linux - fs/buffer.c

3418 lines · 88.3 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/export.h>
33	#include <linux/backing-dev.h>
34	#include <linux/writeback.h>
35	#include <linux/hash.h>
36	#include <linux/suspend.h>
37	#include <linux/buffer_head.h>
38	#include <linux/task_io_accounting_ops.h>
39	#include <linux/bio.h>
40	#include <linux/notifier.h>
41	#include <linux/cpu.h>
42	#include <linux/bitops.h>
43	#include <linux/mpage.h>
44	#include <linux/bit_spinlock.h>
45	#include <trace/events/block.h>
46
47	static int fsync_buffers_list(spinlock_t lock, struct list_head list);
48	static int submit_bh_wbc(int rw, struct buffer_head *bh,
49	unsigned long bio_flags,
50	struct writeback_control *wbc);
51
52	#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
53
54	void init_buffer(struct buffer_head bh, bh_end_io_t handler, void *private)
55	{
56	bh->b_end_io = handler;
57	bh->b_private = private;
58	}
59	EXPORT_SYMBOL(init_buffer);
60
61	inline void touch_buffer(struct buffer_head *bh)
62	{
63	trace_block_touch_buffer(bh);
64	mark_page_accessed(bh->b_page);
65	}
66	EXPORT_SYMBOL(touch_buffer);
67
68	void __lock_buffer(struct buffer_head *bh)
69	{
70	wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
71	}
72	EXPORT_SYMBOL(__lock_buffer);
73
74	void unlock_buffer(struct buffer_head *bh)
75	{
76	clear_bit_unlock(BH_Lock, &bh->b_state);
77	smp_mb__after_atomic();
78	wake_up_bit(&bh->b_state, BH_Lock);
79	}
80	EXPORT_SYMBOL(unlock_buffer);
81
82	/*
83	* Returns if the page has dirty or writeback buffers. If all the buffers
84	* are unlocked and clean then the PageDirty information is stale. If
85	* any of the pages are locked, it is assumed they are locked for IO.
86	*/
87	void buffer_check_dirty_writeback(struct page *page,
88	bool dirty, bool writeback)
89	{
90	struct buffer_head head, bh;
91	*dirty = false;
92	*writeback = false;
93
94	BUG_ON(!PageLocked(page));
95
96	if (!page_has_buffers(page))
97	return;
98
99	if (PageWriteback(page))
100	*writeback = true;
101
102	head = page_buffers(page);
103	bh = head;
104	do {
105	if (buffer_locked(bh))
106	*writeback = true;
107
108	if (buffer_dirty(bh))
109	*dirty = true;
110
111	bh = bh->b_this_page;
112	} while (bh != head);
113	}
114	EXPORT_SYMBOL(buffer_check_dirty_writeback);
115
116	/*
117	* Block until a buffer comes unlocked. This doesn't stop it
118	* from becoming locked again - you have to lock it yourself
119	* if you want to preserve its state.
120	*/
121	void __wait_on_buffer(struct buffer_head * bh)
122	{
123	wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
124	}
125	EXPORT_SYMBOL(__wait_on_buffer);
126
127	static void
128	__clear_page_buffers(struct page *page)
129	{
130	ClearPagePrivate(page);
131	set_page_private(page, 0);
132	put_page(page);
133	}
134
135	static void buffer_io_error(struct buffer_head bh, char msg)
136	{
137	if (!test_bit(BH_Quiet, &bh->b_state))
138	printk_ratelimited(KERN_ERR
139	"Buffer I/O error on dev %pg, logical block %llu%s\n",
140	bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
141	}
142
143	/*
144	* End-of-IO handler helper function which does not touch the bh after
145	* unlocking it.
146	* Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
147	* a race there is benign: unlock_buffer() only use the bh's address for
148	* hashing after unlocking the buffer, so it doesn't actually touch the bh
149	* itself.
150	*/
151	static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
152	{
153	if (uptodate) {
154	set_buffer_uptodate(bh);
155	} else {
156	/* This happens, due to failed READA attempts. */
157	clear_buffer_uptodate(bh);
158	}
159	unlock_buffer(bh);
160	}
161
162	/*
163	* Default synchronous end-of-IO handler.. Just mark it up-to-date and
164	* unlock the buffer. This is what ll_rw_block uses too.
165	*/
166	void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
167	{
168	__end_buffer_read_notouch(bh, uptodate);
169	put_bh(bh);
170	}
171	EXPORT_SYMBOL(end_buffer_read_sync);
172
173	void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
174	{
175	if (uptodate) {
176	set_buffer_uptodate(bh);
177	} else {
178	buffer_io_error(bh, ", lost sync page write");
179	set_buffer_write_io_error(bh);
180	clear_buffer_uptodate(bh);
181	}
182	unlock_buffer(bh);
183	put_bh(bh);
184	}
185	EXPORT_SYMBOL(end_buffer_write_sync);
186
187	/*
188	* Various filesystems appear to want __find_get_block to be non-blocking.
189	* But it's the page lock which protects the buffers. To get around this,
190	* we get exclusion from try_to_free_buffers with the blockdev mapping's
191	* private_lock.
192	*
193	* Hack idea: for the blockdev mapping, i_bufferlist_lock contention
194	* may be quite high. This code could TryLock the page, and if that
195	* succeeds, there is no need to take private_lock. (But if
196	* private_lock is contended then so is mapping->tree_lock).
197	*/
198	static struct buffer_head *
199	__find_get_block_slow(struct block_device *bdev, sector_t block)
200	{
201	struct inode *bd_inode = bdev->bd_inode;
202	struct address_space *bd_mapping = bd_inode->i_mapping;
203	struct buffer_head *ret = NULL;
204	pgoff_t index;
205	struct buffer_head *bh;
206	struct buffer_head *head;
207	struct page *page;
208	int all_mapped = 1;
209
210	index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
211	page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
212	if (!page)
213	goto out;
214
215	spin_lock(&bd_mapping->private_lock);
216	if (!page_has_buffers(page))
217	goto out_unlock;
218	head = page_buffers(page);
219	bh = head;
220	do {
221	if (!buffer_mapped(bh))
222	all_mapped = 0;
223	else if (bh->b_blocknr == block) {
224	ret = bh;
225	get_bh(bh);
226	goto out_unlock;
227	}
228	bh = bh->b_this_page;
229	} while (bh != head);
230
231	/* we might be here because some of the buffers on this page are
232	* not mapped. This is due to various races between
233	* file io on the block device and getblk. It gets dealt with
234	* elsewhere, don't buffer_error if we had some unmapped buffers
235	*/
236	if (all_mapped) {
237	printk("__find_get_block_slow() failed. "
238	"block=%llu, b_blocknr=%llu\n",
239	(unsigned long long)block,
240	(unsigned long long)bh->b_blocknr);
241	printk("b_state=0x%08lx, b_size=%zu\n",
242	bh->b_state, bh->b_size);
243	printk("device %pg blocksize: %d\n", bdev,
244	1 << bd_inode->i_blkbits);
245	}
246	out_unlock:
247	spin_unlock(&bd_mapping->private_lock);
248	put_page(page);
249	out:
250	return ret;
251	}
252
253	/*
254	* Kick the writeback threads then try to free up some ZONE_NORMAL memory.
255	*/
256	static void free_more_memory(void)
257	{
258	struct zoneref *z;
259	int nid;
260
261	wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
262	yield();
263
264	for_each_online_node(nid) {
265
266	z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
267	gfp_zone(GFP_NOFS), NULL);
268	if (z->zone)
269	try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
270	GFP_NOFS, NULL);
271	}
272	}
273
274	/*
275	* I/O completion handler for block_read_full_page() - pages
276	* which come unlocked at the end of I/O.
277	*/
278	static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
279	{
280	unsigned long flags;
281	struct buffer_head *first;
282	struct buffer_head *tmp;
283	struct page *page;
284	int page_uptodate = 1;
285
286	BUG_ON(!buffer_async_read(bh));
287
288	page = bh->b_page;
289	if (uptodate) {
290	set_buffer_uptodate(bh);
291	} else {
292	clear_buffer_uptodate(bh);
293	buffer_io_error(bh, ", async page read");
294	SetPageError(page);
295	}
296
297	/*
298	* Be _very_ careful from here on. Bad things can happen if
299	* two buffer heads end IO at almost the same time and both
300	* decide that the page is now completely done.
301	*/
302	first = page_buffers(page);
303	local_irq_save(flags);
304	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
305	clear_buffer_async_read(bh);
306	unlock_buffer(bh);
307	tmp = bh;
308	do {
309	if (!buffer_uptodate(tmp))
310	page_uptodate = 0;
311	if (buffer_async_read(tmp)) {
312	BUG_ON(!buffer_locked(tmp));
313	goto still_busy;
314	}
315	tmp = tmp->b_this_page;
316	} while (tmp != bh);
317	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
318	local_irq_restore(flags);
319
320	/*
321	* If none of the buffers had errors and they are all
322	* uptodate then we can set the page uptodate.
323	*/
324	if (page_uptodate && !PageError(page))
325	SetPageUptodate(page);
326	unlock_page(page);
327	return;
328
329	still_busy:
330	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
331	local_irq_restore(flags);
332	return;
333	}
334
335	/*
336	* Completion handler for block_write_full_page() - pages which are unlocked
337	* during I/O, and which have PageWriteback cleared upon I/O completion.
338	*/
339	void end_buffer_async_write(struct buffer_head *bh, int uptodate)
340	{
341	unsigned long flags;
342	struct buffer_head *first;
343	struct buffer_head *tmp;
344	struct page *page;
345
346	BUG_ON(!buffer_async_write(bh));
347
348	page = bh->b_page;
349	if (uptodate) {
350	set_buffer_uptodate(bh);
351	} else {
352	buffer_io_error(bh, ", lost async page write");
353	set_bit(AS_EIO, &page->mapping->flags);
354	set_buffer_write_io_error(bh);
355	clear_buffer_uptodate(bh);
356	SetPageError(page);
357	}
358
359	first = page_buffers(page);
360	local_irq_save(flags);
361	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
362
363	clear_buffer_async_write(bh);
364	unlock_buffer(bh);
365	tmp = bh->b_this_page;
366	while (tmp != bh) {
367	if (buffer_async_write(tmp)) {
368	BUG_ON(!buffer_locked(tmp));
369	goto still_busy;
370	}
371	tmp = tmp->b_this_page;
372	}
373	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
374	local_irq_restore(flags);
375	end_page_writeback(page);
376	return;
377
378	still_busy:
379	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
380	local_irq_restore(flags);
381	return;
382	}
383	EXPORT_SYMBOL(end_buffer_async_write);
384
385	/*
386	* If a page's buffers are under async readin (end_buffer_async_read
387	* completion) then there is a possibility that another thread of
388	* control could lock one of the buffers after it has completed
389	* but while some of the other buffers have not completed. This
390	* locked buffer would confuse end_buffer_async_read() into not unlocking
391	* the page. So the absence of BH_Async_Read tells end_buffer_async_read()
392	* that this buffer is not under async I/O.
393	*
394	* The page comes unlocked when it has no locked buffer_async buffers
395	* left.
396	*
397	* PageLocked prevents anyone starting new async I/O reads any of
398	* the buffers.
399	*
400	* PageWriteback is used to prevent simultaneous writeout of the same
401	* page.
402	*
403	* PageLocked prevents anyone from starting writeback of a page which is
404	* under read I/O (PageWriteback is only ever set against a locked page).
405	*/
406	static void mark_buffer_async_read(struct buffer_head *bh)
407	{
408	bh->b_end_io = end_buffer_async_read;
409	set_buffer_async_read(bh);
410	}
411
412	static void mark_buffer_async_write_endio(struct buffer_head *bh,
413	bh_end_io_t *handler)
414	{
415	bh->b_end_io = handler;
416	set_buffer_async_write(bh);
417	}
418
419	void mark_buffer_async_write(struct buffer_head *bh)
420	{
421	mark_buffer_async_write_endio(bh, end_buffer_async_write);
422	}
423	EXPORT_SYMBOL(mark_buffer_async_write);
424
425
426	/*
427	* fs/buffer.c contains helper functions for buffer-backed address space's
428	* fsync functions. A common requirement for buffer-based filesystems is
429	* that certain data from the backing blockdev needs to be written out for
430	* a successful fsync(). For example, ext2 indirect blocks need to be
431	* written back and waited upon before fsync() returns.
432	*
433	* The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
434	* inode_has_buffers() and invalidate_inode_buffers() are provided for the
435	* management of a list of dependent buffers at ->i_mapping->private_list.
436	*
437	* Locking is a little subtle: try_to_free_buffers() will remove buffers
438	* from their controlling inode's queue when they are being freed. But
439	* try_to_free_buffers() will be operating against the blockdev mapping
440	* at the time, not against the S_ISREG file which depends on those buffers.
441	* So the locking for private_list is via the private_lock in the address_space
442	* which backs the buffers. Which is different from the address_space
443	* against which the buffers are listed. So for a particular address_space,
444	* mapping->private_lock does not protect mapping->private_list! In fact,
445	* mapping->private_list will always be protected by the backing blockdev's
446	* ->private_lock.
447	*
448	* Which introduces a requirement: all buffers on an address_space's
449	* ->private_list must be from the same address_space: the blockdev's.
450	*
451	* address_spaces which do not place buffers at ->private_list via these
452	* utility functions are free to use private_lock and private_list for
453	* whatever they want. The only requirement is that list_empty(private_list)
454	* be true at clear_inode() time.
455	*
456	* FIXME: clear_inode should not call invalidate_inode_buffers(). The
457	* filesystems should do that. invalidate_inode_buffers() should just go
458	* BUG_ON(!list_empty).
459	*
460	* FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
461	* take an address_space, not an inode. And it should be called
462	* mark_buffer_dirty_fsync() to clearly define why those buffers are being
463	* queued up.
464	*
465	* FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
466	* list if it is already on a list. Because if the buffer is on a list,
467	* it must already be on the right one. If not, the filesystem is being
468	* silly. This will save a ton of locking. But first we have to ensure
469	* that buffers are taken off the old inode's list when they are freed
470	* (presumably in truncate). That requires careful auditing of all
471	* filesystems (do it inside bforget()). It could also be done by bringing
472	* b_inode back.
473	*/
474
475	/*
476	* The buffer's backing address_space's private_lock must be held
477	*/
478	static void __remove_assoc_queue(struct buffer_head *bh)
479	{
480	list_del_init(&bh->b_assoc_buffers);
481	WARN_ON(!bh->b_assoc_map);
482	if (buffer_write_io_error(bh))
483	set_bit(AS_EIO, &bh->b_assoc_map->flags);
484	bh->b_assoc_map = NULL;
485	}
486
487	int inode_has_buffers(struct inode *inode)
488	{
489	return !list_empty(&inode->i_data.private_list);
490	}
491
492	/*
493	* osync is designed to support O_SYNC io. It waits synchronously for
494	* all already-submitted IO to complete, but does not queue any new
495	* writes to the disk.
496	*
497	* To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
498	* you dirty the buffers, and then use osync_inode_buffers to wait for
499	* completion. Any other dirty buffers which are not yet queued for
500	* write will not be flushed to disk by the osync.
501	*/
502	static int osync_buffers_list(spinlock_t lock, struct list_head list)
503	{
504	struct buffer_head *bh;
505	struct list_head *p;
506	int err = 0;
507
508	spin_lock(lock);
509	repeat:
510	list_for_each_prev(p, list) {
511	bh = BH_ENTRY(p);
512	if (buffer_locked(bh)) {
513	get_bh(bh);
514	spin_unlock(lock);
515	wait_on_buffer(bh);
516	if (!buffer_uptodate(bh))
517	err = -EIO;
518	brelse(bh);
519	spin_lock(lock);
520	goto repeat;
521	}
522	}
523	spin_unlock(lock);
524	return err;
525	}
526
527	static void do_thaw_one(struct super_block sb, void unused)
528	{
529	while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
530	printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
531	}
532
533	static void do_thaw_all(struct work_struct *work)
534	{
535	iterate_supers(do_thaw_one, NULL);
536	kfree(work);
537	printk(KERN_WARNING "Emergency Thaw complete\n");
538	}
539
540	/**
541	* emergency_thaw_all -- forcibly thaw every frozen filesystem
542	*
543	* Used for emergency unfreeze of all filesystems via SysRq
544	*/
545	void emergency_thaw_all(void)
546	{
547	struct work_struct *work;
548
549	work = kmalloc(sizeof(*work), GFP_ATOMIC);
550	if (work) {
551	INIT_WORK(work, do_thaw_all);
552	schedule_work(work);
553	}
554	}
555
556	/**
557	* sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
558	* @mapping: the mapping which wants those buffers written
559	*
560	* Starts I/O against the buffers at mapping->private_list, and waits upon
561	* that I/O.
562	*
563	* Basically, this is a convenience function for fsync().
564	* @mapping is a file or directory which needs those buffers to be written for
565	* a successful fsync().
566	*/
567	int sync_mapping_buffers(struct address_space *mapping)
568	{
569	struct address_space *buffer_mapping = mapping->private_data;
570
571	if (buffer_mapping == NULL \|\| list_empty(&mapping->private_list))
572	return 0;
573
574	return fsync_buffers_list(&buffer_mapping->private_lock,
575	&mapping->private_list);
576	}
577	EXPORT_SYMBOL(sync_mapping_buffers);
578
579	/*
580	* Called when we've recently written block `bblock', and it is known that
581	* `bblock' was for a buffer_boundary() buffer. This means that the block at
582	* `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
583	* dirty, schedule it for IO. So that indirects merge nicely with their data.
584	*/
585	void write_boundary_block(struct block_device *bdev,
586	sector_t bblock, unsigned blocksize)
587	{
588	struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
589	if (bh) {
590	if (buffer_dirty(bh))
591	ll_rw_block(WRITE, 1, &bh);
592	put_bh(bh);
593	}
594	}
595
596	void mark_buffer_dirty_inode(struct buffer_head bh, struct inode inode)
597	{
598	struct address_space *mapping = inode->i_mapping;
599	struct address_space *buffer_mapping = bh->b_page->mapping;
600
601	mark_buffer_dirty(bh);
602	if (!mapping->private_data) {
603	mapping->private_data = buffer_mapping;
604	} else {
605	BUG_ON(mapping->private_data != buffer_mapping);
606	}
607	if (!bh->b_assoc_map) {
608	spin_lock(&buffer_mapping->private_lock);
609	list_move_tail(&bh->b_assoc_buffers,
610	&mapping->private_list);
611	bh->b_assoc_map = mapping;
612	spin_unlock(&buffer_mapping->private_lock);
613	}
614	}
615	EXPORT_SYMBOL(mark_buffer_dirty_inode);
616
617	/*
618	* Mark the page dirty, and set it dirty in the radix tree, and mark the inode
619	* dirty.
620	*
621	* If warn is true, then emit a warning if the page is not uptodate and has
622	* not been truncated.
623	*
624	* The caller must hold lock_page_memcg().
625	*/
626	static void __set_page_dirty(struct page page, struct address_space mapping,
627	int warn)
628	{
629	unsigned long flags;
630
631	spin_lock_irqsave(&mapping->tree_lock, flags);
632	if (page->mapping) { /* Race with truncate? */
633	WARN_ON_ONCE(warn && !PageUptodate(page));
634	account_page_dirtied(page, mapping);
635	radix_tree_tag_set(&mapping->page_tree,
636	page_index(page), PAGECACHE_TAG_DIRTY);
637	}
638	spin_unlock_irqrestore(&mapping->tree_lock, flags);
639	}
640
641	/*
642	* Add a page to the dirty page list.
643	*
644	* It is a sad fact of life that this function is called from several places
645	* deeply under spinlocking. It may not sleep.
646	*
647	* If the page has buffers, the uptodate buffers are set dirty, to preserve
648	* dirty-state coherency between the page and the buffers. It the page does
649	* not have buffers then when they are later attached they will all be set
650	* dirty.
651	*
652	* The buffers are dirtied before the page is dirtied. There's a small race
653	* window in which a writepage caller may see the page cleanness but not the
654	* buffer dirtiness. That's fine. If this code were to set the page dirty
655	* before the buffers, a concurrent writepage caller could clear the page dirty
656	* bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
657	* page on the dirty page list.
658	*
659	* We use private_lock to lock against try_to_free_buffers while using the
660	* page's buffer list. Also use this to protect against clean buffers being
661	* added to the page after it was set dirty.
662	*
663	* FIXME: may need to call ->reservepage here as well. That's rather up to the
664	* address_space though.
665	*/
666	int __set_page_dirty_buffers(struct page *page)
667	{
668	int newly_dirty;
669	struct address_space *mapping = page_mapping(page);
670
671	if (unlikely(!mapping))
672	return !TestSetPageDirty(page);
673
674	spin_lock(&mapping->private_lock);
675	if (page_has_buffers(page)) {
676	struct buffer_head *head = page_buffers(page);
677	struct buffer_head *bh = head;
678
679	do {
680	set_buffer_dirty(bh);
681	bh = bh->b_this_page;
682	} while (bh != head);
683	}
684	/*
685	* Lock out page->mem_cgroup migration to keep PageDirty
686	* synchronized with per-memcg dirty page counters.
687	*/
688	lock_page_memcg(page);
689	newly_dirty = !TestSetPageDirty(page);
690	spin_unlock(&mapping->private_lock);
691
692	if (newly_dirty)
693	__set_page_dirty(page, mapping, 1);
694
695	unlock_page_memcg(page);
696
697	if (newly_dirty)
698	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
699
700	return newly_dirty;
701	}
702	EXPORT_SYMBOL(__set_page_dirty_buffers);
703
704	/*
705	* Write out and wait upon a list of buffers.
706	*
707	* We have conflicting pressures: we want to make sure that all
708	* initially dirty buffers get waited on, but that any subsequently
709	* dirtied buffers don't. After all, we don't want fsync to last
710	* forever if somebody is actively writing to the file.
711	*
712	* Do this in two main stages: first we copy dirty buffers to a
713	* temporary inode list, queueing the writes as we go. Then we clean
714	* up, waiting for those writes to complete.
715	*
716	* During this second stage, any subsequent updates to the file may end
717	* up refiling the buffer on the original inode's dirty list again, so
718	* there is a chance we will end up with a buffer queued for write but
719	* not yet completed on that list. So, as a final cleanup we go through
720	* the osync code to catch these locked, dirty buffers without requeuing
721	* any newly dirty buffers for write.
722	*/
723	static int fsync_buffers_list(spinlock_t lock, struct list_head list)
724	{
725	struct buffer_head *bh;
726	struct list_head tmp;
727	struct address_space *mapping;
728	int err = 0, err2;
729	struct blk_plug plug;
730
731	INIT_LIST_HEAD(&tmp);
732	blk_start_plug(&plug);
733
734	spin_lock(lock);
735	while (!list_empty(list)) {
736	bh = BH_ENTRY(list->next);
737	mapping = bh->b_assoc_map;
738	__remove_assoc_queue(bh);
739	/* Avoid race with mark_buffer_dirty_inode() which does
740	* a lockless check and we rely on seeing the dirty bit */
741	smp_mb();
742	if (buffer_dirty(bh) \|\| buffer_locked(bh)) {
743	list_add(&bh->b_assoc_buffers, &tmp);
744	bh->b_assoc_map = mapping;
745	if (buffer_dirty(bh)) {
746	get_bh(bh);
747	spin_unlock(lock);
748	/*
749	* Ensure any pending I/O completes so that
750	* write_dirty_buffer() actually writes the
751	* current contents - it is a noop if I/O is
752	* still in flight on potentially older
753	* contents.
754	*/
755	write_dirty_buffer(bh, WRITE_SYNC);
756
757	/*
758	* Kick off IO for the previous mapping. Note
759	* that we will not run the very last mapping,
760	* wait_on_buffer() will do that for us
761	* through sync_buffer().
762	*/
763	brelse(bh);
764	spin_lock(lock);
765	}
766	}
767	}
768
769	spin_unlock(lock);
770	blk_finish_plug(&plug);
771	spin_lock(lock);
772
773	while (!list_empty(&tmp)) {
774	bh = BH_ENTRY(tmp.prev);
775	get_bh(bh);
776	mapping = bh->b_assoc_map;
777	__remove_assoc_queue(bh);
778	/* Avoid race with mark_buffer_dirty_inode() which does
779	* a lockless check and we rely on seeing the dirty bit */
780	smp_mb();
781	if (buffer_dirty(bh)) {
782	list_add(&bh->b_assoc_buffers,
783	&mapping->private_list);
784	bh->b_assoc_map = mapping;
785	}
786	spin_unlock(lock);
787	wait_on_buffer(bh);
788	if (!buffer_uptodate(bh))
789	err = -EIO;
790	brelse(bh);
791	spin_lock(lock);
792	}
793
794	spin_unlock(lock);
795	err2 = osync_buffers_list(lock, list);
796	if (err)
797	return err;
798	else
799	return err2;
800	}
801
802	/*
803	* Invalidate any and all dirty buffers on a given inode. We are
804	* probably unmounting the fs, but that doesn't mean we have already
805	* done a sync(). Just drop the buffers from the inode list.
806	*
807	* NOTE: we take the inode's blockdev's mapping's private_lock. Which
808	* assumes that all the buffers are against the blockdev. Not true
809	* for reiserfs.
810	*/
811	void invalidate_inode_buffers(struct inode *inode)
812	{
813	if (inode_has_buffers(inode)) {
814	struct address_space *mapping = &inode->i_data;
815	struct list_head *list = &mapping->private_list;
816	struct address_space *buffer_mapping = mapping->private_data;
817
818	spin_lock(&buffer_mapping->private_lock);
819	while (!list_empty(list))
820	__remove_assoc_queue(BH_ENTRY(list->next));
821	spin_unlock(&buffer_mapping->private_lock);
822	}
823	}
824	EXPORT_SYMBOL(invalidate_inode_buffers);
825
826	/*
827	* Remove any clean buffers from the inode's buffer list. This is called
828	* when we're trying to free the inode itself. Those buffers can pin it.
829	*
830	* Returns true if all buffers were removed.
831	*/
832	int remove_inode_buffers(struct inode *inode)
833	{
834	int ret = 1;
835
836	if (inode_has_buffers(inode)) {
837	struct address_space *mapping = &inode->i_data;
838	struct list_head *list = &mapping->private_list;
839	struct address_space *buffer_mapping = mapping->private_data;
840
841	spin_lock(&buffer_mapping->private_lock);
842	while (!list_empty(list)) {
843	struct buffer_head *bh = BH_ENTRY(list->next);
844	if (buffer_dirty(bh)) {
845	ret = 0;
846	break;
847	}
848	__remove_assoc_queue(bh);
849	}
850	spin_unlock(&buffer_mapping->private_lock);
851	}
852	return ret;
853	}
854
855	/*
856	* Create the appropriate buffers when given a page for data area and
857	* the size of each buffer.. Use the bh->b_this_page linked list to
858	* follow the buffers created. Return NULL if unable to create more
859	* buffers.
860	*
861	* The retry flag is used to differentiate async IO (paging, swapping)
862	* which may not fail from ordinary buffer allocations.
863	*/
864	struct buffer_head alloc_page_buffers(struct page page, unsigned long size,
865	int retry)
866	{
867	struct buffer_head bh, head;
868	long offset;
869
870	try_again:
871	head = NULL;
872	offset = PAGE_SIZE;
873	while ((offset -= size) >= 0) {
874	bh = alloc_buffer_head(GFP_NOFS);
875	if (!bh)
876	goto no_grow;
877
878	bh->b_this_page = head;
879	bh->b_blocknr = -1;
880	head = bh;
881
882	bh->b_size = size;
883
884	/* Link the buffer to its page */
885	set_bh_page(bh, page, offset);
886	}
887	return head;
888	/*
889	* In case anything failed, we just free everything we got.
890	*/
891	no_grow:
892	if (head) {
893	do {
894	bh = head;
895	head = head->b_this_page;
896	free_buffer_head(bh);
897	} while (head);
898	}
899
900	/*
901	* Return failure for non-async IO requests. Async IO requests
902	* are not allowed to fail, so we have to wait until buffer heads
903	* become available. But we don't want tasks sleeping with
904	* partially complete buffers, so all were released above.
905	*/
906	if (!retry)
907	return NULL;
908
909	/* We're _really_ low on memory. Now we just
910	* wait for old buffer heads to become free due to
911	* finishing IO. Since this is an async request and
912	* the reserve list is empty, we're sure there are
913	* async buffer heads in use.
914	*/
915	free_more_memory();
916	goto try_again;
917	}
918	EXPORT_SYMBOL_GPL(alloc_page_buffers);
919
920	static inline void
921	link_dev_buffers(struct page page, struct buffer_head head)
922	{
923	struct buffer_head bh, tail;
924
925	bh = head;
926	do {
927	tail = bh;
928	bh = bh->b_this_page;
929	} while (bh);
930	tail->b_this_page = head;
931	attach_page_buffers(page, head);
932	}
933
934	static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
935	{
936	sector_t retval = ~((sector_t)0);
937	loff_t sz = i_size_read(bdev->bd_inode);
938
939	if (sz) {
940	unsigned int sizebits = blksize_bits(size);
941	retval = (sz >> sizebits);
942	}
943	return retval;
944	}
945
946	/*
947	* Initialise the state of a blockdev page's buffers.
948	*/
949	static sector_t
950	init_page_buffers(struct page page, struct block_device bdev,
951	sector_t block, int size)
952	{
953	struct buffer_head *head = page_buffers(page);
954	struct buffer_head *bh = head;
955	int uptodate = PageUptodate(page);
956	sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
957
958	do {
959	if (!buffer_mapped(bh)) {
960	init_buffer(bh, NULL, NULL);
961	bh->b_bdev = bdev;
962	bh->b_blocknr = block;
963	if (uptodate)
964	set_buffer_uptodate(bh);
965	if (block < end_block)
966	set_buffer_mapped(bh);
967	}
968	block++;
969	bh = bh->b_this_page;
970	} while (bh != head);
971
972	/*
973	* Caller needs to validate requested block against end of device.
974	*/
975	return end_block;
976	}
977
978	/*
979	* Create the page-cache page that contains the requested block.
980	*
981	* This is used purely for blockdev mappings.
982	*/
983	static int
984	grow_dev_page(struct block_device *bdev, sector_t block,
985	pgoff_t index, int size, int sizebits, gfp_t gfp)
986	{
987	struct inode *inode = bdev->bd_inode;
988	struct page *page;
989	struct buffer_head *bh;
990	sector_t end_block;
991	int ret = 0; /* Will call free_more_memory() */
992	gfp_t gfp_mask;
993
994	gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) \| gfp;
995
996	/*
997	* XXX: __getblk_slow() can not really deal with failure and
998	* will endlessly loop on improvised global reclaim. Prefer
999	* looping in the allocator rather than here, at least that
1000	* code knows what it's doing.
1001	*/
1002	gfp_mask \|= __GFP_NOFAIL;
1003
1004	page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1005	if (!page)
1006	return ret;
1007
1008	BUG_ON(!PageLocked(page));
1009
1010	if (page_has_buffers(page)) {
1011	bh = page_buffers(page);
1012	if (bh->b_size == size) {
1013	end_block = init_page_buffers(page, bdev,
1014	(sector_t)index << sizebits,
1015	size);
1016	goto done;
1017	}
1018	if (!try_to_free_buffers(page))
1019	goto failed;
1020	}
1021
1022	/*
1023	* Allocate some buffers for this page
1024	*/
1025	bh = alloc_page_buffers(page, size, 0);
1026	if (!bh)
1027	goto failed;
1028
1029	/*
1030	* Link the page to the buffers and initialise them. Take the
1031	* lock to be atomic wrt __find_get_block(), which does not
1032	* run under the page lock.
1033	*/
1034	spin_lock(&inode->i_mapping->private_lock);
1035	link_dev_buffers(page, bh);
1036	end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1037	size);
1038	spin_unlock(&inode->i_mapping->private_lock);
1039	done:
1040	ret = (block < end_block) ? 1 : -ENXIO;
1041	failed:
1042	unlock_page(page);
1043	put_page(page);
1044	return ret;
1045	}
1046
1047	/*
1048	* Create buffers for the specified block device block's page. If
1049	* that page was dirty, the buffers are set dirty also.
1050	*/
1051	static int
1052	grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1053	{
1054	pgoff_t index;
1055	int sizebits;
1056
1057	sizebits = -1;
1058	do {
1059	sizebits++;
1060	} while ((size << sizebits) < PAGE_SIZE);
1061
1062	index = block >> sizebits;
1063
1064	/*
1065	* Check for a block which wants to lie outside our maximum possible
1066	* pagecache index. (this comparison is done using sector_t types).
1067	*/
1068	if (unlikely(index != block >> sizebits)) {
1069	printk(KERN_ERR "%s: requested out-of-range block %llu for "
1070	"device %pg\n",
1071	__func__, (unsigned long long)block,
1072	bdev);
1073	return -EIO;
1074	}
1075
1076	/* Create a page with the proper size buffers.. */
1077	return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1078	}
1079
1080	struct buffer_head *
1081	__getblk_slow(struct block_device *bdev, sector_t block,
1082	unsigned size, gfp_t gfp)
1083	{
1084	/* Size must be multiple of hard sectorsize */
1085	if (unlikely(size & (bdev_logical_block_size(bdev)-1) \|\|
1086	(size < 512 \|\| size > PAGE_SIZE))) {
1087	printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1088	size);
1089	printk(KERN_ERR "logical block size: %d\n",
1090	bdev_logical_block_size(bdev));
1091
1092	dump_stack();
1093	return NULL;
1094	}
1095
1096	for (;;) {
1097	struct buffer_head *bh;
1098	int ret;
1099
1100	bh = __find_get_block(bdev, block, size);
1101	if (bh)
1102	return bh;
1103
1104	ret = grow_buffers(bdev, block, size, gfp);
1105	if (ret < 0)
1106	return NULL;
1107	if (ret == 0)
1108	free_more_memory();
1109	}
1110	}
1111	EXPORT_SYMBOL(__getblk_slow);
1112
1113	/*
1114	* The relationship between dirty buffers and dirty pages:
1115	*
1116	* Whenever a page has any dirty buffers, the page's dirty bit is set, and
1117	* the page is tagged dirty in its radix tree.
1118	*
1119	* At all times, the dirtiness of the buffers represents the dirtiness of
1120	* subsections of the page. If the page has buffers, the page dirty bit is
1121	* merely a hint about the true dirty state.
1122	*
1123	* When a page is set dirty in its entirety, all its buffers are marked dirty
1124	* (if the page has buffers).
1125	*
1126	* When a buffer is marked dirty, its page is dirtied, but the page's other
1127	* buffers are not.
1128	*
1129	* Also. When blockdev buffers are explicitly read with bread(), they
1130	* individually become uptodate. But their backing page remains not
1131	* uptodate - even if all of its buffers are uptodate. A subsequent
1132	* block_read_full_page() against that page will discover all the uptodate
1133	* buffers, will set the page uptodate and will perform no I/O.
1134	*/
1135
1136	/**
1137	* mark_buffer_dirty - mark a buffer_head as needing writeout
1138	* @bh: the buffer_head to mark dirty
1139	*
1140	* mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1141	* backing page dirty, then tag the page as dirty in its address_space's radix
1142	* tree and then attach the address_space's inode to its superblock's dirty
1143	* inode list.
1144	*
1145	* mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1146	* mapping->tree_lock and mapping->host->i_lock.
1147	*/
1148	void mark_buffer_dirty(struct buffer_head *bh)
1149	{
1150	WARN_ON_ONCE(!buffer_uptodate(bh));
1151
1152	trace_block_dirty_buffer(bh);
1153
1154	/*
1155	* Very carefully optimize the it-is-already-dirty case.
1156	*
1157	* Don't let the final "is it dirty" escape to before we
1158	* perhaps modified the buffer.
1159	*/
1160	if (buffer_dirty(bh)) {
1161	smp_mb();
1162	if (buffer_dirty(bh))
1163	return;
1164	}
1165
1166	if (!test_set_buffer_dirty(bh)) {
1167	struct page *page = bh->b_page;
1168	struct address_space *mapping = NULL;
1169
1170	lock_page_memcg(page);
1171	if (!TestSetPageDirty(page)) {
1172	mapping = page_mapping(page);
1173	if (mapping)
1174	__set_page_dirty(page, mapping, 0);
1175	}
1176	unlock_page_memcg(page);
1177	if (mapping)
1178	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1179	}
1180	}
1181	EXPORT_SYMBOL(mark_buffer_dirty);
1182
1183	/*
1184	* Decrement a buffer_head's reference count. If all buffers against a page
1185	* have zero reference count, are clean and unlocked, and if the page is clean
1186	* and unlocked then try_to_free_buffers() may strip the buffers from the page
1187	* in preparation for freeing it (sometimes, rarely, buffers are removed from
1188	* a page but it ends up not being freed, and buffers may later be reattached).
1189	*/
1190	void __brelse(struct buffer_head * buf)
1191	{
1192	if (atomic_read(&buf->b_count)) {
1193	put_bh(buf);
1194	return;
1195	}
1196	WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1197	}
1198	EXPORT_SYMBOL(__brelse);
1199
1200	/*
1201	* bforget() is like brelse(), except it discards any
1202	* potentially dirty data.
1203	*/
1204	void __bforget(struct buffer_head *bh)
1205	{
1206	clear_buffer_dirty(bh);
1207	if (bh->b_assoc_map) {
1208	struct address_space *buffer_mapping = bh->b_page->mapping;
1209
1210	spin_lock(&buffer_mapping->private_lock);
1211	list_del_init(&bh->b_assoc_buffers);
1212	bh->b_assoc_map = NULL;
1213	spin_unlock(&buffer_mapping->private_lock);
1214	}
1215	__brelse(bh);
1216	}
1217	EXPORT_SYMBOL(__bforget);
1218
1219	static struct buffer_head __bread_slow(struct buffer_head bh)
1220	{
1221	lock_buffer(bh);
1222	if (buffer_uptodate(bh)) {
1223	unlock_buffer(bh);
1224	return bh;
1225	} else {
1226	get_bh(bh);
1227	bh->b_end_io = end_buffer_read_sync;
1228	submit_bh(READ, bh);
1229	wait_on_buffer(bh);
1230	if (buffer_uptodate(bh))
1231	return bh;
1232	}
1233	brelse(bh);
1234	return NULL;
1235	}
1236
1237	/*
1238	* Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1239	* The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1240	* refcount elevated by one when they're in an LRU. A buffer can only appear
1241	* once in a particular CPU's LRU. A single buffer can be present in multiple
1242	* CPU's LRUs at the same time.
1243	*
1244	* This is a transparent caching front-end to sb_bread(), sb_getblk() and
1245	* sb_find_get_block().
1246	*
1247	* The LRUs themselves only need locking against invalidate_bh_lrus. We use
1248	* a local interrupt disable for that.
1249	*/
1250
1251	#define BH_LRU_SIZE 16
1252
1253	struct bh_lru {
1254	struct buffer_head *bhs[BH_LRU_SIZE];
1255	};
1256
1257	static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1258
1259	#ifdef CONFIG_SMP
1260	#define bh_lru_lock() local_irq_disable()
1261	#define bh_lru_unlock() local_irq_enable()
1262	#else
1263	#define bh_lru_lock() preempt_disable()
1264	#define bh_lru_unlock() preempt_enable()
1265	#endif
1266
1267	static inline void check_irqs_on(void)
1268	{
1269	#ifdef irqs_disabled
1270	BUG_ON(irqs_disabled());
1271	#endif
1272	}
1273
1274	/*
1275	* The LRU management algorithm is dopey-but-simple. Sorry.
1276	*/
1277	static void bh_lru_install(struct buffer_head *bh)
1278	{
1279	struct buffer_head *evictee = NULL;
1280
1281	check_irqs_on();
1282	bh_lru_lock();
1283	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1284	struct buffer_head *bhs[BH_LRU_SIZE];
1285	int in;
1286	int out = 0;
1287
1288	get_bh(bh);
1289	bhs[out++] = bh;
1290	for (in = 0; in < BH_LRU_SIZE; in++) {
1291	struct buffer_head *bh2 =
1292	__this_cpu_read(bh_lrus.bhs[in]);
1293
1294	if (bh2 == bh) {
1295	__brelse(bh2);
1296	} else {
1297	if (out >= BH_LRU_SIZE) {
1298	BUG_ON(evictee != NULL);
1299	evictee = bh2;
1300	} else {
1301	bhs[out++] = bh2;
1302	}
1303	}
1304	}
1305	while (out < BH_LRU_SIZE)
1306	bhs[out++] = NULL;
1307	memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1308	}
1309	bh_lru_unlock();
1310
1311	if (evictee)
1312	__brelse(evictee);
1313	}
1314
1315	/*
1316	* Look up the bh in this cpu's LRU. If it's there, move it to the head.
1317	*/
1318	static struct buffer_head *
1319	lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1320	{
1321	struct buffer_head *ret = NULL;
1322	unsigned int i;
1323
1324	check_irqs_on();
1325	bh_lru_lock();
1326	for (i = 0; i < BH_LRU_SIZE; i++) {
1327	struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1328
1329	if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1330	bh->b_size == size) {
1331	if (i) {
1332	while (i) {
1333	__this_cpu_write(bh_lrus.bhs[i],
1334	__this_cpu_read(bh_lrus.bhs[i - 1]));
1335	i--;
1336	}
1337	__this_cpu_write(bh_lrus.bhs[0], bh);
1338	}
1339	get_bh(bh);
1340	ret = bh;
1341	break;
1342	}
1343	}
1344	bh_lru_unlock();
1345	return ret;
1346	}
1347
1348	/*
1349	* Perform a pagecache lookup for the matching buffer. If it's there, refresh
1350	* it in the LRU and mark it as accessed. If it is not present then return
1351	* NULL
1352	*/
1353	struct buffer_head *
1354	__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1355	{
1356	struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1357
1358	if (bh == NULL) {
1359	/* __find_get_block_slow will mark the page accessed */
1360	bh = __find_get_block_slow(bdev, block);
1361	if (bh)
1362	bh_lru_install(bh);
1363	} else
1364	touch_buffer(bh);
1365
1366	return bh;
1367	}
1368	EXPORT_SYMBOL(__find_get_block);
1369
1370	/*
1371	* __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1372	* which corresponds to the passed block_device, block and size. The
1373	* returned buffer has its reference count incremented.
1374	*
1375	* __getblk_gfp() will lock up the machine if grow_dev_page's
1376	* try_to_free_buffers() attempt is failing. FIXME, perhaps?
1377	*/
1378	struct buffer_head *
1379	__getblk_gfp(struct block_device *bdev, sector_t block,
1380	unsigned size, gfp_t gfp)
1381	{
1382	struct buffer_head *bh = __find_get_block(bdev, block, size);
1383
1384	might_sleep();
1385	if (bh == NULL)
1386	bh = __getblk_slow(bdev, block, size, gfp);
1387	return bh;
1388	}
1389	EXPORT_SYMBOL(__getblk_gfp);
1390
1391	/*
1392	* Do async read-ahead on a buffer..
1393	*/
1394	void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1395	{
1396	struct buffer_head *bh = __getblk(bdev, block, size);
1397	if (likely(bh)) {
1398	ll_rw_block(READA, 1, &bh);
1399	brelse(bh);
1400	}
1401	}
1402	EXPORT_SYMBOL(__breadahead);
1403
1404	/**
1405	* __bread_gfp() - reads a specified block and returns the bh
1406	* @bdev: the block_device to read from
1407	* @block: number of block
1408	* @size: size (in bytes) to read
1409	* @gfp: page allocation flag
1410	*
1411	* Reads a specified block, and returns buffer head that contains it.
1412	* The page cache can be allocated from non-movable area
1413	* not to prevent page migration if you set gfp to zero.
1414	* It returns NULL if the block was unreadable.
1415	*/
1416	struct buffer_head *
1417	__bread_gfp(struct block_device *bdev, sector_t block,
1418	unsigned size, gfp_t gfp)
1419	{
1420	struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1421
1422	if (likely(bh) && !buffer_uptodate(bh))
1423	bh = __bread_slow(bh);
1424	return bh;
1425	}
1426	EXPORT_SYMBOL(__bread_gfp);
1427
1428	/*
1429	* invalidate_bh_lrus() is called rarely - but not only at unmount.
1430	* This doesn't race because it runs in each cpu either in irq
1431	* or with preempt disabled.
1432	*/
1433	static void invalidate_bh_lru(void *arg)
1434	{
1435	struct bh_lru *b = &get_cpu_var(bh_lrus);
1436	int i;
1437
1438	for (i = 0; i < BH_LRU_SIZE; i++) {
1439	brelse(b->bhs[i]);
1440	b->bhs[i] = NULL;
1441	}
1442	put_cpu_var(bh_lrus);
1443	}
1444
1445	static bool has_bh_in_lru(int cpu, void *dummy)
1446	{
1447	struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1448	int i;
1449
1450	for (i = 0; i < BH_LRU_SIZE; i++) {
1451	if (b->bhs[i])
1452	return 1;
1453	}
1454
1455	return 0;
1456	}
1457
1458	void invalidate_bh_lrus(void)
1459	{
1460	on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1461	}
1462	EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1463
1464	void set_bh_page(struct buffer_head *bh,
1465	struct page *page, unsigned long offset)
1466	{
1467	bh->b_page = page;
1468	BUG_ON(offset >= PAGE_SIZE);
1469	if (PageHighMem(page))
1470	/*
1471	* This catches illegal uses and preserves the offset:
1472	*/
1473	bh->b_data = (char *)(0 + offset);
1474	else
1475	bh->b_data = page_address(page) + offset;
1476	}
1477	EXPORT_SYMBOL(set_bh_page);
1478
1479	/*
1480	* Called when truncating a buffer on a page completely.
1481	*/
1482
1483	/* Bits that are cleared during an invalidate */
1484	#define BUFFER_FLAGS_DISCARD \
1485	(1 << BH_Mapped \| 1 << BH_New \| 1 << BH_Req \| \
1486	1 << BH_Delay \| 1 << BH_Unwritten)
1487
1488	static void discard_buffer(struct buffer_head * bh)
1489	{
1490	unsigned long b_state, b_state_old;
1491
1492	lock_buffer(bh);
1493	clear_buffer_dirty(bh);
1494	bh->b_bdev = NULL;
1495	b_state = bh->b_state;
1496	for (;;) {
1497	b_state_old = cmpxchg(&bh->b_state, b_state,
1498	(b_state & ~BUFFER_FLAGS_DISCARD));
1499	if (b_state_old == b_state)
1500	break;
1501	b_state = b_state_old;
1502	}
1503	unlock_buffer(bh);
1504	}
1505
1506	/**
1507	* block_invalidatepage - invalidate part or all of a buffer-backed page
1508	*
1509	* @page: the page which is affected
1510	* @offset: start of the range to invalidate
1511	* @length: length of the range to invalidate
1512	*
1513	* block_invalidatepage() is called when all or part of the page has become
1514	* invalidated by a truncate operation.
1515	*
1516	* block_invalidatepage() does not have to release all buffers, but it must
1517	* ensure that no dirty buffer is left outside @offset and that no I/O
1518	* is underway against any of the blocks which are outside the truncation
1519	* point. Because the caller is about to free (and possibly reuse) those
1520	* blocks on-disk.
1521	*/
1522	void block_invalidatepage(struct page *page, unsigned int offset,
1523	unsigned int length)
1524	{
1525	struct buffer_head head, bh, *next;
1526	unsigned int curr_off = 0;
1527	unsigned int stop = length + offset;
1528
1529	BUG_ON(!PageLocked(page));
1530	if (!page_has_buffers(page))
1531	goto out;
1532
1533	/*
1534	* Check for overflow
1535	*/
1536	BUG_ON(stop > PAGE_SIZE \|\| stop < length);
1537
1538	head = page_buffers(page);
1539	bh = head;
1540	do {
1541	unsigned int next_off = curr_off + bh->b_size;
1542	next = bh->b_this_page;
1543
1544	/*
1545	* Are we still fully in range ?
1546	*/
1547	if (next_off > stop)
1548	goto out;
1549
1550	/*
1551	* is this block fully invalidated?
1552	*/
1553	if (offset <= curr_off)
1554	discard_buffer(bh);
1555	curr_off = next_off;
1556	bh = next;
1557	} while (bh != head);
1558
1559	/*
1560	* We release buffers only if the entire page is being invalidated.
1561	* The get_block cached value has been unconditionally invalidated,
1562	* so real IO is not possible anymore.
1563	*/
1564	if (offset == 0)
1565	try_to_release_page(page, 0);
1566	out:
1567	return;
1568	}
1569	EXPORT_SYMBOL(block_invalidatepage);
1570
1571
1572	/*
1573	* We attach and possibly dirty the buffers atomically wrt
1574	* __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1575	* is already excluded via the page lock.
1576	*/
1577	void create_empty_buffers(struct page *page,
1578	unsigned long blocksize, unsigned long b_state)
1579	{
1580	struct buffer_head bh, head, *tail;
1581
1582	head = alloc_page_buffers(page, blocksize, 1);
1583	bh = head;
1584	do {
1585	bh->b_state \|= b_state;
1586	tail = bh;
1587	bh = bh->b_this_page;
1588	} while (bh);
1589	tail->b_this_page = head;
1590
1591	spin_lock(&page->mapping->private_lock);
1592	if (PageUptodate(page) \|\| PageDirty(page)) {
1593	bh = head;
1594	do {
1595	if (PageDirty(page))
1596	set_buffer_dirty(bh);
1597	if (PageUptodate(page))
1598	set_buffer_uptodate(bh);
1599	bh = bh->b_this_page;
1600	} while (bh != head);
1601	}
1602	attach_page_buffers(page, head);
1603	spin_unlock(&page->mapping->private_lock);
1604	}
1605	EXPORT_SYMBOL(create_empty_buffers);
1606
1607	/*
1608	* We are taking a block for data and we don't want any output from any
1609	* buffer-cache aliases starting from return from that function and
1610	* until the moment when something will explicitly mark the buffer
1611	* dirty (hopefully that will not happen until we will free that block ;-)
1612	* We don't even need to mark it not-uptodate - nobody can expect
1613	* anything from a newly allocated buffer anyway. We used to used
1614	* unmap_buffer() for such invalidation, but that was wrong. We definitely
1615	* don't want to mark the alias unmapped, for example - it would confuse
1616	* anyone who might pick it with bread() afterwards...
1617	*
1618	* Also.. Note that bforget() doesn't lock the buffer. So there can
1619	* be writeout I/O going on against recently-freed buffers. We don't
1620	* wait on that I/O in bforget() - it's more efficient to wait on the I/O
1621	* only if we really need to. That happens here.
1622	*/
1623	void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1624	{
1625	struct buffer_head *old_bh;
1626
1627	might_sleep();
1628
1629	old_bh = __find_get_block_slow(bdev, block);
1630	if (old_bh) {
1631	clear_buffer_dirty(old_bh);
1632	wait_on_buffer(old_bh);
1633	clear_buffer_req(old_bh);
1634	__brelse(old_bh);
1635	}
1636	}
1637	EXPORT_SYMBOL(unmap_underlying_metadata);
1638
1639	/*
1640	* Size is a power-of-two in the range 512..PAGE_SIZE,
1641	* and the case we care about most is PAGE_SIZE.
1642	*
1643	* So this could possibly be written with those
1644	* constraints in mind (relevant mostly if some
1645	* architecture has a slow bit-scan instruction)
1646	*/
1647	static inline int block_size_bits(unsigned int blocksize)
1648	{
1649	return ilog2(blocksize);
1650	}
1651
1652	static struct buffer_head create_page_buffers(struct page page, struct inode *inode, unsigned int b_state)
1653	{
1654	BUG_ON(!PageLocked(page));
1655
1656	if (!page_has_buffers(page))
1657	create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1658	return page_buffers(page);
1659	}
1660
1661	/*
1662	* NOTE! All mapped/uptodate combinations are valid:
1663	*
1664	* Mapped Uptodate Meaning
1665	*
1666	* No No "unknown" - must do get_block()
1667	* No Yes "hole" - zero-filled
1668	* Yes No "allocated" - allocated on disk, not read in
1669	* Yes Yes "valid" - allocated and up-to-date in memory.
1670	*
1671	* "Dirty" is valid only with the last case (mapped+uptodate).
1672	*/
1673
1674	/*
1675	* While block_write_full_page is writing back the dirty buffers under
1676	* the page lock, whoever dirtied the buffers may decide to clean them
1677	* again at any time. We handle that by only looking at the buffer
1678	* state inside lock_buffer().
1679	*
1680	* If block_write_full_page() is called for regular writeback
1681	* (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1682	* locked buffer. This only can happen if someone has written the buffer
1683	* directly, with submit_bh(). At the address_space level PageWriteback
1684	* prevents this contention from occurring.
1685	*
1686	* If block_write_full_page() is called with wbc->sync_mode ==
1687	* WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1688	* causes the writes to be flagged as synchronous writes.
1689	*/
1690	static int __block_write_full_page(struct inode inode, struct page page,
1691	get_block_t get_block, struct writeback_control wbc,
1692	bh_end_io_t *handler)
1693	{
1694	int err;
1695	sector_t block;
1696	sector_t last_block;
1697	struct buffer_head bh, head;
1698	unsigned int blocksize, bbits;
1699	int nr_underway = 0;
1700	int write_op = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE);
1701
1702	head = create_page_buffers(page, inode,
1703	(1 << BH_Dirty)\|(1 << BH_Uptodate));
1704
1705	/*
1706	* Be very careful. We have no exclusion from __set_page_dirty_buffers
1707	* here, and the (potentially unmapped) buffers may become dirty at
1708	* any time. If a buffer becomes dirty here after we've inspected it
1709	* then we just miss that fact, and the page stays dirty.
1710	*
1711	* Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1712	* handle that here by just cleaning them.
1713	*/
1714
1715	bh = head;
1716	blocksize = bh->b_size;
1717	bbits = block_size_bits(blocksize);
1718
1719	block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1720	last_block = (i_size_read(inode) - 1) >> bbits;
1721
1722	/*
1723	* Get all the dirty buffers mapped to disk addresses and
1724	* handle any aliases from the underlying blockdev's mapping.
1725	*/
1726	do {
1727	if (block > last_block) {
1728	/*
1729	* mapped buffers outside i_size will occur, because
1730	* this page can be outside i_size when there is a
1731	* truncate in progress.
1732	*/
1733	/*
1734	* The buffer was zeroed by block_write_full_page()
1735	*/
1736	clear_buffer_dirty(bh);
1737	set_buffer_uptodate(bh);
1738	} else if ((!buffer_mapped(bh) \|\| buffer_delay(bh)) &&
1739	buffer_dirty(bh)) {
1740	WARN_ON(bh->b_size != blocksize);
1741	err = get_block(inode, block, bh, 1);
1742	if (err)
1743	goto recover;
1744	clear_buffer_delay(bh);
1745	if (buffer_new(bh)) {
1746	/* blockdev mappings never come here */
1747	clear_buffer_new(bh);
1748	unmap_underlying_metadata(bh->b_bdev,
1749	bh->b_blocknr);
1750	}
1751	}
1752	bh = bh->b_this_page;
1753	block++;
1754	} while (bh != head);
1755
1756	do {
1757	if (!buffer_mapped(bh))
1758	continue;
1759	/*
1760	* If it's a fully non-blocking write attempt and we cannot
1761	* lock the buffer then redirty the page. Note that this can
1762	* potentially cause a busy-wait loop from writeback threads
1763	* and kswapd activity, but those code paths have their own
1764	* higher-level throttling.
1765	*/
1766	if (wbc->sync_mode != WB_SYNC_NONE) {
1767	lock_buffer(bh);
1768	} else if (!trylock_buffer(bh)) {
1769	redirty_page_for_writepage(wbc, page);
1770	continue;
1771	}
1772	if (test_clear_buffer_dirty(bh)) {
1773	mark_buffer_async_write_endio(bh, handler);
1774	} else {
1775	unlock_buffer(bh);
1776	}
1777	} while ((bh = bh->b_this_page) != head);
1778
1779	/*
1780	* The page and its buffers are protected by PageWriteback(), so we can
1781	* drop the bh refcounts early.
1782	*/
1783	BUG_ON(PageWriteback(page));
1784	set_page_writeback(page);
1785
1786	do {
1787	struct buffer_head *next = bh->b_this_page;
1788	if (buffer_async_write(bh)) {
1789	submit_bh_wbc(write_op, bh, 0, wbc);
1790	nr_underway++;
1791	}
1792	bh = next;
1793	} while (bh != head);
1794	unlock_page(page);
1795
1796	err = 0;
1797	done:
1798	if (nr_underway == 0) {
1799	/*
1800	* The page was marked dirty, but the buffers were
1801	* clean. Someone wrote them back by hand with
1802	* ll_rw_block/submit_bh. A rare case.
1803	*/
1804	end_page_writeback(page);
1805
1806	/*
1807	* The page and buffer_heads can be released at any time from
1808	* here on.
1809	*/
1810	}
1811	return err;
1812
1813	recover:
1814	/*
1815	* ENOSPC, or some other error. We may already have added some
1816	* blocks to the file, so we need to write these out to avoid
1817	* exposing stale data.
1818	* The page is currently locked and not marked for writeback
1819	*/
1820	bh = head;
1821	/* Recovery: lock and submit the mapped buffers */
1822	do {
1823	if (buffer_mapped(bh) && buffer_dirty(bh) &&
1824	!buffer_delay(bh)) {
1825	lock_buffer(bh);
1826	mark_buffer_async_write_endio(bh, handler);
1827	} else {
1828	/*
1829	* The buffer may have been set dirty during
1830	* attachment to a dirty page.
1831	*/
1832	clear_buffer_dirty(bh);
1833	}
1834	} while ((bh = bh->b_this_page) != head);
1835	SetPageError(page);
1836	BUG_ON(PageWriteback(page));
1837	mapping_set_error(page->mapping, err);
1838	set_page_writeback(page);
1839	do {
1840	struct buffer_head *next = bh->b_this_page;
1841	if (buffer_async_write(bh)) {
1842	clear_buffer_dirty(bh);
1843	submit_bh_wbc(write_op, bh, 0, wbc);
1844	nr_underway++;
1845	}
1846	bh = next;
1847	} while (bh != head);
1848	unlock_page(page);
1849	goto done;
1850	}
1851
1852	/*
1853	* If a page has any new buffers, zero them out here, and mark them uptodate
1854	* and dirty so they'll be written out (in order to prevent uninitialised
1855	* block data from leaking). And clear the new bit.
1856	*/
1857	void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1858	{
1859	unsigned int block_start, block_end;
1860	struct buffer_head head, bh;
1861
1862	BUG_ON(!PageLocked(page));
1863	if (!page_has_buffers(page))
1864	return;
1865
1866	bh = head = page_buffers(page);
1867	block_start = 0;
1868	do {
1869	block_end = block_start + bh->b_size;
1870
1871	if (buffer_new(bh)) {
1872	if (block_end > from && block_start < to) {
1873	if (!PageUptodate(page)) {
1874	unsigned start, size;
1875
1876	start = max(from, block_start);
1877	size = min(to, block_end) - start;
1878
1879	zero_user(page, start, size);
1880	set_buffer_uptodate(bh);
1881	}
1882
1883	clear_buffer_new(bh);
1884	mark_buffer_dirty(bh);
1885	}
1886	}
1887
1888	block_start = block_end;
1889	bh = bh->b_this_page;
1890	} while (bh != head);
1891	}
1892	EXPORT_SYMBOL(page_zero_new_buffers);
1893
1894	int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1895	get_block_t *get_block)
1896	{
1897	unsigned from = pos & (PAGE_SIZE - 1);
1898	unsigned to = from + len;
1899	struct inode *inode = page->mapping->host;
1900	unsigned block_start, block_end;
1901	sector_t block;
1902	int err = 0;
1903	unsigned blocksize, bbits;
1904	struct buffer_head bh, head, wait[2], *wait_bh=wait;
1905
1906	BUG_ON(!PageLocked(page));
1907	BUG_ON(from > PAGE_SIZE);
1908	BUG_ON(to > PAGE_SIZE);
1909	BUG_ON(from > to);
1910
1911	head = create_page_buffers(page, inode, 0);
1912	blocksize = head->b_size;
1913	bbits = block_size_bits(blocksize);
1914
1915	block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1916
1917	for(bh = head, block_start = 0; bh != head \|\| !block_start;
1918	block++, block_start=block_end, bh = bh->b_this_page) {
1919	block_end = block_start + blocksize;
1920	if (block_end <= from \|\| block_start >= to) {
1921	if (PageUptodate(page)) {
1922	if (!buffer_uptodate(bh))
1923	set_buffer_uptodate(bh);
1924	}
1925	continue;
1926	}
1927	if (buffer_new(bh))
1928	clear_buffer_new(bh);
1929	if (!buffer_mapped(bh)) {
1930	WARN_ON(bh->b_size != blocksize);
1931	err = get_block(inode, block, bh, 1);
1932	if (err)
1933	break;
1934	if (buffer_new(bh)) {
1935	unmap_underlying_metadata(bh->b_bdev,
1936	bh->b_blocknr);
1937	if (PageUptodate(page)) {
1938	clear_buffer_new(bh);
1939	set_buffer_uptodate(bh);
1940	mark_buffer_dirty(bh);
1941	continue;
1942	}
1943	if (block_end > to \|\| block_start < from)
1944	zero_user_segments(page,
1945	to, block_end,
1946	block_start, from);
1947	continue;
1948	}
1949	}
1950	if (PageUptodate(page)) {
1951	if (!buffer_uptodate(bh))
1952	set_buffer_uptodate(bh);
1953	continue;
1954	}
1955	if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1956	!buffer_unwritten(bh) &&
1957	(block_start < from \|\| block_end > to)) {
1958	ll_rw_block(READ, 1, &bh);
1959	*wait_bh++=bh;
1960	}
1961	}
1962	/*
1963	* If we issued read requests - let them complete.
1964	*/
1965	while(wait_bh > wait) {
1966	wait_on_buffer(*--wait_bh);
1967	if (!buffer_uptodate(*wait_bh))
1968	err = -EIO;
1969	}
1970	if (unlikely(err))
1971	page_zero_new_buffers(page, from, to);
1972	return err;
1973	}
1974	EXPORT_SYMBOL(__block_write_begin);
1975
1976	static int __block_commit_write(struct inode inode, struct page page,
1977	unsigned from, unsigned to)
1978	{
1979	unsigned block_start, block_end;
1980	int partial = 0;
1981	unsigned blocksize;
1982	struct buffer_head bh, head;
1983
1984	bh = head = page_buffers(page);
1985	blocksize = bh->b_size;
1986
1987	block_start = 0;
1988	do {
1989	block_end = block_start + blocksize;
1990	if (block_end <= from \|\| block_start >= to) {
1991	if (!buffer_uptodate(bh))
1992	partial = 1;
1993	} else {
1994	set_buffer_uptodate(bh);
1995	mark_buffer_dirty(bh);
1996	}
1997	clear_buffer_new(bh);
1998
1999	block_start = block_end;
2000	bh = bh->b_this_page;
2001	} while (bh != head);
2002
2003	/*
2004	* If this is a partial write which happened to make all buffers
2005	* uptodate then we can optimize away a bogus readpage() for
2006	* the next read(). Here we 'discover' whether the page went
2007	* uptodate as a result of this (potentially partial) write.
2008	*/
2009	if (!partial)
2010	SetPageUptodate(page);
2011	return 0;
2012	}
2013
2014	/*
2015	* block_write_begin takes care of the basic task of block allocation and
2016	* bringing partial write blocks uptodate first.
2017	*
2018	* The filesystem needs to handle block truncation upon failure.
2019	*/
2020	int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2021	unsigned flags, struct page *pagep, get_block_t get_block)
2022	{
2023	pgoff_t index = pos >> PAGE_SHIFT;
2024	struct page *page;
2025	int status;
2026
2027	page = grab_cache_page_write_begin(mapping, index, flags);
2028	if (!page)
2029	return -ENOMEM;
2030
2031	status = __block_write_begin(page, pos, len, get_block);
2032	if (unlikely(status)) {
2033	unlock_page(page);
2034	put_page(page);
2035	page = NULL;
2036	}
2037
2038	*pagep = page;
2039	return status;
2040	}
2041	EXPORT_SYMBOL(block_write_begin);
2042
2043	int block_write_end(struct file file, struct address_space mapping,
2044	loff_t pos, unsigned len, unsigned copied,
2045	struct page page, void fsdata)
2046	{
2047	struct inode *inode = mapping->host;
2048	unsigned start;
2049
2050	start = pos & (PAGE_SIZE - 1);
2051
2052	if (unlikely(copied < len)) {
2053	/*
2054	* The buffers that were written will now be uptodate, so we
2055	* don't have to worry about a readpage reading them and
2056	* overwriting a partial write. However if we have encountered
2057	* a short write and only partially written into a buffer, it
2058	* will not be marked uptodate, so a readpage might come in and
2059	* destroy our partial write.
2060	*
2061	* Do the simplest thing, and just treat any short write to a
2062	* non uptodate page as a zero-length write, and force the
2063	* caller to redo the whole thing.
2064	*/
2065	if (!PageUptodate(page))
2066	copied = 0;
2067
2068	page_zero_new_buffers(page, start+copied, start+len);
2069	}
2070	flush_dcache_page(page);
2071
2072	/* This could be a short (even 0-length) commit */
2073	__block_commit_write(inode, page, start, start+copied);
2074
2075	return copied;
2076	}
2077	EXPORT_SYMBOL(block_write_end);
2078
2079	int generic_write_end(struct file file, struct address_space mapping,
2080	loff_t pos, unsigned len, unsigned copied,
2081	struct page page, void fsdata)
2082	{
2083	struct inode *inode = mapping->host;
2084	loff_t old_size = inode->i_size;
2085	int i_size_changed = 0;
2086
2087	copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2088
2089	/*
2090	* No need to use i_size_read() here, the i_size
2091	* cannot change under us because we hold i_mutex.
2092	*
2093	* But it's important to update i_size while still holding page lock:
2094	* page writeout could otherwise come in and zero beyond i_size.
2095	*/
2096	if (pos+copied > inode->i_size) {
2097	i_size_write(inode, pos+copied);
2098	i_size_changed = 1;
2099	}
2100
2101	unlock_page(page);
2102	put_page(page);
2103
2104	if (old_size < pos)
2105	pagecache_isize_extended(inode, old_size, pos);
2106	/*
2107	* Don't mark the inode dirty under page lock. First, it unnecessarily
2108	* makes the holding time of page lock longer. Second, it forces lock
2109	* ordering of page lock and transaction start for journaling
2110	* filesystems.
2111	*/
2112	if (i_size_changed)
2113	mark_inode_dirty(inode);
2114
2115	return copied;
2116	}
2117	EXPORT_SYMBOL(generic_write_end);
2118
2119	/*
2120	* block_is_partially_uptodate checks whether buffers within a page are
2121	* uptodate or not.
2122	*
2123	* Returns true if all buffers which correspond to a file portion
2124	* we want to read are uptodate.
2125	*/
2126	int block_is_partially_uptodate(struct page *page, unsigned long from,
2127	unsigned long count)
2128	{
2129	unsigned block_start, block_end, blocksize;
2130	unsigned to;
2131	struct buffer_head bh, head;
2132	int ret = 1;
2133
2134	if (!page_has_buffers(page))
2135	return 0;
2136
2137	head = page_buffers(page);
2138	blocksize = head->b_size;
2139	to = min_t(unsigned, PAGE_SIZE - from, count);
2140	to = from + to;
2141	if (from < blocksize && to > PAGE_SIZE - blocksize)
2142	return 0;
2143
2144	bh = head;
2145	block_start = 0;
2146	do {
2147	block_end = block_start + blocksize;
2148	if (block_end > from && block_start < to) {
2149	if (!buffer_uptodate(bh)) {
2150	ret = 0;
2151	break;
2152	}
2153	if (block_end >= to)
2154	break;
2155	}
2156	block_start = block_end;
2157	bh = bh->b_this_page;
2158	} while (bh != head);
2159
2160	return ret;
2161	}
2162	EXPORT_SYMBOL(block_is_partially_uptodate);
2163
2164	/*
2165	* Generic "read page" function for block devices that have the normal
2166	* get_block functionality. This is most of the block device filesystems.
2167	* Reads the page asynchronously --- the unlock_buffer() and
2168	* set/clear_buffer_uptodate() functions propagate buffer state into the
2169	* page struct once IO has completed.
2170	*/
2171	int block_read_full_page(struct page page, get_block_t get_block)
2172	{
2173	struct inode *inode = page->mapping->host;
2174	sector_t iblock, lblock;
2175	struct buffer_head bh, head, *arr[MAX_BUF_PER_PAGE];
2176	unsigned int blocksize, bbits;
2177	int nr, i;
2178	int fully_mapped = 1;
2179
2180	head = create_page_buffers(page, inode, 0);
2181	blocksize = head->b_size;
2182	bbits = block_size_bits(blocksize);
2183
2184	iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2185	lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2186	bh = head;
2187	nr = 0;
2188	i = 0;
2189
2190	do {
2191	if (buffer_uptodate(bh))
2192	continue;
2193
2194	if (!buffer_mapped(bh)) {
2195	int err = 0;
2196
2197	fully_mapped = 0;
2198	if (iblock < lblock) {
2199	WARN_ON(bh->b_size != blocksize);
2200	err = get_block(inode, iblock, bh, 0);
2201	if (err)
2202	SetPageError(page);
2203	}
2204	if (!buffer_mapped(bh)) {
2205	zero_user(page, i * blocksize, blocksize);
2206	if (!err)
2207	set_buffer_uptodate(bh);
2208	continue;
2209	}
2210	/*
2211	* get_block() might have updated the buffer
2212	* synchronously
2213	*/
2214	if (buffer_uptodate(bh))
2215	continue;
2216	}
2217	arr[nr++] = bh;
2218	} while (i++, iblock++, (bh = bh->b_this_page) != head);
2219
2220	if (fully_mapped)
2221	SetPageMappedToDisk(page);
2222
2223	if (!nr) {
2224	/*
2225	* All buffers are uptodate - we can set the page uptodate
2226	* as well. But not if get_block() returned an error.
2227	*/
2228	if (!PageError(page))
2229	SetPageUptodate(page);
2230	unlock_page(page);
2231	return 0;
2232	}
2233
2234	/* Stage two: lock the buffers */
2235	for (i = 0; i < nr; i++) {
2236	bh = arr[i];
2237	lock_buffer(bh);
2238	mark_buffer_async_read(bh);
2239	}
2240
2241	/*
2242	* Stage 3: start the IO. Check for uptodateness
2243	* inside the buffer lock in case another process reading
2244	* the underlying blockdev brought it uptodate (the sct fix).
2245	*/
2246	for (i = 0; i < nr; i++) {
2247	bh = arr[i];
2248	if (buffer_uptodate(bh))
2249	end_buffer_async_read(bh, 1);
2250	else
2251	submit_bh(READ, bh);
2252	}
2253	return 0;
2254	}
2255	EXPORT_SYMBOL(block_read_full_page);
2256
2257	/* utility function for filesystems that need to do work on expanding
2258	* truncates. Uses filesystem pagecache writes to allow the filesystem to
2259	* deal with the hole.
2260	*/
2261	int generic_cont_expand_simple(struct inode *inode, loff_t size)
2262	{
2263	struct address_space *mapping = inode->i_mapping;
2264	struct page *page;
2265	void *fsdata;
2266	int err;
2267
2268	err = inode_newsize_ok(inode, size);
2269	if (err)
2270	goto out;
2271
2272	err = pagecache_write_begin(NULL, mapping, size, 0,
2273	AOP_FLAG_UNINTERRUPTIBLE\|AOP_FLAG_CONT_EXPAND,
2274	&page, &fsdata);
2275	if (err)
2276	goto out;
2277
2278	err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2279	BUG_ON(err > 0);
2280
2281	out:
2282	return err;
2283	}
2284	EXPORT_SYMBOL(generic_cont_expand_simple);
2285
2286	static int cont_expand_zero(struct file file, struct address_space mapping,
2287	loff_t pos, loff_t *bytes)
2288	{
2289	struct inode *inode = mapping->host;
2290	unsigned blocksize = 1 << inode->i_blkbits;
2291	struct page *page;
2292	void *fsdata;
2293	pgoff_t index, curidx;
2294	loff_t curpos;
2295	unsigned zerofrom, offset, len;
2296	int err = 0;
2297
2298	index = pos >> PAGE_SHIFT;
2299	offset = pos & ~PAGE_MASK;
2300
2301	while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2302	zerofrom = curpos & ~PAGE_MASK;
2303	if (zerofrom & (blocksize-1)) {
2304	*bytes \|= (blocksize-1);
2305	(*bytes)++;
2306	}
2307	len = PAGE_SIZE - zerofrom;
2308
2309	err = pagecache_write_begin(file, mapping, curpos, len,
2310	AOP_FLAG_UNINTERRUPTIBLE,
2311	&page, &fsdata);
2312	if (err)
2313	goto out;
2314	zero_user(page, zerofrom, len);
2315	err = pagecache_write_end(file, mapping, curpos, len, len,
2316	page, fsdata);
2317	if (err < 0)
2318	goto out;
2319	BUG_ON(err != len);
2320	err = 0;
2321
2322	balance_dirty_pages_ratelimited(mapping);
2323
2324	if (unlikely(fatal_signal_pending(current))) {
2325	err = -EINTR;
2326	goto out;
2327	}
2328	}
2329
2330	/* page covers the boundary, find the boundary offset */
2331	if (index == curidx) {
2332	zerofrom = curpos & ~PAGE_MASK;
2333	/* if we will expand the thing last block will be filled */
2334	if (offset <= zerofrom) {
2335	goto out;
2336	}
2337	if (zerofrom & (blocksize-1)) {
2338	*bytes \|= (blocksize-1);
2339	(*bytes)++;
2340	}
2341	len = offset - zerofrom;
2342
2343	err = pagecache_write_begin(file, mapping, curpos, len,
2344	AOP_FLAG_UNINTERRUPTIBLE,
2345	&page, &fsdata);
2346	if (err)
2347	goto out;
2348	zero_user(page, zerofrom, len);
2349	err = pagecache_write_end(file, mapping, curpos, len, len,
2350	page, fsdata);
2351	if (err < 0)
2352	goto out;
2353	BUG_ON(err != len);
2354	err = 0;
2355	}
2356	out:
2357	return err;
2358	}
2359
2360	/*
2361	* For moronic filesystems that do not allow holes in file.
2362	* We may have to extend the file.
2363	*/
2364	int cont_write_begin(struct file file, struct address_space mapping,
2365	loff_t pos, unsigned len, unsigned flags,
2366	struct page pagep, void fsdata,
2367	get_block_t get_block, loff_t bytes)
2368	{
2369	struct inode *inode = mapping->host;
2370	unsigned blocksize = 1 << inode->i_blkbits;
2371	unsigned zerofrom;
2372	int err;
2373
2374	err = cont_expand_zero(file, mapping, pos, bytes);
2375	if (err)
2376	return err;
2377
2378	zerofrom = *bytes & ~PAGE_MASK;
2379	if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2380	*bytes \|= (blocksize-1);
2381	(*bytes)++;
2382	}
2383
2384	return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2385	}
2386	EXPORT_SYMBOL(cont_write_begin);
2387
2388	int block_commit_write(struct page *page, unsigned from, unsigned to)
2389	{
2390	struct inode *inode = page->mapping->host;
2391	__block_commit_write(inode,page,from,to);
2392	return 0;
2393	}
2394	EXPORT_SYMBOL(block_commit_write);
2395
2396	/*
2397	* block_page_mkwrite() is not allowed to change the file size as it gets
2398	* called from a page fault handler when a page is first dirtied. Hence we must
2399	* be careful to check for EOF conditions here. We set the page up correctly
2400	* for a written page which means we get ENOSPC checking when writing into
2401	* holes and correct delalloc and unwritten extent mapping on filesystems that
2402	* support these features.
2403	*
2404	* We are not allowed to take the i_mutex here so we have to play games to
2405	* protect against truncate races as the page could now be beyond EOF. Because
2406	* truncate writes the inode size before removing pages, once we have the
2407	* page lock we can determine safely if the page is beyond EOF. If it is not
2408	* beyond EOF, then the page is guaranteed safe against truncation until we
2409	* unlock the page.
2410	*
2411	* Direct callers of this function should protect against filesystem freezing
2412	* using sb_start_pagefault() - sb_end_pagefault() functions.
2413	*/
2414	int block_page_mkwrite(struct vm_area_struct vma, struct vm_fault vmf,
2415	get_block_t get_block)
2416	{
2417	struct page *page = vmf->page;
2418	struct inode *inode = file_inode(vma->vm_file);
2419	unsigned long end;
2420	loff_t size;
2421	int ret;
2422
2423	lock_page(page);
2424	size = i_size_read(inode);
2425	if ((page->mapping != inode->i_mapping) \|\|
2426	(page_offset(page) > size)) {
2427	/* We overload EFAULT to mean page got truncated */
2428	ret = -EFAULT;
2429	goto out_unlock;
2430	}
2431
2432	/* page is wholly or partially inside EOF */
2433	if (((page->index + 1) << PAGE_SHIFT) > size)
2434	end = size & ~PAGE_MASK;
2435	else
2436	end = PAGE_SIZE;
2437
2438	ret = __block_write_begin(page, 0, end, get_block);
2439	if (!ret)
2440	ret = block_commit_write(page, 0, end);
2441
2442	if (unlikely(ret < 0))
2443	goto out_unlock;
2444	set_page_dirty(page);
2445	wait_for_stable_page(page);
2446	return 0;
2447	out_unlock:
2448	unlock_page(page);
2449	return ret;
2450	}
2451	EXPORT_SYMBOL(block_page_mkwrite);
2452
2453	/*
2454	* nobh_write_begin()'s prereads are special: the buffer_heads are freed
2455	* immediately, while under the page lock. So it needs a special end_io
2456	* handler which does not touch the bh after unlocking it.
2457	*/
2458	static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2459	{
2460	__end_buffer_read_notouch(bh, uptodate);
2461	}
2462
2463	/*
2464	* Attach the singly-linked list of buffers created by nobh_write_begin, to
2465	* the page (converting it to circular linked list and taking care of page
2466	* dirty races).
2467	*/
2468	static void attach_nobh_buffers(struct page page, struct buffer_head head)
2469	{
2470	struct buffer_head *bh;
2471
2472	BUG_ON(!PageLocked(page));
2473
2474	spin_lock(&page->mapping->private_lock);
2475	bh = head;
2476	do {
2477	if (PageDirty(page))
2478	set_buffer_dirty(bh);
2479	if (!bh->b_this_page)
2480	bh->b_this_page = head;
2481	bh = bh->b_this_page;
2482	} while (bh != head);
2483	attach_page_buffers(page, head);
2484	spin_unlock(&page->mapping->private_lock);
2485	}
2486
2487	/*
2488	* On entry, the page is fully not uptodate.
2489	* On exit the page is fully uptodate in the areas outside (from,to)
2490	* The filesystem needs to handle block truncation upon failure.
2491	*/
2492	int nobh_write_begin(struct address_space *mapping,
2493	loff_t pos, unsigned len, unsigned flags,
2494	struct page pagep, void fsdata,
2495	get_block_t *get_block)
2496	{
2497	struct inode *inode = mapping->host;
2498	const unsigned blkbits = inode->i_blkbits;
2499	const unsigned blocksize = 1 << blkbits;
2500	struct buffer_head head, bh;
2501	struct page *page;
2502	pgoff_t index;
2503	unsigned from, to;
2504	unsigned block_in_page;
2505	unsigned block_start, block_end;
2506	sector_t block_in_file;
2507	int nr_reads = 0;
2508	int ret = 0;
2509	int is_mapped_to_disk = 1;
2510
2511	index = pos >> PAGE_SHIFT;
2512	from = pos & (PAGE_SIZE - 1);
2513	to = from + len;
2514
2515	page = grab_cache_page_write_begin(mapping, index, flags);
2516	if (!page)
2517	return -ENOMEM;
2518	*pagep = page;
2519	*fsdata = NULL;
2520
2521	if (page_has_buffers(page)) {
2522	ret = __block_write_begin(page, pos, len, get_block);
2523	if (unlikely(ret))
2524	goto out_release;
2525	return ret;
2526	}
2527
2528	if (PageMappedToDisk(page))
2529	return 0;
2530
2531	/*
2532	* Allocate buffers so that we can keep track of state, and potentially
2533	* attach them to the page if an error occurs. In the common case of
2534	* no error, they will just be freed again without ever being attached
2535	* to the page (which is all OK, because we're under the page lock).
2536	*
2537	* Be careful: the buffer linked list is a NULL terminated one, rather
2538	* than the circular one we're used to.
2539	*/
2540	head = alloc_page_buffers(page, blocksize, 0);
2541	if (!head) {
2542	ret = -ENOMEM;
2543	goto out_release;
2544	}
2545
2546	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2547
2548	/*
2549	* We loop across all blocks in the page, whether or not they are
2550	* part of the affected region. This is so we can discover if the
2551	* page is fully mapped-to-disk.
2552	*/
2553	for (block_start = 0, block_in_page = 0, bh = head;
2554	block_start < PAGE_SIZE;
2555	block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2556	int create;
2557
2558	block_end = block_start + blocksize;
2559	bh->b_state = 0;
2560	create = 1;
2561	if (block_start >= to)
2562	create = 0;
2563	ret = get_block(inode, block_in_file + block_in_page,
2564	bh, create);
2565	if (ret)
2566	goto failed;
2567	if (!buffer_mapped(bh))
2568	is_mapped_to_disk = 0;
2569	if (buffer_new(bh))
2570	unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2571	if (PageUptodate(page)) {
2572	set_buffer_uptodate(bh);
2573	continue;
2574	}
2575	if (buffer_new(bh) \|\| !buffer_mapped(bh)) {
2576	zero_user_segments(page, block_start, from,
2577	to, block_end);
2578	continue;
2579	}
2580	if (buffer_uptodate(bh))
2581	continue; /* reiserfs does this */
2582	if (block_start < from \|\| block_end > to) {
2583	lock_buffer(bh);
2584	bh->b_end_io = end_buffer_read_nobh;
2585	submit_bh(READ, bh);
2586	nr_reads++;
2587	}
2588	}
2589
2590	if (nr_reads) {
2591	/*
2592	* The page is locked, so these buffers are protected from
2593	* any VM or truncate activity. Hence we don't need to care
2594	* for the buffer_head refcounts.
2595	*/
2596	for (bh = head; bh; bh = bh->b_this_page) {
2597	wait_on_buffer(bh);
2598	if (!buffer_uptodate(bh))
2599	ret = -EIO;
2600	}
2601	if (ret)
2602	goto failed;
2603	}
2604
2605	if (is_mapped_to_disk)
2606	SetPageMappedToDisk(page);
2607
2608	fsdata = head; / to be released by nobh_write_end */
2609
2610	return 0;
2611
2612	failed:
2613	BUG_ON(!ret);
2614	/*
2615	* Error recovery is a bit difficult. We need to zero out blocks that
2616	* were newly allocated, and dirty them to ensure they get written out.
2617	* Buffers need to be attached to the page at this point, otherwise
2618	* the handling of potential IO errors during writeout would be hard
2619	* (could try doing synchronous writeout, but what if that fails too?)
2620	*/
2621	attach_nobh_buffers(page, head);
2622	page_zero_new_buffers(page, from, to);
2623
2624	out_release:
2625	unlock_page(page);
2626	put_page(page);
2627	*pagep = NULL;
2628
2629	return ret;
2630	}
2631	EXPORT_SYMBOL(nobh_write_begin);
2632
2633	int nobh_write_end(struct file file, struct address_space mapping,
2634	loff_t pos, unsigned len, unsigned copied,
2635	struct page page, void fsdata)
2636	{
2637	struct inode *inode = page->mapping->host;
2638	struct buffer_head *head = fsdata;
2639	struct buffer_head *bh;
2640	BUG_ON(fsdata != NULL && page_has_buffers(page));
2641
2642	if (unlikely(copied < len) && head)
2643	attach_nobh_buffers(page, head);
2644	if (page_has_buffers(page))
2645	return generic_write_end(file, mapping, pos, len,
2646	copied, page, fsdata);
2647
2648	SetPageUptodate(page);
2649	set_page_dirty(page);
2650	if (pos+copied > inode->i_size) {
2651	i_size_write(inode, pos+copied);
2652	mark_inode_dirty(inode);
2653	}
2654
2655	unlock_page(page);
2656	put_page(page);
2657
2658	while (head) {
2659	bh = head;
2660	head = head->b_this_page;
2661	free_buffer_head(bh);
2662	}
2663
2664	return copied;
2665	}
2666	EXPORT_SYMBOL(nobh_write_end);
2667
2668	/*
2669	* nobh_writepage() - based on block_full_write_page() except
2670	* that it tries to operate without attaching bufferheads to
2671	* the page.
2672	*/
2673	int nobh_writepage(struct page page, get_block_t get_block,
2674	struct writeback_control *wbc)
2675	{
2676	struct inode * const inode = page->mapping->host;
2677	loff_t i_size = i_size_read(inode);
2678	const pgoff_t end_index = i_size >> PAGE_SHIFT;
2679	unsigned offset;
2680	int ret;
2681
2682	/* Is the page fully inside i_size? */
2683	if (page->index < end_index)
2684	goto out;
2685
2686	/* Is the page fully outside i_size? (truncate in progress) */
2687	offset = i_size & (PAGE_SIZE-1);
2688	if (page->index >= end_index+1 \|\| !offset) {
2689	/*
2690	* The page may have dirty, unmapped buffers. For example,
2691	* they may have been added in ext3_writepage(). Make them
2692	* freeable here, so the page does not leak.
2693	*/
2694	#if 0
2695	/* Not really sure about this - do we need this ? */
2696	if (page->mapping->a_ops->invalidatepage)
2697	page->mapping->a_ops->invalidatepage(page, offset);
2698	#endif
2699	unlock_page(page);
2700	return 0; /* don't care */
2701	}
2702
2703	/*
2704	* The page straddles i_size. It must be zeroed out on each and every
2705	* writepage invocation because it may be mmapped. "A file is mapped
2706	* in multiples of the page size. For a file that is not a multiple of
2707	* the page size, the remaining memory is zeroed when mapped, and
2708	* writes to that region are not written out to the file."
2709	*/
2710	zero_user_segment(page, offset, PAGE_SIZE);
2711	out:
2712	ret = mpage_writepage(page, get_block, wbc);
2713	if (ret == -EAGAIN)
2714	ret = __block_write_full_page(inode, page, get_block, wbc,
2715	end_buffer_async_write);
2716	return ret;
2717	}
2718	EXPORT_SYMBOL(nobh_writepage);
2719
2720	int nobh_truncate_page(struct address_space *mapping,
2721	loff_t from, get_block_t *get_block)
2722	{
2723	pgoff_t index = from >> PAGE_SHIFT;
2724	unsigned offset = from & (PAGE_SIZE-1);
2725	unsigned blocksize;
2726	sector_t iblock;
2727	unsigned length, pos;
2728	struct inode *inode = mapping->host;
2729	struct page *page;
2730	struct buffer_head map_bh;
2731	int err;
2732
2733	blocksize = 1 << inode->i_blkbits;
2734	length = offset & (blocksize - 1);
2735
2736	/* Block boundary? Nothing to do */
2737	if (!length)
2738	return 0;
2739
2740	length = blocksize - length;
2741	iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2742
2743	page = grab_cache_page(mapping, index);
2744	err = -ENOMEM;
2745	if (!page)
2746	goto out;
2747
2748	if (page_has_buffers(page)) {
2749	has_buffers:
2750	unlock_page(page);
2751	put_page(page);
2752	return block_truncate_page(mapping, from, get_block);
2753	}
2754
2755	/* Find the buffer that contains "offset" */
2756	pos = blocksize;
2757	while (offset >= pos) {
2758	iblock++;
2759	pos += blocksize;
2760	}
2761
2762	map_bh.b_size = blocksize;
2763	map_bh.b_state = 0;
2764	err = get_block(inode, iblock, &map_bh, 0);
2765	if (err)
2766	goto unlock;
2767	/* unmapped? It's a hole - nothing to do */
2768	if (!buffer_mapped(&map_bh))
2769	goto unlock;
2770
2771	/* Ok, it's mapped. Make sure it's up-to-date */
2772	if (!PageUptodate(page)) {
2773	err = mapping->a_ops->readpage(NULL, page);
2774	if (err) {
2775	put_page(page);
2776	goto out;
2777	}
2778	lock_page(page);
2779	if (!PageUptodate(page)) {
2780	err = -EIO;
2781	goto unlock;
2782	}
2783	if (page_has_buffers(page))
2784	goto has_buffers;
2785	}
2786	zero_user(page, offset, length);
2787	set_page_dirty(page);
2788	err = 0;
2789
2790	unlock:
2791	unlock_page(page);
2792	put_page(page);
2793	out:
2794	return err;
2795	}
2796	EXPORT_SYMBOL(nobh_truncate_page);
2797
2798	int block_truncate_page(struct address_space *mapping,
2799	loff_t from, get_block_t *get_block)
2800	{
2801	pgoff_t index = from >> PAGE_SHIFT;
2802	unsigned offset = from & (PAGE_SIZE-1);
2803	unsigned blocksize;
2804	sector_t iblock;
2805	unsigned length, pos;
2806	struct inode *inode = mapping->host;
2807	struct page *page;
2808	struct buffer_head *bh;
2809	int err;
2810
2811	blocksize = 1 << inode->i_blkbits;
2812	length = offset & (blocksize - 1);
2813
2814	/* Block boundary? Nothing to do */
2815	if (!length)
2816	return 0;
2817
2818	length = blocksize - length;
2819	iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2820
2821	page = grab_cache_page(mapping, index);
2822	err = -ENOMEM;
2823	if (!page)
2824	goto out;
2825
2826	if (!page_has_buffers(page))
2827	create_empty_buffers(page, blocksize, 0);
2828
2829	/* Find the buffer that contains "offset" */
2830	bh = page_buffers(page);
2831	pos = blocksize;
2832	while (offset >= pos) {
2833	bh = bh->b_this_page;
2834	iblock++;
2835	pos += blocksize;
2836	}
2837
2838	err = 0;
2839	if (!buffer_mapped(bh)) {
2840	WARN_ON(bh->b_size != blocksize);
2841	err = get_block(inode, iblock, bh, 0);
2842	if (err)
2843	goto unlock;
2844	/* unmapped? It's a hole - nothing to do */
2845	if (!buffer_mapped(bh))
2846	goto unlock;
2847	}
2848
2849	/* Ok, it's mapped. Make sure it's up-to-date */
2850	if (PageUptodate(page))
2851	set_buffer_uptodate(bh);
2852
2853	if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2854	err = -EIO;
2855	ll_rw_block(READ, 1, &bh);
2856	wait_on_buffer(bh);
2857	/* Uhhuh. Read error. Complain and punt. */
2858	if (!buffer_uptodate(bh))
2859	goto unlock;
2860	}
2861
2862	zero_user(page, offset, length);
2863	mark_buffer_dirty(bh);
2864	err = 0;
2865
2866	unlock:
2867	unlock_page(page);
2868	put_page(page);
2869	out:
2870	return err;
2871	}
2872	EXPORT_SYMBOL(block_truncate_page);
2873
2874	/*
2875	* The generic ->writepage function for buffer-backed address_spaces
2876	*/
2877	int block_write_full_page(struct page page, get_block_t get_block,
2878	struct writeback_control *wbc)
2879	{
2880	struct inode * const inode = page->mapping->host;
2881	loff_t i_size = i_size_read(inode);
2882	const pgoff_t end_index = i_size >> PAGE_SHIFT;
2883	unsigned offset;
2884
2885	/* Is the page fully inside i_size? */
2886	if (page->index < end_index)
2887	return __block_write_full_page(inode, page, get_block, wbc,
2888	end_buffer_async_write);
2889
2890	/* Is the page fully outside i_size? (truncate in progress) */
2891	offset = i_size & (PAGE_SIZE-1);
2892	if (page->index >= end_index+1 \|\| !offset) {
2893	/*
2894	* The page may have dirty, unmapped buffers. For example,
2895	* they may have been added in ext3_writepage(). Make them
2896	* freeable here, so the page does not leak.
2897	*/
2898	do_invalidatepage(page, 0, PAGE_SIZE);
2899	unlock_page(page);
2900	return 0; /* don't care */
2901	}
2902
2903	/*
2904	* The page straddles i_size. It must be zeroed out on each and every
2905	* writepage invocation because it may be mmapped. "A file is mapped
2906	* in multiples of the page size. For a file that is not a multiple of
2907	* the page size, the remaining memory is zeroed when mapped, and
2908	* writes to that region are not written out to the file."
2909	*/
2910	zero_user_segment(page, offset, PAGE_SIZE);
2911	return __block_write_full_page(inode, page, get_block, wbc,
2912	end_buffer_async_write);
2913	}
2914	EXPORT_SYMBOL(block_write_full_page);
2915
2916	sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2917	get_block_t *get_block)
2918	{
2919	struct buffer_head tmp;
2920	struct inode *inode = mapping->host;
2921	tmp.b_state = 0;
2922	tmp.b_blocknr = 0;
2923	tmp.b_size = 1 << inode->i_blkbits;
2924	get_block(inode, block, &tmp, 0);
2925	return tmp.b_blocknr;
2926	}
2927	EXPORT_SYMBOL(generic_block_bmap);
2928
2929	static void end_bio_bh_io_sync(struct bio *bio)
2930	{
2931	struct buffer_head *bh = bio->bi_private;
2932
2933	if (unlikely(bio_flagged(bio, BIO_QUIET)))
2934	set_bit(BH_Quiet, &bh->b_state);
2935
2936	bh->b_end_io(bh, !bio->bi_error);
2937	bio_put(bio);
2938	}
2939
2940	/*
2941	* This allows us to do IO even on the odd last sectors
2942	* of a device, even if the block size is some multiple
2943	* of the physical sector size.
2944	*
2945	* We'll just truncate the bio to the size of the device,
2946	* and clear the end of the buffer head manually.
2947	*
2948	* Truly out-of-range accesses will turn into actual IO
2949	* errors, this only handles the "we need to be able to
2950	* do IO at the final sector" case.
2951	*/
2952	void guard_bio_eod(int rw, struct bio *bio)
2953	{
2954	sector_t maxsector;
2955	struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
2956	unsigned truncated_bytes;
2957
2958	maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2959	if (!maxsector)
2960	return;
2961
2962	/*
2963	* If the whole IO is past the end of the device,
2964	* let it through, and the IO layer will turn it into
2965	* an EIO.
2966	*/
2967	if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2968	return;
2969
2970	maxsector -= bio->bi_iter.bi_sector;
2971	if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
2972	return;
2973
2974	/* Uhhuh. We've got a bio that straddles the device size! */
2975	truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
2976
2977	/* Truncate the bio.. */
2978	bio->bi_iter.bi_size -= truncated_bytes;
2979	bvec->bv_len -= truncated_bytes;
2980
2981	/* ..and clear the end of the buffer for reads */
2982	if ((rw & RW_MASK) == READ) {
2983	zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
2984	truncated_bytes);
2985	}
2986	}
2987
2988	static int submit_bh_wbc(int rw, struct buffer_head *bh,
2989	unsigned long bio_flags, struct writeback_control *wbc)
2990	{
2991	struct bio *bio;
2992
2993	BUG_ON(!buffer_locked(bh));
2994	BUG_ON(!buffer_mapped(bh));
2995	BUG_ON(!bh->b_end_io);
2996	BUG_ON(buffer_delay(bh));
2997	BUG_ON(buffer_unwritten(bh));
2998
2999	/*
3000	* Only clear out a write error when rewriting
3001	*/
3002	if (test_set_buffer_req(bh) && (rw & WRITE))
3003	clear_buffer_write_io_error(bh);
3004
3005	/*
3006	* from here on down, it's all bio -- do the initial mapping,
3007	* submit_bio -> generic_make_request may further map this bio around
3008	*/
3009	bio = bio_alloc(GFP_NOIO, 1);
3010
3011	if (wbc) {
3012	wbc_init_bio(wbc, bio);
3013	wbc_account_io(wbc, bh->b_page, bh->b_size);
3014	}
3015
3016	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3017	bio->bi_bdev = bh->b_bdev;
3018
3019	bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3020	BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3021
3022	bio->bi_end_io = end_bio_bh_io_sync;
3023	bio->bi_private = bh;
3024	bio->bi_flags \|= bio_flags;
3025
3026	/* Take care of bh's that straddle the end of the device */
3027	guard_bio_eod(rw, bio);
3028
3029	if (buffer_meta(bh))
3030	rw \|= REQ_META;
3031	if (buffer_prio(bh))
3032	rw \|= REQ_PRIO;
3033
3034	submit_bio(rw, bio);
3035	return 0;
3036	}
3037
3038	int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3039	{
3040	return submit_bh_wbc(rw, bh, bio_flags, NULL);
3041	}
3042	EXPORT_SYMBOL_GPL(_submit_bh);
3043
3044	int submit_bh(int rw, struct buffer_head *bh)
3045	{
3046	return submit_bh_wbc(rw, bh, 0, NULL);
3047	}
3048	EXPORT_SYMBOL(submit_bh);
3049
3050	/**
3051	* ll_rw_block: low-level access to block devices (DEPRECATED)
3052	* @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3053	* @nr: number of &struct buffer_heads in the array
3054	* @bhs: array of pointers to &struct buffer_head
3055	*
3056	* ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3057	* requests an I/O operation on them, either a %READ or a %WRITE. The third
3058	* %READA option is described in the documentation for generic_make_request()
3059	* which ll_rw_block() calls.
3060	*
3061	* This function drops any buffer that it cannot get a lock on (with the
3062	* BH_Lock state bit), any buffer that appears to be clean when doing a write
3063	* request, and any buffer that appears to be up-to-date when doing read
3064	* request. Further it marks as clean buffers that are processed for
3065	* writing (the buffer cache won't assume that they are actually clean
3066	* until the buffer gets unlocked).
3067	*
3068	* ll_rw_block sets b_end_io to simple completion handler that marks
3069	* the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3070	* any waiters.
3071	*
3072	* All of the buffers must be for the same device, and must also be a
3073	* multiple of the current approved size for the device.
3074	*/
3075	void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3076	{
3077	int i;
3078
3079	for (i = 0; i < nr; i++) {
3080	struct buffer_head *bh = bhs[i];
3081
3082	if (!trylock_buffer(bh))
3083	continue;
3084	if (rw == WRITE) {
3085	if (test_clear_buffer_dirty(bh)) {
3086	bh->b_end_io = end_buffer_write_sync;
3087	get_bh(bh);
3088	submit_bh(WRITE, bh);
3089	continue;
3090	}
3091	} else {
3092	if (!buffer_uptodate(bh)) {
3093	bh->b_end_io = end_buffer_read_sync;
3094	get_bh(bh);
3095	submit_bh(rw, bh);
3096	continue;
3097	}
3098	}
3099	unlock_buffer(bh);
3100	}
3101	}
3102	EXPORT_SYMBOL(ll_rw_block);
3103
3104	void write_dirty_buffer(struct buffer_head *bh, int rw)
3105	{
3106	lock_buffer(bh);
3107	if (!test_clear_buffer_dirty(bh)) {
3108	unlock_buffer(bh);
3109	return;
3110	}
3111	bh->b_end_io = end_buffer_write_sync;
3112	get_bh(bh);
3113	submit_bh(rw, bh);
3114	}
3115	EXPORT_SYMBOL(write_dirty_buffer);
3116
3117	/*
3118	* For a data-integrity writeout, we need to wait upon any in-progress I/O
3119	* and then start new I/O and then wait upon it. The caller must have a ref on
3120	* the buffer_head.
3121	*/
3122	int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3123	{
3124	int ret = 0;
3125
3126	WARN_ON(atomic_read(&bh->b_count) < 1);
3127	lock_buffer(bh);
3128	if (test_clear_buffer_dirty(bh)) {
3129	get_bh(bh);
3130	bh->b_end_io = end_buffer_write_sync;
3131	ret = submit_bh(rw, bh);
3132	wait_on_buffer(bh);
3133	if (!ret && !buffer_uptodate(bh))
3134	ret = -EIO;
3135	} else {
3136	unlock_buffer(bh);
3137	}
3138	return ret;
3139	}
3140	EXPORT_SYMBOL(__sync_dirty_buffer);
3141
3142	int sync_dirty_buffer(struct buffer_head *bh)
3143	{
3144	return __sync_dirty_buffer(bh, WRITE_SYNC);
3145	}
3146	EXPORT_SYMBOL(sync_dirty_buffer);
3147
3148	/*
3149	* try_to_free_buffers() checks if all the buffers on this particular page
3150	* are unused, and releases them if so.
3151	*
3152	* Exclusion against try_to_free_buffers may be obtained by either
3153	* locking the page or by holding its mapping's private_lock.
3154	*
3155	* If the page is dirty but all the buffers are clean then we need to
3156	* be sure to mark the page clean as well. This is because the page
3157	* may be against a block device, and a later reattachment of buffers
3158	* to a dirty page will set all buffers dirty. Which would corrupt
3159	* filesystem data on the same device.
3160	*
3161	* The same applies to regular filesystem pages: if all the buffers are
3162	* clean then we set the page clean and proceed. To do that, we require
3163	* total exclusion from __set_page_dirty_buffers(). That is obtained with
3164	* private_lock.
3165	*
3166	* try_to_free_buffers() is non-blocking.
3167	*/
3168	static inline int buffer_busy(struct buffer_head *bh)
3169	{
3170	return atomic_read(&bh->b_count) \|
3171	(bh->b_state & ((1 << BH_Dirty) \| (1 << BH_Lock)));
3172	}
3173
3174	static int
3175	drop_buffers(struct page page, struct buffer_head *buffers_to_free)
3176	{
3177	struct buffer_head *head = page_buffers(page);
3178	struct buffer_head *bh;
3179
3180	bh = head;
3181	do {
3182	if (buffer_write_io_error(bh) && page->mapping)
3183	set_bit(AS_EIO, &page->mapping->flags);
3184	if (buffer_busy(bh))
3185	goto failed;
3186	bh = bh->b_this_page;
3187	} while (bh != head);
3188
3189	do {
3190	struct buffer_head *next = bh->b_this_page;
3191
3192	if (bh->b_assoc_map)
3193	__remove_assoc_queue(bh);
3194	bh = next;
3195	} while (bh != head);
3196	*buffers_to_free = head;
3197	__clear_page_buffers(page);
3198	return 1;
3199	failed:
3200	return 0;
3201	}
3202
3203	int try_to_free_buffers(struct page *page)
3204	{
3205	struct address_space * const mapping = page->mapping;
3206	struct buffer_head *buffers_to_free = NULL;
3207	int ret = 0;
3208
3209	BUG_ON(!PageLocked(page));
3210	if (PageWriteback(page))
3211	return 0;
3212
3213	if (mapping == NULL) { /* can this still happen? */
3214	ret = drop_buffers(page, &buffers_to_free);
3215	goto out;
3216	}
3217
3218	spin_lock(&mapping->private_lock);
3219	ret = drop_buffers(page, &buffers_to_free);
3220
3221	/*
3222	* If the filesystem writes its buffers by hand (eg ext3)
3223	* then we can have clean buffers against a dirty page. We
3224	* clean the page here; otherwise the VM will never notice
3225	* that the filesystem did any IO at all.
3226	*
3227	* Also, during truncate, discard_buffer will have marked all
3228	* the page's buffers clean. We discover that here and clean
3229	* the page also.
3230	*
3231	* private_lock must be held over this entire operation in order
3232	* to synchronise against __set_page_dirty_buffers and prevent the
3233	* dirty bit from being lost.
3234	*/
3235	if (ret)
3236	cancel_dirty_page(page);
3237	spin_unlock(&mapping->private_lock);
3238	out:
3239	if (buffers_to_free) {
3240	struct buffer_head *bh = buffers_to_free;
3241
3242	do {
3243	struct buffer_head *next = bh->b_this_page;
3244	free_buffer_head(bh);
3245	bh = next;
3246	} while (bh != buffers_to_free);
3247	}
3248	return ret;
3249	}
3250	EXPORT_SYMBOL(try_to_free_buffers);
3251
3252	/*
3253	* There are no bdflush tunables left. But distributions are
3254	* still running obsolete flush daemons, so we terminate them here.
3255	*
3256	* Use of bdflush() is deprecated and will be removed in a future kernel.
3257	* The `flush-X' kernel threads fully replace bdflush daemons and this call.
3258	*/
3259	SYSCALL_DEFINE2(bdflush, int, func, long, data)
3260	{
3261	static int msg_count;
3262
3263	if (!capable(CAP_SYS_ADMIN))
3264	return -EPERM;
3265
3266	if (msg_count < 5) {
3267	msg_count++;
3268	printk(KERN_INFO
3269	"warning: process `%s' used the obsolete bdflush"
3270	" system call\n", current->comm);
3271	printk(KERN_INFO "Fix your initscripts?\n");
3272	}
3273
3274	if (func == 1)
3275	do_exit(0);
3276	return 0;
3277	}
3278
3279	/*
3280	* Buffer-head allocation
3281	*/
3282	static struct kmem_cache *bh_cachep __read_mostly;
3283
3284	/*
3285	* Once the number of bh's in the machine exceeds this level, we start
3286	* stripping them in writeback.
3287	*/
3288	static unsigned long max_buffer_heads;
3289
3290	int buffer_heads_over_limit;
3291
3292	struct bh_accounting {
3293	int nr; /* Number of live bh's */
3294	int ratelimit; /* Limit cacheline bouncing */
3295	};
3296
3297	static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3298
3299	static void recalc_bh_state(void)
3300	{
3301	int i;
3302	int tot = 0;
3303
3304	if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3305	return;
3306	__this_cpu_write(bh_accounting.ratelimit, 0);
3307	for_each_online_cpu(i)
3308	tot += per_cpu(bh_accounting, i).nr;
3309	buffer_heads_over_limit = (tot > max_buffer_heads);
3310	}
3311
3312	struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3313	{
3314	struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3315	if (ret) {
3316	INIT_LIST_HEAD(&ret->b_assoc_buffers);
3317	preempt_disable();
3318	__this_cpu_inc(bh_accounting.nr);
3319	recalc_bh_state();
3320	preempt_enable();
3321	}
3322	return ret;
3323	}
3324	EXPORT_SYMBOL(alloc_buffer_head);
3325
3326	void free_buffer_head(struct buffer_head *bh)
3327	{
3328	BUG_ON(!list_empty(&bh->b_assoc_buffers));
3329	kmem_cache_free(bh_cachep, bh);
3330	preempt_disable();
3331	__this_cpu_dec(bh_accounting.nr);
3332	recalc_bh_state();
3333	preempt_enable();
3334	}
3335	EXPORT_SYMBOL(free_buffer_head);
3336
3337	static void buffer_exit_cpu(int cpu)
3338	{
3339	int i;
3340	struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3341
3342	for (i = 0; i < BH_LRU_SIZE; i++) {
3343	brelse(b->bhs[i]);
3344	b->bhs[i] = NULL;
3345	}
3346	this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3347	per_cpu(bh_accounting, cpu).nr = 0;
3348	}
3349
3350	static int buffer_cpu_notify(struct notifier_block *self,
3351	unsigned long action, void *hcpu)
3352	{
3353	if (action == CPU_DEAD \|\| action == CPU_DEAD_FROZEN)
3354	buffer_exit_cpu((unsigned long)hcpu);
3355	return NOTIFY_OK;
3356	}
3357
3358	/**
3359	* bh_uptodate_or_lock - Test whether the buffer is uptodate
3360	* @bh: struct buffer_head
3361	*
3362	* Return true if the buffer is up-to-date and false,
3363	* with the buffer locked, if not.
3364	*/
3365	int bh_uptodate_or_lock(struct buffer_head *bh)
3366	{
3367	if (!buffer_uptodate(bh)) {
3368	lock_buffer(bh);
3369	if (!buffer_uptodate(bh))
3370	return 0;
3371	unlock_buffer(bh);
3372	}
3373	return 1;
3374	}
3375	EXPORT_SYMBOL(bh_uptodate_or_lock);
3376
3377	/**
3378	* bh_submit_read - Submit a locked buffer for reading
3379	* @bh: struct buffer_head
3380	*
3381	* Returns zero on success and -EIO on error.
3382	*/
3383	int bh_submit_read(struct buffer_head *bh)
3384	{
3385	BUG_ON(!buffer_locked(bh));
3386
3387	if (buffer_uptodate(bh)) {
3388	unlock_buffer(bh);
3389	return 0;
3390	}
3391
3392	get_bh(bh);
3393	bh->b_end_io = end_buffer_read_sync;
3394	submit_bh(READ, bh);
3395	wait_on_buffer(bh);
3396	if (buffer_uptodate(bh))
3397	return 0;
3398	return -EIO;
3399	}
3400	EXPORT_SYMBOL(bh_submit_read);
3401
3402	void __init buffer_init(void)
3403	{
3404	unsigned long nrpages;
3405
3406	bh_cachep = kmem_cache_create("buffer_head",
3407	sizeof(struct buffer_head), 0,
3408	(SLAB_RECLAIM_ACCOUNT\|SLAB_PANIC\|
3409	SLAB_MEM_SPREAD),
3410	NULL);
3411
3412	/*
3413	* Limit the bh occupancy to 10% of ZONE_NORMAL
3414	*/
3415	nrpages = (nr_free_buffer_pages() * 10) / 100;
3416	max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3417	hotcpu_notifier(buffer_cpu_notify, 0);
3418	}