path: root/fs/buffer.c (plain)
blob: a89be9741d125ee5b567ed7a282f5f1a1d1b830b

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 <andrea@suse.de>
19	*/
20
21	#include <linux/kernel.h>
22	#include <linux/syscalls.h>
23	#include <linux/fs.h>
24	#include <linux/iomap.h>
25	#include <linux/mm.h>
26	#include <linux/percpu.h>
27	#include <linux/slab.h>
28	#include <linux/capability.h>
29	#include <linux/blkdev.h>
30	#include <linux/file.h>
31	#include <linux/quotaops.h>
32	#include <linux/highmem.h>
33	#include <linux/export.h>
34	#include <linux/backing-dev.h>
35	#include <linux/writeback.h>
36	#include <linux/hash.h>
37	#include <linux/suspend.h>
38	#include <linux/buffer_head.h>
39	#include <linux/task_io_accounting_ops.h>
40	#include <linux/bio.h>
41	#include <linux/notifier.h>
42	#include <linux/cpu.h>
43	#include <linux/bitops.h>
44	#include <linux/mpage.h>
45	#include <linux/bit_spinlock.h>
46	#include <trace/events/block.h>
47
48	static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
49	static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
50	unsigned long bio_flags,
51	struct writeback_control *wbc);
52
53	#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
54
55	void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
56	{
57	bh->b_end_io = handler;
58	bh->b_private = private;
59	}
60	EXPORT_SYMBOL(init_buffer);
61
62	inline void touch_buffer(struct buffer_head *bh)
63	{
64	trace_block_touch_buffer(bh);
65	mark_page_accessed(bh->b_page);
66	}
67	EXPORT_SYMBOL(touch_buffer);
68
69	void __lock_buffer(struct buffer_head *bh)
70	{
71	wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
72	}
73	EXPORT_SYMBOL(__lock_buffer);
74
75	void unlock_buffer(struct buffer_head *bh)
76	{
77	clear_bit_unlock(BH_Lock, &bh->b_state);
78	smp_mb__after_atomic();
79	wake_up_bit(&bh->b_state, BH_Lock);
80	}
81	EXPORT_SYMBOL(unlock_buffer);
82
83	/*
84	* Returns if the page has dirty or writeback buffers. If all the buffers
85	* are unlocked and clean then the PageDirty information is stale. If
86	* any of the pages are locked, it is assumed they are locked for IO.
87	*/
88	void buffer_check_dirty_writeback(struct page *page,
89	bool *dirty, bool *writeback)
90	{
91	struct buffer_head *head, *bh;
92	*dirty = false;
93	*writeback = false;
94
95	BUG_ON(!PageLocked(page));
96
97	if (!page_has_buffers(page))
98	return;
99
100	if (PageWriteback(page))
101	*writeback = true;
102
103	head = page_buffers(page);
104	bh = head;
105	do {
106	if (buffer_locked(bh))
107	*writeback = true;
108
109	if (buffer_dirty(bh))
110	*dirty = true;
111
112	bh = bh->b_this_page;
113	} while (bh != head);
114	}
115	EXPORT_SYMBOL(buffer_check_dirty_writeback);
116
117	/*
118	* Block until a buffer comes unlocked. This doesn't stop it
119	* from becoming locked again - you have to lock it yourself
120	* if you want to preserve its state.
121	*/
122	void __wait_on_buffer(struct buffer_head * bh)
123	{
124	wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
125	}
126	EXPORT_SYMBOL(__wait_on_buffer);
127
128	static void
129	__clear_page_buffers(struct page *page)
130	{
131	ClearPagePrivate(page);
132	set_page_private(page, 0);
133	put_page(page);
134	}
135
136	static void buffer_io_error(struct buffer_head *bh, char *msg)
137	{
138	if (!test_bit(BH_Quiet, &bh->b_state))
139	printk_ratelimited(KERN_ERR
140	"Buffer I/O error on dev %pg, logical block %llu%s\n",
141	bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
142	}
143
144	/*
145	* End-of-IO handler helper function which does not touch the bh after
146	* unlocking it.
147	* Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
148	* a race there is benign: unlock_buffer() only use the bh's address for
149	* hashing after unlocking the buffer, so it doesn't actually touch the bh
150	* itself.
151	*/
152	static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
153	{
154	if (uptodate) {
155	set_buffer_uptodate(bh);
156	} else {
157	/* This happens, due to failed read-ahead attempts. */
158	clear_buffer_uptodate(bh);
159	}
160	unlock_buffer(bh);
161	}
162
163	/*
164	* Default synchronous end-of-IO handler.. Just mark it up-to-date and
165	* unlock the buffer. This is what ll_rw_block uses too.
166	*/
167	void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
168	{
169	__end_buffer_read_notouch(bh, uptodate);
170	put_bh(bh);
171	}
172	EXPORT_SYMBOL(end_buffer_read_sync);
173
174	void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
175	{
176	if (uptodate) {
177	set_buffer_uptodate(bh);
178	} else {
179	buffer_io_error(bh, ", lost sync page write");
180	set_buffer_write_io_error(bh);
181	clear_buffer_uptodate(bh);
182	}
183	unlock_buffer(bh);
184	put_bh(bh);
185	}
186	EXPORT_SYMBOL(end_buffer_write_sync);
187
188	/*
189	* Various filesystems appear to want __find_get_block to be non-blocking.
190	* But it's the page lock which protects the buffers. To get around this,
191	* we get exclusion from try_to_free_buffers with the blockdev mapping's
192	* private_lock.
193	*
194	* Hack idea: for the blockdev mapping, i_bufferlist_lock contention
195	* may be quite high. This code could TryLock the page, and if that
196	* succeeds, there is no need to take private_lock. (But if
197	* private_lock is contended then so is mapping->tree_lock).
198	*/
199	static struct buffer_head *
200	__find_get_block_slow(struct block_device *bdev, sector_t block)
201	{
202	struct inode *bd_inode = bdev->bd_inode;
203	struct address_space *bd_mapping = bd_inode->i_mapping;
204	struct buffer_head *ret = NULL;
205	pgoff_t index;
206	struct buffer_head *bh;
207	struct buffer_head *head;
208	struct page *page;
209	int all_mapped = 1;
210	static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
211
212	index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
213	page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
214	if (!page)
215	goto out;
216
217	spin_lock(&bd_mapping->private_lock);
218	if (!page_has_buffers(page))
219	goto out_unlock;
220	head = page_buffers(page);
221	bh = head;
222	do {
223	if (!buffer_mapped(bh))
224	all_mapped = 0;
225	else if (bh->b_blocknr == block) {
226	ret = bh;
227	get_bh(bh);
228	goto out_unlock;
229	}
230	bh = bh->b_this_page;
231	} while (bh != head);
232
233	/* we might be here because some of the buffers on this page are
234	* not mapped. This is due to various races between
235	* file io on the block device and getblk. It gets dealt with
236	* elsewhere, don't buffer_error if we had some unmapped buffers
237	*/
238	ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
239	if (all_mapped && __ratelimit(&last_warned)) {
240	printk("__find_get_block_slow() failed. block=%llu, "
241	"b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
242	"device %pg blocksize: %d\n",
243	(unsigned long long)block,
244	(unsigned long long)bh->b_blocknr,
245	bh->b_state, bh->b_size, bdev,
246	1 << bd_inode->i_blkbits);
247	}
248	out_unlock:
249	spin_unlock(&bd_mapping->private_lock);
250	put_page(page);
251	out:
252	return ret;
253	}
254
255	/*
256	* Kick the writeback threads then try to free up some ZONE_NORMAL memory.
257	*/
258	static void free_more_memory(void)
259	{
260	struct zoneref *z;
261	int nid;
262
263	wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
264	yield();
265
266	for_each_online_node(nid) {
267
268	z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
269	gfp_zone(GFP_NOFS), NULL);
270	if (z->zone)
271	try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
272	GFP_NOFS, NULL);
273	}
274	}
275
276	/*
277	* I/O completion handler for block_read_full_page() - pages
278	* which come unlocked at the end of I/O.
279	*/
280	static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
281	{
282	unsigned long flags;
283	struct buffer_head *first;
284	struct buffer_head *tmp;
285	struct page *page;
286	int page_uptodate = 1;
287
288	BUG_ON(!buffer_async_read(bh));
289
290	page = bh->b_page;
291	if (uptodate) {
292	set_buffer_uptodate(bh);
293	} else {
294	clear_buffer_uptodate(bh);
295	buffer_io_error(bh, ", async page read");
296	SetPageError(page);
297	}
298
299	/*
300	* Be _very_ careful from here on. Bad things can happen if
301	* two buffer heads end IO at almost the same time and both
302	* decide that the page is now completely done.
303	*/
304	first = page_buffers(page);
305	local_irq_save(flags);
306	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
307	clear_buffer_async_read(bh);
308	unlock_buffer(bh);
309	tmp = bh;
310	do {
311	if (!buffer_uptodate(tmp))
312	page_uptodate = 0;
313	if (buffer_async_read(tmp)) {
314	BUG_ON(!buffer_locked(tmp));
315	goto still_busy;
316	}
317	tmp = tmp->b_this_page;
318	} while (tmp != bh);
319	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
320	local_irq_restore(flags);
321
322	/*
323	* If none of the buffers had errors and they are all
324	* uptodate then we can set the page uptodate.
325	*/
326	if (page_uptodate && !PageError(page))
327	SetPageUptodate(page);
328	unlock_page(page);
329	return;
330
331	still_busy:
332	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
333	local_irq_restore(flags);
334	return;
335	}
336
337	/*
338	* Completion handler for block_write_full_page() - pages which are unlocked
339	* during I/O, and which have PageWriteback cleared upon I/O completion.
340	*/
341	void end_buffer_async_write(struct buffer_head *bh, int uptodate)
342	{
343	unsigned long flags;
344	struct buffer_head *first;
345	struct buffer_head *tmp;
346	struct page *page;
347
348	BUG_ON(!buffer_async_write(bh));
349
350	page = bh->b_page;
351	if (uptodate) {
352	set_buffer_uptodate(bh);
353	} else {
354	buffer_io_error(bh, ", lost async page write");
355	mapping_set_error(page->mapping, -EIO);
356	set_buffer_write_io_error(bh);
357	clear_buffer_uptodate(bh);
358	SetPageError(page);
359	}
360
361	first = page_buffers(page);
362	local_irq_save(flags);
363	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
364
365	clear_buffer_async_write(bh);
366	unlock_buffer(bh);
367	tmp = bh->b_this_page;
368	while (tmp != bh) {
369	if (buffer_async_write(tmp)) {
370	BUG_ON(!buffer_locked(tmp));
371	goto still_busy;
372	}
373	tmp = tmp->b_this_page;
374	}
375	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
376	local_irq_restore(flags);
377	end_page_writeback(page);
378	return;
379
380	still_busy:
381	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
382	local_irq_restore(flags);
383	return;
384	}
385	EXPORT_SYMBOL(end_buffer_async_write);
386
387	/*
388	* If a page's buffers are under async readin (end_buffer_async_read
389	* completion) then there is a possibility that another thread of
390	* control could lock one of the buffers after it has completed
391	* but while some of the other buffers have not completed. This
392	* locked buffer would confuse end_buffer_async_read() into not unlocking
393	* the page. So the absence of BH_Async_Read tells end_buffer_async_read()
394	* that this buffer is not under async I/O.
395	*
396	* The page comes unlocked when it has no locked buffer_async buffers
397	* left.
398	*
399	* PageLocked prevents anyone starting new async I/O reads any of
400	* the buffers.
401	*
402	* PageWriteback is used to prevent simultaneous writeout of the same
403	* page.
404	*
405	* PageLocked prevents anyone from starting writeback of a page which is
406	* under read I/O (PageWriteback is only ever set against a locked page).
407	*/
408	static void mark_buffer_async_read(struct buffer_head *bh)
409	{
410	bh->b_end_io = end_buffer_async_read;
411	set_buffer_async_read(bh);
412	}
413
414	static void mark_buffer_async_write_endio(struct buffer_head *bh,
415	bh_end_io_t *handler)
416	{
417	bh->b_end_io = handler;
418	set_buffer_async_write(bh);
419	}
420
421	void mark_buffer_async_write(struct buffer_head *bh)
422	{
423	mark_buffer_async_write_endio(bh, end_buffer_async_write);
424	}
425	EXPORT_SYMBOL(mark_buffer_async_write);
426
427
428	/*
429	* fs/buffer.c contains helper functions for buffer-backed address space's
430	* fsync functions. A common requirement for buffer-based filesystems is
431	* that certain data from the backing blockdev needs to be written out for
432	* a successful fsync(). For example, ext2 indirect blocks need to be
433	* written back and waited upon before fsync() returns.
434	*
435	* The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
436	* inode_has_buffers() and invalidate_inode_buffers() are provided for the
437	* management of a list of dependent buffers at ->i_mapping->private_list.
438	*
439	* Locking is a little subtle: try_to_free_buffers() will remove buffers
440	* from their controlling inode's queue when they are being freed. But
441	* try_to_free_buffers() will be operating against the *blockdev* mapping
442	* at the time, not against the S_ISREG file which depends on those buffers.
443	* So the locking for private_list is via the private_lock in the address_space
444	* which backs the buffers. Which is different from the address_space
445	* against which the buffers are listed. So for a particular address_space,
446	* mapping->private_lock does *not* protect mapping->private_list! In fact,
447	* mapping->private_list will always be protected by the backing blockdev's
448	* ->private_lock.
449	*
450	* Which introduces a requirement: all buffers on an address_space's
451	* ->private_list must be from the same address_space: the blockdev's.
452	*
453	* address_spaces which do not place buffers at ->private_list via these
454	* utility functions are free to use private_lock and private_list for
455	* whatever they want. The only requirement is that list_empty(private_list)
456	* be true at clear_inode() time.
457	*
458	* FIXME: clear_inode should not call invalidate_inode_buffers(). The
459	* filesystems should do that. invalidate_inode_buffers() should just go
460	* BUG_ON(!list_empty).
461	*
462	* FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
463	* take an address_space, not an inode. And it should be called
464	* mark_buffer_dirty_fsync() to clearly define why those buffers are being
465	* queued up.
466	*
467	* FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
468	* list if it is already on a list. Because if the buffer is on a list,
469	* it *must* already be on the right one. If not, the filesystem is being
470	* silly. This will save a ton of locking. But first we have to ensure
471	* that buffers are taken *off* the old inode's list when they are freed
472	* (presumably in truncate). That requires careful auditing of all
473	* filesystems (do it inside bforget()). It could also be done by bringing
474	* b_inode back.
475	*/
476
477	/*
478	* The buffer's backing address_space's private_lock must be held
479	*/
480	static void __remove_assoc_queue(struct buffer_head *bh)
481	{
482	list_del_init(&bh->b_assoc_buffers);
483	WARN_ON(!bh->b_assoc_map);
484	if (buffer_write_io_error(bh))
485	set_bit(AS_EIO, &bh->b_assoc_map->flags);
486	bh->b_assoc_map = NULL;
487	}
488
489	int inode_has_buffers(struct inode *inode)
490	{
491	return !list_empty(&inode->i_data.private_list);
492	}
493
494	/*
495	* osync is designed to support O_SYNC io. It waits synchronously for
496	* all already-submitted IO to complete, but does not queue any new
497	* writes to the disk.
498	*
499	* To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
500	* you dirty the buffers, and then use osync_inode_buffers to wait for
501	* completion. Any other dirty buffers which are not yet queued for
502	* write will not be flushed to disk by the osync.
503	*/
504	static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
505	{
506	struct buffer_head *bh;
507	struct list_head *p;
508	int err = 0;
509
510	spin_lock(lock);
511	repeat:
512	list_for_each_prev(p, list) {
513	bh = BH_ENTRY(p);
514	if (buffer_locked(bh)) {
515	get_bh(bh);
516	spin_unlock(lock);
517	wait_on_buffer(bh);
518	if (!buffer_uptodate(bh))
519	err = -EIO;
520	brelse(bh);
521	spin_lock(lock);
522	goto repeat;
523	}
524	}
525	spin_unlock(lock);
526	return err;
527	}
528
529	static void do_thaw_one(struct super_block *sb, void *unused)
530	{
531	while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
532	printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
533	}
534
535	static void do_thaw_all(struct work_struct *work)
536	{
537	iterate_supers(do_thaw_one, NULL);
538	kfree(work);
539	printk(KERN_WARNING "Emergency Thaw complete\n");
540	}
541
542	/**
543	* emergency_thaw_all -- forcibly thaw every frozen filesystem
544	*
545	* Used for emergency unfreeze of all filesystems via SysRq
546	*/
547	void emergency_thaw_all(void)
548	{
549	struct work_struct *work;
550
551	work = kmalloc(sizeof(*work), GFP_ATOMIC);
552	if (work) {
553	INIT_WORK(work, do_thaw_all);
554	schedule_work(work);
555	}
556	}
557
558	/**
559	* sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
560	* @mapping: the mapping which wants those buffers written
561	*
562	* Starts I/O against the buffers at mapping->private_list, and waits upon
563	* that I/O.
564	*
565	* Basically, this is a convenience function for fsync().
566	* @mapping is a file or directory which needs those buffers to be written for
567	* a successful fsync().
568	*/
569	int sync_mapping_buffers(struct address_space *mapping)
570	{
571	struct address_space *buffer_mapping = mapping->private_data;
572
573	if (buffer_mapping == NULL || list_empty(&mapping->private_list))
574	return 0;
575
576	return fsync_buffers_list(&buffer_mapping->private_lock,
577	&mapping->private_list);
578	}
579	EXPORT_SYMBOL(sync_mapping_buffers);
580
581	/*
582	* Called when we've recently written block `bblock', and it is known that
583	* `bblock' was for a buffer_boundary() buffer. This means that the block at
584	* `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
585	* dirty, schedule it for IO. So that indirects merge nicely with their data.
586	*/
587	void write_boundary_block(struct block_device *bdev,
588	sector_t bblock, unsigned blocksize)
589	{
590	struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
591	if (bh) {
592	if (buffer_dirty(bh))
593	ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
594	put_bh(bh);
595	}
596	}
597
598	void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
599	{
600	struct address_space *mapping = inode->i_mapping;
601	struct address_space *buffer_mapping = bh->b_page->mapping;
602
603	mark_buffer_dirty(bh);
604	if (!mapping->private_data) {
605	mapping->private_data = buffer_mapping;
606	} else {
607	BUG_ON(mapping->private_data != buffer_mapping);
608	}
609	if (!bh->b_assoc_map) {
610	spin_lock(&buffer_mapping->private_lock);
611	list_move_tail(&bh->b_assoc_buffers,
612	&mapping->private_list);
613	bh->b_assoc_map = mapping;
614	spin_unlock(&buffer_mapping->private_lock);
615	}
616	}
617	EXPORT_SYMBOL(mark_buffer_dirty_inode);
618
619	/*
620	* Mark the page dirty, and set it dirty in the radix tree, and mark the inode
621	* dirty.
622	*
623	* If warn is true, then emit a warning if the page is not uptodate and has
624	* not been truncated.
625	*
626	* The caller must hold lock_page_memcg().
627	*/
628	static void __set_page_dirty(struct page *page, struct address_space *mapping,
629	int warn)
630	{
631	unsigned long flags;
632
633	spin_lock_irqsave(&mapping->tree_lock, flags);
634	if (page->mapping) { /* Race with truncate? */
635	WARN_ON_ONCE(warn && !PageUptodate(page));
636	account_page_dirtied(page, mapping);
637	radix_tree_tag_set(&mapping->page_tree,
638	page_index(page), PAGECACHE_TAG_DIRTY);
639	}
640	spin_unlock_irqrestore(&mapping->tree_lock, flags);
641	}
642
643	/*
644	* Add a page to the dirty page list.
645	*
646	* It is a sad fact of life that this function is called from several places
647	* deeply under spinlocking. It may not sleep.
648	*
649	* If the page has buffers, the uptodate buffers are set dirty, to preserve
650	* dirty-state coherency between the page and the buffers. It the page does
651	* not have buffers then when they are later attached they will all be set
652	* dirty.
653	*
654	* The buffers are dirtied before the page is dirtied. There's a small race
655	* window in which a writepage caller may see the page cleanness but not the
656	* buffer dirtiness. That's fine. If this code were to set the page dirty
657	* before the buffers, a concurrent writepage caller could clear the page dirty
658	* bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
659	* page on the dirty page list.
660	*
661	* We use private_lock to lock against try_to_free_buffers while using the
662	* page's buffer list. Also use this to protect against clean buffers being
663	* added to the page after it was set dirty.
664	*
665	* FIXME: may need to call ->reservepage here as well. That's rather up to the
666	* address_space though.
667	*/
668	int __set_page_dirty_buffers(struct page *page)
669	{
670	int newly_dirty;
671	struct address_space *mapping = page_mapping(page);
672
673	if (unlikely(!mapping))
674	return !TestSetPageDirty(page);
675
676	spin_lock(&mapping->private_lock);
677	if (page_has_buffers(page)) {
678	struct buffer_head *head = page_buffers(page);
679	struct buffer_head *bh = head;
680
681	do {
682	set_buffer_dirty(bh);
683	bh = bh->b_this_page;
684	} while (bh != head);
685	}
686	/*
687	* Lock out page->mem_cgroup migration to keep PageDirty
688	* synchronized with per-memcg dirty page counters.
689	*/
690	lock_page_memcg(page);
691	newly_dirty = !TestSetPageDirty(page);
692	spin_unlock(&mapping->private_lock);
693
694	if (newly_dirty)
695	__set_page_dirty(page, mapping, 1);
696
697	unlock_page_memcg(page);
698
699	if (newly_dirty)
700	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
701
702	return newly_dirty;
703	}
704	EXPORT_SYMBOL(__set_page_dirty_buffers);
705
706	/*
707	* Write out and wait upon a list of buffers.
708	*
709	* We have conflicting pressures: we want to make sure that all
710	* initially dirty buffers get waited on, but that any subsequently
711	* dirtied buffers don't. After all, we don't want fsync to last
712	* forever if somebody is actively writing to the file.
713	*
714	* Do this in two main stages: first we copy dirty buffers to a
715	* temporary inode list, queueing the writes as we go. Then we clean
716	* up, waiting for those writes to complete.
717	*
718	* During this second stage, any subsequent updates to the file may end
719	* up refiling the buffer on the original inode's dirty list again, so
720	* there is a chance we will end up with a buffer queued for write but
721	* not yet completed on that list. So, as a final cleanup we go through
722	* the osync code to catch these locked, dirty buffers without requeuing
723	* any newly dirty buffers for write.
724	*/
725	static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
726	{
727	struct buffer_head *bh;
728	struct list_head tmp;
729	struct address_space *mapping;
730	int err = 0, err2;
731	struct blk_plug plug;
732
733	INIT_LIST_HEAD(&tmp);
734	blk_start_plug(&plug);
735
736	spin_lock(lock);
737	while (!list_empty(list)) {
738	bh = BH_ENTRY(list->next);
739	mapping = bh->b_assoc_map;
740	__remove_assoc_queue(bh);
741	/* Avoid race with mark_buffer_dirty_inode() which does
742	* a lockless check and we rely on seeing the dirty bit */
743	smp_mb();
744	if (buffer_dirty(bh) || buffer_locked(bh)) {
745	list_add(&bh->b_assoc_buffers, &tmp);
746	bh->b_assoc_map = mapping;
747	if (buffer_dirty(bh)) {
748	get_bh(bh);
749	spin_unlock(lock);
750	/*
751	* Ensure any pending I/O completes so that
752	* write_dirty_buffer() actually writes the
753	* current contents - it is a noop if I/O is
754	* still in flight on potentially older
755	* contents.
756	*/
757	write_dirty_buffer(bh, WRITE_SYNC);
758
759	/*
760	* Kick off IO for the previous mapping. Note
761	* that we will not run the very last mapping,
762	* wait_on_buffer() will do that for us
763	* through sync_buffer().
764	*/
765	brelse(bh);
766	spin_lock(lock);
767	}
768	}
769	}
770
771	spin_unlock(lock);
772	blk_finish_plug(&plug);
773	spin_lock(lock);
774
775	while (!list_empty(&tmp)) {
776	bh = BH_ENTRY(tmp.prev);
777	get_bh(bh);
778	mapping = bh->b_assoc_map;
779	__remove_assoc_queue(bh);
780	/* Avoid race with mark_buffer_dirty_inode() which does
781	* a lockless check and we rely on seeing the dirty bit */
782	smp_mb();
783	if (buffer_dirty(bh)) {
784	list_add(&bh->b_assoc_buffers,
785	&mapping->private_list);
786	bh->b_assoc_map = mapping;
787	}
788	spin_unlock(lock);
789	wait_on_buffer(bh);
790	if (!buffer_uptodate(bh))
791	err = -EIO;
792	brelse(bh);
793	spin_lock(lock);
794	}
795
796	spin_unlock(lock);
797	err2 = osync_buffers_list(lock, list);
798	if (err)
799	return err;
800	else
801	return err2;
802	}
803
804	/*
805	* Invalidate any and all dirty buffers on a given inode. We are
806	* probably unmounting the fs, but that doesn't mean we have already
807	* done a sync(). Just drop the buffers from the inode list.
808	*
809	* NOTE: we take the inode's blockdev's mapping's private_lock. Which
810	* assumes that all the buffers are against the blockdev. Not true
811	* for reiserfs.
812	*/
813	void invalidate_inode_buffers(struct inode *inode)
814	{
815	if (inode_has_buffers(inode)) {
816	struct address_space *mapping = &inode->i_data;
817	struct list_head *list = &mapping->private_list;
818	struct address_space *buffer_mapping = mapping->private_data;
819
820	spin_lock(&buffer_mapping->private_lock);
821	while (!list_empty(list))
822	__remove_assoc_queue(BH_ENTRY(list->next));
823	spin_unlock(&buffer_mapping->private_lock);
824	}
825	}
826	EXPORT_SYMBOL(invalidate_inode_buffers);
827
828	/*
829	* Remove any clean buffers from the inode's buffer list. This is called
830	* when we're trying to free the inode itself. Those buffers can pin it.
831	*
832	* Returns true if all buffers were removed.
833	*/
834	int remove_inode_buffers(struct inode *inode)
835	{
836	int ret = 1;
837
838	if (inode_has_buffers(inode)) {
839	struct address_space *mapping = &inode->i_data;
840	struct list_head *list = &mapping->private_list;
841	struct address_space *buffer_mapping = mapping->private_data;
842
843	spin_lock(&buffer_mapping->private_lock);
844	while (!list_empty(list)) {
845	struct buffer_head *bh = BH_ENTRY(list->next);
846	if (buffer_dirty(bh)) {
847	ret = 0;
848	break;
849	}
850	__remove_assoc_queue(bh);
851	}
852	spin_unlock(&buffer_mapping->private_lock);
853	}
854	return ret;
855	}
856
857	/*
858	* Create the appropriate buffers when given a page for data area and
859	* the size of each buffer.. Use the bh->b_this_page linked list to
860	* follow the buffers created. Return NULL if unable to create more
861	* buffers.
862	*
863	* The retry flag is used to differentiate async IO (paging, swapping)
864	* which may not fail from ordinary buffer allocations.
865	*/
866	struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
867	int retry)
868	{
869	struct buffer_head *bh, *head;
870	long offset;
871
872	try_again:
873	head = NULL;
874	offset = PAGE_SIZE;
875	while ((offset -= size) >= 0) {
876	bh = alloc_buffer_head(GFP_NOFS);
877	if (!bh)
878	goto no_grow;
879
880	bh->b_this_page = head;
881	bh->b_blocknr = -1;
882	head = bh;
883
884	bh->b_size = size;
885
886	/* Link the buffer to its page */
887	set_bh_page(bh, page, offset);
888	}
889	return head;
890	/*
891	* In case anything failed, we just free everything we got.
892	*/
893	no_grow:
894	if (head) {
895	do {
896	bh = head;
897	head = head->b_this_page;
898	free_buffer_head(bh);
899	} while (head);
900	}
901
902	/*
903	* Return failure for non-async IO requests. Async IO requests
904	* are not allowed to fail, so we have to wait until buffer heads
905	* become available. But we don't want tasks sleeping with
906	* partially complete buffers, so all were released above.
907	*/
908	if (!retry)
909	return NULL;
910
911	/* We're _really_ low on memory. Now we just
912	* wait for old buffer heads to become free due to
913	* finishing IO. Since this is an async request and
914	* the reserve list is empty, we're sure there are
915	* async buffer heads in use.
916	*/
917	free_more_memory();
918	goto try_again;
919	}
920	EXPORT_SYMBOL_GPL(alloc_page_buffers);
921
922	static inline void
923	link_dev_buffers(struct page *page, struct buffer_head *head)
924	{
925	struct buffer_head *bh, *tail;
926
927	bh = head;
928	do {
929	tail = bh;
930	bh = bh->b_this_page;
931	} while (bh);
932	tail->b_this_page = head;
933	attach_page_buffers(page, head);
934	}
935
936	static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
937	{
938	sector_t retval = ~((sector_t)0);
939	loff_t sz = i_size_read(bdev->bd_inode);
940
941	if (sz) {
942	unsigned int sizebits = blksize_bits(size);
943	retval = (sz >> sizebits);
944	}
945	return retval;
946	}
947
948	/*
949	* Initialise the state of a blockdev page's buffers.
950	*/
951	static sector_t
952	init_page_buffers(struct page *page, struct block_device *bdev,
953	sector_t block, int size)
954	{
955	struct buffer_head *head = page_buffers(page);
956	struct buffer_head *bh = head;
957	int uptodate = PageUptodate(page);
958	sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
959
960	do {
961	if (!buffer_mapped(bh)) {
962	init_buffer(bh, NULL, NULL);
963	bh->b_bdev = bdev;
964	bh->b_blocknr = block;
965	if (uptodate)
966	set_buffer_uptodate(bh);
967	if (block < end_block)
968	set_buffer_mapped(bh);
969	}
970	block++;
971	bh = bh->b_this_page;
972	} while (bh != head);
973
974	/*
975	* Caller needs to validate requested block against end of device.
976	*/
977	return end_block;
978	}
979
980	/*
981	* Create the page-cache page that contains the requested block.
982	*
983	* This is used purely for blockdev mappings.
984	*/
985	static int
986	grow_dev_page(struct block_device *bdev, sector_t block,
987	pgoff_t index, int size, int sizebits, gfp_t gfp)
988	{
989	struct inode *inode = bdev->bd_inode;
990	struct page *page;
991	struct buffer_head *bh;
992	sector_t end_block;
993	int ret = 0; /* Will call free_more_memory() */
994	gfp_t gfp_mask;
995
996	gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
997
998	/*
999	* XXX: __getblk_slow() can not really deal with failure and
1000	* will endlessly loop on improvised global reclaim. Prefer
1001	* looping in the allocator rather than here, at least that
1002	* code knows what it's doing.
1003	*/
1004	gfp_mask |= __GFP_NOFAIL;
1005
1006	page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1007	if (!page)
1008	return ret;
1009
1010	BUG_ON(!PageLocked(page));
1011
1012	if (page_has_buffers(page)) {
1013	bh = page_buffers(page);
1014	if (bh->b_size == size) {
1015	end_block = init_page_buffers(page, bdev,
1016	(sector_t)index << sizebits,
1017	size);
1018	goto done;
1019	}
1020	if (!try_to_free_buffers(page))
1021	goto failed;
1022	}
1023
1024	/*
1025	* Allocate some buffers for this page
1026	*/
1027	bh = alloc_page_buffers(page, size, 0);
1028	if (!bh)
1029	goto failed;
1030
1031	/*
1032	* Link the page to the buffers and initialise them. Take the
1033	* lock to be atomic wrt __find_get_block(), which does not
1034	* run under the page lock.
1035	*/
1036	spin_lock(&inode->i_mapping->private_lock);
1037	link_dev_buffers(page, bh);
1038	end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1039	size);
1040	spin_unlock(&inode->i_mapping->private_lock);
1041	done:
1042	ret = (block < end_block) ? 1 : -ENXIO;
1043	failed:
1044	unlock_page(page);
1045	put_page(page);
1046	return ret;
1047	}
1048
1049	/*
1050	* Create buffers for the specified block device block's page. If
1051	* that page was dirty, the buffers are set dirty also.
1052	*/
1053	static int
1054	grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1055	{
1056	pgoff_t index;
1057	int sizebits;
1058
1059	sizebits = -1;
1060	do {
1061	sizebits++;
1062	} while ((size << sizebits) < PAGE_SIZE);
1063
1064	index = block >> sizebits;
1065
1066	/*
1067	* Check for a block which wants to lie outside our maximum possible
1068	* pagecache index. (this comparison is done using sector_t types).
1069	*/
1070	if (unlikely(index != block >> sizebits)) {
1071	printk(KERN_ERR "%s: requested out-of-range block %llu for "
1072	"device %pg\n",
1073	__func__, (unsigned long long)block,
1074	bdev);
1075	return -EIO;
1076	}
1077
1078	/* Create a page with the proper size buffers.. */
1079	return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1080	}
1081
1082	static struct buffer_head *
1083	__getblk_slow(struct block_device *bdev, sector_t block,
1084	unsigned size, gfp_t gfp)
1085	{
1086	/* Size must be multiple of hard sectorsize */
1087	if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1088	(size < 512 || size > PAGE_SIZE))) {
1089	printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1090	size);
1091	printk(KERN_ERR "logical block size: %d\n",
1092	bdev_logical_block_size(bdev));
1093
1094	dump_stack();
1095	return NULL;
1096	}
1097
1098	for (;;) {
1099	struct buffer_head *bh;
1100	int ret;
1101
1102	bh = __find_get_block(bdev, block, size);
1103	if (bh)
1104	return bh;
1105
1106	ret = grow_buffers(bdev, block, size, gfp);
1107	if (ret < 0)
1108	return NULL;
1109	if (ret == 0)
1110	free_more_memory();
1111	}
1112	}
1113
1114	/*
1115	* The relationship between dirty buffers and dirty pages:
1116	*
1117	* Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118	* the page is tagged dirty in its radix tree.
1119	*
1120	* At all times, the dirtiness of the buffers represents the dirtiness of
1121	* subsections of the page. If the page has buffers, the page dirty bit is
1122	* merely a hint about the true dirty state.
1123	*
1124	* When a page is set dirty in its entirety, all its buffers are marked dirty
1125	* (if the page has buffers).
1126	*
1127	* When a buffer is marked dirty, its page is dirtied, but the page's other
1128	* buffers are not.
1129	*
1130	* Also. When blockdev buffers are explicitly read with bread(), they
1131	* individually become uptodate. But their backing page remains not
1132	* uptodate - even if all of its buffers are uptodate. A subsequent
1133	* block_read_full_page() against that page will discover all the uptodate
1134	* buffers, will set the page uptodate and will perform no I/O.
1135	*/
1136
1137	/**
1138	* mark_buffer_dirty - mark a buffer_head as needing writeout
1139	* @bh: the buffer_head to mark dirty
1140	*
1141	* mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142	* backing page dirty, then tag the page as dirty in its address_space's radix
1143	* tree and then attach the address_space's inode to its superblock's dirty
1144	* inode list.
1145	*
1146	* mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1147	* mapping->tree_lock and mapping->host->i_lock.
1148	*/
1149	void mark_buffer_dirty(struct buffer_head *bh)
1150	{
1151	WARN_ON_ONCE(!buffer_uptodate(bh));
1152
1153	trace_block_dirty_buffer(bh);
1154
1155	/*
1156	* Very *carefully* optimize the it-is-already-dirty case.
1157	*
1158	* Don't let the final "is it dirty" escape to before we
1159	* perhaps modified the buffer.
1160	*/
1161	if (buffer_dirty(bh)) {
1162	smp_mb();
1163	if (buffer_dirty(bh))
1164	return;
1165	}
1166
1167	if (!test_set_buffer_dirty(bh)) {
1168	struct page *page = bh->b_page;
1169	struct address_space *mapping = NULL;
1170
1171	lock_page_memcg(page);
1172	if (!TestSetPageDirty(page)) {
1173	mapping = page_mapping(page);
1174	if (mapping)
1175	__set_page_dirty(page, mapping, 0);
1176	}
1177	unlock_page_memcg(page);
1178	if (mapping)
1179	__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1180	}
1181	}
1182	EXPORT_SYMBOL(mark_buffer_dirty);
1183
1184	/*
1185	* Decrement a buffer_head's reference count. If all buffers against a page
1186	* have zero reference count, are clean and unlocked, and if the page is clean
1187	* and unlocked then try_to_free_buffers() may strip the buffers from the page
1188	* in preparation for freeing it (sometimes, rarely, buffers are removed from
1189	* a page but it ends up not being freed, and buffers may later be reattached).
1190	*/
1191	void __brelse(struct buffer_head * buf)
1192	{
1193	if (atomic_read(&buf->b_count)) {
1194	put_bh(buf);
1195	return;
1196	}
1197	WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1198	}
1199	EXPORT_SYMBOL(__brelse);
1200
1201	/*
1202	* bforget() is like brelse(), except it discards any
1203	* potentially dirty data.
1204	*/
1205	void __bforget(struct buffer_head *bh)
1206	{
1207	clear_buffer_dirty(bh);
1208	if (bh->b_assoc_map) {
1209	struct address_space *buffer_mapping = bh->b_page->mapping;
1210
1211	spin_lock(&buffer_mapping->private_lock);
1212	list_del_init(&bh->b_assoc_buffers);
1213	bh->b_assoc_map = NULL;
1214	spin_unlock(&buffer_mapping->private_lock);
1215	}
1216	__brelse(bh);
1217	}
1218	EXPORT_SYMBOL(__bforget);
1219
1220	static struct buffer_head *__bread_slow(struct buffer_head *bh)
1221	{
1222	lock_buffer(bh);
1223	if (buffer_uptodate(bh)) {
1224	unlock_buffer(bh);
1225	return bh;
1226	} else {
1227	get_bh(bh);
1228	bh->b_end_io = end_buffer_read_sync;
1229	submit_bh(REQ_OP_READ, 0, bh);
1230	wait_on_buffer(bh);
1231	if (buffer_uptodate(bh))
1232	return bh;
1233	}
1234	brelse(bh);
1235	return NULL;
1236	}
1237
1238	/*
1239	* Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1240	* The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1241	* refcount elevated by one when they're in an LRU. A buffer can only appear
1242	* once in a particular CPU's LRU. A single buffer can be present in multiple
1243	* CPU's LRUs at the same time.
1244	*
1245	* This is a transparent caching front-end to sb_bread(), sb_getblk() and
1246	* sb_find_get_block().
1247	*
1248	* The LRUs themselves only need locking against invalidate_bh_lrus. We use
1249	* a local interrupt disable for that.
1250	*/
1251
1252	#define BH_LRU_SIZE 16
1253
1254	struct bh_lru {
1255	struct buffer_head *bhs[BH_LRU_SIZE];
1256	};
1257
1258	static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1259
1260	#ifdef CONFIG_SMP
1261	#define bh_lru_lock() local_irq_disable()
1262	#define bh_lru_unlock() local_irq_enable()
1263	#else
1264	#define bh_lru_lock() preempt_disable()
1265	#define bh_lru_unlock() preempt_enable()
1266	#endif
1267
1268	static inline void check_irqs_on(void)
1269	{
1270	#ifdef irqs_disabled
1271	BUG_ON(irqs_disabled());
1272	#endif
1273	}
1274
1275	/*
1276	* The LRU management algorithm is dopey-but-simple. Sorry.
1277	*/
1278	static void bh_lru_install(struct buffer_head *bh)
1279	{
1280	struct buffer_head *evictee = NULL;
1281
1282	check_irqs_on();
1283	bh_lru_lock();
1284	if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1285	struct buffer_head *bhs[BH_LRU_SIZE];
1286	int in;
1287	int out = 0;
1288
1289	get_bh(bh);
1290	bhs[out++] = bh;
1291	for (in = 0; in < BH_LRU_SIZE; in++) {
1292	struct buffer_head *bh2 =
1293	__this_cpu_read(bh_lrus.bhs[in]);
1294
1295	if (bh2 == bh) {
1296	__brelse(bh2);
1297	} else {
1298	if (out >= BH_LRU_SIZE) {
1299	BUG_ON(evictee != NULL);
1300	evictee = bh2;
1301	} else {
1302	bhs[out++] = bh2;
1303	}
1304	}
1305	}
1306	while (out < BH_LRU_SIZE)
1307	bhs[out++] = NULL;
1308	memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1309	}
1310	bh_lru_unlock();
1311
1312	if (evictee)
1313	__brelse(evictee);
1314	}
1315
1316	/*
1317	* Look up the bh in this cpu's LRU. If it's there, move it to the head.
1318	*/
1319	static struct buffer_head *
1320	lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1321	{
1322	struct buffer_head *ret = NULL;
1323	unsigned int i;
1324
1325	check_irqs_on();
1326	bh_lru_lock();
1327	for (i = 0; i < BH_LRU_SIZE; i++) {
1328	struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1329
1330	if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1331	bh->b_size == size) {
1332	if (i) {
1333	while (i) {
1334	__this_cpu_write(bh_lrus.bhs[i],
1335	__this_cpu_read(bh_lrus.bhs[i - 1]));
1336	i--;
1337	}
1338	__this_cpu_write(bh_lrus.bhs[0], bh);
1339	}
1340	get_bh(bh);
1341	ret = bh;
1342	break;
1343	}
1344	}
1345	bh_lru_unlock();
1346	return ret;
1347	}
1348
1349	/*
1350	* Perform a pagecache lookup for the matching buffer. If it's there, refresh
1351	* it in the LRU and mark it as accessed. If it is not present then return
1352	* NULL
1353	*/
1354	struct buffer_head *
1355	__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1356	{
1357	struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1358
1359	if (bh == NULL) {
1360	/* __find_get_block_slow will mark the page accessed */
1361	bh = __find_get_block_slow(bdev, block);
1362	if (bh)
1363	bh_lru_install(bh);
1364	} else
1365	touch_buffer(bh);
1366
1367	return bh;
1368	}
1369	EXPORT_SYMBOL(__find_get_block);
1370
1371	/*
1372	* __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1373	* which corresponds to the passed block_device, block and size. The
1374	* returned buffer has its reference count incremented.
1375	*
1376	* __getblk_gfp() will lock up the machine if grow_dev_page's
1377	* try_to_free_buffers() attempt is failing. FIXME, perhaps?
1378	*/
1379	struct buffer_head *
1380	__getblk_gfp(struct block_device *bdev, sector_t block,
1381	unsigned size, gfp_t gfp)
1382	{
1383	struct buffer_head *bh = __find_get_block(bdev, block, size);
1384
1385	might_sleep();
1386	if (bh == NULL)
1387	bh = __getblk_slow(bdev, block, size, gfp);
1388	return bh;
1389	}
1390	EXPORT_SYMBOL(__getblk_gfp);
1391
1392	/*
1393	* Do async read-ahead on a buffer..
1394	*/
1395	void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1396	{
1397	struct buffer_head *bh = __getblk(bdev, block, size);
1398	if (likely(bh)) {
1399	ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1400	brelse(bh);
1401	}
1402	}
1403	EXPORT_SYMBOL(__breadahead);
1404
1405	/**
1406	* __bread_gfp() - reads a specified block and returns the bh
1407	* @bdev: the block_device to read from
1408	* @block: number of block
1409	* @size: size (in bytes) to read
1410	* @gfp: page allocation flag
1411	*
1412	* Reads a specified block, and returns buffer head that contains it.
1413	* The page cache can be allocated from non-movable area
1414	* not to prevent page migration if you set gfp to zero.
1415	* It returns NULL if the block was unreadable.
1416	*/
1417	struct buffer_head *
1418	__bread_gfp(struct block_device *bdev, sector_t block,
1419	unsigned size, gfp_t gfp)
1420	{
1421	struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1422
1423	if (likely(bh) && !buffer_uptodate(bh))
1424	bh = __bread_slow(bh);
1425	return bh;
1426	}
1427	EXPORT_SYMBOL(__bread_gfp);
1428
1429	/*
1430	* invalidate_bh_lrus() is called rarely - but not only at unmount.
1431	* This doesn't race because it runs in each cpu either in irq
1432	* or with preempt disabled.
1433	*/
1434	static void invalidate_bh_lru(void *arg)
1435	{
1436	struct bh_lru *b = &get_cpu_var(bh_lrus);
1437	int i;
1438
1439	for (i = 0; i < BH_LRU_SIZE; i++) {
1440	brelse(b->bhs[i]);
1441	b->bhs[i] = NULL;
1442	}
1443	put_cpu_var(bh_lrus);
1444	}
1445
1446	static bool has_bh_in_lru(int cpu, void *dummy)
1447	{
1448	struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1449	int i;
1450
1451	for (i = 0; i < BH_LRU_SIZE; i++) {
1452	if (b->bhs[i])
1453	return 1;
1454	}
1455
1456	return 0;
1457	}
1458
1459	void invalidate_bh_lrus(void)
1460	{
1461	on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1462	}
1463	EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1464
1465	void set_bh_page(struct buffer_head *bh,
1466	struct page *page, unsigned long offset)
1467	{
1468	bh->b_page = page;
1469	BUG_ON(offset >= PAGE_SIZE);
1470	if (PageHighMem(page))
1471	/*
1472	* This catches illegal uses and preserves the offset:
1473	*/
1474	bh->b_data = (char *)(0 + offset);
1475	else
1476	bh->b_data = page_address(page) + offset;
1477	}
1478	EXPORT_SYMBOL(set_bh_page);
1479
1480	/*
1481	* Called when truncating a buffer on a page completely.
1482	*/
1483
1484	/* Bits that are cleared during an invalidate */
1485	#define BUFFER_FLAGS_DISCARD \
1486	(1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1487	1 << BH_Delay | 1 << BH_Unwritten)
1488
1489	static void discard_buffer(struct buffer_head * bh)
1490	{
1491	unsigned long b_state, b_state_old;
1492
1493	lock_buffer(bh);
1494	clear_buffer_dirty(bh);
1495	bh->b_bdev = NULL;
1496	b_state = bh->b_state;
1497	for (;;) {
1498	b_state_old = cmpxchg(&bh->b_state, b_state,
1499	(b_state & ~BUFFER_FLAGS_DISCARD));
1500	if (b_state_old == b_state)
1501	break;
1502	b_state = b_state_old;
1503	}
1504	unlock_buffer(bh);
1505	}
1506
1507	/**
1508	* block_invalidatepage - invalidate part or all of a buffer-backed page
1509	*
1510	* @page: the page which is affected
1511	* @offset: start of the range to invalidate
1512	* @length: length of the range to invalidate
1513	*
1514	* block_invalidatepage() is called when all or part of the page has become
1515	* invalidated by a truncate operation.
1516	*
1517	* block_invalidatepage() does not have to release all buffers, but it must
1518	* ensure that no dirty buffer is left outside @offset and that no I/O
1519	* is underway against any of the blocks which are outside the truncation
1520	* point. Because the caller is about to free (and possibly reuse) those
1521	* blocks on-disk.
1522	*/
1523	void block_invalidatepage(struct page *page, unsigned int offset,
1524	unsigned int length)
1525	{
1526	struct buffer_head *head, *bh, *next;
1527	unsigned int curr_off = 0;
1528	unsigned int stop = length + offset;
1529
1530	BUG_ON(!PageLocked(page));
1531	if (!page_has_buffers(page))
1532	goto out;
1533
1534	/*
1535	* Check for overflow
1536	*/
1537	BUG_ON(stop > PAGE_SIZE || stop < length);
1538
1539	head = page_buffers(page);
1540	bh = head;
1541	do {
1542	unsigned int next_off = curr_off + bh->b_size;
1543	next = bh->b_this_page;
1544
1545	/*
1546	* Are we still fully in range ?
1547	*/
1548	if (next_off > stop)
1549	goto out;
1550
1551	/*
1552	* is this block fully invalidated?
1553	*/
1554	if (offset <= curr_off)
1555	discard_buffer(bh);
1556	curr_off = next_off;
1557	bh = next;
1558	} while (bh != head);
1559
1560	/*
1561	* We release buffers only if the entire page is being invalidated.
1562	* The get_block cached value has been unconditionally invalidated,
1563	* so real IO is not possible anymore.
1564	*/
1565	if (offset == 0)
1566	try_to_release_page(page, 0);
1567	out:
1568	return;
1569	}
1570	EXPORT_SYMBOL(block_invalidatepage);
1571
1572
1573	/*
1574	* We attach and possibly dirty the buffers atomically wrt
1575	* __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1576	* is already excluded via the page lock.
1577	*/
1578	void create_empty_buffers(struct page *page,
1579	unsigned long blocksize, unsigned long b_state)
1580	{
1581	struct buffer_head *bh, *head, *tail;
1582
1583	head = alloc_page_buffers(page, blocksize, 1);
1584	bh = head;
1585	do {
1586	bh->b_state |= b_state;
1587	tail = bh;
1588	bh = bh->b_this_page;
1589	} while (bh);
1590	tail->b_this_page = head;
1591
1592	spin_lock(&page->mapping->private_lock);
1593	if (PageUptodate(page) || PageDirty(page)) {
1594	bh = head;
1595	do {
1596	if (PageDirty(page))
1597	set_buffer_dirty(bh);
1598	if (PageUptodate(page))
1599	set_buffer_uptodate(bh);
1600	bh = bh->b_this_page;
1601	} while (bh != head);
1602	}
1603	attach_page_buffers(page, head);
1604	spin_unlock(&page->mapping->private_lock);
1605	}
1606	EXPORT_SYMBOL(create_empty_buffers);
1607
1608	/*
1609	* We are taking a block for data and we don't want any output from any
1610	* buffer-cache aliases starting from return from that function and
1611	* until the moment when something will explicitly mark the buffer
1612	* dirty (hopefully that will not happen until we will free that block ;-)
1613	* We don't even need to mark it not-uptodate - nobody can expect
1614	* anything from a newly allocated buffer anyway. We used to used
1615	* unmap_buffer() for such invalidation, but that was wrong. We definitely
1616	* don't want to mark the alias unmapped, for example - it would confuse
1617	* anyone who might pick it with bread() afterwards...
1618	*
1619	* Also.. Note that bforget() doesn't lock the buffer. So there can
1620	* be writeout I/O going on against recently-freed buffers. We don't
1621	* wait on that I/O in bforget() - it's more efficient to wait on the I/O
1622	* only if we really need to. That happens here.
1623	*/
1624	void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1625	{
1626	struct buffer_head *old_bh;
1627
1628	might_sleep();
1629
1630	old_bh = __find_get_block_slow(bdev, block);
1631	if (old_bh) {
1632	clear_buffer_dirty(old_bh);
1633	wait_on_buffer(old_bh);
1634	clear_buffer_req(old_bh);
1635	__brelse(old_bh);
1636	}
1637	}
1638	EXPORT_SYMBOL(unmap_underlying_metadata);
1639
1640	/*
1641	* Size is a power-of-two in the range 512..PAGE_SIZE,
1642	* and the case we care about most is PAGE_SIZE.
1643	*
1644	* So this *could* possibly be written with those
1645	* constraints in mind (relevant mostly if some
1646	* architecture has a slow bit-scan instruction)
1647	*/
1648	static inline int block_size_bits(unsigned int blocksize)
1649	{
1650	return ilog2(blocksize);
1651	}
1652
1653	static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1654	{
1655	BUG_ON(!PageLocked(page));
1656
1657	if (!page_has_buffers(page))
1658	create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1659	return page_buffers(page);
1660	}
1661
1662	/*
1663	* NOTE! All mapped/uptodate combinations are valid:
1664	*
1665	* Mapped Uptodate Meaning
1666	*
1667	* No No "unknown" - must do get_block()
1668	* No Yes "hole" - zero-filled
1669	* Yes No "allocated" - allocated on disk, not read in
1670	* Yes Yes "valid" - allocated and up-to-date in memory.
1671	*
1672	* "Dirty" is valid only with the last case (mapped+uptodate).
1673	*/
1674
1675	/*
1676	* While block_write_full_page is writing back the dirty buffers under
1677	* the page lock, whoever dirtied the buffers may decide to clean them
1678	* again at any time. We handle that by only looking at the buffer
1679	* state inside lock_buffer().
1680	*
1681	* If block_write_full_page() is called for regular writeback
1682	* (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1683	* locked buffer. This only can happen if someone has written the buffer
1684	* directly, with submit_bh(). At the address_space level PageWriteback
1685	* prevents this contention from occurring.
1686	*
1687	* If block_write_full_page() is called with wbc->sync_mode ==
1688	* WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1689	* causes the writes to be flagged as synchronous writes.
1690	*/
1691	int __block_write_full_page(struct inode *inode, struct page *page,
1692	get_block_t *get_block, struct writeback_control *wbc,
1693	bh_end_io_t *handler)
1694	{
1695	int err;
1696	sector_t block;
1697	sector_t last_block;
1698	struct buffer_head *bh, *head;
1699	unsigned int blocksize, bbits;
1700	int nr_underway = 0;
1701	int write_flags = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : 0);
1702
1703	head = create_page_buffers(page, inode,
1704	(1 << BH_Dirty)|(1 << BH_Uptodate));
1705
1706	/*
1707	* Be very careful. We have no exclusion from __set_page_dirty_buffers
1708	* here, and the (potentially unmapped) buffers may become dirty at
1709	* any time. If a buffer becomes dirty here after we've inspected it
1710	* then we just miss that fact, and the page stays dirty.
1711	*
1712	* Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1713	* handle that here by just cleaning them.
1714	*/
1715
1716	bh = head;
1717	blocksize = bh->b_size;
1718	bbits = block_size_bits(blocksize);
1719
1720	block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1721	last_block = (i_size_read(inode) - 1) >> bbits;
1722
1723	/*
1724	* Get all the dirty buffers mapped to disk addresses and
1725	* handle any aliases from the underlying blockdev's mapping.
1726	*/
1727	do {
1728	if (block > last_block) {
1729	/*
1730	* mapped buffers outside i_size will occur, because
1731	* this page can be outside i_size when there is a
1732	* truncate in progress.
1733	*/
1734	/*
1735	* The buffer was zeroed by block_write_full_page()
1736	*/
1737	clear_buffer_dirty(bh);
1738	set_buffer_uptodate(bh);
1739	} else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1740	buffer_dirty(bh)) {
1741	WARN_ON(bh->b_size != blocksize);
1742	err = get_block(inode, block, bh, 1);
1743	if (err)
1744	goto recover;
1745	clear_buffer_delay(bh);
1746	if (buffer_new(bh)) {
1747	/* blockdev mappings never come here */
1748	clear_buffer_new(bh);
1749	unmap_underlying_metadata(bh->b_bdev,
1750	bh->b_blocknr);
1751	}
1752	}
1753	bh = bh->b_this_page;
1754	block++;
1755	} while (bh != head);
1756
1757	do {
1758	if (!buffer_mapped(bh))
1759	continue;
1760	/*
1761	* If it's a fully non-blocking write attempt and we cannot
1762	* lock the buffer then redirty the page. Note that this can
1763	* potentially cause a busy-wait loop from writeback threads
1764	* and kswapd activity, but those code paths have their own
1765	* higher-level throttling.
1766	*/
1767	if (wbc->sync_mode != WB_SYNC_NONE) {
1768	lock_buffer(bh);
1769	} else if (!trylock_buffer(bh)) {
1770	redirty_page_for_writepage(wbc, page);
1771	continue;
1772	}
1773	if (test_clear_buffer_dirty(bh)) {
1774	mark_buffer_async_write_endio(bh, handler);
1775	} else {
1776	unlock_buffer(bh);
1777	}
1778	} while ((bh = bh->b_this_page) != head);
1779
1780	/*
1781	* The page and its buffers are protected by PageWriteback(), so we can
1782	* drop the bh refcounts early.
1783	*/
1784	BUG_ON(PageWriteback(page));
1785	set_page_writeback(page);
1786
1787	do {
1788	struct buffer_head *next = bh->b_this_page;
1789	if (buffer_async_write(bh)) {
1790	submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1791	nr_underway++;
1792	}
1793	bh = next;
1794	} while (bh != head);
1795	unlock_page(page);
1796
1797	err = 0;
1798	done:
1799	if (nr_underway == 0) {
1800	/*
1801	* The page was marked dirty, but the buffers were
1802	* clean. Someone wrote them back by hand with
1803	* ll_rw_block/submit_bh. A rare case.
1804	*/
1805	end_page_writeback(page);
1806
1807	/*
1808	* The page and buffer_heads can be released at any time from
1809	* here on.
1810	*/
1811	}
1812	return err;
1813
1814	recover:
1815	/*
1816	* ENOSPC, or some other error. We may already have added some
1817	* blocks to the file, so we need to write these out to avoid
1818	* exposing stale data.
1819	* The page is currently locked and not marked for writeback
1820	*/
1821	bh = head;
1822	/* Recovery: lock and submit the mapped buffers */
1823	do {
1824	if (buffer_mapped(bh) && buffer_dirty(bh) &&
1825	!buffer_delay(bh)) {
1826	lock_buffer(bh);
1827	mark_buffer_async_write_endio(bh, handler);
1828	} else {
1829	/*
1830	* The buffer may have been set dirty during
1831	* attachment to a dirty page.
1832	*/
1833	clear_buffer_dirty(bh);
1834	}
1835	} while ((bh = bh->b_this_page) != head);
1836	SetPageError(page);
1837	BUG_ON(PageWriteback(page));
1838	mapping_set_error(page->mapping, err);
1839	set_page_writeback(page);
1840	do {
1841	struct buffer_head *next = bh->b_this_page;
1842	if (buffer_async_write(bh)) {
1843	clear_buffer_dirty(bh);
1844	submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1845	nr_underway++;
1846	}
1847	bh = next;
1848	} while (bh != head);
1849	unlock_page(page);
1850	goto done;
1851	}
1852	EXPORT_SYMBOL(__block_write_full_page);
1853
1854	/*
1855	* If a page has any new buffers, zero them out here, and mark them uptodate
1856	* and dirty so they'll be written out (in order to prevent uninitialised
1857	* block data from leaking). And clear the new bit.
1858	*/
1859	void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1860	{
1861	unsigned int block_start, block_end;
1862	struct buffer_head *head, *bh;
1863
1864	BUG_ON(!PageLocked(page));
1865	if (!page_has_buffers(page))
1866	return;
1867
1868	bh = head = page_buffers(page);
1869	block_start = 0;
1870	do {
1871	block_end = block_start + bh->b_size;
1872
1873	if (buffer_new(bh)) {
1874	if (block_end > from && block_start < to) {
1875	if (!PageUptodate(page)) {
1876	unsigned start, size;
1877
1878	start = max(from, block_start);
1879	size = min(to, block_end) - start;
1880
1881	zero_user(page, start, size);
1882	set_buffer_uptodate(bh);
1883	}
1884
1885	clear_buffer_new(bh);
1886	mark_buffer_dirty(bh);
1887	}
1888	}
1889
1890	block_start = block_end;
1891	bh = bh->b_this_page;
1892	} while (bh != head);
1893	}
1894	EXPORT_SYMBOL(page_zero_new_buffers);
1895
1896	static void
1897	iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1898	struct iomap *iomap)
1899	{
1900	loff_t offset = block << inode->i_blkbits;
1901
1902	bh->b_bdev = iomap->bdev;
1903
1904	/*
1905	* Block points to offset in file we need to map, iomap contains
1906	* the offset at which the map starts. If the map ends before the
1907	* current block, then do not map the buffer and let the caller
1908	* handle it.
1909	*/
1910	BUG_ON(offset >= iomap->offset + iomap->length);
1911
1912	switch (iomap->type) {
1913	case IOMAP_HOLE:
1914	/*
1915	* If the buffer is not up to date or beyond the current EOF,
1916	* we need to mark it as new to ensure sub-block zeroing is
1917	* executed if necessary.
1918	*/
1919	if (!buffer_uptodate(bh) ||
1920	(offset >= i_size_read(inode)))
1921	set_buffer_new(bh);
1922	break;
1923	case IOMAP_DELALLOC:
1924	if (!buffer_uptodate(bh) ||
1925	(offset >= i_size_read(inode)))
1926	set_buffer_new(bh);
1927	set_buffer_uptodate(bh);
1928	set_buffer_mapped(bh);
1929	set_buffer_delay(bh);
1930	break;
1931	case IOMAP_UNWRITTEN:
1932	/*
1933	* For unwritten regions, we always need to ensure that
1934	* sub-block writes cause the regions in the block we are not
1935	* writing to are zeroed. Set the buffer as new to ensure this.
1936	*/
1937	set_buffer_new(bh);
1938	set_buffer_unwritten(bh);
1939	/* FALLTHRU */
1940	case IOMAP_MAPPED:
1941	if (offset >= i_size_read(inode))
1942	set_buffer_new(bh);
1943	bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1944	((offset - iomap->offset) >> inode->i_blkbits);
1945	set_buffer_mapped(bh);
1946	break;
1947	}
1948	}
1949
1950	int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1951	get_block_t *get_block, struct iomap *iomap)
1952	{
1953	unsigned from = pos & (PAGE_SIZE - 1);
1954	unsigned to = from + len;
1955	struct inode *inode = page->mapping->host;
1956	unsigned block_start, block_end;
1957	sector_t block;
1958	int err = 0;
1959	unsigned blocksize, bbits;
1960	struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1961
1962	BUG_ON(!PageLocked(page));
1963	BUG_ON(from > PAGE_SIZE);
1964	BUG_ON(to > PAGE_SIZE);
1965	BUG_ON(from > to);
1966
1967	head = create_page_buffers(page, inode, 0);
1968	blocksize = head->b_size;
1969	bbits = block_size_bits(blocksize);
1970
1971	block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1972
1973	for(bh = head, block_start = 0; bh != head || !block_start;
1974	block++, block_start=block_end, bh = bh->b_this_page) {
1975	block_end = block_start + blocksize;
1976	if (block_end <= from || block_start >= to) {
1977	if (PageUptodate(page)) {
1978	if (!buffer_uptodate(bh))
1979	set_buffer_uptodate(bh);
1980	}
1981	continue;
1982	}
1983	if (buffer_new(bh))
1984	clear_buffer_new(bh);
1985	if (!buffer_mapped(bh)) {
1986	WARN_ON(bh->b_size != blocksize);
1987	if (get_block) {
1988	err = get_block(inode, block, bh, 1);
1989	if (err)
1990	break;
1991	} else {
1992	iomap_to_bh(inode, block, bh, iomap);
1993	}
1994
1995	if (buffer_new(bh)) {
1996	unmap_underlying_metadata(bh->b_bdev,
1997	bh->b_blocknr);
1998	if (PageUptodate(page)) {
1999	clear_buffer_new(bh);
2000	set_buffer_uptodate(bh);
2001	mark_buffer_dirty(bh);
2002	continue;
2003	}
2004	if (block_end > to || block_start < from)
2005	zero_user_segments(page,
2006	to, block_end,
2007	block_start, from);
2008	continue;
2009	}
2010	}
2011	if (PageUptodate(page)) {
2012	if (!buffer_uptodate(bh))
2013	set_buffer_uptodate(bh);
2014	continue;
2015	}
2016	if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2017	!buffer_unwritten(bh) &&
2018	(block_start < from || block_end > to)) {
2019	ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2020	*wait_bh++=bh;
2021	}
2022	}
2023	/*
2024	* If we issued read requests - let them complete.
2025	*/
2026	while(wait_bh > wait) {
2027	wait_on_buffer(*--wait_bh);
2028	if (!buffer_uptodate(*wait_bh))
2029	err = -EIO;
2030	}
2031	if (unlikely(err))
2032	page_zero_new_buffers(page, from, to);
2033	return err;
2034	}
2035
2036	int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2037	get_block_t *get_block)
2038	{
2039	return __block_write_begin_int(page, pos, len, get_block, NULL);
2040	}
2041	EXPORT_SYMBOL(__block_write_begin);
2042
2043	static int __block_commit_write(struct inode *inode, struct page *page,
2044	unsigned from, unsigned to)
2045	{
2046	unsigned block_start, block_end;
2047	int partial = 0;
2048	unsigned blocksize;
2049	struct buffer_head *bh, *head;
2050
2051	bh = head = page_buffers(page);
2052	blocksize = bh->b_size;
2053
2054	block_start = 0;
2055	do {
2056	block_end = block_start + blocksize;
2057	if (block_end <= from || block_start >= to) {
2058	if (!buffer_uptodate(bh))
2059	partial = 1;
2060	} else {
2061	set_buffer_uptodate(bh);
2062	mark_buffer_dirty(bh);
2063	}
2064	clear_buffer_new(bh);
2065
2066	block_start = block_end;
2067	bh = bh->b_this_page;
2068	} while (bh != head);
2069
2070	/*
2071	* If this is a partial write which happened to make all buffers
2072	* uptodate then we can optimize away a bogus readpage() for
2073	* the next read(). Here we 'discover' whether the page went
2074	* uptodate as a result of this (potentially partial) write.
2075	*/
2076	if (!partial)
2077	SetPageUptodate(page);
2078	return 0;
2079	}
2080
2081	/*
2082	* block_write_begin takes care of the basic task of block allocation and
2083	* bringing partial write blocks uptodate first.
2084	*
2085	* The filesystem needs to handle block truncation upon failure.
2086	*/
2087	int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2088	unsigned flags, struct page **pagep, get_block_t *get_block)
2089	{
2090	pgoff_t index = pos >> PAGE_SHIFT;
2091	struct page *page;
2092	int status;
2093
2094	page = grab_cache_page_write_begin(mapping, index, flags);
2095	if (!page)
2096	return -ENOMEM;
2097
2098	status = __block_write_begin(page, pos, len, get_block);
2099	if (unlikely(status)) {
2100	unlock_page(page);
2101	put_page(page);
2102	page = NULL;
2103	}
2104
2105	*pagep = page;
2106	return status;
2107	}
2108	EXPORT_SYMBOL(block_write_begin);
2109
2110	int block_write_end(struct file *file, struct address_space *mapping,
2111	loff_t pos, unsigned len, unsigned copied,
2112	struct page *page, void *fsdata)
2113	{
2114	struct inode *inode = mapping->host;
2115	unsigned start;
2116
2117	start = pos & (PAGE_SIZE - 1);
2118
2119	if (unlikely(copied < len)) {
2120	/*
2121	* The buffers that were written will now be uptodate, so we
2122	* don't have to worry about a readpage reading them and
2123	* overwriting a partial write. However if we have encountered
2124	* a short write and only partially written into a buffer, it
2125	* will not be marked uptodate, so a readpage might come in and
2126	* destroy our partial write.
2127	*
2128	* Do the simplest thing, and just treat any short write to a
2129	* non uptodate page as a zero-length write, and force the
2130	* caller to redo the whole thing.
2131	*/
2132	if (!PageUptodate(page))
2133	copied = 0;
2134
2135	page_zero_new_buffers(page, start+copied, start+len);
2136	}
2137	flush_dcache_page(page);
2138
2139	/* This could be a short (even 0-length) commit */
2140	__block_commit_write(inode, page, start, start+copied);
2141
2142	return copied;
2143	}
2144	EXPORT_SYMBOL(block_write_end);
2145
2146	int generic_write_end(struct file *file, struct address_space *mapping,
2147	loff_t pos, unsigned len, unsigned copied,
2148	struct page *page, void *fsdata)
2149	{
2150	struct inode *inode = mapping->host;
2151	loff_t old_size = inode->i_size;
2152	int i_size_changed = 0;
2153
2154	copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2155
2156	/*
2157	* No need to use i_size_read() here, the i_size
2158	* cannot change under us because we hold i_mutex.
2159	*
2160	* But it's important to update i_size while still holding page lock:
2161	* page writeout could otherwise come in and zero beyond i_size.
2162	*/
2163	if (pos+copied > inode->i_size) {
2164	i_size_write(inode, pos+copied);
2165	i_size_changed = 1;
2166	}
2167
2168	unlock_page(page);
2169	put_page(page);
2170
2171	if (old_size < pos)
2172	pagecache_isize_extended(inode, old_size, pos);
2173	/*
2174	* Don't mark the inode dirty under page lock. First, it unnecessarily
2175	* makes the holding time of page lock longer. Second, it forces lock
2176	* ordering of page lock and transaction start for journaling
2177	* filesystems.
2178	*/
2179	if (i_size_changed)
2180	mark_inode_dirty(inode);
2181
2182	return copied;
2183	}
2184	EXPORT_SYMBOL(generic_write_end);
2185
2186	/*
2187	* block_is_partially_uptodate checks whether buffers within a page are
2188	* uptodate or not.
2189	*
2190	* Returns true if all buffers which correspond to a file portion
2191	* we want to read are uptodate.
2192	*/
2193	int block_is_partially_uptodate(struct page *page, unsigned long from,
2194	unsigned long count)
2195	{
2196	unsigned block_start, block_end, blocksize;
2197	unsigned to;
2198	struct buffer_head *bh, *head;
2199	int ret = 1;
2200
2201	if (!page_has_buffers(page))
2202	return 0;
2203
2204	head = page_buffers(page);
2205	blocksize = head->b_size;
2206	to = min_t(unsigned, PAGE_SIZE - from, count);
2207	to = from + to;
2208	if (from < blocksize && to > PAGE_SIZE - blocksize)
2209	return 0;
2210
2211	bh = head;
2212	block_start = 0;
2213	do {
2214	block_end = block_start + blocksize;
2215	if (block_end > from && block_start < to) {
2216	if (!buffer_uptodate(bh)) {
2217	ret = 0;
2218	break;
2219	}
2220	if (block_end >= to)
2221	break;
2222	}
2223	block_start = block_end;
2224	bh = bh->b_this_page;
2225	} while (bh != head);
2226
2227	return ret;
2228	}
2229	EXPORT_SYMBOL(block_is_partially_uptodate);
2230
2231	/*
2232	* Generic "read page" function for block devices that have the normal
2233	* get_block functionality. This is most of the block device filesystems.
2234	* Reads the page asynchronously --- the unlock_buffer() and
2235	* set/clear_buffer_uptodate() functions propagate buffer state into the
2236	* page struct once IO has completed.
2237	*/
2238	int block_read_full_page(struct page *page, get_block_t *get_block)
2239	{
2240	struct inode *inode = page->mapping->host;
2241	sector_t iblock, lblock;
2242	struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2243	unsigned int blocksize, bbits;
2244	int nr, i;
2245	int fully_mapped = 1;
2246
2247	head = create_page_buffers(page, inode, 0);
2248	blocksize = head->b_size;
2249	bbits = block_size_bits(blocksize);
2250
2251	iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2252	lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2253	bh = head;
2254	nr = 0;
2255	i = 0;
2256
2257	do {
2258	if (buffer_uptodate(bh))
2259	continue;
2260
2261	if (!buffer_mapped(bh)) {
2262	int err = 0;
2263
2264	fully_mapped = 0;
2265	if (iblock < lblock) {
2266	WARN_ON(bh->b_size != blocksize);
2267	err = get_block(inode, iblock, bh, 0);
2268	if (err)
2269	SetPageError(page);
2270	}
2271	if (!buffer_mapped(bh)) {
2272	zero_user(page, i * blocksize, blocksize);
2273	if (!err)
2274	set_buffer_uptodate(bh);
2275	continue;
2276	}
2277	/*
2278	* get_block() might have updated the buffer
2279	* synchronously
2280	*/
2281	if (buffer_uptodate(bh))
2282	continue;
2283	}
2284	arr[nr++] = bh;
2285	} while (i++, iblock++, (bh = bh->b_this_page) != head);
2286
2287	if (fully_mapped)
2288	SetPageMappedToDisk(page);
2289
2290	if (!nr) {
2291	/*
2292	* All buffers are uptodate - we can set the page uptodate
2293	* as well. But not if get_block() returned an error.
2294	*/
2295	if (!PageError(page))
2296	SetPageUptodate(page);
2297	unlock_page(page);
2298	return 0;
2299	}
2300
2301	/* Stage two: lock the buffers */
2302	for (i = 0; i < nr; i++) {
2303	bh = arr[i];
2304	lock_buffer(bh);
2305	mark_buffer_async_read(bh);
2306	}
2307
2308	/*
2309	* Stage 3: start the IO. Check for uptodateness
2310	* inside the buffer lock in case another process reading
2311	* the underlying blockdev brought it uptodate (the sct fix).
2312	*/
2313	for (i = 0; i < nr; i++) {
2314	bh = arr[i];
2315	if (buffer_uptodate(bh))
2316	end_buffer_async_read(bh, 1);
2317	else
2318	submit_bh(REQ_OP_READ, 0, bh);
2319	}
2320	return 0;
2321	}
2322	EXPORT_SYMBOL(block_read_full_page);
2323
2324	/* utility function for filesystems that need to do work on expanding
2325	* truncates. Uses filesystem pagecache writes to allow the filesystem to
2326	* deal with the hole.
2327	*/
2328	int generic_cont_expand_simple(struct inode *inode, loff_t size)
2329	{
2330	struct address_space *mapping = inode->i_mapping;
2331	struct page *page;
2332	void *fsdata;
2333	int err;
2334
2335	err = inode_newsize_ok(inode, size);
2336	if (err)
2337	goto out;
2338
2339	err = pagecache_write_begin(NULL, mapping, size, 0,
2340	AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2341	&page, &fsdata);
2342	if (err)
2343	goto out;
2344
2345	err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2346	BUG_ON(err > 0);
2347
2348	out:
2349	return err;
2350	}
2351	EXPORT_SYMBOL(generic_cont_expand_simple);
2352
2353	static int cont_expand_zero(struct file *file, struct address_space *mapping,
2354	loff_t pos, loff_t *bytes)
2355	{
2356	struct inode *inode = mapping->host;
2357	unsigned int blocksize = i_blocksize(inode);
2358	struct page *page;
2359	void *fsdata;
2360	pgoff_t index, curidx;
2361	loff_t curpos;
2362	unsigned zerofrom, offset, len;
2363	int err = 0;
2364
2365	index = pos >> PAGE_SHIFT;
2366	offset = pos & ~PAGE_MASK;
2367
2368	while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2369	zerofrom = curpos & ~PAGE_MASK;
2370	if (zerofrom & (blocksize-1)) {
2371	*bytes |= (blocksize-1);
2372	(*bytes)++;
2373	}
2374	len = PAGE_SIZE - zerofrom;
2375
2376	err = pagecache_write_begin(file, mapping, curpos, len,
2377	AOP_FLAG_UNINTERRUPTIBLE,
2378	&page, &fsdata);
2379	if (err)
2380	goto out;
2381	zero_user(page, zerofrom, len);
2382	err = pagecache_write_end(file, mapping, curpos, len, len,
2383	page, fsdata);
2384	if (err < 0)
2385	goto out;
2386	BUG_ON(err != len);
2387	err = 0;
2388
2389	balance_dirty_pages_ratelimited(mapping);
2390
2391	if (unlikely(fatal_signal_pending(current))) {
2392	err = -EINTR;
2393	goto out;
2394	}
2395	}
2396
2397	/* page covers the boundary, find the boundary offset */
2398	if (index == curidx) {
2399	zerofrom = curpos & ~PAGE_MASK;
2400	/* if we will expand the thing last block will be filled */
2401	if (offset <= zerofrom) {
2402	goto out;
2403	}
2404	if (zerofrom & (blocksize-1)) {
2405	*bytes |= (blocksize-1);
2406	(*bytes)++;
2407	}
2408	len = offset - zerofrom;
2409
2410	err = pagecache_write_begin(file, mapping, curpos, len,
2411	AOP_FLAG_UNINTERRUPTIBLE,
2412	&page, &fsdata);
2413	if (err)
2414	goto out;
2415	zero_user(page, zerofrom, len);
2416	err = pagecache_write_end(file, mapping, curpos, len, len,
2417	page, fsdata);
2418	if (err < 0)
2419	goto out;
2420	BUG_ON(err != len);
2421	err = 0;
2422	}
2423	out:
2424	return err;
2425	}
2426
2427	/*
2428	* For moronic filesystems that do not allow holes in file.
2429	* We may have to extend the file.
2430	*/
2431	int cont_write_begin(struct file *file, struct address_space *mapping,
2432	loff_t pos, unsigned len, unsigned flags,
2433	struct page **pagep, void **fsdata,
2434	get_block_t *get_block, loff_t *bytes)
2435	{
2436	struct inode *inode = mapping->host;
2437	unsigned int blocksize = i_blocksize(inode);
2438	unsigned int zerofrom;
2439	int err;
2440
2441	err = cont_expand_zero(file, mapping, pos, bytes);
2442	if (err)
2443	return err;
2444
2445	zerofrom = *bytes & ~PAGE_MASK;
2446	if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2447	*bytes |= (blocksize-1);
2448	(*bytes)++;
2449	}
2450
2451	return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2452	}
2453	EXPORT_SYMBOL(cont_write_begin);
2454
2455	int block_commit_write(struct page *page, unsigned from, unsigned to)
2456	{
2457	struct inode *inode = page->mapping->host;
2458	__block_commit_write(inode,page,from,to);
2459	return 0;
2460	}
2461	EXPORT_SYMBOL(block_commit_write);
2462
2463	/*
2464	* block_page_mkwrite() is not allowed to change the file size as it gets
2465	* called from a page fault handler when a page is first dirtied. Hence we must
2466	* be careful to check for EOF conditions here. We set the page up correctly
2467	* for a written page which means we get ENOSPC checking when writing into
2468	* holes and correct delalloc and unwritten extent mapping on filesystems that
2469	* support these features.
2470	*
2471	* We are not allowed to take the i_mutex here so we have to play games to
2472	* protect against truncate races as the page could now be beyond EOF. Because
2473	* truncate writes the inode size before removing pages, once we have the
2474	* page lock we can determine safely if the page is beyond EOF. If it is not
2475	* beyond EOF, then the page is guaranteed safe against truncation until we
2476	* unlock the page.
2477	*
2478	* Direct callers of this function should protect against filesystem freezing
2479	* using sb_start_pagefault() - sb_end_pagefault() functions.
2480	*/
2481	int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2482	get_block_t get_block)
2483	{
2484	struct page *page = vmf->page;
2485	struct inode *inode = file_inode(vma->vm_file);
2486	unsigned long end;
2487	loff_t size;
2488	int ret;
2489
2490	lock_page(page);
2491	size = i_size_read(inode);
2492	if ((page->mapping != inode->i_mapping) ||
2493	(page_offset(page) > size)) {
2494	/* We overload EFAULT to mean page got truncated */
2495	ret = -EFAULT;
2496	goto out_unlock;
2497	}
2498
2499	/* page is wholly or partially inside EOF */
2500	if (((page->index + 1) << PAGE_SHIFT) > size)
2501	end = size & ~PAGE_MASK;
2502	else
2503	end = PAGE_SIZE;
2504
2505	ret = __block_write_begin(page, 0, end, get_block);
2506	if (!ret)
2507	ret = block_commit_write(page, 0, end);
2508
2509	if (unlikely(ret < 0))
2510	goto out_unlock;
2511	set_page_dirty(page);
2512	wait_for_stable_page(page);
2513	return 0;
2514	out_unlock:
2515	unlock_page(page);
2516	return ret;
2517	}
2518	EXPORT_SYMBOL(block_page_mkwrite);
2519
2520	/*
2521	* nobh_write_begin()'s prereads are special: the buffer_heads are freed
2522	* immediately, while under the page lock. So it needs a special end_io
2523	* handler which does not touch the bh after unlocking it.
2524	*/
2525	static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2526	{
2527	__end_buffer_read_notouch(bh, uptodate);
2528	}
2529
2530	/*
2531	* Attach the singly-linked list of buffers created by nobh_write_begin, to
2532	* the page (converting it to circular linked list and taking care of page
2533	* dirty races).
2534	*/
2535	static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2536	{
2537	struct buffer_head *bh;
2538
2539	BUG_ON(!PageLocked(page));
2540
2541	spin_lock(&page->mapping->private_lock);
2542	bh = head;
2543	do {
2544	if (PageDirty(page))
2545	set_buffer_dirty(bh);
2546	if (!bh->b_this_page)
2547	bh->b_this_page = head;
2548	bh = bh->b_this_page;
2549	} while (bh != head);
2550	attach_page_buffers(page, head);
2551	spin_unlock(&page->mapping->private_lock);
2552	}
2553
2554	/*
2555	* On entry, the page is fully not uptodate.
2556	* On exit the page is fully uptodate in the areas outside (from,to)
2557	* The filesystem needs to handle block truncation upon failure.
2558	*/
2559	int nobh_write_begin(struct address_space *mapping,
2560	loff_t pos, unsigned len, unsigned flags,
2561	struct page **pagep, void **fsdata,
2562	get_block_t *get_block)
2563	{
2564	struct inode *inode = mapping->host;
2565	const unsigned blkbits = inode->i_blkbits;
2566	const unsigned blocksize = 1 << blkbits;
2567	struct buffer_head *head, *bh;
2568	struct page *page;
2569	pgoff_t index;
2570	unsigned from, to;
2571	unsigned block_in_page;
2572	unsigned block_start, block_end;
2573	sector_t block_in_file;
2574	int nr_reads = 0;
2575	int ret = 0;
2576	int is_mapped_to_disk = 1;
2577
2578	index = pos >> PAGE_SHIFT;
2579	from = pos & (PAGE_SIZE - 1);
2580	to = from + len;
2581
2582	page = grab_cache_page_write_begin(mapping, index, flags);
2583	if (!page)
2584	return -ENOMEM;
2585	*pagep = page;
2586	*fsdata = NULL;
2587
2588	if (page_has_buffers(page)) {
2589	ret = __block_write_begin(page, pos, len, get_block);
2590	if (unlikely(ret))
2591	goto out_release;
2592	return ret;
2593	}
2594
2595	if (PageMappedToDisk(page))
2596	return 0;
2597
2598	/*
2599	* Allocate buffers so that we can keep track of state, and potentially
2600	* attach them to the page if an error occurs. In the common case of
2601	* no error, they will just be freed again without ever being attached
2602	* to the page (which is all OK, because we're under the page lock).
2603	*
2604	* Be careful: the buffer linked list is a NULL terminated one, rather
2605	* than the circular one we're used to.
2606	*/
2607	head = alloc_page_buffers(page, blocksize, 0);
2608	if (!head) {
2609	ret = -ENOMEM;
2610	goto out_release;
2611	}
2612
2613	block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2614
2615	/*
2616	* We loop across all blocks in the page, whether or not they are
2617	* part of the affected region. This is so we can discover if the
2618	* page is fully mapped-to-disk.
2619	*/
2620	for (block_start = 0, block_in_page = 0, bh = head;
2621	block_start < PAGE_SIZE;
2622	block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2623	int create;
2624
2625	block_end = block_start + blocksize;
2626	bh->b_state = 0;
2627	create = 1;
2628	if (block_start >= to)
2629	create = 0;
2630	ret = get_block(inode, block_in_file + block_in_page,
2631	bh, create);
2632	if (ret)
2633	goto failed;
2634	if (!buffer_mapped(bh))
2635	is_mapped_to_disk = 0;
2636	if (buffer_new(bh))
2637	unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2638	if (PageUptodate(page)) {
2639	set_buffer_uptodate(bh);
2640	continue;
2641	}
2642	if (buffer_new(bh) || !buffer_mapped(bh)) {
2643	zero_user_segments(page, block_start, from,
2644	to, block_end);
2645	continue;
2646	}
2647	if (buffer_uptodate(bh))
2648	continue; /* reiserfs does this */
2649	if (block_start < from || block_end > to) {
2650	lock_buffer(bh);
2651	bh->b_end_io = end_buffer_read_nobh;
2652	submit_bh(REQ_OP_READ, 0, bh);
2653	nr_reads++;
2654	}
2655	}
2656
2657	if (nr_reads) {
2658	/*
2659	* The page is locked, so these buffers are protected from
2660	* any VM or truncate activity. Hence we don't need to care
2661	* for the buffer_head refcounts.
2662	*/
2663	for (bh = head; bh; bh = bh->b_this_page) {
2664	wait_on_buffer(bh);
2665	if (!buffer_uptodate(bh))
2666	ret = -EIO;
2667	}
2668	if (ret)
2669	goto failed;
2670	}
2671
2672	if (is_mapped_to_disk)
2673	SetPageMappedToDisk(page);
2674
2675	*fsdata = head; /* to be released by nobh_write_end */
2676
2677	return 0;
2678
2679	failed:
2680	BUG_ON(!ret);
2681	/*
2682	* Error recovery is a bit difficult. We need to zero out blocks that
2683	* were newly allocated, and dirty them to ensure they get written out.
2684	* Buffers need to be attached to the page at this point, otherwise
2685	* the handling of potential IO errors during writeout would be hard
2686	* (could try doing synchronous writeout, but what if that fails too?)
2687	*/
2688	attach_nobh_buffers(page, head);
2689	page_zero_new_buffers(page, from, to);
2690
2691	out_release:
2692	unlock_page(page);
2693	put_page(page);
2694	*pagep = NULL;
2695
2696	return ret;
2697	}
2698	EXPORT_SYMBOL(nobh_write_begin);
2699
2700	int nobh_write_end(struct file *file, struct address_space *mapping,
2701	loff_t pos, unsigned len, unsigned copied,
2702	struct page *page, void *fsdata)
2703	{
2704	struct inode *inode = page->mapping->host;
2705	struct buffer_head *head = fsdata;
2706	struct buffer_head *bh;
2707	BUG_ON(fsdata != NULL && page_has_buffers(page));
2708
2709	if (unlikely(copied < len) && head)
2710	attach_nobh_buffers(page, head);
2711	if (page_has_buffers(page))
2712	return generic_write_end(file, mapping, pos, len,
2713	copied, page, fsdata);
2714
2715	SetPageUptodate(page);
2716	set_page_dirty(page);
2717	if (pos+copied > inode->i_size) {
2718	i_size_write(inode, pos+copied);
2719	mark_inode_dirty(inode);
2720	}
2721
2722	unlock_page(page);
2723	put_page(page);
2724
2725	while (head) {
2726	bh = head;
2727	head = head->b_this_page;
2728	free_buffer_head(bh);
2729	}
2730
2731	return copied;
2732	}
2733	EXPORT_SYMBOL(nobh_write_end);
2734
2735	/*
2736	* nobh_writepage() - based on block_full_write_page() except
2737	* that it tries to operate without attaching bufferheads to
2738	* the page.
2739	*/
2740	int nobh_writepage(struct page *page, get_block_t *get_block,
2741	struct writeback_control *wbc)
2742	{
2743	struct inode * const inode = page->mapping->host;
2744	loff_t i_size = i_size_read(inode);
2745	const pgoff_t end_index = i_size >> PAGE_SHIFT;
2746	unsigned offset;
2747	int ret;
2748
2749	/* Is the page fully inside i_size? */
2750	if (page->index < end_index)
2751	goto out;
2752
2753	/* Is the page fully outside i_size? (truncate in progress) */
2754	offset = i_size & (PAGE_SIZE-1);
2755	if (page->index >= end_index+1 || !offset) {
2756	/*
2757	* The page may have dirty, unmapped buffers. For example,
2758	* they may have been added in ext3_writepage(). Make them
2759	* freeable here, so the page does not leak.
2760	*/
2761	#if 0
2762	/* Not really sure about this - do we need this ? */
2763	if (page->mapping->a_ops->invalidatepage)
2764	page->mapping->a_ops->invalidatepage(page, offset);
2765	#endif
2766	unlock_page(page);
2767	return 0; /* don't care */
2768	}
2769
2770	/*
2771	* The page straddles i_size. It must be zeroed out on each and every
2772	* writepage invocation because it may be mmapped. "A file is mapped
2773	* in multiples of the page size. For a file that is not a multiple of
2774	* the page size, the remaining memory is zeroed when mapped, and
2775	* writes to that region are not written out to the file."
2776	*/
2777	zero_user_segment(page, offset, PAGE_SIZE);
2778	out:
2779	ret = mpage_writepage(page, get_block, wbc);
2780	if (ret == -EAGAIN)
2781	ret = __block_write_full_page(inode, page, get_block, wbc,
2782	end_buffer_async_write);
2783	return ret;
2784	}
2785	EXPORT_SYMBOL(nobh_writepage);
2786
2787	int nobh_truncate_page(struct address_space *mapping,
2788	loff_t from, get_block_t *get_block)
2789	{
2790	pgoff_t index = from >> PAGE_SHIFT;
2791	unsigned offset = from & (PAGE_SIZE-1);
2792	unsigned blocksize;
2793	sector_t iblock;
2794	unsigned length, pos;
2795	struct inode *inode = mapping->host;
2796	struct page *page;
2797	struct buffer_head map_bh;
2798	int err;
2799
2800	blocksize = i_blocksize(inode);
2801	length = offset & (blocksize - 1);
2802
2803	/* Block boundary? Nothing to do */
2804	if (!length)
2805	return 0;
2806
2807	length = blocksize - length;
2808	iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2809
2810	page = grab_cache_page(mapping, index);
2811	err = -ENOMEM;
2812	if (!page)
2813	goto out;
2814
2815	if (page_has_buffers(page)) {
2816	has_buffers:
2817	unlock_page(page);
2818	put_page(page);
2819	return block_truncate_page(mapping, from, get_block);
2820	}
2821
2822	/* Find the buffer that contains "offset" */
2823	pos = blocksize;
2824	while (offset >= pos) {
2825	iblock++;
2826	pos += blocksize;
2827	}
2828
2829	map_bh.b_size = blocksize;
2830	map_bh.b_state = 0;
2831	err = get_block(inode, iblock, &map_bh, 0);
2832	if (err)
2833	goto unlock;
2834	/* unmapped? It's a hole - nothing to do */
2835	if (!buffer_mapped(&map_bh))
2836	goto unlock;
2837
2838	/* Ok, it's mapped. Make sure it's up-to-date */
2839	if (!PageUptodate(page)) {
2840	err = mapping->a_ops->readpage(NULL, page);
2841	if (err) {
2842	put_page(page);
2843	goto out;
2844	}
2845	lock_page(page);
2846	if (!PageUptodate(page)) {
2847	err = -EIO;
2848	goto unlock;
2849	}
2850	if (page_has_buffers(page))
2851	goto has_buffers;
2852	}
2853	zero_user(page, offset, length);
2854	set_page_dirty(page);
2855	err = 0;
2856
2857	unlock:
2858	unlock_page(page);
2859	put_page(page);
2860	out:
2861	return err;
2862	}
2863	EXPORT_SYMBOL(nobh_truncate_page);
2864
2865	int block_truncate_page(struct address_space *mapping,
2866	loff_t from, get_block_t *get_block)
2867	{
2868	pgoff_t index = from >> PAGE_SHIFT;
2869	unsigned offset = from & (PAGE_SIZE-1);
2870	unsigned blocksize;
2871	sector_t iblock;
2872	unsigned length, pos;
2873	struct inode *inode = mapping->host;
2874	struct page *page;
2875	struct buffer_head *bh;
2876	int err;
2877
2878	blocksize = i_blocksize(inode);
2879	length = offset & (blocksize - 1);
2880
2881	/* Block boundary? Nothing to do */
2882	if (!length)
2883	return 0;
2884
2885	length = blocksize - length;
2886	iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2887
2888	page = grab_cache_page(mapping, index);
2889	err = -ENOMEM;
2890	if (!page)
2891	goto out;
2892
2893	if (!page_has_buffers(page))
2894	create_empty_buffers(page, blocksize, 0);
2895
2896	/* Find the buffer that contains "offset" */
2897	bh = page_buffers(page);
2898	pos = blocksize;
2899	while (offset >= pos) {
2900	bh = bh->b_this_page;
2901	iblock++;
2902	pos += blocksize;
2903	}
2904
2905	err = 0;
2906	if (!buffer_mapped(bh)) {
2907	WARN_ON(bh->b_size != blocksize);
2908	err = get_block(inode, iblock, bh, 0);
2909	if (err)
2910	goto unlock;
2911	/* unmapped? It's a hole - nothing to do */
2912	if (!buffer_mapped(bh))
2913	goto unlock;
2914	}
2915
2916	/* Ok, it's mapped. Make sure it's up-to-date */
2917	if (PageUptodate(page))
2918	set_buffer_uptodate(bh);
2919
2920	if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2921	err = -EIO;
2922	ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2923	wait_on_buffer(bh);
2924	/* Uhhuh. Read error. Complain and punt. */
2925	if (!buffer_uptodate(bh))
2926	goto unlock;
2927	}
2928
2929	zero_user(page, offset, length);
2930	mark_buffer_dirty(bh);
2931	err = 0;
2932
2933	unlock:
2934	unlock_page(page);
2935	put_page(page);
2936	out:
2937	return err;
2938	}
2939	EXPORT_SYMBOL(block_truncate_page);
2940
2941	/*
2942	* The generic ->writepage function for buffer-backed address_spaces
2943	*/
2944	int block_write_full_page(struct page *page, get_block_t *get_block,
2945	struct writeback_control *wbc)
2946	{
2947	struct inode * const inode = page->mapping->host;
2948	loff_t i_size = i_size_read(inode);
2949	const pgoff_t end_index = i_size >> PAGE_SHIFT;
2950	unsigned offset;
2951
2952	/* Is the page fully inside i_size? */
2953	if (page->index < end_index)
2954	return __block_write_full_page(inode, page, get_block, wbc,
2955	end_buffer_async_write);
2956
2957	/* Is the page fully outside i_size? (truncate in progress) */
2958	offset = i_size & (PAGE_SIZE-1);
2959	if (page->index >= end_index+1 || !offset) {
2960	/*
2961	* The page may have dirty, unmapped buffers. For example,
2962	* they may have been added in ext3_writepage(). Make them
2963	* freeable here, so the page does not leak.
2964	*/
2965	do_invalidatepage(page, 0, PAGE_SIZE);
2966	unlock_page(page);
2967	return 0; /* don't care */
2968	}
2969
2970	/*
2971	* The page straddles i_size. It must be zeroed out on each and every
2972	* writepage invocation because it may be mmapped. "A file is mapped
2973	* in multiples of the page size. For a file that is not a multiple of
2974	* the page size, the remaining memory is zeroed when mapped, and
2975	* writes to that region are not written out to the file."
2976	*/
2977	zero_user_segment(page, offset, PAGE_SIZE);
2978	return __block_write_full_page(inode, page, get_block, wbc,
2979	end_buffer_async_write);
2980	}
2981	EXPORT_SYMBOL(block_write_full_page);
2982
2983	sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2984	get_block_t *get_block)
2985	{
2986	struct buffer_head tmp;
2987	struct inode *inode = mapping->host;
2988	tmp.b_state = 0;
2989	tmp.b_blocknr = 0;
2990	tmp.b_size = i_blocksize(inode);
2991	get_block(inode, block, &tmp, 0);
2992	return tmp.b_blocknr;
2993	}
2994	EXPORT_SYMBOL(generic_block_bmap);
2995
2996	static void end_bio_bh_io_sync(struct bio *bio)
2997	{
2998	struct buffer_head *bh = bio->bi_private;
2999
3000	if (unlikely(bio_flagged(bio, BIO_QUIET)))
3001	set_bit(BH_Quiet, &bh->b_state);
3002
3003	bh->b_end_io(bh, !bio->bi_error);
3004	bio_put(bio);
3005	}
3006
3007	/*
3008	* This allows us to do IO even on the odd last sectors
3009	* of a device, even if the block size is some multiple
3010	* of the physical sector size.
3011	*
3012	* We'll just truncate the bio to the size of the device,
3013	* and clear the end of the buffer head manually.
3014	*
3015	* Truly out-of-range accesses will turn into actual IO
3016	* errors, this only handles the "we need to be able to
3017	* do IO at the final sector" case.
3018	*/
3019	void guard_bio_eod(int op, struct bio *bio)
3020	{
3021	sector_t maxsector;
3022	struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3023	unsigned truncated_bytes;
3024
3025	maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3026	if (!maxsector)
3027	return;
3028
3029	/*
3030	* If the *whole* IO is past the end of the device,
3031	* let it through, and the IO layer will turn it into
3032	* an EIO.
3033	*/
3034	if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3035	return;
3036
3037	maxsector -= bio->bi_iter.bi_sector;
3038	if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3039	return;
3040
3041	/* Uhhuh. We've got a bio that straddles the device size! */
3042	truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3043
3044	/*
3045	* The bio contains more than one segment which spans EOD, just return
3046	* and let IO layer turn it into an EIO
3047	*/
3048	if (truncated_bytes > bvec->bv_len)
3049	return;
3050
3051	/* Truncate the bio.. */
3052	bio->bi_iter.bi_size -= truncated_bytes;
3053	bvec->bv_len -= truncated_bytes;
3054
3055	/* ..and clear the end of the buffer for reads */
3056	if (op == REQ_OP_READ) {
3057	zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3058	truncated_bytes);
3059	}
3060	}
3061
3062	static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3063	unsigned long bio_flags, struct writeback_control *wbc)
3064	{
3065	struct bio *bio;
3066
3067	BUG_ON(!buffer_locked(bh));
3068	BUG_ON(!buffer_mapped(bh));
3069	BUG_ON(!bh->b_end_io);
3070	BUG_ON(buffer_delay(bh));
3071	BUG_ON(buffer_unwritten(bh));
3072
3073	/*
3074	* Only clear out a write error when rewriting
3075	*/
3076	if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3077	clear_buffer_write_io_error(bh);
3078
3079	/*
3080	* from here on down, it's all bio -- do the initial mapping,
3081	* submit_bio -> generic_make_request may further map this bio around
3082	*/
3083	bio = bio_alloc(GFP_NOIO, 1);
3084
3085	if (wbc) {
3086	wbc_init_bio(wbc, bio);
3087	wbc_account_io(wbc, bh->b_page, bh->b_size);
3088	}
3089
3090	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3091	bio->bi_bdev = bh->b_bdev;
3092
3093	bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3094	BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3095
3096	bio->bi_end_io = end_bio_bh_io_sync;
3097	bio->bi_private = bh;
3098	bio->bi_flags |= bio_flags;
3099
3100	/* Take care of bh's that straddle the end of the device */
3101	guard_bio_eod(op, bio);
3102
3103	if (buffer_meta(bh))
3104	op_flags |= REQ_META;
3105	if (buffer_prio(bh))
3106	op_flags |= REQ_PRIO;
3107	bio_set_op_attrs(bio, op, op_flags);
3108
3109	submit_bio(bio);
3110	return 0;
3111	}
3112
3113	int _submit_bh(int op, int op_flags, struct buffer_head *bh,
3114	unsigned long bio_flags)
3115	{
3116	return submit_bh_wbc(op, op_flags, bh, bio_flags, NULL);
3117	}
3118	EXPORT_SYMBOL_GPL(_submit_bh);
3119
3120	int submit_bh(int op, int op_flags, struct buffer_head *bh)
3121	{
3122	return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3123	}
3124	EXPORT_SYMBOL(submit_bh);
3125
3126	/**
3127	* ll_rw_block: low-level access to block devices (DEPRECATED)
3128	* @op: whether to %READ or %WRITE
3129	* @op_flags: rq_flag_bits
3130	* @nr: number of &struct buffer_heads in the array
3131	* @bhs: array of pointers to &struct buffer_head
3132	*
3133	* ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3134	* requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3135	* @op_flags contains flags modifying the detailed I/O behavior, most notably
3136	* %REQ_RAHEAD.
3137	*
3138	* This function drops any buffer that it cannot get a lock on (with the
3139	* BH_Lock state bit), any buffer that appears to be clean when doing a write
3140	* request, and any buffer that appears to be up-to-date when doing read
3141	* request. Further it marks as clean buffers that are processed for
3142	* writing (the buffer cache won't assume that they are actually clean
3143	* until the buffer gets unlocked).
3144	*
3145	* ll_rw_block sets b_end_io to simple completion handler that marks
3146	* the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3147	* any waiters.
3148	*
3149	* All of the buffers must be for the same device, and must also be a
3150	* multiple of the current approved size for the device.
3151	*/
3152	void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3153	{
3154	int i;
3155
3156	for (i = 0; i < nr; i++) {
3157	struct buffer_head *bh = bhs[i];
3158
3159	if (!trylock_buffer(bh))
3160	continue;
3161	if (op == WRITE) {
3162	if (test_clear_buffer_dirty(bh)) {
3163	bh->b_end_io = end_buffer_write_sync;
3164	get_bh(bh);
3165	submit_bh(op, op_flags, bh);
3166	continue;
3167	}
3168	} else {
3169	if (!buffer_uptodate(bh)) {
3170	bh->b_end_io = end_buffer_read_sync;
3171	get_bh(bh);
3172	submit_bh(op, op_flags, bh);
3173	continue;
3174	}
3175	}
3176	unlock_buffer(bh);
3177	}
3178	}
3179	EXPORT_SYMBOL(ll_rw_block);
3180
3181	void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3182	{
3183	lock_buffer(bh);
3184	if (!test_clear_buffer_dirty(bh)) {
3185	unlock_buffer(bh);
3186	return;
3187	}
3188	bh->b_end_io = end_buffer_write_sync;
3189	get_bh(bh);
3190	submit_bh(REQ_OP_WRITE, op_flags, bh);
3191	}
3192	EXPORT_SYMBOL(write_dirty_buffer);
3193
3194	/*
3195	* For a data-integrity writeout, we need to wait upon any in-progress I/O
3196	* and then start new I/O and then wait upon it. The caller must have a ref on
3197	* the buffer_head.
3198	*/
3199	int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3200	{
3201	int ret = 0;
3202
3203	WARN_ON(atomic_read(&bh->b_count) < 1);
3204	lock_buffer(bh);
3205	if (test_clear_buffer_dirty(bh)) {
3206	get_bh(bh);
3207	bh->b_end_io = end_buffer_write_sync;
3208	ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3209	wait_on_buffer(bh);
3210	if (!ret && !buffer_uptodate(bh))
3211	ret = -EIO;
3212	} else {
3213	unlock_buffer(bh);
3214	}
3215	return ret;
3216	}
3217	EXPORT_SYMBOL(__sync_dirty_buffer);
3218
3219	int sync_dirty_buffer(struct buffer_head *bh)
3220	{
3221	return __sync_dirty_buffer(bh, WRITE_SYNC);
3222	}
3223	EXPORT_SYMBOL(sync_dirty_buffer);
3224
3225	/*
3226	* try_to_free_buffers() checks if all the buffers on this particular page
3227	* are unused, and releases them if so.
3228	*
3229	* Exclusion against try_to_free_buffers may be obtained by either
3230	* locking the page or by holding its mapping's private_lock.
3231	*
3232	* If the page is dirty but all the buffers are clean then we need to
3233	* be sure to mark the page clean as well. This is because the page
3234	* may be against a block device, and a later reattachment of buffers
3235	* to a dirty page will set *all* buffers dirty. Which would corrupt
3236	* filesystem data on the same device.
3237	*
3238	* The same applies to regular filesystem pages: if all the buffers are
3239	* clean then we set the page clean and proceed. To do that, we require
3240	* total exclusion from __set_page_dirty_buffers(). That is obtained with
3241	* private_lock.
3242	*
3243	* try_to_free_buffers() is non-blocking.
3244	*/
3245	static inline int buffer_busy(struct buffer_head *bh)
3246	{
3247	return atomic_read(&bh->b_count) |
3248	(bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3249	}
3250
3251	static int
3252	drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3253	{
3254	struct buffer_head *head = page_buffers(page);
3255	struct buffer_head *bh;
3256
3257	bh = head;
3258	do {
3259	if (buffer_write_io_error(bh) && page->mapping)
3260	mapping_set_error(page->mapping, -EIO);
3261	if (buffer_busy(bh))
3262	goto failed;
3263	bh = bh->b_this_page;
3264	} while (bh != head);
3265
3266	do {
3267	struct buffer_head *next = bh->b_this_page;
3268
3269	if (bh->b_assoc_map)
3270	__remove_assoc_queue(bh);
3271	bh = next;
3272	} while (bh != head);
3273	*buffers_to_free = head;
3274	__clear_page_buffers(page);
3275	return 1;
3276	failed:
3277	return 0;
3278	}
3279
3280	int try_to_free_buffers(struct page *page)
3281	{
3282	struct address_space * const mapping = page->mapping;
3283	struct buffer_head *buffers_to_free = NULL;
3284	int ret = 0;
3285
3286	BUG_ON(!PageLocked(page));
3287	if (PageWriteback(page))
3288	return 0;
3289
3290	if (mapping == NULL) { /* can this still happen? */
3291	ret = drop_buffers(page, &buffers_to_free);
3292	goto out;
3293	}
3294
3295	spin_lock(&mapping->private_lock);
3296	ret = drop_buffers(page, &buffers_to_free);
3297
3298	/*
3299	* If the filesystem writes its buffers by hand (eg ext3)
3300	* then we can have clean buffers against a dirty page. We
3301	* clean the page here; otherwise the VM will never notice
3302	* that the filesystem did any IO at all.
3303	*
3304	* Also, during truncate, discard_buffer will have marked all
3305	* the page's buffers clean. We discover that here and clean
3306	* the page also.
3307	*
3308	* private_lock must be held over this entire operation in order
3309	* to synchronise against __set_page_dirty_buffers and prevent the
3310	* dirty bit from being lost.
3311	*/
3312	if (ret)
3313	cancel_dirty_page(page);
3314	spin_unlock(&mapping->private_lock);
3315	out:
3316	if (buffers_to_free) {
3317	struct buffer_head *bh = buffers_to_free;
3318
3319	do {
3320	struct buffer_head *next = bh->b_this_page;
3321	free_buffer_head(bh);
3322	bh = next;
3323	} while (bh != buffers_to_free);
3324	}
3325	return ret;
3326	}
3327	EXPORT_SYMBOL(try_to_free_buffers);
3328
3329	/*
3330	* There are no bdflush tunables left. But distributions are
3331	* still running obsolete flush daemons, so we terminate them here.
3332	*
3333	* Use of bdflush() is deprecated and will be removed in a future kernel.
3334	* The `flush-X' kernel threads fully replace bdflush daemons and this call.
3335	*/
3336	SYSCALL_DEFINE2(bdflush, int, func, long, data)
3337	{
3338	static int msg_count;
3339
3340	if (!capable(CAP_SYS_ADMIN))
3341	return -EPERM;
3342
3343	if (msg_count < 5) {
3344	msg_count++;
3345	printk(KERN_INFO
3346	"warning: process `%s' used the obsolete bdflush"
3347	" system call\n", current->comm);
3348	printk(KERN_INFO "Fix your initscripts?\n");
3349	}
3350
3351	if (func == 1)
3352	do_exit(0);
3353	return 0;
3354	}
3355
3356	/*
3357	* Buffer-head allocation
3358	*/
3359	static struct kmem_cache *bh_cachep __read_mostly;
3360
3361	/*
3362	* Once the number of bh's in the machine exceeds this level, we start
3363	* stripping them in writeback.
3364	*/
3365	static unsigned long max_buffer_heads;
3366
3367	int buffer_heads_over_limit;
3368
3369	struct bh_accounting {
3370	int nr; /* Number of live bh's */
3371	int ratelimit; /* Limit cacheline bouncing */
3372	};
3373
3374	static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3375
3376	static void recalc_bh_state(void)
3377	{
3378	int i;
3379	int tot = 0;
3380
3381	if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3382	return;
3383	__this_cpu_write(bh_accounting.ratelimit, 0);
3384	for_each_online_cpu(i)
3385	tot += per_cpu(bh_accounting, i).nr;
3386	buffer_heads_over_limit = (tot > max_buffer_heads);
3387	}
3388
3389	struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3390	{
3391	struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3392	if (ret) {
3393	INIT_LIST_HEAD(&ret->b_assoc_buffers);
3394	preempt_disable();
3395	__this_cpu_inc(bh_accounting.nr);
3396	recalc_bh_state();
3397	preempt_enable();
3398	}
3399	return ret;
3400	}
3401	EXPORT_SYMBOL(alloc_buffer_head);
3402
3403	void free_buffer_head(struct buffer_head *bh)
3404	{
3405	BUG_ON(!list_empty(&bh->b_assoc_buffers));
3406	kmem_cache_free(bh_cachep, bh);
3407	preempt_disable();
3408	__this_cpu_dec(bh_accounting.nr);
3409	recalc_bh_state();
3410	preempt_enable();
3411	}
3412	EXPORT_SYMBOL(free_buffer_head);
3413
3414	static void buffer_exit_cpu(int cpu)
3415	{
3416	int i;
3417	struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3418
3419	for (i = 0; i < BH_LRU_SIZE; i++) {
3420	brelse(b->bhs[i]);
3421	b->bhs[i] = NULL;
3422	}
3423	this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3424	per_cpu(bh_accounting, cpu).nr = 0;
3425	}
3426
3427	static int buffer_cpu_notify(struct notifier_block *self,
3428	unsigned long action, void *hcpu)
3429	{
3430	if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3431	buffer_exit_cpu((unsigned long)hcpu);
3432	return NOTIFY_OK;
3433	}
3434
3435	/**
3436	* bh_uptodate_or_lock - Test whether the buffer is uptodate
3437	* @bh: struct buffer_head
3438	*
3439	* Return true if the buffer is up-to-date and false,
3440	* with the buffer locked, if not.
3441	*/
3442	int bh_uptodate_or_lock(struct buffer_head *bh)
3443	{
3444	if (!buffer_uptodate(bh)) {
3445	lock_buffer(bh);
3446	if (!buffer_uptodate(bh))
3447	return 0;
3448	unlock_buffer(bh);
3449	}
3450	return 1;
3451	}
3452	EXPORT_SYMBOL(bh_uptodate_or_lock);
3453
3454	/**
3455	* bh_submit_read - Submit a locked buffer for reading
3456	* @bh: struct buffer_head
3457	*
3458	* Returns zero on success and -EIO on error.
3459	*/
3460	int bh_submit_read(struct buffer_head *bh)
3461	{
3462	BUG_ON(!buffer_locked(bh));
3463
3464	if (buffer_uptodate(bh)) {
3465	unlock_buffer(bh);
3466	return 0;
3467	}
3468
3469	get_bh(bh);
3470	bh->b_end_io = end_buffer_read_sync;
3471	submit_bh(REQ_OP_READ, 0, bh);
3472	wait_on_buffer(bh);
3473	if (buffer_uptodate(bh))
3474	return 0;
3475	return -EIO;
3476	}
3477	EXPORT_SYMBOL(bh_submit_read);
3478
3479	void __init buffer_init(void)
3480	{
3481	unsigned long nrpages;
3482
3483	bh_cachep = kmem_cache_create("buffer_head",
3484	sizeof(struct buffer_head), 0,
3485	(SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3486	SLAB_MEM_SPREAD),
3487	NULL);
3488
3489	/*
3490	* Limit the bh occupancy to 10% of ZONE_NORMAL
3491	*/
3492	nrpages = (nr_free_buffer_pages() * 10) / 100;
3493	max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3494	hotcpu_notifier(buffer_cpu_notify, 0);
3495	}
3496