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