linux - fs/buffer.c

3436 lines · 90.5 KiB

1	/*
2	* linux/fs/buffer.c
3	*
4	* Copyright (C) 1991, 1992, 2002 Linus Torvalds
5	*/
6
7	/*
8	* Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
9	*
10	* Removed a lot of unnecessary code and simplified things now that
11	* the buffer cache isn't our primary cache - Andrew Tridgell 12/96
12	*
13	* Speed up hash, lru, and free list operations. Use gfp() for allocating
14	* hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
15	*
16	* Added 32k buffer block sizes - these are required older ARM systems. - RMK
17	*
18	* async buffer flushing, 1999 Andrea Arcangeli <[email protected]>
19	*/
20
21	#include <linux/kernel.h>
22	#include <linux/syscalls.h>
23	#include <linux/fs.h>
24	#include <linux/mm.h>
25	#include <linux/percpu.h>
26	#include <linux/slab.h>
27	#include <linux/capability.h>
28	#include <linux/blkdev.h>
29	#include <linux/file.h>
30	#include <linux/quotaops.h>
31	#include <linux/highmem.h>
32	#include <linux/module.h>
33	#include <linux/writeback.h>
34	#include <linux/hash.h>
35	#include <linux/suspend.h>
36	#include <linux/buffer_head.h>
37	#include <linux/task_io_accounting_ops.h>
38	#include <linux/bio.h>
39	#include <linux/notifier.h>
40	#include <linux/cpu.h>
41	#include <linux/bitops.h>
42	#include <linux/mpage.h>
43	#include <linux/bit_spinlock.h>
44
45	static int fsync_buffers_list(spinlock_t lock, struct list_head list);
46
47	#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
48
49	inline void
50	init_buffer(struct buffer_head bh, bh_end_io_t handler, void *private)
51	{
52	bh->b_end_io = handler;
53	bh->b_private = private;
54	}
55	EXPORT_SYMBOL(init_buffer);
56
57	static int sync_buffer(void *word)
58	{
59	struct block_device *bd;
60	struct buffer_head *bh
61	= container_of(word, struct buffer_head, b_state);
62
63	smp_mb();
64	bd = bh->b_bdev;
65	if (bd)
66	blk_run_address_space(bd->bd_inode->i_mapping);
67	io_schedule();
68	return 0;
69	}
70
71	void __lock_buffer(struct buffer_head *bh)
72	{
73	wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
74	TASK_UNINTERRUPTIBLE);
75	}
76	EXPORT_SYMBOL(__lock_buffer);
77
78	void unlock_buffer(struct buffer_head *bh)
79	{
80	clear_bit_unlock(BH_Lock, &bh->b_state);
81	smp_mb__after_clear_bit();
82	wake_up_bit(&bh->b_state, BH_Lock);
83	}
84	EXPORT_SYMBOL(unlock_buffer);
85
86	/*
87	* Block until a buffer comes unlocked. This doesn't stop it
88	* from becoming locked again - you have to lock it yourself
89	* if you want to preserve its state.
90	*/
91	void __wait_on_buffer(struct buffer_head * bh)
92	{
93	wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
94	}
95	EXPORT_SYMBOL(__wait_on_buffer);
96
97	static void
98	__clear_page_buffers(struct page *page)
99	{
100	ClearPagePrivate(page);
101	set_page_private(page, 0);
102	page_cache_release(page);
103	}
104
105
106	static int quiet_error(struct buffer_head *bh)
107	{
108	if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
109	return 0;
110	return 1;
111	}
112
113
114	static void buffer_io_error(struct buffer_head *bh)
115	{
116	char b[BDEVNAME_SIZE];
117	printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
118	bdevname(bh->b_bdev, b),
119	(unsigned long long)bh->b_blocknr);
120	}
121
122	/*
123	* End-of-IO handler helper function which does not touch the bh after
124	* unlocking it.
125	* Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
126	* a race there is benign: unlock_buffer() only use the bh's address for
127	* hashing after unlocking the buffer, so it doesn't actually touch the bh
128	* itself.
129	*/
130	static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
131	{
132	if (uptodate) {
133	set_buffer_uptodate(bh);
134	} else {
135	/* This happens, due to failed READA attempts. */
136	clear_buffer_uptodate(bh);
137	}
138	unlock_buffer(bh);
139	}
140
141	/*
142	* Default synchronous end-of-IO handler.. Just mark it up-to-date and
143	* unlock the buffer. This is what ll_rw_block uses too.
144	*/
145	void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
146	{
147	__end_buffer_read_notouch(bh, uptodate);
148	put_bh(bh);
149	}
150	EXPORT_SYMBOL(end_buffer_read_sync);
151
152	void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
153	{
154	char b[BDEVNAME_SIZE];
155
156	if (uptodate) {
157	set_buffer_uptodate(bh);
158	} else {
159	if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
160	buffer_io_error(bh);
161	printk(KERN_WARNING "lost page write due to "
162	"I/O error on %s\n",
163	bdevname(bh->b_bdev, b));
164	}
165	set_buffer_write_io_error(bh);
166	clear_buffer_uptodate(bh);
167	}
168	unlock_buffer(bh);
169	put_bh(bh);
170	}
171	EXPORT_SYMBOL(end_buffer_write_sync);
172
173	/*
174	* Various filesystems appear to want __find_get_block to be non-blocking.
175	* But it's the page lock which protects the buffers. To get around this,
176	* we get exclusion from try_to_free_buffers with the blockdev mapping's
177	* private_lock.
178	*
179	* Hack idea: for the blockdev mapping, i_bufferlist_lock contention
180	* may be quite high. This code could TryLock the page, and if that
181	* succeeds, there is no need to take private_lock. (But if
182	* private_lock is contended then so is mapping->tree_lock).
183	*/
184	static struct buffer_head *
185	__find_get_block_slow(struct block_device *bdev, sector_t block)
186	{
187	struct inode *bd_inode = bdev->bd_inode;
188	struct address_space *bd_mapping = bd_inode->i_mapping;
189	struct buffer_head *ret = NULL;
190	pgoff_t index;
191	struct buffer_head *bh;
192	struct buffer_head *head;
193	struct page *page;
194	int all_mapped = 1;
195
196	index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
197	page = find_get_page(bd_mapping, index);
198	if (!page)
199	goto out;
200
201	spin_lock(&bd_mapping->private_lock);
202	if (!page_has_buffers(page))
203	goto out_unlock;
204	head = page_buffers(page);
205	bh = head;
206	do {
207	if (!buffer_mapped(bh))
208	all_mapped = 0;
209	else if (bh->b_blocknr == block) {
210	ret = bh;
211	get_bh(bh);
212	goto out_unlock;
213	}
214	bh = bh->b_this_page;
215	} while (bh != head);
216
217	/* we might be here because some of the buffers on this page are
218	* not mapped. This is due to various races between
219	* file io on the block device and getblk. It gets dealt with
220	* elsewhere, don't buffer_error if we had some unmapped buffers
221	*/
222	if (all_mapped) {
223	printk("__find_get_block_slow() failed. "
224	"block=%llu, b_blocknr=%llu\n",
225	(unsigned long long)block,
226	(unsigned long long)bh->b_blocknr);
227	printk("b_state=0x%08lx, b_size=%zu\n",
228	bh->b_state, bh->b_size);
229	printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
230	}
231	out_unlock:
232	spin_unlock(&bd_mapping->private_lock);
233	page_cache_release(page);
234	out:
235	return ret;
236	}
237
238	/* If invalidate_buffers() will trash dirty buffers, it means some kind
239	of fs corruption is going on. Trashing dirty data always imply losing
240	information that was supposed to be just stored on the physical layer
241	by the user.
242
243	Thus invalidate_buffers in general usage is not allwowed to trash
244	dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
245	be preserved. These buffers are simply skipped.
246
247	We also skip buffers which are still in use. For example this can
248	happen if a userspace program is reading the block device.
249
250	NOTE: In the case where the user removed a removable-media-disk even if
251	there's still dirty data not synced on disk (due a bug in the device driver
252	or due an error of the user), by not destroying the dirty buffers we could
253	generate corruption also on the next media inserted, thus a parameter is
254	necessary to handle this case in the most safe way possible (trying
255	to not corrupt also the new disk inserted with the data belonging to
256	the old now corrupted disk). Also for the ramdisk the natural thing
257	to do in order to release the ramdisk memory is to destroy dirty buffers.
258
259	These are two special cases. Normal usage imply the device driver
260	to issue a sync on the device (without waiting I/O completion) and
261	then an invalidate_buffers call that doesn't trash dirty buffers.
262
263	For handling cache coherency with the blkdev pagecache the 'update' case
264	is been introduced. It is needed to re-read from disk any pinned
265	buffer. NOTE: re-reading from disk is destructive so we can do it only
266	when we assume nobody is changing the buffercache under our I/O and when
267	we think the disk contains more recent information than the buffercache.
268	The update == 1 pass marks the buffers we need to update, the update == 2
269	pass does the actual I/O. */
270	void invalidate_bdev(struct block_device *bdev)
271	{
272	struct address_space *mapping = bdev->bd_inode->i_mapping;
273
274	if (mapping->nrpages == 0)
275	return;
276
277	invalidate_bh_lrus();
278	lru_add_drain_all(); /* make sure all lru add caches are flushed */
279	invalidate_mapping_pages(mapping, 0, -1);
280	}
281	EXPORT_SYMBOL(invalidate_bdev);
282
283	/*
284	* Kick the writeback threads then try to free up some ZONE_NORMAL memory.
285	*/
286	static void free_more_memory(void)
287	{
288	struct zone *zone;
289	int nid;
290
291	wakeup_flusher_threads(1024);
292	yield();
293
294	for_each_online_node(nid) {
295	(void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
296	gfp_zone(GFP_NOFS), NULL,
297	&zone);
298	if (zone)
299	try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
300	GFP_NOFS, NULL);
301	}
302	}
303
304	/*
305	* I/O completion handler for block_read_full_page() - pages
306	* which come unlocked at the end of I/O.
307	*/
308	static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
309	{
310	unsigned long flags;
311	struct buffer_head *first;
312	struct buffer_head *tmp;
313	struct page *page;
314	int page_uptodate = 1;
315
316	BUG_ON(!buffer_async_read(bh));
317
318	page = bh->b_page;
319	if (uptodate) {
320	set_buffer_uptodate(bh);
321	} else {
322	clear_buffer_uptodate(bh);
323	if (!quiet_error(bh))
324	buffer_io_error(bh);
325	SetPageError(page);
326	}
327
328	/*
329	* Be _very_ careful from here on. Bad things can happen if
330	* two buffer heads end IO at almost the same time and both
331	* decide that the page is now completely done.
332	*/
333	first = page_buffers(page);
334	local_irq_save(flags);
335	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336	clear_buffer_async_read(bh);
337	unlock_buffer(bh);
338	tmp = bh;
339	do {
340	if (!buffer_uptodate(tmp))
341	page_uptodate = 0;
342	if (buffer_async_read(tmp)) {
343	BUG_ON(!buffer_locked(tmp));
344	goto still_busy;
345	}
346	tmp = tmp->b_this_page;
347	} while (tmp != bh);
348	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
349	local_irq_restore(flags);
350
351	/*
352	* If none of the buffers had errors and they are all
353	* uptodate then we can set the page uptodate.
354	*/
355	if (page_uptodate && !PageError(page))
356	SetPageUptodate(page);
357	unlock_page(page);
358	return;
359
360	still_busy:
361	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
362	local_irq_restore(flags);
363	return;
364	}
365
366	/*
367	* Completion handler for block_write_full_page() - pages which are unlocked
368	* during I/O, and which have PageWriteback cleared upon I/O completion.
369	*/
370	void end_buffer_async_write(struct buffer_head *bh, int uptodate)
371	{
372	char b[BDEVNAME_SIZE];
373	unsigned long flags;
374	struct buffer_head *first;
375	struct buffer_head *tmp;
376	struct page *page;
377
378	BUG_ON(!buffer_async_write(bh));
379
380	page = bh->b_page;
381	if (uptodate) {
382	set_buffer_uptodate(bh);
383	} else {
384	if (!quiet_error(bh)) {
385	buffer_io_error(bh);
386	printk(KERN_WARNING "lost page write due to "
387	"I/O error on %s\n",
388	bdevname(bh->b_bdev, b));
389	}
390	set_bit(AS_EIO, &page->mapping->flags);
391	set_buffer_write_io_error(bh);
392	clear_buffer_uptodate(bh);
393	SetPageError(page);
394	}
395
396	first = page_buffers(page);
397	local_irq_save(flags);
398	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
399
400	clear_buffer_async_write(bh);
401	unlock_buffer(bh);
402	tmp = bh->b_this_page;
403	while (tmp != bh) {
404	if (buffer_async_write(tmp)) {
405	BUG_ON(!buffer_locked(tmp));
406	goto still_busy;
407	}
408	tmp = tmp->b_this_page;
409	}
410	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
411	local_irq_restore(flags);
412	end_page_writeback(page);
413	return;
414
415	still_busy:
416	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
417	local_irq_restore(flags);
418	return;
419	}
420	EXPORT_SYMBOL(end_buffer_async_write);
421
422	/*
423	* If a page's buffers are under async readin (end_buffer_async_read
424	* completion) then there is a possibility that another thread of
425	* control could lock one of the buffers after it has completed
426	* but while some of the other buffers have not completed. This
427	* locked buffer would confuse end_buffer_async_read() into not unlocking
428	* the page. So the absence of BH_Async_Read tells end_buffer_async_read()
429	* that this buffer is not under async I/O.
430	*
431	* The page comes unlocked when it has no locked buffer_async buffers
432	* left.
433	*
434	* PageLocked prevents anyone starting new async I/O reads any of
435	* the buffers.
436	*
437	* PageWriteback is used to prevent simultaneous writeout of the same
438	* page.
439	*
440	* PageLocked prevents anyone from starting writeback of a page which is
441	* under read I/O (PageWriteback is only ever set against a locked page).
442	*/
443	static void mark_buffer_async_read(struct buffer_head *bh)
444	{
445	bh->b_end_io = end_buffer_async_read;
446	set_buffer_async_read(bh);
447	}
448
449	static void mark_buffer_async_write_endio(struct buffer_head *bh,
450	bh_end_io_t *handler)
451	{
452	bh->b_end_io = handler;
453	set_buffer_async_write(bh);
454	}
455
456	void mark_buffer_async_write(struct buffer_head *bh)
457	{
458	mark_buffer_async_write_endio(bh, end_buffer_async_write);
459	}
460	EXPORT_SYMBOL(mark_buffer_async_write);
461
462
463	/*
464	* fs/buffer.c contains helper functions for buffer-backed address space's
465	* fsync functions. A common requirement for buffer-based filesystems is
466	* that certain data from the backing blockdev needs to be written out for
467	* a successful fsync(). For example, ext2 indirect blocks need to be
468	* written back and waited upon before fsync() returns.
469	*
470	* The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
471	* inode_has_buffers() and invalidate_inode_buffers() are provided for the
472	* management of a list of dependent buffers at ->i_mapping->private_list.
473	*
474	* Locking is a little subtle: try_to_free_buffers() will remove buffers
475	* from their controlling inode's queue when they are being freed. But
476	* try_to_free_buffers() will be operating against the blockdev mapping
477	* at the time, not against the S_ISREG file which depends on those buffers.
478	* So the locking for private_list is via the private_lock in the address_space
479	* which backs the buffers. Which is different from the address_space
480	* against which the buffers are listed. So for a particular address_space,
481	* mapping->private_lock does not protect mapping->private_list! In fact,
482	* mapping->private_list will always be protected by the backing blockdev's
483	* ->private_lock.
484	*
485	* Which introduces a requirement: all buffers on an address_space's
486	* ->private_list must be from the same address_space: the blockdev's.
487	*
488	* address_spaces which do not place buffers at ->private_list via these
489	* utility functions are free to use private_lock and private_list for
490	* whatever they want. The only requirement is that list_empty(private_list)
491	* be true at clear_inode() time.
492	*
493	* FIXME: clear_inode should not call invalidate_inode_buffers(). The
494	* filesystems should do that. invalidate_inode_buffers() should just go
495	* BUG_ON(!list_empty).
496	*
497	* FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
498	* take an address_space, not an inode. And it should be called
499	* mark_buffer_dirty_fsync() to clearly define why those buffers are being
500	* queued up.
501	*
502	* FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
503	* list if it is already on a list. Because if the buffer is on a list,
504	* it must already be on the right one. If not, the filesystem is being
505	* silly. This will save a ton of locking. But first we have to ensure
506	* that buffers are taken off the old inode's list when they are freed
507	* (presumably in truncate). That requires careful auditing of all
508	* filesystems (do it inside bforget()). It could also be done by bringing
509	* b_inode back.
510	*/
511
512	/*
513	* The buffer's backing address_space's private_lock must be held
514	*/
515	static void __remove_assoc_queue(struct buffer_head *bh)
516	{
517	list_del_init(&bh->b_assoc_buffers);
518	WARN_ON(!bh->b_assoc_map);
519	if (buffer_write_io_error(bh))
520	set_bit(AS_EIO, &bh->b_assoc_map->flags);
521	bh->b_assoc_map = NULL;
522	}
523
524	int inode_has_buffers(struct inode *inode)
525	{
526	return !list_empty(&inode->i_data.private_list);
527	}
528
529	/*
530	* osync is designed to support O_SYNC io. It waits synchronously for
531	* all already-submitted IO to complete, but does not queue any new
532	* writes to the disk.
533	*
534	* To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
535	* you dirty the buffers, and then use osync_inode_buffers to wait for
536	* completion. Any other dirty buffers which are not yet queued for
537	* write will not be flushed to disk by the osync.
538	*/
539	static int osync_buffers_list(spinlock_t lock, struct list_head list)
540	{
541	struct buffer_head *bh;
542	struct list_head *p;
543	int err = 0;
544
545	spin_lock(lock);
546	repeat:
547	list_for_each_prev(p, list) {
548	bh = BH_ENTRY(p);
549	if (buffer_locked(bh)) {
550	get_bh(bh);
551	spin_unlock(lock);
552	wait_on_buffer(bh);
553	if (!buffer_uptodate(bh))
554	err = -EIO;
555	brelse(bh);
556	spin_lock(lock);
557	goto repeat;
558	}
559	}
560	spin_unlock(lock);
561	return err;
562	}
563
564	static void do_thaw_one(struct super_block sb, void unused)
565	{
566	char b[BDEVNAME_SIZE];
567	while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
568	printk(KERN_WARNING "Emergency Thaw on %s\n",
569	bdevname(sb->s_bdev, b));
570	}
571
572	static void do_thaw_all(struct work_struct *work)
573	{
574	iterate_supers(do_thaw_one, NULL);
575	kfree(work);
576	printk(KERN_WARNING "Emergency Thaw complete\n");
577	}
578
579	/**
580	* emergency_thaw_all -- forcibly thaw every frozen filesystem
581	*
582	* Used for emergency unfreeze of all filesystems via SysRq
583	*/
584	void emergency_thaw_all(void)
585	{
586	struct work_struct *work;
587
588	work = kmalloc(sizeof(*work), GFP_ATOMIC);
589	if (work) {
590	INIT_WORK(work, do_thaw_all);
591	schedule_work(work);
592	}
593	}
594
595	/**
596	* sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
597	* @mapping: the mapping which wants those buffers written
598	*
599	* Starts I/O against the buffers at mapping->private_list, and waits upon
600	* that I/O.
601	*
602	* Basically, this is a convenience function for fsync().
603	* @mapping is a file or directory which needs those buffers to be written for
604	* a successful fsync().
605	*/
606	int sync_mapping_buffers(struct address_space *mapping)
607	{
608	struct address_space *buffer_mapping = mapping->assoc_mapping;
609
610	if (buffer_mapping == NULL \|\| list_empty(&mapping->private_list))
611	return 0;
612
613	return fsync_buffers_list(&buffer_mapping->private_lock,
614	&mapping->private_list);
615	}
616	EXPORT_SYMBOL(sync_mapping_buffers);
617
618	/*
619	* Called when we've recently written block `bblock', and it is known that
620	* `bblock' was for a buffer_boundary() buffer. This means that the block at
621	* `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
622	* dirty, schedule it for IO. So that indirects merge nicely with their data.
623	*/
624	void write_boundary_block(struct block_device *bdev,
625	sector_t bblock, unsigned blocksize)
626	{
627	struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
628	if (bh) {
629	if (buffer_dirty(bh))
630	ll_rw_block(WRITE, 1, &bh);
631	put_bh(bh);
632	}
633	}
634
635	void mark_buffer_dirty_inode(struct buffer_head bh, struct inode inode)
636	{
637	struct address_space *mapping = inode->i_mapping;
638	struct address_space *buffer_mapping = bh->b_page->mapping;
639
640	mark_buffer_dirty(bh);
641	if (!mapping->assoc_mapping) {
642	mapping->assoc_mapping = buffer_mapping;
643	} else {
644	BUG_ON(mapping->assoc_mapping != buffer_mapping);
645	}
646	if (!bh->b_assoc_map) {
647	spin_lock(&buffer_mapping->private_lock);
648	list_move_tail(&bh->b_assoc_buffers,
649	&mapping->private_list);
650	bh->b_assoc_map = mapping;
651	spin_unlock(&buffer_mapping->private_lock);
652	}
653	}
654	EXPORT_SYMBOL(mark_buffer_dirty_inode);
655
656	/*
657	* Mark the page dirty, and set it dirty in the radix tree, and mark the inode
658	* dirty.
659	*
660	* If warn is true, then emit a warning if the page is not uptodate and has
661	* not been truncated.
662	*/
663	static void __set_page_dirty(struct page *page,
664	struct address_space *mapping, int warn)
665	{
666	spin_lock_irq(&mapping->tree_lock);
667	if (page->mapping) { /* Race with truncate? */
668	WARN_ON_ONCE(warn && !PageUptodate(page));
669	account_page_dirtied(page, mapping);
670	radix_tree_tag_set(&mapping->page_tree,
671	page_index(page), PAGECACHE_TAG_DIRTY);
672	}
673	spin_unlock_irq(&mapping->tree_lock);
674	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
675	}
676
677	/*
678	* Add a page to the dirty page list.
679	*
680	* It is a sad fact of life that this function is called from several places
681	* deeply under spinlocking. It may not sleep.
682	*
683	* If the page has buffers, the uptodate buffers are set dirty, to preserve
684	* dirty-state coherency between the page and the buffers. It the page does
685	* not have buffers then when they are later attached they will all be set
686	* dirty.
687	*
688	* The buffers are dirtied before the page is dirtied. There's a small race
689	* window in which a writepage caller may see the page cleanness but not the
690	* buffer dirtiness. That's fine. If this code were to set the page dirty
691	* before the buffers, a concurrent writepage caller could clear the page dirty
692	* bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
693	* page on the dirty page list.
694	*
695	* We use private_lock to lock against try_to_free_buffers while using the
696	* page's buffer list. Also use this to protect against clean buffers being
697	* added to the page after it was set dirty.
698	*
699	* FIXME: may need to call ->reservepage here as well. That's rather up to the
700	* address_space though.
701	*/
702	int __set_page_dirty_buffers(struct page *page)
703	{
704	int newly_dirty;
705	struct address_space *mapping = page_mapping(page);
706
707	if (unlikely(!mapping))
708	return !TestSetPageDirty(page);
709
710	spin_lock(&mapping->private_lock);
711	if (page_has_buffers(page)) {
712	struct buffer_head *head = page_buffers(page);
713	struct buffer_head *bh = head;
714
715	do {
716	set_buffer_dirty(bh);
717	bh = bh->b_this_page;
718	} while (bh != head);
719	}
720	newly_dirty = !TestSetPageDirty(page);
721	spin_unlock(&mapping->private_lock);
722
723	if (newly_dirty)
724	__set_page_dirty(page, mapping, 1);
725	return newly_dirty;
726	}
727	EXPORT_SYMBOL(__set_page_dirty_buffers);
728
729	/*
730	* Write out and wait upon a list of buffers.
731	*
732	* We have conflicting pressures: we want to make sure that all
733	* initially dirty buffers get waited on, but that any subsequently
734	* dirtied buffers don't. After all, we don't want fsync to last
735	* forever if somebody is actively writing to the file.
736	*
737	* Do this in two main stages: first we copy dirty buffers to a
738	* temporary inode list, queueing the writes as we go. Then we clean
739	* up, waiting for those writes to complete.
740	*
741	* During this second stage, any subsequent updates to the file may end
742	* up refiling the buffer on the original inode's dirty list again, so
743	* there is a chance we will end up with a buffer queued for write but
744	* not yet completed on that list. So, as a final cleanup we go through
745	* the osync code to catch these locked, dirty buffers without requeuing
746	* any newly dirty buffers for write.
747	*/
748	static int fsync_buffers_list(spinlock_t lock, struct list_head list)
749	{
750	struct buffer_head *bh;
751	struct list_head tmp;
752	struct address_space mapping, prev_mapping = NULL;
753	int err = 0, err2;
754
755	INIT_LIST_HEAD(&tmp);
756
757	spin_lock(lock);
758	while (!list_empty(list)) {
759	bh = BH_ENTRY(list->next);
760	mapping = bh->b_assoc_map;
761	__remove_assoc_queue(bh);
762	/* Avoid race with mark_buffer_dirty_inode() which does
763	* a lockless check and we rely on seeing the dirty bit */
764	smp_mb();
765	if (buffer_dirty(bh) \|\| buffer_locked(bh)) {
766	list_add(&bh->b_assoc_buffers, &tmp);
767	bh->b_assoc_map = mapping;
768	if (buffer_dirty(bh)) {
769	get_bh(bh);
770	spin_unlock(lock);
771	/*
772	* Ensure any pending I/O completes so that
773	* ll_rw_block() actually writes the current
774	* contents - it is a noop if I/O is still in
775	* flight on potentially older contents.
776	*/
777	ll_rw_block(SWRITE_SYNC_PLUG, 1, &bh);
778
779	/*
780	* Kick off IO for the previous mapping. Note
781	* that we will not run the very last mapping,
782	* wait_on_buffer() will do that for us
783	* through sync_buffer().
784	*/
785	if (prev_mapping && prev_mapping != mapping)
786	blk_run_address_space(prev_mapping);
787	prev_mapping = mapping;
788
789	brelse(bh);
790	spin_lock(lock);
791	}
792	}
793	}
794
795	while (!list_empty(&tmp)) {
796	bh = BH_ENTRY(tmp.prev);
797	get_bh(bh);
798	mapping = bh->b_assoc_map;
799	__remove_assoc_queue(bh);
800	/* Avoid race with mark_buffer_dirty_inode() which does
801	* a lockless check and we rely on seeing the dirty bit */
802	smp_mb();
803	if (buffer_dirty(bh)) {
804	list_add(&bh->b_assoc_buffers,
805	&mapping->private_list);
806	bh->b_assoc_map = mapping;
807	}
808	spin_unlock(lock);
809	wait_on_buffer(bh);
810	if (!buffer_uptodate(bh))
811	err = -EIO;
812	brelse(bh);
813	spin_lock(lock);
814	}
815
816	spin_unlock(lock);
817	err2 = osync_buffers_list(lock, list);
818	if (err)
819	return err;
820	else
821	return err2;
822	}
823
824	/*
825	* Invalidate any and all dirty buffers on a given inode. We are
826	* probably unmounting the fs, but that doesn't mean we have already
827	* done a sync(). Just drop the buffers from the inode list.
828	*
829	* NOTE: we take the inode's blockdev's mapping's private_lock. Which
830	* assumes that all the buffers are against the blockdev. Not true
831	* for reiserfs.
832	*/
833	void invalidate_inode_buffers(struct inode *inode)
834	{
835	if (inode_has_buffers(inode)) {
836	struct address_space *mapping = &inode->i_data;
837	struct list_head *list = &mapping->private_list;
838	struct address_space *buffer_mapping = mapping->assoc_mapping;
839
840	spin_lock(&buffer_mapping->private_lock);
841	while (!list_empty(list))
842	__remove_assoc_queue(BH_ENTRY(list->next));
843	spin_unlock(&buffer_mapping->private_lock);
844	}
845	}
846	EXPORT_SYMBOL(invalidate_inode_buffers);
847
848	/*
849	* Remove any clean buffers from the inode's buffer list. This is called
850	* when we're trying to free the inode itself. Those buffers can pin it.
851	*
852	* Returns true if all buffers were removed.
853	*/
854	int remove_inode_buffers(struct inode *inode)
855	{
856	int ret = 1;
857
858	if (inode_has_buffers(inode)) {
859	struct address_space *mapping = &inode->i_data;
860	struct list_head *list = &mapping->private_list;
861	struct address_space *buffer_mapping = mapping->assoc_mapping;
862
863	spin_lock(&buffer_mapping->private_lock);
864	while (!list_empty(list)) {
865	struct buffer_head *bh = BH_ENTRY(list->next);
866	if (buffer_dirty(bh)) {
867	ret = 0;
868	break;
869	}
870	__remove_assoc_queue(bh);
871	}
872	spin_unlock(&buffer_mapping->private_lock);
873	}
874	return ret;
875	}
876
877	/*
878	* Create the appropriate buffers when given a page for data area and
879	* the size of each buffer.. Use the bh->b_this_page linked list to
880	* follow the buffers created. Return NULL if unable to create more
881	* buffers.
882	*
883	* The retry flag is used to differentiate async IO (paging, swapping)
884	* which may not fail from ordinary buffer allocations.
885	*/
886	struct buffer_head alloc_page_buffers(struct page page, unsigned long size,
887	int retry)
888	{
889	struct buffer_head bh, head;
890	long offset;
891
892	try_again:
893	head = NULL;
894	offset = PAGE_SIZE;
895	while ((offset -= size) >= 0) {
896	bh = alloc_buffer_head(GFP_NOFS);
897	if (!bh)
898	goto no_grow;
899
900	bh->b_bdev = NULL;
901	bh->b_this_page = head;
902	bh->b_blocknr = -1;
903	head = bh;
904
905	bh->b_state = 0;
906	atomic_set(&bh->b_count, 0);
907	bh->b_private = NULL;
908	bh->b_size = size;
909
910	/* Link the buffer to its page */
911	set_bh_page(bh, page, offset);
912
913	init_buffer(bh, NULL, NULL);
914	}
915	return head;
916	/*
917	* In case anything failed, we just free everything we got.
918	*/
919	no_grow:
920	if (head) {
921	do {
922	bh = head;
923	head = head->b_this_page;
924	free_buffer_head(bh);
925	} while (head);
926	}
927
928	/*
929	* Return failure for non-async IO requests. Async IO requests
930	* are not allowed to fail, so we have to wait until buffer heads
931	* become available. But we don't want tasks sleeping with
932	* partially complete buffers, so all were released above.
933	*/
934	if (!retry)
935	return NULL;
936
937	/* We're _really_ low on memory. Now we just
938	* wait for old buffer heads to become free due to
939	* finishing IO. Since this is an async request and
940	* the reserve list is empty, we're sure there are
941	* async buffer heads in use.
942	*/
943	free_more_memory();
944	goto try_again;
945	}
946	EXPORT_SYMBOL_GPL(alloc_page_buffers);
947
948	static inline void
949	link_dev_buffers(struct page page, struct buffer_head head)
950	{
951	struct buffer_head bh, tail;
952
953	bh = head;
954	do {
955	tail = bh;
956	bh = bh->b_this_page;
957	} while (bh);
958	tail->b_this_page = head;
959	attach_page_buffers(page, head);
960	}
961
962	/*
963	* Initialise the state of a blockdev page's buffers.
964	*/
965	static void
966	init_page_buffers(struct page page, struct block_device bdev,
967	sector_t block, int size)
968	{
969	struct buffer_head *head = page_buffers(page);
970	struct buffer_head *bh = head;
971	int uptodate = PageUptodate(page);
972
973	do {
974	if (!buffer_mapped(bh)) {
975	init_buffer(bh, NULL, NULL);
976	bh->b_bdev = bdev;
977	bh->b_blocknr = block;
978	if (uptodate)
979	set_buffer_uptodate(bh);
980	set_buffer_mapped(bh);
981	}
982	block++;
983	bh = bh->b_this_page;
984	} while (bh != head);
985	}
986
987	/*
988	* Create the page-cache page that contains the requested block.
989	*
990	* This is user purely for blockdev mappings.
991	*/
992	static struct page *
993	grow_dev_page(struct block_device *bdev, sector_t block,
994	pgoff_t index, int size)
995	{
996	struct inode *inode = bdev->bd_inode;
997	struct page *page;
998	struct buffer_head *bh;
999
1000	page = find_or_create_page(inode->i_mapping, index,
1001	(mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)\|__GFP_MOVABLE);
1002	if (!page)
1003	return NULL;
1004
1005	BUG_ON(!PageLocked(page));
1006
1007	if (page_has_buffers(page)) {
1008	bh = page_buffers(page);
1009	if (bh->b_size == size) {
1010	init_page_buffers(page, bdev, block, size);
1011	return page;
1012	}
1013	if (!try_to_free_buffers(page))
1014	goto failed;
1015	}
1016
1017	/*
1018	* Allocate some buffers for this page
1019	*/
1020	bh = alloc_page_buffers(page, size, 0);
1021	if (!bh)
1022	goto failed;
1023
1024	/*
1025	* Link the page to the buffers and initialise them. Take the
1026	* lock to be atomic wrt __find_get_block(), which does not
1027	* run under the page lock.
1028	*/
1029	spin_lock(&inode->i_mapping->private_lock);
1030	link_dev_buffers(page, bh);
1031	init_page_buffers(page, bdev, block, size);
1032	spin_unlock(&inode->i_mapping->private_lock);
1033	return page;
1034
1035	failed:
1036	BUG();
1037	unlock_page(page);
1038	page_cache_release(page);
1039	return NULL;
1040	}
1041
1042	/*
1043	* Create buffers for the specified block device block's page. If
1044	* that page was dirty, the buffers are set dirty also.
1045	*/
1046	static int
1047	grow_buffers(struct block_device *bdev, sector_t block, int size)
1048	{
1049	struct page *page;
1050	pgoff_t index;
1051	int sizebits;
1052
1053	sizebits = -1;
1054	do {
1055	sizebits++;
1056	} while ((size << sizebits) < PAGE_SIZE);
1057
1058	index = block >> sizebits;
1059
1060	/*
1061	* Check for a block which wants to lie outside our maximum possible
1062	* pagecache index. (this comparison is done using sector_t types).
1063	*/
1064	if (unlikely(index != block >> sizebits)) {
1065	char b[BDEVNAME_SIZE];
1066
1067	printk(KERN_ERR "%s: requested out-of-range block %llu for "
1068	"device %s\n",
1069	__func__, (unsigned long long)block,
1070	bdevname(bdev, b));
1071	return -EIO;
1072	}
1073	block = index << sizebits;
1074	/* Create a page with the proper size buffers.. */
1075	page = grow_dev_page(bdev, block, index, size);
1076	if (!page)
1077	return 0;
1078	unlock_page(page);
1079	page_cache_release(page);
1080	return 1;
1081	}
1082
1083	static struct buffer_head *
1084	__getblk_slow(struct block_device *bdev, sector_t block, int size)
1085	{
1086	/* Size must be multiple of hard sectorsize */
1087	if (unlikely(size & (bdev_logical_block_size(bdev)-1) \|\|
1088	(size < 512 \|\| size > PAGE_SIZE))) {
1089	printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1090	size);
1091	printk(KERN_ERR "logical block size: %d\n",
1092	bdev_logical_block_size(bdev));
1093
1094	dump_stack();
1095	return NULL;
1096	}
1097
1098	for (;;) {
1099	struct buffer_head * bh;
1100	int ret;
1101
1102	bh = __find_get_block(bdev, block, size);
1103	if (bh)
1104	return bh;
1105
1106	ret = grow_buffers(bdev, block, size);
1107	if (ret < 0)
1108	return NULL;
1109	if (ret == 0)
1110	free_more_memory();
1111	}
1112	}
1113
1114	/*
1115	* The relationship between dirty buffers and dirty pages:
1116	*
1117	* Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118	* the page is tagged dirty in its radix tree.
1119	*
1120	* At all times, the dirtiness of the buffers represents the dirtiness of
1121	* subsections of the page. If the page has buffers, the page dirty bit is
1122	* merely a hint about the true dirty state.
1123	*
1124	* When a page is set dirty in its entirety, all its buffers are marked dirty
1125	* (if the page has buffers).
1126	*
1127	* When a buffer is marked dirty, its page is dirtied, but the page's other
1128	* buffers are not.
1129	*
1130	* Also. When blockdev buffers are explicitly read with bread(), they
1131	* individually become uptodate. But their backing page remains not
1132	* uptodate - even if all of its buffers are uptodate. A subsequent
1133	* block_read_full_page() against that page will discover all the uptodate
1134	* buffers, will set the page uptodate and will perform no I/O.
1135	*/
1136
1137	/**
1138	* mark_buffer_dirty - mark a buffer_head as needing writeout
1139	* @bh: the buffer_head to mark dirty
1140	*
1141	* mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142	* backing page dirty, then tag the page as dirty in its address_space's radix
1143	* tree and then attach the address_space's inode to its superblock's dirty
1144	* inode list.
1145	*
1146	* mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1147	* mapping->tree_lock and the global inode_lock.
1148	*/
1149	void mark_buffer_dirty(struct buffer_head *bh)
1150	{
1151	WARN_ON_ONCE(!buffer_uptodate(bh));
1152
1153	/*
1154	* Very carefully optimize the it-is-already-dirty case.
1155	*
1156	* Don't let the final "is it dirty" escape to before we
1157	* perhaps modified the buffer.
1158	*/
1159	if (buffer_dirty(bh)) {
1160	smp_mb();
1161	if (buffer_dirty(bh))
1162	return;
1163	}
1164
1165	if (!test_set_buffer_dirty(bh)) {
1166	struct page *page = bh->b_page;
1167	if (!TestSetPageDirty(page)) {
1168	struct address_space *mapping = page_mapping(page);
1169	if (mapping)
1170	__set_page_dirty(page, mapping, 0);
1171	}
1172	}
1173	}
1174	EXPORT_SYMBOL(mark_buffer_dirty);
1175
1176	/*
1177	* Decrement a buffer_head's reference count. If all buffers against a page
1178	* have zero reference count, are clean and unlocked, and if the page is clean
1179	* and unlocked then try_to_free_buffers() may strip the buffers from the page
1180	* in preparation for freeing it (sometimes, rarely, buffers are removed from
1181	* a page but it ends up not being freed, and buffers may later be reattached).
1182	*/
1183	void __brelse(struct buffer_head * buf)
1184	{
1185	if (atomic_read(&buf->b_count)) {
1186	put_bh(buf);
1187	return;
1188	}
1189	WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1190	}
1191	EXPORT_SYMBOL(__brelse);
1192
1193	/*
1194	* bforget() is like brelse(), except it discards any
1195	* potentially dirty data.
1196	*/
1197	void __bforget(struct buffer_head *bh)
1198	{
1199	clear_buffer_dirty(bh);
1200	if (bh->b_assoc_map) {
1201	struct address_space *buffer_mapping = bh->b_page->mapping;
1202
1203	spin_lock(&buffer_mapping->private_lock);
1204	list_del_init(&bh->b_assoc_buffers);
1205	bh->b_assoc_map = NULL;
1206	spin_unlock(&buffer_mapping->private_lock);
1207	}
1208	__brelse(bh);
1209	}
1210	EXPORT_SYMBOL(__bforget);
1211
1212	static struct buffer_head __bread_slow(struct buffer_head bh)
1213	{
1214	lock_buffer(bh);
1215	if (buffer_uptodate(bh)) {
1216	unlock_buffer(bh);
1217	return bh;
1218	} else {
1219	get_bh(bh);
1220	bh->b_end_io = end_buffer_read_sync;
1221	submit_bh(READ, bh);
1222	wait_on_buffer(bh);
1223	if (buffer_uptodate(bh))
1224	return bh;
1225	}
1226	brelse(bh);
1227	return NULL;
1228	}
1229
1230	/*
1231	* Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1232	* The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1233	* refcount elevated by one when they're in an LRU. A buffer can only appear
1234	* once in a particular CPU's LRU. A single buffer can be present in multiple
1235	* CPU's LRUs at the same time.
1236	*
1237	* This is a transparent caching front-end to sb_bread(), sb_getblk() and
1238	* sb_find_get_block().
1239	*
1240	* The LRUs themselves only need locking against invalidate_bh_lrus. We use
1241	* a local interrupt disable for that.
1242	*/
1243
1244	#define BH_LRU_SIZE 8
1245
1246	struct bh_lru {
1247	struct buffer_head *bhs[BH_LRU_SIZE];
1248	};
1249
1250	static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1251
1252	#ifdef CONFIG_SMP
1253	#define bh_lru_lock() local_irq_disable()
1254	#define bh_lru_unlock() local_irq_enable()
1255	#else
1256	#define bh_lru_lock() preempt_disable()
1257	#define bh_lru_unlock() preempt_enable()
1258	#endif
1259
1260	static inline void check_irqs_on(void)
1261	{
1262	#ifdef irqs_disabled
1263	BUG_ON(irqs_disabled());
1264	#endif
1265	}
1266
1267	/*
1268	* The LRU management algorithm is dopey-but-simple. Sorry.
1269	*/
1270	static void bh_lru_install(struct buffer_head *bh)
1271	{
1272	struct buffer_head *evictee = NULL;
1273	struct bh_lru *lru;
1274
1275	check_irqs_on();
1276	bh_lru_lock();
1277	lru = &__get_cpu_var(bh_lrus);
1278	if (lru->bhs[0] != bh) {
1279	struct buffer_head *bhs[BH_LRU_SIZE];
1280	int in;
1281	int out = 0;
1282
1283	get_bh(bh);
1284	bhs[out++] = bh;
1285	for (in = 0; in < BH_LRU_SIZE; in++) {
1286	struct buffer_head *bh2 = lru->bhs[in];
1287
1288	if (bh2 == bh) {
1289	__brelse(bh2);
1290	} else {
1291	if (out >= BH_LRU_SIZE) {
1292	BUG_ON(evictee != NULL);
1293	evictee = bh2;
1294	} else {
1295	bhs[out++] = bh2;
1296	}
1297	}
1298	}
1299	while (out < BH_LRU_SIZE)
1300	bhs[out++] = NULL;
1301	memcpy(lru->bhs, bhs, sizeof(bhs));
1302	}
1303	bh_lru_unlock();
1304
1305	if (evictee)
1306	__brelse(evictee);
1307	}
1308
1309	/*
1310	* Look up the bh in this cpu's LRU. If it's there, move it to the head.
1311	*/
1312	static struct buffer_head *
1313	lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1314	{
1315	struct buffer_head *ret = NULL;
1316	struct bh_lru *lru;
1317	unsigned int i;
1318
1319	check_irqs_on();
1320	bh_lru_lock();
1321	lru = &__get_cpu_var(bh_lrus);
1322	for (i = 0; i < BH_LRU_SIZE; i++) {
1323	struct buffer_head *bh = lru->bhs[i];
1324
1325	if (bh && bh->b_bdev == bdev &&
1326	bh->b_blocknr == block && bh->b_size == size) {
1327	if (i) {
1328	while (i) {
1329	lru->bhs[i] = lru->bhs[i - 1];
1330	i--;
1331	}
1332	lru->bhs[0] = bh;
1333	}
1334	get_bh(bh);
1335	ret = bh;
1336	break;
1337	}
1338	}
1339	bh_lru_unlock();
1340	return ret;
1341	}
1342
1343	/*
1344	* Perform a pagecache lookup for the matching buffer. If it's there, refresh
1345	* it in the LRU and mark it as accessed. If it is not present then return
1346	* NULL
1347	*/
1348	struct buffer_head *
1349	__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1350	{
1351	struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1352
1353	if (bh == NULL) {
1354	bh = __find_get_block_slow(bdev, block);
1355	if (bh)
1356	bh_lru_install(bh);
1357	}
1358	if (bh)
1359	touch_buffer(bh);
1360	return bh;
1361	}
1362	EXPORT_SYMBOL(__find_get_block);
1363
1364	/*
1365	* __getblk will locate (and, if necessary, create) the buffer_head
1366	* which corresponds to the passed block_device, block and size. The
1367	* returned buffer has its reference count incremented.
1368	*
1369	* __getblk() cannot fail - it just keeps trying. If you pass it an
1370	* illegal block number, __getblk() will happily return a buffer_head
1371	* which represents the non-existent block. Very weird.
1372	*
1373	* __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1374	* attempt is failing. FIXME, perhaps?
1375	*/
1376	struct buffer_head *
1377	__getblk(struct block_device *bdev, sector_t block, unsigned size)
1378	{
1379	struct buffer_head *bh = __find_get_block(bdev, block, size);
1380
1381	might_sleep();
1382	if (bh == NULL)
1383	bh = __getblk_slow(bdev, block, size);
1384	return bh;
1385	}
1386	EXPORT_SYMBOL(__getblk);
1387
1388	/*
1389	* Do async read-ahead on a buffer..
1390	*/
1391	void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1392	{
1393	struct buffer_head *bh = __getblk(bdev, block, size);
1394	if (likely(bh)) {
1395	ll_rw_block(READA, 1, &bh);
1396	brelse(bh);
1397	}
1398	}
1399	EXPORT_SYMBOL(__breadahead);
1400
1401	/**
1402	* __bread() - reads a specified block and returns the bh
1403	* @bdev: the block_device to read from
1404	* @block: number of block
1405	* @size: size (in bytes) to read
1406	*
1407	* Reads a specified block, and returns buffer head that contains it.
1408	* It returns NULL if the block was unreadable.
1409	*/
1410	struct buffer_head *
1411	__bread(struct block_device *bdev, sector_t block, unsigned size)
1412	{
1413	struct buffer_head *bh = __getblk(bdev, block, size);
1414
1415	if (likely(bh) && !buffer_uptodate(bh))
1416	bh = __bread_slow(bh);
1417	return bh;
1418	}
1419	EXPORT_SYMBOL(__bread);
1420
1421	/*
1422	* invalidate_bh_lrus() is called rarely - but not only at unmount.
1423	* This doesn't race because it runs in each cpu either in irq
1424	* or with preempt disabled.
1425	*/
1426	static void invalidate_bh_lru(void *arg)
1427	{
1428	struct bh_lru *b = &get_cpu_var(bh_lrus);
1429	int i;
1430
1431	for (i = 0; i < BH_LRU_SIZE; i++) {
1432	brelse(b->bhs[i]);
1433	b->bhs[i] = NULL;
1434	}
1435	put_cpu_var(bh_lrus);
1436	}
1437
1438	void invalidate_bh_lrus(void)
1439	{
1440	on_each_cpu(invalidate_bh_lru, NULL, 1);
1441	}
1442	EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1443
1444	void set_bh_page(struct buffer_head *bh,
1445	struct page *page, unsigned long offset)
1446	{
1447	bh->b_page = page;
1448	BUG_ON(offset >= PAGE_SIZE);
1449	if (PageHighMem(page))
1450	/*
1451	* This catches illegal uses and preserves the offset:
1452	*/
1453	bh->b_data = (char *)(0 + offset);
1454	else
1455	bh->b_data = page_address(page) + offset;
1456	}
1457	EXPORT_SYMBOL(set_bh_page);
1458
1459	/*
1460	* Called when truncating a buffer on a page completely.
1461	*/
1462	static void discard_buffer(struct buffer_head * bh)
1463	{
1464	lock_buffer(bh);
1465	clear_buffer_dirty(bh);
1466	bh->b_bdev = NULL;
1467	clear_buffer_mapped(bh);
1468	clear_buffer_req(bh);
1469	clear_buffer_new(bh);
1470	clear_buffer_delay(bh);
1471	clear_buffer_unwritten(bh);
1472	unlock_buffer(bh);
1473	}
1474
1475	/**
1476	* block_invalidatepage - invalidate part of all of a buffer-backed page
1477	*
1478	* @page: the page which is affected
1479	* @offset: the index of the truncation point
1480	*
1481	* block_invalidatepage() is called when all or part of the page has become
1482	* invalidatedby a truncate operation.
1483	*
1484	* block_invalidatepage() does not have to release all buffers, but it must
1485	* ensure that no dirty buffer is left outside @offset and that no I/O
1486	* is underway against any of the blocks which are outside the truncation
1487	* point. Because the caller is about to free (and possibly reuse) those
1488	* blocks on-disk.
1489	*/
1490	void block_invalidatepage(struct page *page, unsigned long offset)
1491	{
1492	struct buffer_head head, bh, *next;
1493	unsigned int curr_off = 0;
1494
1495	BUG_ON(!PageLocked(page));
1496	if (!page_has_buffers(page))
1497	goto out;
1498
1499	head = page_buffers(page);
1500	bh = head;
1501	do {
1502	unsigned int next_off = curr_off + bh->b_size;
1503	next = bh->b_this_page;
1504
1505	/*
1506	* is this block fully invalidated?
1507	*/
1508	if (offset <= curr_off)
1509	discard_buffer(bh);
1510	curr_off = next_off;
1511	bh = next;
1512	} while (bh != head);
1513
1514	/*
1515	* We release buffers only if the entire page is being invalidated.
1516	* The get_block cached value has been unconditionally invalidated,
1517	* so real IO is not possible anymore.
1518	*/
1519	if (offset == 0)
1520	try_to_release_page(page, 0);
1521	out:
1522	return;
1523	}
1524	EXPORT_SYMBOL(block_invalidatepage);
1525
1526	/*
1527	* We attach and possibly dirty the buffers atomically wrt
1528	* __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1529	* is already excluded via the page lock.
1530	*/
1531	void create_empty_buffers(struct page *page,
1532	unsigned long blocksize, unsigned long b_state)
1533	{
1534	struct buffer_head bh, head, *tail;
1535
1536	head = alloc_page_buffers(page, blocksize, 1);
1537	bh = head;
1538	do {
1539	bh->b_state \|= b_state;
1540	tail = bh;
1541	bh = bh->b_this_page;
1542	} while (bh);
1543	tail->b_this_page = head;
1544
1545	spin_lock(&page->mapping->private_lock);
1546	if (PageUptodate(page) \|\| PageDirty(page)) {
1547	bh = head;
1548	do {
1549	if (PageDirty(page))
1550	set_buffer_dirty(bh);
1551	if (PageUptodate(page))
1552	set_buffer_uptodate(bh);
1553	bh = bh->b_this_page;
1554	} while (bh != head);
1555	}
1556	attach_page_buffers(page, head);
1557	spin_unlock(&page->mapping->private_lock);
1558	}
1559	EXPORT_SYMBOL(create_empty_buffers);
1560
1561	/*
1562	* We are taking a block for data and we don't want any output from any
1563	* buffer-cache aliases starting from return from that function and
1564	* until the moment when something will explicitly mark the buffer
1565	* dirty (hopefully that will not happen until we will free that block ;-)
1566	* We don't even need to mark it not-uptodate - nobody can expect
1567	* anything from a newly allocated buffer anyway. We used to used
1568	* unmap_buffer() for such invalidation, but that was wrong. We definitely
1569	* don't want to mark the alias unmapped, for example - it would confuse
1570	* anyone who might pick it with bread() afterwards...
1571	*
1572	* Also.. Note that bforget() doesn't lock the buffer. So there can
1573	* be writeout I/O going on against recently-freed buffers. We don't
1574	* wait on that I/O in bforget() - it's more efficient to wait on the I/O
1575	* only if we really need to. That happens here.
1576	*/
1577	void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1578	{
1579	struct buffer_head *old_bh;
1580
1581	might_sleep();
1582
1583	old_bh = __find_get_block_slow(bdev, block);
1584	if (old_bh) {
1585	clear_buffer_dirty(old_bh);
1586	wait_on_buffer(old_bh);
1587	clear_buffer_req(old_bh);
1588	__brelse(old_bh);
1589	}
1590	}
1591	EXPORT_SYMBOL(unmap_underlying_metadata);
1592
1593	/*
1594	* NOTE! All mapped/uptodate combinations are valid:
1595	*
1596	* Mapped Uptodate Meaning
1597	*
1598	* No No "unknown" - must do get_block()
1599	* No Yes "hole" - zero-filled
1600	* Yes No "allocated" - allocated on disk, not read in
1601	* Yes Yes "valid" - allocated and up-to-date in memory.
1602	*
1603	* "Dirty" is valid only with the last case (mapped+uptodate).
1604	*/
1605
1606	/*
1607	* While block_write_full_page is writing back the dirty buffers under
1608	* the page lock, whoever dirtied the buffers may decide to clean them
1609	* again at any time. We handle that by only looking at the buffer
1610	* state inside lock_buffer().
1611	*
1612	* If block_write_full_page() is called for regular writeback
1613	* (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1614	* locked buffer. This only can happen if someone has written the buffer
1615	* directly, with submit_bh(). At the address_space level PageWriteback
1616	* prevents this contention from occurring.
1617	*
1618	* If block_write_full_page() is called with wbc->sync_mode ==
1619	* WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1620	* causes the writes to be flagged as synchronous writes, but the
1621	* block device queue will NOT be unplugged, since usually many pages
1622	* will be pushed to the out before the higher-level caller actually
1623	* waits for the writes to be completed. The various wait functions,
1624	* such as wait_on_writeback_range() will ultimately call sync_page()
1625	* which will ultimately call blk_run_backing_dev(), which will end up
1626	* unplugging the device queue.
1627	*/
1628	static int __block_write_full_page(struct inode inode, struct page page,
1629	get_block_t get_block, struct writeback_control wbc,
1630	bh_end_io_t *handler)
1631	{
1632	int err;
1633	sector_t block;
1634	sector_t last_block;
1635	struct buffer_head bh, head;
1636	const unsigned blocksize = 1 << inode->i_blkbits;
1637	int nr_underway = 0;
1638	int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1639	WRITE_SYNC_PLUG : WRITE);
1640
1641	BUG_ON(!PageLocked(page));
1642
1643	last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1644
1645	if (!page_has_buffers(page)) {
1646	create_empty_buffers(page, blocksize,
1647	(1 << BH_Dirty)\|(1 << BH_Uptodate));
1648	}
1649
1650	/*
1651	* Be very careful. We have no exclusion from __set_page_dirty_buffers
1652	* here, and the (potentially unmapped) buffers may become dirty at
1653	* any time. If a buffer becomes dirty here after we've inspected it
1654	* then we just miss that fact, and the page stays dirty.
1655	*
1656	* Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1657	* handle that here by just cleaning them.
1658	*/
1659
1660	block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1661	head = page_buffers(page);
1662	bh = head;
1663
1664	/*
1665	* Get all the dirty buffers mapped to disk addresses and
1666	* handle any aliases from the underlying blockdev's mapping.
1667	*/
1668	do {
1669	if (block > last_block) {
1670	/*
1671	* mapped buffers outside i_size will occur, because
1672	* this page can be outside i_size when there is a
1673	* truncate in progress.
1674	*/
1675	/*
1676	* The buffer was zeroed by block_write_full_page()
1677	*/
1678	clear_buffer_dirty(bh);
1679	set_buffer_uptodate(bh);
1680	} else if ((!buffer_mapped(bh) \|\| buffer_delay(bh)) &&
1681	buffer_dirty(bh)) {
1682	WARN_ON(bh->b_size != blocksize);
1683	err = get_block(inode, block, bh, 1);
1684	if (err)
1685	goto recover;
1686	clear_buffer_delay(bh);
1687	if (buffer_new(bh)) {
1688	/* blockdev mappings never come here */
1689	clear_buffer_new(bh);
1690	unmap_underlying_metadata(bh->b_bdev,
1691	bh->b_blocknr);
1692	}
1693	}
1694	bh = bh->b_this_page;
1695	block++;
1696	} while (bh != head);
1697
1698	do {
1699	if (!buffer_mapped(bh))
1700	continue;
1701	/*
1702	* If it's a fully non-blocking write attempt and we cannot
1703	* lock the buffer then redirty the page. Note that this can
1704	* potentially cause a busy-wait loop from writeback threads
1705	* and kswapd activity, but those code paths have their own
1706	* higher-level throttling.
1707	*/
1708	if (wbc->sync_mode != WB_SYNC_NONE \|\| !wbc->nonblocking) {
1709	lock_buffer(bh);
1710	} else if (!trylock_buffer(bh)) {
1711	redirty_page_for_writepage(wbc, page);
1712	continue;
1713	}
1714	if (test_clear_buffer_dirty(bh)) {
1715	mark_buffer_async_write_endio(bh, handler);
1716	} else {
1717	unlock_buffer(bh);
1718	}
1719	} while ((bh = bh->b_this_page) != head);
1720
1721	/*
1722	* The page and its buffers are protected by PageWriteback(), so we can
1723	* drop the bh refcounts early.
1724	*/
1725	BUG_ON(PageWriteback(page));
1726	set_page_writeback(page);
1727
1728	do {
1729	struct buffer_head *next = bh->b_this_page;
1730	if (buffer_async_write(bh)) {
1731	submit_bh(write_op, bh);
1732	nr_underway++;
1733	}
1734	bh = next;
1735	} while (bh != head);
1736	unlock_page(page);
1737
1738	err = 0;
1739	done:
1740	if (nr_underway == 0) {
1741	/*
1742	* The page was marked dirty, but the buffers were
1743	* clean. Someone wrote them back by hand with
1744	* ll_rw_block/submit_bh. A rare case.
1745	*/
1746	end_page_writeback(page);
1747
1748	/*
1749	* The page and buffer_heads can be released at any time from
1750	* here on.
1751	*/
1752	}
1753	return err;
1754
1755	recover:
1756	/*
1757	* ENOSPC, or some other error. We may already have added some
1758	* blocks to the file, so we need to write these out to avoid
1759	* exposing stale data.
1760	* The page is currently locked and not marked for writeback
1761	*/
1762	bh = head;
1763	/* Recovery: lock and submit the mapped buffers */
1764	do {
1765	if (buffer_mapped(bh) && buffer_dirty(bh) &&
1766	!buffer_delay(bh)) {
1767	lock_buffer(bh);
1768	mark_buffer_async_write_endio(bh, handler);
1769	} else {
1770	/*
1771	* The buffer may have been set dirty during
1772	* attachment to a dirty page.
1773	*/
1774	clear_buffer_dirty(bh);
1775	}
1776	} while ((bh = bh->b_this_page) != head);
1777	SetPageError(page);
1778	BUG_ON(PageWriteback(page));
1779	mapping_set_error(page->mapping, err);
1780	set_page_writeback(page);
1781	do {
1782	struct buffer_head *next = bh->b_this_page;
1783	if (buffer_async_write(bh)) {
1784	clear_buffer_dirty(bh);
1785	submit_bh(write_op, bh);
1786	nr_underway++;
1787	}
1788	bh = next;
1789	} while (bh != head);
1790	unlock_page(page);
1791	goto done;
1792	}
1793
1794	/*
1795	* If a page has any new buffers, zero them out here, and mark them uptodate
1796	* and dirty so they'll be written out (in order to prevent uninitialised
1797	* block data from leaking). And clear the new bit.
1798	*/
1799	void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1800	{
1801	unsigned int block_start, block_end;
1802	struct buffer_head head, bh;
1803
1804	BUG_ON(!PageLocked(page));
1805	if (!page_has_buffers(page))
1806	return;
1807
1808	bh = head = page_buffers(page);
1809	block_start = 0;
1810	do {
1811	block_end = block_start + bh->b_size;
1812
1813	if (buffer_new(bh)) {
1814	if (block_end > from && block_start < to) {
1815	if (!PageUptodate(page)) {
1816	unsigned start, size;
1817
1818	start = max(from, block_start);
1819	size = min(to, block_end) - start;
1820
1821	zero_user(page, start, size);
1822	set_buffer_uptodate(bh);
1823	}
1824
1825	clear_buffer_new(bh);
1826	mark_buffer_dirty(bh);
1827	}
1828	}
1829
1830	block_start = block_end;
1831	bh = bh->b_this_page;
1832	} while (bh != head);
1833	}
1834	EXPORT_SYMBOL(page_zero_new_buffers);
1835
1836	static int __block_prepare_write(struct inode inode, struct page page,
1837	unsigned from, unsigned to, get_block_t *get_block)
1838	{
1839	unsigned block_start, block_end;
1840	sector_t block;
1841	int err = 0;
1842	unsigned blocksize, bbits;
1843	struct buffer_head bh, head, wait[2], *wait_bh=wait;
1844
1845	BUG_ON(!PageLocked(page));
1846	BUG_ON(from > PAGE_CACHE_SIZE);
1847	BUG_ON(to > PAGE_CACHE_SIZE);
1848	BUG_ON(from > to);
1849
1850	blocksize = 1 << inode->i_blkbits;
1851	if (!page_has_buffers(page))
1852	create_empty_buffers(page, blocksize, 0);
1853	head = page_buffers(page);
1854
1855	bbits = inode->i_blkbits;
1856	block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1857
1858	for(bh = head, block_start = 0; bh != head \|\| !block_start;
1859	block++, block_start=block_end, bh = bh->b_this_page) {
1860	block_end = block_start + blocksize;
1861	if (block_end <= from \|\| block_start >= to) {
1862	if (PageUptodate(page)) {
1863	if (!buffer_uptodate(bh))
1864	set_buffer_uptodate(bh);
1865	}
1866	continue;
1867	}
1868	if (buffer_new(bh))
1869	clear_buffer_new(bh);
1870	if (!buffer_mapped(bh)) {
1871	WARN_ON(bh->b_size != blocksize);
1872	err = get_block(inode, block, bh, 1);
1873	if (err)
1874	break;
1875	if (buffer_new(bh)) {
1876	unmap_underlying_metadata(bh->b_bdev,
1877	bh->b_blocknr);
1878	if (PageUptodate(page)) {
1879	clear_buffer_new(bh);
1880	set_buffer_uptodate(bh);
1881	mark_buffer_dirty(bh);
1882	continue;
1883	}
1884	if (block_end > to \|\| block_start < from)
1885	zero_user_segments(page,
1886	to, block_end,
1887	block_start, from);
1888	continue;
1889	}
1890	}
1891	if (PageUptodate(page)) {
1892	if (!buffer_uptodate(bh))
1893	set_buffer_uptodate(bh);
1894	continue;
1895	}
1896	if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1897	!buffer_unwritten(bh) &&
1898	(block_start < from \|\| block_end > to)) {
1899	ll_rw_block(READ, 1, &bh);
1900	*wait_bh++=bh;
1901	}
1902	}
1903	/*
1904	* If we issued read requests - let them complete.
1905	*/
1906	while(wait_bh > wait) {
1907	wait_on_buffer(*--wait_bh);
1908	if (!buffer_uptodate(*wait_bh))
1909	err = -EIO;
1910	}
1911	if (unlikely(err))
1912	page_zero_new_buffers(page, from, to);
1913	return err;
1914	}
1915
1916	static int __block_commit_write(struct inode inode, struct page page,
1917	unsigned from, unsigned to)
1918	{
1919	unsigned block_start, block_end;
1920	int partial = 0;
1921	unsigned blocksize;
1922	struct buffer_head bh, head;
1923
1924	blocksize = 1 << inode->i_blkbits;
1925
1926	for(bh = head = page_buffers(page), block_start = 0;
1927	bh != head \|\| !block_start;
1928	block_start=block_end, bh = bh->b_this_page) {
1929	block_end = block_start + blocksize;
1930	if (block_end <= from \|\| block_start >= to) {
1931	if (!buffer_uptodate(bh))
1932	partial = 1;
1933	} else {
1934	set_buffer_uptodate(bh);
1935	mark_buffer_dirty(bh);
1936	}
1937	clear_buffer_new(bh);
1938	}
1939
1940	/*
1941	* If this is a partial write which happened to make all buffers
1942	* uptodate then we can optimize away a bogus readpage() for
1943	* the next read(). Here we 'discover' whether the page went
1944	* uptodate as a result of this (potentially partial) write.
1945	*/
1946	if (!partial)
1947	SetPageUptodate(page);
1948	return 0;
1949	}
1950
1951	/*
1952	* Filesystems implementing the new truncate sequence should use the
1953	* _newtrunc postfix variant which won't incorrectly call vmtruncate.
1954	* The filesystem needs to handle block truncation upon failure.
1955	*/
1956	int block_write_begin_newtrunc(struct file file, struct address_space mapping,
1957	loff_t pos, unsigned len, unsigned flags,
1958	struct page pagep, void fsdata,
1959	get_block_t *get_block)
1960	{
1961	struct inode *inode = mapping->host;
1962	int status = 0;
1963	struct page *page;
1964	pgoff_t index;
1965	unsigned start, end;
1966	int ownpage = 0;
1967
1968	index = pos >> PAGE_CACHE_SHIFT;
1969	start = pos & (PAGE_CACHE_SIZE - 1);
1970	end = start + len;
1971
1972	page = *pagep;
1973	if (page == NULL) {
1974	ownpage = 1;
1975	page = grab_cache_page_write_begin(mapping, index, flags);
1976	if (!page) {
1977	status = -ENOMEM;
1978	goto out;
1979	}
1980	*pagep = page;
1981	} else
1982	BUG_ON(!PageLocked(page));
1983
1984	status = __block_prepare_write(inode, page, start, end, get_block);
1985	if (unlikely(status)) {
1986	ClearPageUptodate(page);
1987
1988	if (ownpage) {
1989	unlock_page(page);
1990	page_cache_release(page);
1991	*pagep = NULL;
1992	}
1993	}
1994
1995	out:
1996	return status;
1997	}
1998	EXPORT_SYMBOL(block_write_begin_newtrunc);
1999
2000	/*
2001	* block_write_begin takes care of the basic task of block allocation and
2002	* bringing partial write blocks uptodate first.
2003	*
2004	* If *pagep is not NULL, then block_write_begin uses the locked page
2005	* at *pagep rather than allocating its own. In this case, the page will
2006	* not be unlocked or deallocated on failure.
2007	*/
2008	int block_write_begin(struct file file, struct address_space mapping,
2009	loff_t pos, unsigned len, unsigned flags,
2010	struct page pagep, void fsdata,
2011	get_block_t *get_block)
2012	{
2013	int ret;
2014
2015	ret = block_write_begin_newtrunc(file, mapping, pos, len, flags,
2016	pagep, fsdata, get_block);
2017
2018	/*
2019	* prepare_write() may have instantiated a few blocks
2020	* outside i_size. Trim these off again. Don't need
2021	* i_size_read because we hold i_mutex.
2022	*
2023	* Filesystems which pass down their own page also cannot
2024	* call into vmtruncate here because it would lead to lock
2025	* inversion problems (*pagep is locked). This is a further
2026	* example of where the old truncate sequence is inadequate.
2027	*/
2028	if (unlikely(ret) && *pagep == NULL) {
2029	loff_t isize = mapping->host->i_size;
2030	if (pos + len > isize)
2031	vmtruncate(mapping->host, isize);
2032	}
2033
2034	return ret;
2035	}
2036	EXPORT_SYMBOL(block_write_begin);
2037
2038	int block_write_end(struct file file, struct address_space mapping,
2039	loff_t pos, unsigned len, unsigned copied,
2040	struct page page, void fsdata)
2041	{
2042	struct inode *inode = mapping->host;
2043	unsigned start;
2044
2045	start = pos & (PAGE_CACHE_SIZE - 1);
2046
2047	if (unlikely(copied < len)) {
2048	/*
2049	* The buffers that were written will now be uptodate, so we
2050	* don't have to worry about a readpage reading them and
2051	* overwriting a partial write. However if we have encountered
2052	* a short write and only partially written into a buffer, it
2053	* will not be marked uptodate, so a readpage might come in and
2054	* destroy our partial write.
2055	*
2056	* Do the simplest thing, and just treat any short write to a
2057	* non uptodate page as a zero-length write, and force the
2058	* caller to redo the whole thing.
2059	*/
2060	if (!PageUptodate(page))
2061	copied = 0;
2062
2063	page_zero_new_buffers(page, start+copied, start+len);
2064	}
2065	flush_dcache_page(page);
2066
2067	/* This could be a short (even 0-length) commit */
2068	__block_commit_write(inode, page, start, start+copied);
2069
2070	return copied;
2071	}
2072	EXPORT_SYMBOL(block_write_end);
2073
2074	int generic_write_end(struct file file, struct address_space mapping,
2075	loff_t pos, unsigned len, unsigned copied,
2076	struct page page, void fsdata)
2077	{
2078	struct inode *inode = mapping->host;
2079	int i_size_changed = 0;
2080
2081	copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2082
2083	/*
2084	* No need to use i_size_read() here, the i_size
2085	* cannot change under us because we hold i_mutex.
2086	*
2087	* But it's important to update i_size while still holding page lock:
2088	* page writeout could otherwise come in and zero beyond i_size.
2089	*/
2090	if (pos+copied > inode->i_size) {
2091	i_size_write(inode, pos+copied);
2092	i_size_changed = 1;
2093	}
2094
2095	unlock_page(page);
2096	page_cache_release(page);
2097
2098	/*
2099	* Don't mark the inode dirty under page lock. First, it unnecessarily
2100	* makes the holding time of page lock longer. Second, it forces lock
2101	* ordering of page lock and transaction start for journaling
2102	* filesystems.
2103	*/
2104	if (i_size_changed)
2105	mark_inode_dirty(inode);
2106
2107	return copied;
2108	}
2109	EXPORT_SYMBOL(generic_write_end);
2110
2111	/*
2112	* block_is_partially_uptodate checks whether buffers within a page are
2113	* uptodate or not.
2114	*
2115	* Returns true if all buffers which correspond to a file portion
2116	* we want to read are uptodate.
2117	*/
2118	int block_is_partially_uptodate(struct page page, read_descriptor_t desc,
2119	unsigned long from)
2120	{
2121	struct inode *inode = page->mapping->host;
2122	unsigned block_start, block_end, blocksize;
2123	unsigned to;
2124	struct buffer_head bh, head;
2125	int ret = 1;
2126
2127	if (!page_has_buffers(page))
2128	return 0;
2129
2130	blocksize = 1 << inode->i_blkbits;
2131	to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2132	to = from + to;
2133	if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2134	return 0;
2135
2136	head = page_buffers(page);
2137	bh = head;
2138	block_start = 0;
2139	do {
2140	block_end = block_start + blocksize;
2141	if (block_end > from && block_start < to) {
2142	if (!buffer_uptodate(bh)) {
2143	ret = 0;
2144	break;
2145	}
2146	if (block_end >= to)
2147	break;
2148	}
2149	block_start = block_end;
2150	bh = bh->b_this_page;
2151	} while (bh != head);
2152
2153	return ret;
2154	}
2155	EXPORT_SYMBOL(block_is_partially_uptodate);
2156
2157	/*
2158	* Generic "read page" function for block devices that have the normal
2159	* get_block functionality. This is most of the block device filesystems.
2160	* Reads the page asynchronously --- the unlock_buffer() and
2161	* set/clear_buffer_uptodate() functions propagate buffer state into the
2162	* page struct once IO has completed.
2163	*/
2164	int block_read_full_page(struct page page, get_block_t get_block)
2165	{
2166	struct inode *inode = page->mapping->host;
2167	sector_t iblock, lblock;
2168	struct buffer_head bh, head, *arr[MAX_BUF_PER_PAGE];
2169	unsigned int blocksize;
2170	int nr, i;
2171	int fully_mapped = 1;
2172
2173	BUG_ON(!PageLocked(page));
2174	blocksize = 1 << inode->i_blkbits;
2175	if (!page_has_buffers(page))
2176	create_empty_buffers(page, blocksize, 0);
2177	head = page_buffers(page);
2178
2179	iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2180	lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2181	bh = head;
2182	nr = 0;
2183	i = 0;
2184
2185	do {
2186	if (buffer_uptodate(bh))
2187	continue;
2188
2189	if (!buffer_mapped(bh)) {
2190	int err = 0;
2191
2192	fully_mapped = 0;
2193	if (iblock < lblock) {
2194	WARN_ON(bh->b_size != blocksize);
2195	err = get_block(inode, iblock, bh, 0);
2196	if (err)
2197	SetPageError(page);
2198	}
2199	if (!buffer_mapped(bh)) {
2200	zero_user(page, i * blocksize, blocksize);
2201	if (!err)
2202	set_buffer_uptodate(bh);
2203	continue;
2204	}
2205	/*
2206	* get_block() might have updated the buffer
2207	* synchronously
2208	*/
2209	if (buffer_uptodate(bh))
2210	continue;
2211	}
2212	arr[nr++] = bh;
2213	} while (i++, iblock++, (bh = bh->b_this_page) != head);
2214
2215	if (fully_mapped)
2216	SetPageMappedToDisk(page);
2217
2218	if (!nr) {
2219	/*
2220	* All buffers are uptodate - we can set the page uptodate
2221	* as well. But not if get_block() returned an error.
2222	*/
2223	if (!PageError(page))
2224	SetPageUptodate(page);
2225	unlock_page(page);
2226	return 0;
2227	}
2228
2229	/* Stage two: lock the buffers */
2230	for (i = 0; i < nr; i++) {
2231	bh = arr[i];
2232	lock_buffer(bh);
2233	mark_buffer_async_read(bh);
2234	}
2235
2236	/*
2237	* Stage 3: start the IO. Check for uptodateness
2238	* inside the buffer lock in case another process reading
2239	* the underlying blockdev brought it uptodate (the sct fix).
2240	*/
2241	for (i = 0; i < nr; i++) {
2242	bh = arr[i];
2243	if (buffer_uptodate(bh))
2244	end_buffer_async_read(bh, 1);
2245	else
2246	submit_bh(READ, bh);
2247	}
2248	return 0;
2249	}
2250	EXPORT_SYMBOL(block_read_full_page);
2251
2252	/* utility function for filesystems that need to do work on expanding
2253	* truncates. Uses filesystem pagecache writes to allow the filesystem to
2254	* deal with the hole.
2255	*/
2256	int generic_cont_expand_simple(struct inode *inode, loff_t size)
2257	{
2258	struct address_space *mapping = inode->i_mapping;
2259	struct page *page;
2260	void *fsdata;
2261	int err;
2262
2263	err = inode_newsize_ok(inode, size);
2264	if (err)
2265	goto out;
2266
2267	err = pagecache_write_begin(NULL, mapping, size, 0,
2268	AOP_FLAG_UNINTERRUPTIBLE\|AOP_FLAG_CONT_EXPAND,
2269	&page, &fsdata);
2270	if (err)
2271	goto out;
2272
2273	err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2274	BUG_ON(err > 0);
2275
2276	out:
2277	return err;
2278	}
2279	EXPORT_SYMBOL(generic_cont_expand_simple);
2280
2281	static int cont_expand_zero(struct file file, struct address_space mapping,
2282	loff_t pos, loff_t *bytes)
2283	{
2284	struct inode *inode = mapping->host;
2285	unsigned blocksize = 1 << inode->i_blkbits;
2286	struct page *page;
2287	void *fsdata;
2288	pgoff_t index, curidx;
2289	loff_t curpos;
2290	unsigned zerofrom, offset, len;
2291	int err = 0;
2292
2293	index = pos >> PAGE_CACHE_SHIFT;
2294	offset = pos & ~PAGE_CACHE_MASK;
2295
2296	while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2297	zerofrom = curpos & ~PAGE_CACHE_MASK;
2298	if (zerofrom & (blocksize-1)) {
2299	*bytes \|= (blocksize-1);
2300	(*bytes)++;
2301	}
2302	len = PAGE_CACHE_SIZE - zerofrom;
2303
2304	err = pagecache_write_begin(file, mapping, curpos, len,
2305	AOP_FLAG_UNINTERRUPTIBLE,
2306	&page, &fsdata);
2307	if (err)
2308	goto out;
2309	zero_user(page, zerofrom, len);
2310	err = pagecache_write_end(file, mapping, curpos, len, len,
2311	page, fsdata);
2312	if (err < 0)
2313	goto out;
2314	BUG_ON(err != len);
2315	err = 0;
2316
2317	balance_dirty_pages_ratelimited(mapping);
2318	}
2319
2320	/* page covers the boundary, find the boundary offset */
2321	if (index == curidx) {
2322	zerofrom = curpos & ~PAGE_CACHE_MASK;
2323	/* if we will expand the thing last block will be filled */
2324	if (offset <= zerofrom) {
2325	goto out;
2326	}
2327	if (zerofrom & (blocksize-1)) {
2328	*bytes \|= (blocksize-1);
2329	(*bytes)++;
2330	}
2331	len = offset - zerofrom;
2332
2333	err = pagecache_write_begin(file, mapping, curpos, len,
2334	AOP_FLAG_UNINTERRUPTIBLE,
2335	&page, &fsdata);
2336	if (err)
2337	goto out;
2338	zero_user(page, zerofrom, len);
2339	err = pagecache_write_end(file, mapping, curpos, len, len,
2340	page, fsdata);
2341	if (err < 0)
2342	goto out;
2343	BUG_ON(err != len);
2344	err = 0;
2345	}
2346	out:
2347	return err;
2348	}
2349
2350	/*
2351	* For moronic filesystems that do not allow holes in file.
2352	* We may have to extend the file.
2353	*/
2354	int cont_write_begin_newtrunc(struct file file, struct address_space mapping,
2355	loff_t pos, unsigned len, unsigned flags,
2356	struct page pagep, void fsdata,
2357	get_block_t get_block, loff_t bytes)
2358	{
2359	struct inode *inode = mapping->host;
2360	unsigned blocksize = 1 << inode->i_blkbits;
2361	unsigned zerofrom;
2362	int err;
2363
2364	err = cont_expand_zero(file, mapping, pos, bytes);
2365	if (err)
2366	goto out;
2367
2368	zerofrom = *bytes & ~PAGE_CACHE_MASK;
2369	if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2370	*bytes \|= (blocksize-1);
2371	(*bytes)++;
2372	}
2373
2374	*pagep = NULL;
2375	err = block_write_begin_newtrunc(file, mapping, pos, len,
2376	flags, pagep, fsdata, get_block);
2377	out:
2378	return err;
2379	}
2380	EXPORT_SYMBOL(cont_write_begin_newtrunc);
2381
2382	int cont_write_begin(struct file file, struct address_space mapping,
2383	loff_t pos, unsigned len, unsigned flags,
2384	struct page pagep, void fsdata,
2385	get_block_t get_block, loff_t bytes)
2386	{
2387	int ret;
2388
2389	ret = cont_write_begin_newtrunc(file, mapping, pos, len, flags,
2390	pagep, fsdata, get_block, bytes);
2391	if (unlikely(ret)) {
2392	loff_t isize = mapping->host->i_size;
2393	if (pos + len > isize)
2394	vmtruncate(mapping->host, isize);
2395	}
2396
2397	return ret;
2398	}
2399	EXPORT_SYMBOL(cont_write_begin);
2400
2401	int block_prepare_write(struct page *page, unsigned from, unsigned to,
2402	get_block_t *get_block)
2403	{
2404	struct inode *inode = page->mapping->host;
2405	int err = __block_prepare_write(inode, page, from, to, get_block);
2406	if (err)
2407	ClearPageUptodate(page);
2408	return err;
2409	}
2410	EXPORT_SYMBOL(block_prepare_write);
2411
2412	int block_commit_write(struct page *page, unsigned from, unsigned to)
2413	{
2414	struct inode *inode = page->mapping->host;
2415	__block_commit_write(inode,page,from,to);
2416	return 0;
2417	}
2418	EXPORT_SYMBOL(block_commit_write);
2419
2420	/*
2421	* block_page_mkwrite() is not allowed to change the file size as it gets
2422	* called from a page fault handler when a page is first dirtied. Hence we must
2423	* be careful to check for EOF conditions here. We set the page up correctly
2424	* for a written page which means we get ENOSPC checking when writing into
2425	* holes and correct delalloc and unwritten extent mapping on filesystems that
2426	* support these features.
2427	*
2428	* We are not allowed to take the i_mutex here so we have to play games to
2429	* protect against truncate races as the page could now be beyond EOF. Because
2430	* truncate writes the inode size before removing pages, once we have the
2431	* page lock we can determine safely if the page is beyond EOF. If it is not
2432	* beyond EOF, then the page is guaranteed safe against truncation until we
2433	* unlock the page.
2434	*/
2435	int
2436	block_page_mkwrite(struct vm_area_struct vma, struct vm_fault vmf,
2437	get_block_t get_block)
2438	{
2439	struct page *page = vmf->page;
2440	struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2441	unsigned long end;
2442	loff_t size;
2443	int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2444
2445	lock_page(page);
2446	size = i_size_read(inode);
2447	if ((page->mapping != inode->i_mapping) \|\|
2448	(page_offset(page) > size)) {
2449	/* page got truncated out from underneath us */
2450	unlock_page(page);
2451	goto out;
2452	}
2453
2454	/* page is wholly or partially inside EOF */
2455	if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2456	end = size & ~PAGE_CACHE_MASK;
2457	else
2458	end = PAGE_CACHE_SIZE;
2459
2460	ret = block_prepare_write(page, 0, end, get_block);
2461	if (!ret)
2462	ret = block_commit_write(page, 0, end);
2463
2464	if (unlikely(ret)) {
2465	unlock_page(page);
2466	if (ret == -ENOMEM)
2467	ret = VM_FAULT_OOM;
2468	else /* -ENOSPC, -EIO, etc */
2469	ret = VM_FAULT_SIGBUS;
2470	} else
2471	ret = VM_FAULT_LOCKED;
2472
2473	out:
2474	return ret;
2475	}
2476	EXPORT_SYMBOL(block_page_mkwrite);
2477
2478	/*
2479	* nobh_write_begin()'s prereads are special: the buffer_heads are freed
2480	* immediately, while under the page lock. So it needs a special end_io
2481	* handler which does not touch the bh after unlocking it.
2482	*/
2483	static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2484	{
2485	__end_buffer_read_notouch(bh, uptodate);
2486	}
2487
2488	/*
2489	* Attach the singly-linked list of buffers created by nobh_write_begin, to
2490	* the page (converting it to circular linked list and taking care of page
2491	* dirty races).
2492	*/
2493	static void attach_nobh_buffers(struct page page, struct buffer_head head)
2494	{
2495	struct buffer_head *bh;
2496
2497	BUG_ON(!PageLocked(page));
2498
2499	spin_lock(&page->mapping->private_lock);
2500	bh = head;
2501	do {
2502	if (PageDirty(page))
2503	set_buffer_dirty(bh);
2504	if (!bh->b_this_page)
2505	bh->b_this_page = head;
2506	bh = bh->b_this_page;
2507	} while (bh != head);
2508	attach_page_buffers(page, head);
2509	spin_unlock(&page->mapping->private_lock);
2510	}
2511
2512	/*
2513	* Filesystems implementing the new truncate sequence should use the
2514	* _newtrunc postfix variant which won't incorrectly call vmtruncate.
2515	* The filesystem needs to handle block truncation upon failure.
2516	*/
2517	int nobh_write_begin_newtrunc(struct file file, struct address_space mapping,
2518	loff_t pos, unsigned len, unsigned flags,
2519	struct page pagep, void fsdata,
2520	get_block_t *get_block)
2521	{
2522	struct inode *inode = mapping->host;
2523	const unsigned blkbits = inode->i_blkbits;
2524	const unsigned blocksize = 1 << blkbits;
2525	struct buffer_head head, bh;
2526	struct page *page;
2527	pgoff_t index;
2528	unsigned from, to;
2529	unsigned block_in_page;
2530	unsigned block_start, block_end;
2531	sector_t block_in_file;
2532	int nr_reads = 0;
2533	int ret = 0;
2534	int is_mapped_to_disk = 1;
2535
2536	index = pos >> PAGE_CACHE_SHIFT;
2537	from = pos & (PAGE_CACHE_SIZE - 1);
2538	to = from + len;
2539
2540	page = grab_cache_page_write_begin(mapping, index, flags);
2541	if (!page)
2542	return -ENOMEM;
2543	*pagep = page;
2544	*fsdata = NULL;
2545
2546	if (page_has_buffers(page)) {
2547	unlock_page(page);
2548	page_cache_release(page);
2549	*pagep = NULL;
2550	return block_write_begin_newtrunc(file, mapping, pos, len,
2551	flags, pagep, fsdata, get_block);
2552	}
2553
2554	if (PageMappedToDisk(page))
2555	return 0;
2556
2557	/*
2558	* Allocate buffers so that we can keep track of state, and potentially
2559	* attach them to the page if an error occurs. In the common case of
2560	* no error, they will just be freed again without ever being attached
2561	* to the page (which is all OK, because we're under the page lock).
2562	*
2563	* Be careful: the buffer linked list is a NULL terminated one, rather
2564	* than the circular one we're used to.
2565	*/
2566	head = alloc_page_buffers(page, blocksize, 0);
2567	if (!head) {
2568	ret = -ENOMEM;
2569	goto out_release;
2570	}
2571
2572	block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2573
2574	/*
2575	* We loop across all blocks in the page, whether or not they are
2576	* part of the affected region. This is so we can discover if the
2577	* page is fully mapped-to-disk.
2578	*/
2579	for (block_start = 0, block_in_page = 0, bh = head;
2580	block_start < PAGE_CACHE_SIZE;
2581	block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2582	int create;
2583
2584	block_end = block_start + blocksize;
2585	bh->b_state = 0;
2586	create = 1;
2587	if (block_start >= to)
2588	create = 0;
2589	ret = get_block(inode, block_in_file + block_in_page,
2590	bh, create);
2591	if (ret)
2592	goto failed;
2593	if (!buffer_mapped(bh))
2594	is_mapped_to_disk = 0;
2595	if (buffer_new(bh))
2596	unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2597	if (PageUptodate(page)) {
2598	set_buffer_uptodate(bh);
2599	continue;
2600	}
2601	if (buffer_new(bh) \|\| !buffer_mapped(bh)) {
2602	zero_user_segments(page, block_start, from,
2603	to, block_end);
2604	continue;
2605	}
2606	if (buffer_uptodate(bh))
2607	continue; /* reiserfs does this */
2608	if (block_start < from \|\| block_end > to) {
2609	lock_buffer(bh);
2610	bh->b_end_io = end_buffer_read_nobh;
2611	submit_bh(READ, bh);
2612	nr_reads++;
2613	}
2614	}
2615
2616	if (nr_reads) {
2617	/*
2618	* The page is locked, so these buffers are protected from
2619	* any VM or truncate activity. Hence we don't need to care
2620	* for the buffer_head refcounts.
2621	*/
2622	for (bh = head; bh; bh = bh->b_this_page) {
2623	wait_on_buffer(bh);
2624	if (!buffer_uptodate(bh))
2625	ret = -EIO;
2626	}
2627	if (ret)
2628	goto failed;
2629	}
2630
2631	if (is_mapped_to_disk)
2632	SetPageMappedToDisk(page);
2633
2634	fsdata = head; / to be released by nobh_write_end */
2635
2636	return 0;
2637
2638	failed:
2639	BUG_ON(!ret);
2640	/*
2641	* Error recovery is a bit difficult. We need to zero out blocks that
2642	* were newly allocated, and dirty them to ensure they get written out.
2643	* Buffers need to be attached to the page at this point, otherwise
2644	* the handling of potential IO errors during writeout would be hard
2645	* (could try doing synchronous writeout, but what if that fails too?)
2646	*/
2647	attach_nobh_buffers(page, head);
2648	page_zero_new_buffers(page, from, to);
2649
2650	out_release:
2651	unlock_page(page);
2652	page_cache_release(page);
2653	*pagep = NULL;
2654
2655	return ret;
2656	}
2657	EXPORT_SYMBOL(nobh_write_begin_newtrunc);
2658
2659	/*
2660	* On entry, the page is fully not uptodate.
2661	* On exit the page is fully uptodate in the areas outside (from,to)
2662	*/
2663	int nobh_write_begin(struct file file, struct address_space mapping,
2664	loff_t pos, unsigned len, unsigned flags,
2665	struct page pagep, void fsdata,
2666	get_block_t *get_block)
2667	{
2668	int ret;
2669
2670	ret = nobh_write_begin_newtrunc(file, mapping, pos, len, flags,
2671	pagep, fsdata, get_block);
2672
2673	/*
2674	* prepare_write() may have instantiated a few blocks
2675	* outside i_size. Trim these off again. Don't need
2676	* i_size_read because we hold i_mutex.
2677	*/
2678	if (unlikely(ret)) {
2679	loff_t isize = mapping->host->i_size;
2680	if (pos + len > isize)
2681	vmtruncate(mapping->host, isize);
2682	}
2683
2684	return ret;
2685	}
2686	EXPORT_SYMBOL(nobh_write_begin);
2687
2688	int nobh_write_end(struct file file, struct address_space mapping,
2689	loff_t pos, unsigned len, unsigned copied,
2690	struct page page, void fsdata)
2691	{
2692	struct inode *inode = page->mapping->host;
2693	struct buffer_head *head = fsdata;
2694	struct buffer_head *bh;
2695	BUG_ON(fsdata != NULL && page_has_buffers(page));
2696
2697	if (unlikely(copied < len) && head)
2698	attach_nobh_buffers(page, head);
2699	if (page_has_buffers(page))
2700	return generic_write_end(file, mapping, pos, len,
2701	copied, page, fsdata);
2702
2703	SetPageUptodate(page);
2704	set_page_dirty(page);
2705	if (pos+copied > inode->i_size) {
2706	i_size_write(inode, pos+copied);
2707	mark_inode_dirty(inode);
2708	}
2709
2710	unlock_page(page);
2711	page_cache_release(page);
2712
2713	while (head) {
2714	bh = head;
2715	head = head->b_this_page;
2716	free_buffer_head(bh);
2717	}
2718
2719	return copied;
2720	}
2721	EXPORT_SYMBOL(nobh_write_end);
2722
2723	/*
2724	* nobh_writepage() - based on block_full_write_page() except
2725	* that it tries to operate without attaching bufferheads to
2726	* the page.
2727	*/
2728	int nobh_writepage(struct page page, get_block_t get_block,
2729	struct writeback_control *wbc)
2730	{
2731	struct inode * const inode = page->mapping->host;
2732	loff_t i_size = i_size_read(inode);
2733	const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2734	unsigned offset;
2735	int ret;
2736
2737	/* Is the page fully inside i_size? */
2738	if (page->index < end_index)
2739	goto out;
2740
2741	/* Is the page fully outside i_size? (truncate in progress) */
2742	offset = i_size & (PAGE_CACHE_SIZE-1);
2743	if (page->index >= end_index+1 \|\| !offset) {
2744	/*
2745	* The page may have dirty, unmapped buffers. For example,
2746	* they may have been added in ext3_writepage(). Make them
2747	* freeable here, so the page does not leak.
2748	*/
2749	#if 0
2750	/* Not really sure about this - do we need this ? */
2751	if (page->mapping->a_ops->invalidatepage)
2752	page->mapping->a_ops->invalidatepage(page, offset);
2753	#endif
2754	unlock_page(page);
2755	return 0; /* don't care */
2756	}
2757
2758	/*
2759	* The page straddles i_size. It must be zeroed out on each and every
2760	* writepage invocation because it may be mmapped. "A file is mapped
2761	* in multiples of the page size. For a file that is not a multiple of
2762	* the page size, the remaining memory is zeroed when mapped, and
2763	* writes to that region are not written out to the file."
2764	*/
2765	zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2766	out:
2767	ret = mpage_writepage(page, get_block, wbc);
2768	if (ret == -EAGAIN)
2769	ret = __block_write_full_page(inode, page, get_block, wbc,
2770	end_buffer_async_write);
2771	return ret;
2772	}
2773	EXPORT_SYMBOL(nobh_writepage);
2774
2775	int nobh_truncate_page(struct address_space *mapping,
2776	loff_t from, get_block_t *get_block)
2777	{
2778	pgoff_t index = from >> PAGE_CACHE_SHIFT;
2779	unsigned offset = from & (PAGE_CACHE_SIZE-1);
2780	unsigned blocksize;
2781	sector_t iblock;
2782	unsigned length, pos;
2783	struct inode *inode = mapping->host;
2784	struct page *page;
2785	struct buffer_head map_bh;
2786	int err;
2787
2788	blocksize = 1 << inode->i_blkbits;
2789	length = offset & (blocksize - 1);
2790
2791	/* Block boundary? Nothing to do */
2792	if (!length)
2793	return 0;
2794
2795	length = blocksize - length;
2796	iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2797
2798	page = grab_cache_page(mapping, index);
2799	err = -ENOMEM;
2800	if (!page)
2801	goto out;
2802
2803	if (page_has_buffers(page)) {
2804	has_buffers:
2805	unlock_page(page);
2806	page_cache_release(page);
2807	return block_truncate_page(mapping, from, get_block);
2808	}
2809
2810	/* Find the buffer that contains "offset" */
2811	pos = blocksize;
2812	while (offset >= pos) {
2813	iblock++;
2814	pos += blocksize;
2815	}
2816
2817	map_bh.b_size = blocksize;
2818	map_bh.b_state = 0;
2819	err = get_block(inode, iblock, &map_bh, 0);
2820	if (err)
2821	goto unlock;
2822	/* unmapped? It's a hole - nothing to do */
2823	if (!buffer_mapped(&map_bh))
2824	goto unlock;
2825
2826	/* Ok, it's mapped. Make sure it's up-to-date */
2827	if (!PageUptodate(page)) {
2828	err = mapping->a_ops->readpage(NULL, page);
2829	if (err) {
2830	page_cache_release(page);
2831	goto out;
2832	}
2833	lock_page(page);
2834	if (!PageUptodate(page)) {
2835	err = -EIO;
2836	goto unlock;
2837	}
2838	if (page_has_buffers(page))
2839	goto has_buffers;
2840	}
2841	zero_user(page, offset, length);
2842	set_page_dirty(page);
2843	err = 0;
2844
2845	unlock:
2846	unlock_page(page);
2847	page_cache_release(page);
2848	out:
2849	return err;
2850	}
2851	EXPORT_SYMBOL(nobh_truncate_page);
2852
2853	int block_truncate_page(struct address_space *mapping,
2854	loff_t from, get_block_t *get_block)
2855	{
2856	pgoff_t index = from >> PAGE_CACHE_SHIFT;
2857	unsigned offset = from & (PAGE_CACHE_SIZE-1);
2858	unsigned blocksize;
2859	sector_t iblock;
2860	unsigned length, pos;
2861	struct inode *inode = mapping->host;
2862	struct page *page;
2863	struct buffer_head *bh;
2864	int err;
2865
2866	blocksize = 1 << inode->i_blkbits;
2867	length = offset & (blocksize - 1);
2868
2869	/* Block boundary? Nothing to do */
2870	if (!length)
2871	return 0;
2872
2873	length = blocksize - length;
2874	iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2875
2876	page = grab_cache_page(mapping, index);
2877	err = -ENOMEM;
2878	if (!page)
2879	goto out;
2880
2881	if (!page_has_buffers(page))
2882	create_empty_buffers(page, blocksize, 0);
2883
2884	/* Find the buffer that contains "offset" */
2885	bh = page_buffers(page);
2886	pos = blocksize;
2887	while (offset >= pos) {
2888	bh = bh->b_this_page;
2889	iblock++;
2890	pos += blocksize;
2891	}
2892
2893	err = 0;
2894	if (!buffer_mapped(bh)) {
2895	WARN_ON(bh->b_size != blocksize);
2896	err = get_block(inode, iblock, bh, 0);
2897	if (err)
2898	goto unlock;
2899	/* unmapped? It's a hole - nothing to do */
2900	if (!buffer_mapped(bh))
2901	goto unlock;
2902	}
2903
2904	/* Ok, it's mapped. Make sure it's up-to-date */
2905	if (PageUptodate(page))
2906	set_buffer_uptodate(bh);
2907
2908	if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2909	err = -EIO;
2910	ll_rw_block(READ, 1, &bh);
2911	wait_on_buffer(bh);
2912	/* Uhhuh. Read error. Complain and punt. */
2913	if (!buffer_uptodate(bh))
2914	goto unlock;
2915	}
2916
2917	zero_user(page, offset, length);
2918	mark_buffer_dirty(bh);
2919	err = 0;
2920
2921	unlock:
2922	unlock_page(page);
2923	page_cache_release(page);
2924	out:
2925	return err;
2926	}
2927	EXPORT_SYMBOL(block_truncate_page);
2928
2929	/*
2930	* The generic ->writepage function for buffer-backed address_spaces
2931	* this form passes in the end_io handler used to finish the IO.
2932	*/
2933	int block_write_full_page_endio(struct page page, get_block_t get_block,
2934	struct writeback_control wbc, bh_end_io_t handler)
2935	{
2936	struct inode * const inode = page->mapping->host;
2937	loff_t i_size = i_size_read(inode);
2938	const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2939	unsigned offset;
2940
2941	/* Is the page fully inside i_size? */
2942	if (page->index < end_index)
2943	return __block_write_full_page(inode, page, get_block, wbc,
2944	handler);
2945
2946	/* Is the page fully outside i_size? (truncate in progress) */
2947	offset = i_size & (PAGE_CACHE_SIZE-1);
2948	if (page->index >= end_index+1 \|\| !offset) {
2949	/*
2950	* The page may have dirty, unmapped buffers. For example,
2951	* they may have been added in ext3_writepage(). Make them
2952	* freeable here, so the page does not leak.
2953	*/
2954	do_invalidatepage(page, 0);
2955	unlock_page(page);
2956	return 0; /* don't care */
2957	}
2958
2959	/*
2960	* The page straddles i_size. It must be zeroed out on each and every
2961	* writepage invocation because it may be mmapped. "A file is mapped
2962	* in multiples of the page size. For a file that is not a multiple of
2963	* the page size, the remaining memory is zeroed when mapped, and
2964	* writes to that region are not written out to the file."
2965	*/
2966	zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2967	return __block_write_full_page(inode, page, get_block, wbc, handler);
2968	}
2969	EXPORT_SYMBOL(block_write_full_page_endio);
2970
2971	/*
2972	* The generic ->writepage function for buffer-backed address_spaces
2973	*/
2974	int block_write_full_page(struct page page, get_block_t get_block,
2975	struct writeback_control *wbc)
2976	{
2977	return block_write_full_page_endio(page, get_block, wbc,
2978	end_buffer_async_write);
2979	}
2980	EXPORT_SYMBOL(block_write_full_page);
2981
2982	sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2983	get_block_t *get_block)
2984	{
2985	struct buffer_head tmp;
2986	struct inode *inode = mapping->host;
2987	tmp.b_state = 0;
2988	tmp.b_blocknr = 0;
2989	tmp.b_size = 1 << inode->i_blkbits;
2990	get_block(inode, block, &tmp, 0);
2991	return tmp.b_blocknr;
2992	}
2993	EXPORT_SYMBOL(generic_block_bmap);
2994
2995	static void end_bio_bh_io_sync(struct bio *bio, int err)
2996	{
2997	struct buffer_head *bh = bio->bi_private;
2998
2999	if (err == -EOPNOTSUPP) {
3000	set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
3001	set_bit(BH_Eopnotsupp, &bh->b_state);
3002	}
3003
3004	if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
3005	set_bit(BH_Quiet, &bh->b_state);
3006
3007	bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
3008	bio_put(bio);
3009	}
3010
3011	int submit_bh(int rw, struct buffer_head * bh)
3012	{
3013	struct bio *bio;
3014	int ret = 0;
3015
3016	BUG_ON(!buffer_locked(bh));
3017	BUG_ON(!buffer_mapped(bh));
3018	BUG_ON(!bh->b_end_io);
3019	BUG_ON(buffer_delay(bh));
3020	BUG_ON(buffer_unwritten(bh));
3021
3022	/*
3023	* Mask in barrier bit for a write (could be either a WRITE or a
3024	* WRITE_SYNC
3025	*/
3026	if (buffer_ordered(bh) && (rw & WRITE))
3027	rw \|= WRITE_BARRIER;
3028
3029	/*
3030	* Only clear out a write error when rewriting
3031	*/
3032	if (test_set_buffer_req(bh) && (rw & WRITE))
3033	clear_buffer_write_io_error(bh);
3034
3035	/*
3036	* from here on down, it's all bio -- do the initial mapping,
3037	* submit_bio -> generic_make_request may further map this bio around
3038	*/
3039	bio = bio_alloc(GFP_NOIO, 1);
3040
3041	bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3042	bio->bi_bdev = bh->b_bdev;
3043	bio->bi_io_vec[0].bv_page = bh->b_page;
3044	bio->bi_io_vec[0].bv_len = bh->b_size;
3045	bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3046
3047	bio->bi_vcnt = 1;
3048	bio->bi_idx = 0;
3049	bio->bi_size = bh->b_size;
3050
3051	bio->bi_end_io = end_bio_bh_io_sync;
3052	bio->bi_private = bh;
3053
3054	bio_get(bio);
3055	submit_bio(rw, bio);
3056
3057	if (bio_flagged(bio, BIO_EOPNOTSUPP))
3058	ret = -EOPNOTSUPP;
3059
3060	bio_put(bio);
3061	return ret;
3062	}
3063	EXPORT_SYMBOL(submit_bh);
3064
3065	/**
3066	* ll_rw_block: low-level access to block devices (DEPRECATED)
3067	* @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
3068	* @nr: number of &struct buffer_heads in the array
3069	* @bhs: array of pointers to &struct buffer_head
3070	*
3071	* ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3072	* requests an I/O operation on them, either a %READ or a %WRITE. The third
3073	* %SWRITE is like %WRITE only we make sure that the current data in buffers
3074	* are sent to disk. The fourth %READA option is described in the documentation
3075	* for generic_make_request() which ll_rw_block() calls.
3076	*
3077	* This function drops any buffer that it cannot get a lock on (with the
3078	* BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
3079	* clean when doing a write request, and any buffer that appears to be
3080	* up-to-date when doing read request. Further it marks as clean buffers that
3081	* are processed for writing (the buffer cache won't assume that they are
3082	* actually clean until the buffer gets unlocked).
3083	*
3084	* ll_rw_block sets b_end_io to simple completion handler that marks
3085	* the buffer up-to-date (if approriate), unlocks the buffer and wakes
3086	* any waiters.
3087	*
3088	* All of the buffers must be for the same device, and must also be a
3089	* multiple of the current approved size for the device.
3090	*/
3091	void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3092	{
3093	int i;
3094
3095	for (i = 0; i < nr; i++) {
3096	struct buffer_head *bh = bhs[i];
3097
3098	if (rw == SWRITE \|\| rw == SWRITE_SYNC \|\| rw == SWRITE_SYNC_PLUG)
3099	lock_buffer(bh);
3100	else if (!trylock_buffer(bh))
3101	continue;
3102
3103	if (rw == WRITE \|\| rw == SWRITE \|\| rw == SWRITE_SYNC \|\|
3104	rw == SWRITE_SYNC_PLUG) {
3105	if (test_clear_buffer_dirty(bh)) {
3106	bh->b_end_io = end_buffer_write_sync;
3107	get_bh(bh);
3108	if (rw == SWRITE_SYNC)
3109	submit_bh(WRITE_SYNC, bh);
3110	else
3111	submit_bh(WRITE, bh);
3112	continue;
3113	}
3114	} else {
3115	if (!buffer_uptodate(bh)) {
3116	bh->b_end_io = end_buffer_read_sync;
3117	get_bh(bh);
3118	submit_bh(rw, bh);
3119	continue;
3120	}
3121	}
3122	unlock_buffer(bh);
3123	}
3124	}
3125	EXPORT_SYMBOL(ll_rw_block);
3126
3127	/*
3128	* For a data-integrity writeout, we need to wait upon any in-progress I/O
3129	* and then start new I/O and then wait upon it. The caller must have a ref on
3130	* the buffer_head.
3131	*/
3132	int sync_dirty_buffer(struct buffer_head *bh)
3133	{
3134	int ret = 0;
3135
3136	WARN_ON(atomic_read(&bh->b_count) < 1);
3137	lock_buffer(bh);
3138	if (test_clear_buffer_dirty(bh)) {
3139	get_bh(bh);
3140	bh->b_end_io = end_buffer_write_sync;
3141	ret = submit_bh(WRITE_SYNC, bh);
3142	wait_on_buffer(bh);
3143	if (buffer_eopnotsupp(bh)) {
3144	clear_buffer_eopnotsupp(bh);
3145	ret = -EOPNOTSUPP;
3146	}
3147	if (!ret && !buffer_uptodate(bh))
3148	ret = -EIO;
3149	} else {
3150	unlock_buffer(bh);
3151	}
3152	return ret;
3153	}
3154	EXPORT_SYMBOL(sync_dirty_buffer);
3155
3156	/*
3157	* try_to_free_buffers() checks if all the buffers on this particular page
3158	* are unused, and releases them if so.
3159	*
3160	* Exclusion against try_to_free_buffers may be obtained by either
3161	* locking the page or by holding its mapping's private_lock.
3162	*
3163	* If the page is dirty but all the buffers are clean then we need to
3164	* be sure to mark the page clean as well. This is because the page
3165	* may be against a block device, and a later reattachment of buffers
3166	* to a dirty page will set all buffers dirty. Which would corrupt
3167	* filesystem data on the same device.
3168	*
3169	* The same applies to regular filesystem pages: if all the buffers are
3170	* clean then we set the page clean and proceed. To do that, we require
3171	* total exclusion from __set_page_dirty_buffers(). That is obtained with
3172	* private_lock.
3173	*
3174	* try_to_free_buffers() is non-blocking.
3175	*/
3176	static inline int buffer_busy(struct buffer_head *bh)
3177	{
3178	return atomic_read(&bh->b_count) \|
3179	(bh->b_state & ((1 << BH_Dirty) \| (1 << BH_Lock)));
3180	}
3181
3182	static int
3183	drop_buffers(struct page page, struct buffer_head *buffers_to_free)
3184	{
3185	struct buffer_head *head = page_buffers(page);
3186	struct buffer_head *bh;
3187
3188	bh = head;
3189	do {
3190	if (buffer_write_io_error(bh) && page->mapping)
3191	set_bit(AS_EIO, &page->mapping->flags);
3192	if (buffer_busy(bh))
3193	goto failed;
3194	bh = bh->b_this_page;
3195	} while (bh != head);
3196
3197	do {
3198	struct buffer_head *next = bh->b_this_page;
3199
3200	if (bh->b_assoc_map)
3201	__remove_assoc_queue(bh);
3202	bh = next;
3203	} while (bh != head);
3204	*buffers_to_free = head;
3205	__clear_page_buffers(page);
3206	return 1;
3207	failed:
3208	return 0;
3209	}
3210
3211	int try_to_free_buffers(struct page *page)
3212	{
3213	struct address_space * const mapping = page->mapping;
3214	struct buffer_head *buffers_to_free = NULL;
3215	int ret = 0;
3216
3217	BUG_ON(!PageLocked(page));
3218	if (PageWriteback(page))
3219	return 0;
3220
3221	if (mapping == NULL) { /* can this still happen? */
3222	ret = drop_buffers(page, &buffers_to_free);
3223	goto out;
3224	}
3225
3226	spin_lock(&mapping->private_lock);
3227	ret = drop_buffers(page, &buffers_to_free);
3228
3229	/*
3230	* If the filesystem writes its buffers by hand (eg ext3)
3231	* then we can have clean buffers against a dirty page. We
3232	* clean the page here; otherwise the VM will never notice
3233	* that the filesystem did any IO at all.
3234	*
3235	* Also, during truncate, discard_buffer will have marked all
3236	* the page's buffers clean. We discover that here and clean
3237	* the page also.
3238	*
3239	* private_lock must be held over this entire operation in order
3240	* to synchronise against __set_page_dirty_buffers and prevent the
3241	* dirty bit from being lost.
3242	*/
3243	if (ret)
3244	cancel_dirty_page(page, PAGE_CACHE_SIZE);
3245	spin_unlock(&mapping->private_lock);
3246	out:
3247	if (buffers_to_free) {
3248	struct buffer_head *bh = buffers_to_free;
3249
3250	do {
3251	struct buffer_head *next = bh->b_this_page;
3252	free_buffer_head(bh);
3253	bh = next;
3254	} while (bh != buffers_to_free);
3255	}
3256	return ret;
3257	}
3258	EXPORT_SYMBOL(try_to_free_buffers);
3259
3260	void block_sync_page(struct page *page)
3261	{
3262	struct address_space *mapping;
3263
3264	smp_mb();
3265	mapping = page_mapping(page);
3266	if (mapping)
3267	blk_run_backing_dev(mapping->backing_dev_info, page);
3268	}
3269	EXPORT_SYMBOL(block_sync_page);
3270
3271	/*
3272	* There are no bdflush tunables left. But distributions are
3273	* still running obsolete flush daemons, so we terminate them here.
3274	*
3275	* Use of bdflush() is deprecated and will be removed in a future kernel.
3276	* The `flush-X' kernel threads fully replace bdflush daemons and this call.
3277	*/
3278	SYSCALL_DEFINE2(bdflush, int, func, long, data)
3279	{
3280	static int msg_count;
3281
3282	if (!capable(CAP_SYS_ADMIN))
3283	return -EPERM;
3284
3285	if (msg_count < 5) {
3286	msg_count++;
3287	printk(KERN_INFO
3288	"warning: process `%s' used the obsolete bdflush"
3289	" system call\n", current->comm);
3290	printk(KERN_INFO "Fix your initscripts?\n");
3291	}
3292
3293	if (func == 1)
3294	do_exit(0);
3295	return 0;
3296	}
3297
3298	/*
3299	* Buffer-head allocation
3300	*/
3301	static struct kmem_cache *bh_cachep;
3302
3303	/*
3304	* Once the number of bh's in the machine exceeds this level, we start
3305	* stripping them in writeback.
3306	*/
3307	static int max_buffer_heads;
3308
3309	int buffer_heads_over_limit;
3310
3311	struct bh_accounting {
3312	int nr; /* Number of live bh's */
3313	int ratelimit; /* Limit cacheline bouncing */
3314	};
3315
3316	static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3317
3318	static void recalc_bh_state(void)
3319	{
3320	int i;
3321	int tot = 0;
3322
3323	if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3324	return;
3325	__get_cpu_var(bh_accounting).ratelimit = 0;
3326	for_each_online_cpu(i)
3327	tot += per_cpu(bh_accounting, i).nr;
3328	buffer_heads_over_limit = (tot > max_buffer_heads);
3329	}
3330
3331	struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3332	{
3333	struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3334	if (ret) {
3335	INIT_LIST_HEAD(&ret->b_assoc_buffers);
3336	get_cpu_var(bh_accounting).nr++;
3337	recalc_bh_state();
3338	put_cpu_var(bh_accounting);
3339	}
3340	return ret;
3341	}
3342	EXPORT_SYMBOL(alloc_buffer_head);
3343
3344	void free_buffer_head(struct buffer_head *bh)
3345	{
3346	BUG_ON(!list_empty(&bh->b_assoc_buffers));
3347	kmem_cache_free(bh_cachep, bh);
3348	get_cpu_var(bh_accounting).nr--;
3349	recalc_bh_state();
3350	put_cpu_var(bh_accounting);
3351	}
3352	EXPORT_SYMBOL(free_buffer_head);
3353
3354	static void buffer_exit_cpu(int cpu)
3355	{
3356	int i;
3357	struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3358
3359	for (i = 0; i < BH_LRU_SIZE; i++) {
3360	brelse(b->bhs[i]);
3361	b->bhs[i] = NULL;
3362	}
3363	get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3364	per_cpu(bh_accounting, cpu).nr = 0;
3365	put_cpu_var(bh_accounting);
3366	}
3367
3368	static int buffer_cpu_notify(struct notifier_block *self,
3369	unsigned long action, void *hcpu)
3370	{
3371	if (action == CPU_DEAD \|\| action == CPU_DEAD_FROZEN)
3372	buffer_exit_cpu((unsigned long)hcpu);
3373	return NOTIFY_OK;
3374	}
3375
3376	/**
3377	* bh_uptodate_or_lock - Test whether the buffer is uptodate
3378	* @bh: struct buffer_head
3379	*
3380	* Return true if the buffer is up-to-date and false,
3381	* with the buffer locked, if not.
3382	*/
3383	int bh_uptodate_or_lock(struct buffer_head *bh)
3384	{
3385	if (!buffer_uptodate(bh)) {
3386	lock_buffer(bh);
3387	if (!buffer_uptodate(bh))
3388	return 0;
3389	unlock_buffer(bh);
3390	}
3391	return 1;
3392	}
3393	EXPORT_SYMBOL(bh_uptodate_or_lock);
3394
3395	/**
3396	* bh_submit_read - Submit a locked buffer for reading
3397	* @bh: struct buffer_head
3398	*
3399	* Returns zero on success and -EIO on error.
3400	*/
3401	int bh_submit_read(struct buffer_head *bh)
3402	{
3403	BUG_ON(!buffer_locked(bh));
3404
3405	if (buffer_uptodate(bh)) {
3406	unlock_buffer(bh);
3407	return 0;
3408	}
3409
3410	get_bh(bh);
3411	bh->b_end_io = end_buffer_read_sync;
3412	submit_bh(READ, bh);
3413	wait_on_buffer(bh);
3414	if (buffer_uptodate(bh))
3415	return 0;
3416	return -EIO;
3417	}
3418	EXPORT_SYMBOL(bh_submit_read);
3419
3420	void __init buffer_init(void)
3421	{
3422	int nrpages;
3423
3424	bh_cachep = kmem_cache_create("buffer_head",
3425	sizeof(struct buffer_head), 0,
3426	(SLAB_RECLAIM_ACCOUNT\|SLAB_PANIC\|
3427	SLAB_MEM_SPREAD),
3428	NULL);
3429
3430	/*
3431	* Limit the bh occupancy to 10% of ZONE_NORMAL
3432	*/
3433	nrpages = (nr_free_buffer_pages() * 10) / 100;
3434	max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3435	hotcpu_notifier(buffer_cpu_notify, 0);
3436	}