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