do_generic_file_read: s/EINTR/EIO/ if lock_page_killable() fails
[deliverable/linux.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/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.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 "internal.h"
37
38 /*
39 * FIXME: remove all knowledge of the buffer layer from the core VM
40 */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
45
46 /*
47 * Shared mappings implemented 30.11.1994. It's not fully working yet,
48 * though.
49 *
50 * Shared mappings now work. 15.8.1995 Bruno.
51 *
52 * finished 'unifying' the page and buffer cache and SMP-threaded the
53 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54 *
55 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56 */
57
58 /*
59 * Lock ordering:
60 *
61 * ->i_mmap_lock (vmtruncate)
62 * ->private_lock (__free_pte->__set_page_dirty_buffers)
63 * ->swap_lock (exclusive_swap_page, others)
64 * ->mapping->tree_lock
65 *
66 * ->i_mutex
67 * ->i_mmap_lock (truncate->unmap_mapping_range)
68 *
69 * ->mmap_sem
70 * ->i_mmap_lock
71 * ->page_table_lock or pte_lock (various, mainly in memory.c)
72 * ->mapping->tree_lock (arch-dependent flush_dcache_mmap_lock)
73 *
74 * ->mmap_sem
75 * ->lock_page (access_process_vm)
76 *
77 * ->i_mutex (generic_file_buffered_write)
78 * ->mmap_sem (fault_in_pages_readable->do_page_fault)
79 *
80 * ->i_mutex
81 * ->i_alloc_sem (various)
82 *
83 * ->inode_lock
84 * ->sb_lock (fs/fs-writeback.c)
85 * ->mapping->tree_lock (__sync_single_inode)
86 *
87 * ->i_mmap_lock
88 * ->anon_vma.lock (vma_adjust)
89 *
90 * ->anon_vma.lock
91 * ->page_table_lock or pte_lock (anon_vma_prepare and various)
92 *
93 * ->page_table_lock or pte_lock
94 * ->swap_lock (try_to_unmap_one)
95 * ->private_lock (try_to_unmap_one)
96 * ->tree_lock (try_to_unmap_one)
97 * ->zone.lru_lock (follow_page->mark_page_accessed)
98 * ->zone.lru_lock (check_pte_range->isolate_lru_page)
99 * ->private_lock (page_remove_rmap->set_page_dirty)
100 * ->tree_lock (page_remove_rmap->set_page_dirty)
101 * ->inode_lock (page_remove_rmap->set_page_dirty)
102 * ->inode_lock (zap_pte_range->set_page_dirty)
103 * ->private_lock (zap_pte_range->__set_page_dirty_buffers)
104 *
105 * ->task->proc_lock
106 * ->dcache_lock (proc_pid_lookup)
107 */
108
109 /*
110 * Remove a page from the page cache and free it. Caller has to make
111 * sure the page is locked and that nobody else uses it - or that usage
112 * is safe. The caller must hold the mapping's tree_lock.
113 */
114 void __remove_from_page_cache(struct page *page)
115 {
116 struct address_space *mapping = page->mapping;
117
118 mem_cgroup_uncharge_cache_page(page);
119 radix_tree_delete(&mapping->page_tree, page->index);
120 page->mapping = NULL;
121 mapping->nrpages--;
122 __dec_zone_page_state(page, NR_FILE_PAGES);
123 BUG_ON(page_mapped(page));
124
125 /*
126 * Some filesystems seem to re-dirty the page even after
127 * the VM has canceled the dirty bit (eg ext3 journaling).
128 *
129 * Fix it up by doing a final dirty accounting check after
130 * having removed the page entirely.
131 */
132 if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
133 dec_zone_page_state(page, NR_FILE_DIRTY);
134 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
135 }
136 }
137
138 void remove_from_page_cache(struct page *page)
139 {
140 struct address_space *mapping = page->mapping;
141
142 BUG_ON(!PageLocked(page));
143
144 spin_lock_irq(&mapping->tree_lock);
145 __remove_from_page_cache(page);
146 spin_unlock_irq(&mapping->tree_lock);
147 }
148
149 static int sync_page(void *word)
150 {
151 struct address_space *mapping;
152 struct page *page;
153
154 page = container_of((unsigned long *)word, struct page, flags);
155
156 /*
157 * page_mapping() is being called without PG_locked held.
158 * Some knowledge of the state and use of the page is used to
159 * reduce the requirements down to a memory barrier.
160 * The danger here is of a stale page_mapping() return value
161 * indicating a struct address_space different from the one it's
162 * associated with when it is associated with one.
163 * After smp_mb(), it's either the correct page_mapping() for
164 * the page, or an old page_mapping() and the page's own
165 * page_mapping() has gone NULL.
166 * The ->sync_page() address_space operation must tolerate
167 * page_mapping() going NULL. By an amazing coincidence,
168 * this comes about because none of the users of the page
169 * in the ->sync_page() methods make essential use of the
170 * page_mapping(), merely passing the page down to the backing
171 * device's unplug functions when it's non-NULL, which in turn
172 * ignore it for all cases but swap, where only page_private(page) is
173 * of interest. When page_mapping() does go NULL, the entire
174 * call stack gracefully ignores the page and returns.
175 * -- wli
176 */
177 smp_mb();
178 mapping = page_mapping(page);
179 if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
180 mapping->a_ops->sync_page(page);
181 io_schedule();
182 return 0;
183 }
184
185 static int sync_page_killable(void *word)
186 {
187 sync_page(word);
188 return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 /**
192 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
193 * @mapping: address space structure to write
194 * @start: offset in bytes where the range starts
195 * @end: offset in bytes where the range ends (inclusive)
196 * @sync_mode: enable synchronous operation
197 *
198 * Start writeback against all of a mapping's dirty pages that lie
199 * within the byte offsets <start, end> inclusive.
200 *
201 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
202 * opposed to a regular memory cleansing writeback. The difference between
203 * these two operations is that if a dirty page/buffer is encountered, it must
204 * be waited upon, and not just skipped over.
205 */
206 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
207 loff_t end, int sync_mode)
208 {
209 int ret;
210 struct writeback_control wbc = {
211 .sync_mode = sync_mode,
212 .nr_to_write = mapping->nrpages * 2,
213 .range_start = start,
214 .range_end = end,
215 };
216
217 if (!mapping_cap_writeback_dirty(mapping))
218 return 0;
219
220 ret = do_writepages(mapping, &wbc);
221 return ret;
222 }
223
224 static inline int __filemap_fdatawrite(struct address_space *mapping,
225 int sync_mode)
226 {
227 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
228 }
229
230 int filemap_fdatawrite(struct address_space *mapping)
231 {
232 return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
233 }
234 EXPORT_SYMBOL(filemap_fdatawrite);
235
236 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
237 loff_t end)
238 {
239 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite_range);
242
243 /**
244 * filemap_flush - mostly a non-blocking flush
245 * @mapping: target address_space
246 *
247 * This is a mostly non-blocking flush. Not suitable for data-integrity
248 * purposes - I/O may not be started against all dirty pages.
249 */
250 int filemap_flush(struct address_space *mapping)
251 {
252 return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
253 }
254 EXPORT_SYMBOL(filemap_flush);
255
256 /**
257 * wait_on_page_writeback_range - wait for writeback to complete
258 * @mapping: target address_space
259 * @start: beginning page index
260 * @end: ending page index
261 *
262 * Wait for writeback to complete against pages indexed by start->end
263 * inclusive
264 */
265 int wait_on_page_writeback_range(struct address_space *mapping,
266 pgoff_t start, pgoff_t end)
267 {
268 struct pagevec pvec;
269 int nr_pages;
270 int ret = 0;
271 pgoff_t index;
272
273 if (end < start)
274 return 0;
275
276 pagevec_init(&pvec, 0);
277 index = start;
278 while ((index <= end) &&
279 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
280 PAGECACHE_TAG_WRITEBACK,
281 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
282 unsigned i;
283
284 for (i = 0; i < nr_pages; i++) {
285 struct page *page = pvec.pages[i];
286
287 /* until radix tree lookup accepts end_index */
288 if (page->index > end)
289 continue;
290
291 wait_on_page_writeback(page);
292 if (PageError(page))
293 ret = -EIO;
294 }
295 pagevec_release(&pvec);
296 cond_resched();
297 }
298
299 /* Check for outstanding write errors */
300 if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
301 ret = -ENOSPC;
302 if (test_and_clear_bit(AS_EIO, &mapping->flags))
303 ret = -EIO;
304
305 return ret;
306 }
307
308 /**
309 * sync_page_range - write and wait on all pages in the passed range
310 * @inode: target inode
311 * @mapping: target address_space
312 * @pos: beginning offset in pages to write
313 * @count: number of bytes to write
314 *
315 * Write and wait upon all the pages in the passed range. This is a "data
316 * integrity" operation. It waits upon in-flight writeout before starting and
317 * waiting upon new writeout. If there was an IO error, return it.
318 *
319 * We need to re-take i_mutex during the generic_osync_inode list walk because
320 * it is otherwise livelockable.
321 */
322 int sync_page_range(struct inode *inode, struct address_space *mapping,
323 loff_t pos, loff_t count)
324 {
325 pgoff_t start = pos >> PAGE_CACHE_SHIFT;
326 pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
327 int ret;
328
329 if (!mapping_cap_writeback_dirty(mapping) || !count)
330 return 0;
331 ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
332 if (ret == 0) {
333 mutex_lock(&inode->i_mutex);
334 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
335 mutex_unlock(&inode->i_mutex);
336 }
337 if (ret == 0)
338 ret = wait_on_page_writeback_range(mapping, start, end);
339 return ret;
340 }
341 EXPORT_SYMBOL(sync_page_range);
342
343 /**
344 * sync_page_range_nolock - write & wait on all pages in the passed range without locking
345 * @inode: target inode
346 * @mapping: target address_space
347 * @pos: beginning offset in pages to write
348 * @count: number of bytes to write
349 *
350 * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
351 * as it forces O_SYNC writers to different parts of the same file
352 * to be serialised right until io completion.
353 */
354 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
355 loff_t pos, loff_t count)
356 {
357 pgoff_t start = pos >> PAGE_CACHE_SHIFT;
358 pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
359 int ret;
360
361 if (!mapping_cap_writeback_dirty(mapping) || !count)
362 return 0;
363 ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
364 if (ret == 0)
365 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
366 if (ret == 0)
367 ret = wait_on_page_writeback_range(mapping, start, end);
368 return ret;
369 }
370 EXPORT_SYMBOL(sync_page_range_nolock);
371
372 /**
373 * filemap_fdatawait - wait for all under-writeback pages to complete
374 * @mapping: address space structure to wait for
375 *
376 * Walk the list of under-writeback pages of the given address space
377 * and wait for all of them.
378 */
379 int filemap_fdatawait(struct address_space *mapping)
380 {
381 loff_t i_size = i_size_read(mapping->host);
382
383 if (i_size == 0)
384 return 0;
385
386 return wait_on_page_writeback_range(mapping, 0,
387 (i_size - 1) >> PAGE_CACHE_SHIFT);
388 }
389 EXPORT_SYMBOL(filemap_fdatawait);
390
391 int filemap_write_and_wait(struct address_space *mapping)
392 {
393 int err = 0;
394
395 if (mapping->nrpages) {
396 err = filemap_fdatawrite(mapping);
397 /*
398 * Even if the above returned error, the pages may be
399 * written partially (e.g. -ENOSPC), so we wait for it.
400 * But the -EIO is special case, it may indicate the worst
401 * thing (e.g. bug) happened, so we avoid waiting for it.
402 */
403 if (err != -EIO) {
404 int err2 = filemap_fdatawait(mapping);
405 if (!err)
406 err = err2;
407 }
408 }
409 return err;
410 }
411 EXPORT_SYMBOL(filemap_write_and_wait);
412
413 /**
414 * filemap_write_and_wait_range - write out & wait on a file range
415 * @mapping: the address_space for the pages
416 * @lstart: offset in bytes where the range starts
417 * @lend: offset in bytes where the range ends (inclusive)
418 *
419 * Write out and wait upon file offsets lstart->lend, inclusive.
420 *
421 * Note that `lend' is inclusive (describes the last byte to be written) so
422 * that this function can be used to write to the very end-of-file (end = -1).
423 */
424 int filemap_write_and_wait_range(struct address_space *mapping,
425 loff_t lstart, loff_t lend)
426 {
427 int err = 0;
428
429 if (mapping->nrpages) {
430 err = __filemap_fdatawrite_range(mapping, lstart, lend,
431 WB_SYNC_ALL);
432 /* See comment of filemap_write_and_wait() */
433 if (err != -EIO) {
434 int err2 = wait_on_page_writeback_range(mapping,
435 lstart >> PAGE_CACHE_SHIFT,
436 lend >> PAGE_CACHE_SHIFT);
437 if (!err)
438 err = err2;
439 }
440 }
441 return err;
442 }
443
444 /**
445 * add_to_page_cache_locked - add a locked page to the pagecache
446 * @page: page to add
447 * @mapping: the page's address_space
448 * @offset: page index
449 * @gfp_mask: page allocation mode
450 *
451 * This function is used to add a page to the pagecache. It must be locked.
452 * This function does not add the page to the LRU. The caller must do that.
453 */
454 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
455 pgoff_t offset, gfp_t gfp_mask)
456 {
457 int error;
458
459 VM_BUG_ON(!PageLocked(page));
460
461 error = mem_cgroup_cache_charge(page, current->mm,
462 gfp_mask & ~__GFP_HIGHMEM);
463 if (error)
464 goto out;
465
466 error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
467 if (error == 0) {
468 page_cache_get(page);
469 page->mapping = mapping;
470 page->index = offset;
471
472 spin_lock_irq(&mapping->tree_lock);
473 error = radix_tree_insert(&mapping->page_tree, offset, page);
474 if (likely(!error)) {
475 mapping->nrpages++;
476 __inc_zone_page_state(page, NR_FILE_PAGES);
477 } else {
478 page->mapping = NULL;
479 mem_cgroup_uncharge_cache_page(page);
480 page_cache_release(page);
481 }
482
483 spin_unlock_irq(&mapping->tree_lock);
484 radix_tree_preload_end();
485 } else
486 mem_cgroup_uncharge_cache_page(page);
487 out:
488 return error;
489 }
490 EXPORT_SYMBOL(add_to_page_cache_locked);
491
492 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
493 pgoff_t offset, gfp_t gfp_mask)
494 {
495 int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
496 if (ret == 0)
497 lru_cache_add(page);
498 return ret;
499 }
500
501 #ifdef CONFIG_NUMA
502 struct page *__page_cache_alloc(gfp_t gfp)
503 {
504 if (cpuset_do_page_mem_spread()) {
505 int n = cpuset_mem_spread_node();
506 return alloc_pages_node(n, gfp, 0);
507 }
508 return alloc_pages(gfp, 0);
509 }
510 EXPORT_SYMBOL(__page_cache_alloc);
511 #endif
512
513 static int __sleep_on_page_lock(void *word)
514 {
515 io_schedule();
516 return 0;
517 }
518
519 /*
520 * In order to wait for pages to become available there must be
521 * waitqueues associated with pages. By using a hash table of
522 * waitqueues where the bucket discipline is to maintain all
523 * waiters on the same queue and wake all when any of the pages
524 * become available, and for the woken contexts to check to be
525 * sure the appropriate page became available, this saves space
526 * at a cost of "thundering herd" phenomena during rare hash
527 * collisions.
528 */
529 static wait_queue_head_t *page_waitqueue(struct page *page)
530 {
531 const struct zone *zone = page_zone(page);
532
533 return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
534 }
535
536 static inline void wake_up_page(struct page *page, int bit)
537 {
538 __wake_up_bit(page_waitqueue(page), &page->flags, bit);
539 }
540
541 void wait_on_page_bit(struct page *page, int bit_nr)
542 {
543 DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
544
545 if (test_bit(bit_nr, &page->flags))
546 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
547 TASK_UNINTERRUPTIBLE);
548 }
549 EXPORT_SYMBOL(wait_on_page_bit);
550
551 /**
552 * unlock_page - unlock a locked page
553 * @page: the page
554 *
555 * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
556 * Also wakes sleepers in wait_on_page_writeback() because the wakeup
557 * mechananism between PageLocked pages and PageWriteback pages is shared.
558 * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
559 *
560 * The first mb is necessary to safely close the critical section opened by the
561 * test_and_set_bit() to lock the page; the second mb is necessary to enforce
562 * ordering between the clear_bit and the read of the waitqueue (to avoid SMP
563 * races with a parallel wait_on_page_locked()).
564 */
565 void unlock_page(struct page *page)
566 {
567 smp_mb__before_clear_bit();
568 if (!test_and_clear_bit(PG_locked, &page->flags))
569 BUG();
570 smp_mb__after_clear_bit();
571 wake_up_page(page, PG_locked);
572 }
573 EXPORT_SYMBOL(unlock_page);
574
575 /**
576 * end_page_writeback - end writeback against a page
577 * @page: the page
578 */
579 void end_page_writeback(struct page *page)
580 {
581 if (TestClearPageReclaim(page))
582 rotate_reclaimable_page(page);
583
584 if (!test_clear_page_writeback(page))
585 BUG();
586
587 smp_mb__after_clear_bit();
588 wake_up_page(page, PG_writeback);
589 }
590 EXPORT_SYMBOL(end_page_writeback);
591
592 /**
593 * __lock_page - get a lock on the page, assuming we need to sleep to get it
594 * @page: the page to lock
595 *
596 * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary. If some
597 * random driver's requestfn sets TASK_RUNNING, we could busywait. However
598 * chances are that on the second loop, the block layer's plug list is empty,
599 * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
600 */
601 void __lock_page(struct page *page)
602 {
603 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
604
605 __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
606 TASK_UNINTERRUPTIBLE);
607 }
608 EXPORT_SYMBOL(__lock_page);
609
610 int __lock_page_killable(struct page *page)
611 {
612 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
613
614 return __wait_on_bit_lock(page_waitqueue(page), &wait,
615 sync_page_killable, TASK_KILLABLE);
616 }
617
618 /**
619 * __lock_page_nosync - get a lock on the page, without calling sync_page()
620 * @page: the page to lock
621 *
622 * Variant of lock_page that does not require the caller to hold a reference
623 * on the page's mapping.
624 */
625 void __lock_page_nosync(struct page *page)
626 {
627 DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
628 __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
629 TASK_UNINTERRUPTIBLE);
630 }
631
632 /**
633 * find_get_page - find and get a page reference
634 * @mapping: the address_space to search
635 * @offset: the page index
636 *
637 * Is there a pagecache struct page at the given (mapping, offset) tuple?
638 * If yes, increment its refcount and return it; if no, return NULL.
639 */
640 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
641 {
642 void **pagep;
643 struct page *page;
644
645 rcu_read_lock();
646 repeat:
647 page = NULL;
648 pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
649 if (pagep) {
650 page = radix_tree_deref_slot(pagep);
651 if (unlikely(!page || page == RADIX_TREE_RETRY))
652 goto repeat;
653
654 if (!page_cache_get_speculative(page))
655 goto repeat;
656
657 /*
658 * Has the page moved?
659 * This is part of the lockless pagecache protocol. See
660 * include/linux/pagemap.h for details.
661 */
662 if (unlikely(page != *pagep)) {
663 page_cache_release(page);
664 goto repeat;
665 }
666 }
667 rcu_read_unlock();
668
669 return page;
670 }
671 EXPORT_SYMBOL(find_get_page);
672
673 /**
674 * find_lock_page - locate, pin and lock a pagecache page
675 * @mapping: the address_space to search
676 * @offset: the page index
677 *
678 * Locates the desired pagecache page, locks it, increments its reference
679 * count and returns its address.
680 *
681 * Returns zero if the page was not present. find_lock_page() may sleep.
682 */
683 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
684 {
685 struct page *page;
686
687 repeat:
688 page = find_get_page(mapping, offset);
689 if (page) {
690 lock_page(page);
691 /* Has the page been truncated? */
692 if (unlikely(page->mapping != mapping)) {
693 unlock_page(page);
694 page_cache_release(page);
695 goto repeat;
696 }
697 VM_BUG_ON(page->index != offset);
698 }
699 return page;
700 }
701 EXPORT_SYMBOL(find_lock_page);
702
703 /**
704 * find_or_create_page - locate or add a pagecache page
705 * @mapping: the page's address_space
706 * @index: the page's index into the mapping
707 * @gfp_mask: page allocation mode
708 *
709 * Locates a page in the pagecache. If the page is not present, a new page
710 * is allocated using @gfp_mask and is added to the pagecache and to the VM's
711 * LRU list. The returned page is locked and has its reference count
712 * incremented.
713 *
714 * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
715 * allocation!
716 *
717 * find_or_create_page() returns the desired page's address, or zero on
718 * memory exhaustion.
719 */
720 struct page *find_or_create_page(struct address_space *mapping,
721 pgoff_t index, gfp_t gfp_mask)
722 {
723 struct page *page;
724 int err;
725 repeat:
726 page = find_lock_page(mapping, index);
727 if (!page) {
728 page = __page_cache_alloc(gfp_mask);
729 if (!page)
730 return NULL;
731 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
732 if (unlikely(err)) {
733 page_cache_release(page);
734 page = NULL;
735 if (err == -EEXIST)
736 goto repeat;
737 }
738 }
739 return page;
740 }
741 EXPORT_SYMBOL(find_or_create_page);
742
743 /**
744 * find_get_pages - gang pagecache lookup
745 * @mapping: The address_space to search
746 * @start: The starting page index
747 * @nr_pages: The maximum number of pages
748 * @pages: Where the resulting pages are placed
749 *
750 * find_get_pages() will search for and return a group of up to
751 * @nr_pages pages in the mapping. The pages are placed at @pages.
752 * find_get_pages() takes a reference against the returned pages.
753 *
754 * The search returns a group of mapping-contiguous pages with ascending
755 * indexes. There may be holes in the indices due to not-present pages.
756 *
757 * find_get_pages() returns the number of pages which were found.
758 */
759 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
760 unsigned int nr_pages, struct page **pages)
761 {
762 unsigned int i;
763 unsigned int ret;
764 unsigned int nr_found;
765
766 rcu_read_lock();
767 restart:
768 nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
769 (void ***)pages, start, nr_pages);
770 ret = 0;
771 for (i = 0; i < nr_found; i++) {
772 struct page *page;
773 repeat:
774 page = radix_tree_deref_slot((void **)pages[i]);
775 if (unlikely(!page))
776 continue;
777 /*
778 * this can only trigger if nr_found == 1, making livelock
779 * a non issue.
780 */
781 if (unlikely(page == RADIX_TREE_RETRY))
782 goto restart;
783
784 if (!page_cache_get_speculative(page))
785 goto repeat;
786
787 /* Has the page moved? */
788 if (unlikely(page != *((void **)pages[i]))) {
789 page_cache_release(page);
790 goto repeat;
791 }
792
793 pages[ret] = page;
794 ret++;
795 }
796 rcu_read_unlock();
797 return ret;
798 }
799
800 /**
801 * find_get_pages_contig - gang contiguous pagecache lookup
802 * @mapping: The address_space to search
803 * @index: The starting page index
804 * @nr_pages: The maximum number of pages
805 * @pages: Where the resulting pages are placed
806 *
807 * find_get_pages_contig() works exactly like find_get_pages(), except
808 * that the returned number of pages are guaranteed to be contiguous.
809 *
810 * find_get_pages_contig() returns the number of pages which were found.
811 */
812 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
813 unsigned int nr_pages, struct page **pages)
814 {
815 unsigned int i;
816 unsigned int ret;
817 unsigned int nr_found;
818
819 rcu_read_lock();
820 restart:
821 nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
822 (void ***)pages, index, nr_pages);
823 ret = 0;
824 for (i = 0; i < nr_found; i++) {
825 struct page *page;
826 repeat:
827 page = radix_tree_deref_slot((void **)pages[i]);
828 if (unlikely(!page))
829 continue;
830 /*
831 * this can only trigger if nr_found == 1, making livelock
832 * a non issue.
833 */
834 if (unlikely(page == RADIX_TREE_RETRY))
835 goto restart;
836
837 if (page->mapping == NULL || page->index != index)
838 break;
839
840 if (!page_cache_get_speculative(page))
841 goto repeat;
842
843 /* Has the page moved? */
844 if (unlikely(page != *((void **)pages[i]))) {
845 page_cache_release(page);
846 goto repeat;
847 }
848
849 pages[ret] = page;
850 ret++;
851 index++;
852 }
853 rcu_read_unlock();
854 return ret;
855 }
856 EXPORT_SYMBOL(find_get_pages_contig);
857
858 /**
859 * find_get_pages_tag - find and return pages that match @tag
860 * @mapping: the address_space to search
861 * @index: the starting page index
862 * @tag: the tag index
863 * @nr_pages: the maximum number of pages
864 * @pages: where the resulting pages are placed
865 *
866 * Like find_get_pages, except we only return pages which are tagged with
867 * @tag. We update @index to index the next page for the traversal.
868 */
869 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
870 int tag, 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_tag_slot(&mapping->page_tree,
879 (void ***)pages, *index, nr_pages, tag);
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 * this can only trigger if nr_found == 1, making livelock
889 * a non issue.
890 */
891 if (unlikely(page == RADIX_TREE_RETRY))
892 goto restart;
893
894 if (!page_cache_get_speculative(page))
895 goto repeat;
896
897 /* Has the page moved? */
898 if (unlikely(page != *((void **)pages[i]))) {
899 page_cache_release(page);
900 goto repeat;
901 }
902
903 pages[ret] = page;
904 ret++;
905 }
906 rcu_read_unlock();
907
908 if (ret)
909 *index = pages[ret - 1]->index + 1;
910
911 return ret;
912 }
913 EXPORT_SYMBOL(find_get_pages_tag);
914
915 /**
916 * grab_cache_page_nowait - returns locked page at given index in given cache
917 * @mapping: target address_space
918 * @index: the page index
919 *
920 * Same as grab_cache_page(), but do not wait if the page is unavailable.
921 * This is intended for speculative data generators, where the data can
922 * be regenerated if the page couldn't be grabbed. This routine should
923 * be safe to call while holding the lock for another page.
924 *
925 * Clear __GFP_FS when allocating the page to avoid recursion into the fs
926 * and deadlock against the caller's locked page.
927 */
928 struct page *
929 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
930 {
931 struct page *page = find_get_page(mapping, index);
932
933 if (page) {
934 if (trylock_page(page))
935 return page;
936 page_cache_release(page);
937 return NULL;
938 }
939 page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
940 if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
941 page_cache_release(page);
942 page = NULL;
943 }
944 return page;
945 }
946 EXPORT_SYMBOL(grab_cache_page_nowait);
947
948 /*
949 * CD/DVDs are error prone. When a medium error occurs, the driver may fail
950 * a _large_ part of the i/o request. Imagine the worst scenario:
951 *
952 * ---R__________________________________________B__________
953 * ^ reading here ^ bad block(assume 4k)
954 *
955 * read(R) => miss => readahead(R...B) => media error => frustrating retries
956 * => failing the whole request => read(R) => read(R+1) =>
957 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
958 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
959 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
960 *
961 * It is going insane. Fix it by quickly scaling down the readahead size.
962 */
963 static void shrink_readahead_size_eio(struct file *filp,
964 struct file_ra_state *ra)
965 {
966 if (!ra->ra_pages)
967 return;
968
969 ra->ra_pages /= 4;
970 }
971
972 /**
973 * do_generic_file_read - generic file read routine
974 * @filp: the file to read
975 * @ppos: current file position
976 * @desc: read_descriptor
977 * @actor: read method
978 *
979 * This is a generic file read routine, and uses the
980 * mapping->a_ops->readpage() function for the actual low-level stuff.
981 *
982 * This is really ugly. But the goto's actually try to clarify some
983 * of the logic when it comes to error handling etc.
984 */
985 static void do_generic_file_read(struct file *filp, loff_t *ppos,
986 read_descriptor_t *desc, read_actor_t actor)
987 {
988 struct address_space *mapping = filp->f_mapping;
989 struct inode *inode = mapping->host;
990 struct file_ra_state *ra = &filp->f_ra;
991 pgoff_t index;
992 pgoff_t last_index;
993 pgoff_t prev_index;
994 unsigned long offset; /* offset into pagecache page */
995 unsigned int prev_offset;
996 int error;
997
998 index = *ppos >> PAGE_CACHE_SHIFT;
999 prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1000 prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1001 last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1002 offset = *ppos & ~PAGE_CACHE_MASK;
1003
1004 for (;;) {
1005 struct page *page;
1006 pgoff_t end_index;
1007 loff_t isize;
1008 unsigned long nr, ret;
1009
1010 cond_resched();
1011 find_page:
1012 page = find_get_page(mapping, index);
1013 if (!page) {
1014 page_cache_sync_readahead(mapping,
1015 ra, filp,
1016 index, last_index - index);
1017 page = find_get_page(mapping, index);
1018 if (unlikely(page == NULL))
1019 goto no_cached_page;
1020 }
1021 if (PageReadahead(page)) {
1022 page_cache_async_readahead(mapping,
1023 ra, filp, page,
1024 index, last_index - index);
1025 }
1026 if (!PageUptodate(page)) {
1027 if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1028 !mapping->a_ops->is_partially_uptodate)
1029 goto page_not_up_to_date;
1030 if (!trylock_page(page))
1031 goto page_not_up_to_date;
1032 if (!mapping->a_ops->is_partially_uptodate(page,
1033 desc, offset))
1034 goto page_not_up_to_date_locked;
1035 unlock_page(page);
1036 }
1037 page_ok:
1038 /*
1039 * i_size must be checked after we know the page is Uptodate.
1040 *
1041 * Checking i_size after the check allows us to calculate
1042 * the correct value for "nr", which means the zero-filled
1043 * part of the page is not copied back to userspace (unless
1044 * another truncate extends the file - this is desired though).
1045 */
1046
1047 isize = i_size_read(inode);
1048 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1049 if (unlikely(!isize || index > end_index)) {
1050 page_cache_release(page);
1051 goto out;
1052 }
1053
1054 /* nr is the maximum number of bytes to copy from this page */
1055 nr = PAGE_CACHE_SIZE;
1056 if (index == end_index) {
1057 nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1058 if (nr <= offset) {
1059 page_cache_release(page);
1060 goto out;
1061 }
1062 }
1063 nr = nr - offset;
1064
1065 /* If users can be writing to this page using arbitrary
1066 * virtual addresses, take care about potential aliasing
1067 * before reading the page on the kernel side.
1068 */
1069 if (mapping_writably_mapped(mapping))
1070 flush_dcache_page(page);
1071
1072 /*
1073 * When a sequential read accesses a page several times,
1074 * only mark it as accessed the first time.
1075 */
1076 if (prev_index != index || offset != prev_offset)
1077 mark_page_accessed(page);
1078 prev_index = index;
1079
1080 /*
1081 * Ok, we have the page, and it's up-to-date, so
1082 * now we can copy it to user space...
1083 *
1084 * The actor routine returns how many bytes were actually used..
1085 * NOTE! This may not be the same as how much of a user buffer
1086 * we filled up (we may be padding etc), so we can only update
1087 * "pos" here (the actor routine has to update the user buffer
1088 * pointers and the remaining count).
1089 */
1090 ret = actor(desc, page, offset, nr);
1091 offset += ret;
1092 index += offset >> PAGE_CACHE_SHIFT;
1093 offset &= ~PAGE_CACHE_MASK;
1094 prev_offset = offset;
1095
1096 page_cache_release(page);
1097 if (ret == nr && desc->count)
1098 continue;
1099 goto out;
1100
1101 page_not_up_to_date:
1102 /* Get exclusive access to the page ... */
1103 error = lock_page_killable(page);
1104 if (unlikely(error))
1105 goto readpage_error;
1106
1107 page_not_up_to_date_locked:
1108 /* Did it get truncated before we got the lock? */
1109 if (!page->mapping) {
1110 unlock_page(page);
1111 page_cache_release(page);
1112 continue;
1113 }
1114
1115 /* Did somebody else fill it already? */
1116 if (PageUptodate(page)) {
1117 unlock_page(page);
1118 goto page_ok;
1119 }
1120
1121 readpage:
1122 /* Start the actual read. The read will unlock the page. */
1123 error = mapping->a_ops->readpage(filp, page);
1124
1125 if (unlikely(error)) {
1126 if (error == AOP_TRUNCATED_PAGE) {
1127 page_cache_release(page);
1128 goto find_page;
1129 }
1130 goto readpage_error;
1131 }
1132
1133 if (!PageUptodate(page)) {
1134 error = lock_page_killable(page);
1135 if (unlikely(error))
1136 goto readpage_error;
1137 if (!PageUptodate(page)) {
1138 if (page->mapping == NULL) {
1139 /*
1140 * invalidate_inode_pages got it
1141 */
1142 unlock_page(page);
1143 page_cache_release(page);
1144 goto find_page;
1145 }
1146 unlock_page(page);
1147 shrink_readahead_size_eio(filp, ra);
1148 error = -EIO;
1149 goto readpage_error;
1150 }
1151 unlock_page(page);
1152 }
1153
1154 goto page_ok;
1155
1156 readpage_error:
1157 /* UHHUH! A synchronous read error occurred. Report it */
1158 desc->error = error;
1159 page_cache_release(page);
1160 goto out;
1161
1162 no_cached_page:
1163 /*
1164 * Ok, it wasn't cached, so we need to create a new
1165 * page..
1166 */
1167 page = page_cache_alloc_cold(mapping);
1168 if (!page) {
1169 desc->error = -ENOMEM;
1170 goto out;
1171 }
1172 error = add_to_page_cache_lru(page, mapping,
1173 index, GFP_KERNEL);
1174 if (error) {
1175 page_cache_release(page);
1176 if (error == -EEXIST)
1177 goto find_page;
1178 desc->error = error;
1179 goto out;
1180 }
1181 goto readpage;
1182 }
1183
1184 out:
1185 ra->prev_pos = prev_index;
1186 ra->prev_pos <<= PAGE_CACHE_SHIFT;
1187 ra->prev_pos |= prev_offset;
1188
1189 *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1190 if (filp)
1191 file_accessed(filp);
1192 }
1193
1194 int file_read_actor(read_descriptor_t *desc, struct page *page,
1195 unsigned long offset, unsigned long size)
1196 {
1197 char *kaddr;
1198 unsigned long left, count = desc->count;
1199
1200 if (size > count)
1201 size = count;
1202
1203 /*
1204 * Faults on the destination of a read are common, so do it before
1205 * taking the kmap.
1206 */
1207 if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1208 kaddr = kmap_atomic(page, KM_USER0);
1209 left = __copy_to_user_inatomic(desc->arg.buf,
1210 kaddr + offset, size);
1211 kunmap_atomic(kaddr, KM_USER0);
1212 if (left == 0)
1213 goto success;
1214 }
1215
1216 /* Do it the slow way */
1217 kaddr = kmap(page);
1218 left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1219 kunmap(page);
1220
1221 if (left) {
1222 size -= left;
1223 desc->error = -EFAULT;
1224 }
1225 success:
1226 desc->count = count - size;
1227 desc->written += size;
1228 desc->arg.buf += size;
1229 return size;
1230 }
1231
1232 /*
1233 * Performs necessary checks before doing a write
1234 * @iov: io vector request
1235 * @nr_segs: number of segments in the iovec
1236 * @count: number of bytes to write
1237 * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1238 *
1239 * Adjust number of segments and amount of bytes to write (nr_segs should be
1240 * properly initialized first). Returns appropriate error code that caller
1241 * should return or zero in case that write should be allowed.
1242 */
1243 int generic_segment_checks(const struct iovec *iov,
1244 unsigned long *nr_segs, size_t *count, int access_flags)
1245 {
1246 unsigned long seg;
1247 size_t cnt = 0;
1248 for (seg = 0; seg < *nr_segs; seg++) {
1249 const struct iovec *iv = &iov[seg];
1250
1251 /*
1252 * If any segment has a negative length, or the cumulative
1253 * length ever wraps negative then return -EINVAL.
1254 */
1255 cnt += iv->iov_len;
1256 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1257 return -EINVAL;
1258 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1259 continue;
1260 if (seg == 0)
1261 return -EFAULT;
1262 *nr_segs = seg;
1263 cnt -= iv->iov_len; /* This segment is no good */
1264 break;
1265 }
1266 *count = cnt;
1267 return 0;
1268 }
1269 EXPORT_SYMBOL(generic_segment_checks);
1270
1271 /**
1272 * generic_file_aio_read - generic filesystem read routine
1273 * @iocb: kernel I/O control block
1274 * @iov: io vector request
1275 * @nr_segs: number of segments in the iovec
1276 * @pos: current file position
1277 *
1278 * This is the "read()" routine for all filesystems
1279 * that can use the page cache directly.
1280 */
1281 ssize_t
1282 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1283 unsigned long nr_segs, loff_t pos)
1284 {
1285 struct file *filp = iocb->ki_filp;
1286 ssize_t retval;
1287 unsigned long seg;
1288 size_t count;
1289 loff_t *ppos = &iocb->ki_pos;
1290
1291 count = 0;
1292 retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1293 if (retval)
1294 return retval;
1295
1296 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1297 if (filp->f_flags & O_DIRECT) {
1298 loff_t size;
1299 struct address_space *mapping;
1300 struct inode *inode;
1301
1302 mapping = filp->f_mapping;
1303 inode = mapping->host;
1304 if (!count)
1305 goto out; /* skip atime */
1306 size = i_size_read(inode);
1307 if (pos < size) {
1308 retval = filemap_write_and_wait(mapping);
1309 if (!retval) {
1310 retval = mapping->a_ops->direct_IO(READ, iocb,
1311 iov, pos, nr_segs);
1312 }
1313 if (retval > 0)
1314 *ppos = pos + retval;
1315 if (retval) {
1316 file_accessed(filp);
1317 goto out;
1318 }
1319 }
1320 }
1321
1322 for (seg = 0; seg < nr_segs; seg++) {
1323 read_descriptor_t desc;
1324
1325 desc.written = 0;
1326 desc.arg.buf = iov[seg].iov_base;
1327 desc.count = iov[seg].iov_len;
1328 if (desc.count == 0)
1329 continue;
1330 desc.error = 0;
1331 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1332 retval += desc.written;
1333 if (desc.error) {
1334 retval = retval ?: desc.error;
1335 break;
1336 }
1337 if (desc.count > 0)
1338 break;
1339 }
1340 out:
1341 return retval;
1342 }
1343 EXPORT_SYMBOL(generic_file_aio_read);
1344
1345 static ssize_t
1346 do_readahead(struct address_space *mapping, struct file *filp,
1347 pgoff_t index, unsigned long nr)
1348 {
1349 if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1350 return -EINVAL;
1351
1352 force_page_cache_readahead(mapping, filp, index,
1353 max_sane_readahead(nr));
1354 return 0;
1355 }
1356
1357 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1358 {
1359 ssize_t ret;
1360 struct file *file;
1361
1362 ret = -EBADF;
1363 file = fget(fd);
1364 if (file) {
1365 if (file->f_mode & FMODE_READ) {
1366 struct address_space *mapping = file->f_mapping;
1367 pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1368 pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1369 unsigned long len = end - start + 1;
1370 ret = do_readahead(mapping, file, start, len);
1371 }
1372 fput(file);
1373 }
1374 return ret;
1375 }
1376
1377 #ifdef CONFIG_MMU
1378 /**
1379 * page_cache_read - adds requested page to the page cache if not already there
1380 * @file: file to read
1381 * @offset: page index
1382 *
1383 * This adds the requested page to the page cache if it isn't already there,
1384 * and schedules an I/O to read in its contents from disk.
1385 */
1386 static int page_cache_read(struct file *file, pgoff_t offset)
1387 {
1388 struct address_space *mapping = file->f_mapping;
1389 struct page *page;
1390 int ret;
1391
1392 do {
1393 page = page_cache_alloc_cold(mapping);
1394 if (!page)
1395 return -ENOMEM;
1396
1397 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1398 if (ret == 0)
1399 ret = mapping->a_ops->readpage(file, page);
1400 else if (ret == -EEXIST)
1401 ret = 0; /* losing race to add is OK */
1402
1403 page_cache_release(page);
1404
1405 } while (ret == AOP_TRUNCATED_PAGE);
1406
1407 return ret;
1408 }
1409
1410 #define MMAP_LOTSAMISS (100)
1411
1412 /**
1413 * filemap_fault - read in file data for page fault handling
1414 * @vma: vma in which the fault was taken
1415 * @vmf: struct vm_fault containing details of the fault
1416 *
1417 * filemap_fault() is invoked via the vma operations vector for a
1418 * mapped memory region to read in file data during a page fault.
1419 *
1420 * The goto's are kind of ugly, but this streamlines the normal case of having
1421 * it in the page cache, and handles the special cases reasonably without
1422 * having a lot of duplicated code.
1423 */
1424 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1425 {
1426 int error;
1427 struct file *file = vma->vm_file;
1428 struct address_space *mapping = file->f_mapping;
1429 struct file_ra_state *ra = &file->f_ra;
1430 struct inode *inode = mapping->host;
1431 struct page *page;
1432 pgoff_t size;
1433 int did_readaround = 0;
1434 int ret = 0;
1435
1436 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1437 if (vmf->pgoff >= size)
1438 return VM_FAULT_SIGBUS;
1439
1440 /* If we don't want any read-ahead, don't bother */
1441 if (VM_RandomReadHint(vma))
1442 goto no_cached_page;
1443
1444 /*
1445 * Do we have something in the page cache already?
1446 */
1447 retry_find:
1448 page = find_lock_page(mapping, vmf->pgoff);
1449 /*
1450 * For sequential accesses, we use the generic readahead logic.
1451 */
1452 if (VM_SequentialReadHint(vma)) {
1453 if (!page) {
1454 page_cache_sync_readahead(mapping, ra, file,
1455 vmf->pgoff, 1);
1456 page = find_lock_page(mapping, vmf->pgoff);
1457 if (!page)
1458 goto no_cached_page;
1459 }
1460 if (PageReadahead(page)) {
1461 page_cache_async_readahead(mapping, ra, file, page,
1462 vmf->pgoff, 1);
1463 }
1464 }
1465
1466 if (!page) {
1467 unsigned long ra_pages;
1468
1469 ra->mmap_miss++;
1470
1471 /*
1472 * Do we miss much more than hit in this file? If so,
1473 * stop bothering with read-ahead. It will only hurt.
1474 */
1475 if (ra->mmap_miss > MMAP_LOTSAMISS)
1476 goto no_cached_page;
1477
1478 /*
1479 * To keep the pgmajfault counter straight, we need to
1480 * check did_readaround, as this is an inner loop.
1481 */
1482 if (!did_readaround) {
1483 ret = VM_FAULT_MAJOR;
1484 count_vm_event(PGMAJFAULT);
1485 }
1486 did_readaround = 1;
1487 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1488 if (ra_pages) {
1489 pgoff_t start = 0;
1490
1491 if (vmf->pgoff > ra_pages / 2)
1492 start = vmf->pgoff - ra_pages / 2;
1493 do_page_cache_readahead(mapping, file, start, ra_pages);
1494 }
1495 page = find_lock_page(mapping, vmf->pgoff);
1496 if (!page)
1497 goto no_cached_page;
1498 }
1499
1500 if (!did_readaround)
1501 ra->mmap_miss--;
1502
1503 /*
1504 * We have a locked page in the page cache, now we need to check
1505 * that it's up-to-date. If not, it is going to be due to an error.
1506 */
1507 if (unlikely(!PageUptodate(page)))
1508 goto page_not_uptodate;
1509
1510 /* Must recheck i_size under page lock */
1511 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1512 if (unlikely(vmf->pgoff >= size)) {
1513 unlock_page(page);
1514 page_cache_release(page);
1515 return VM_FAULT_SIGBUS;
1516 }
1517
1518 /*
1519 * Found the page and have a reference on it.
1520 */
1521 mark_page_accessed(page);
1522 ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1523 vmf->page = page;
1524 return ret | VM_FAULT_LOCKED;
1525
1526 no_cached_page:
1527 /*
1528 * We're only likely to ever get here if MADV_RANDOM is in
1529 * effect.
1530 */
1531 error = page_cache_read(file, vmf->pgoff);
1532
1533 /*
1534 * The page we want has now been added to the page cache.
1535 * In the unlikely event that someone removed it in the
1536 * meantime, we'll just come back here and read it again.
1537 */
1538 if (error >= 0)
1539 goto retry_find;
1540
1541 /*
1542 * An error return from page_cache_read can result if the
1543 * system is low on memory, or a problem occurs while trying
1544 * to schedule I/O.
1545 */
1546 if (error == -ENOMEM)
1547 return VM_FAULT_OOM;
1548 return VM_FAULT_SIGBUS;
1549
1550 page_not_uptodate:
1551 /* IO error path */
1552 if (!did_readaround) {
1553 ret = VM_FAULT_MAJOR;
1554 count_vm_event(PGMAJFAULT);
1555 }
1556
1557 /*
1558 * Umm, take care of errors if the page isn't up-to-date.
1559 * Try to re-read it _once_. We do this synchronously,
1560 * because there really aren't any performance issues here
1561 * and we need to check for errors.
1562 */
1563 ClearPageError(page);
1564 error = mapping->a_ops->readpage(file, page);
1565 if (!error) {
1566 wait_on_page_locked(page);
1567 if (!PageUptodate(page))
1568 error = -EIO;
1569 }
1570 page_cache_release(page);
1571
1572 if (!error || error == AOP_TRUNCATED_PAGE)
1573 goto retry_find;
1574
1575 /* Things didn't work out. Return zero to tell the mm layer so. */
1576 shrink_readahead_size_eio(file, ra);
1577 return VM_FAULT_SIGBUS;
1578 }
1579 EXPORT_SYMBOL(filemap_fault);
1580
1581 struct vm_operations_struct generic_file_vm_ops = {
1582 .fault = filemap_fault,
1583 };
1584
1585 /* This is used for a general mmap of a disk file */
1586
1587 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1588 {
1589 struct address_space *mapping = file->f_mapping;
1590
1591 if (!mapping->a_ops->readpage)
1592 return -ENOEXEC;
1593 file_accessed(file);
1594 vma->vm_ops = &generic_file_vm_ops;
1595 vma->vm_flags |= VM_CAN_NONLINEAR;
1596 return 0;
1597 }
1598
1599 /*
1600 * This is for filesystems which do not implement ->writepage.
1601 */
1602 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1603 {
1604 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1605 return -EINVAL;
1606 return generic_file_mmap(file, vma);
1607 }
1608 #else
1609 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1610 {
1611 return -ENOSYS;
1612 }
1613 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1614 {
1615 return -ENOSYS;
1616 }
1617 #endif /* CONFIG_MMU */
1618
1619 EXPORT_SYMBOL(generic_file_mmap);
1620 EXPORT_SYMBOL(generic_file_readonly_mmap);
1621
1622 static struct page *__read_cache_page(struct address_space *mapping,
1623 pgoff_t index,
1624 int (*filler)(void *,struct page*),
1625 void *data)
1626 {
1627 struct page *page;
1628 int err;
1629 repeat:
1630 page = find_get_page(mapping, index);
1631 if (!page) {
1632 page = page_cache_alloc_cold(mapping);
1633 if (!page)
1634 return ERR_PTR(-ENOMEM);
1635 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1636 if (unlikely(err)) {
1637 page_cache_release(page);
1638 if (err == -EEXIST)
1639 goto repeat;
1640 /* Presumably ENOMEM for radix tree node */
1641 return ERR_PTR(err);
1642 }
1643 err = filler(data, page);
1644 if (err < 0) {
1645 page_cache_release(page);
1646 page = ERR_PTR(err);
1647 }
1648 }
1649 return page;
1650 }
1651
1652 /**
1653 * read_cache_page_async - read into page cache, fill it if needed
1654 * @mapping: the page's address_space
1655 * @index: the page index
1656 * @filler: function to perform the read
1657 * @data: destination for read data
1658 *
1659 * Same as read_cache_page, but don't wait for page to become unlocked
1660 * after submitting it to the filler.
1661 *
1662 * Read into the page cache. If a page already exists, and PageUptodate() is
1663 * not set, try to fill the page but don't wait for it to become unlocked.
1664 *
1665 * If the page does not get brought uptodate, return -EIO.
1666 */
1667 struct page *read_cache_page_async(struct address_space *mapping,
1668 pgoff_t index,
1669 int (*filler)(void *,struct page*),
1670 void *data)
1671 {
1672 struct page *page;
1673 int err;
1674
1675 retry:
1676 page = __read_cache_page(mapping, index, filler, data);
1677 if (IS_ERR(page))
1678 return page;
1679 if (PageUptodate(page))
1680 goto out;
1681
1682 lock_page(page);
1683 if (!page->mapping) {
1684 unlock_page(page);
1685 page_cache_release(page);
1686 goto retry;
1687 }
1688 if (PageUptodate(page)) {
1689 unlock_page(page);
1690 goto out;
1691 }
1692 err = filler(data, page);
1693 if (err < 0) {
1694 page_cache_release(page);
1695 return ERR_PTR(err);
1696 }
1697 out:
1698 mark_page_accessed(page);
1699 return page;
1700 }
1701 EXPORT_SYMBOL(read_cache_page_async);
1702
1703 /**
1704 * read_cache_page - read into page cache, fill it if needed
1705 * @mapping: the page's address_space
1706 * @index: the page index
1707 * @filler: function to perform the read
1708 * @data: destination for read data
1709 *
1710 * Read into the page cache. If a page already exists, and PageUptodate() is
1711 * not set, try to fill the page then wait for it to become unlocked.
1712 *
1713 * If the page does not get brought uptodate, return -EIO.
1714 */
1715 struct page *read_cache_page(struct address_space *mapping,
1716 pgoff_t index,
1717 int (*filler)(void *,struct page*),
1718 void *data)
1719 {
1720 struct page *page;
1721
1722 page = read_cache_page_async(mapping, index, filler, data);
1723 if (IS_ERR(page))
1724 goto out;
1725 wait_on_page_locked(page);
1726 if (!PageUptodate(page)) {
1727 page_cache_release(page);
1728 page = ERR_PTR(-EIO);
1729 }
1730 out:
1731 return page;
1732 }
1733 EXPORT_SYMBOL(read_cache_page);
1734
1735 /*
1736 * The logic we want is
1737 *
1738 * if suid or (sgid and xgrp)
1739 * remove privs
1740 */
1741 int should_remove_suid(struct dentry *dentry)
1742 {
1743 mode_t mode = dentry->d_inode->i_mode;
1744 int kill = 0;
1745
1746 /* suid always must be killed */
1747 if (unlikely(mode & S_ISUID))
1748 kill = ATTR_KILL_SUID;
1749
1750 /*
1751 * sgid without any exec bits is just a mandatory locking mark; leave
1752 * it alone. If some exec bits are set, it's a real sgid; kill it.
1753 */
1754 if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1755 kill |= ATTR_KILL_SGID;
1756
1757 if (unlikely(kill && !capable(CAP_FSETID)))
1758 return kill;
1759
1760 return 0;
1761 }
1762 EXPORT_SYMBOL(should_remove_suid);
1763
1764 static int __remove_suid(struct dentry *dentry, int kill)
1765 {
1766 struct iattr newattrs;
1767
1768 newattrs.ia_valid = ATTR_FORCE | kill;
1769 return notify_change(dentry, &newattrs);
1770 }
1771
1772 int file_remove_suid(struct file *file)
1773 {
1774 struct dentry *dentry = file->f_path.dentry;
1775 int killsuid = should_remove_suid(dentry);
1776 int killpriv = security_inode_need_killpriv(dentry);
1777 int error = 0;
1778
1779 if (killpriv < 0)
1780 return killpriv;
1781 if (killpriv)
1782 error = security_inode_killpriv(dentry);
1783 if (!error && killsuid)
1784 error = __remove_suid(dentry, killsuid);
1785
1786 return error;
1787 }
1788 EXPORT_SYMBOL(file_remove_suid);
1789
1790 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1791 const struct iovec *iov, size_t base, size_t bytes)
1792 {
1793 size_t copied = 0, left = 0;
1794
1795 while (bytes) {
1796 char __user *buf = iov->iov_base + base;
1797 int copy = min(bytes, iov->iov_len - base);
1798
1799 base = 0;
1800 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1801 copied += copy;
1802 bytes -= copy;
1803 vaddr += copy;
1804 iov++;
1805
1806 if (unlikely(left))
1807 break;
1808 }
1809 return copied - left;
1810 }
1811
1812 /*
1813 * Copy as much as we can into the page and return the number of bytes which
1814 * were sucessfully copied. If a fault is encountered then return the number of
1815 * bytes which were copied.
1816 */
1817 size_t iov_iter_copy_from_user_atomic(struct page *page,
1818 struct iov_iter *i, unsigned long offset, size_t bytes)
1819 {
1820 char *kaddr;
1821 size_t copied;
1822
1823 BUG_ON(!in_atomic());
1824 kaddr = kmap_atomic(page, KM_USER0);
1825 if (likely(i->nr_segs == 1)) {
1826 int left;
1827 char __user *buf = i->iov->iov_base + i->iov_offset;
1828 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1829 buf, bytes);
1830 copied = bytes - left;
1831 } else {
1832 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1833 i->iov, i->iov_offset, bytes);
1834 }
1835 kunmap_atomic(kaddr, KM_USER0);
1836
1837 return copied;
1838 }
1839 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1840
1841 /*
1842 * This has the same sideeffects and return value as
1843 * iov_iter_copy_from_user_atomic().
1844 * The difference is that it attempts to resolve faults.
1845 * Page must not be locked.
1846 */
1847 size_t iov_iter_copy_from_user(struct page *page,
1848 struct iov_iter *i, unsigned long offset, size_t bytes)
1849 {
1850 char *kaddr;
1851 size_t copied;
1852
1853 kaddr = kmap(page);
1854 if (likely(i->nr_segs == 1)) {
1855 int left;
1856 char __user *buf = i->iov->iov_base + i->iov_offset;
1857 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1858 copied = bytes - left;
1859 } else {
1860 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1861 i->iov, i->iov_offset, bytes);
1862 }
1863 kunmap(page);
1864 return copied;
1865 }
1866 EXPORT_SYMBOL(iov_iter_copy_from_user);
1867
1868 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1869 {
1870 BUG_ON(i->count < bytes);
1871
1872 if (likely(i->nr_segs == 1)) {
1873 i->iov_offset += bytes;
1874 i->count -= bytes;
1875 } else {
1876 const struct iovec *iov = i->iov;
1877 size_t base = i->iov_offset;
1878
1879 /*
1880 * The !iov->iov_len check ensures we skip over unlikely
1881 * zero-length segments (without overruning the iovec).
1882 */
1883 while (bytes || unlikely(i->count && !iov->iov_len)) {
1884 int copy;
1885
1886 copy = min(bytes, iov->iov_len - base);
1887 BUG_ON(!i->count || i->count < copy);
1888 i->count -= copy;
1889 bytes -= copy;
1890 base += copy;
1891 if (iov->iov_len == base) {
1892 iov++;
1893 base = 0;
1894 }
1895 }
1896 i->iov = iov;
1897 i->iov_offset = base;
1898 }
1899 }
1900 EXPORT_SYMBOL(iov_iter_advance);
1901
1902 /*
1903 * Fault in the first iovec of the given iov_iter, to a maximum length
1904 * of bytes. Returns 0 on success, or non-zero if the memory could not be
1905 * accessed (ie. because it is an invalid address).
1906 *
1907 * writev-intensive code may want this to prefault several iovecs -- that
1908 * would be possible (callers must not rely on the fact that _only_ the
1909 * first iovec will be faulted with the current implementation).
1910 */
1911 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1912 {
1913 char __user *buf = i->iov->iov_base + i->iov_offset;
1914 bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1915 return fault_in_pages_readable(buf, bytes);
1916 }
1917 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1918
1919 /*
1920 * Return the count of just the current iov_iter segment.
1921 */
1922 size_t iov_iter_single_seg_count(struct iov_iter *i)
1923 {
1924 const struct iovec *iov = i->iov;
1925 if (i->nr_segs == 1)
1926 return i->count;
1927 else
1928 return min(i->count, iov->iov_len - i->iov_offset);
1929 }
1930 EXPORT_SYMBOL(iov_iter_single_seg_count);
1931
1932 /*
1933 * Performs necessary checks before doing a write
1934 *
1935 * Can adjust writing position or amount of bytes to write.
1936 * Returns appropriate error code that caller should return or
1937 * zero in case that write should be allowed.
1938 */
1939 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1940 {
1941 struct inode *inode = file->f_mapping->host;
1942 unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1943
1944 if (unlikely(*pos < 0))
1945 return -EINVAL;
1946
1947 if (!isblk) {
1948 /* FIXME: this is for backwards compatibility with 2.4 */
1949 if (file->f_flags & O_APPEND)
1950 *pos = i_size_read(inode);
1951
1952 if (limit != RLIM_INFINITY) {
1953 if (*pos >= limit) {
1954 send_sig(SIGXFSZ, current, 0);
1955 return -EFBIG;
1956 }
1957 if (*count > limit - (typeof(limit))*pos) {
1958 *count = limit - (typeof(limit))*pos;
1959 }
1960 }
1961 }
1962
1963 /*
1964 * LFS rule
1965 */
1966 if (unlikely(*pos + *count > MAX_NON_LFS &&
1967 !(file->f_flags & O_LARGEFILE))) {
1968 if (*pos >= MAX_NON_LFS) {
1969 return -EFBIG;
1970 }
1971 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1972 *count = MAX_NON_LFS - (unsigned long)*pos;
1973 }
1974 }
1975
1976 /*
1977 * Are we about to exceed the fs block limit ?
1978 *
1979 * If we have written data it becomes a short write. If we have
1980 * exceeded without writing data we send a signal and return EFBIG.
1981 * Linus frestrict idea will clean these up nicely..
1982 */
1983 if (likely(!isblk)) {
1984 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1985 if (*count || *pos > inode->i_sb->s_maxbytes) {
1986 return -EFBIG;
1987 }
1988 /* zero-length writes at ->s_maxbytes are OK */
1989 }
1990
1991 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1992 *count = inode->i_sb->s_maxbytes - *pos;
1993 } else {
1994 #ifdef CONFIG_BLOCK
1995 loff_t isize;
1996 if (bdev_read_only(I_BDEV(inode)))
1997 return -EPERM;
1998 isize = i_size_read(inode);
1999 if (*pos >= isize) {
2000 if (*count || *pos > isize)
2001 return -ENOSPC;
2002 }
2003
2004 if (*pos + *count > isize)
2005 *count = isize - *pos;
2006 #else
2007 return -EPERM;
2008 #endif
2009 }
2010 return 0;
2011 }
2012 EXPORT_SYMBOL(generic_write_checks);
2013
2014 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2015 loff_t pos, unsigned len, unsigned flags,
2016 struct page **pagep, void **fsdata)
2017 {
2018 const struct address_space_operations *aops = mapping->a_ops;
2019
2020 if (aops->write_begin) {
2021 return aops->write_begin(file, mapping, pos, len, flags,
2022 pagep, fsdata);
2023 } else {
2024 int ret;
2025 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2026 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2027 struct inode *inode = mapping->host;
2028 struct page *page;
2029 again:
2030 page = __grab_cache_page(mapping, index);
2031 *pagep = page;
2032 if (!page)
2033 return -ENOMEM;
2034
2035 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
2036 /*
2037 * There is no way to resolve a short write situation
2038 * for a !Uptodate page (except by double copying in
2039 * the caller done by generic_perform_write_2copy).
2040 *
2041 * Instead, we have to bring it uptodate here.
2042 */
2043 ret = aops->readpage(file, page);
2044 page_cache_release(page);
2045 if (ret) {
2046 if (ret == AOP_TRUNCATED_PAGE)
2047 goto again;
2048 return ret;
2049 }
2050 goto again;
2051 }
2052
2053 ret = aops->prepare_write(file, page, offset, offset+len);
2054 if (ret) {
2055 unlock_page(page);
2056 page_cache_release(page);
2057 if (pos + len > inode->i_size)
2058 vmtruncate(inode, inode->i_size);
2059 }
2060 return ret;
2061 }
2062 }
2063 EXPORT_SYMBOL(pagecache_write_begin);
2064
2065 int pagecache_write_end(struct file *file, struct address_space *mapping,
2066 loff_t pos, unsigned len, unsigned copied,
2067 struct page *page, void *fsdata)
2068 {
2069 const struct address_space_operations *aops = mapping->a_ops;
2070 int ret;
2071
2072 if (aops->write_end) {
2073 mark_page_accessed(page);
2074 ret = aops->write_end(file, mapping, pos, len, copied,
2075 page, fsdata);
2076 } else {
2077 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2078 struct inode *inode = mapping->host;
2079
2080 flush_dcache_page(page);
2081 ret = aops->commit_write(file, page, offset, offset+len);
2082 unlock_page(page);
2083 mark_page_accessed(page);
2084 page_cache_release(page);
2085
2086 if (ret < 0) {
2087 if (pos + len > inode->i_size)
2088 vmtruncate(inode, inode->i_size);
2089 } else if (ret > 0)
2090 ret = min_t(size_t, copied, ret);
2091 else
2092 ret = copied;
2093 }
2094
2095 return ret;
2096 }
2097 EXPORT_SYMBOL(pagecache_write_end);
2098
2099 ssize_t
2100 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2101 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2102 size_t count, size_t ocount)
2103 {
2104 struct file *file = iocb->ki_filp;
2105 struct address_space *mapping = file->f_mapping;
2106 struct inode *inode = mapping->host;
2107 ssize_t written;
2108 size_t write_len;
2109 pgoff_t end;
2110
2111 if (count != ocount)
2112 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2113
2114 /*
2115 * Unmap all mmappings of the file up-front.
2116 *
2117 * This will cause any pte dirty bits to be propagated into the
2118 * pageframes for the subsequent filemap_write_and_wait().
2119 */
2120 write_len = iov_length(iov, *nr_segs);
2121 end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2122 if (mapping_mapped(mapping))
2123 unmap_mapping_range(mapping, pos, write_len, 0);
2124
2125 written = filemap_write_and_wait(mapping);
2126 if (written)
2127 goto out;
2128
2129 /*
2130 * After a write we want buffered reads to be sure to go to disk to get
2131 * the new data. We invalidate clean cached page from the region we're
2132 * about to write. We do this *before* the write so that we can return
2133 * without clobbering -EIOCBQUEUED from ->direct_IO().
2134 */
2135 if (mapping->nrpages) {
2136 written = invalidate_inode_pages2_range(mapping,
2137 pos >> PAGE_CACHE_SHIFT, end);
2138 /*
2139 * If a page can not be invalidated, return 0 to fall back
2140 * to buffered write.
2141 */
2142 if (written) {
2143 if (written == -EBUSY)
2144 return 0;
2145 goto out;
2146 }
2147 }
2148
2149 written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2150
2151 /*
2152 * Finally, try again to invalidate clean pages which might have been
2153 * cached by non-direct readahead, or faulted in by get_user_pages()
2154 * if the source of the write was an mmap'ed region of the file
2155 * we're writing. Either one is a pretty crazy thing to do,
2156 * so we don't support it 100%. If this invalidation
2157 * fails, tough, the write still worked...
2158 */
2159 if (mapping->nrpages) {
2160 invalidate_inode_pages2_range(mapping,
2161 pos >> PAGE_CACHE_SHIFT, end);
2162 }
2163
2164 if (written > 0) {
2165 loff_t end = pos + written;
2166 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2167 i_size_write(inode, end);
2168 mark_inode_dirty(inode);
2169 }
2170 *ppos = end;
2171 }
2172
2173 /*
2174 * Sync the fs metadata but not the minor inode changes and
2175 * of course not the data as we did direct DMA for the IO.
2176 * i_mutex is held, which protects generic_osync_inode() from
2177 * livelocking. AIO O_DIRECT ops attempt to sync metadata here.
2178 */
2179 out:
2180 if ((written >= 0 || written == -EIOCBQUEUED) &&
2181 ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2182 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2183 if (err < 0)
2184 written = err;
2185 }
2186 return written;
2187 }
2188 EXPORT_SYMBOL(generic_file_direct_write);
2189
2190 /*
2191 * Find or create a page at the given pagecache position. Return the locked
2192 * page. This function is specifically for buffered writes.
2193 */
2194 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2195 {
2196 int status;
2197 struct page *page;
2198 repeat:
2199 page = find_lock_page(mapping, index);
2200 if (likely(page))
2201 return page;
2202
2203 page = page_cache_alloc(mapping);
2204 if (!page)
2205 return NULL;
2206 status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2207 if (unlikely(status)) {
2208 page_cache_release(page);
2209 if (status == -EEXIST)
2210 goto repeat;
2211 return NULL;
2212 }
2213 return page;
2214 }
2215 EXPORT_SYMBOL(__grab_cache_page);
2216
2217 static ssize_t generic_perform_write_2copy(struct file *file,
2218 struct iov_iter *i, loff_t pos)
2219 {
2220 struct address_space *mapping = file->f_mapping;
2221 const struct address_space_operations *a_ops = mapping->a_ops;
2222 struct inode *inode = mapping->host;
2223 long status = 0;
2224 ssize_t written = 0;
2225
2226 do {
2227 struct page *src_page;
2228 struct page *page;
2229 pgoff_t index; /* Pagecache index for current page */
2230 unsigned long offset; /* Offset into pagecache page */
2231 unsigned long bytes; /* Bytes to write to page */
2232 size_t copied; /* Bytes copied from user */
2233
2234 offset = (pos & (PAGE_CACHE_SIZE - 1));
2235 index = pos >> PAGE_CACHE_SHIFT;
2236 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2237 iov_iter_count(i));
2238
2239 /*
2240 * a non-NULL src_page indicates that we're doing the
2241 * copy via get_user_pages and kmap.
2242 */
2243 src_page = NULL;
2244
2245 /*
2246 * Bring in the user page that we will copy from _first_.
2247 * Otherwise there's a nasty deadlock on copying from the
2248 * same page as we're writing to, without it being marked
2249 * up-to-date.
2250 *
2251 * Not only is this an optimisation, but it is also required
2252 * to check that the address is actually valid, when atomic
2253 * usercopies are used, below.
2254 */
2255 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2256 status = -EFAULT;
2257 break;
2258 }
2259
2260 page = __grab_cache_page(mapping, index);
2261 if (!page) {
2262 status = -ENOMEM;
2263 break;
2264 }
2265
2266 /*
2267 * non-uptodate pages cannot cope with short copies, and we
2268 * cannot take a pagefault with the destination page locked.
2269 * So pin the source page to copy it.
2270 */
2271 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2272 unlock_page(page);
2273
2274 src_page = alloc_page(GFP_KERNEL);
2275 if (!src_page) {
2276 page_cache_release(page);
2277 status = -ENOMEM;
2278 break;
2279 }
2280
2281 /*
2282 * Cannot get_user_pages with a page locked for the
2283 * same reason as we can't take a page fault with a
2284 * page locked (as explained below).
2285 */
2286 copied = iov_iter_copy_from_user(src_page, i,
2287 offset, bytes);
2288 if (unlikely(copied == 0)) {
2289 status = -EFAULT;
2290 page_cache_release(page);
2291 page_cache_release(src_page);
2292 break;
2293 }
2294 bytes = copied;
2295
2296 lock_page(page);
2297 /*
2298 * Can't handle the page going uptodate here, because
2299 * that means we would use non-atomic usercopies, which
2300 * zero out the tail of the page, which can cause
2301 * zeroes to become transiently visible. We could just
2302 * use a non-zeroing copy, but the APIs aren't too
2303 * consistent.
2304 */
2305 if (unlikely(!page->mapping || PageUptodate(page))) {
2306 unlock_page(page);
2307 page_cache_release(page);
2308 page_cache_release(src_page);
2309 continue;
2310 }
2311 }
2312
2313 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2314 if (unlikely(status))
2315 goto fs_write_aop_error;
2316
2317 if (!src_page) {
2318 /*
2319 * Must not enter the pagefault handler here, because
2320 * we hold the page lock, so we might recursively
2321 * deadlock on the same lock, or get an ABBA deadlock
2322 * against a different lock, or against the mmap_sem
2323 * (which nests outside the page lock). So increment
2324 * preempt count, and use _atomic usercopies.
2325 *
2326 * The page is uptodate so we are OK to encounter a
2327 * short copy: if unmodified parts of the page are
2328 * marked dirty and written out to disk, it doesn't
2329 * really matter.
2330 */
2331 pagefault_disable();
2332 copied = iov_iter_copy_from_user_atomic(page, i,
2333 offset, bytes);
2334 pagefault_enable();
2335 } else {
2336 void *src, *dst;
2337 src = kmap_atomic(src_page, KM_USER0);
2338 dst = kmap_atomic(page, KM_USER1);
2339 memcpy(dst + offset, src + offset, bytes);
2340 kunmap_atomic(dst, KM_USER1);
2341 kunmap_atomic(src, KM_USER0);
2342 copied = bytes;
2343 }
2344 flush_dcache_page(page);
2345
2346 status = a_ops->commit_write(file, page, offset, offset+bytes);
2347 if (unlikely(status < 0))
2348 goto fs_write_aop_error;
2349 if (unlikely(status > 0)) /* filesystem did partial write */
2350 copied = min_t(size_t, copied, status);
2351
2352 unlock_page(page);
2353 mark_page_accessed(page);
2354 page_cache_release(page);
2355 if (src_page)
2356 page_cache_release(src_page);
2357
2358 iov_iter_advance(i, copied);
2359 pos += copied;
2360 written += copied;
2361
2362 balance_dirty_pages_ratelimited(mapping);
2363 cond_resched();
2364 continue;
2365
2366 fs_write_aop_error:
2367 unlock_page(page);
2368 page_cache_release(page);
2369 if (src_page)
2370 page_cache_release(src_page);
2371
2372 /*
2373 * prepare_write() may have instantiated a few blocks
2374 * outside i_size. Trim these off again. Don't need
2375 * i_size_read because we hold i_mutex.
2376 */
2377 if (pos + bytes > inode->i_size)
2378 vmtruncate(inode, inode->i_size);
2379 break;
2380 } while (iov_iter_count(i));
2381
2382 return written ? written : status;
2383 }
2384
2385 static ssize_t generic_perform_write(struct file *file,
2386 struct iov_iter *i, loff_t pos)
2387 {
2388 struct address_space *mapping = file->f_mapping;
2389 const struct address_space_operations *a_ops = mapping->a_ops;
2390 long status = 0;
2391 ssize_t written = 0;
2392 unsigned int flags = 0;
2393
2394 /*
2395 * Copies from kernel address space cannot fail (NFSD is a big user).
2396 */
2397 if (segment_eq(get_fs(), KERNEL_DS))
2398 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2399
2400 do {
2401 struct page *page;
2402 pgoff_t index; /* Pagecache index for current page */
2403 unsigned long offset; /* Offset into pagecache page */
2404 unsigned long bytes; /* Bytes to write to page */
2405 size_t copied; /* Bytes copied from user */
2406 void *fsdata;
2407
2408 offset = (pos & (PAGE_CACHE_SIZE - 1));
2409 index = pos >> PAGE_CACHE_SHIFT;
2410 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2411 iov_iter_count(i));
2412
2413 again:
2414
2415 /*
2416 * Bring in the user page that we will copy from _first_.
2417 * Otherwise there's a nasty deadlock on copying from the
2418 * same page as we're writing to, without it being marked
2419 * up-to-date.
2420 *
2421 * Not only is this an optimisation, but it is also required
2422 * to check that the address is actually valid, when atomic
2423 * usercopies are used, below.
2424 */
2425 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2426 status = -EFAULT;
2427 break;
2428 }
2429
2430 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2431 &page, &fsdata);
2432 if (unlikely(status))
2433 break;
2434
2435 pagefault_disable();
2436 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2437 pagefault_enable();
2438 flush_dcache_page(page);
2439
2440 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2441 page, fsdata);
2442 if (unlikely(status < 0))
2443 break;
2444 copied = status;
2445
2446 cond_resched();
2447
2448 iov_iter_advance(i, copied);
2449 if (unlikely(copied == 0)) {
2450 /*
2451 * If we were unable to copy any data at all, we must
2452 * fall back to a single segment length write.
2453 *
2454 * If we didn't fallback here, we could livelock
2455 * because not all segments in the iov can be copied at
2456 * once without a pagefault.
2457 */
2458 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2459 iov_iter_single_seg_count(i));
2460 goto again;
2461 }
2462 pos += copied;
2463 written += copied;
2464
2465 balance_dirty_pages_ratelimited(mapping);
2466
2467 } while (iov_iter_count(i));
2468
2469 return written ? written : status;
2470 }
2471
2472 ssize_t
2473 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2474 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2475 size_t count, ssize_t written)
2476 {
2477 struct file *file = iocb->ki_filp;
2478 struct address_space *mapping = file->f_mapping;
2479 const struct address_space_operations *a_ops = mapping->a_ops;
2480 struct inode *inode = mapping->host;
2481 ssize_t status;
2482 struct iov_iter i;
2483
2484 iov_iter_init(&i, iov, nr_segs, count, written);
2485 if (a_ops->write_begin)
2486 status = generic_perform_write(file, &i, pos);
2487 else
2488 status = generic_perform_write_2copy(file, &i, pos);
2489
2490 if (likely(status >= 0)) {
2491 written += status;
2492 *ppos = pos + status;
2493
2494 /*
2495 * For now, when the user asks for O_SYNC, we'll actually give
2496 * O_DSYNC
2497 */
2498 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2499 if (!a_ops->writepage || !is_sync_kiocb(iocb))
2500 status = generic_osync_inode(inode, mapping,
2501 OSYNC_METADATA|OSYNC_DATA);
2502 }
2503 }
2504
2505 /*
2506 * If we get here for O_DIRECT writes then we must have fallen through
2507 * to buffered writes (block instantiation inside i_size). So we sync
2508 * the file data here, to try to honour O_DIRECT expectations.
2509 */
2510 if (unlikely(file->f_flags & O_DIRECT) && written)
2511 status = filemap_write_and_wait(mapping);
2512
2513 return written ? written : status;
2514 }
2515 EXPORT_SYMBOL(generic_file_buffered_write);
2516
2517 static ssize_t
2518 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2519 unsigned long nr_segs, loff_t *ppos)
2520 {
2521 struct file *file = iocb->ki_filp;
2522 struct address_space * mapping = file->f_mapping;
2523 size_t ocount; /* original count */
2524 size_t count; /* after file limit checks */
2525 struct inode *inode = mapping->host;
2526 loff_t pos;
2527 ssize_t written;
2528 ssize_t err;
2529
2530 ocount = 0;
2531 err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2532 if (err)
2533 return err;
2534
2535 count = ocount;
2536 pos = *ppos;
2537
2538 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2539
2540 /* We can write back this queue in page reclaim */
2541 current->backing_dev_info = mapping->backing_dev_info;
2542 written = 0;
2543
2544 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2545 if (err)
2546 goto out;
2547
2548 if (count == 0)
2549 goto out;
2550
2551 err = file_remove_suid(file);
2552 if (err)
2553 goto out;
2554
2555 file_update_time(file);
2556
2557 /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2558 if (unlikely(file->f_flags & O_DIRECT)) {
2559 loff_t endbyte;
2560 ssize_t written_buffered;
2561
2562 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2563 ppos, count, ocount);
2564 if (written < 0 || written == count)
2565 goto out;
2566 /*
2567 * direct-io write to a hole: fall through to buffered I/O
2568 * for completing the rest of the request.
2569 */
2570 pos += written;
2571 count -= written;
2572 written_buffered = generic_file_buffered_write(iocb, iov,
2573 nr_segs, pos, ppos, count,
2574 written);
2575 /*
2576 * If generic_file_buffered_write() retuned a synchronous error
2577 * then we want to return the number of bytes which were
2578 * direct-written, or the error code if that was zero. Note
2579 * that this differs from normal direct-io semantics, which
2580 * will return -EFOO even if some bytes were written.
2581 */
2582 if (written_buffered < 0) {
2583 err = written_buffered;
2584 goto out;
2585 }
2586
2587 /*
2588 * We need to ensure that the page cache pages are written to
2589 * disk and invalidated to preserve the expected O_DIRECT
2590 * semantics.
2591 */
2592 endbyte = pos + written_buffered - written - 1;
2593 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2594 SYNC_FILE_RANGE_WAIT_BEFORE|
2595 SYNC_FILE_RANGE_WRITE|
2596 SYNC_FILE_RANGE_WAIT_AFTER);
2597 if (err == 0) {
2598 written = written_buffered;
2599 invalidate_mapping_pages(mapping,
2600 pos >> PAGE_CACHE_SHIFT,
2601 endbyte >> PAGE_CACHE_SHIFT);
2602 } else {
2603 /*
2604 * We don't know how much we wrote, so just return
2605 * the number of bytes which were direct-written
2606 */
2607 }
2608 } else {
2609 written = generic_file_buffered_write(iocb, iov, nr_segs,
2610 pos, ppos, count, written);
2611 }
2612 out:
2613 current->backing_dev_info = NULL;
2614 return written ? written : err;
2615 }
2616
2617 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2618 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2619 {
2620 struct file *file = iocb->ki_filp;
2621 struct address_space *mapping = file->f_mapping;
2622 struct inode *inode = mapping->host;
2623 ssize_t ret;
2624
2625 BUG_ON(iocb->ki_pos != pos);
2626
2627 ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2628 &iocb->ki_pos);
2629
2630 if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2631 ssize_t err;
2632
2633 err = sync_page_range_nolock(inode, mapping, pos, ret);
2634 if (err < 0)
2635 ret = err;
2636 }
2637 return ret;
2638 }
2639 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2640
2641 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2642 unsigned long nr_segs, loff_t pos)
2643 {
2644 struct file *file = iocb->ki_filp;
2645 struct address_space *mapping = file->f_mapping;
2646 struct inode *inode = mapping->host;
2647 ssize_t ret;
2648
2649 BUG_ON(iocb->ki_pos != pos);
2650
2651 mutex_lock(&inode->i_mutex);
2652 ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2653 &iocb->ki_pos);
2654 mutex_unlock(&inode->i_mutex);
2655
2656 if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2657 ssize_t err;
2658
2659 err = sync_page_range(inode, mapping, pos, ret);
2660 if (err < 0)
2661 ret = err;
2662 }
2663 return ret;
2664 }
2665 EXPORT_SYMBOL(generic_file_aio_write);
2666
2667 /**
2668 * try_to_release_page() - release old fs-specific metadata on a page
2669 *
2670 * @page: the page which the kernel is trying to free
2671 * @gfp_mask: memory allocation flags (and I/O mode)
2672 *
2673 * The address_space is to try to release any data against the page
2674 * (presumably at page->private). If the release was successful, return `1'.
2675 * Otherwise return zero.
2676 *
2677 * The @gfp_mask argument specifies whether I/O may be performed to release
2678 * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2679 *
2680 */
2681 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2682 {
2683 struct address_space * const mapping = page->mapping;
2684
2685 BUG_ON(!PageLocked(page));
2686 if (PageWriteback(page))
2687 return 0;
2688
2689 if (mapping && mapping->a_ops->releasepage)
2690 return mapping->a_ops->releasepage(page, gfp_mask);
2691 return try_to_free_buffers(page);
2692 }
2693
2694 EXPORT_SYMBOL(try_to_release_page);
This page took 0.121572 seconds and 5 git commands to generate.