linux - fs/fs-writeback.c

2851 lines · 82.1 KiB

1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* fs/fs-writeback.c
4	*
5	* Copyright (C) 2002, Linus Torvalds.
6	*
7	* Contains all the functions related to writing back and waiting
8	* upon dirty inodes against superblocks, and writing back dirty
9	* pages against inodes. ie: data writeback. Writeout of the
10	* inode itself is not handled here.
11	*
12	* 10Apr2002 Andrew Morton
13	* Split out of fs/inode.c
14	* Additions for address_space-based writeback
15	*/
16
17	#include <linux/kernel.h>
18	#include <linux/export.h>
19	#include <linux/spinlock.h>
20	#include <linux/slab.h>
21	#include <linux/sched.h>
22	#include <linux/fs.h>
23	#include <linux/mm.h>
24	#include <linux/pagemap.h>
25	#include <linux/kthread.h>
26	#include <linux/writeback.h>
27	#include <linux/blkdev.h>
28	#include <linux/backing-dev.h>
29	#include <linux/tracepoint.h>
30	#include <linux/device.h>
31	#include <linux/memcontrol.h>
32	#include "internal.h"
33
34	/*
35	* 4MB minimal write chunk size
36	*/
37	#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38
39	/*
40	* Passed into wb_writeback(), essentially a subset of writeback_control
41	*/
42	struct wb_writeback_work {
43	long nr_pages;
44	struct super_block *sb;
45	enum writeback_sync_modes sync_mode;
46	unsigned int tagged_writepages:1;
47	unsigned int for_kupdate:1;
48	unsigned int range_cyclic:1;
49	unsigned int for_background:1;
50	unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
51	unsigned int auto_free:1; /* free on completion */
52	enum wb_reason reason; /* why was writeback initiated? */
53
54	struct list_head list; /* pending work list */
55	struct wb_completion done; / set if the caller waits */
56	};
57
58	/*
59	* If an inode is constantly having its pages dirtied, but then the
60	* updates stop dirtytime_expire_interval seconds in the past, it's
61	* possible for the worst case time between when an inode has its
62	* timestamps updated and when they finally get written out to be two
63	* dirtytime_expire_intervals. We set the default to 12 hours (in
64	* seconds), which means most of the time inodes will have their
65	* timestamps written to disk after 12 hours, but in the worst case a
66	* few inodes might not their timestamps updated for 24 hours.
67	*/
68	unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
70	static inline struct inode wb_inode(struct list_head head)
71	{
72	return list_entry(head, struct inode, i_io_list);
73	}
74
75	/*
76	* Include the creation of the trace points after defining the
77	* wb_writeback_work structure and inline functions so that the definition
78	* remains local to this file.
79	*/
80	#define CREATE_TRACE_POINTS
81	#include <trace/events/writeback.h>
82
83	EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
85	static bool wb_io_lists_populated(struct bdi_writeback *wb)
86	{
87	if (wb_has_dirty_io(wb)) {
88	return false;
89	} else {
90	set_bit(WB_has_dirty_io, &wb->state);
91	WARN_ON_ONCE(!wb->avg_write_bandwidth);
92	atomic_long_add(wb->avg_write_bandwidth,
93	&wb->bdi->tot_write_bandwidth);
94	return true;
95	}
96	}
97
98	static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99	{
100	if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101	list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102	clear_bit(WB_has_dirty_io, &wb->state);
103	WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104	&wb->bdi->tot_write_bandwidth) < 0);
105	}
106	}
107
108	/**
109	* inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110	* @inode: inode to be moved
111	* @wb: target bdi_writeback
112	* @head: one of @wb->b_{dirty\|io\|more_io\|dirty_time}
113	*
114	* Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115	* Returns %true if @inode is the first occupant of the !dirty_time IO
116	* lists; otherwise, %false.
117	*/
118	static bool inode_io_list_move_locked(struct inode *inode,
119	struct bdi_writeback *wb,
120	struct list_head *head)
121	{
122	assert_spin_locked(&wb->list_lock);
123	assert_spin_locked(&inode->i_lock);
124	WARN_ON_ONCE(inode->i_state & I_FREEING);
125
126	list_move(&inode->i_io_list, head);
127
128	/* dirty_time doesn't count as dirty_io until expiration */
129	if (head != &wb->b_dirty_time)
130	return wb_io_lists_populated(wb);
131
132	wb_io_lists_depopulated(wb);
133	return false;
134	}
135
136	static void wb_wakeup(struct bdi_writeback *wb)
137	{
138	spin_lock_irq(&wb->work_lock);
139	if (test_bit(WB_registered, &wb->state))
140	mod_delayed_work(bdi_wq, &wb->dwork, 0);
141	spin_unlock_irq(&wb->work_lock);
142	}
143
144	/*
145	* This function is used when the first inode for this wb is marked dirty. It
146	* wakes-up the corresponding bdi thread which should then take care of the
147	* periodic background write-out of dirty inodes. Since the write-out would
148	* starts only 'dirty_writeback_interval' centisecs from now anyway, we just
149	* set up a timer which wakes the bdi thread up later.
150	*
151	* Note, we wouldn't bother setting up the timer, but this function is on the
152	* fast-path (used by '__mark_inode_dirty()'), so we save few context switches
153	* by delaying the wake-up.
154	*
155	* We have to be careful not to postpone flush work if it is scheduled for
156	* earlier. Thus we use queue_delayed_work().
157	*/
158	static void wb_wakeup_delayed(struct bdi_writeback *wb)
159	{
160	unsigned long timeout;
161
162	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
163	spin_lock_irq(&wb->work_lock);
164	if (test_bit(WB_registered, &wb->state))
165	queue_delayed_work(bdi_wq, &wb->dwork, timeout);
166	spin_unlock_irq(&wb->work_lock);
167	}
168
169	static void finish_writeback_work(struct wb_writeback_work *work)
170	{
171	struct wb_completion *done = work->done;
172
173	if (work->auto_free)
174	kfree(work);
175	if (done) {
176	wait_queue_head_t *waitq = done->waitq;
177
178	/* @done can't be accessed after the following dec */
179	if (atomic_dec_and_test(&done->cnt))
180	wake_up_all(waitq);
181	}
182	}
183
184	static void wb_queue_work(struct bdi_writeback *wb,
185	struct wb_writeback_work *work)
186	{
187	trace_writeback_queue(wb, work);
188
189	if (work->done)
190	atomic_inc(&work->done->cnt);
191
192	spin_lock_irq(&wb->work_lock);
193
194	if (test_bit(WB_registered, &wb->state)) {
195	list_add_tail(&work->list, &wb->work_list);
196	mod_delayed_work(bdi_wq, &wb->dwork, 0);
197	} else
198	finish_writeback_work(work);
199
200	spin_unlock_irq(&wb->work_lock);
201	}
202
203	/**
204	* wb_wait_for_completion - wait for completion of bdi_writeback_works
205	* @done: target wb_completion
206	*
207	* Wait for one or more work items issued to @bdi with their ->done field
208	* set to @done, which should have been initialized with
209	* DEFINE_WB_COMPLETION(). This function returns after all such work items
210	* are completed. Work items which are waited upon aren't freed
211	* automatically on completion.
212	*/
213	void wb_wait_for_completion(struct wb_completion *done)
214	{
215	atomic_dec(&done->cnt); /* put down the initial count */
216	wait_event(*done->waitq, !atomic_read(&done->cnt));
217	}
218
219	#ifdef CONFIG_CGROUP_WRITEBACK
220
221	/*
222	* Parameters for foreign inode detection, see wbc_detach_inode() to see
223	* how they're used.
224	*
225	* These paramters are inherently heuristical as the detection target
226	* itself is fuzzy. All we want to do is detaching an inode from the
227	* current owner if it's being written to by some other cgroups too much.
228	*
229	* The current cgroup writeback is built on the assumption that multiple
230	* cgroups writing to the same inode concurrently is very rare and a mode
231	* of operation which isn't well supported. As such, the goal is not
232	* taking too long when a different cgroup takes over an inode while
233	* avoiding too aggressive flip-flops from occasional foreign writes.
234	*
235	* We record, very roughly, 2s worth of IO time history and if more than
236	* half of that is foreign, trigger the switch. The recording is quantized
237	* to 16 slots. To avoid tiny writes from swinging the decision too much,
238	* writes smaller than 1/8 of avg size are ignored.
239	*/
240	#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
241	#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
242	#define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
243	#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
244
245	#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
246	#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
247	/* each slot's duration is 2s / 16 */
248	#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
249	/* if foreign slots >= 8, switch */
250	#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
251	/* one round can affect upto 5 slots */
252	#define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
253
254	/*
255	* Maximum inodes per isw. A specific value has been chosen to make
256	* struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
257	*/
258	#define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
259	/ sizeof(struct inode *))
260
261	static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
262	static struct workqueue_struct *isw_wq;
263
264	void __inode_attach_wb(struct inode inode, struct folio folio)
265	{
266	struct backing_dev_info *bdi = inode_to_bdi(inode);
267	struct bdi_writeback *wb = NULL;
268
269	if (inode_cgwb_enabled(inode)) {
270	struct cgroup_subsys_state *memcg_css;
271
272	if (folio) {
273	memcg_css = mem_cgroup_css_from_folio(folio);
274	wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
275	} else {
276	/* must pin memcg_css, see wb_get_create() */
277	memcg_css = task_get_css(current, memory_cgrp_id);
278	wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
279	css_put(memcg_css);
280	}
281	}
282
283	if (!wb)
284	wb = &bdi->wb;
285
286	/*
287	* There may be multiple instances of this function racing to
288	* update the same inode. Use cmpxchg() to tell the winner.
289	*/
290	if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
291	wb_put(wb);
292	}
293	EXPORT_SYMBOL_GPL(__inode_attach_wb);
294
295	/**
296	* inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
297	* @inode: inode of interest with i_lock held
298	* @wb: target bdi_writeback
299	*
300	* Remove the inode from wb's io lists and if necessarily put onto b_attached
301	* list. Only inodes attached to cgwb's are kept on this list.
302	*/
303	static void inode_cgwb_move_to_attached(struct inode *inode,
304	struct bdi_writeback *wb)
305	{
306	assert_spin_locked(&wb->list_lock);
307	assert_spin_locked(&inode->i_lock);
308	WARN_ON_ONCE(inode->i_state & I_FREEING);
309
310	inode->i_state &= ~I_SYNC_QUEUED;
311	if (wb != &wb->bdi->wb)
312	list_move(&inode->i_io_list, &wb->b_attached);
313	else
314	list_del_init(&inode->i_io_list);
315	wb_io_lists_depopulated(wb);
316	}
317
318	/**
319	* locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
320	* @inode: inode of interest with i_lock held
321	*
322	* Returns @inode's wb with its list_lock held. @inode->i_lock must be
323	* held on entry and is released on return. The returned wb is guaranteed
324	* to stay @inode's associated wb until its list_lock is released.
325	*/
326	static struct bdi_writeback *
327	locked_inode_to_wb_and_lock_list(struct inode *inode)
328	__releases(&inode->i_lock)
329	__acquires(&wb->list_lock)
330	{
331	while (true) {
332	struct bdi_writeback *wb = inode_to_wb(inode);
333
334	/*
335	* inode_to_wb() association is protected by both
336	* @inode->i_lock and @wb->list_lock but list_lock nests
337	* outside i_lock. Drop i_lock and verify that the
338	* association hasn't changed after acquiring list_lock.
339	*/
340	wb_get(wb);
341	spin_unlock(&inode->i_lock);
342	spin_lock(&wb->list_lock);
343
344	/* i_wb may have changed inbetween, can't use inode_to_wb() */
345	if (likely(wb == inode->i_wb)) {
346	wb_put(wb); /* @inode already has ref */
347	return wb;
348	}
349
350	spin_unlock(&wb->list_lock);
351	wb_put(wb);
352	cpu_relax();
353	spin_lock(&inode->i_lock);
354	}
355	}
356
357	/**
358	* inode_to_wb_and_lock_list - determine an inode's wb and lock it
359	* @inode: inode of interest
360	*
361	* Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
362	* on entry.
363	*/
364	static struct bdi_writeback inode_to_wb_and_lock_list(struct inode inode)
365	__acquires(&wb->list_lock)
366	{
367	spin_lock(&inode->i_lock);
368	return locked_inode_to_wb_and_lock_list(inode);
369	}
370
371	struct inode_switch_wbs_context {
372	struct rcu_work work;
373
374	/*
375	* Multiple inodes can be switched at once. The switching procedure
376	* consists of two parts, separated by a RCU grace period. To make
377	* sure that the second part is executed for each inode gone through
378	* the first part, all inode pointers are placed into a NULL-terminated
379	* array embedded into struct inode_switch_wbs_context. Otherwise
380	* an inode could be left in a non-consistent state.
381	*/
382	struct bdi_writeback *new_wb;
383	struct inode *inodes[];
384	};
385
386	static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
387	{
388	down_write(&bdi->wb_switch_rwsem);
389	}
390
391	static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
392	{
393	up_write(&bdi->wb_switch_rwsem);
394	}
395
396	static bool inode_do_switch_wbs(struct inode *inode,
397	struct bdi_writeback *old_wb,
398	struct bdi_writeback *new_wb)
399	{
400	struct address_space *mapping = inode->i_mapping;
401	XA_STATE(xas, &mapping->i_pages, 0);
402	struct folio *folio;
403	bool switched = false;
404
405	spin_lock(&inode->i_lock);
406	xa_lock_irq(&mapping->i_pages);
407
408	/*
409	* Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
410	* path owns the inode and we shouldn't modify ->i_io_list.
411	*/
412	if (unlikely(inode->i_state & (I_FREEING \| I_WILL_FREE)))
413	goto skip_switch;
414
415	trace_inode_switch_wbs(inode, old_wb, new_wb);
416
417	/*
418	* Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
419	* to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
420	* folios actually under writeback.
421	*/
422	xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
423	if (folio_test_dirty(folio)) {
424	long nr = folio_nr_pages(folio);
425	wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
426	wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
427	}
428	}
429
430	xas_set(&xas, 0);
431	xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
432	long nr = folio_nr_pages(folio);
433	WARN_ON_ONCE(!folio_test_writeback(folio));
434	wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
435	wb_stat_mod(new_wb, WB_WRITEBACK, nr);
436	}
437
438	if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
439	atomic_dec(&old_wb->writeback_inodes);
440	atomic_inc(&new_wb->writeback_inodes);
441	}
442
443	wb_get(new_wb);
444
445	/*
446	* Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
447	* the specific list @inode was on is ignored and the @inode is put on
448	* ->b_dirty which is always correct including from ->b_dirty_time.
449	* The transfer preserves @inode->dirtied_when ordering. If the @inode
450	* was clean, it means it was on the b_attached list, so move it onto
451	* the b_attached list of @new_wb.
452	*/
453	if (!list_empty(&inode->i_io_list)) {
454	inode->i_wb = new_wb;
455
456	if (inode->i_state & I_DIRTY_ALL) {
457	struct inode *pos;
458
459	list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
460	if (time_after_eq(inode->dirtied_when,
461	pos->dirtied_when))
462	break;
463	inode_io_list_move_locked(inode, new_wb,
464	pos->i_io_list.prev);
465	} else {
466	inode_cgwb_move_to_attached(inode, new_wb);
467	}
468	} else {
469	inode->i_wb = new_wb;
470	}
471
472	/* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
473	inode->i_wb_frn_winner = 0;
474	inode->i_wb_frn_avg_time = 0;
475	inode->i_wb_frn_history = 0;
476	switched = true;
477	skip_switch:
478	/*
479	* Paired with load_acquire in unlocked_inode_to_wb_begin() and
480	* ensures that the new wb is visible if they see !I_WB_SWITCH.
481	*/
482	smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
483
484	xa_unlock_irq(&mapping->i_pages);
485	spin_unlock(&inode->i_lock);
486
487	return switched;
488	}
489
490	static void inode_switch_wbs_work_fn(struct work_struct *work)
491	{
492	struct inode_switch_wbs_context *isw =
493	container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
494	struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
495	struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
496	struct bdi_writeback *new_wb = isw->new_wb;
497	unsigned long nr_switched = 0;
498	struct inode **inodep;
499
500	/*
501	* If @inode switches cgwb membership while sync_inodes_sb() is
502	* being issued, sync_inodes_sb() might miss it. Synchronize.
503	*/
504	down_read(&bdi->wb_switch_rwsem);
505
506	/*
507	* By the time control reaches here, RCU grace period has passed
508	* since I_WB_SWITCH assertion and all wb stat update transactions
509	* between unlocked_inode_to_wb_begin/end() are guaranteed to be
510	* synchronizing against the i_pages lock.
511	*
512	* Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
513	* gives us exclusion against all wb related operations on @inode
514	* including IO list manipulations and stat updates.
515	*/
516	if (old_wb < new_wb) {
517	spin_lock(&old_wb->list_lock);
518	spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
519	} else {
520	spin_lock(&new_wb->list_lock);
521	spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
522	}
523
524	for (inodep = isw->inodes; *inodep; inodep++) {
525	WARN_ON_ONCE((*inodep)->i_wb != old_wb);
526	if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
527	nr_switched++;
528	}
529
530	spin_unlock(&new_wb->list_lock);
531	spin_unlock(&old_wb->list_lock);
532
533	up_read(&bdi->wb_switch_rwsem);
534
535	if (nr_switched) {
536	wb_wakeup(new_wb);
537	wb_put_many(old_wb, nr_switched);
538	}
539
540	for (inodep = isw->inodes; *inodep; inodep++)
541	iput(*inodep);
542	wb_put(new_wb);
543	kfree(isw);
544	atomic_dec(&isw_nr_in_flight);
545	}
546
547	static bool inode_prepare_wbs_switch(struct inode *inode,
548	struct bdi_writeback *new_wb)
549	{
550	/*
551	* Paired with smp_mb() in cgroup_writeback_umount().
552	* isw_nr_in_flight must be increased before checking SB_ACTIVE and
553	* grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
554	* in cgroup_writeback_umount() and the isw_wq will be not flushed.
555	*/
556	smp_mb();
557
558	if (IS_DAX(inode))
559	return false;
560
561	/* while holding I_WB_SWITCH, no one else can update the association */
562	spin_lock(&inode->i_lock);
563	if (!(inode->i_sb->s_flags & SB_ACTIVE) \|\|
564	inode->i_state & (I_WB_SWITCH \| I_FREEING \| I_WILL_FREE) \|\|
565	inode_to_wb(inode) == new_wb) {
566	spin_unlock(&inode->i_lock);
567	return false;
568	}
569	inode->i_state \|= I_WB_SWITCH;
570	__iget(inode);
571	spin_unlock(&inode->i_lock);
572
573	return true;
574	}
575
576	/**
577	* inode_switch_wbs - change the wb association of an inode
578	* @inode: target inode
579	* @new_wb_id: ID of the new wb
580	*
581	* Switch @inode's wb association to the wb identified by @new_wb_id. The
582	* switching is performed asynchronously and may fail silently.
583	*/
584	static void inode_switch_wbs(struct inode *inode, int new_wb_id)
585	{
586	struct backing_dev_info *bdi = inode_to_bdi(inode);
587	struct cgroup_subsys_state *memcg_css;
588	struct inode_switch_wbs_context *isw;
589
590	/* noop if seems to be already in progress */
591	if (inode->i_state & I_WB_SWITCH)
592	return;
593
594	/* avoid queueing a new switch if too many are already in flight */
595	if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
596	return;
597
598	isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
599	if (!isw)
600	return;
601
602	atomic_inc(&isw_nr_in_flight);
603
604	/* find and pin the new wb */
605	rcu_read_lock();
606	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
607	if (memcg_css && !css_tryget(memcg_css))
608	memcg_css = NULL;
609	rcu_read_unlock();
610	if (!memcg_css)
611	goto out_free;
612
613	isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
614	css_put(memcg_css);
615	if (!isw->new_wb)
616	goto out_free;
617
618	if (!inode_prepare_wbs_switch(inode, isw->new_wb))
619	goto out_free;
620
621	isw->inodes[0] = inode;
622
623	/*
624	* In addition to synchronizing among switchers, I_WB_SWITCH tells
625	* the RCU protected stat update paths to grab the i_page
626	* lock so that stat transfer can synchronize against them.
627	* Let's continue after I_WB_SWITCH is guaranteed to be visible.
628	*/
629	INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
630	queue_rcu_work(isw_wq, &isw->work);
631	return;
632
633	out_free:
634	atomic_dec(&isw_nr_in_flight);
635	if (isw->new_wb)
636	wb_put(isw->new_wb);
637	kfree(isw);
638	}
639
640	static bool isw_prepare_wbs_switch(struct inode_switch_wbs_context *isw,
641	struct list_head list, int nr)
642	{
643	struct inode *inode;
644
645	list_for_each_entry(inode, list, i_io_list) {
646	if (!inode_prepare_wbs_switch(inode, isw->new_wb))
647	continue;
648
649	isw->inodes[*nr] = inode;
650	(*nr)++;
651
652	if (*nr >= WB_MAX_INODES_PER_ISW - 1)
653	return true;
654	}
655	return false;
656	}
657
658	/**
659	* cleanup_offline_cgwb - detach associated inodes
660	* @wb: target wb
661	*
662	* Switch all inodes attached to @wb to a nearest living ancestor's wb in order
663	* to eventually release the dying @wb. Returns %true if not all inodes were
664	* switched and the function has to be restarted.
665	*/
666	bool cleanup_offline_cgwb(struct bdi_writeback *wb)
667	{
668	struct cgroup_subsys_state *memcg_css;
669	struct inode_switch_wbs_context *isw;
670	int nr;
671	bool restart = false;
672
673	isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
674	GFP_KERNEL);
675	if (!isw)
676	return restart;
677
678	atomic_inc(&isw_nr_in_flight);
679
680	for (memcg_css = wb->memcg_css->parent; memcg_css;
681	memcg_css = memcg_css->parent) {
682	isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
683	if (isw->new_wb)
684	break;
685	}
686	if (unlikely(!isw->new_wb))
687	isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
688
689	nr = 0;
690	spin_lock(&wb->list_lock);
691	/*
692	* In addition to the inodes that have completed writeback, also switch
693	* cgwbs for those inodes only with dirty timestamps. Otherwise, those
694	* inodes won't be written back for a long time when lazytime is
695	* enabled, and thus pinning the dying cgwbs. It won't break the
696	* bandwidth restrictions, as writeback of inode metadata is not
697	* accounted for.
698	*/
699	restart = isw_prepare_wbs_switch(isw, &wb->b_attached, &nr);
700	if (!restart)
701	restart = isw_prepare_wbs_switch(isw, &wb->b_dirty_time, &nr);
702	spin_unlock(&wb->list_lock);
703
704	/* no attached inodes? bail out */
705	if (nr == 0) {
706	atomic_dec(&isw_nr_in_flight);
707	wb_put(isw->new_wb);
708	kfree(isw);
709	return restart;
710	}
711
712	/*
713	* In addition to synchronizing among switchers, I_WB_SWITCH tells
714	* the RCU protected stat update paths to grab the i_page
715	* lock so that stat transfer can synchronize against them.
716	* Let's continue after I_WB_SWITCH is guaranteed to be visible.
717	*/
718	INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
719	queue_rcu_work(isw_wq, &isw->work);
720
721	return restart;
722	}
723
724	/**
725	* wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
726	* @wbc: writeback_control of interest
727	* @inode: target inode
728	*
729	* @inode is locked and about to be written back under the control of @wbc.
730	* Record @inode's writeback context into @wbc and unlock the i_lock. On
731	* writeback completion, wbc_detach_inode() should be called. This is used
732	* to track the cgroup writeback context.
733	*/
734	void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
735	struct inode *inode)
736	{
737	if (!inode_cgwb_enabled(inode)) {
738	spin_unlock(&inode->i_lock);
739	return;
740	}
741
742	wbc->wb = inode_to_wb(inode);
743	wbc->inode = inode;
744
745	wbc->wb_id = wbc->wb->memcg_css->id;
746	wbc->wb_lcand_id = inode->i_wb_frn_winner;
747	wbc->wb_tcand_id = 0;
748	wbc->wb_bytes = 0;
749	wbc->wb_lcand_bytes = 0;
750	wbc->wb_tcand_bytes = 0;
751
752	wb_get(wbc->wb);
753	spin_unlock(&inode->i_lock);
754
755	/*
756	* A dying wb indicates that either the blkcg associated with the
757	* memcg changed or the associated memcg is dying. In the first
758	* case, a replacement wb should already be available and we should
759	* refresh the wb immediately. In the second case, trying to
760	* refresh will keep failing.
761	*/
762	if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
763	inode_switch_wbs(inode, wbc->wb_id);
764	}
765	EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
766
767	/**
768	* wbc_detach_inode - disassociate wbc from inode and perform foreign detection
769	* @wbc: writeback_control of the just finished writeback
770	*
771	* To be called after a writeback attempt of an inode finishes and undoes
772	* wbc_attach_and_unlock_inode(). Can be called under any context.
773	*
774	* As concurrent write sharing of an inode is expected to be very rare and
775	* memcg only tracks page ownership on first-use basis severely confining
776	* the usefulness of such sharing, cgroup writeback tracks ownership
777	* per-inode. While the support for concurrent write sharing of an inode
778	* is deemed unnecessary, an inode being written to by different cgroups at
779	* different points in time is a lot more common, and, more importantly,
780	* charging only by first-use can too readily lead to grossly incorrect
781	* behaviors (single foreign page can lead to gigabytes of writeback to be
782	* incorrectly attributed).
783	*
784	* To resolve this issue, cgroup writeback detects the majority dirtier of
785	* an inode and transfers the ownership to it. To avoid unnecessary
786	* oscillation, the detection mechanism keeps track of history and gives
787	* out the switch verdict only if the foreign usage pattern is stable over
788	* a certain amount of time and/or writeback attempts.
789	*
790	* On each writeback attempt, @wbc tries to detect the majority writer
791	* using Boyer-Moore majority vote algorithm. In addition to the byte
792	* count from the majority voting, it also counts the bytes written for the
793	* current wb and the last round's winner wb (max of last round's current
794	* wb, the winner from two rounds ago, and the last round's majority
795	* candidate). Keeping track of the historical winner helps the algorithm
796	* to semi-reliably detect the most active writer even when it's not the
797	* absolute majority.
798	*
799	* Once the winner of the round is determined, whether the winner is
800	* foreign or not and how much IO time the round consumed is recorded in
801	* inode->i_wb_frn_history. If the amount of recorded foreign IO time is
802	* over a certain threshold, the switch verdict is given.
803	*/
804	void wbc_detach_inode(struct writeback_control *wbc)
805	{
806	struct bdi_writeback *wb = wbc->wb;
807	struct inode *inode = wbc->inode;
808	unsigned long avg_time, max_bytes, max_time;
809	u16 history;
810	int max_id;
811
812	if (!wb)
813	return;
814
815	history = inode->i_wb_frn_history;
816	avg_time = inode->i_wb_frn_avg_time;
817
818	/* pick the winner of this round */
819	if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
820	wbc->wb_bytes >= wbc->wb_tcand_bytes) {
821	max_id = wbc->wb_id;
822	max_bytes = wbc->wb_bytes;
823	} else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
824	max_id = wbc->wb_lcand_id;
825	max_bytes = wbc->wb_lcand_bytes;
826	} else {
827	max_id = wbc->wb_tcand_id;
828	max_bytes = wbc->wb_tcand_bytes;
829	}
830
831	/*
832	* Calculate the amount of IO time the winner consumed and fold it
833	* into the running average kept per inode. If the consumed IO
834	* time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
835	* deciding whether to switch or not. This is to prevent one-off
836	* small dirtiers from skewing the verdict.
837	*/
838	max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
839	wb->avg_write_bandwidth);
840	if (avg_time)
841	avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
842	(avg_time >> WB_FRN_TIME_AVG_SHIFT);
843	else
844	avg_time = max_time; /* immediate catch up on first run */
845
846	if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
847	int slots;
848
849	/*
850	* The switch verdict is reached if foreign wb's consume
851	* more than a certain proportion of IO time in a
852	* WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
853	* history mask where each bit represents one sixteenth of
854	* the period. Determine the number of slots to shift into
855	* history from @max_time.
856	*/
857	slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
858	(unsigned long)WB_FRN_HIST_MAX_SLOTS);
859	history <<= slots;
860	if (wbc->wb_id != max_id)
861	history \|= (1U << slots) - 1;
862
863	if (history)
864	trace_inode_foreign_history(inode, wbc, history);
865
866	/*
867	* Switch if the current wb isn't the consistent winner.
868	* If there are multiple closely competing dirtiers, the
869	* inode may switch across them repeatedly over time, which
870	* is okay. The main goal is avoiding keeping an inode on
871	* the wrong wb for an extended period of time.
872	*/
873	if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
874	inode_switch_wbs(inode, max_id);
875	}
876
877	/*
878	* Multiple instances of this function may race to update the
879	* following fields but we don't mind occassional inaccuracies.
880	*/
881	inode->i_wb_frn_winner = max_id;
882	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
883	inode->i_wb_frn_history = history;
884
885	wb_put(wbc->wb);
886	wbc->wb = NULL;
887	}
888	EXPORT_SYMBOL_GPL(wbc_detach_inode);
889
890	/**
891	* wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
892	* @wbc: writeback_control of the writeback in progress
893	* @page: page being written out
894	* @bytes: number of bytes being written out
895	*
896	* @bytes from @page are about to written out during the writeback
897	* controlled by @wbc. Keep the book for foreign inode detection. See
898	* wbc_detach_inode().
899	*/
900	void wbc_account_cgroup_owner(struct writeback_control wbc, struct page page,
901	size_t bytes)
902	{
903	struct folio *folio;
904	struct cgroup_subsys_state *css;
905	int id;
906
907	/*
908	* pageout() path doesn't attach @wbc to the inode being written
909	* out. This is intentional as we don't want the function to block
910	* behind a slow cgroup. Ultimately, we want pageout() to kick off
911	* regular writeback instead of writing things out itself.
912	*/
913	if (!wbc->wb \|\| wbc->no_cgroup_owner)
914	return;
915
916	folio = page_folio(page);
917	css = mem_cgroup_css_from_folio(folio);
918	/* dead cgroups shouldn't contribute to inode ownership arbitration */
919	if (!(css->flags & CSS_ONLINE))
920	return;
921
922	id = css->id;
923
924	if (id == wbc->wb_id) {
925	wbc->wb_bytes += bytes;
926	return;
927	}
928
929	if (id == wbc->wb_lcand_id)
930	wbc->wb_lcand_bytes += bytes;
931
932	/* Boyer-Moore majority vote algorithm */
933	if (!wbc->wb_tcand_bytes)
934	wbc->wb_tcand_id = id;
935	if (id == wbc->wb_tcand_id)
936	wbc->wb_tcand_bytes += bytes;
937	else
938	wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
939	}
940	EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
941
942	/**
943	* wb_split_bdi_pages - split nr_pages to write according to bandwidth
944	* @wb: target bdi_writeback to split @nr_pages to
945	* @nr_pages: number of pages to write for the whole bdi
946	*
947	* Split @wb's portion of @nr_pages according to @wb's write bandwidth in
948	* relation to the total write bandwidth of all wb's w/ dirty inodes on
949	* @wb->bdi.
950	*/
951	static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
952	{
953	unsigned long this_bw = wb->avg_write_bandwidth;
954	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
955
956	if (nr_pages == LONG_MAX)
957	return LONG_MAX;
958
959	/*
960	* This may be called on clean wb's and proportional distribution
961	* may not make sense, just use the original @nr_pages in those
962	* cases. In general, we wanna err on the side of writing more.
963	*/
964	if (!tot_bw \|\| this_bw >= tot_bw)
965	return nr_pages;
966	else
967	return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
968	}
969
970	/**
971	* bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
972	* @bdi: target backing_dev_info
973	* @base_work: wb_writeback_work to issue
974	* @skip_if_busy: skip wb's which already have writeback in progress
975	*
976	* Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
977	* have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
978	* distributed to the busy wbs according to each wb's proportion in the
979	* total active write bandwidth of @bdi.
980	*/
981	static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
982	struct wb_writeback_work *base_work,
983	bool skip_if_busy)
984	{
985	struct bdi_writeback *last_wb = NULL;
986	struct bdi_writeback *wb = list_entry(&bdi->wb_list,
987	struct bdi_writeback, bdi_node);
988
989	might_sleep();
990	restart:
991	rcu_read_lock();
992	list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
993	DEFINE_WB_COMPLETION(fallback_work_done, bdi);
994	struct wb_writeback_work fallback_work;
995	struct wb_writeback_work *work;
996	long nr_pages;
997
998	if (last_wb) {
999	wb_put(last_wb);
1000	last_wb = NULL;
1001	}
1002
1003	/* SYNC_ALL writes out I_DIRTY_TIME too */
1004	if (!wb_has_dirty_io(wb) &&
1005	(base_work->sync_mode == WB_SYNC_NONE \|\|
1006	list_empty(&wb->b_dirty_time)))
1007	continue;
1008	if (skip_if_busy && writeback_in_progress(wb))
1009	continue;
1010
1011	nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1012
1013	work = kmalloc(sizeof(*work), GFP_ATOMIC);
1014	if (work) {
1015	work = base_work;
1016	work->nr_pages = nr_pages;
1017	work->auto_free = 1;
1018	wb_queue_work(wb, work);
1019	continue;
1020	}
1021
1022	/*
1023	* If wb_tryget fails, the wb has been shutdown, skip it.
1024	*
1025	* Pin @wb so that it stays on @bdi->wb_list. This allows
1026	* continuing iteration from @wb after dropping and
1027	* regrabbing rcu read lock.
1028	*/
1029	if (!wb_tryget(wb))
1030	continue;
1031
1032	/* alloc failed, execute synchronously using on-stack fallback */
1033	work = &fallback_work;
1034	work = base_work;
1035	work->nr_pages = nr_pages;
1036	work->auto_free = 0;
1037	work->done = &fallback_work_done;
1038
1039	wb_queue_work(wb, work);
1040	last_wb = wb;
1041
1042	rcu_read_unlock();
1043	wb_wait_for_completion(&fallback_work_done);
1044	goto restart;
1045	}
1046	rcu_read_unlock();
1047
1048	if (last_wb)
1049	wb_put(last_wb);
1050	}
1051
1052	/**
1053	* cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1054	* @bdi_id: target bdi id
1055	* @memcg_id: target memcg css id
1056	* @reason: reason why some writeback work initiated
1057	* @done: target wb_completion
1058	*
1059	* Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1060	* with the specified parameters.
1061	*/
1062	int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1063	enum wb_reason reason, struct wb_completion *done)
1064	{
1065	struct backing_dev_info *bdi;
1066	struct cgroup_subsys_state *memcg_css;
1067	struct bdi_writeback *wb;
1068	struct wb_writeback_work *work;
1069	unsigned long dirty;
1070	int ret;
1071
1072	/* lookup bdi and memcg */
1073	bdi = bdi_get_by_id(bdi_id);
1074	if (!bdi)
1075	return -ENOENT;
1076
1077	rcu_read_lock();
1078	memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1079	if (memcg_css && !css_tryget(memcg_css))
1080	memcg_css = NULL;
1081	rcu_read_unlock();
1082	if (!memcg_css) {
1083	ret = -ENOENT;
1084	goto out_bdi_put;
1085	}
1086
1087	/*
1088	* And find the associated wb. If the wb isn't there already
1089	* there's nothing to flush, don't create one.
1090	*/
1091	wb = wb_get_lookup(bdi, memcg_css);
1092	if (!wb) {
1093	ret = -ENOENT;
1094	goto out_css_put;
1095	}
1096
1097	/*
1098	* The caller is attempting to write out most of
1099	* the currently dirty pages. Let's take the current dirty page
1100	* count and inflate it by 25% which should be large enough to
1101	* flush out most dirty pages while avoiding getting livelocked by
1102	* concurrent dirtiers.
1103	*
1104	* BTW the memcg stats are flushed periodically and this is best-effort
1105	* estimation, so some potential error is ok.
1106	*/
1107	dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1108	dirty = dirty * 10 / 8;
1109
1110	/* issue the writeback work */
1111	work = kzalloc(sizeof(*work), GFP_NOWAIT \| __GFP_NOWARN);
1112	if (work) {
1113	work->nr_pages = dirty;
1114	work->sync_mode = WB_SYNC_NONE;
1115	work->range_cyclic = 1;
1116	work->reason = reason;
1117	work->done = done;
1118	work->auto_free = 1;
1119	wb_queue_work(wb, work);
1120	ret = 0;
1121	} else {
1122	ret = -ENOMEM;
1123	}
1124
1125	wb_put(wb);
1126	out_css_put:
1127	css_put(memcg_css);
1128	out_bdi_put:
1129	bdi_put(bdi);
1130	return ret;
1131	}
1132
1133	/**
1134	* cgroup_writeback_umount - flush inode wb switches for umount
1135	* @sb: target super_block
1136	*
1137	* This function is called when a super_block is about to be destroyed and
1138	* flushes in-flight inode wb switches. An inode wb switch goes through
1139	* RCU and then workqueue, so the two need to be flushed in order to ensure
1140	* that all previously scheduled switches are finished. As wb switches are
1141	* rare occurrences and synchronize_rcu() can take a while, perform
1142	* flushing iff wb switches are in flight.
1143	*/
1144	void cgroup_writeback_umount(struct super_block *sb)
1145	{
1146
1147	if (!(sb->s_bdi->capabilities & BDI_CAP_WRITEBACK))
1148	return;
1149
1150	/*
1151	* SB_ACTIVE should be reliably cleared before checking
1152	* isw_nr_in_flight, see generic_shutdown_super().
1153	*/
1154	smp_mb();
1155
1156	if (atomic_read(&isw_nr_in_flight)) {
1157	/*
1158	* Use rcu_barrier() to wait for all pending callbacks to
1159	* ensure that all in-flight wb switches are in the workqueue.
1160	*/
1161	rcu_barrier();
1162	flush_workqueue(isw_wq);
1163	}
1164	}
1165
1166	static int __init cgroup_writeback_init(void)
1167	{
1168	isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1169	if (!isw_wq)
1170	return -ENOMEM;
1171	return 0;
1172	}
1173	fs_initcall(cgroup_writeback_init);
1174
1175	#else /* CONFIG_CGROUP_WRITEBACK */
1176
1177	static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1178	static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1179
1180	static void inode_cgwb_move_to_attached(struct inode *inode,
1181	struct bdi_writeback *wb)
1182	{
1183	assert_spin_locked(&wb->list_lock);
1184	assert_spin_locked(&inode->i_lock);
1185	WARN_ON_ONCE(inode->i_state & I_FREEING);
1186
1187	inode->i_state &= ~I_SYNC_QUEUED;
1188	list_del_init(&inode->i_io_list);
1189	wb_io_lists_depopulated(wb);
1190	}
1191
1192	static struct bdi_writeback *
1193	locked_inode_to_wb_and_lock_list(struct inode *inode)
1194	__releases(&inode->i_lock)
1195	__acquires(&wb->list_lock)
1196	{
1197	struct bdi_writeback *wb = inode_to_wb(inode);
1198
1199	spin_unlock(&inode->i_lock);
1200	spin_lock(&wb->list_lock);
1201	return wb;
1202	}
1203
1204	static struct bdi_writeback inode_to_wb_and_lock_list(struct inode inode)
1205	__acquires(&wb->list_lock)
1206	{
1207	struct bdi_writeback *wb = inode_to_wb(inode);
1208
1209	spin_lock(&wb->list_lock);
1210	return wb;
1211	}
1212
1213	static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1214	{
1215	return nr_pages;
1216	}
1217
1218	static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1219	struct wb_writeback_work *base_work,
1220	bool skip_if_busy)
1221	{
1222	might_sleep();
1223
1224	if (!skip_if_busy \|\| !writeback_in_progress(&bdi->wb)) {
1225	base_work->auto_free = 0;
1226	wb_queue_work(&bdi->wb, base_work);
1227	}
1228	}
1229
1230	#endif /* CONFIG_CGROUP_WRITEBACK */
1231
1232	/*
1233	* Add in the number of potentially dirty inodes, because each inode
1234	* write can dirty pagecache in the underlying blockdev.
1235	*/
1236	static unsigned long get_nr_dirty_pages(void)
1237	{
1238	return global_node_page_state(NR_FILE_DIRTY) +
1239	get_nr_dirty_inodes();
1240	}
1241
1242	static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1243	{
1244	if (!wb_has_dirty_io(wb))
1245	return;
1246
1247	/*
1248	* All callers of this function want to start writeback of all
1249	* dirty pages. Places like vmscan can call this at a very
1250	* high frequency, causing pointless allocations of tons of
1251	* work items and keeping the flusher threads busy retrieving
1252	* that work. Ensure that we only allow one of them pending and
1253	* inflight at the time.
1254	*/
1255	if (test_bit(WB_start_all, &wb->state) \|\|
1256	test_and_set_bit(WB_start_all, &wb->state))
1257	return;
1258
1259	wb->start_all_reason = reason;
1260	wb_wakeup(wb);
1261	}
1262
1263	/**
1264	* wb_start_background_writeback - start background writeback
1265	* @wb: bdi_writback to write from
1266	*
1267	* Description:
1268	* This makes sure WB_SYNC_NONE background writeback happens. When
1269	* this function returns, it is only guaranteed that for given wb
1270	* some IO is happening if we are over background dirty threshold.
1271	* Caller need not hold sb s_umount semaphore.
1272	*/
1273	void wb_start_background_writeback(struct bdi_writeback *wb)
1274	{
1275	/*
1276	* We just wake up the flusher thread. It will perform background
1277	* writeback as soon as there is no other work to do.
1278	*/
1279	trace_writeback_wake_background(wb);
1280	wb_wakeup(wb);
1281	}
1282
1283	/*
1284	* Remove the inode from the writeback list it is on.
1285	*/
1286	void inode_io_list_del(struct inode *inode)
1287	{
1288	struct bdi_writeback *wb;
1289
1290	wb = inode_to_wb_and_lock_list(inode);
1291	spin_lock(&inode->i_lock);
1292
1293	inode->i_state &= ~I_SYNC_QUEUED;
1294	list_del_init(&inode->i_io_list);
1295	wb_io_lists_depopulated(wb);
1296
1297	spin_unlock(&inode->i_lock);
1298	spin_unlock(&wb->list_lock);
1299	}
1300	EXPORT_SYMBOL(inode_io_list_del);
1301
1302	/*
1303	* mark an inode as under writeback on the sb
1304	*/
1305	void sb_mark_inode_writeback(struct inode *inode)
1306	{
1307	struct super_block *sb = inode->i_sb;
1308	unsigned long flags;
1309
1310	if (list_empty(&inode->i_wb_list)) {
1311	spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1312	if (list_empty(&inode->i_wb_list)) {
1313	list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1314	trace_sb_mark_inode_writeback(inode);
1315	}
1316	spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1317	}
1318	}
1319
1320	/*
1321	* clear an inode as under writeback on the sb
1322	*/
1323	void sb_clear_inode_writeback(struct inode *inode)
1324	{
1325	struct super_block *sb = inode->i_sb;
1326	unsigned long flags;
1327
1328	if (!list_empty(&inode->i_wb_list)) {
1329	spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1330	if (!list_empty(&inode->i_wb_list)) {
1331	list_del_init(&inode->i_wb_list);
1332	trace_sb_clear_inode_writeback(inode);
1333	}
1334	spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1335	}
1336	}
1337
1338	/*
1339	* Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1340	* furthest end of its superblock's dirty-inode list.
1341	*
1342	* Before stamping the inode's ->dirtied_when, we check to see whether it is
1343	* already the most-recently-dirtied inode on the b_dirty list. If that is
1344	* the case then the inode must have been redirtied while it was being written
1345	* out and we don't reset its dirtied_when.
1346	*/
1347	static void redirty_tail_locked(struct inode inode, struct bdi_writeback wb)
1348	{
1349	assert_spin_locked(&inode->i_lock);
1350
1351	inode->i_state &= ~I_SYNC_QUEUED;
1352	/*
1353	* When the inode is being freed just don't bother with dirty list
1354	* tracking. Flush worker will ignore this inode anyway and it will
1355	* trigger assertions in inode_io_list_move_locked().
1356	*/
1357	if (inode->i_state & I_FREEING) {
1358	list_del_init(&inode->i_io_list);
1359	wb_io_lists_depopulated(wb);
1360	return;
1361	}
1362	if (!list_empty(&wb->b_dirty)) {
1363	struct inode *tail;
1364
1365	tail = wb_inode(wb->b_dirty.next);
1366	if (time_before(inode->dirtied_when, tail->dirtied_when))
1367	inode->dirtied_when = jiffies;
1368	}
1369	inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1370	}
1371
1372	static void redirty_tail(struct inode inode, struct bdi_writeback wb)
1373	{
1374	spin_lock(&inode->i_lock);
1375	redirty_tail_locked(inode, wb);
1376	spin_unlock(&inode->i_lock);
1377	}
1378
1379	/*
1380	* requeue inode for re-scanning after bdi->b_io list is exhausted.
1381	*/
1382	static void requeue_io(struct inode inode, struct bdi_writeback wb)
1383	{
1384	inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1385	}
1386
1387	static void inode_sync_complete(struct inode *inode)
1388	{
1389	assert_spin_locked(&inode->i_lock);
1390
1391	inode->i_state &= ~I_SYNC;
1392	/* If inode is clean an unused, put it into LRU now... */
1393	inode_add_lru(inode);
1394	/* Called with inode->i_lock which ensures memory ordering. */
1395	inode_wake_up_bit(inode, __I_SYNC);
1396	}
1397
1398	static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1399	{
1400	bool ret = time_after(inode->dirtied_when, t);
1401	#ifndef CONFIG_64BIT
1402	/*
1403	* For inodes being constantly redirtied, dirtied_when can get stuck.
1404	* It _appears_ to be in the future, but is actually in distant past.
1405	* This test is necessary to prevent such wrapped-around relative times
1406	* from permanently stopping the whole bdi writeback.
1407	*/
1408	ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1409	#endif
1410	return ret;
1411	}
1412
1413	/*
1414	* Move expired (dirtied before dirtied_before) dirty inodes from
1415	* @delaying_queue to @dispatch_queue.
1416	*/
1417	static int move_expired_inodes(struct list_head *delaying_queue,
1418	struct list_head *dispatch_queue,
1419	unsigned long dirtied_before)
1420	{
1421	LIST_HEAD(tmp);
1422	struct list_head pos, node;
1423	struct super_block *sb = NULL;
1424	struct inode *inode;
1425	int do_sb_sort = 0;
1426	int moved = 0;
1427
1428	while (!list_empty(delaying_queue)) {
1429	inode = wb_inode(delaying_queue->prev);
1430	if (inode_dirtied_after(inode, dirtied_before))
1431	break;
1432	spin_lock(&inode->i_lock);
1433	list_move(&inode->i_io_list, &tmp);
1434	moved++;
1435	inode->i_state \|= I_SYNC_QUEUED;
1436	spin_unlock(&inode->i_lock);
1437	if (sb_is_blkdev_sb(inode->i_sb))
1438	continue;
1439	if (sb && sb != inode->i_sb)
1440	do_sb_sort = 1;
1441	sb = inode->i_sb;
1442	}
1443
1444	/* just one sb in list, splice to dispatch_queue and we're done */
1445	if (!do_sb_sort) {
1446	list_splice(&tmp, dispatch_queue);
1447	goto out;
1448	}
1449
1450	/*
1451	* Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1452	* we don't take inode->i_lock here because it is just a pointless overhead.
1453	* Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1454	* fully under our control.
1455	*/
1456	while (!list_empty(&tmp)) {
1457	sb = wb_inode(tmp.prev)->i_sb;
1458	list_for_each_prev_safe(pos, node, &tmp) {
1459	inode = wb_inode(pos);
1460	if (inode->i_sb == sb)
1461	list_move(&inode->i_io_list, dispatch_queue);
1462	}
1463	}
1464	out:
1465	return moved;
1466	}
1467
1468	/*
1469	* Queue all expired dirty inodes for io, eldest first.
1470	* Before
1471	* newly dirtied b_dirty b_io b_more_io
1472	* =============> gf edc BA
1473	* After
1474	* newly dirtied b_dirty b_io b_more_io
1475	* =============> g fBAedc
1476	* \|
1477	* +--> dequeue for IO
1478	*/
1479	static void queue_io(struct bdi_writeback wb, struct wb_writeback_work work,
1480	unsigned long dirtied_before)
1481	{
1482	int moved;
1483	unsigned long time_expire_jif = dirtied_before;
1484
1485	assert_spin_locked(&wb->list_lock);
1486	list_splice_init(&wb->b_more_io, &wb->b_io);
1487	moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1488	if (!work->for_sync)
1489	time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1490	moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1491	time_expire_jif);
1492	if (moved)
1493	wb_io_lists_populated(wb);
1494	trace_writeback_queue_io(wb, work, dirtied_before, moved);
1495	}
1496
1497	static int write_inode(struct inode inode, struct writeback_control wbc)
1498	{
1499	int ret;
1500
1501	if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1502	trace_writeback_write_inode_start(inode, wbc);
1503	ret = inode->i_sb->s_op->write_inode(inode, wbc);
1504	trace_writeback_write_inode(inode, wbc);
1505	return ret;
1506	}
1507	return 0;
1508	}
1509
1510	/*
1511	* Wait for writeback on an inode to complete. Called with i_lock held.
1512	* Caller must make sure inode cannot go away when we drop i_lock.
1513	*/
1514	void inode_wait_for_writeback(struct inode *inode)
1515	{
1516	struct wait_bit_queue_entry wqe;
1517	struct wait_queue_head *wq_head;
1518
1519	assert_spin_locked(&inode->i_lock);
1520
1521	if (!(inode->i_state & I_SYNC))
1522	return;
1523
1524	wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1525	for (;;) {
1526	prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1527	/* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1528	if (!(inode->i_state & I_SYNC))
1529	break;
1530	spin_unlock(&inode->i_lock);
1531	schedule();
1532	spin_lock(&inode->i_lock);
1533	}
1534	finish_wait(wq_head, &wqe.wq_entry);
1535	}
1536
1537	/*
1538	* Sleep until I_SYNC is cleared. This function must be called with i_lock
1539	* held and drops it. It is aimed for callers not holding any inode reference
1540	* so once i_lock is dropped, inode can go away.
1541	*/
1542	static void inode_sleep_on_writeback(struct inode *inode)
1543	__releases(inode->i_lock)
1544	{
1545	struct wait_bit_queue_entry wqe;
1546	struct wait_queue_head *wq_head;
1547	bool sleep;
1548
1549	assert_spin_locked(&inode->i_lock);
1550
1551	wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1552	prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1553	/* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1554	sleep = !!(inode->i_state & I_SYNC);
1555	spin_unlock(&inode->i_lock);
1556	if (sleep)
1557	schedule();
1558	finish_wait(wq_head, &wqe.wq_entry);
1559	}
1560
1561	/*
1562	* Find proper writeback list for the inode depending on its current state and
1563	* possibly also change of its state while we were doing writeback. Here we
1564	* handle things such as livelock prevention or fairness of writeback among
1565	* inodes. This function can be called only by flusher thread - noone else
1566	* processes all inodes in writeback lists and requeueing inodes behind flusher
1567	* thread's back can have unexpected consequences.
1568	*/
1569	static void requeue_inode(struct inode inode, struct bdi_writeback wb,
1570	struct writeback_control *wbc,
1571	unsigned long dirtied_before)
1572	{
1573	if (inode->i_state & I_FREEING)
1574	return;
1575
1576	/*
1577	* Sync livelock prevention. Each inode is tagged and synced in one
1578	* shot. If still dirty, it will be redirty_tail()'ed below. Update
1579	* the dirty time to prevent enqueue and sync it again.
1580	*/
1581	if ((inode->i_state & I_DIRTY) &&
1582	(wbc->sync_mode == WB_SYNC_ALL \|\| wbc->tagged_writepages))
1583	inode->dirtied_when = jiffies;
1584
1585	if (wbc->pages_skipped) {
1586	/*
1587	* Writeback is not making progress due to locked buffers.
1588	* Skip this inode for now. Although having skipped pages
1589	* is odd for clean inodes, it can happen for some
1590	* filesystems so handle that gracefully.
1591	*/
1592	if (inode->i_state & I_DIRTY_ALL)
1593	redirty_tail_locked(inode, wb);
1594	else
1595	inode_cgwb_move_to_attached(inode, wb);
1596	return;
1597	}
1598
1599	if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1600	/*
1601	* We didn't write back all the pages. nfs_writepages()
1602	* sometimes bales out without doing anything.
1603	*/
1604	if (wbc->nr_to_write <= 0 &&
1605	!inode_dirtied_after(inode, dirtied_before)) {
1606	/* Slice used up. Queue for next turn. */
1607	requeue_io(inode, wb);
1608	} else {
1609	/*
1610	* Writeback blocked by something other than
1611	* congestion. Delay the inode for some time to
1612	* avoid spinning on the CPU (100% iowait)
1613	* retrying writeback of the dirty page/inode
1614	* that cannot be performed immediately.
1615	*/
1616	redirty_tail_locked(inode, wb);
1617	}
1618	} else if (inode->i_state & I_DIRTY) {
1619	/*
1620	* Filesystems can dirty the inode during writeback operations,
1621	* such as delayed allocation during submission or metadata
1622	* updates after data IO completion.
1623	*/
1624	redirty_tail_locked(inode, wb);
1625	} else if (inode->i_state & I_DIRTY_TIME) {
1626	inode->dirtied_when = jiffies;
1627	inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1628	inode->i_state &= ~I_SYNC_QUEUED;
1629	} else {
1630	/* The inode is clean. Remove from writeback lists. */
1631	inode_cgwb_move_to_attached(inode, wb);
1632	}
1633	}
1634
1635	/*
1636	* Write out an inode and its dirty pages (or some of its dirty pages, depending
1637	* on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1638	*
1639	* This doesn't remove the inode from the writeback list it is on, except
1640	* potentially to move it from b_dirty_time to b_dirty due to timestamp
1641	* expiration. The caller is otherwise responsible for writeback list handling.
1642	*
1643	* The caller is also responsible for setting the I_SYNC flag beforehand and
1644	* calling inode_sync_complete() to clear it afterwards.
1645	*/
1646	static int
1647	__writeback_single_inode(struct inode inode, struct writeback_control wbc)
1648	{
1649	struct address_space *mapping = inode->i_mapping;
1650	long nr_to_write = wbc->nr_to_write;
1651	unsigned dirty;
1652	int ret;
1653
1654	WARN_ON(!(inode->i_state & I_SYNC));
1655
1656	trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1657
1658	ret = do_writepages(mapping, wbc);
1659
1660	/*
1661	* Make sure to wait on the data before writing out the metadata.
1662	* This is important for filesystems that modify metadata on data
1663	* I/O completion. We don't do it for sync(2) writeback because it has a
1664	* separate, external IO completion path and ->sync_fs for guaranteeing
1665	* inode metadata is written back correctly.
1666	*/
1667	if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1668	int err = filemap_fdatawait(mapping);
1669	if (ret == 0)
1670	ret = err;
1671	}
1672
1673	/*
1674	* If the inode has dirty timestamps and we need to write them, call
1675	* mark_inode_dirty_sync() to notify the filesystem about it and to
1676	* change I_DIRTY_TIME into I_DIRTY_SYNC.
1677	*/
1678	if ((inode->i_state & I_DIRTY_TIME) &&
1679	(wbc->sync_mode == WB_SYNC_ALL \|\|
1680	time_after(jiffies, inode->dirtied_time_when +
1681	dirtytime_expire_interval * HZ))) {
1682	trace_writeback_lazytime(inode);
1683	mark_inode_dirty_sync(inode);
1684	}
1685
1686	/*
1687	* Get and clear the dirty flags from i_state. This needs to be done
1688	* after calling writepages because some filesystems may redirty the
1689	* inode during writepages due to delalloc. It also needs to be done
1690	* after handling timestamp expiration, as that may dirty the inode too.
1691	*/
1692	spin_lock(&inode->i_lock);
1693	dirty = inode->i_state & I_DIRTY;
1694	inode->i_state &= ~dirty;
1695
1696	/*
1697	* Paired with smp_mb() in __mark_inode_dirty(). This allows
1698	* __mark_inode_dirty() to test i_state without grabbing i_lock -
1699	* either they see the I_DIRTY bits cleared or we see the dirtied
1700	* inode.
1701	*
1702	* I_DIRTY_PAGES is always cleared together above even if @mapping
1703	* still has dirty pages. The flag is reinstated after smp_mb() if
1704	* necessary. This guarantees that either __mark_inode_dirty()
1705	* sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1706	*/
1707	smp_mb();
1708
1709	if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1710	inode->i_state \|= I_DIRTY_PAGES;
1711	else if (unlikely(inode->i_state & I_PINNING_NETFS_WB)) {
1712	if (!(inode->i_state & I_DIRTY_PAGES)) {
1713	inode->i_state &= ~I_PINNING_NETFS_WB;
1714	wbc->unpinned_netfs_wb = true;
1715	dirty \|= I_PINNING_NETFS_WB; /* Cause write_inode */
1716	}
1717	}
1718
1719	spin_unlock(&inode->i_lock);
1720
1721	/* Don't write the inode if only I_DIRTY_PAGES was set */
1722	if (dirty & ~I_DIRTY_PAGES) {
1723	int err = write_inode(inode, wbc);
1724	if (ret == 0)
1725	ret = err;
1726	}
1727	wbc->unpinned_netfs_wb = false;
1728	trace_writeback_single_inode(inode, wbc, nr_to_write);
1729	return ret;
1730	}
1731
1732	/*
1733	* Write out an inode's dirty data and metadata on-demand, i.e. separately from
1734	* the regular batched writeback done by the flusher threads in
1735	* writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1736	* whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1737	*
1738	* To prevent the inode from going away, either the caller must have a reference
1739	* to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1740	*/
1741	static int writeback_single_inode(struct inode *inode,
1742	struct writeback_control *wbc)
1743	{
1744	struct bdi_writeback *wb;
1745	int ret = 0;
1746
1747	spin_lock(&inode->i_lock);
1748	if (!atomic_read(&inode->i_count))
1749	WARN_ON(!(inode->i_state & (I_WILL_FREE\|I_FREEING)));
1750	else
1751	WARN_ON(inode->i_state & I_WILL_FREE);
1752
1753	if (inode->i_state & I_SYNC) {
1754	/*
1755	* Writeback is already running on the inode. For WB_SYNC_NONE,
1756	* that's enough and we can just return. For WB_SYNC_ALL, we
1757	* must wait for the existing writeback to complete, then do
1758	* writeback again if there's anything left.
1759	*/
1760	if (wbc->sync_mode != WB_SYNC_ALL)
1761	goto out;
1762	inode_wait_for_writeback(inode);
1763	}
1764	WARN_ON(inode->i_state & I_SYNC);
1765	/*
1766	* If the inode is already fully clean, then there's nothing to do.
1767	*
1768	* For data-integrity syncs we also need to check whether any pages are
1769	* still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1770	* there are any such pages, we'll need to wait for them.
1771	*/
1772	if (!(inode->i_state & I_DIRTY_ALL) &&
1773	(wbc->sync_mode != WB_SYNC_ALL \|\|
1774	!mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1775	goto out;
1776	inode->i_state \|= I_SYNC;
1777	wbc_attach_and_unlock_inode(wbc, inode);
1778
1779	ret = __writeback_single_inode(inode, wbc);
1780
1781	wbc_detach_inode(wbc);
1782
1783	wb = inode_to_wb_and_lock_list(inode);
1784	spin_lock(&inode->i_lock);
1785	/*
1786	* If the inode is freeing, its i_io_list shoudn't be updated
1787	* as it can be finally deleted at this moment.
1788	*/
1789	if (!(inode->i_state & I_FREEING)) {
1790	/*
1791	* If the inode is now fully clean, then it can be safely
1792	* removed from its writeback list (if any). Otherwise the
1793	* flusher threads are responsible for the writeback lists.
1794	*/
1795	if (!(inode->i_state & I_DIRTY_ALL))
1796	inode_cgwb_move_to_attached(inode, wb);
1797	else if (!(inode->i_state & I_SYNC_QUEUED)) {
1798	if ((inode->i_state & I_DIRTY))
1799	redirty_tail_locked(inode, wb);
1800	else if (inode->i_state & I_DIRTY_TIME) {
1801	inode->dirtied_when = jiffies;
1802	inode_io_list_move_locked(inode,
1803	wb,
1804	&wb->b_dirty_time);
1805	}
1806	}
1807	}
1808
1809	spin_unlock(&wb->list_lock);
1810	inode_sync_complete(inode);
1811	out:
1812	spin_unlock(&inode->i_lock);
1813	return ret;
1814	}
1815
1816	static long writeback_chunk_size(struct bdi_writeback *wb,
1817	struct wb_writeback_work *work)
1818	{
1819	long pages;
1820
1821	/*
1822	* WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1823	* inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1824	* here avoids calling into writeback_inodes_wb() more than once.
1825	*
1826	* The intended call sequence for WB_SYNC_ALL writeback is:
1827	*
1828	* wb_writeback()
1829	* writeback_sb_inodes() <== called only once
1830	* write_cache_pages() <== called once for each inode
1831	* (quickly) tag currently dirty pages
1832	* (maybe slowly) sync all tagged pages
1833	*/
1834	if (work->sync_mode == WB_SYNC_ALL \|\| work->tagged_writepages)
1835	pages = LONG_MAX;
1836	else {
1837	pages = min(wb->avg_write_bandwidth / 2,
1838	global_wb_domain.dirty_limit / DIRTY_SCOPE);
1839	pages = min(pages, work->nr_pages);
1840	pages = round_down(pages + MIN_WRITEBACK_PAGES,
1841	MIN_WRITEBACK_PAGES);
1842	}
1843
1844	return pages;
1845	}
1846
1847	/*
1848	* Write a portion of b_io inodes which belong to @sb.
1849	*
1850	* Return the number of pages and/or inodes written.
1851	*
1852	* NOTE! This is called with wb->list_lock held, and will
1853	* unlock and relock that for each inode it ends up doing
1854	* IO for.
1855	*/
1856	static long writeback_sb_inodes(struct super_block *sb,
1857	struct bdi_writeback *wb,
1858	struct wb_writeback_work *work)
1859	{
1860	struct writeback_control wbc = {
1861	.sync_mode = work->sync_mode,
1862	.tagged_writepages = work->tagged_writepages,
1863	.for_kupdate = work->for_kupdate,
1864	.for_background = work->for_background,
1865	.for_sync = work->for_sync,
1866	.range_cyclic = work->range_cyclic,
1867	.range_start = 0,
1868	.range_end = LLONG_MAX,
1869	};
1870	unsigned long start_time = jiffies;
1871	long write_chunk;
1872	long total_wrote = 0; /* count both pages and inodes */
1873	unsigned long dirtied_before = jiffies;
1874
1875	if (work->for_kupdate)
1876	dirtied_before = jiffies -
1877	msecs_to_jiffies(dirty_expire_interval * 10);
1878
1879	while (!list_empty(&wb->b_io)) {
1880	struct inode *inode = wb_inode(wb->b_io.prev);
1881	struct bdi_writeback *tmp_wb;
1882	long wrote;
1883
1884	if (inode->i_sb != sb) {
1885	if (work->sb) {
1886	/*
1887	* We only want to write back data for this
1888	* superblock, move all inodes not belonging
1889	* to it back onto the dirty list.
1890	*/
1891	redirty_tail(inode, wb);
1892	continue;
1893	}
1894
1895	/*
1896	* The inode belongs to a different superblock.
1897	* Bounce back to the caller to unpin this and
1898	* pin the next superblock.
1899	*/
1900	break;
1901	}
1902
1903	/*
1904	* Don't bother with new inodes or inodes being freed, first
1905	* kind does not need periodic writeout yet, and for the latter
1906	* kind writeout is handled by the freer.
1907	*/
1908	spin_lock(&inode->i_lock);
1909	if (inode->i_state & (I_NEW \| I_FREEING \| I_WILL_FREE)) {
1910	redirty_tail_locked(inode, wb);
1911	spin_unlock(&inode->i_lock);
1912	continue;
1913	}
1914	if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1915	/*
1916	* If this inode is locked for writeback and we are not
1917	* doing writeback-for-data-integrity, move it to
1918	* b_more_io so that writeback can proceed with the
1919	* other inodes on s_io.
1920	*
1921	* We'll have another go at writing back this inode
1922	* when we completed a full scan of b_io.
1923	*/
1924	requeue_io(inode, wb);
1925	spin_unlock(&inode->i_lock);
1926	trace_writeback_sb_inodes_requeue(inode);
1927	continue;
1928	}
1929	spin_unlock(&wb->list_lock);
1930
1931	/*
1932	* We already requeued the inode if it had I_SYNC set and we
1933	* are doing WB_SYNC_NONE writeback. So this catches only the
1934	* WB_SYNC_ALL case.
1935	*/
1936	if (inode->i_state & I_SYNC) {
1937	/* Wait for I_SYNC. This function drops i_lock... */
1938	inode_sleep_on_writeback(inode);
1939	/* Inode may be gone, start again */
1940	spin_lock(&wb->list_lock);
1941	continue;
1942	}
1943	inode->i_state \|= I_SYNC;
1944	wbc_attach_and_unlock_inode(&wbc, inode);
1945
1946	write_chunk = writeback_chunk_size(wb, work);
1947	wbc.nr_to_write = write_chunk;
1948	wbc.pages_skipped = 0;
1949
1950	/*
1951	* We use I_SYNC to pin the inode in memory. While it is set
1952	* evict_inode() will wait so the inode cannot be freed.
1953	*/
1954	__writeback_single_inode(inode, &wbc);
1955
1956	wbc_detach_inode(&wbc);
1957	work->nr_pages -= write_chunk - wbc.nr_to_write;
1958	wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1959	wrote = wrote < 0 ? 0 : wrote;
1960	total_wrote += wrote;
1961
1962	if (need_resched()) {
1963	/*
1964	* We're trying to balance between building up a nice
1965	* long list of IOs to improve our merge rate, and
1966	* getting those IOs out quickly for anyone throttling
1967	* in balance_dirty_pages(). cond_resched() doesn't
1968	* unplug, so get our IOs out the door before we
1969	* give up the CPU.
1970	*/
1971	blk_flush_plug(current->plug, false);
1972	cond_resched();
1973	}
1974
1975	/*
1976	* Requeue @inode if still dirty. Be careful as @inode may
1977	* have been switched to another wb in the meantime.
1978	*/
1979	tmp_wb = inode_to_wb_and_lock_list(inode);
1980	spin_lock(&inode->i_lock);
1981	if (!(inode->i_state & I_DIRTY_ALL))
1982	total_wrote++;
1983	requeue_inode(inode, tmp_wb, &wbc, dirtied_before);
1984	inode_sync_complete(inode);
1985	spin_unlock(&inode->i_lock);
1986
1987	if (unlikely(tmp_wb != wb)) {
1988	spin_unlock(&tmp_wb->list_lock);
1989	spin_lock(&wb->list_lock);
1990	}
1991
1992	/*
1993	* bail out to wb_writeback() often enough to check
1994	* background threshold and other termination conditions.
1995	*/
1996	if (total_wrote) {
1997	if (time_is_before_jiffies(start_time + HZ / 10UL))
1998	break;
1999	if (work->nr_pages <= 0)
2000	break;
2001	}
2002	}
2003	return total_wrote;
2004	}
2005
2006	static long __writeback_inodes_wb(struct bdi_writeback *wb,
2007	struct wb_writeback_work *work)
2008	{
2009	unsigned long start_time = jiffies;
2010	long wrote = 0;
2011
2012	while (!list_empty(&wb->b_io)) {
2013	struct inode *inode = wb_inode(wb->b_io.prev);
2014	struct super_block *sb = inode->i_sb;
2015
2016	if (!super_trylock_shared(sb)) {
2017	/*
2018	* super_trylock_shared() may fail consistently due to
2019	* s_umount being grabbed by someone else. Don't use
2020	* requeue_io() to avoid busy retrying the inode/sb.
2021	*/
2022	redirty_tail(inode, wb);
2023	continue;
2024	}
2025	wrote += writeback_sb_inodes(sb, wb, work);
2026	up_read(&sb->s_umount);
2027
2028	/* refer to the same tests at the end of writeback_sb_inodes */
2029	if (wrote) {
2030	if (time_is_before_jiffies(start_time + HZ / 10UL))
2031	break;
2032	if (work->nr_pages <= 0)
2033	break;
2034	}
2035	}
2036	/* Leave any unwritten inodes on b_io */
2037	return wrote;
2038	}
2039
2040	static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
2041	enum wb_reason reason)
2042	{
2043	struct wb_writeback_work work = {
2044	.nr_pages = nr_pages,
2045	.sync_mode = WB_SYNC_NONE,
2046	.range_cyclic = 1,
2047	.reason = reason,
2048	};
2049	struct blk_plug plug;
2050
2051	blk_start_plug(&plug);
2052	spin_lock(&wb->list_lock);
2053	if (list_empty(&wb->b_io))
2054	queue_io(wb, &work, jiffies);
2055	__writeback_inodes_wb(wb, &work);
2056	spin_unlock(&wb->list_lock);
2057	blk_finish_plug(&plug);
2058
2059	return nr_pages - work.nr_pages;
2060	}
2061
2062	/*
2063	* Explicit flushing or periodic writeback of "old" data.
2064	*
2065	* Define "old": the first time one of an inode's pages is dirtied, we mark the
2066	* dirtying-time in the inode's address_space. So this periodic writeback code
2067	* just walks the superblock inode list, writing back any inodes which are
2068	* older than a specific point in time.
2069	*
2070	* Try to run once per dirty_writeback_interval. But if a writeback event
2071	* takes longer than a dirty_writeback_interval interval, then leave a
2072	* one-second gap.
2073	*
2074	* dirtied_before takes precedence over nr_to_write. So we'll only write back
2075	* all dirty pages if they are all attached to "old" mappings.
2076	*/
2077	static long wb_writeback(struct bdi_writeback *wb,
2078	struct wb_writeback_work *work)
2079	{
2080	long nr_pages = work->nr_pages;
2081	unsigned long dirtied_before = jiffies;
2082	struct inode *inode;
2083	long progress;
2084	struct blk_plug plug;
2085	bool queued = false;
2086
2087	blk_start_plug(&plug);
2088	for (;;) {
2089	/*
2090	* Stop writeback when nr_pages has been consumed
2091	*/
2092	if (work->nr_pages <= 0)
2093	break;
2094
2095	/*
2096	* Background writeout and kupdate-style writeback may
2097	* run forever. Stop them if there is other work to do
2098	* so that e.g. sync can proceed. They'll be restarted
2099	* after the other works are all done.
2100	*/
2101	if ((work->for_background \|\| work->for_kupdate) &&
2102	!list_empty(&wb->work_list))
2103	break;
2104
2105	/*
2106	* For background writeout, stop when we are below the
2107	* background dirty threshold
2108	*/
2109	if (work->for_background && !wb_over_bg_thresh(wb))
2110	break;
2111
2112
2113	spin_lock(&wb->list_lock);
2114
2115	trace_writeback_start(wb, work);
2116	if (list_empty(&wb->b_io)) {
2117	/*
2118	* Kupdate and background works are special and we want
2119	* to include all inodes that need writing. Livelock
2120	* avoidance is handled by these works yielding to any
2121	* other work so we are safe.
2122	*/
2123	if (work->for_kupdate) {
2124	dirtied_before = jiffies -
2125	msecs_to_jiffies(dirty_expire_interval *
2126	10);
2127	} else if (work->for_background)
2128	dirtied_before = jiffies;
2129
2130	queue_io(wb, work, dirtied_before);
2131	queued = true;
2132	}
2133	if (work->sb)
2134	progress = writeback_sb_inodes(work->sb, wb, work);
2135	else
2136	progress = __writeback_inodes_wb(wb, work);
2137	trace_writeback_written(wb, work);
2138
2139	/*
2140	* Did we write something? Try for more
2141	*
2142	* Dirty inodes are moved to b_io for writeback in batches.
2143	* The completion of the current batch does not necessarily
2144	* mean the overall work is done. So we keep looping as long
2145	* as made some progress on cleaning pages or inodes.
2146	*/
2147	if (progress \|\| !queued) {
2148	spin_unlock(&wb->list_lock);
2149	continue;
2150	}
2151
2152	/*
2153	* No more inodes for IO, bail
2154	*/
2155	if (list_empty(&wb->b_more_io)) {
2156	spin_unlock(&wb->list_lock);
2157	break;
2158	}
2159
2160	/*
2161	* Nothing written. Wait for some inode to
2162	* become available for writeback. Otherwise
2163	* we'll just busyloop.
2164	*/
2165	trace_writeback_wait(wb, work);
2166	inode = wb_inode(wb->b_more_io.prev);
2167	spin_lock(&inode->i_lock);
2168	spin_unlock(&wb->list_lock);
2169	/* This function drops i_lock... */
2170	inode_sleep_on_writeback(inode);
2171	}
2172	blk_finish_plug(&plug);
2173
2174	return nr_pages - work->nr_pages;
2175	}
2176
2177	/*
2178	* Return the next wb_writeback_work struct that hasn't been processed yet.
2179	*/
2180	static struct wb_writeback_work get_next_work_item(struct bdi_writeback wb)
2181	{
2182	struct wb_writeback_work *work = NULL;
2183
2184	spin_lock_irq(&wb->work_lock);
2185	if (!list_empty(&wb->work_list)) {
2186	work = list_entry(wb->work_list.next,
2187	struct wb_writeback_work, list);
2188	list_del_init(&work->list);
2189	}
2190	spin_unlock_irq(&wb->work_lock);
2191	return work;
2192	}
2193
2194	static long wb_check_background_flush(struct bdi_writeback *wb)
2195	{
2196	if (wb_over_bg_thresh(wb)) {
2197
2198	struct wb_writeback_work work = {
2199	.nr_pages = LONG_MAX,
2200	.sync_mode = WB_SYNC_NONE,
2201	.for_background = 1,
2202	.range_cyclic = 1,
2203	.reason = WB_REASON_BACKGROUND,
2204	};
2205
2206	return wb_writeback(wb, &work);
2207	}
2208
2209	return 0;
2210	}
2211
2212	static long wb_check_old_data_flush(struct bdi_writeback *wb)
2213	{
2214	unsigned long expired;
2215	long nr_pages;
2216
2217	/*
2218	* When set to zero, disable periodic writeback
2219	*/
2220	if (!dirty_writeback_interval)
2221	return 0;
2222
2223	expired = wb->last_old_flush +
2224	msecs_to_jiffies(dirty_writeback_interval * 10);
2225	if (time_before(jiffies, expired))
2226	return 0;
2227
2228	wb->last_old_flush = jiffies;
2229	nr_pages = get_nr_dirty_pages();
2230
2231	if (nr_pages) {
2232	struct wb_writeback_work work = {
2233	.nr_pages = nr_pages,
2234	.sync_mode = WB_SYNC_NONE,
2235	.for_kupdate = 1,
2236	.range_cyclic = 1,
2237	.reason = WB_REASON_PERIODIC,
2238	};
2239
2240	return wb_writeback(wb, &work);
2241	}
2242
2243	return 0;
2244	}
2245
2246	static long wb_check_start_all(struct bdi_writeback *wb)
2247	{
2248	long nr_pages;
2249
2250	if (!test_bit(WB_start_all, &wb->state))
2251	return 0;
2252
2253	nr_pages = get_nr_dirty_pages();
2254	if (nr_pages) {
2255	struct wb_writeback_work work = {
2256	.nr_pages = wb_split_bdi_pages(wb, nr_pages),
2257	.sync_mode = WB_SYNC_NONE,
2258	.range_cyclic = 1,
2259	.reason = wb->start_all_reason,
2260	};
2261
2262	nr_pages = wb_writeback(wb, &work);
2263	}
2264
2265	clear_bit(WB_start_all, &wb->state);
2266	return nr_pages;
2267	}
2268
2269
2270	/*
2271	* Retrieve work items and do the writeback they describe
2272	*/
2273	static long wb_do_writeback(struct bdi_writeback *wb)
2274	{
2275	struct wb_writeback_work *work;
2276	long wrote = 0;
2277
2278	set_bit(WB_writeback_running, &wb->state);
2279	while ((work = get_next_work_item(wb)) != NULL) {
2280	trace_writeback_exec(wb, work);
2281	wrote += wb_writeback(wb, work);
2282	finish_writeback_work(work);
2283	}
2284
2285	/*
2286	* Check for a flush-everything request
2287	*/
2288	wrote += wb_check_start_all(wb);
2289
2290	/*
2291	* Check for periodic writeback, kupdated() style
2292	*/
2293	wrote += wb_check_old_data_flush(wb);
2294	wrote += wb_check_background_flush(wb);
2295	clear_bit(WB_writeback_running, &wb->state);
2296
2297	return wrote;
2298	}
2299
2300	/*
2301	* Handle writeback of dirty data for the device backed by this bdi. Also
2302	* reschedules periodically and does kupdated style flushing.
2303	*/
2304	void wb_workfn(struct work_struct *work)
2305	{
2306	struct bdi_writeback *wb = container_of(to_delayed_work(work),
2307	struct bdi_writeback, dwork);
2308	long pages_written;
2309
2310	set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2311
2312	if (likely(!current_is_workqueue_rescuer() \|\|
2313	!test_bit(WB_registered, &wb->state))) {
2314	/*
2315	* The normal path. Keep writing back @wb until its
2316	* work_list is empty. Note that this path is also taken
2317	* if @wb is shutting down even when we're running off the
2318	* rescuer as work_list needs to be drained.
2319	*/
2320	do {
2321	pages_written = wb_do_writeback(wb);
2322	trace_writeback_pages_written(pages_written);
2323	} while (!list_empty(&wb->work_list));
2324	} else {
2325	/*
2326	* bdi_wq can't get enough workers and we're running off
2327	* the emergency worker. Don't hog it. Hopefully, 1024 is
2328	* enough for efficient IO.
2329	*/
2330	pages_written = writeback_inodes_wb(wb, 1024,
2331	WB_REASON_FORKER_THREAD);
2332	trace_writeback_pages_written(pages_written);
2333	}
2334
2335	if (!list_empty(&wb->work_list))
2336	wb_wakeup(wb);
2337	else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2338	wb_wakeup_delayed(wb);
2339	}
2340
2341	/*
2342	* Start writeback of all dirty pages on this bdi.
2343	*/
2344	static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2345	enum wb_reason reason)
2346	{
2347	struct bdi_writeback *wb;
2348
2349	if (!bdi_has_dirty_io(bdi))
2350	return;
2351
2352	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2353	wb_start_writeback(wb, reason);
2354	}
2355
2356	void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2357	enum wb_reason reason)
2358	{
2359	rcu_read_lock();
2360	__wakeup_flusher_threads_bdi(bdi, reason);
2361	rcu_read_unlock();
2362	}
2363
2364	/*
2365	* Wakeup the flusher threads to start writeback of all currently dirty pages
2366	*/
2367	void wakeup_flusher_threads(enum wb_reason reason)
2368	{
2369	struct backing_dev_info *bdi;
2370
2371	/*
2372	* If we are expecting writeback progress we must submit plugged IO.
2373	*/
2374	blk_flush_plug(current->plug, true);
2375
2376	rcu_read_lock();
2377	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2378	__wakeup_flusher_threads_bdi(bdi, reason);
2379	rcu_read_unlock();
2380	}
2381
2382	/*
2383	* Wake up bdi's periodically to make sure dirtytime inodes gets
2384	* written back periodically. We deliberately do not check the
2385	* b_dirtytime list in wb_has_dirty_io(), since this would cause the
2386	* kernel to be constantly waking up once there are any dirtytime
2387	* inodes on the system. So instead we define a separate delayed work
2388	* function which gets called much more rarely. (By default, only
2389	* once every 12 hours.)
2390	*
2391	* If there is any other write activity going on in the file system,
2392	* this function won't be necessary. But if the only thing that has
2393	* happened on the file system is a dirtytime inode caused by an atime
2394	* update, we need this infrastructure below to make sure that inode
2395	* eventually gets pushed out to disk.
2396	*/
2397	static void wakeup_dirtytime_writeback(struct work_struct *w);
2398	static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2399
2400	static void wakeup_dirtytime_writeback(struct work_struct *w)
2401	{
2402	struct backing_dev_info *bdi;
2403
2404	rcu_read_lock();
2405	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2406	struct bdi_writeback *wb;
2407
2408	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2409	if (!list_empty(&wb->b_dirty_time))
2410	wb_wakeup(wb);
2411	}
2412	rcu_read_unlock();
2413	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2414	}
2415
2416	static int __init start_dirtytime_writeback(void)
2417	{
2418	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2419	return 0;
2420	}
2421	__initcall(start_dirtytime_writeback);
2422
2423	int dirtytime_interval_handler(const struct ctl_table *table, int write,
2424	void buffer, size_t lenp, loff_t *ppos)
2425	{
2426	int ret;
2427
2428	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2429	if (ret == 0 && write)
2430	mod_delayed_work(system_wq, &dirtytime_work, 0);
2431	return ret;
2432	}
2433
2434	/**
2435	* __mark_inode_dirty - internal function to mark an inode dirty
2436	*
2437	* @inode: inode to mark
2438	* @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2439	* multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2440	* with I_DIRTY_PAGES.
2441	*
2442	* Mark an inode as dirty. We notify the filesystem, then update the inode's
2443	* dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2444	*
2445	* Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2446	* instead of calling this directly.
2447	*
2448	* CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2449	* refers to a blockdev. Unhashed inodes will never be added to the dirty list
2450	* even if they are later hashed, as they will have been marked dirty already.
2451	*
2452	* In short, ensure you hash any inodes _before_ you start marking them dirty.
2453	*
2454	* Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2455	* the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2456	* the kernel-internal blockdev inode represents the dirtying time of the
2457	* blockdev's pages. This is why for I_DIRTY_PAGES we always use
2458	* page->mapping->host, so the page-dirtying time is recorded in the internal
2459	* blockdev inode.
2460	*/
2461	void __mark_inode_dirty(struct inode *inode, int flags)
2462	{
2463	struct super_block *sb = inode->i_sb;
2464	int dirtytime = 0;
2465	struct bdi_writeback *wb = NULL;
2466
2467	trace_writeback_mark_inode_dirty(inode, flags);
2468
2469	if (flags & I_DIRTY_INODE) {
2470	/*
2471	* Inode timestamp update will piggback on this dirtying.
2472	* We tell ->dirty_inode callback that timestamps need to
2473	* be updated by setting I_DIRTY_TIME in flags.
2474	*/
2475	if (inode->i_state & I_DIRTY_TIME) {
2476	spin_lock(&inode->i_lock);
2477	if (inode->i_state & I_DIRTY_TIME) {
2478	inode->i_state &= ~I_DIRTY_TIME;
2479	flags \|= I_DIRTY_TIME;
2480	}
2481	spin_unlock(&inode->i_lock);
2482	}
2483
2484	/*
2485	* Notify the filesystem about the inode being dirtied, so that
2486	* (if needed) it can update on-disk fields and journal the
2487	* inode. This is only needed when the inode itself is being
2488	* dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2489	* for just I_DIRTY_PAGES or I_DIRTY_TIME.
2490	*/
2491	trace_writeback_dirty_inode_start(inode, flags);
2492	if (sb->s_op->dirty_inode)
2493	sb->s_op->dirty_inode(inode,
2494	flags & (I_DIRTY_INODE \| I_DIRTY_TIME));
2495	trace_writeback_dirty_inode(inode, flags);
2496
2497	/* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2498	flags &= ~I_DIRTY_TIME;
2499	} else {
2500	/*
2501	* Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2502	* (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2503	* in one call to __mark_inode_dirty().)
2504	*/
2505	dirtytime = flags & I_DIRTY_TIME;
2506	WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2507	}
2508
2509	/*
2510	* Paired with smp_mb() in __writeback_single_inode() for the
2511	* following lockless i_state test. See there for details.
2512	*/
2513	smp_mb();
2514
2515	if ((inode->i_state & flags) == flags)
2516	return;
2517
2518	spin_lock(&inode->i_lock);
2519	if ((inode->i_state & flags) != flags) {
2520	const int was_dirty = inode->i_state & I_DIRTY;
2521
2522	inode_attach_wb(inode, NULL);
2523
2524	inode->i_state \|= flags;
2525
2526	/*
2527	* Grab inode's wb early because it requires dropping i_lock and we
2528	* need to make sure following checks happen atomically with dirty
2529	* list handling so that we don't move inodes under flush worker's
2530	* hands.
2531	*/
2532	if (!was_dirty) {
2533	wb = locked_inode_to_wb_and_lock_list(inode);
2534	spin_lock(&inode->i_lock);
2535	}
2536
2537	/*
2538	* If the inode is queued for writeback by flush worker, just
2539	* update its dirty state. Once the flush worker is done with
2540	* the inode it will place it on the appropriate superblock
2541	* list, based upon its state.
2542	*/
2543	if (inode->i_state & I_SYNC_QUEUED)
2544	goto out_unlock;
2545
2546	/*
2547	* Only add valid (hashed) inodes to the superblock's
2548	* dirty list. Add blockdev inodes as well.
2549	*/
2550	if (!S_ISBLK(inode->i_mode)) {
2551	if (inode_unhashed(inode))
2552	goto out_unlock;
2553	}
2554	if (inode->i_state & I_FREEING)
2555	goto out_unlock;
2556
2557	/*
2558	* If the inode was already on b_dirty/b_io/b_more_io, don't
2559	* reposition it (that would break b_dirty time-ordering).
2560	*/
2561	if (!was_dirty) {
2562	struct list_head *dirty_list;
2563	bool wakeup_bdi = false;
2564
2565	inode->dirtied_when = jiffies;
2566	if (dirtytime)
2567	inode->dirtied_time_when = jiffies;
2568
2569	if (inode->i_state & I_DIRTY)
2570	dirty_list = &wb->b_dirty;
2571	else
2572	dirty_list = &wb->b_dirty_time;
2573
2574	wakeup_bdi = inode_io_list_move_locked(inode, wb,
2575	dirty_list);
2576
2577	spin_unlock(&wb->list_lock);
2578	spin_unlock(&inode->i_lock);
2579	trace_writeback_dirty_inode_enqueue(inode);
2580
2581	/*
2582	* If this is the first dirty inode for this bdi,
2583	* we have to wake-up the corresponding bdi thread
2584	* to make sure background write-back happens
2585	* later.
2586	*/
2587	if (wakeup_bdi &&
2588	(wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2589	wb_wakeup_delayed(wb);
2590	return;
2591	}
2592	}
2593	out_unlock:
2594	if (wb)
2595	spin_unlock(&wb->list_lock);
2596	spin_unlock(&inode->i_lock);
2597	}
2598	EXPORT_SYMBOL(__mark_inode_dirty);
2599
2600	/*
2601	* The @s_sync_lock is used to serialise concurrent sync operations
2602	* to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2603	* Concurrent callers will block on the s_sync_lock rather than doing contending
2604	* walks. The queueing maintains sync(2) required behaviour as all the IO that
2605	* has been issued up to the time this function is enter is guaranteed to be
2606	* completed by the time we have gained the lock and waited for all IO that is
2607	* in progress regardless of the order callers are granted the lock.
2608	*/
2609	static void wait_sb_inodes(struct super_block *sb)
2610	{
2611	LIST_HEAD(sync_list);
2612
2613	/*
2614	* We need to be protected against the filesystem going from
2615	* r/o to r/w or vice versa.
2616	*/
2617	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2618
2619	mutex_lock(&sb->s_sync_lock);
2620
2621	/*
2622	* Splice the writeback list onto a temporary list to avoid waiting on
2623	* inodes that have started writeback after this point.
2624	*
2625	* Use rcu_read_lock() to keep the inodes around until we have a
2626	* reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2627	* the local list because inodes can be dropped from either by writeback
2628	* completion.
2629	*/
2630	rcu_read_lock();
2631	spin_lock_irq(&sb->s_inode_wblist_lock);
2632	list_splice_init(&sb->s_inodes_wb, &sync_list);
2633
2634	/*
2635	* Data integrity sync. Must wait for all pages under writeback, because
2636	* there may have been pages dirtied before our sync call, but which had
2637	* writeout started before we write it out. In which case, the inode
2638	* may not be on the dirty list, but we still have to wait for that
2639	* writeout.
2640	*/
2641	while (!list_empty(&sync_list)) {
2642	struct inode *inode = list_first_entry(&sync_list, struct inode,
2643	i_wb_list);
2644	struct address_space *mapping = inode->i_mapping;
2645
2646	/*
2647	* Move each inode back to the wb list before we drop the lock
2648	* to preserve consistency between i_wb_list and the mapping
2649	* writeback tag. Writeback completion is responsible to remove
2650	* the inode from either list once the writeback tag is cleared.
2651	*/
2652	list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2653
2654	/*
2655	* The mapping can appear untagged while still on-list since we
2656	* do not have the mapping lock. Skip it here, wb completion
2657	* will remove it.
2658	*/
2659	if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2660	continue;
2661
2662	spin_unlock_irq(&sb->s_inode_wblist_lock);
2663
2664	spin_lock(&inode->i_lock);
2665	if (inode->i_state & (I_FREEING\|I_WILL_FREE\|I_NEW)) {
2666	spin_unlock(&inode->i_lock);
2667
2668	spin_lock_irq(&sb->s_inode_wblist_lock);
2669	continue;
2670	}
2671	__iget(inode);
2672	spin_unlock(&inode->i_lock);
2673	rcu_read_unlock();
2674
2675	/*
2676	* We keep the error status of individual mapping so that
2677	* applications can catch the writeback error using fsync(2).
2678	* See filemap_fdatawait_keep_errors() for details.
2679	*/
2680	filemap_fdatawait_keep_errors(mapping);
2681
2682	cond_resched();
2683
2684	iput(inode);
2685
2686	rcu_read_lock();
2687	spin_lock_irq(&sb->s_inode_wblist_lock);
2688	}
2689	spin_unlock_irq(&sb->s_inode_wblist_lock);
2690	rcu_read_unlock();
2691	mutex_unlock(&sb->s_sync_lock);
2692	}
2693
2694	static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2695	enum wb_reason reason, bool skip_if_busy)
2696	{
2697	struct backing_dev_info *bdi = sb->s_bdi;
2698	DEFINE_WB_COMPLETION(done, bdi);
2699	struct wb_writeback_work work = {
2700	.sb = sb,
2701	.sync_mode = WB_SYNC_NONE,
2702	.tagged_writepages = 1,
2703	.done = &done,
2704	.nr_pages = nr,
2705	.reason = reason,
2706	};
2707
2708	if (!bdi_has_dirty_io(bdi) \|\| bdi == &noop_backing_dev_info)
2709	return;
2710	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2711
2712	bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2713	wb_wait_for_completion(&done);
2714	}
2715
2716	/**
2717	* writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2718	* @sb: the superblock
2719	* @nr: the number of pages to write
2720	* @reason: reason why some writeback work initiated
2721	*
2722	* Start writeback on some inodes on this super_block. No guarantees are made
2723	* on how many (if any) will be written, and this function does not wait
2724	* for IO completion of submitted IO.
2725	*/
2726	void writeback_inodes_sb_nr(struct super_block *sb,
2727	unsigned long nr,
2728	enum wb_reason reason)
2729	{
2730	__writeback_inodes_sb_nr(sb, nr, reason, false);
2731	}
2732	EXPORT_SYMBOL(writeback_inodes_sb_nr);
2733
2734	/**
2735	* writeback_inodes_sb - writeback dirty inodes from given super_block
2736	* @sb: the superblock
2737	* @reason: reason why some writeback work was initiated
2738	*
2739	* Start writeback on some inodes on this super_block. No guarantees are made
2740	* on how many (if any) will be written, and this function does not wait
2741	* for IO completion of submitted IO.
2742	*/
2743	void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2744	{
2745	writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2746	}
2747	EXPORT_SYMBOL(writeback_inodes_sb);
2748
2749	/**
2750	* try_to_writeback_inodes_sb - try to start writeback if none underway
2751	* @sb: the superblock
2752	* @reason: reason why some writeback work was initiated
2753	*
2754	* Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2755	*/
2756	void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2757	{
2758	if (!down_read_trylock(&sb->s_umount))
2759	return;
2760
2761	__writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2762	up_read(&sb->s_umount);
2763	}
2764	EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2765
2766	/**
2767	* sync_inodes_sb - sync sb inode pages
2768	* @sb: the superblock
2769	*
2770	* This function writes and waits on any dirty inode belonging to this
2771	* super_block.
2772	*/
2773	void sync_inodes_sb(struct super_block *sb)
2774	{
2775	struct backing_dev_info *bdi = sb->s_bdi;
2776	DEFINE_WB_COMPLETION(done, bdi);
2777	struct wb_writeback_work work = {
2778	.sb = sb,
2779	.sync_mode = WB_SYNC_ALL,
2780	.nr_pages = LONG_MAX,
2781	.range_cyclic = 0,
2782	.done = &done,
2783	.reason = WB_REASON_SYNC,
2784	.for_sync = 1,
2785	};
2786
2787	/*
2788	* Can't skip on !bdi_has_dirty() because we should wait for !dirty
2789	* inodes under writeback and I_DIRTY_TIME inodes ignored by
2790	* bdi_has_dirty() need to be written out too.
2791	*/
2792	if (bdi == &noop_backing_dev_info)
2793	return;
2794	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2795
2796	/* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2797	bdi_down_write_wb_switch_rwsem(bdi);
2798	bdi_split_work_to_wbs(bdi, &work, false);
2799	wb_wait_for_completion(&done);
2800	bdi_up_write_wb_switch_rwsem(bdi);
2801
2802	wait_sb_inodes(sb);
2803	}
2804	EXPORT_SYMBOL(sync_inodes_sb);
2805
2806	/**
2807	* write_inode_now - write an inode to disk
2808	* @inode: inode to write to disk
2809	* @sync: whether the write should be synchronous or not
2810	*
2811	* This function commits an inode to disk immediately if it is dirty. This is
2812	* primarily needed by knfsd.
2813	*
2814	* The caller must either have a ref on the inode or must have set I_WILL_FREE.
2815	*/
2816	int write_inode_now(struct inode *inode, int sync)
2817	{
2818	struct writeback_control wbc = {
2819	.nr_to_write = LONG_MAX,
2820	.sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2821	.range_start = 0,
2822	.range_end = LLONG_MAX,
2823	};
2824
2825	if (!mapping_can_writeback(inode->i_mapping))
2826	wbc.nr_to_write = 0;
2827
2828	might_sleep();
2829	return writeback_single_inode(inode, &wbc);
2830	}
2831	EXPORT_SYMBOL(write_inode_now);
2832
2833	/**
2834	* sync_inode_metadata - write an inode to disk
2835	* @inode: the inode to sync
2836	* @wait: wait for I/O to complete.
2837	*
2838	* Write an inode to disk and adjust its dirty state after completion.
2839	*
2840	* Note: only writes the actual inode, no associated data or other metadata.
2841	*/
2842	int sync_inode_metadata(struct inode *inode, int wait)
2843	{
2844	struct writeback_control wbc = {
2845	.sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2846	.nr_to_write = 0, /* metadata-only */
2847	};
2848
2849	return writeback_single_inode(inode, &wbc);
2850	}
2851	EXPORT_SYMBOL(sync_inode_metadata);