tcp: enforce tcp_min_snd_mss in tcp_mtu_probing()
[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                 if (fatal_signal_pending(current)) {
1121                         error = -EINTR;
1122                         goto out;
1123                 }
1124
1125                 page = find_get_page(mapping, index);
1126                 if (!page) {
1127                         page_cache_sync_readahead(mapping,
1128                                         ra, filp,
1129                                         index, last_index - index);
1130                         page = find_get_page(mapping, index);
1131                         if (unlikely(page == NULL))
1132                                 goto no_cached_page;
1133                 }
1134                 if (PageReadahead(page)) {
1135                         page_cache_async_readahead(mapping,
1136                                         ra, filp, page,
1137                                         index, last_index - index);
1138                 }
1139                 if (!PageUptodate(page)) {
1140                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1141                                         !mapping->a_ops->is_partially_uptodate)
1142                                 goto page_not_up_to_date;
1143                         if (!trylock_page(page))
1144                                 goto page_not_up_to_date;
1145                         /* Did it get truncated before we got the lock? */
1146                         if (!page->mapping)
1147                                 goto page_not_up_to_date_locked;
1148                         if (!mapping->a_ops->is_partially_uptodate(page,
1149                                                                 desc, offset))
1150                                 goto page_not_up_to_date_locked;
1151                         unlock_page(page);
1152                 }
1153 page_ok:
1154                 /*
1155                  * i_size must be checked after we know the page is Uptodate.
1156                  *
1157                  * Checking i_size after the check allows us to calculate
1158                  * the correct value for "nr", which means the zero-filled
1159                  * part of the page is not copied back to userspace (unless
1160                  * another truncate extends the file - this is desired though).
1161                  */
1162
1163                 isize = i_size_read(inode);
1164                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1165                 if (unlikely(!isize || index > end_index)) {
1166                         page_cache_release(page);
1167                         goto out;
1168                 }
1169
1170                 /* nr is the maximum number of bytes to copy from this page */
1171                 nr = PAGE_CACHE_SIZE;
1172                 if (index == end_index) {
1173                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1174                         if (nr <= offset) {
1175                                 page_cache_release(page);
1176                                 goto out;
1177                         }
1178                 }
1179                 nr = nr - offset;
1180
1181                 /* If users can be writing to this page using arbitrary
1182                  * virtual addresses, take care about potential aliasing
1183                  * before reading the page on the kernel side.
1184                  */
1185                 if (mapping_writably_mapped(mapping))
1186                         flush_dcache_page(page);
1187
1188                 /*
1189                  * When a sequential read accesses a page several times,
1190                  * only mark it as accessed the first time.
1191                  */
1192                 if (prev_index != index || offset != prev_offset)
1193                         mark_page_accessed(page);
1194                 prev_index = index;
1195
1196                 /*
1197                  * Ok, we have the page, and it's up-to-date, so
1198                  * now we can copy it to user space...
1199                  *
1200                  * The actor routine returns how many bytes were actually used..
1201                  * NOTE! This may not be the same as how much of a user buffer
1202                  * we filled up (we may be padding etc), so we can only update
1203                  * "pos" here (the actor routine has to update the user buffer
1204                  * pointers and the remaining count).
1205                  */
1206                 ret = actor(desc, page, offset, nr);
1207                 offset += ret;
1208                 index += offset >> PAGE_CACHE_SHIFT;
1209                 offset &= ~PAGE_CACHE_MASK;
1210                 prev_offset = offset;
1211
1212                 page_cache_release(page);
1213                 if (ret == nr && desc->count)
1214                         continue;
1215                 goto out;
1216
1217 page_not_up_to_date:
1218                 /* Get exclusive access to the page ... */
1219                 error = lock_page_killable(page);
1220                 if (unlikely(error))
1221                         goto readpage_error;
1222
1223 page_not_up_to_date_locked:
1224                 /* Did it get truncated before we got the lock? */
1225                 if (!page->mapping) {
1226                         unlock_page(page);
1227                         page_cache_release(page);
1228                         continue;
1229                 }
1230
1231                 /* Did somebody else fill it already? */
1232                 if (PageUptodate(page)) {
1233                         unlock_page(page);
1234                         goto page_ok;
1235                 }
1236
1237 readpage:
1238                 /*
1239                  * A previous I/O error may have been due to temporary
1240                  * failures, eg. multipath errors.
1241                  * PG_error will be set again if readpage fails.
1242                  */
1243                 ClearPageError(page);
1244                 /* Start the actual read. The read will unlock the page. */
1245                 error = mapping->a_ops->readpage(filp, page);
1246
1247                 if (unlikely(error)) {
1248                         if (error == AOP_TRUNCATED_PAGE) {
1249                                 page_cache_release(page);
1250                                 goto find_page;
1251                         }
1252                         goto readpage_error;
1253                 }
1254
1255                 if (!PageUptodate(page)) {
1256                         error = lock_page_killable(page);
1257                         if (unlikely(error))
1258                                 goto readpage_error;
1259                         if (!PageUptodate(page)) {
1260                                 if (page->mapping == NULL) {
1261                                         /*
1262                                          * invalidate_mapping_pages got it
1263                                          */
1264                                         unlock_page(page);
1265                                         page_cache_release(page);
1266                                         goto find_page;
1267                                 }
1268                                 unlock_page(page);
1269                                 shrink_readahead_size_eio(filp, ra);
1270                                 error = -EIO;
1271                                 goto readpage_error;
1272                         }
1273                         unlock_page(page);
1274                 }
1275
1276                 goto page_ok;
1277
1278 readpage_error:
1279                 /* UHHUH! A synchronous read error occurred. Report it */
1280                 desc->error = error;
1281                 page_cache_release(page);
1282                 goto out;
1283
1284 no_cached_page:
1285                 /*
1286                  * Ok, it wasn't cached, so we need to create a new
1287                  * page..
1288                  */
1289                 page = page_cache_alloc_cold(mapping);
1290                 if (!page) {
1291                         desc->error = -ENOMEM;
1292                         goto out;
1293                 }
1294                 error = add_to_page_cache_lru(page, mapping,
1295                                                 index, GFP_KERNEL);
1296                 if (error) {
1297                         page_cache_release(page);
1298                         if (error == -EEXIST)
1299                                 goto find_page;
1300                         desc->error = error;
1301                         goto out;
1302                 }
1303                 goto readpage;
1304         }
1305
1306 out:
1307         ra->prev_pos = prev_index;
1308         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1309         ra->prev_pos |= prev_offset;
1310
1311         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1312         file_accessed(filp);
1313 }
1314
1315 int file_read_actor(read_descriptor_t *desc, struct page *page,
1316                         unsigned long offset, unsigned long size)
1317 {
1318         char *kaddr;
1319         unsigned long left, count = desc->count;
1320
1321         if (size > count)
1322                 size = count;
1323
1324         /*
1325          * Faults on the destination of a read are common, so do it before
1326          * taking the kmap.
1327          */
1328         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1329                 kaddr = kmap_atomic(page, KM_USER0);
1330                 left = __copy_to_user_inatomic(desc->arg.buf,
1331                                                 kaddr + offset, size);
1332                 kunmap_atomic(kaddr, KM_USER0);
1333                 if (left == 0)
1334                         goto success;
1335         }
1336
1337         /* Do it the slow way */
1338         kaddr = kmap(page);
1339         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1340         kunmap(page);
1341
1342         if (left) {
1343                 size -= left;
1344                 desc->error = -EFAULT;
1345         }
1346 success:
1347         desc->count = count - size;
1348         desc->written += size;
1349         desc->arg.buf += size;
1350         return size;
1351 }
1352
1353 /*
1354  * Performs necessary checks before doing a write
1355  * @iov:        io vector request
1356  * @nr_segs:    number of segments in the iovec
1357  * @count:      number of bytes to write
1358  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1359  *
1360  * Adjust number of segments and amount of bytes to write (nr_segs should be
1361  * properly initialized first). Returns appropriate error code that caller
1362  * should return or zero in case that write should be allowed.
1363  */
1364 int generic_segment_checks(const struct iovec *iov,
1365                         unsigned long *nr_segs, size_t *count, int access_flags)
1366 {
1367         unsigned long   seg;
1368         size_t cnt = 0;
1369         for (seg = 0; seg < *nr_segs; seg++) {
1370                 const struct iovec *iv = &iov[seg];
1371
1372                 /*
1373                  * If any segment has a negative length, or the cumulative
1374                  * length ever wraps negative then return -EINVAL.
1375                  */
1376                 cnt += iv->iov_len;
1377                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1378                         return -EINVAL;
1379                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1380                         continue;
1381                 if (seg == 0)
1382                         return -EFAULT;
1383                 *nr_segs = seg;
1384                 cnt -= iv->iov_len;     /* This segment is no good */
1385                 break;
1386         }
1387         *count = cnt;
1388         return 0;
1389 }
1390 EXPORT_SYMBOL(generic_segment_checks);
1391
1392 /**
1393  * generic_file_aio_read - generic filesystem read routine
1394  * @iocb:       kernel I/O control block
1395  * @iov:        io vector request
1396  * @nr_segs:    number of segments in the iovec
1397  * @pos:        current file position
1398  *
1399  * This is the "read()" routine for all filesystems
1400  * that can use the page cache directly.
1401  */
1402 ssize_t
1403 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1404                 unsigned long nr_segs, loff_t pos)
1405 {
1406         struct file *filp = iocb->ki_filp;
1407         ssize_t retval;
1408         unsigned long seg = 0;
1409         size_t count;
1410         loff_t *ppos = &iocb->ki_pos;
1411
1412         count = 0;
1413         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1414         if (retval)
1415                 return retval;
1416
1417         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1418         if (filp->f_flags & O_DIRECT) {
1419                 loff_t size;
1420                 struct address_space *mapping;
1421                 struct inode *inode;
1422
1423                 mapping = filp->f_mapping;
1424                 inode = mapping->host;
1425                 if (!count)
1426                         goto out; /* skip atime */
1427                 size = i_size_read(inode);
1428                 if (pos < size) {
1429                         retval = filemap_write_and_wait_range(mapping, pos,
1430                                         pos + iov_length(iov, nr_segs) - 1);
1431                         if (!retval) {
1432                                 struct blk_plug plug;
1433
1434                                 blk_start_plug(&plug);
1435                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1436                                                         iov, pos, nr_segs);
1437                                 blk_finish_plug(&plug);
1438                         }
1439                         if (retval > 0) {
1440                                 *ppos = pos + retval;
1441                                 count -= retval;
1442                         }
1443
1444                         /*
1445                          * Btrfs can have a short DIO read if we encounter
1446                          * compressed extents, so if there was an error, or if
1447                          * we've already read everything we wanted to, or if
1448                          * there was a short read because we hit EOF, go ahead
1449                          * and return.  Otherwise fallthrough to buffered io for
1450                          * the rest of the read.
1451                          */
1452                         if (retval < 0 || !count || *ppos >= size) {
1453                                 file_accessed(filp);
1454                                 goto out;
1455                         }
1456                 }
1457         }
1458
1459         count = retval;
1460         for (seg = 0; seg < nr_segs; seg++) {
1461                 read_descriptor_t desc;
1462                 loff_t offset = 0;
1463
1464                 /*
1465                  * If we did a short DIO read we need to skip the section of the
1466                  * iov that we've already read data into.
1467                  */
1468                 if (count) {
1469                         if (count > iov[seg].iov_len) {
1470                                 count -= iov[seg].iov_len;
1471                                 continue;
1472                         }
1473                         offset = count;
1474                         count = 0;
1475                 }
1476
1477                 desc.written = 0;
1478                 desc.arg.buf = iov[seg].iov_base + offset;
1479                 desc.count = iov[seg].iov_len - offset;
1480                 if (desc.count == 0)
1481                         continue;
1482                 desc.error = 0;
1483                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1484                 retval += desc.written;
1485                 if (desc.error) {
1486                         retval = retval ?: desc.error;
1487                         break;
1488                 }
1489                 if (desc.count > 0)
1490                         break;
1491         }
1492 out:
1493         return retval;
1494 }
1495 EXPORT_SYMBOL(generic_file_aio_read);
1496
1497 static ssize_t
1498 do_readahead(struct address_space *mapping, struct file *filp,
1499              pgoff_t index, unsigned long nr)
1500 {
1501         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1502                 return -EINVAL;
1503
1504         force_page_cache_readahead(mapping, filp, index, nr);
1505         return 0;
1506 }
1507
1508 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1509 {
1510         ssize_t ret;
1511         struct file *file;
1512
1513         ret = -EBADF;
1514         file = fget(fd);
1515         if (file) {
1516                 if (file->f_mode & FMODE_READ) {
1517                         struct address_space *mapping = file->f_mapping;
1518                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1519                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1520                         unsigned long len = end - start + 1;
1521                         ret = do_readahead(mapping, file, start, len);
1522                 }
1523                 fput(file);
1524         }
1525         return ret;
1526 }
1527 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1528 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1529 {
1530         return SYSC_readahead((int) fd, offset, (size_t) count);
1531 }
1532 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1533 #endif
1534
1535 #ifdef CONFIG_MMU
1536 /**
1537  * page_cache_read - adds requested page to the page cache if not already there
1538  * @file:       file to read
1539  * @offset:     page index
1540  *
1541  * This adds the requested page to the page cache if it isn't already there,
1542  * and schedules an I/O to read in its contents from disk.
1543  */
1544 static int page_cache_read(struct file *file, pgoff_t offset)
1545 {
1546         struct address_space *mapping = file->f_mapping;
1547         struct page *page; 
1548         int ret;
1549
1550         do {
1551                 page = page_cache_alloc_cold(mapping);
1552                 if (!page)
1553                         return -ENOMEM;
1554
1555                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1556                 if (ret == 0)
1557                         ret = mapping->a_ops->readpage(file, page);
1558                 else if (ret == -EEXIST)
1559                         ret = 0; /* losing race to add is OK */
1560
1561                 page_cache_release(page);
1562
1563         } while (ret == AOP_TRUNCATED_PAGE);
1564                 
1565         return ret;
1566 }
1567
1568 #define MMAP_LOTSAMISS  (100)
1569
1570 /*
1571  * Synchronous readahead happens when we don't even find
1572  * a page in the page cache at all.
1573  */
1574 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1575                                    struct file_ra_state *ra,
1576                                    struct file *file,
1577                                    pgoff_t offset)
1578 {
1579         unsigned long ra_pages;
1580         struct address_space *mapping = file->f_mapping;
1581
1582         /* If we don't want any read-ahead, don't bother */
1583         if (VM_RandomReadHint(vma))
1584                 return;
1585         if (!ra->ra_pages)
1586                 return;
1587
1588         if (VM_SequentialReadHint(vma)) {
1589                 page_cache_sync_readahead(mapping, ra, file, offset,
1590                                           ra->ra_pages);
1591                 return;
1592         }
1593
1594         /* Avoid banging the cache line if not needed */
1595         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1596                 ra->mmap_miss++;
1597
1598         /*
1599          * Do we miss much more than hit in this file? If so,
1600          * stop bothering with read-ahead. It will only hurt.
1601          */
1602         if (ra->mmap_miss > MMAP_LOTSAMISS)
1603                 return;
1604
1605         /*
1606          * mmap read-around
1607          */
1608         ra_pages = max_sane_readahead(ra->ra_pages);
1609         ra->start = max_t(long, 0, offset - ra_pages / 2);
1610         ra->size = ra_pages;
1611         ra->async_size = ra_pages / 4;
1612         ra_submit(ra, mapping, file);
1613 }
1614
1615 /*
1616  * Asynchronous readahead happens when we find the page and PG_readahead,
1617  * so we want to possibly extend the readahead further..
1618  */
1619 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1620                                     struct file_ra_state *ra,
1621                                     struct file *file,
1622                                     struct page *page,
1623                                     pgoff_t offset)
1624 {
1625         struct address_space *mapping = file->f_mapping;
1626
1627         /* If we don't want any read-ahead, don't bother */
1628         if (VM_RandomReadHint(vma))
1629                 return;
1630         if (ra->mmap_miss > 0)
1631                 ra->mmap_miss--;
1632         if (PageReadahead(page))
1633                 page_cache_async_readahead(mapping, ra, file,
1634                                            page, offset, ra->ra_pages);
1635 }
1636
1637 /**
1638  * filemap_fault - read in file data for page fault handling
1639  * @vma:        vma in which the fault was taken
1640  * @vmf:        struct vm_fault containing details of the fault
1641  *
1642  * filemap_fault() is invoked via the vma operations vector for a
1643  * mapped memory region to read in file data during a page fault.
1644  *
1645  * The goto's are kind of ugly, but this streamlines the normal case of having
1646  * it in the page cache, and handles the special cases reasonably without
1647  * having a lot of duplicated code.
1648  */
1649 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1650 {
1651         int error;
1652         struct file *file = vma->vm_file;
1653         struct address_space *mapping = file->f_mapping;
1654         struct file_ra_state *ra = &file->f_ra;
1655         struct inode *inode = mapping->host;
1656         pgoff_t offset = vmf->pgoff;
1657         struct page *page;
1658         pgoff_t size;
1659         int ret = 0;
1660
1661         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1662         if (offset >= size)
1663                 return VM_FAULT_SIGBUS;
1664
1665         /*
1666          * Do we have something in the page cache already?
1667          */
1668         page = find_get_page(mapping, offset);
1669         if (likely(page)) {
1670                 /*
1671                  * We found the page, so try async readahead before
1672                  * waiting for the lock.
1673                  */
1674                 do_async_mmap_readahead(vma, ra, file, page, offset);
1675         } else {
1676                 /* No page in the page cache at all */
1677                 do_sync_mmap_readahead(vma, ra, file, offset);
1678                 count_vm_event(PGMAJFAULT);
1679                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1680                 ret = VM_FAULT_MAJOR;
1681 retry_find:
1682                 page = find_get_page(mapping, offset);
1683                 if (!page)
1684                         goto no_cached_page;
1685         }
1686
1687         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1688                 page_cache_release(page);
1689                 return ret | VM_FAULT_RETRY;
1690         }
1691
1692         /* Did it get truncated? */
1693         if (unlikely(page->mapping != mapping)) {
1694                 unlock_page(page);
1695                 put_page(page);
1696                 goto retry_find;
1697         }
1698         VM_BUG_ON(page->index != offset);
1699
1700         /*
1701          * We have a locked page in the page cache, now we need to check
1702          * that it's up-to-date. If not, it is going to be due to an error.
1703          */
1704         if (unlikely(!PageUptodate(page)))
1705                 goto page_not_uptodate;
1706
1707         /*
1708          * Found the page and have a reference on it.
1709          * We must recheck i_size under page lock.
1710          */
1711         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1712         if (unlikely(offset >= size)) {
1713                 unlock_page(page);
1714                 page_cache_release(page);
1715                 return VM_FAULT_SIGBUS;
1716         }
1717
1718         vmf->page = page;
1719         return ret | VM_FAULT_LOCKED;
1720
1721 no_cached_page:
1722         /*
1723          * We're only likely to ever get here if MADV_RANDOM is in
1724          * effect.
1725          */
1726         error = page_cache_read(file, offset);
1727
1728         /*
1729          * The page we want has now been added to the page cache.
1730          * In the unlikely event that someone removed it in the
1731          * meantime, we'll just come back here and read it again.
1732          */
1733         if (error >= 0)
1734                 goto retry_find;
1735
1736         /*
1737          * An error return from page_cache_read can result if the
1738          * system is low on memory, or a problem occurs while trying
1739          * to schedule I/O.
1740          */
1741         if (error == -ENOMEM)
1742                 return VM_FAULT_OOM;
1743         return VM_FAULT_SIGBUS;
1744
1745 page_not_uptodate:
1746         /*
1747          * Umm, take care of errors if the page isn't up-to-date.
1748          * Try to re-read it _once_. We do this synchronously,
1749          * because there really aren't any performance issues here
1750          * and we need to check for errors.
1751          */
1752         ClearPageError(page);
1753         error = mapping->a_ops->readpage(file, page);
1754         if (!error) {
1755                 wait_on_page_locked(page);
1756                 if (!PageUptodate(page))
1757                         error = -EIO;
1758         }
1759         page_cache_release(page);
1760
1761         if (!error || error == AOP_TRUNCATED_PAGE)
1762                 goto retry_find;
1763
1764         /* Things didn't work out. Return zero to tell the mm layer so. */
1765         shrink_readahead_size_eio(file, ra);
1766         return VM_FAULT_SIGBUS;
1767 }
1768 EXPORT_SYMBOL(filemap_fault);
1769
1770 const struct vm_operations_struct generic_file_vm_ops = {
1771         .fault          = filemap_fault,
1772 };
1773
1774 /* This is used for a general mmap of a disk file */
1775
1776 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1777 {
1778         struct address_space *mapping = file->f_mapping;
1779
1780         if (!mapping->a_ops->readpage)
1781                 return -ENOEXEC;
1782         file_accessed(file);
1783         vma->vm_ops = &generic_file_vm_ops;
1784         vma->vm_flags |= VM_CAN_NONLINEAR;
1785         return 0;
1786 }
1787
1788 /*
1789  * This is for filesystems which do not implement ->writepage.
1790  */
1791 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1792 {
1793         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1794                 return -EINVAL;
1795         return generic_file_mmap(file, vma);
1796 }
1797 #else
1798 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1799 {
1800         return -ENOSYS;
1801 }
1802 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1803 {
1804         return -ENOSYS;
1805 }
1806 #endif /* CONFIG_MMU */
1807
1808 EXPORT_SYMBOL(generic_file_mmap);
1809 EXPORT_SYMBOL(generic_file_readonly_mmap);
1810
1811 static struct page *__read_cache_page(struct address_space *mapping,
1812                                 pgoff_t index,
1813                                 int (*filler)(void *, struct page *),
1814                                 void *data,
1815                                 gfp_t gfp)
1816 {
1817         struct page *page;
1818         int err;
1819 repeat:
1820         page = find_get_page(mapping, index);
1821         if (!page) {
1822                 page = __page_cache_alloc(gfp | __GFP_COLD);
1823                 if (!page)
1824                         return ERR_PTR(-ENOMEM);
1825                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1826                 if (unlikely(err)) {
1827                         page_cache_release(page);
1828                         if (err == -EEXIST)
1829                                 goto repeat;
1830                         /* Presumably ENOMEM for radix tree node */
1831                         return ERR_PTR(err);
1832                 }
1833                 err = filler(data, page);
1834                 if (err < 0) {
1835                         page_cache_release(page);
1836                         page = ERR_PTR(err);
1837                 }
1838         }
1839         return page;
1840 }
1841
1842 static struct page *do_read_cache_page(struct address_space *mapping,
1843                                 pgoff_t index,
1844                                 int (*filler)(void *, struct page *),
1845                                 void *data,
1846                                 gfp_t gfp)
1847
1848 {
1849         struct page *page;
1850         int err;
1851
1852 retry:
1853         page = __read_cache_page(mapping, index, filler, data, gfp);
1854         if (IS_ERR(page))
1855                 return page;
1856         if (PageUptodate(page))
1857                 goto out;
1858
1859         lock_page(page);
1860         if (!page->mapping) {
1861                 unlock_page(page);
1862                 page_cache_release(page);
1863                 goto retry;
1864         }
1865         if (PageUptodate(page)) {
1866                 unlock_page(page);
1867                 goto out;
1868         }
1869         err = filler(data, page);
1870         if (err < 0) {
1871                 page_cache_release(page);
1872                 return ERR_PTR(err);
1873         }
1874 out:
1875         mark_page_accessed(page);
1876         return page;
1877 }
1878
1879 /**
1880  * read_cache_page_async - read into page cache, fill it if needed
1881  * @mapping:    the page's address_space
1882  * @index:      the page index
1883  * @filler:     function to perform the read
1884  * @data:       first arg to filler(data, page) function, often left as NULL
1885  *
1886  * Same as read_cache_page, but don't wait for page to become unlocked
1887  * after submitting it to the filler.
1888  *
1889  * Read into the page cache. If a page already exists, and PageUptodate() is
1890  * not set, try to fill the page but don't wait for it to become unlocked.
1891  *
1892  * If the page does not get brought uptodate, return -EIO.
1893  */
1894 struct page *read_cache_page_async(struct address_space *mapping,
1895                                 pgoff_t index,
1896                                 int (*filler)(void *, struct page *),
1897                                 void *data)
1898 {
1899         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1900 }
1901 EXPORT_SYMBOL(read_cache_page_async);
1902
1903 static struct page *wait_on_page_read(struct page *page)
1904 {
1905         if (!IS_ERR(page)) {
1906                 wait_on_page_locked(page);
1907                 if (!PageUptodate(page)) {
1908                         page_cache_release(page);
1909                         page = ERR_PTR(-EIO);
1910                 }
1911         }
1912         return page;
1913 }
1914
1915 /**
1916  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1917  * @mapping:    the page's address_space
1918  * @index:      the page index
1919  * @gfp:        the page allocator flags to use if allocating
1920  *
1921  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1922  * any new page allocations done using the specified allocation flags.
1923  *
1924  * If the page does not get brought uptodate, return -EIO.
1925  */
1926 struct page *read_cache_page_gfp(struct address_space *mapping,
1927                                 pgoff_t index,
1928                                 gfp_t gfp)
1929 {
1930         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1931
1932         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1933 }
1934 EXPORT_SYMBOL(read_cache_page_gfp);
1935
1936 /**
1937  * read_cache_page - read into page cache, fill it if needed
1938  * @mapping:    the page's address_space
1939  * @index:      the page index
1940  * @filler:     function to perform the read
1941  * @data:       first arg to filler(data, page) function, often left as NULL
1942  *
1943  * Read into the page cache. If a page already exists, and PageUptodate() is
1944  * not set, try to fill the page then wait for it to become unlocked.
1945  *
1946  * If the page does not get brought uptodate, return -EIO.
1947  */
1948 struct page *read_cache_page(struct address_space *mapping,
1949                                 pgoff_t index,
1950                                 int (*filler)(void *, struct page *),
1951                                 void *data)
1952 {
1953         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1954 }
1955 EXPORT_SYMBOL(read_cache_page);
1956
1957 /*
1958  * The logic we want is
1959  *
1960  *      if suid or (sgid and xgrp)
1961  *              remove privs
1962  */
1963 int should_remove_suid(struct dentry *dentry)
1964 {
1965         mode_t mode = dentry->d_inode->i_mode;
1966         int kill = 0;
1967
1968         /* suid always must be killed */
1969         if (unlikely(mode & S_ISUID))
1970                 kill = ATTR_KILL_SUID;
1971
1972         /*
1973          * sgid without any exec bits is just a mandatory locking mark; leave
1974          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1975          */
1976         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1977                 kill |= ATTR_KILL_SGID;
1978
1979         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1980                 return kill;
1981
1982         return 0;
1983 }
1984 EXPORT_SYMBOL(should_remove_suid);
1985
1986 static int __remove_suid(struct dentry *dentry, int kill)
1987 {
1988         struct iattr newattrs;
1989
1990         newattrs.ia_valid = ATTR_FORCE | kill;
1991         return notify_change(dentry, &newattrs);
1992 }
1993
1994 int file_remove_suid(struct file *file)
1995 {
1996         struct dentry *dentry = file->f_path.dentry;
1997         struct inode *inode = dentry->d_inode;
1998         int killsuid;
1999         int killpriv;
2000         int error = 0;
2001
2002         /* Fast path for nothing security related */
2003         if (IS_NOSEC(inode))
2004                 return 0;
2005
2006         killsuid = should_remove_suid(dentry);
2007         killpriv = security_inode_need_killpriv(dentry);
2008
2009         if (killpriv < 0)
2010                 return killpriv;
2011         if (killpriv)
2012                 error = security_inode_killpriv(dentry);
2013         if (!error && killsuid)
2014                 error = __remove_suid(dentry, killsuid);
2015         if (!error)
2016                 inode_has_no_xattr(inode);
2017
2018         return error;
2019 }
2020 EXPORT_SYMBOL(file_remove_suid);
2021
2022 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
2023                         const struct iovec *iov, size_t base, size_t bytes)
2024 {
2025         size_t copied = 0, left = 0;
2026
2027         while (bytes) {
2028                 char __user *buf = iov->iov_base + base;
2029                 int copy = min(bytes, iov->iov_len - base);
2030
2031                 base = 0;
2032                 left = __copy_from_user_inatomic(vaddr, buf, copy);
2033                 copied += copy;
2034                 bytes -= copy;
2035                 vaddr += copy;
2036                 iov++;
2037
2038                 if (unlikely(left))
2039                         break;
2040         }
2041         return copied - left;
2042 }
2043
2044 /*
2045  * Copy as much as we can into the page and return the number of bytes which
2046  * were successfully copied.  If a fault is encountered then return the number of
2047  * bytes which were copied.
2048  */
2049 size_t iov_iter_copy_from_user_atomic(struct page *page,
2050                 struct iov_iter *i, unsigned long offset, size_t bytes)
2051 {
2052         char *kaddr;
2053         size_t copied;
2054
2055         BUG_ON(!in_atomic());
2056         kaddr = kmap_atomic(page, KM_USER0);
2057         if (likely(i->nr_segs == 1)) {
2058                 int left;
2059                 char __user *buf = i->iov->iov_base + i->iov_offset;
2060                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
2061                 copied = bytes - left;
2062         } else {
2063                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2064                                                 i->iov, i->iov_offset, bytes);
2065         }
2066         kunmap_atomic(kaddr, KM_USER0);
2067
2068         return copied;
2069 }
2070 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2071
2072 /*
2073  * This has the same sideeffects and return value as
2074  * iov_iter_copy_from_user_atomic().
2075  * The difference is that it attempts to resolve faults.
2076  * Page must not be locked.
2077  */
2078 size_t iov_iter_copy_from_user(struct page *page,
2079                 struct iov_iter *i, unsigned long offset, size_t bytes)
2080 {
2081         char *kaddr;
2082         size_t copied;
2083
2084         kaddr = kmap(page);
2085         if (likely(i->nr_segs == 1)) {
2086                 int left;
2087                 char __user *buf = i->iov->iov_base + i->iov_offset;
2088                 left = __copy_from_user(kaddr + offset, buf, bytes);
2089                 copied = bytes - left;
2090         } else {
2091                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2092                                                 i->iov, i->iov_offset, bytes);
2093         }
2094         kunmap(page);
2095         return copied;
2096 }
2097 EXPORT_SYMBOL(iov_iter_copy_from_user);
2098
2099 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2100 {
2101         BUG_ON(i->count < bytes);
2102
2103         if (likely(i->nr_segs == 1)) {
2104                 i->iov_offset += bytes;
2105                 i->count -= bytes;
2106         } else {
2107                 const struct iovec *iov = i->iov;
2108                 size_t base = i->iov_offset;
2109                 unsigned long nr_segs = i->nr_segs;
2110
2111                 /*
2112                  * The !iov->iov_len check ensures we skip over unlikely
2113                  * zero-length segments (without overruning the iovec).
2114                  */
2115                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2116                         int copy;
2117
2118                         copy = min(bytes, iov->iov_len - base);
2119                         BUG_ON(!i->count || i->count < copy);
2120                         i->count -= copy;
2121                         bytes -= copy;
2122                         base += copy;
2123                         if (iov->iov_len == base) {
2124                                 iov++;
2125                                 nr_segs--;
2126                                 base = 0;
2127                         }
2128                 }
2129                 i->iov = iov;
2130                 i->iov_offset = base;
2131                 i->nr_segs = nr_segs;
2132         }
2133 }
2134 EXPORT_SYMBOL(iov_iter_advance);
2135
2136 /*
2137  * Fault in the first iovec of the given iov_iter, to a maximum length
2138  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2139  * accessed (ie. because it is an invalid address).
2140  *
2141  * writev-intensive code may want this to prefault several iovecs -- that
2142  * would be possible (callers must not rely on the fact that _only_ the
2143  * first iovec will be faulted with the current implementation).
2144  */
2145 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2146 {
2147         char __user *buf = i->iov->iov_base + i->iov_offset;
2148         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2149         return fault_in_pages_readable(buf, bytes);
2150 }
2151 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2152
2153 /*
2154  * Return the count of just the current iov_iter segment.
2155  */
2156 size_t iov_iter_single_seg_count(struct iov_iter *i)
2157 {
2158         const struct iovec *iov = i->iov;
2159         if (i->nr_segs == 1)
2160                 return i->count;
2161         else
2162                 return min(i->count, iov->iov_len - i->iov_offset);
2163 }
2164 EXPORT_SYMBOL(iov_iter_single_seg_count);
2165
2166 /*
2167  * Performs necessary checks before doing a write
2168  *
2169  * Can adjust writing position or amount of bytes to write.
2170  * Returns appropriate error code that caller should return or
2171  * zero in case that write should be allowed.
2172  */
2173 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2174 {
2175         struct inode *inode = file->f_mapping->host;
2176         unsigned long limit = rlimit(RLIMIT_FSIZE);
2177
2178         if (unlikely(*pos < 0))
2179                 return -EINVAL;
2180
2181         if (!isblk) {
2182                 /* FIXME: this is for backwards compatibility with 2.4 */
2183                 if (file->f_flags & O_APPEND)
2184                         *pos = i_size_read(inode);
2185
2186                 if (limit != RLIM_INFINITY) {
2187                         if (*pos >= limit) {
2188                                 send_sig(SIGXFSZ, current, 0);
2189                                 return -EFBIG;
2190                         }
2191                         if (*count > limit - (typeof(limit))*pos) {
2192                                 *count = limit - (typeof(limit))*pos;
2193                         }
2194                 }
2195         }
2196
2197         /*
2198          * LFS rule
2199          */
2200         if (unlikely(*pos + *count > MAX_NON_LFS &&
2201                                 !(file->f_flags & O_LARGEFILE))) {
2202                 if (*pos >= MAX_NON_LFS) {
2203                         return -EFBIG;
2204                 }
2205                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2206                         *count = MAX_NON_LFS - (unsigned long)*pos;
2207                 }
2208         }
2209
2210         /*
2211          * Are we about to exceed the fs block limit ?
2212          *
2213          * If we have written data it becomes a short write.  If we have
2214          * exceeded without writing data we send a signal and return EFBIG.
2215          * Linus frestrict idea will clean these up nicely..
2216          */
2217         if (likely(!isblk)) {
2218                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2219                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2220                                 return -EFBIG;
2221                         }
2222                         /* zero-length writes at ->s_maxbytes are OK */
2223                 }
2224
2225                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2226                         *count = inode->i_sb->s_maxbytes - *pos;
2227         } else {
2228 #ifdef CONFIG_BLOCK
2229                 loff_t isize;
2230                 if (bdev_read_only(I_BDEV(inode)))
2231                         return -EPERM;
2232                 isize = i_size_read(inode);
2233                 if (*pos >= isize) {
2234                         if (*count || *pos > isize)
2235                                 return -ENOSPC;
2236                 }
2237
2238                 if (*pos + *count > isize)
2239                         *count = isize - *pos;
2240 #else
2241                 return -EPERM;
2242 #endif
2243         }
2244         return 0;
2245 }
2246 EXPORT_SYMBOL(generic_write_checks);
2247
2248 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2249                                 loff_t pos, unsigned len, unsigned flags,
2250                                 struct page **pagep, void **fsdata)
2251 {
2252         const struct address_space_operations *aops = mapping->a_ops;
2253
2254         return aops->write_begin(file, mapping, pos, len, flags,
2255                                                         pagep, fsdata);
2256 }
2257 EXPORT_SYMBOL(pagecache_write_begin);
2258
2259 int pagecache_write_end(struct file *file, struct address_space *mapping,
2260                                 loff_t pos, unsigned len, unsigned copied,
2261                                 struct page *page, void *fsdata)
2262 {
2263         const struct address_space_operations *aops = mapping->a_ops;
2264
2265         mark_page_accessed(page);
2266         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2267 }
2268 EXPORT_SYMBOL(pagecache_write_end);
2269
2270 ssize_t
2271 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2272                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2273                 size_t count, size_t ocount)
2274 {
2275         struct file     *file = iocb->ki_filp;
2276         struct address_space *mapping = file->f_mapping;
2277         struct inode    *inode = mapping->host;
2278         ssize_t         written;
2279         size_t          write_len;
2280         pgoff_t         end;
2281
2282         if (count != ocount)
2283                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2284
2285         write_len = iov_length(iov, *nr_segs);
2286         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2287
2288         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2289         if (written)
2290                 goto out;
2291
2292         /*
2293          * After a write we want buffered reads to be sure to go to disk to get
2294          * the new data.  We invalidate clean cached page from the region we're
2295          * about to write.  We do this *before* the write so that we can return
2296          * without clobbering -EIOCBQUEUED from ->direct_IO().
2297          */
2298         if (mapping->nrpages) {
2299                 written = invalidate_inode_pages2_range(mapping,
2300                                         pos >> PAGE_CACHE_SHIFT, end);
2301                 /*
2302                  * If a page can not be invalidated, return 0 to fall back
2303                  * to buffered write.
2304                  */
2305                 if (written) {
2306                         if (written == -EBUSY)
2307                                 return 0;
2308                         goto out;
2309                 }
2310         }
2311
2312         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2313
2314         /*
2315          * Finally, try again to invalidate clean pages which might have been
2316          * cached by non-direct readahead, or faulted in by get_user_pages()
2317          * if the source of the write was an mmap'ed region of the file
2318          * we're writing.  Either one is a pretty crazy thing to do,
2319          * so we don't support it 100%.  If this invalidation
2320          * fails, tough, the write still worked...
2321          */
2322         if (mapping->nrpages) {
2323                 invalidate_inode_pages2_range(mapping,
2324                                               pos >> PAGE_CACHE_SHIFT, end);
2325         }
2326
2327         if (written > 0) {
2328                 pos += written;
2329                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2330                         i_size_write(inode, pos);
2331                         mark_inode_dirty(inode);
2332                 }
2333                 *ppos = pos;
2334         }
2335 out:
2336         return written;
2337 }
2338 EXPORT_SYMBOL(generic_file_direct_write);
2339
2340 /*
2341  * Find or create a page at the given pagecache position. Return the locked
2342  * page. This function is specifically for buffered writes.
2343  */
2344 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2345                                         pgoff_t index, unsigned flags)
2346 {
2347         int status;
2348         struct page *page;
2349         gfp_t gfp_notmask = 0;
2350         if (flags & AOP_FLAG_NOFS)
2351                 gfp_notmask = __GFP_FS;
2352 repeat:
2353         page = find_lock_page(mapping, index);
2354         if (page)
2355                 goto found;
2356
2357         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2358         if (!page)
2359                 return NULL;
2360         status = add_to_page_cache_lru(page, mapping, index,
2361                                                 GFP_KERNEL & ~gfp_notmask);
2362         if (unlikely(status)) {
2363                 page_cache_release(page);
2364                 if (status == -EEXIST)
2365                         goto repeat;
2366                 return NULL;
2367         }
2368 found:
2369         wait_on_page_writeback(page);
2370         return page;
2371 }
2372 EXPORT_SYMBOL(grab_cache_page_write_begin);
2373
2374 static ssize_t generic_perform_write(struct file *file,
2375                                 struct iov_iter *i, loff_t pos)
2376 {
2377         struct address_space *mapping = file->f_mapping;
2378         const struct address_space_operations *a_ops = mapping->a_ops;
2379         long status = 0;
2380         ssize_t written = 0;
2381         unsigned int flags = 0;
2382
2383         /*
2384          * Copies from kernel address space cannot fail (NFSD is a big user).
2385          */
2386         if (segment_eq(get_fs(), KERNEL_DS))
2387                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2388
2389         do {
2390                 struct page *page;
2391                 unsigned long offset;   /* Offset into pagecache page */
2392                 unsigned long bytes;    /* Bytes to write to page */
2393                 size_t copied;          /* Bytes copied from user */
2394                 void *fsdata;
2395
2396                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2397                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2398                                                 iov_iter_count(i));
2399
2400 again:
2401                 /*
2402                  * Bring in the user page that we will copy from _first_.
2403                  * Otherwise there's a nasty deadlock on copying from the
2404                  * same page as we're writing to, without it being marked
2405                  * up-to-date.
2406                  *
2407                  * Not only is this an optimisation, but it is also required
2408                  * to check that the address is actually valid, when atomic
2409                  * usercopies are used, below.
2410                  */
2411                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2412                         status = -EFAULT;
2413                         break;
2414                 }
2415
2416                 if (fatal_signal_pending(current)) {
2417                         status = -EINTR;
2418                         break;
2419                 }
2420
2421                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2422                                                 &page, &fsdata);
2423                 if (unlikely(status))
2424                         break;
2425
2426                 if (mapping_writably_mapped(mapping))
2427                         flush_dcache_page(page);
2428
2429                 pagefault_disable();
2430                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2431                 pagefault_enable();
2432                 flush_dcache_page(page);
2433
2434                 mark_page_accessed(page);
2435                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2436                                                 page, fsdata);
2437                 if (unlikely(status < 0))
2438                         break;
2439                 copied = status;
2440
2441                 cond_resched();
2442
2443                 iov_iter_advance(i, copied);
2444                 if (unlikely(copied == 0)) {
2445                         /*
2446                          * If we were unable to copy any data at all, we must
2447                          * fall back to a single segment length write.
2448                          *
2449                          * If we didn't fallback here, we could livelock
2450                          * because not all segments in the iov can be copied at
2451                          * once without a pagefault.
2452                          */
2453                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2454                                                 iov_iter_single_seg_count(i));
2455                         goto again;
2456                 }
2457                 pos += copied;
2458                 written += copied;
2459
2460                 balance_dirty_pages_ratelimited(mapping);
2461         } while (iov_iter_count(i));
2462
2463         return written ? written : status;
2464 }
2465
2466 ssize_t
2467 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2468                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2469                 size_t count, ssize_t written)
2470 {
2471         struct file *file = iocb->ki_filp;
2472         ssize_t status;
2473         struct iov_iter i;
2474
2475         iov_iter_init(&i, iov, nr_segs, count, written);
2476         status = generic_perform_write(file, &i, pos);
2477
2478         if (likely(status >= 0)) {
2479                 written += status;
2480                 *ppos = pos + status;
2481         }
2482         
2483         return written ? written : status;
2484 }
2485 EXPORT_SYMBOL(generic_file_buffered_write);
2486
2487 /**
2488  * __generic_file_aio_write - write data to a file
2489  * @iocb:       IO state structure (file, offset, etc.)
2490  * @iov:        vector with data to write
2491  * @nr_segs:    number of segments in the vector
2492  * @ppos:       position where to write
2493  *
2494  * This function does all the work needed for actually writing data to a
2495  * file. It does all basic checks, removes SUID from the file, updates
2496  * modification times and calls proper subroutines depending on whether we
2497  * do direct IO or a standard buffered write.
2498  *
2499  * It expects i_mutex to be grabbed unless we work on a block device or similar
2500  * object which does not need locking at all.
2501  *
2502  * This function does *not* take care of syncing data in case of O_SYNC write.
2503  * A caller has to handle it. This is mainly due to the fact that we want to
2504  * avoid syncing under i_mutex.
2505  */
2506 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2507                                  unsigned long nr_segs, loff_t *ppos)
2508 {
2509         struct file *file = iocb->ki_filp;
2510         struct address_space * mapping = file->f_mapping;
2511         size_t ocount;          /* original count */
2512         size_t count;           /* after file limit checks */
2513         struct inode    *inode = mapping->host;
2514         loff_t          pos;
2515         ssize_t         written;
2516         ssize_t         err;
2517
2518         ocount = 0;
2519         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2520         if (err)
2521                 return err;
2522
2523         count = ocount;
2524         pos = *ppos;
2525
2526         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2527
2528         /* We can write back this queue in page reclaim */
2529         current->backing_dev_info = mapping->backing_dev_info;
2530         written = 0;
2531
2532         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2533         if (err)
2534                 goto out;
2535
2536         if (count == 0)
2537                 goto out;
2538
2539         err = file_remove_suid(file);
2540         if (err)
2541                 goto out;
2542
2543         file_update_time(file);
2544
2545         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2546         if (unlikely(file->f_flags & O_DIRECT)) {
2547                 loff_t endbyte;
2548                 ssize_t written_buffered;
2549
2550                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2551                                                         ppos, count, ocount);
2552                 if (written < 0 || written == count)
2553                         goto out;
2554                 /*
2555                  * direct-io write to a hole: fall through to buffered I/O
2556                  * for completing the rest of the request.
2557                  */
2558                 pos += written;
2559                 count -= written;
2560                 written_buffered = generic_file_buffered_write(iocb, iov,
2561                                                 nr_segs, pos, ppos, count,
2562                                                 written);
2563                 /*
2564                  * If generic_file_buffered_write() retuned a synchronous error
2565                  * then we want to return the number of bytes which were
2566                  * direct-written, or the error code if that was zero.  Note
2567                  * that this differs from normal direct-io semantics, which
2568                  * will return -EFOO even if some bytes were written.
2569                  */
2570                 if (written_buffered < 0) {
2571                         err = written_buffered;
2572                         goto out;
2573                 }
2574
2575                 /*
2576                  * We need to ensure that the page cache pages are written to
2577                  * disk and invalidated to preserve the expected O_DIRECT
2578                  * semantics.
2579                  */
2580                 endbyte = pos + written_buffered - written - 1;
2581                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2582                 if (err == 0) {
2583                         written = written_buffered;
2584                         invalidate_mapping_pages(mapping,
2585                                                  pos >> PAGE_CACHE_SHIFT,
2586                                                  endbyte >> PAGE_CACHE_SHIFT);
2587                 } else {
2588                         /*
2589                          * We don't know how much we wrote, so just return
2590                          * the number of bytes which were direct-written
2591                          */
2592                 }
2593         } else {
2594                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2595                                 pos, ppos, count, written);
2596         }
2597 out:
2598         current->backing_dev_info = NULL;
2599         return written ? written : err;
2600 }
2601 EXPORT_SYMBOL(__generic_file_aio_write);
2602
2603 /**
2604  * generic_file_aio_write - write data to a file
2605  * @iocb:       IO state structure
2606  * @iov:        vector with data to write
2607  * @nr_segs:    number of segments in the vector
2608  * @pos:        position in file where to write
2609  *
2610  * This is a wrapper around __generic_file_aio_write() to be used by most
2611  * filesystems. It takes care of syncing the file in case of O_SYNC file
2612  * and acquires i_mutex as needed.
2613  */
2614 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2615                 unsigned long nr_segs, loff_t pos)
2616 {
2617         struct file *file = iocb->ki_filp;
2618         struct inode *inode = file->f_mapping->host;
2619         struct blk_plug plug;
2620         ssize_t ret;
2621
2622         BUG_ON(iocb->ki_pos != pos);
2623
2624         mutex_lock(&inode->i_mutex);
2625         blk_start_plug(&plug);
2626         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2627         mutex_unlock(&inode->i_mutex);
2628
2629         if (ret > 0 || ret == -EIOCBQUEUED) {
2630                 ssize_t err;
2631
2632                 err = generic_write_sync(file, pos, ret);
2633                 if (err < 0 && ret > 0)
2634                         ret = err;
2635         }
2636         blk_finish_plug(&plug);
2637         return ret;
2638 }
2639 EXPORT_SYMBOL(generic_file_aio_write);
2640
2641 /**
2642  * try_to_release_page() - release old fs-specific metadata on a page
2643  *
2644  * @page: the page which the kernel is trying to free
2645  * @gfp_mask: memory allocation flags (and I/O mode)
2646  *
2647  * The address_space is to try to release any data against the page
2648  * (presumably at page->private).  If the release was successful, return `1'.
2649  * Otherwise return zero.
2650  *
2651  * This may also be called if PG_fscache is set on a page, indicating that the
2652  * page is known to the local caching routines.
2653  *
2654  * The @gfp_mask argument specifies whether I/O may be performed to release
2655  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2656  *
2657  */
2658 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2659 {
2660         struct address_space * const mapping = page->mapping;
2661
2662         BUG_ON(!PageLocked(page));
2663         if (PageWriteback(page))
2664                 return 0;
2665
2666         if (mapping && mapping->a_ops->releasepage)
2667                 return mapping->a_ops->releasepage(page, gfp_mask);
2668         return try_to_free_buffers(page);
2669 }
2670
2671 EXPORT_SYMBOL(try_to_release_page);