                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2279
2280         write_len = iov_length(iov, *nr_segs);
2281         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2282
2283         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2284         if (written)
2285                 goto out;
2286
2287         /*
2288          * After a write we want buffered reads to be sure to go to disk to get
2289          * the new data.  We invalidate clean cached page from the region we're
2290          * about to write.  We do this *before* the write so that we can return
2291          * without clobbering -EIOCBQUEUED from ->direct_IO().
2292          */
2293         if (mapping->nrpages) {
2294                 written = invalidate_inode_pages2_range(mapping,
2295                                         pos >> PAGE_CACHE_SHIFT, end);
2296                 /*
2297                  * If a page can not be invalidated, return 0 to fall back
2298                  * to buffered write.
2299                  */
2300                 if (written) {
2301                         if (written == -EBUSY)
2302                                 return 0;
2303                         goto out;
2304                 }
2305         }
2306
2307         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2308
2309         /*
2310          * Finally, try again to invalidate clean pages which might have been
2311          * cached by non-direct readahead, or faulted in by get_user_pages()
2312          * if the source of the write was an mmap'ed region of the file
2313          * we're writing.  Either one is a pretty crazy thing to do,
2314          * so we don't support it 100%.  If this invalidation
2315          * fails, tough, the write still worked...
2316          */
2317         if (mapping->nrpages) {
2318                 invalidate_inode_pages2_range(mapping,
2319                                               pos >> PAGE_CACHE_SHIFT, end);
2320         }
2321
2322         if (written > 0) {
2323                 pos += written;
2324                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2325                         i_size_write(inode, pos);
2326                         mark_inode_dirty(inode);
2327                 }
2328                 *ppos = pos;
2329         }
2330 out:
2331         return written;
2332 }
2333 EXPORT_SYMBOL(generic_file_direct_write);
2334
2335 /*
2336  * Find or create a page at the given pagecache position. Return the locked
2337  * page. This function is specifically for buffered writes.
2338  */
2339 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2340                                         pgoff_t index, unsigned flags)
2341 {
2342         int status;
2343         struct page *page;
2344         gfp_t gfp_notmask = 0;
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(mapping_gfp_mask(mapping) & ~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);