path: root/fs/fs-writeback.c (plain)
blob: 6faf93b52c7cd6b4c46b1d0d312f5ad19fca1682

1	/*
2	* fs/fs-writeback.c
3	*
4	* Copyright (C) 2002, Linus Torvalds.
5	*
6	* Contains all the functions related to writing back and waiting
7	* upon dirty inodes against superblocks, and writing back dirty
8	* pages against inodes. ie: data writeback. Writeout of the
9	* inode itself is not handled here.
10	*
11	* 10Apr2002 Andrew Morton
12	* Split out of fs/inode.c
13	* Additions for address_space-based writeback
14	*/
15
16	#include <linux/kernel.h>
17	#include <linux/export.h>
18	#include <linux/spinlock.h>
19	#include <linux/slab.h>
20	#include <linux/sched.h>
21	#include <linux/fs.h>
22	#include <linux/mm.h>
23	#include <linux/pagemap.h>
24	#include <linux/kthread.h>
25	#include <linux/writeback.h>
26	#include <linux/blkdev.h>
27	#include <linux/backing-dev.h>
28	#include <linux/tracepoint.h>
29	#include <linux/device.h>
30	#include <linux/memcontrol.h>
31	#include "internal.h"
32
33	/*
34	* 4MB minimal write chunk size
35	*/
36	#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
37
38	struct wb_completion {
39	atomic_t cnt;
40	};
41
42	/*
43	* Passed into wb_writeback(), essentially a subset of writeback_control
44	*/
45	struct wb_writeback_work {
46	long nr_pages;
47	struct super_block *sb;
48	unsigned long *older_than_this;
49	enum writeback_sync_modes sync_mode;
50	unsigned int tagged_writepages:1;
51	unsigned int for_kupdate:1;
52	unsigned int range_cyclic:1;
53	unsigned int for_background:1;
54	unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
55	unsigned int auto_free:1; /* free on completion */
56	enum wb_reason reason; /* why was writeback initiated? */
57
58	struct list_head list; /* pending work list */
59	struct wb_completion *done; /* set if the caller waits */
60	};
61
62	/*
63	* If one wants to wait for one or more wb_writeback_works, each work's
64	* ->done should be set to a wb_completion defined using the following
65	* macro. Once all work items are issued with wb_queue_work(), the caller
66	* can wait for the completion of all using wb_wait_for_completion(). Work
67	* items which are waited upon aren't freed automatically on completion.
68	*/
69	#define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
70	struct wb_completion cmpl = { \
71	.cnt = ATOMIC_INIT(1), \
72	}
73
74
75	/*
76	* If an inode is constantly having its pages dirtied, but then the
77	* updates stop dirtytime_expire_interval seconds in the past, it's
78	* possible for the worst case time between when an inode has its
79	* timestamps updated and when they finally get written out to be two
80	* dirtytime_expire_intervals. We set the default to 12 hours (in
81	* seconds), which means most of the time inodes will have their
82	* timestamps written to disk after 12 hours, but in the worst case a
83	* few inodes might not their timestamps updated for 24 hours.
84	*/
85	unsigned int dirtytime_expire_interval = 12 * 60 * 60;
86
87	static inline struct inode *wb_inode(struct list_head *head)
88	{
89	return list_entry(head, struct inode, i_io_list);
90	}
91
92	/*
93	* Include the creation of the trace points after defining the
94	* wb_writeback_work structure and inline functions so that the definition
95	* remains local to this file.
96	*/
97	#define CREATE_TRACE_POINTS
98	#include <trace/events/writeback.h>
99
100	EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
101
102	static bool wb_io_lists_populated(struct bdi_writeback *wb)
103	{
104	if (wb_has_dirty_io(wb)) {
105	return false;
106	} else {
107	set_bit(WB_has_dirty_io, &wb->state);
108	WARN_ON_ONCE(!wb->avg_write_bandwidth);
109	atomic_long_add(wb->avg_write_bandwidth,
110	&wb->bdi->tot_write_bandwidth);
111	return true;
112	}
113	}
114
115	static void wb_io_lists_depopulated(struct bdi_writeback *wb)
116	{
117	if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
118	list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
119	clear_bit(WB_has_dirty_io, &wb->state);
120	WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
121	&wb->bdi->tot_write_bandwidth) < 0);
122	}
123	}
124
125	/**
126	* inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
127	* @inode: inode to be moved
128	* @wb: target bdi_writeback
129	* @head: one of @wb->b_{dirty|io|more_io}
130	*
131	* Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
132	* Returns %true if @inode is the first occupant of the !dirty_time IO
133	* lists; otherwise, %false.
134	*/
135	static bool inode_io_list_move_locked(struct inode *inode,
136	struct bdi_writeback *wb,
137	struct list_head *head)
138	{
139	assert_spin_locked(&wb->list_lock);
140
141	list_move(&inode->i_io_list, head);
142
143	/* dirty_time doesn't count as dirty_io until expiration */
144	if (head != &wb->b_dirty_time)
145	return wb_io_lists_populated(wb);
146
147	wb_io_lists_depopulated(wb);
148	return false;
149	}
150
151	/**
152	* inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
153	* @inode: inode to be removed
154	* @wb: bdi_writeback @inode is being removed from
155	*
156	* Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157	* clear %WB_has_dirty_io if all are empty afterwards.
158	*/
159	static void inode_io_list_del_locked(struct inode *inode,
160	struct bdi_writeback *wb)
161	{
162	assert_spin_locked(&wb->list_lock);
163
164	list_del_init(&inode->i_io_list);
165	wb_io_lists_depopulated(wb);
166	}
167
168	static void wb_wakeup(struct bdi_writeback *wb)
169	{
170	spin_lock_bh(&wb->work_lock);
171	if (test_bit(WB_registered, &wb->state))
172	mod_delayed_work(bdi_wq, &wb->dwork, 0);
173	spin_unlock_bh(&wb->work_lock);
174	}
175
176	static void finish_writeback_work(struct bdi_writeback *wb,
177	struct wb_writeback_work *work)
178	{
179	struct wb_completion *done = work->done;
180
181	if (work->auto_free)
182	kfree(work);
183	if (done && atomic_dec_and_test(&done->cnt))
184	wake_up_all(&wb->bdi->wb_waitq);
185	}
186
187	static void wb_queue_work(struct bdi_writeback *wb,
188	struct wb_writeback_work *work)
189	{
190	trace_writeback_queue(wb, work);
191
192	if (work->done)
193	atomic_inc(&work->done->cnt);
194
195	spin_lock_bh(&wb->work_lock);
196
197	if (test_bit(WB_registered, &wb->state)) {
198	list_add_tail(&work->list, &wb->work_list);
199	mod_delayed_work(bdi_wq, &wb->dwork, 0);
200	} else
201	finish_writeback_work(wb, work);
202
203	spin_unlock_bh(&wb->work_lock);
204	}
205
206	/**
207	* wb_wait_for_completion - wait for completion of bdi_writeback_works
208	* @bdi: bdi work items were issued to
209	* @done: target wb_completion
210	*
211	* Wait for one or more work items issued to @bdi with their ->done field
212	* set to @done, which should have been defined with
213	* DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
214	* work items are completed. Work items which are waited upon aren't freed
215	* automatically on completion.
216	*/
217	static void wb_wait_for_completion(struct backing_dev_info *bdi,
218	struct wb_completion *done)
219	{
220	atomic_dec(&done->cnt); /* put down the initial count */
221	wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
222	}
223
224	#ifdef CONFIG_CGROUP_WRITEBACK
225
226	/* parameters for foreign inode detection, see wb_detach_inode() */
227	#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
228	#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
229	#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
230	#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
231
232	#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
233	#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
234	/* each slot's duration is 2s / 16 */
235	#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
236	/* if foreign slots >= 8, switch */
237	#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
238	/* one round can affect upto 5 slots */
239
240	static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
241	static struct workqueue_struct *isw_wq;
242
243	void __inode_attach_wb(struct inode *inode, struct page *page)
244	{
245	struct backing_dev_info *bdi = inode_to_bdi(inode);
246	struct bdi_writeback *wb = NULL;
247
248	if (inode_cgwb_enabled(inode)) {
249	struct cgroup_subsys_state *memcg_css;
250
251	if (page) {
252	memcg_css = mem_cgroup_css_from_page(page);
253	wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254	} else {
255	/* must pin memcg_css, see wb_get_create() */
256	memcg_css = task_get_css(current, memory_cgrp_id);
257	wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
258	css_put(memcg_css);
259	}
260	}
261
262	if (!wb)
263	wb = &bdi->wb;
264
265	/*
266	* There may be multiple instances of this function racing to
267	* update the same inode. Use cmpxchg() to tell the winner.
268	*/
269	if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
270	wb_put(wb);
271	}
272
273	/**
274	* locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
275	* @inode: inode of interest with i_lock held
276	*
277	* Returns @inode's wb with its list_lock held. @inode->i_lock must be
278	* held on entry and is released on return. The returned wb is guaranteed
279	* to stay @inode's associated wb until its list_lock is released.
280	*/
281	static struct bdi_writeback *
282	locked_inode_to_wb_and_lock_list(struct inode *inode)
283	__releases(&inode->i_lock)
284	__acquires(&wb->list_lock)
285	{
286	while (true) {
287	struct bdi_writeback *wb = inode_to_wb(inode);
288
289	/*
290	* inode_to_wb() association is protected by both
291	* @inode->i_lock and @wb->list_lock but list_lock nests
292	* outside i_lock. Drop i_lock and verify that the
293	* association hasn't changed after acquiring list_lock.
294	*/
295	wb_get(wb);
296	spin_unlock(&inode->i_lock);
297	spin_lock(&wb->list_lock);
298
299	/* i_wb may have changed inbetween, can't use inode_to_wb() */
300	if (likely(wb == inode->i_wb)) {
301	wb_put(wb); /* @inode already has ref */
302	return wb;
303	}
304
305	spin_unlock(&wb->list_lock);
306	wb_put(wb);
307	cpu_relax();
308	spin_lock(&inode->i_lock);
309	}
310	}
311
312	/**
313	* inode_to_wb_and_lock_list - determine an inode's wb and lock it
314	* @inode: inode of interest
315	*
316	* Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
317	* on entry.
318	*/
319	static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
320	__acquires(&wb->list_lock)
321	{
322	spin_lock(&inode->i_lock);
323	return locked_inode_to_wb_and_lock_list(inode);
324	}
325
326	struct inode_switch_wbs_context {
327	struct inode *inode;
328	struct bdi_writeback *new_wb;
329
330	struct rcu_head rcu_head;
331	struct work_struct work;
332	};
333
334	static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
335	{
336	down_write(&bdi->wb_switch_rwsem);
337	}
338
339	static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
340	{
341	up_write(&bdi->wb_switch_rwsem);
342	}
343
344	static void inode_switch_wbs_work_fn(struct work_struct *work)
345	{
346	struct inode_switch_wbs_context *isw =
347	container_of(work, struct inode_switch_wbs_context, work);
348	struct inode *inode = isw->inode;
349	struct backing_dev_info *bdi = inode_to_bdi(inode);
350	struct address_space *mapping = inode->i_mapping;
351	struct bdi_writeback *old_wb = inode->i_wb;
352	struct bdi_writeback *new_wb = isw->new_wb;
353	struct radix_tree_iter iter;
354	bool switched = false;
355	void **slot;
356
357	/*
358	* If @inode switches cgwb membership while sync_inodes_sb() is
359	* being issued, sync_inodes_sb() might miss it. Synchronize.
360	*/
361	down_read(&bdi->wb_switch_rwsem);
362
363	/*
364	* By the time control reaches here, RCU grace period has passed
365	* since I_WB_SWITCH assertion and all wb stat update transactions
366	* between unlocked_inode_to_wb_begin/end() are guaranteed to be
367	* synchronizing against mapping->tree_lock.
368	*
369	* Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
370	* gives us exclusion against all wb related operations on @inode
371	* including IO list manipulations and stat updates.
372	*/
373	if (old_wb < new_wb) {
374	spin_lock(&old_wb->list_lock);
375	spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
376	} else {
377	spin_lock(&new_wb->list_lock);
378	spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
379	}
380	spin_lock(&inode->i_lock);
381	spin_lock_irq(&mapping->tree_lock);
382
383	/*
384	* Once I_FREEING is visible under i_lock, the eviction path owns
385	* the inode and we shouldn't modify ->i_io_list.
386	*/
387	if (unlikely(inode->i_state & I_FREEING))
388	goto skip_switch;
389
390	/*
391	* Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392	* to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393	* pages actually under underwriteback.
394	*/
395	radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
396	PAGECACHE_TAG_DIRTY) {
397	struct page *page = radix_tree_deref_slot_protected(slot,
398	&mapping->tree_lock);
399	if (likely(page) && PageDirty(page)) {
400	__dec_wb_stat(old_wb, WB_RECLAIMABLE);
401	__inc_wb_stat(new_wb, WB_RECLAIMABLE);
402	}
403	}
404
405	radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
406	PAGECACHE_TAG_WRITEBACK) {
407	struct page *page = radix_tree_deref_slot_protected(slot,
408	&mapping->tree_lock);
409	if (likely(page)) {
410	WARN_ON_ONCE(!PageWriteback(page));
411	__dec_wb_stat(old_wb, WB_WRITEBACK);
412	__inc_wb_stat(new_wb, WB_WRITEBACK);
413	}
414	}
415
416	wb_get(new_wb);
417
418	/*
419	* Transfer to @new_wb's IO list if necessary. The specific list
420	* @inode was on is ignored and the inode is put on ->b_dirty which
421	* is always correct including from ->b_dirty_time. The transfer
422	* preserves @inode->dirtied_when ordering.
423	*/
424	if (!list_empty(&inode->i_io_list)) {
425	struct inode *pos;
426
427	inode_io_list_del_locked(inode, old_wb);
428	inode->i_wb = new_wb;
429	list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
430	if (time_after_eq(inode->dirtied_when,
431	pos->dirtied_when))
432	break;
433	inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
434	} else {
435	inode->i_wb = new_wb;
436	}
437
438	/* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
439	inode->i_wb_frn_winner = 0;
440	inode->i_wb_frn_avg_time = 0;
441	inode->i_wb_frn_history = 0;
442	switched = true;
443	skip_switch:
444	/*
445	* Paired with load_acquire in unlocked_inode_to_wb_begin() and
446	* ensures that the new wb is visible if they see !I_WB_SWITCH.
447	*/
448	smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
449
450	spin_unlock_irq(&mapping->tree_lock);
451	spin_unlock(&inode->i_lock);
452	spin_unlock(&new_wb->list_lock);
453	spin_unlock(&old_wb->list_lock);
454
455	up_read(&bdi->wb_switch_rwsem);
456
457	if (switched) {
458	wb_wakeup(new_wb);
459	wb_put(old_wb);
460	}
461	wb_put(new_wb);
462
463	iput(inode);
464	kfree(isw);
465
466	atomic_dec(&isw_nr_in_flight);
467	}
468
469	static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
470	{
471	struct inode_switch_wbs_context *isw = container_of(rcu_head,
472	struct inode_switch_wbs_context, rcu_head);
473
474	/* needs to grab bh-unsafe locks, bounce to work item */
475	INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
476	queue_work(isw_wq, &isw->work);
477	}
478
479	/**
480	* inode_switch_wbs - change the wb association of an inode
481	* @inode: target inode
482	* @new_wb_id: ID of the new wb
483	*
484	* Switch @inode's wb association to the wb identified by @new_wb_id. The
485	* switching is performed asynchronously and may fail silently.
486	*/
487	static void inode_switch_wbs(struct inode *inode, int new_wb_id)
488	{
489	struct backing_dev_info *bdi = inode_to_bdi(inode);
490	struct cgroup_subsys_state *memcg_css;
491	struct inode_switch_wbs_context *isw;
492
493	/* noop if seems to be already in progress */
494	if (inode->i_state & I_WB_SWITCH)
495	return;
496
497	/*
498	* Avoid starting new switches while sync_inodes_sb() is in
499	* progress. Otherwise, if the down_write protected issue path
500	* blocks heavily, we might end up starting a large number of
501	* switches which will block on the rwsem.
502	*/
503	if (!down_read_trylock(&bdi->wb_switch_rwsem))
504	return;
505
506	isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
507	if (!isw)
508	goto out_unlock;
509
510	/* find and pin the new wb */
511	rcu_read_lock();
512	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
513	if (memcg_css)
514	isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
515	rcu_read_unlock();
516	if (!isw->new_wb)
517	goto out_free;
518
519	/* while holding I_WB_SWITCH, no one else can update the association */
520	spin_lock(&inode->i_lock);
521	if (!(inode->i_sb->s_flags & MS_ACTIVE) ||
522	inode->i_state & (I_WB_SWITCH | I_FREEING) ||
523	inode_to_wb(inode) == isw->new_wb) {
524	spin_unlock(&inode->i_lock);
525	goto out_free;
526	}
527	inode->i_state |= I_WB_SWITCH;
528	__iget(inode);
529	spin_unlock(&inode->i_lock);
530
531	isw->inode = inode;
532
533	/*
534	* In addition to synchronizing among switchers, I_WB_SWITCH tells
535	* the RCU protected stat update paths to grab the mapping's
536	* tree_lock so that stat transfer can synchronize against them.
537	* Let's continue after I_WB_SWITCH is guaranteed to be visible.
538	*/
539	call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
540
541	atomic_inc(&isw_nr_in_flight);
542
543	goto out_unlock;
544
545	out_free:
546	if (isw->new_wb)
547	wb_put(isw->new_wb);
548	kfree(isw);
549	out_unlock:
550	up_read(&bdi->wb_switch_rwsem);
551	}
552
553	/**
554	* wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
555	* @wbc: writeback_control of interest
556	* @inode: target inode
557	*
558	* @inode is locked and about to be written back under the control of @wbc.
559	* Record @inode's writeback context into @wbc and unlock the i_lock. On
560	* writeback completion, wbc_detach_inode() should be called. This is used
561	* to track the cgroup writeback context.
562	*/
563	void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
564	struct inode *inode)
565	{
566	if (!inode_cgwb_enabled(inode)) {
567	spin_unlock(&inode->i_lock);
568	return;
569	}
570
571	wbc->wb = inode_to_wb(inode);
572	wbc->inode = inode;
573
574	wbc->wb_id = wbc->wb->memcg_css->id;
575	wbc->wb_lcand_id = inode->i_wb_frn_winner;
576	wbc->wb_tcand_id = 0;
577	wbc->wb_bytes = 0;
578	wbc->wb_lcand_bytes = 0;
579	wbc->wb_tcand_bytes = 0;
580
581	wb_get(wbc->wb);
582	spin_unlock(&inode->i_lock);
583
584	/*
585	* A dying wb indicates that the memcg-blkcg mapping has changed
586	* and a new wb is already serving the memcg. Switch immediately.
587	*/
588	if (unlikely(wb_dying(wbc->wb)))
589	inode_switch_wbs(inode, wbc->wb_id);
590	}
591
592	/**
593	* wbc_detach_inode - disassociate wbc from inode and perform foreign detection
594	* @wbc: writeback_control of the just finished writeback
595	*
596	* To be called after a writeback attempt of an inode finishes and undoes
597	* wbc_attach_and_unlock_inode(). Can be called under any context.
598	*
599	* As concurrent write sharing of an inode is expected to be very rare and
600	* memcg only tracks page ownership on first-use basis severely confining
601	* the usefulness of such sharing, cgroup writeback tracks ownership
602	* per-inode. While the support for concurrent write sharing of an inode
603	* is deemed unnecessary, an inode being written to by different cgroups at
604	* different points in time is a lot more common, and, more importantly,
605	* charging only by first-use can too readily lead to grossly incorrect
606	* behaviors (single foreign page can lead to gigabytes of writeback to be
607	* incorrectly attributed).
608	*
609	* To resolve this issue, cgroup writeback detects the majority dirtier of
610	* an inode and transfers the ownership to it. To avoid unnnecessary
611	* oscillation, the detection mechanism keeps track of history and gives
612	* out the switch verdict only if the foreign usage pattern is stable over
613	* a certain amount of time and/or writeback attempts.
614	*
615	* On each writeback attempt, @wbc tries to detect the majority writer
616	* using Boyer-Moore majority vote algorithm. In addition to the byte
617	* count from the majority voting, it also counts the bytes written for the
618	* current wb and the last round's winner wb (max of last round's current
619	* wb, the winner from two rounds ago, and the last round's majority
620	* candidate). Keeping track of the historical winner helps the algorithm
621	* to semi-reliably detect the most active writer even when it's not the
622	* absolute majority.
623	*
624	* Once the winner of the round is determined, whether the winner is
625	* foreign or not and how much IO time the round consumed is recorded in
626	* inode->i_wb_frn_history. If the amount of recorded foreign IO time is
627	* over a certain threshold, the switch verdict is given.
628	*/
629	void wbc_detach_inode(struct writeback_control *wbc)
630	{
631	struct bdi_writeback *wb = wbc->wb;
632	struct inode *inode = wbc->inode;
633	unsigned long avg_time, max_bytes, max_time;
634	u16 history;
635	int max_id;
636
637	if (!wb)
638	return;
639
640	history = inode->i_wb_frn_history;
641	avg_time = inode->i_wb_frn_avg_time;
642
643	/* pick the winner of this round */
644	if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
645	wbc->wb_bytes >= wbc->wb_tcand_bytes) {
646	max_id = wbc->wb_id;
647	max_bytes = wbc->wb_bytes;
648	} else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
649	max_id = wbc->wb_lcand_id;
650	max_bytes = wbc->wb_lcand_bytes;
651	} else {
652	max_id = wbc->wb_tcand_id;
653	max_bytes = wbc->wb_tcand_bytes;
654	}
655
656	/*
657	* Calculate the amount of IO time the winner consumed and fold it
658	* into the running average kept per inode. If the consumed IO
659	* time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
660	* deciding whether to switch or not. This is to prevent one-off
661	* small dirtiers from skewing the verdict.
662	*/
663	max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
664	wb->avg_write_bandwidth);
665	if (avg_time)
666	avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
667	(avg_time >> WB_FRN_TIME_AVG_SHIFT);
668	else
669	avg_time = max_time; /* immediate catch up on first run */
670
671	if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
672	int slots;
673
674	/*
675	* The switch verdict is reached if foreign wb's consume
676	* more than a certain proportion of IO time in a
677	* WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
678	* history mask where each bit represents one sixteenth of
679	* the period. Determine the number of slots to shift into
680	* history from @max_time.
681	*/
682	slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
683	(unsigned long)WB_FRN_HIST_MAX_SLOTS);
684	history <<= slots;
685	if (wbc->wb_id != max_id)
686	history |= (1U << slots) - 1;
687
688	/*
689	* Switch if the current wb isn't the consistent winner.
690	* If there are multiple closely competing dirtiers, the
691	* inode may switch across them repeatedly over time, which
692	* is okay. The main goal is avoiding keeping an inode on
693	* the wrong wb for an extended period of time.
694	*/
695	if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
696	inode_switch_wbs(inode, max_id);
697	}
698
699	/*
700	* Multiple instances of this function may race to update the
701	* following fields but we don't mind occassional inaccuracies.
702	*/
703	inode->i_wb_frn_winner = max_id;
704	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
705	inode->i_wb_frn_history = history;
706
707	wb_put(wbc->wb);
708	wbc->wb = NULL;
709	}
710
711	/**
712	* wbc_account_io - account IO issued during writeback
713	* @wbc: writeback_control of the writeback in progress
714	* @page: page being written out
715	* @bytes: number of bytes being written out
716	*
717	* @bytes from @page are about to written out during the writeback
718	* controlled by @wbc. Keep the book for foreign inode detection. See
719	* wbc_detach_inode().
720	*/
721	void wbc_account_io(struct writeback_control *wbc, struct page *page,
722	size_t bytes)
723	{
724	int id;
725
726	/*
727	* pageout() path doesn't attach @wbc to the inode being written
728	* out. This is intentional as we don't want the function to block
729	* behind a slow cgroup. Ultimately, we want pageout() to kick off
730	* regular writeback instead of writing things out itself.
731	*/
732	if (!wbc->wb)
733	return;
734
735	id = mem_cgroup_css_from_page(page)->id;
736
737	if (id == wbc->wb_id) {
738	wbc->wb_bytes += bytes;
739	return;
740	}
741
742	if (id == wbc->wb_lcand_id)
743	wbc->wb_lcand_bytes += bytes;
744
745	/* Boyer-Moore majority vote algorithm */
746	if (!wbc->wb_tcand_bytes)
747	wbc->wb_tcand_id = id;
748	if (id == wbc->wb_tcand_id)
749	wbc->wb_tcand_bytes += bytes;
750	else
751	wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
752	}
753	EXPORT_SYMBOL_GPL(wbc_account_io);
754
755	/**
756	* inode_congested - test whether an inode is congested
757	* @inode: inode to test for congestion (may be NULL)
758	* @cong_bits: mask of WB_[a]sync_congested bits to test
759	*
760	* Tests whether @inode is congested. @cong_bits is the mask of congestion
761	* bits to test and the return value is the mask of set bits.
762	*
763	* If cgroup writeback is enabled for @inode, the congestion state is
764	* determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
765	* associated with @inode is congested; otherwise, the root wb's congestion
766	* state is used.
767	*
768	* @inode is allowed to be NULL as this function is often called on
769	* mapping->host which is NULL for the swapper space.
770	*/
771	int inode_congested(struct inode *inode, int cong_bits)
772	{
773	/*
774	* Once set, ->i_wb never becomes NULL while the inode is alive.
775	* Start transaction iff ->i_wb is visible.
776	*/
777	if (inode && inode_to_wb_is_valid(inode)) {
778	struct bdi_writeback *wb;
779	struct wb_lock_cookie lock_cookie = {};
780	bool congested;
781
782	wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
783	congested = wb_congested(wb, cong_bits);
784	unlocked_inode_to_wb_end(inode, &lock_cookie);
785	return congested;
786	}
787
788	return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
789	}
790	EXPORT_SYMBOL_GPL(inode_congested);
791
792	/**
793	* wb_split_bdi_pages - split nr_pages to write according to bandwidth
794	* @wb: target bdi_writeback to split @nr_pages to
795	* @nr_pages: number of pages to write for the whole bdi
796	*
797	* Split @wb's portion of @nr_pages according to @wb's write bandwidth in
798	* relation to the total write bandwidth of all wb's w/ dirty inodes on
799	* @wb->bdi.
800	*/
801	static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
802	{
803	unsigned long this_bw = wb->avg_write_bandwidth;
804	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
805
806	if (nr_pages == LONG_MAX)
807	return LONG_MAX;
808
809	/*
810	* This may be called on clean wb's and proportional distribution
811	* may not make sense, just use the original @nr_pages in those
812	* cases. In general, we wanna err on the side of writing more.
813	*/
814	if (!tot_bw || this_bw >= tot_bw)
815	return nr_pages;
816	else
817	return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
818	}
819
820	/**
821	* bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
822	* @bdi: target backing_dev_info
823	* @base_work: wb_writeback_work to issue
824	* @skip_if_busy: skip wb's which already have writeback in progress
825	*
826	* Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
827	* have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
828	* distributed to the busy wbs according to each wb's proportion in the
829	* total active write bandwidth of @bdi.
830	*/
831	static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
832	struct wb_writeback_work *base_work,
833	bool skip_if_busy)
834	{
835	struct bdi_writeback *last_wb = NULL;
836	struct bdi_writeback *wb = list_entry(&bdi->wb_list,
837	struct bdi_writeback, bdi_node);
838
839	might_sleep();
840	restart:
841	rcu_read_lock();
842	list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
843	DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
844	struct wb_writeback_work fallback_work;
845	struct wb_writeback_work *work;
846	long nr_pages;
847
848	if (last_wb) {
849	wb_put(last_wb);
850	last_wb = NULL;
851	}
852
853	/* SYNC_ALL writes out I_DIRTY_TIME too */
854	if (!wb_has_dirty_io(wb) &&
855	(base_work->sync_mode == WB_SYNC_NONE ||
856	list_empty(&wb->b_dirty_time)))
857	continue;
858	if (skip_if_busy && writeback_in_progress(wb))
859	continue;
860
861	nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
862
863	work = kmalloc(sizeof(*work), GFP_ATOMIC);
864	if (work) {
865	*work = *base_work;
866	work->nr_pages = nr_pages;
867	work->auto_free = 1;
868	wb_queue_work(wb, work);
869	continue;
870	}
871
872	/* alloc failed, execute synchronously using on-stack fallback */
873	work = &fallback_work;
874	*work = *base_work;
875	work->nr_pages = nr_pages;
876	work->auto_free = 0;
877	work->done = &fallback_work_done;
878
879	wb_queue_work(wb, work);
880
881	/*
882	* Pin @wb so that it stays on @bdi->wb_list. This allows
883	* continuing iteration from @wb after dropping and
884	* regrabbing rcu read lock.
885	*/
886	wb_get(wb);
887	last_wb = wb;
888
889	rcu_read_unlock();
890	wb_wait_for_completion(bdi, &fallback_work_done);
891	goto restart;
892	}
893	rcu_read_unlock();
894
895	if (last_wb)
896	wb_put(last_wb);
897	}
898
899	/**
900	* cgroup_writeback_umount - flush inode wb switches for umount
901	*
902	* This function is called when a super_block is about to be destroyed and
903	* flushes in-flight inode wb switches. An inode wb switch goes through
904	* RCU and then workqueue, so the two need to be flushed in order to ensure
905	* that all previously scheduled switches are finished. As wb switches are
906	* rare occurrences and synchronize_rcu() can take a while, perform
907	* flushing iff wb switches are in flight.
908	*/
909	void cgroup_writeback_umount(void)
910	{
911	if (atomic_read(&isw_nr_in_flight)) {
912	/*
913	* Use rcu_barrier() to wait for all pending callbacks to
914	* ensure that all in-flight wb switches are in the workqueue.
915	*/
916	rcu_barrier();
917	flush_workqueue(isw_wq);
918	}
919	}
920
921	static int __init cgroup_writeback_init(void)
922	{
923	isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
924	if (!isw_wq)
925	return -ENOMEM;
926	return 0;
927	}
928	fs_initcall(cgroup_writeback_init);
929
930	#else /* CONFIG_CGROUP_WRITEBACK */
931
932	static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
933	static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
934
935	static struct bdi_writeback *
936	locked_inode_to_wb_and_lock_list(struct inode *inode)
937	__releases(&inode->i_lock)
938	__acquires(&wb->list_lock)
939	{
940	struct bdi_writeback *wb = inode_to_wb(inode);
941
942	spin_unlock(&inode->i_lock);
943	spin_lock(&wb->list_lock);
944	return wb;
945	}
946
947	static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
948	__acquires(&wb->list_lock)
949	{
950	struct bdi_writeback *wb = inode_to_wb(inode);
951
952	spin_lock(&wb->list_lock);
953	return wb;
954	}
955
956	static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
957	{
958	return nr_pages;
959	}
960
961	static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
962	struct wb_writeback_work *base_work,
963	bool skip_if_busy)
964	{
965	might_sleep();
966
967	if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
968	base_work->auto_free = 0;
969	wb_queue_work(&bdi->wb, base_work);
970	}
971	}
972
973	#endif /* CONFIG_CGROUP_WRITEBACK */
974
975	void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
976	bool range_cyclic, enum wb_reason reason)
977	{
978	struct wb_writeback_work *work;
979
980	if (!wb_has_dirty_io(wb))
981	return;
982
983	/*
984	* This is WB_SYNC_NONE writeback, so if allocation fails just
985	* wakeup the thread for old dirty data writeback
986	*/
987	work = kzalloc(sizeof(*work),
988	GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
989	if (!work) {
990	trace_writeback_nowork(wb);
991	wb_wakeup(wb);
992	return;
993	}
994
995	work->sync_mode = WB_SYNC_NONE;
996	work->nr_pages = nr_pages;
997	work->range_cyclic = range_cyclic;
998	work->reason = reason;
999	work->auto_free = 1;
1000
1001	wb_queue_work(wb, work);
1002	}
1003
1004	/**
1005	* wb_start_background_writeback - start background writeback
1006	* @wb: bdi_writback to write from
1007	*
1008	* Description:
1009	* This makes sure WB_SYNC_NONE background writeback happens. When
1010	* this function returns, it is only guaranteed that for given wb
1011	* some IO is happening if we are over background dirty threshold.
1012	* Caller need not hold sb s_umount semaphore.
1013	*/
1014	void wb_start_background_writeback(struct bdi_writeback *wb)
1015	{
1016	/*
1017	* We just wake up the flusher thread. It will perform background
1018	* writeback as soon as there is no other work to do.
1019	*/
1020	trace_writeback_wake_background(wb);
1021	wb_wakeup(wb);
1022	}
1023
1024	/*
1025	* Remove the inode from the writeback list it is on.
1026	*/
1027	void inode_io_list_del(struct inode *inode)
1028	{
1029	struct bdi_writeback *wb;
1030
1031	wb = inode_to_wb_and_lock_list(inode);
1032	inode_io_list_del_locked(inode, wb);
1033	spin_unlock(&wb->list_lock);
1034	}
1035
1036	/*
1037	* mark an inode as under writeback on the sb
1038	*/
1039	void sb_mark_inode_writeback(struct inode *inode)
1040	{
1041	struct super_block *sb = inode->i_sb;
1042	unsigned long flags;
1043
1044	if (list_empty(&inode->i_wb_list)) {
1045	spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1046	if (list_empty(&inode->i_wb_list)) {
1047	list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1048	trace_sb_mark_inode_writeback(inode);
1049	}
1050	spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1051	}
1052	}
1053
1054	/*
1055	* clear an inode as under writeback on the sb
1056	*/
1057	void sb_clear_inode_writeback(struct inode *inode)
1058	{
1059	struct super_block *sb = inode->i_sb;
1060	unsigned long flags;
1061
1062	if (!list_empty(&inode->i_wb_list)) {
1063	spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1064	if (!list_empty(&inode->i_wb_list)) {
1065	list_del_init(&inode->i_wb_list);
1066	trace_sb_clear_inode_writeback(inode);
1067	}
1068	spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1069	}
1070	}
1071
1072	/*
1073	* Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1074	* furthest end of its superblock's dirty-inode list.
1075	*
1076	* Before stamping the inode's ->dirtied_when, we check to see whether it is
1077	* already the most-recently-dirtied inode on the b_dirty list. If that is
1078	* the case then the inode must have been redirtied while it was being written
1079	* out and we don't reset its dirtied_when.
1080	*/
1081	static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1082	{
1083	if (!list_empty(&wb->b_dirty)) {
1084	struct inode *tail;
1085
1086	tail = wb_inode(wb->b_dirty.next);
1087	if (time_before(inode->dirtied_when, tail->dirtied_when))
1088	inode->dirtied_when = jiffies;
1089	}
1090	inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1091	}
1092
1093	/*
1094	* requeue inode for re-scanning after bdi->b_io list is exhausted.
1095	*/
1096	static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1097	{
1098	inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1099	}
1100
1101	static void inode_sync_complete(struct inode *inode)
1102	{
1103	inode->i_state &= ~I_SYNC;
1104	/* If inode is clean an unused, put it into LRU now... */
1105	inode_add_lru(inode);
1106	/* Waiters must see I_SYNC cleared before being woken up */
1107	smp_mb();
1108	wake_up_bit(&inode->i_state, __I_SYNC);
1109	}
1110
1111	static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1112	{
1113	bool ret = time_after(inode->dirtied_when, t);
1114	#ifndef CONFIG_64BIT
1115	/*
1116	* For inodes being constantly redirtied, dirtied_when can get stuck.
1117	* It _appears_ to be in the future, but is actually in distant past.
1118	* This test is necessary to prevent such wrapped-around relative times
1119	* from permanently stopping the whole bdi writeback.
1120	*/
1121	ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1122	#endif
1123	return ret;
1124	}
1125
1126	#define EXPIRE_DIRTY_ATIME 0x0001
1127
1128	/*
1129	* Move expired (dirtied before work->older_than_this) dirty inodes from
1130	* @delaying_queue to @dispatch_queue.
1131	*/
1132	static int move_expired_inodes(struct list_head *delaying_queue,
1133	struct list_head *dispatch_queue,
1134	int flags,
1135	struct wb_writeback_work *work)
1136	{
1137	unsigned long *older_than_this = NULL;
1138	unsigned long expire_time;
1139	LIST_HEAD(tmp);
1140	struct list_head *pos, *node;
1141	struct super_block *sb = NULL;
1142	struct inode *inode;
1143	int do_sb_sort = 0;
1144	int moved = 0;
1145
1146	if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1147	older_than_this = work->older_than_this;
1148	else if (!work->for_sync) {
1149	expire_time = jiffies - (dirtytime_expire_interval * HZ);
1150	older_than_this = &expire_time;
1151	}
1152	while (!list_empty(delaying_queue)) {
1153	inode = wb_inode(delaying_queue->prev);
1154	if (older_than_this &&
1155	inode_dirtied_after(inode, *older_than_this))
1156	break;
1157	list_move(&inode->i_io_list, &tmp);
1158	moved++;
1159	if (flags & EXPIRE_DIRTY_ATIME)
1160	set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1161	if (sb_is_blkdev_sb(inode->i_sb))
1162	continue;
1163	if (sb && sb != inode->i_sb)
1164	do_sb_sort = 1;
1165	sb = inode->i_sb;
1166	}
1167
1168	/* just one sb in list, splice to dispatch_queue and we're done */
1169	if (!do_sb_sort) {
1170	list_splice(&tmp, dispatch_queue);
1171	goto out;
1172	}
1173
1174	/* Move inodes from one superblock together */
1175	while (!list_empty(&tmp)) {
1176	sb = wb_inode(tmp.prev)->i_sb;
1177	list_for_each_prev_safe(pos, node, &tmp) {
1178	inode = wb_inode(pos);
1179	if (inode->i_sb == sb)
1180	list_move(&inode->i_io_list, dispatch_queue);
1181	}
1182	}
1183	out:
1184	return moved;
1185	}
1186
1187	/*
1188	* Queue all expired dirty inodes for io, eldest first.
1189	* Before
1190	* newly dirtied b_dirty b_io b_more_io
1191	* =============> gf edc BA
1192	* After
1193	* newly dirtied b_dirty b_io b_more_io
1194	* =============> g fBAedc
1195	* |
1196	* +--> dequeue for IO
1197	*/
1198	static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1199	{
1200	int moved;
1201
1202	assert_spin_locked(&wb->list_lock);
1203	list_splice_init(&wb->b_more_io, &wb->b_io);
1204	moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1205	moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1206	EXPIRE_DIRTY_ATIME, work);
1207	if (moved)
1208	wb_io_lists_populated(wb);
1209	trace_writeback_queue_io(wb, work, moved);
1210	}
1211
1212	static int write_inode(struct inode *inode, struct writeback_control *wbc)
1213	{
1214	int ret;
1215
1216	if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1217	trace_writeback_write_inode_start(inode, wbc);
1218	ret = inode->i_sb->s_op->write_inode(inode, wbc);
1219	trace_writeback_write_inode(inode, wbc);
1220	return ret;
1221	}
1222	return 0;
1223	}
1224
1225	/*
1226	* Wait for writeback on an inode to complete. Called with i_lock held.
1227	* Caller must make sure inode cannot go away when we drop i_lock.
1228	*/
1229	static void __inode_wait_for_writeback(struct inode *inode)
1230	__releases(inode->i_lock)
1231	__acquires(inode->i_lock)
1232	{
1233	DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1234	wait_queue_head_t *wqh;
1235
1236	wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1237	while (inode->i_state & I_SYNC) {
1238	spin_unlock(&inode->i_lock);
1239	__wait_on_bit(wqh, &wq, bit_wait,
1240	TASK_UNINTERRUPTIBLE);
1241	spin_lock(&inode->i_lock);
1242	}
1243	}
1244
1245	/*
1246	* Wait for writeback on an inode to complete. Caller must have inode pinned.
1247	*/
1248	void inode_wait_for_writeback(struct inode *inode)
1249	{
1250	spin_lock(&inode->i_lock);
1251	__inode_wait_for_writeback(inode);
1252	spin_unlock(&inode->i_lock);
1253	}
1254
1255	/*
1256	* Sleep until I_SYNC is cleared. This function must be called with i_lock
1257	* held and drops it. It is aimed for callers not holding any inode reference
1258	* so once i_lock is dropped, inode can go away.
1259	*/
1260	static void inode_sleep_on_writeback(struct inode *inode)
1261	__releases(inode->i_lock)
1262	{
1263	DEFINE_WAIT(wait);
1264	wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1265	int sleep;
1266
1267	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1268	sleep = inode->i_state & I_SYNC;
1269	spin_unlock(&inode->i_lock);
1270	if (sleep)
1271	schedule();
1272	finish_wait(wqh, &wait);
1273	}
1274
1275	/*
1276	* Find proper writeback list for the inode depending on its current state and
1277	* possibly also change of its state while we were doing writeback. Here we
1278	* handle things such as livelock prevention or fairness of writeback among
1279	* inodes. This function can be called only by flusher thread - noone else
1280	* processes all inodes in writeback lists and requeueing inodes behind flusher
1281	* thread's back can have unexpected consequences.
1282	*/
1283	static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1284	struct writeback_control *wbc)
1285	{
1286	if (inode->i_state & I_FREEING)
1287	return;
1288
1289	/*
1290	* Sync livelock prevention. Each inode is tagged and synced in one
1291	* shot. If still dirty, it will be redirty_tail()'ed below. Update
1292	* the dirty time to prevent enqueue and sync it again.
1293	*/
1294	if ((inode->i_state & I_DIRTY) &&
1295	(wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1296	inode->dirtied_when = jiffies;
1297
1298	if (wbc->pages_skipped) {
1299	/*
1300	* writeback is not making progress due to locked
1301	* buffers. Skip this inode for now.
1302	*/
1303	redirty_tail(inode, wb);
1304	return;
1305	}
1306
1307	if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1308	/*
1309	* We didn't write back all the pages. nfs_writepages()
1310	* sometimes bales out without doing anything.
1311	*/
1312	if (wbc->nr_to_write <= 0) {
1313	/* Slice used up. Queue for next turn. */
1314	requeue_io(inode, wb);
1315	} else {
1316	/*
1317	* Writeback blocked by something other than
1318	* congestion. Delay the inode for some time to
1319	* avoid spinning on the CPU (100% iowait)
1320	* retrying writeback of the dirty page/inode
1321	* that cannot be performed immediately.
1322	*/
1323	redirty_tail(inode, wb);
1324	}
1325	} else if (inode->i_state & I_DIRTY) {
1326	/*
1327	* Filesystems can dirty the inode during writeback operations,
1328	* such as delayed allocation during submission or metadata
1329	* updates after data IO completion.
1330	*/
1331	redirty_tail(inode, wb);
1332	} else if (inode->i_state & I_DIRTY_TIME) {
1333	inode->dirtied_when = jiffies;
1334	inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1335	} else {
1336	/* The inode is clean. Remove from writeback lists. */
1337	inode_io_list_del_locked(inode, wb);
1338	}
1339	}
1340
1341	/*
1342	* Write out an inode and its dirty pages. Do not update the writeback list
1343	* linkage. That is left to the caller. The caller is also responsible for
1344	* setting I_SYNC flag and calling inode_sync_complete() to clear it.
1345	*/
1346	static int
1347	__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1348	{
1349	struct address_space *mapping = inode->i_mapping;
1350	long nr_to_write = wbc->nr_to_write;
1351	unsigned dirty;
1352	int ret;
1353
1354	WARN_ON(!(inode->i_state & I_SYNC));
1355
1356	trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1357
1358	ret = do_writepages(mapping, wbc);
1359
1360	/*
1361	* Make sure to wait on the data before writing out the metadata.
1362	* This is important for filesystems that modify metadata on data
1363	* I/O completion. We don't do it for sync(2) writeback because it has a
1364	* separate, external IO completion path and ->sync_fs for guaranteeing
1365	* inode metadata is written back correctly.
1366	*/
1367	if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1368	int err = filemap_fdatawait(mapping);
1369	if (ret == 0)
1370	ret = err;
1371	}
1372
1373	/*
1374	* Some filesystems may redirty the inode during the writeback
1375	* due to delalloc, clear dirty metadata flags right before
1376	* write_inode()
1377	*/
1378	spin_lock(&inode->i_lock);
1379
1380	dirty = inode->i_state & I_DIRTY;
1381	if (inode->i_state & I_DIRTY_TIME) {
1382	if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1383	wbc->sync_mode == WB_SYNC_ALL ||
1384	unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1385	unlikely(time_after(jiffies,
1386	(inode->dirtied_time_when +
1387	dirtytime_expire_interval * HZ)))) {
1388	dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1389	trace_writeback_lazytime(inode);
1390	}
1391	} else
1392	inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1393	inode->i_state &= ~dirty;
1394
1395	/*
1396	* Paired with smp_mb() in __mark_inode_dirty(). This allows
1397	* __mark_inode_dirty() to test i_state without grabbing i_lock -
1398	* either they see the I_DIRTY bits cleared or we see the dirtied
1399	* inode.
1400	*
1401	* I_DIRTY_PAGES is always cleared together above even if @mapping
1402	* still has dirty pages. The flag is reinstated after smp_mb() if
1403	* necessary. This guarantees that either __mark_inode_dirty()
1404	* sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1405	*/
1406	smp_mb();
1407
1408	if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1409	inode->i_state |= I_DIRTY_PAGES;
1410
1411	spin_unlock(&inode->i_lock);
1412
1413	if (dirty & I_DIRTY_TIME)
1414	mark_inode_dirty_sync(inode);
1415	/* Don't write the inode if only I_DIRTY_PAGES was set */
1416	if (dirty & ~I_DIRTY_PAGES) {
1417	int err = write_inode(inode, wbc);
1418	if (ret == 0)
1419	ret = err;
1420	}
1421	trace_writeback_single_inode(inode, wbc, nr_to_write);
1422	return ret;
1423	}
1424
1425	/*
1426	* Write out an inode's dirty pages. Either the caller has an active reference
1427	* on the inode or the inode has I_WILL_FREE set.
1428	*
1429	* This function is designed to be called for writing back one inode which
1430	* we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1431	* and does more profound writeback list handling in writeback_sb_inodes().
1432	*/
1433	static int writeback_single_inode(struct inode *inode,
1434	struct writeback_control *wbc)
1435	{
1436	struct bdi_writeback *wb;
1437	int ret = 0;
1438
1439	spin_lock(&inode->i_lock);
1440	if (!atomic_read(&inode->i_count))
1441	WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1442	else
1443	WARN_ON(inode->i_state & I_WILL_FREE);
1444
1445	if (inode->i_state & I_SYNC) {
1446	if (wbc->sync_mode != WB_SYNC_ALL)
1447	goto out;
1448	/*
1449	* It's a data-integrity sync. We must wait. Since callers hold
1450	* inode reference or inode has I_WILL_FREE set, it cannot go
1451	* away under us.
1452	*/
1453	__inode_wait_for_writeback(inode);
1454	}
1455	WARN_ON(inode->i_state & I_SYNC);
1456	/*
1457	* Skip inode if it is clean and we have no outstanding writeback in
1458	* WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1459	* function since flusher thread may be doing for example sync in
1460	* parallel and if we move the inode, it could get skipped. So here we
1461	* make sure inode is on some writeback list and leave it there unless
1462	* we have completely cleaned the inode.
1463	*/
1464	if (!(inode->i_state & I_DIRTY_ALL) &&
1465	(wbc->sync_mode != WB_SYNC_ALL ||
1466	!mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1467	goto out;
1468	inode->i_state |= I_SYNC;
1469	wbc_attach_and_unlock_inode(wbc, inode);
1470
1471	ret = __writeback_single_inode(inode, wbc);
1472
1473	wbc_detach_inode(wbc);
1474
1475	wb = inode_to_wb_and_lock_list(inode);
1476	spin_lock(&inode->i_lock);
1477	/*
1478	* If inode is clean, remove it from writeback lists. Otherwise don't
1479	* touch it. See comment above for explanation.
1480	*/
1481	if (!(inode->i_state & I_DIRTY_ALL))
1482	inode_io_list_del_locked(inode, wb);
1483	spin_unlock(&wb->list_lock);
1484	inode_sync_complete(inode);
1485	out:
1486	spin_unlock(&inode->i_lock);
1487	return ret;
1488	}
1489
1490	static long writeback_chunk_size(struct bdi_writeback *wb,
1491	struct wb_writeback_work *work)
1492	{
1493	long pages;
1494
1495	/*
1496	* WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1497	* inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1498	* here avoids calling into writeback_inodes_wb() more than once.
1499	*
1500	* The intended call sequence for WB_SYNC_ALL writeback is:
1501	*
1502	* wb_writeback()
1503	* writeback_sb_inodes() <== called only once
1504	* write_cache_pages() <== called once for each inode
1505	* (quickly) tag currently dirty pages
1506	* (maybe slowly) sync all tagged pages
1507	*/
1508	if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1509	pages = LONG_MAX;
1510	else {
1511	pages = min(wb->avg_write_bandwidth / 2,
1512	global_wb_domain.dirty_limit / DIRTY_SCOPE);
1513	pages = min(pages, work->nr_pages);
1514	pages = round_down(pages + MIN_WRITEBACK_PAGES,
1515	MIN_WRITEBACK_PAGES);
1516	}
1517
1518	return pages;
1519	}
1520
1521	/*
1522	* Write a portion of b_io inodes which belong to @sb.
1523	*
1524	* Return the number of pages and/or inodes written.
1525	*
1526	* NOTE! This is called with wb->list_lock held, and will
1527	* unlock and relock that for each inode it ends up doing
1528	* IO for.
1529	*/
1530	static long writeback_sb_inodes(struct super_block *sb,
1531	struct bdi_writeback *wb,
1532	struct wb_writeback_work *work)
1533	{
1534	struct writeback_control wbc = {
1535	.sync_mode = work->sync_mode,
1536	.tagged_writepages = work->tagged_writepages,
1537	.for_kupdate = work->for_kupdate,
1538	.for_background = work->for_background,
1539	.for_sync = work->for_sync,
1540	.range_cyclic = work->range_cyclic,
1541	.range_start = 0,
1542	.range_end = LLONG_MAX,
1543	};
1544	unsigned long start_time = jiffies;
1545	long write_chunk;
1546	long wrote = 0; /* count both pages and inodes */
1547
1548	while (!list_empty(&wb->b_io)) {
1549	struct inode *inode = wb_inode(wb->b_io.prev);
1550	struct bdi_writeback *tmp_wb;
1551
1552	if (inode->i_sb != sb) {
1553	if (work->sb) {
1554	/*
1555	* We only want to write back data for this
1556	* superblock, move all inodes not belonging
1557	* to it back onto the dirty list.
1558	*/
1559	redirty_tail(inode, wb);
1560	continue;
1561	}
1562
1563	/*
1564	* The inode belongs to a different superblock.
1565	* Bounce back to the caller to unpin this and
1566	* pin the next superblock.
1567	*/
1568	break;
1569	}
1570
1571	/*
1572	* Don't bother with new inodes or inodes being freed, first
1573	* kind does not need periodic writeout yet, and for the latter
1574	* kind writeout is handled by the freer.
1575	*/
1576	spin_lock(&inode->i_lock);
1577	if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1578	spin_unlock(&inode->i_lock);
1579	redirty_tail(inode, wb);
1580	continue;
1581	}
1582	if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1583	/*
1584	* If this inode is locked for writeback and we are not
1585	* doing writeback-for-data-integrity, move it to
1586	* b_more_io so that writeback can proceed with the
1587	* other inodes on s_io.
1588	*
1589	* We'll have another go at writing back this inode
1590	* when we completed a full scan of b_io.
1591	*/
1592	spin_unlock(&inode->i_lock);
1593	requeue_io(inode, wb);
1594	trace_writeback_sb_inodes_requeue(inode);
1595	continue;
1596	}
1597	spin_unlock(&wb->list_lock);
1598
1599	/*
1600	* We already requeued the inode if it had I_SYNC set and we
1601	* are doing WB_SYNC_NONE writeback. So this catches only the
1602	* WB_SYNC_ALL case.
1603	*/
1604	if (inode->i_state & I_SYNC) {
1605	/* Wait for I_SYNC. This function drops i_lock... */
1606	inode_sleep_on_writeback(inode);
1607	/* Inode may be gone, start again */
1608	spin_lock(&wb->list_lock);
1609	continue;
1610	}
1611	inode->i_state |= I_SYNC;
1612	wbc_attach_and_unlock_inode(&wbc, inode);
1613
1614	write_chunk = writeback_chunk_size(wb, work);
1615	wbc.nr_to_write = write_chunk;
1616	wbc.pages_skipped = 0;
1617
1618	/*
1619	* We use I_SYNC to pin the inode in memory. While it is set
1620	* evict_inode() will wait so the inode cannot be freed.
1621	*/
1622	__writeback_single_inode(inode, &wbc);
1623
1624	wbc_detach_inode(&wbc);
1625	work->nr_pages -= write_chunk - wbc.nr_to_write;
1626	wrote += write_chunk - wbc.nr_to_write;
1627
1628	if (need_resched()) {
1629	/*
1630	* We're trying to balance between building up a nice
1631	* long list of IOs to improve our merge rate, and
1632	* getting those IOs out quickly for anyone throttling
1633	* in balance_dirty_pages(). cond_resched() doesn't
1634	* unplug, so get our IOs out the door before we
1635	* give up the CPU.
1636	*/
1637	blk_flush_plug(current);
1638	cond_resched();
1639	}
1640
1641	/*
1642	* Requeue @inode if still dirty. Be careful as @inode may
1643	* have been switched to another wb in the meantime.
1644	*/
1645	tmp_wb = inode_to_wb_and_lock_list(inode);
1646	spin_lock(&inode->i_lock);
1647	if (!(inode->i_state & I_DIRTY_ALL))
1648	wrote++;
1649	requeue_inode(inode, tmp_wb, &wbc);
1650	inode_sync_complete(inode);
1651	spin_unlock(&inode->i_lock);
1652
1653	if (unlikely(tmp_wb != wb)) {
1654	spin_unlock(&tmp_wb->list_lock);
1655	spin_lock(&wb->list_lock);
1656	}
1657
1658	/*
1659	* bail out to wb_writeback() often enough to check
1660	* background threshold and other termination conditions.
1661	*/
1662	if (wrote) {
1663	if (time_is_before_jiffies(start_time + HZ / 10UL))
1664	break;
1665	if (work->nr_pages <= 0)
1666	break;
1667	}
1668	}
1669	return wrote;
1670	}
1671
1672	static long __writeback_inodes_wb(struct bdi_writeback *wb,
1673	struct wb_writeback_work *work)
1674	{
1675	unsigned long start_time = jiffies;
1676	long wrote = 0;
1677
1678	while (!list_empty(&wb->b_io)) {
1679	struct inode *inode = wb_inode(wb->b_io.prev);
1680	struct super_block *sb = inode->i_sb;
1681
1682	if (!trylock_super(sb)) {
1683	/*
1684	* trylock_super() may fail consistently due to
1685	* s_umount being grabbed by someone else. Don't use
1686	* requeue_io() to avoid busy retrying the inode/sb.
1687	*/
1688	redirty_tail(inode, wb);
1689	continue;
1690	}
1691	wrote += writeback_sb_inodes(sb, wb, work);
1692	up_read(&sb->s_umount);
1693
1694	/* refer to the same tests at the end of writeback_sb_inodes */
1695	if (wrote) {
1696	if (time_is_before_jiffies(start_time + HZ / 10UL))
1697	break;
1698	if (work->nr_pages <= 0)
1699	break;
1700	}
1701	}
1702	/* Leave any unwritten inodes on b_io */
1703	return wrote;
1704	}
1705
1706	static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1707	enum wb_reason reason)
1708	{
1709	struct wb_writeback_work work = {
1710	.nr_pages = nr_pages,
1711	.sync_mode = WB_SYNC_NONE,
1712	.range_cyclic = 1,
1713	.reason = reason,
1714	};
1715	struct blk_plug plug;
1716
1717	blk_start_plug(&plug);
1718	spin_lock(&wb->list_lock);
1719	if (list_empty(&wb->b_io))
1720	queue_io(wb, &work);
1721	__writeback_inodes_wb(wb, &work);
1722	spin_unlock(&wb->list_lock);
1723	blk_finish_plug(&plug);
1724
1725	return nr_pages - work.nr_pages;
1726	}
1727
1728	/*
1729	* Explicit flushing or periodic writeback of "old" data.
1730	*
1731	* Define "old": the first time one of an inode's pages is dirtied, we mark the
1732	* dirtying-time in the inode's address_space. So this periodic writeback code
1733	* just walks the superblock inode list, writing back any inodes which are
1734	* older than a specific point in time.
1735	*
1736	* Try to run once per dirty_writeback_interval. But if a writeback event
1737	* takes longer than a dirty_writeback_interval interval, then leave a
1738	* one-second gap.
1739	*
1740	* older_than_this takes precedence over nr_to_write. So we'll only write back
1741	* all dirty pages if they are all attached to "old" mappings.
1742	*/
1743	static long wb_writeback(struct bdi_writeback *wb,
1744	struct wb_writeback_work *work)
1745	{
1746	unsigned long wb_start = jiffies;
1747	long nr_pages = work->nr_pages;
1748	unsigned long oldest_jif;
1749	struct inode *inode;
1750	long progress;
1751	struct blk_plug plug;
1752
1753	oldest_jif = jiffies;
1754	work->older_than_this = &oldest_jif;
1755
1756	blk_start_plug(&plug);
1757	spin_lock(&wb->list_lock);
1758	for (;;) {
1759	/*
1760	* Stop writeback when nr_pages has been consumed
1761	*/
1762	if (work->nr_pages <= 0)
1763	break;
1764
1765	/*
1766	* Background writeout and kupdate-style writeback may
1767	* run forever. Stop them if there is other work to do
1768	* so that e.g. sync can proceed. They'll be restarted
1769	* after the other works are all done.
1770	*/
1771	if ((work->for_background || work->for_kupdate) &&
1772	!list_empty(&wb->work_list))
1773	break;
1774
1775	/*
1776	* For background writeout, stop when we are below the
1777	* background dirty threshold
1778	*/
1779	if (work->for_background && !wb_over_bg_thresh(wb))
1780	break;
1781
1782	/*
1783	* Kupdate and background works are special and we want to
1784	* include all inodes that need writing. Livelock avoidance is
1785	* handled by these works yielding to any other work so we are
1786	* safe.
1787	*/
1788	if (work->for_kupdate) {
1789	oldest_jif = jiffies -
1790	msecs_to_jiffies(dirty_expire_interval * 10);
1791	} else if (work->for_background)
1792	oldest_jif = jiffies;
1793
1794	trace_writeback_start(wb, work);
1795	if (list_empty(&wb->b_io))
1796	queue_io(wb, work);
1797	if (work->sb)
1798	progress = writeback_sb_inodes(work->sb, wb, work);
1799	else
1800	progress = __writeback_inodes_wb(wb, work);
1801	trace_writeback_written(wb, work);
1802
1803	wb_update_bandwidth(wb, wb_start);
1804
1805	/*
1806	* Did we write something? Try for more
1807	*
1808	* Dirty inodes are moved to b_io for writeback in batches.
1809	* The completion of the current batch does not necessarily
1810	* mean the overall work is done. So we keep looping as long
1811	* as made some progress on cleaning pages or inodes.
1812	*/
1813	if (progress)
1814	continue;
1815	/*
1816	* No more inodes for IO, bail
1817	*/
1818	if (list_empty(&wb->b_more_io))
1819	break;
1820	/*
1821	* Nothing written. Wait for some inode to
1822	* become available for writeback. Otherwise
1823	* we'll just busyloop.
1824	*/
1825	if (!list_empty(&wb->b_more_io)) {
1826	trace_writeback_wait(wb, work);
1827	inode = wb_inode(wb->b_more_io.prev);
1828	spin_lock(&inode->i_lock);
1829	spin_unlock(&wb->list_lock);
1830	/* This function drops i_lock... */
1831	inode_sleep_on_writeback(inode);
1832	spin_lock(&wb->list_lock);
1833	}
1834	}
1835	spin_unlock(&wb->list_lock);
1836	blk_finish_plug(&plug);
1837
1838	return nr_pages - work->nr_pages;
1839	}
1840
1841	/*
1842	* Return the next wb_writeback_work struct that hasn't been processed yet.
1843	*/
1844	static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1845	{
1846	struct wb_writeback_work *work = NULL;
1847
1848	spin_lock_bh(&wb->work_lock);
1849	if (!list_empty(&wb->work_list)) {
1850	work = list_entry(wb->work_list.next,
1851	struct wb_writeback_work, list);
1852	list_del_init(&work->list);
1853	}
1854	spin_unlock_bh(&wb->work_lock);
1855	return work;
1856	}
1857
1858	/*
1859	* Add in the number of potentially dirty inodes, because each inode
1860	* write can dirty pagecache in the underlying blockdev.
1861	*/
1862	static unsigned long get_nr_dirty_pages(void)
1863	{
1864	return global_node_page_state(NR_FILE_DIRTY) +
1865	global_node_page_state(NR_UNSTABLE_NFS) +
1866	get_nr_dirty_inodes();
1867	}
1868
1869	static long wb_check_background_flush(struct bdi_writeback *wb)
1870	{
1871	if (wb_over_bg_thresh(wb)) {
1872
1873	struct wb_writeback_work work = {
1874	.nr_pages = LONG_MAX,
1875	.sync_mode = WB_SYNC_NONE,
1876	.for_background = 1,
1877	.range_cyclic = 1,
1878	.reason = WB_REASON_BACKGROUND,
1879	};
1880
1881	return wb_writeback(wb, &work);
1882	}
1883
1884	return 0;
1885	}
1886
1887	static long wb_check_old_data_flush(struct bdi_writeback *wb)
1888	{
1889	unsigned long expired;
1890	long nr_pages;
1891
1892	/*
1893	* When set to zero, disable periodic writeback
1894	*/
1895	if (!dirty_writeback_interval)
1896	return 0;
1897
1898	expired = wb->last_old_flush +
1899	msecs_to_jiffies(dirty_writeback_interval * 10);
1900	if (time_before(jiffies, expired))
1901	return 0;
1902
1903	wb->last_old_flush = jiffies;
1904	nr_pages = get_nr_dirty_pages();
1905
1906	if (nr_pages) {
1907	struct wb_writeback_work work = {
1908	.nr_pages = nr_pages,
1909	.sync_mode = WB_SYNC_NONE,
1910	.for_kupdate = 1,
1911	.range_cyclic = 1,
1912	.reason = WB_REASON_PERIODIC,
1913	};
1914
1915	return wb_writeback(wb, &work);
1916	}
1917
1918	return 0;
1919	}
1920
1921	/*
1922	* Retrieve work items and do the writeback they describe
1923	*/
1924	static long wb_do_writeback(struct bdi_writeback *wb)
1925	{
1926	struct wb_writeback_work *work;
1927	long wrote = 0;
1928
1929	set_bit(WB_writeback_running, &wb->state);
1930	while ((work = get_next_work_item(wb)) != NULL) {
1931	trace_writeback_exec(wb, work);
1932	wrote += wb_writeback(wb, work);
1933	finish_writeback_work(wb, work);
1934	}
1935
1936	/*
1937	* Check for periodic writeback, kupdated() style
1938	*/
1939	wrote += wb_check_old_data_flush(wb);
1940	wrote += wb_check_background_flush(wb);
1941	clear_bit(WB_writeback_running, &wb->state);
1942
1943	return wrote;
1944	}
1945
1946	/*
1947	* Handle writeback of dirty data for the device backed by this bdi. Also
1948	* reschedules periodically and does kupdated style flushing.
1949	*/
1950	void wb_workfn(struct work_struct *work)
1951	{
1952	struct bdi_writeback *wb = container_of(to_delayed_work(work),
1953	struct bdi_writeback, dwork);
1954	long pages_written;
1955
1956	set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1957	current->flags |= PF_SWAPWRITE;
1958
1959	if (likely(!current_is_workqueue_rescuer() ||
1960	!test_bit(WB_registered, &wb->state))) {
1961	/*
1962	* The normal path. Keep writing back @wb until its
1963	* work_list is empty. Note that this path is also taken
1964	* if @wb is shutting down even when we're running off the
1965	* rescuer as work_list needs to be drained.
1966	*/
1967	do {
1968	pages_written = wb_do_writeback(wb);
1969	trace_writeback_pages_written(pages_written);
1970	} while (!list_empty(&wb->work_list));
1971	} else {
1972	/*
1973	* bdi_wq can't get enough workers and we're running off
1974	* the emergency worker. Don't hog it. Hopefully, 1024 is
1975	* enough for efficient IO.
1976	*/
1977	pages_written = writeback_inodes_wb(wb, 1024,
1978	WB_REASON_FORKER_THREAD);
1979	trace_writeback_pages_written(pages_written);
1980	}
1981
1982	if (!list_empty(&wb->work_list))
1983	wb_wakeup(wb);
1984	else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1985	wb_wakeup_delayed(wb);
1986
1987	current->flags &= ~PF_SWAPWRITE;
1988	}
1989
1990	/*
1991	* Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1992	* the whole world.
1993	*/
1994	void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1995	{
1996	struct backing_dev_info *bdi;
1997
1998	/*
1999	* If we are expecting writeback progress we must submit plugged IO.
2000	*/
2001	if (blk_needs_flush_plug(current))
2002	blk_schedule_flush_plug(current);
2003
2004	if (!nr_pages)
2005	nr_pages = get_nr_dirty_pages();
2006
2007	rcu_read_lock();
2008	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2009	struct bdi_writeback *wb;
2010
2011	if (!bdi_has_dirty_io(bdi))
2012	continue;
2013
2014	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2015	wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
2016	false, reason);
2017	}
2018	rcu_read_unlock();
2019	}
2020
2021	/*
2022	* Wake up bdi's periodically to make sure dirtytime inodes gets
2023	* written back periodically. We deliberately do *not* check the
2024	* b_dirtytime list in wb_has_dirty_io(), since this would cause the
2025	* kernel to be constantly waking up once there are any dirtytime
2026	* inodes on the system. So instead we define a separate delayed work
2027	* function which gets called much more rarely. (By default, only
2028	* once every 12 hours.)
2029	*
2030	* If there is any other write activity going on in the file system,
2031	* this function won't be necessary. But if the only thing that has
2032	* happened on the file system is a dirtytime inode caused by an atime
2033	* update, we need this infrastructure below to make sure that inode
2034	* eventually gets pushed out to disk.
2035	*/
2036	static void wakeup_dirtytime_writeback(struct work_struct *w);
2037	static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2038
2039	static void wakeup_dirtytime_writeback(struct work_struct *w)
2040	{
2041	struct backing_dev_info *bdi;
2042
2043	rcu_read_lock();
2044	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2045	struct bdi_writeback *wb;
2046
2047	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2048	if (!list_empty(&wb->b_dirty_time))
2049	wb_wakeup(wb);
2050	}
2051	rcu_read_unlock();
2052	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2053	}
2054
2055	static int __init start_dirtytime_writeback(void)
2056	{
2057	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2058	return 0;
2059	}
2060	__initcall(start_dirtytime_writeback);
2061
2062	int dirtytime_interval_handler(struct ctl_table *table, int write,
2063	void __user *buffer, size_t *lenp, loff_t *ppos)
2064	{
2065	int ret;
2066
2067	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2068	if (ret == 0 && write)
2069	mod_delayed_work(system_wq, &dirtytime_work, 0);
2070	return ret;
2071	}
2072
2073	static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2074	{
2075	if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2076	struct dentry *dentry;
2077	const char *name = "?";
2078
2079	dentry = d_find_alias(inode);
2080	if (dentry) {
2081	spin_lock(&dentry->d_lock);
2082	name = (const char *) dentry->d_name.name;
2083	}
2084	printk(KERN_DEBUG
2085	"%s(%d): dirtied inode %lu (%s) on %s\n",
2086	current->comm, task_pid_nr(current), inode->i_ino,
2087	name, inode->i_sb->s_id);
2088	if (dentry) {
2089	spin_unlock(&dentry->d_lock);
2090	dput(dentry);
2091	}
2092	}
2093	}
2094
2095	/**
2096	* __mark_inode_dirty - internal function
2097	* @inode: inode to mark
2098	* @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2099	* Mark an inode as dirty. Callers should use mark_inode_dirty or
2100	* mark_inode_dirty_sync.
2101	*
2102	* Put the inode on the super block's dirty list.
2103	*
2104	* CAREFUL! We mark it dirty unconditionally, but move it onto the
2105	* dirty list only if it is hashed or if it refers to a blockdev.
2106	* If it was not hashed, it will never be added to the dirty list
2107	* even if it is later hashed, as it will have been marked dirty already.
2108	*
2109	* In short, make sure you hash any inodes _before_ you start marking
2110	* them dirty.
2111	*
2112	* Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2113	* the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2114	* the kernel-internal blockdev inode represents the dirtying time of the
2115	* blockdev's pages. This is why for I_DIRTY_PAGES we always use
2116	* page->mapping->host, so the page-dirtying time is recorded in the internal
2117	* blockdev inode.
2118	*/
2119	void __mark_inode_dirty(struct inode *inode, int flags)
2120	{
2121	#define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
2122	struct super_block *sb = inode->i_sb;
2123	int dirtytime;
2124
2125	trace_writeback_mark_inode_dirty(inode, flags);
2126
2127	/*
2128	* Don't do this for I_DIRTY_PAGES - that doesn't actually
2129	* dirty the inode itself
2130	*/
2131	if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
2132	trace_writeback_dirty_inode_start(inode, flags);
2133
2134	if (sb->s_op->dirty_inode)
2135	sb->s_op->dirty_inode(inode, flags);
2136
2137	trace_writeback_dirty_inode(inode, flags);
2138	}
2139	if (flags & I_DIRTY_INODE)
2140	flags &= ~I_DIRTY_TIME;
2141	dirtytime = flags & I_DIRTY_TIME;
2142
2143	/*
2144	* Paired with smp_mb() in __writeback_single_inode() for the
2145	* following lockless i_state test. See there for details.
2146	*/
2147	smp_mb();
2148
2149	if (((inode->i_state & flags) == flags) ||
2150	(dirtytime && (inode->i_state & I_DIRTY_INODE)))
2151	return;
2152
2153	if (unlikely(block_dump > 1))
2154	block_dump___mark_inode_dirty(inode);
2155
2156	spin_lock(&inode->i_lock);
2157	if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2158	goto out_unlock_inode;
2159	if ((inode->i_state & flags) != flags) {
2160	const int was_dirty = inode->i_state & I_DIRTY;
2161
2162	inode_attach_wb(inode, NULL);
2163
2164	if (flags & I_DIRTY_INODE)
2165	inode->i_state &= ~I_DIRTY_TIME;
2166	inode->i_state |= flags;
2167
2168	/*
2169	* If the inode is being synced, just update its dirty state.
2170	* The unlocker will place the inode on the appropriate
2171	* superblock list, based upon its state.
2172	*/
2173	if (inode->i_state & I_SYNC)
2174	goto out_unlock_inode;
2175
2176	/*
2177	* Only add valid (hashed) inodes to the superblock's
2178	* dirty list. Add blockdev inodes as well.
2179	*/
2180	if (!S_ISBLK(inode->i_mode)) {
2181	if (inode_unhashed(inode))
2182	goto out_unlock_inode;
2183	}
2184	if (inode->i_state & I_FREEING)
2185	goto out_unlock_inode;
2186
2187	/*
2188	* If the inode was already on b_dirty/b_io/b_more_io, don't
2189	* reposition it (that would break b_dirty time-ordering).
2190	*/
2191	if (!was_dirty) {
2192	struct bdi_writeback *wb;
2193	struct list_head *dirty_list;
2194	bool wakeup_bdi = false;
2195
2196	wb = locked_inode_to_wb_and_lock_list(inode);
2197
2198	WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2199	!test_bit(WB_registered, &wb->state),
2200	"bdi-%s not registered\n", wb->bdi->name);
2201
2202	inode->dirtied_when = jiffies;
2203	if (dirtytime)
2204	inode->dirtied_time_when = jiffies;
2205
2206	if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2207	dirty_list = &wb->b_dirty;
2208	else
2209	dirty_list = &wb->b_dirty_time;
2210
2211	wakeup_bdi = inode_io_list_move_locked(inode, wb,
2212	dirty_list);
2213
2214	spin_unlock(&wb->list_lock);
2215	trace_writeback_dirty_inode_enqueue(inode);
2216
2217	/*
2218	* If this is the first dirty inode for this bdi,
2219	* we have to wake-up the corresponding bdi thread
2220	* to make sure background write-back happens
2221	* later.
2222	*/
2223	if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2224	wb_wakeup_delayed(wb);
2225	return;
2226	}
2227	}
2228	out_unlock_inode:
2229	spin_unlock(&inode->i_lock);
2230
2231	#undef I_DIRTY_INODE
2232	}
2233	EXPORT_SYMBOL(__mark_inode_dirty);
2234
2235	/*
2236	* The @s_sync_lock is used to serialise concurrent sync operations
2237	* to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2238	* Concurrent callers will block on the s_sync_lock rather than doing contending
2239	* walks. The queueing maintains sync(2) required behaviour as all the IO that
2240	* has been issued up to the time this function is enter is guaranteed to be
2241	* completed by the time we have gained the lock and waited for all IO that is
2242	* in progress regardless of the order callers are granted the lock.
2243	*/
2244	static void wait_sb_inodes(struct super_block *sb)
2245	{
2246	LIST_HEAD(sync_list);
2247
2248	/*
2249	* We need to be protected against the filesystem going from
2250	* r/o to r/w or vice versa.
2251	*/
2252	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2253
2254	mutex_lock(&sb->s_sync_lock);
2255
2256	/*
2257	* Splice the writeback list onto a temporary list to avoid waiting on
2258	* inodes that have started writeback after this point.
2259	*
2260	* Use rcu_read_lock() to keep the inodes around until we have a
2261	* reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2262	* the local list because inodes can be dropped from either by writeback
2263	* completion.
2264	*/
2265	rcu_read_lock();
2266	spin_lock_irq(&sb->s_inode_wblist_lock);
2267	list_splice_init(&sb->s_inodes_wb, &sync_list);
2268
2269	/*
2270	* Data integrity sync. Must wait for all pages under writeback, because
2271	* there may have been pages dirtied before our sync call, but which had
2272	* writeout started before we write it out. In which case, the inode
2273	* may not be on the dirty list, but we still have to wait for that
2274	* writeout.
2275	*/
2276	while (!list_empty(&sync_list)) {
2277	struct inode *inode = list_first_entry(&sync_list, struct inode,
2278	i_wb_list);
2279	struct address_space *mapping = inode->i_mapping;
2280
2281	/*
2282	* Move each inode back to the wb list before we drop the lock
2283	* to preserve consistency between i_wb_list and the mapping
2284	* writeback tag. Writeback completion is responsible to remove
2285	* the inode from either list once the writeback tag is cleared.
2286	*/
2287	list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2288
2289	/*
2290	* The mapping can appear untagged while still on-list since we
2291	* do not have the mapping lock. Skip it here, wb completion
2292	* will remove it.
2293	*/
2294	if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2295	continue;
2296
2297	spin_unlock_irq(&sb->s_inode_wblist_lock);
2298
2299	spin_lock(&inode->i_lock);
2300	if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2301	spin_unlock(&inode->i_lock);
2302
2303	spin_lock_irq(&sb->s_inode_wblist_lock);
2304	continue;
2305	}
2306	__iget(inode);
2307	spin_unlock(&inode->i_lock);
2308	rcu_read_unlock();
2309
2310	/*
2311	* We keep the error status of individual mapping so that
2312	* applications can catch the writeback error using fsync(2).
2313	* See filemap_fdatawait_keep_errors() for details.
2314	*/
2315	filemap_fdatawait_keep_errors(mapping);
2316
2317	cond_resched();
2318
2319	iput(inode);
2320
2321	rcu_read_lock();
2322	spin_lock_irq(&sb->s_inode_wblist_lock);
2323	}
2324	spin_unlock_irq(&sb->s_inode_wblist_lock);
2325	rcu_read_unlock();
2326	mutex_unlock(&sb->s_sync_lock);
2327	}
2328
2329	static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2330	enum wb_reason reason, bool skip_if_busy)
2331	{
2332	DEFINE_WB_COMPLETION_ONSTACK(done);
2333	struct wb_writeback_work work = {
2334	.sb = sb,
2335	.sync_mode = WB_SYNC_NONE,
2336	.tagged_writepages = 1,
2337	.done = &done,
2338	.nr_pages = nr,
2339	.reason = reason,
2340	};
2341	struct backing_dev_info *bdi = sb->s_bdi;
2342
2343	if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2344	return;
2345	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2346
2347	bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2348	wb_wait_for_completion(bdi, &done);
2349	}
2350
2351	/**
2352	* writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2353	* @sb: the superblock
2354	* @nr: the number of pages to write
2355	* @reason: reason why some writeback work initiated
2356	*
2357	* Start writeback on some inodes on this super_block. No guarantees are made
2358	* on how many (if any) will be written, and this function does not wait
2359	* for IO completion of submitted IO.
2360	*/
2361	void writeback_inodes_sb_nr(struct super_block *sb,
2362	unsigned long nr,
2363	enum wb_reason reason)
2364	{
2365	__writeback_inodes_sb_nr(sb, nr, reason, false);
2366	}
2367	EXPORT_SYMBOL(writeback_inodes_sb_nr);
2368
2369	/**
2370	* writeback_inodes_sb - writeback dirty inodes from given super_block
2371	* @sb: the superblock
2372	* @reason: reason why some writeback work was initiated
2373	*
2374	* Start writeback on some inodes on this super_block. No guarantees are made
2375	* on how many (if any) will be written, and this function does not wait
2376	* for IO completion of submitted IO.
2377	*/
2378	void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2379	{
2380	return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2381	}
2382	EXPORT_SYMBOL(writeback_inodes_sb);
2383
2384	/**
2385	* try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2386	* @sb: the superblock
2387	* @nr: the number of pages to write
2388	* @reason: the reason of writeback
2389	*
2390	* Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2391	* Returns 1 if writeback was started, 0 if not.
2392	*/
2393	bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2394	enum wb_reason reason)
2395	{
2396	if (!down_read_trylock(&sb->s_umount))
2397	return false;
2398
2399	__writeback_inodes_sb_nr(sb, nr, reason, true);
2400	up_read(&sb->s_umount);
2401	return true;
2402	}
2403	EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2404
2405	/**
2406	* try_to_writeback_inodes_sb - try to start writeback if none underway
2407	* @sb: the superblock
2408	* @reason: reason why some writeback work was initiated
2409	*
2410	* Implement by try_to_writeback_inodes_sb_nr()
2411	* Returns 1 if writeback was started, 0 if not.
2412	*/
2413	bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2414	{
2415	return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2416	}
2417	EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2418
2419	/**
2420	* sync_inodes_sb - sync sb inode pages
2421	* @sb: the superblock
2422	*
2423	* This function writes and waits on any dirty inode belonging to this
2424	* super_block.
2425	*/
2426	void sync_inodes_sb(struct super_block *sb)
2427	{
2428	DEFINE_WB_COMPLETION_ONSTACK(done);
2429	struct wb_writeback_work work = {
2430	.sb = sb,
2431	.sync_mode = WB_SYNC_ALL,
2432	.nr_pages = LONG_MAX,
2433	.range_cyclic = 0,
2434	.done = &done,
2435	.reason = WB_REASON_SYNC,
2436	.for_sync = 1,
2437	};
2438	struct backing_dev_info *bdi = sb->s_bdi;
2439
2440	/*
2441	* Can't skip on !bdi_has_dirty() because we should wait for !dirty
2442	* inodes under writeback and I_DIRTY_TIME inodes ignored by
2443	* bdi_has_dirty() need to be written out too.
2444	*/
2445	if (bdi == &noop_backing_dev_info)
2446	return;
2447	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2448
2449	/* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2450	bdi_down_write_wb_switch_rwsem(bdi);
2451	bdi_split_work_to_wbs(bdi, &work, false);
2452	wb_wait_for_completion(bdi, &done);
2453	bdi_up_write_wb_switch_rwsem(bdi);
2454
2455	wait_sb_inodes(sb);
2456	}
2457	EXPORT_SYMBOL(sync_inodes_sb);
2458
2459	/**
2460	* write_inode_now - write an inode to disk
2461	* @inode: inode to write to disk
2462	* @sync: whether the write should be synchronous or not
2463	*
2464	* This function commits an inode to disk immediately if it is dirty. This is
2465	* primarily needed by knfsd.
2466	*
2467	* The caller must either have a ref on the inode or must have set I_WILL_FREE.
2468	*/
2469	int write_inode_now(struct inode *inode, int sync)
2470	{
2471	struct writeback_control wbc = {
2472	.nr_to_write = LONG_MAX,
2473	.sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2474	.range_start = 0,
2475	.range_end = LLONG_MAX,
2476	};
2477
2478	if (!mapping_cap_writeback_dirty(inode->i_mapping))
2479	wbc.nr_to_write = 0;
2480
2481	might_sleep();
2482	return writeback_single_inode(inode, &wbc);
2483	}
2484	EXPORT_SYMBOL(write_inode_now);
2485
2486	/**
2487	* sync_inode - write an inode and its pages to disk.
2488	* @inode: the inode to sync
2489	* @wbc: controls the writeback mode
2490	*
2491	* sync_inode() will write an inode and its pages to disk. It will also
2492	* correctly update the inode on its superblock's dirty inode lists and will
2493	* update inode->i_state.
2494	*
2495	* The caller must have a ref on the inode.
2496	*/
2497	int sync_inode(struct inode *inode, struct writeback_control *wbc)
2498	{
2499	return writeback_single_inode(inode, wbc);
2500	}
2501	EXPORT_SYMBOL(sync_inode);
2502
2503	/**
2504	* sync_inode_metadata - write an inode to disk
2505	* @inode: the inode to sync
2506	* @wait: wait for I/O to complete.
2507	*
2508	* Write an inode to disk and adjust its dirty state after completion.
2509	*
2510	* Note: only writes the actual inode, no associated data or other metadata.
2511	*/
2512	int sync_inode_metadata(struct inode *inode, int wait)
2513	{
2514	struct writeback_control wbc = {
2515	.sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2516	.nr_to_write = 0, /* metadata-only */
2517	};
2518
2519	return sync_inode(inode, &wbc);
2520	}
2521	EXPORT_SYMBOL(sync_inode_metadata);
2522