fs-writeback.c revision 633a2abb
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 */
42struct 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 */
68unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
70static 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
83EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
85static 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
98static 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 */
118static 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
124	list_move(&inode->i_io_list, head);
125
126	/* dirty_time doesn't count as dirty_io until expiration */
127	if (head != &wb->b_dirty_time)
128		return wb_io_lists_populated(wb);
129
130	wb_io_lists_depopulated(wb);
131	return false;
132}
133
134static void wb_wakeup(struct bdi_writeback *wb)
135{
136	spin_lock_bh(&wb->work_lock);
137	if (test_bit(WB_registered, &wb->state))
138		mod_delayed_work(bdi_wq, &wb->dwork, 0);
139	spin_unlock_bh(&wb->work_lock);
140}
141
142static void finish_writeback_work(struct bdi_writeback *wb,
143				  struct wb_writeback_work *work)
144{
145	struct wb_completion *done = work->done;
146
147	if (work->auto_free)
148		kfree(work);
149	if (done) {
150		wait_queue_head_t *waitq = done->waitq;
151
152		/* @done can't be accessed after the following dec */
153		if (atomic_dec_and_test(&done->cnt))
154			wake_up_all(waitq);
155	}
156}
157
158static void wb_queue_work(struct bdi_writeback *wb,
159			  struct wb_writeback_work *work)
160{
161	trace_writeback_queue(wb, work);
162
163	if (work->done)
164		atomic_inc(&work->done->cnt);
165
166	spin_lock_bh(&wb->work_lock);
167
168	if (test_bit(WB_registered, &wb->state)) {
169		list_add_tail(&work->list, &wb->work_list);
170		mod_delayed_work(bdi_wq, &wb->dwork, 0);
171	} else
172		finish_writeback_work(wb, work);
173
174	spin_unlock_bh(&wb->work_lock);
175}
176
177/**
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
179 * @done: target wb_completion
180 *
181 * Wait for one or more work items issued to @bdi with their ->done field
182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION().  This function returns after all such work items
184 * are completed.  Work items which are waited upon aren't freed
185 * automatically on completion.
186 */
187void wb_wait_for_completion(struct wb_completion *done)
188{
189	atomic_dec(&done->cnt);		/* put down the initial count */
190	wait_event(*done->waitq, !atomic_read(&done->cnt));
191}
192
193#ifdef CONFIG_CGROUP_WRITEBACK
194
195/*
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
197 * how they're used.
198 *
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy.  All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
202 *
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported.  As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
208 *
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch.  The recording is quantized
211 * to 16 slots.  To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
213 */
214#define WB_FRN_TIME_SHIFT	13	/* 1s = 2^13, upto 8 secs w/ 16bit */
215#define WB_FRN_TIME_AVG_SHIFT	3	/* avg = avg * 7/8 + new * 1/8 */
216#define WB_FRN_TIME_CUT_DIV	8	/* ignore rounds < avg / 8 */
217#define WB_FRN_TIME_PERIOD	(2 * (1 << WB_FRN_TIME_SHIFT))	/* 2s */
218
219#define WB_FRN_HIST_SLOTS	16	/* inode->i_wb_frn_history is 16bit */
220#define WB_FRN_HIST_UNIT	(WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221					/* each slot's duration is 2s / 16 */
222#define WB_FRN_HIST_THR_SLOTS	(WB_FRN_HIST_SLOTS / 2)
223					/* if foreign slots >= 8, switch */
224#define WB_FRN_HIST_MAX_SLOTS	(WB_FRN_HIST_THR_SLOTS / 2 + 1)
225					/* one round can affect upto 5 slots */
226#define WB_FRN_MAX_IN_FLIGHT	1024	/* don't queue too many concurrently */
227
228/*
229 * Maximum inodes per isw.  A specific value has been chosen to make
230 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
231 */
232#define WB_MAX_INODES_PER_ISW  ((1024UL - sizeof(struct inode_switch_wbs_context)) \
233                                / sizeof(struct inode *))
234
235static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
236static struct workqueue_struct *isw_wq;
237
238void __inode_attach_wb(struct inode *inode, struct page *page)
239{
240	struct backing_dev_info *bdi = inode_to_bdi(inode);
241	struct bdi_writeback *wb = NULL;
242
243	if (inode_cgwb_enabled(inode)) {
244		struct cgroup_subsys_state *memcg_css;
245
246		if (page) {
247			memcg_css = mem_cgroup_css_from_page(page);
248			wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
249		} else {
250			/* must pin memcg_css, see wb_get_create() */
251			memcg_css = task_get_css(current, memory_cgrp_id);
252			wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
253			css_put(memcg_css);
254		}
255	}
256
257	if (!wb)
258		wb = &bdi->wb;
259
260	/*
261	 * There may be multiple instances of this function racing to
262	 * update the same inode.  Use cmpxchg() to tell the winner.
263	 */
264	if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
265		wb_put(wb);
266}
267EXPORT_SYMBOL_GPL(__inode_attach_wb);
268
269/**
270 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
271 * @inode: inode of interest with i_lock held
272 * @wb: target bdi_writeback
273 *
274 * Remove the inode from wb's io lists and if necessarily put onto b_attached
275 * list.  Only inodes attached to cgwb's are kept on this list.
276 */
277static void inode_cgwb_move_to_attached(struct inode *inode,
278					struct bdi_writeback *wb)
279{
280	assert_spin_locked(&wb->list_lock);
281	assert_spin_locked(&inode->i_lock);
282
283	inode->i_state &= ~I_SYNC_QUEUED;
284	if (wb != &wb->bdi->wb)
285		list_move(&inode->i_io_list, &wb->b_attached);
286	else
287		list_del_init(&inode->i_io_list);
288	wb_io_lists_depopulated(wb);
289}
290
291/**
292 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
293 * @inode: inode of interest with i_lock held
294 *
295 * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
296 * held on entry and is released on return.  The returned wb is guaranteed
297 * to stay @inode's associated wb until its list_lock is released.
298 */
299static struct bdi_writeback *
300locked_inode_to_wb_and_lock_list(struct inode *inode)
301	__releases(&inode->i_lock)
302	__acquires(&wb->list_lock)
303{
304	while (true) {
305		struct bdi_writeback *wb = inode_to_wb(inode);
306
307		/*
308		 * inode_to_wb() association is protected by both
309		 * @inode->i_lock and @wb->list_lock but list_lock nests
310		 * outside i_lock.  Drop i_lock and verify that the
311		 * association hasn't changed after acquiring list_lock.
312		 */
313		wb_get(wb);
314		spin_unlock(&inode->i_lock);
315		spin_lock(&wb->list_lock);
316
317		/* i_wb may have changed inbetween, can't use inode_to_wb() */
318		if (likely(wb == inode->i_wb)) {
319			wb_put(wb);	/* @inode already has ref */
320			return wb;
321		}
322
323		spin_unlock(&wb->list_lock);
324		wb_put(wb);
325		cpu_relax();
326		spin_lock(&inode->i_lock);
327	}
328}
329
330/**
331 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
332 * @inode: inode of interest
333 *
334 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
335 * on entry.
336 */
337static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
338	__acquires(&wb->list_lock)
339{
340	spin_lock(&inode->i_lock);
341	return locked_inode_to_wb_and_lock_list(inode);
342}
343
344struct inode_switch_wbs_context {
345	struct rcu_work		work;
346
347	/*
348	 * Multiple inodes can be switched at once.  The switching procedure
349	 * consists of two parts, separated by a RCU grace period.  To make
350	 * sure that the second part is executed for each inode gone through
351	 * the first part, all inode pointers are placed into a NULL-terminated
352	 * array embedded into struct inode_switch_wbs_context.  Otherwise
353	 * an inode could be left in a non-consistent state.
354	 */
355	struct bdi_writeback	*new_wb;
356	struct inode		*inodes[];
357};
358
359static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
360{
361	down_write(&bdi->wb_switch_rwsem);
362}
363
364static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
365{
366	up_write(&bdi->wb_switch_rwsem);
367}
368
369static bool inode_do_switch_wbs(struct inode *inode,
370				struct bdi_writeback *old_wb,
371				struct bdi_writeback *new_wb)
372{
373	struct address_space *mapping = inode->i_mapping;
374	XA_STATE(xas, &mapping->i_pages, 0);
375	struct page *page;
376	bool switched = false;
377
378	spin_lock(&inode->i_lock);
379	xa_lock_irq(&mapping->i_pages);
380
381	/*
382	 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
383	 * path owns the inode and we shouldn't modify ->i_io_list.
384	 */
385	if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
386		goto skip_switch;
387
388	trace_inode_switch_wbs(inode, old_wb, new_wb);
389
390	/*
391	 * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
392	 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393	 * pages actually under writeback.
394	 */
395	xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
396		if (PageDirty(page)) {
397			dec_wb_stat(old_wb, WB_RECLAIMABLE);
398			inc_wb_stat(new_wb, WB_RECLAIMABLE);
399		}
400	}
401
402	xas_set(&xas, 0);
403	xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
404		WARN_ON_ONCE(!PageWriteback(page));
405		dec_wb_stat(old_wb, WB_WRITEBACK);
406		inc_wb_stat(new_wb, WB_WRITEBACK);
407	}
408
409	if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
410		atomic_dec(&old_wb->writeback_inodes);
411		atomic_inc(&new_wb->writeback_inodes);
412	}
413
414	wb_get(new_wb);
415
416	/*
417	 * Transfer to @new_wb's IO list if necessary.  If the @inode is dirty,
418	 * the specific list @inode was on is ignored and the @inode is put on
419	 * ->b_dirty which is always correct including from ->b_dirty_time.
420	 * The transfer preserves @inode->dirtied_when ordering.  If the @inode
421	 * was clean, it means it was on the b_attached list, so move it onto
422	 * the b_attached list of @new_wb.
423	 */
424	if (!list_empty(&inode->i_io_list)) {
425		inode->i_wb = new_wb;
426
427		if (inode->i_state & I_DIRTY_ALL) {
428			struct inode *pos;
429
430			list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
431				if (time_after_eq(inode->dirtied_when,
432						  pos->dirtied_when))
433					break;
434			inode_io_list_move_locked(inode, new_wb,
435						  pos->i_io_list.prev);
436		} else {
437			inode_cgwb_move_to_attached(inode, new_wb);
438		}
439	} else {
440		inode->i_wb = new_wb;
441	}
442
443	/* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
444	inode->i_wb_frn_winner = 0;
445	inode->i_wb_frn_avg_time = 0;
446	inode->i_wb_frn_history = 0;
447	switched = true;
448skip_switch:
449	/*
450	 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
451	 * ensures that the new wb is visible if they see !I_WB_SWITCH.
452	 */
453	smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
454
455	xa_unlock_irq(&mapping->i_pages);
456	spin_unlock(&inode->i_lock);
457
458	return switched;
459}
460
461static void inode_switch_wbs_work_fn(struct work_struct *work)
462{
463	struct inode_switch_wbs_context *isw =
464		container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
465	struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
466	struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
467	struct bdi_writeback *new_wb = isw->new_wb;
468	unsigned long nr_switched = 0;
469	struct inode **inodep;
470
471	/*
472	 * If @inode switches cgwb membership while sync_inodes_sb() is
473	 * being issued, sync_inodes_sb() might miss it.  Synchronize.
474	 */
475	down_read(&bdi->wb_switch_rwsem);
476
477	/*
478	 * By the time control reaches here, RCU grace period has passed
479	 * since I_WB_SWITCH assertion and all wb stat update transactions
480	 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
481	 * synchronizing against the i_pages lock.
482	 *
483	 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
484	 * gives us exclusion against all wb related operations on @inode
485	 * including IO list manipulations and stat updates.
486	 */
487	if (old_wb < new_wb) {
488		spin_lock(&old_wb->list_lock);
489		spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
490	} else {
491		spin_lock(&new_wb->list_lock);
492		spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
493	}
494
495	for (inodep = isw->inodes; *inodep; inodep++) {
496		WARN_ON_ONCE((*inodep)->i_wb != old_wb);
497		if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
498			nr_switched++;
499	}
500
501	spin_unlock(&new_wb->list_lock);
502	spin_unlock(&old_wb->list_lock);
503
504	up_read(&bdi->wb_switch_rwsem);
505
506	if (nr_switched) {
507		wb_wakeup(new_wb);
508		wb_put_many(old_wb, nr_switched);
509	}
510
511	for (inodep = isw->inodes; *inodep; inodep++)
512		iput(*inodep);
513	wb_put(new_wb);
514	kfree(isw);
515	atomic_dec(&isw_nr_in_flight);
516}
517
518static bool inode_prepare_wbs_switch(struct inode *inode,
519				     struct bdi_writeback *new_wb)
520{
521	/*
522	 * Paired with smp_mb() in cgroup_writeback_umount().
523	 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
524	 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
525	 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
526	 */
527	smp_mb();
528
529	if (IS_DAX(inode))
530		return false;
531
532	/* while holding I_WB_SWITCH, no one else can update the association */
533	spin_lock(&inode->i_lock);
534	if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
535	    inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
536	    inode_to_wb(inode) == new_wb) {
537		spin_unlock(&inode->i_lock);
538		return false;
539	}
540	inode->i_state |= I_WB_SWITCH;
541	__iget(inode);
542	spin_unlock(&inode->i_lock);
543
544	return true;
545}
546
547/**
548 * inode_switch_wbs - change the wb association of an inode
549 * @inode: target inode
550 * @new_wb_id: ID of the new wb
551 *
552 * Switch @inode's wb association to the wb identified by @new_wb_id.  The
553 * switching is performed asynchronously and may fail silently.
554 */
555static void inode_switch_wbs(struct inode *inode, int new_wb_id)
556{
557	struct backing_dev_info *bdi = inode_to_bdi(inode);
558	struct cgroup_subsys_state *memcg_css;
559	struct inode_switch_wbs_context *isw;
560
561	/* noop if seems to be already in progress */
562	if (inode->i_state & I_WB_SWITCH)
563		return;
564
565	/* avoid queueing a new switch if too many are already in flight */
566	if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
567		return;
568
569	isw = kzalloc(sizeof(*isw) + 2 * sizeof(struct inode *), GFP_ATOMIC);
570	if (!isw)
571		return;
572
573	atomic_inc(&isw_nr_in_flight);
574
575	/* find and pin the new wb */
576	rcu_read_lock();
577	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
578	if (memcg_css && !css_tryget(memcg_css))
579		memcg_css = NULL;
580	rcu_read_unlock();
581	if (!memcg_css)
582		goto out_free;
583
584	isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
585	css_put(memcg_css);
586	if (!isw->new_wb)
587		goto out_free;
588
589	if (!inode_prepare_wbs_switch(inode, isw->new_wb))
590		goto out_free;
591
592	isw->inodes[0] = inode;
593
594	/*
595	 * In addition to synchronizing among switchers, I_WB_SWITCH tells
596	 * the RCU protected stat update paths to grab the i_page
597	 * lock so that stat transfer can synchronize against them.
598	 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
599	 */
600	INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
601	queue_rcu_work(isw_wq, &isw->work);
602	return;
603
604out_free:
605	atomic_dec(&isw_nr_in_flight);
606	if (isw->new_wb)
607		wb_put(isw->new_wb);
608	kfree(isw);
609}
610
611/**
612 * cleanup_offline_cgwb - detach associated inodes
613 * @wb: target wb
614 *
615 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
616 * to eventually release the dying @wb.  Returns %true if not all inodes were
617 * switched and the function has to be restarted.
618 */
619bool cleanup_offline_cgwb(struct bdi_writeback *wb)
620{
621	struct cgroup_subsys_state *memcg_css;
622	struct inode_switch_wbs_context *isw;
623	struct inode *inode;
624	int nr;
625	bool restart = false;
626
627	isw = kzalloc(sizeof(*isw) + WB_MAX_INODES_PER_ISW *
628		      sizeof(struct inode *), GFP_KERNEL);
629	if (!isw)
630		return restart;
631
632	atomic_inc(&isw_nr_in_flight);
633
634	for (memcg_css = wb->memcg_css->parent; memcg_css;
635	     memcg_css = memcg_css->parent) {
636		isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
637		if (isw->new_wb)
638			break;
639	}
640	if (unlikely(!isw->new_wb))
641		isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
642
643	nr = 0;
644	spin_lock(&wb->list_lock);
645	list_for_each_entry(inode, &wb->b_attached, i_io_list) {
646		if (!inode_prepare_wbs_switch(inode, isw->new_wb))
647			continue;
648
649		isw->inodes[nr++] = inode;
650
651		if (nr >= WB_MAX_INODES_PER_ISW - 1) {
652			restart = true;
653			break;
654		}
655	}
656	spin_unlock(&wb->list_lock);
657
658	/* no attached inodes? bail out */
659	if (nr == 0) {
660		atomic_dec(&isw_nr_in_flight);
661		wb_put(isw->new_wb);
662		kfree(isw);
663		return restart;
664	}
665
666	/*
667	 * In addition to synchronizing among switchers, I_WB_SWITCH tells
668	 * the RCU protected stat update paths to grab the i_page
669	 * lock so that stat transfer can synchronize against them.
670	 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
671	 */
672	INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
673	queue_rcu_work(isw_wq, &isw->work);
674
675	return restart;
676}
677
678/**
679 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
680 * @wbc: writeback_control of interest
681 * @inode: target inode
682 *
683 * @inode is locked and about to be written back under the control of @wbc.
684 * Record @inode's writeback context into @wbc and unlock the i_lock.  On
685 * writeback completion, wbc_detach_inode() should be called.  This is used
686 * to track the cgroup writeback context.
687 */
688void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
689				 struct inode *inode)
690{
691	if (!inode_cgwb_enabled(inode)) {
692		spin_unlock(&inode->i_lock);
693		return;
694	}
695
696	wbc->wb = inode_to_wb(inode);
697	wbc->inode = inode;
698
699	wbc->wb_id = wbc->wb->memcg_css->id;
700	wbc->wb_lcand_id = inode->i_wb_frn_winner;
701	wbc->wb_tcand_id = 0;
702	wbc->wb_bytes = 0;
703	wbc->wb_lcand_bytes = 0;
704	wbc->wb_tcand_bytes = 0;
705
706	wb_get(wbc->wb);
707	spin_unlock(&inode->i_lock);
708
709	/*
710	 * A dying wb indicates that either the blkcg associated with the
711	 * memcg changed or the associated memcg is dying.  In the first
712	 * case, a replacement wb should already be available and we should
713	 * refresh the wb immediately.  In the second case, trying to
714	 * refresh will keep failing.
715	 */
716	if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
717		inode_switch_wbs(inode, wbc->wb_id);
718}
719EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
720
721/**
722 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
723 * @wbc: writeback_control of the just finished writeback
724 *
725 * To be called after a writeback attempt of an inode finishes and undoes
726 * wbc_attach_and_unlock_inode().  Can be called under any context.
727 *
728 * As concurrent write sharing of an inode is expected to be very rare and
729 * memcg only tracks page ownership on first-use basis severely confining
730 * the usefulness of such sharing, cgroup writeback tracks ownership
731 * per-inode.  While the support for concurrent write sharing of an inode
732 * is deemed unnecessary, an inode being written to by different cgroups at
733 * different points in time is a lot more common, and, more importantly,
734 * charging only by first-use can too readily lead to grossly incorrect
735 * behaviors (single foreign page can lead to gigabytes of writeback to be
736 * incorrectly attributed).
737 *
738 * To resolve this issue, cgroup writeback detects the majority dirtier of
739 * an inode and transfers the ownership to it.  To avoid unnnecessary
740 * oscillation, the detection mechanism keeps track of history and gives
741 * out the switch verdict only if the foreign usage pattern is stable over
742 * a certain amount of time and/or writeback attempts.
743 *
744 * On each writeback attempt, @wbc tries to detect the majority writer
745 * using Boyer-Moore majority vote algorithm.  In addition to the byte
746 * count from the majority voting, it also counts the bytes written for the
747 * current wb and the last round's winner wb (max of last round's current
748 * wb, the winner from two rounds ago, and the last round's majority
749 * candidate).  Keeping track of the historical winner helps the algorithm
750 * to semi-reliably detect the most active writer even when it's not the
751 * absolute majority.
752 *
753 * Once the winner of the round is determined, whether the winner is
754 * foreign or not and how much IO time the round consumed is recorded in
755 * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
756 * over a certain threshold, the switch verdict is given.
757 */
758void wbc_detach_inode(struct writeback_control *wbc)
759{
760	struct bdi_writeback *wb = wbc->wb;
761	struct inode *inode = wbc->inode;
762	unsigned long avg_time, max_bytes, max_time;
763	u16 history;
764	int max_id;
765
766	if (!wb)
767		return;
768
769	history = inode->i_wb_frn_history;
770	avg_time = inode->i_wb_frn_avg_time;
771
772	/* pick the winner of this round */
773	if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
774	    wbc->wb_bytes >= wbc->wb_tcand_bytes) {
775		max_id = wbc->wb_id;
776		max_bytes = wbc->wb_bytes;
777	} else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
778		max_id = wbc->wb_lcand_id;
779		max_bytes = wbc->wb_lcand_bytes;
780	} else {
781		max_id = wbc->wb_tcand_id;
782		max_bytes = wbc->wb_tcand_bytes;
783	}
784
785	/*
786	 * Calculate the amount of IO time the winner consumed and fold it
787	 * into the running average kept per inode.  If the consumed IO
788	 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
789	 * deciding whether to switch or not.  This is to prevent one-off
790	 * small dirtiers from skewing the verdict.
791	 */
792	max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
793				wb->avg_write_bandwidth);
794	if (avg_time)
795		avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
796			    (avg_time >> WB_FRN_TIME_AVG_SHIFT);
797	else
798		avg_time = max_time;	/* immediate catch up on first run */
799
800	if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
801		int slots;
802
803		/*
804		 * The switch verdict is reached if foreign wb's consume
805		 * more than a certain proportion of IO time in a
806		 * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
807		 * history mask where each bit represents one sixteenth of
808		 * the period.  Determine the number of slots to shift into
809		 * history from @max_time.
810		 */
811		slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
812			    (unsigned long)WB_FRN_HIST_MAX_SLOTS);
813		history <<= slots;
814		if (wbc->wb_id != max_id)
815			history |= (1U << slots) - 1;
816
817		if (history)
818			trace_inode_foreign_history(inode, wbc, history);
819
820		/*
821		 * Switch if the current wb isn't the consistent winner.
822		 * If there are multiple closely competing dirtiers, the
823		 * inode may switch across them repeatedly over time, which
824		 * is okay.  The main goal is avoiding keeping an inode on
825		 * the wrong wb for an extended period of time.
826		 */
827		if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
828			inode_switch_wbs(inode, max_id);
829	}
830
831	/*
832	 * Multiple instances of this function may race to update the
833	 * following fields but we don't mind occassional inaccuracies.
834	 */
835	inode->i_wb_frn_winner = max_id;
836	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
837	inode->i_wb_frn_history = history;
838
839	wb_put(wbc->wb);
840	wbc->wb = NULL;
841}
842EXPORT_SYMBOL_GPL(wbc_detach_inode);
843
844/**
845 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
846 * @wbc: writeback_control of the writeback in progress
847 * @page: page being written out
848 * @bytes: number of bytes being written out
849 *
850 * @bytes from @page are about to written out during the writeback
851 * controlled by @wbc.  Keep the book for foreign inode detection.  See
852 * wbc_detach_inode().
853 */
854void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
855			      size_t bytes)
856{
857	struct cgroup_subsys_state *css;
858	int id;
859
860	/*
861	 * pageout() path doesn't attach @wbc to the inode being written
862	 * out.  This is intentional as we don't want the function to block
863	 * behind a slow cgroup.  Ultimately, we want pageout() to kick off
864	 * regular writeback instead of writing things out itself.
865	 */
866	if (!wbc->wb || wbc->no_cgroup_owner)
867		return;
868
869	css = mem_cgroup_css_from_page(page);
870	/* dead cgroups shouldn't contribute to inode ownership arbitration */
871	if (!(css->flags & CSS_ONLINE))
872		return;
873
874	id = css->id;
875
876	if (id == wbc->wb_id) {
877		wbc->wb_bytes += bytes;
878		return;
879	}
880
881	if (id == wbc->wb_lcand_id)
882		wbc->wb_lcand_bytes += bytes;
883
884	/* Boyer-Moore majority vote algorithm */
885	if (!wbc->wb_tcand_bytes)
886		wbc->wb_tcand_id = id;
887	if (id == wbc->wb_tcand_id)
888		wbc->wb_tcand_bytes += bytes;
889	else
890		wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
891}
892EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
893
894/**
895 * inode_congested - test whether an inode is congested
896 * @inode: inode to test for congestion (may be NULL)
897 * @cong_bits: mask of WB_[a]sync_congested bits to test
898 *
899 * Tests whether @inode is congested.  @cong_bits is the mask of congestion
900 * bits to test and the return value is the mask of set bits.
901 *
902 * If cgroup writeback is enabled for @inode, the congestion state is
903 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
904 * associated with @inode is congested; otherwise, the root wb's congestion
905 * state is used.
906 *
907 * @inode is allowed to be NULL as this function is often called on
908 * mapping->host which is NULL for the swapper space.
909 */
910int inode_congested(struct inode *inode, int cong_bits)
911{
912	/*
913	 * Once set, ->i_wb never becomes NULL while the inode is alive.
914	 * Start transaction iff ->i_wb is visible.
915	 */
916	if (inode && inode_to_wb_is_valid(inode)) {
917		struct bdi_writeback *wb;
918		struct wb_lock_cookie lock_cookie = {};
919		bool congested;
920
921		wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
922		congested = wb_congested(wb, cong_bits);
923		unlocked_inode_to_wb_end(inode, &lock_cookie);
924		return congested;
925	}
926
927	return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
928}
929EXPORT_SYMBOL_GPL(inode_congested);
930
931/**
932 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
933 * @wb: target bdi_writeback to split @nr_pages to
934 * @nr_pages: number of pages to write for the whole bdi
935 *
936 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
937 * relation to the total write bandwidth of all wb's w/ dirty inodes on
938 * @wb->bdi.
939 */
940static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
941{
942	unsigned long this_bw = wb->avg_write_bandwidth;
943	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
944
945	if (nr_pages == LONG_MAX)
946		return LONG_MAX;
947
948	/*
949	 * This may be called on clean wb's and proportional distribution
950	 * may not make sense, just use the original @nr_pages in those
951	 * cases.  In general, we wanna err on the side of writing more.
952	 */
953	if (!tot_bw || this_bw >= tot_bw)
954		return nr_pages;
955	else
956		return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
957}
958
959/**
960 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
961 * @bdi: target backing_dev_info
962 * @base_work: wb_writeback_work to issue
963 * @skip_if_busy: skip wb's which already have writeback in progress
964 *
965 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
966 * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
967 * distributed to the busy wbs according to each wb's proportion in the
968 * total active write bandwidth of @bdi.
969 */
970static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
971				  struct wb_writeback_work *base_work,
972				  bool skip_if_busy)
973{
974	struct bdi_writeback *last_wb = NULL;
975	struct bdi_writeback *wb = list_entry(&bdi->wb_list,
976					      struct bdi_writeback, bdi_node);
977
978	might_sleep();
979restart:
980	rcu_read_lock();
981	list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
982		DEFINE_WB_COMPLETION(fallback_work_done, bdi);
983		struct wb_writeback_work fallback_work;
984		struct wb_writeback_work *work;
985		long nr_pages;
986
987		if (last_wb) {
988			wb_put(last_wb);
989			last_wb = NULL;
990		}
991
992		/* SYNC_ALL writes out I_DIRTY_TIME too */
993		if (!wb_has_dirty_io(wb) &&
994		    (base_work->sync_mode == WB_SYNC_NONE ||
995		     list_empty(&wb->b_dirty_time)))
996			continue;
997		if (skip_if_busy && writeback_in_progress(wb))
998			continue;
999
1000		nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1001
1002		work = kmalloc(sizeof(*work), GFP_ATOMIC);
1003		if (work) {
1004			*work = *base_work;
1005			work->nr_pages = nr_pages;
1006			work->auto_free = 1;
1007			wb_queue_work(wb, work);
1008			continue;
1009		}
1010
1011		/* alloc failed, execute synchronously using on-stack fallback */
1012		work = &fallback_work;
1013		*work = *base_work;
1014		work->nr_pages = nr_pages;
1015		work->auto_free = 0;
1016		work->done = &fallback_work_done;
1017
1018		wb_queue_work(wb, work);
1019
1020		/*
1021		 * Pin @wb so that it stays on @bdi->wb_list.  This allows
1022		 * continuing iteration from @wb after dropping and
1023		 * regrabbing rcu read lock.
1024		 */
1025		wb_get(wb);
1026		last_wb = wb;
1027
1028		rcu_read_unlock();
1029		wb_wait_for_completion(&fallback_work_done);
1030		goto restart;
1031	}
1032	rcu_read_unlock();
1033
1034	if (last_wb)
1035		wb_put(last_wb);
1036}
1037
1038/**
1039 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1040 * @bdi_id: target bdi id
1041 * @memcg_id: target memcg css id
1042 * @nr: number of pages to write, 0 for best-effort dirty flushing
1043 * @reason: reason why some writeback work initiated
1044 * @done: target wb_completion
1045 *
1046 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1047 * with the specified parameters.
1048 */
1049int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
1050			   enum wb_reason reason, struct wb_completion *done)
1051{
1052	struct backing_dev_info *bdi;
1053	struct cgroup_subsys_state *memcg_css;
1054	struct bdi_writeback *wb;
1055	struct wb_writeback_work *work;
1056	int ret;
1057
1058	/* lookup bdi and memcg */
1059	bdi = bdi_get_by_id(bdi_id);
1060	if (!bdi)
1061		return -ENOENT;
1062
1063	rcu_read_lock();
1064	memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1065	if (memcg_css && !css_tryget(memcg_css))
1066		memcg_css = NULL;
1067	rcu_read_unlock();
1068	if (!memcg_css) {
1069		ret = -ENOENT;
1070		goto out_bdi_put;
1071	}
1072
1073	/*
1074	 * And find the associated wb.  If the wb isn't there already
1075	 * there's nothing to flush, don't create one.
1076	 */
1077	wb = wb_get_lookup(bdi, memcg_css);
1078	if (!wb) {
1079		ret = -ENOENT;
1080		goto out_css_put;
1081	}
1082
1083	/*
1084	 * If @nr is zero, the caller is attempting to write out most of
1085	 * the currently dirty pages.  Let's take the current dirty page
1086	 * count and inflate it by 25% which should be large enough to
1087	 * flush out most dirty pages while avoiding getting livelocked by
1088	 * concurrent dirtiers.
1089	 */
1090	if (!nr) {
1091		unsigned long filepages, headroom, dirty, writeback;
1092
1093		mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
1094				      &writeback);
1095		nr = dirty * 10 / 8;
1096	}
1097
1098	/* issue the writeback work */
1099	work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1100	if (work) {
1101		work->nr_pages = nr;
1102		work->sync_mode = WB_SYNC_NONE;
1103		work->range_cyclic = 1;
1104		work->reason = reason;
1105		work->done = done;
1106		work->auto_free = 1;
1107		wb_queue_work(wb, work);
1108		ret = 0;
1109	} else {
1110		ret = -ENOMEM;
1111	}
1112
1113	wb_put(wb);
1114out_css_put:
1115	css_put(memcg_css);
1116out_bdi_put:
1117	bdi_put(bdi);
1118	return ret;
1119}
1120
1121/**
1122 * cgroup_writeback_umount - flush inode wb switches for umount
1123 *
1124 * This function is called when a super_block is about to be destroyed and
1125 * flushes in-flight inode wb switches.  An inode wb switch goes through
1126 * RCU and then workqueue, so the two need to be flushed in order to ensure
1127 * that all previously scheduled switches are finished.  As wb switches are
1128 * rare occurrences and synchronize_rcu() can take a while, perform
1129 * flushing iff wb switches are in flight.
1130 */
1131void cgroup_writeback_umount(void)
1132{
1133	/*
1134	 * SB_ACTIVE should be reliably cleared before checking
1135	 * isw_nr_in_flight, see generic_shutdown_super().
1136	 */
1137	smp_mb();
1138
1139	if (atomic_read(&isw_nr_in_flight)) {
1140		/*
1141		 * Use rcu_barrier() to wait for all pending callbacks to
1142		 * ensure that all in-flight wb switches are in the workqueue.
1143		 */
1144		rcu_barrier();
1145		flush_workqueue(isw_wq);
1146	}
1147}
1148
1149static int __init cgroup_writeback_init(void)
1150{
1151	isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1152	if (!isw_wq)
1153		return -ENOMEM;
1154	return 0;
1155}
1156fs_initcall(cgroup_writeback_init);
1157
1158#else	/* CONFIG_CGROUP_WRITEBACK */
1159
1160static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1161static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1162
1163static void inode_cgwb_move_to_attached(struct inode *inode,
1164					struct bdi_writeback *wb)
1165{
1166	assert_spin_locked(&wb->list_lock);
1167	assert_spin_locked(&inode->i_lock);
1168
1169	inode->i_state &= ~I_SYNC_QUEUED;
1170	list_del_init(&inode->i_io_list);
1171	wb_io_lists_depopulated(wb);
1172}
1173
1174static struct bdi_writeback *
1175locked_inode_to_wb_and_lock_list(struct inode *inode)
1176	__releases(&inode->i_lock)
1177	__acquires(&wb->list_lock)
1178{
1179	struct bdi_writeback *wb = inode_to_wb(inode);
1180
1181	spin_unlock(&inode->i_lock);
1182	spin_lock(&wb->list_lock);
1183	return wb;
1184}
1185
1186static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1187	__acquires(&wb->list_lock)
1188{
1189	struct bdi_writeback *wb = inode_to_wb(inode);
1190
1191	spin_lock(&wb->list_lock);
1192	return wb;
1193}
1194
1195static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1196{
1197	return nr_pages;
1198}
1199
1200static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1201				  struct wb_writeback_work *base_work,
1202				  bool skip_if_busy)
1203{
1204	might_sleep();
1205
1206	if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1207		base_work->auto_free = 0;
1208		wb_queue_work(&bdi->wb, base_work);
1209	}
1210}
1211
1212#endif	/* CONFIG_CGROUP_WRITEBACK */
1213
1214/*
1215 * Add in the number of potentially dirty inodes, because each inode
1216 * write can dirty pagecache in the underlying blockdev.
1217 */
1218static unsigned long get_nr_dirty_pages(void)
1219{
1220	return global_node_page_state(NR_FILE_DIRTY) +
1221		get_nr_dirty_inodes();
1222}
1223
1224static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1225{
1226	if (!wb_has_dirty_io(wb))
1227		return;
1228
1229	/*
1230	 * All callers of this function want to start writeback of all
1231	 * dirty pages. Places like vmscan can call this at a very
1232	 * high frequency, causing pointless allocations of tons of
1233	 * work items and keeping the flusher threads busy retrieving
1234	 * that work. Ensure that we only allow one of them pending and
1235	 * inflight at the time.
1236	 */
1237	if (test_bit(WB_start_all, &wb->state) ||
1238	    test_and_set_bit(WB_start_all, &wb->state))
1239		return;
1240
1241	wb->start_all_reason = reason;
1242	wb_wakeup(wb);
1243}
1244
1245/**
1246 * wb_start_background_writeback - start background writeback
1247 * @wb: bdi_writback to write from
1248 *
1249 * Description:
1250 *   This makes sure WB_SYNC_NONE background writeback happens. When
1251 *   this function returns, it is only guaranteed that for given wb
1252 *   some IO is happening if we are over background dirty threshold.
1253 *   Caller need not hold sb s_umount semaphore.
1254 */
1255void wb_start_background_writeback(struct bdi_writeback *wb)
1256{
1257	/*
1258	 * We just wake up the flusher thread. It will perform background
1259	 * writeback as soon as there is no other work to do.
1260	 */
1261	trace_writeback_wake_background(wb);
1262	wb_wakeup(wb);
1263}
1264
1265/*
1266 * Remove the inode from the writeback list it is on.
1267 */
1268void inode_io_list_del(struct inode *inode)
1269{
1270	struct bdi_writeback *wb;
1271
1272	wb = inode_to_wb_and_lock_list(inode);
1273	spin_lock(&inode->i_lock);
1274
1275	inode->i_state &= ~I_SYNC_QUEUED;
1276	list_del_init(&inode->i_io_list);
1277	wb_io_lists_depopulated(wb);
1278
1279	spin_unlock(&inode->i_lock);
1280	spin_unlock(&wb->list_lock);
1281}
1282EXPORT_SYMBOL(inode_io_list_del);
1283
1284/*
1285 * mark an inode as under writeback on the sb
1286 */
1287void sb_mark_inode_writeback(struct inode *inode)
1288{
1289	struct super_block *sb = inode->i_sb;
1290	unsigned long flags;
1291
1292	if (list_empty(&inode->i_wb_list)) {
1293		spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1294		if (list_empty(&inode->i_wb_list)) {
1295			list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1296			trace_sb_mark_inode_writeback(inode);
1297		}
1298		spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1299	}
1300}
1301
1302/*
1303 * clear an inode as under writeback on the sb
1304 */
1305void sb_clear_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_del_init(&inode->i_wb_list);
1314			trace_sb_clear_inode_writeback(inode);
1315		}
1316		spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1317	}
1318}
1319
1320/*
1321 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1322 * furthest end of its superblock's dirty-inode list.
1323 *
1324 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1325 * already the most-recently-dirtied inode on the b_dirty list.  If that is
1326 * the case then the inode must have been redirtied while it was being written
1327 * out and we don't reset its dirtied_when.
1328 */
1329static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1330{
1331	assert_spin_locked(&inode->i_lock);
1332
1333	if (!list_empty(&wb->b_dirty)) {
1334		struct inode *tail;
1335
1336		tail = wb_inode(wb->b_dirty.next);
1337		if (time_before(inode->dirtied_when, tail->dirtied_when))
1338			inode->dirtied_when = jiffies;
1339	}
1340	inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1341	inode->i_state &= ~I_SYNC_QUEUED;
1342}
1343
1344static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1345{
1346	spin_lock(&inode->i_lock);
1347	redirty_tail_locked(inode, wb);
1348	spin_unlock(&inode->i_lock);
1349}
1350
1351/*
1352 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1353 */
1354static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1355{
1356	inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1357}
1358
1359static void inode_sync_complete(struct inode *inode)
1360{
1361	inode->i_state &= ~I_SYNC;
1362	/* If inode is clean an unused, put it into LRU now... */
1363	inode_add_lru(inode);
1364	/* Waiters must see I_SYNC cleared before being woken up */
1365	smp_mb();
1366	wake_up_bit(&inode->i_state, __I_SYNC);
1367}
1368
1369static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1370{
1371	bool ret = time_after(inode->dirtied_when, t);
1372#ifndef CONFIG_64BIT
1373	/*
1374	 * For inodes being constantly redirtied, dirtied_when can get stuck.
1375	 * It _appears_ to be in the future, but is actually in distant past.
1376	 * This test is necessary to prevent such wrapped-around relative times
1377	 * from permanently stopping the whole bdi writeback.
1378	 */
1379	ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1380#endif
1381	return ret;
1382}
1383
1384#define EXPIRE_DIRTY_ATIME 0x0001
1385
1386/*
1387 * Move expired (dirtied before dirtied_before) dirty inodes from
1388 * @delaying_queue to @dispatch_queue.
1389 */
1390static int move_expired_inodes(struct list_head *delaying_queue,
1391			       struct list_head *dispatch_queue,
1392			       unsigned long dirtied_before)
1393{
1394	LIST_HEAD(tmp);
1395	struct list_head *pos, *node;
1396	struct super_block *sb = NULL;
1397	struct inode *inode;
1398	int do_sb_sort = 0;
1399	int moved = 0;
1400
1401	while (!list_empty(delaying_queue)) {
1402		inode = wb_inode(delaying_queue->prev);
1403		if (inode_dirtied_after(inode, dirtied_before))
1404			break;
1405		list_move(&inode->i_io_list, &tmp);
1406		moved++;
1407		spin_lock(&inode->i_lock);
1408		inode->i_state |= I_SYNC_QUEUED;
1409		spin_unlock(&inode->i_lock);
1410		if (sb_is_blkdev_sb(inode->i_sb))
1411			continue;
1412		if (sb && sb != inode->i_sb)
1413			do_sb_sort = 1;
1414		sb = inode->i_sb;
1415	}
1416
1417	/* just one sb in list, splice to dispatch_queue and we're done */
1418	if (!do_sb_sort) {
1419		list_splice(&tmp, dispatch_queue);
1420		goto out;
1421	}
1422
1423	/* Move inodes from one superblock together */
1424	while (!list_empty(&tmp)) {
1425		sb = wb_inode(tmp.prev)->i_sb;
1426		list_for_each_prev_safe(pos, node, &tmp) {
1427			inode = wb_inode(pos);
1428			if (inode->i_sb == sb)
1429				list_move(&inode->i_io_list, dispatch_queue);
1430		}
1431	}
1432out:
1433	return moved;
1434}
1435
1436/*
1437 * Queue all expired dirty inodes for io, eldest first.
1438 * Before
1439 *         newly dirtied     b_dirty    b_io    b_more_io
1440 *         =============>    gf         edc     BA
1441 * After
1442 *         newly dirtied     b_dirty    b_io    b_more_io
1443 *         =============>    g          fBAedc
1444 *                                           |
1445 *                                           +--> dequeue for IO
1446 */
1447static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1448		     unsigned long dirtied_before)
1449{
1450	int moved;
1451	unsigned long time_expire_jif = dirtied_before;
1452
1453	assert_spin_locked(&wb->list_lock);
1454	list_splice_init(&wb->b_more_io, &wb->b_io);
1455	moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1456	if (!work->for_sync)
1457		time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1458	moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1459				     time_expire_jif);
1460	if (moved)
1461		wb_io_lists_populated(wb);
1462	trace_writeback_queue_io(wb, work, dirtied_before, moved);
1463}
1464
1465static int write_inode(struct inode *inode, struct writeback_control *wbc)
1466{
1467	int ret;
1468
1469	if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1470		trace_writeback_write_inode_start(inode, wbc);
1471		ret = inode->i_sb->s_op->write_inode(inode, wbc);
1472		trace_writeback_write_inode(inode, wbc);
1473		return ret;
1474	}
1475	return 0;
1476}
1477
1478/*
1479 * Wait for writeback on an inode to complete. Called with i_lock held.
1480 * Caller must make sure inode cannot go away when we drop i_lock.
1481 */
1482static void __inode_wait_for_writeback(struct inode *inode)
1483	__releases(inode->i_lock)
1484	__acquires(inode->i_lock)
1485{
1486	DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1487	wait_queue_head_t *wqh;
1488
1489	wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1490	while (inode->i_state & I_SYNC) {
1491		spin_unlock(&inode->i_lock);
1492		__wait_on_bit(wqh, &wq, bit_wait,
1493			      TASK_UNINTERRUPTIBLE);
1494		spin_lock(&inode->i_lock);
1495	}
1496}
1497
1498/*
1499 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1500 */
1501void inode_wait_for_writeback(struct inode *inode)
1502{
1503	spin_lock(&inode->i_lock);
1504	__inode_wait_for_writeback(inode);
1505	spin_unlock(&inode->i_lock);
1506}
1507
1508/*
1509 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1510 * held and drops it. It is aimed for callers not holding any inode reference
1511 * so once i_lock is dropped, inode can go away.
1512 */
1513static void inode_sleep_on_writeback(struct inode *inode)
1514	__releases(inode->i_lock)
1515{
1516	DEFINE_WAIT(wait);
1517	wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1518	int sleep;
1519
1520	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1521	sleep = inode->i_state & I_SYNC;
1522	spin_unlock(&inode->i_lock);
1523	if (sleep)
1524		schedule();
1525	finish_wait(wqh, &wait);
1526}
1527
1528/*
1529 * Find proper writeback list for the inode depending on its current state and
1530 * possibly also change of its state while we were doing writeback.  Here we
1531 * handle things such as livelock prevention or fairness of writeback among
1532 * inodes. This function can be called only by flusher thread - noone else
1533 * processes all inodes in writeback lists and requeueing inodes behind flusher
1534 * thread's back can have unexpected consequences.
1535 */
1536static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1537			  struct writeback_control *wbc)
1538{
1539	if (inode->i_state & I_FREEING)
1540		return;
1541
1542	/*
1543	 * Sync livelock prevention. Each inode is tagged and synced in one
1544	 * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1545	 * the dirty time to prevent enqueue and sync it again.
1546	 */
1547	if ((inode->i_state & I_DIRTY) &&
1548	    (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1549		inode->dirtied_when = jiffies;
1550
1551	if (wbc->pages_skipped) {
1552		/*
1553		 * writeback is not making progress due to locked
1554		 * buffers. Skip this inode for now.
1555		 */
1556		redirty_tail_locked(inode, wb);
1557		return;
1558	}
1559
1560	if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1561		/*
1562		 * We didn't write back all the pages.  nfs_writepages()
1563		 * sometimes bales out without doing anything.
1564		 */
1565		if (wbc->nr_to_write <= 0) {
1566			/* Slice used up. Queue for next turn. */
1567			requeue_io(inode, wb);
1568		} else {
1569			/*
1570			 * Writeback blocked by something other than
1571			 * congestion. Delay the inode for some time to
1572			 * avoid spinning on the CPU (100% iowait)
1573			 * retrying writeback of the dirty page/inode
1574			 * that cannot be performed immediately.
1575			 */
1576			redirty_tail_locked(inode, wb);
1577		}
1578	} else if (inode->i_state & I_DIRTY) {
1579		/*
1580		 * Filesystems can dirty the inode during writeback operations,
1581		 * such as delayed allocation during submission or metadata
1582		 * updates after data IO completion.
1583		 */
1584		redirty_tail_locked(inode, wb);
1585	} else if (inode->i_state & I_DIRTY_TIME) {
1586		inode->dirtied_when = jiffies;
1587		inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1588		inode->i_state &= ~I_SYNC_QUEUED;
1589	} else {
1590		/* The inode is clean. Remove from writeback lists. */
1591		inode_cgwb_move_to_attached(inode, wb);
1592	}
1593}
1594
1595/*
1596 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1597 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1598 *
1599 * This doesn't remove the inode from the writeback list it is on, except
1600 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1601 * expiration.  The caller is otherwise responsible for writeback list handling.
1602 *
1603 * The caller is also responsible for setting the I_SYNC flag beforehand and
1604 * calling inode_sync_complete() to clear it afterwards.
1605 */
1606static int
1607__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1608{
1609	struct address_space *mapping = inode->i_mapping;
1610	long nr_to_write = wbc->nr_to_write;
1611	unsigned dirty;
1612	int ret;
1613
1614	WARN_ON(!(inode->i_state & I_SYNC));
1615
1616	trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1617
1618	ret = do_writepages(mapping, wbc);
1619
1620	/*
1621	 * Make sure to wait on the data before writing out the metadata.
1622	 * This is important for filesystems that modify metadata on data
1623	 * I/O completion. We don't do it for sync(2) writeback because it has a
1624	 * separate, external IO completion path and ->sync_fs for guaranteeing
1625	 * inode metadata is written back correctly.
1626	 */
1627	if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1628		int err = filemap_fdatawait(mapping);
1629		if (ret == 0)
1630			ret = err;
1631	}
1632
1633	/*
1634	 * If the inode has dirty timestamps and we need to write them, call
1635	 * mark_inode_dirty_sync() to notify the filesystem about it and to
1636	 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1637	 */
1638	if ((inode->i_state & I_DIRTY_TIME) &&
1639	    (wbc->sync_mode == WB_SYNC_ALL ||
1640	     time_after(jiffies, inode->dirtied_time_when +
1641			dirtytime_expire_interval * HZ))) {
1642		trace_writeback_lazytime(inode);
1643		mark_inode_dirty_sync(inode);
1644	}
1645
1646	/*
1647	 * Get and clear the dirty flags from i_state.  This needs to be done
1648	 * after calling writepages because some filesystems may redirty the
1649	 * inode during writepages due to delalloc.  It also needs to be done
1650	 * after handling timestamp expiration, as that may dirty the inode too.
1651	 */
1652	spin_lock(&inode->i_lock);
1653	dirty = inode->i_state & I_DIRTY;
1654	inode->i_state &= ~dirty;
1655
1656	/*
1657	 * Paired with smp_mb() in __mark_inode_dirty().  This allows
1658	 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1659	 * either they see the I_DIRTY bits cleared or we see the dirtied
1660	 * inode.
1661	 *
1662	 * I_DIRTY_PAGES is always cleared together above even if @mapping
1663	 * still has dirty pages.  The flag is reinstated after smp_mb() if
1664	 * necessary.  This guarantees that either __mark_inode_dirty()
1665	 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1666	 */
1667	smp_mb();
1668
1669	if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1670		inode->i_state |= I_DIRTY_PAGES;
1671
1672	spin_unlock(&inode->i_lock);
1673
1674	/* Don't write the inode if only I_DIRTY_PAGES was set */
1675	if (dirty & ~I_DIRTY_PAGES) {
1676		int err = write_inode(inode, wbc);
1677		if (ret == 0)
1678			ret = err;
1679	}
1680	trace_writeback_single_inode(inode, wbc, nr_to_write);
1681	return ret;
1682}
1683
1684/*
1685 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1686 * the regular batched writeback done by the flusher threads in
1687 * writeback_sb_inodes().  @wbc controls various aspects of the write, such as
1688 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1689 *
1690 * To prevent the inode from going away, either the caller must have a reference
1691 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1692 */
1693static int writeback_single_inode(struct inode *inode,
1694				  struct writeback_control *wbc)
1695{
1696	struct bdi_writeback *wb;
1697	int ret = 0;
1698
1699	spin_lock(&inode->i_lock);
1700	if (!atomic_read(&inode->i_count))
1701		WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1702	else
1703		WARN_ON(inode->i_state & I_WILL_FREE);
1704
1705	if (inode->i_state & I_SYNC) {
1706		/*
1707		 * Writeback is already running on the inode.  For WB_SYNC_NONE,
1708		 * that's enough and we can just return.  For WB_SYNC_ALL, we
1709		 * must wait for the existing writeback to complete, then do
1710		 * writeback again if there's anything left.
1711		 */
1712		if (wbc->sync_mode != WB_SYNC_ALL)
1713			goto out;
1714		__inode_wait_for_writeback(inode);
1715	}
1716	WARN_ON(inode->i_state & I_SYNC);
1717	/*
1718	 * If the inode is already fully clean, then there's nothing to do.
1719	 *
1720	 * For data-integrity syncs we also need to check whether any pages are
1721	 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback.  If
1722	 * there are any such pages, we'll need to wait for them.
1723	 */
1724	if (!(inode->i_state & I_DIRTY_ALL) &&
1725	    (wbc->sync_mode != WB_SYNC_ALL ||
1726	     !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1727		goto out;
1728	inode->i_state |= I_SYNC;
1729	wbc_attach_and_unlock_inode(wbc, inode);
1730
1731	ret = __writeback_single_inode(inode, wbc);
1732
1733	wbc_detach_inode(wbc);
1734
1735	wb = inode_to_wb_and_lock_list(inode);
1736	spin_lock(&inode->i_lock);
1737	/*
1738	 * If the inode is now fully clean, then it can be safely removed from
1739	 * its writeback list (if any).  Otherwise the flusher threads are
1740	 * responsible for the writeback lists.
1741	 */
1742	if (!(inode->i_state & I_DIRTY_ALL))
1743		inode_cgwb_move_to_attached(inode, wb);
1744	spin_unlock(&wb->list_lock);
1745	inode_sync_complete(inode);
1746out:
1747	spin_unlock(&inode->i_lock);
1748	return ret;
1749}
1750
1751static long writeback_chunk_size(struct bdi_writeback *wb,
1752				 struct wb_writeback_work *work)
1753{
1754	long pages;
1755
1756	/*
1757	 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1758	 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1759	 * here avoids calling into writeback_inodes_wb() more than once.
1760	 *
1761	 * The intended call sequence for WB_SYNC_ALL writeback is:
1762	 *
1763	 *      wb_writeback()
1764	 *          writeback_sb_inodes()       <== called only once
1765	 *              write_cache_pages()     <== called once for each inode
1766	 *                   (quickly) tag currently dirty pages
1767	 *                   (maybe slowly) sync all tagged pages
1768	 */
1769	if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1770		pages = LONG_MAX;
1771	else {
1772		pages = min(wb->avg_write_bandwidth / 2,
1773			    global_wb_domain.dirty_limit / DIRTY_SCOPE);
1774		pages = min(pages, work->nr_pages);
1775		pages = round_down(pages + MIN_WRITEBACK_PAGES,
1776				   MIN_WRITEBACK_PAGES);
1777	}
1778
1779	return pages;
1780}
1781
1782/*
1783 * Write a portion of b_io inodes which belong to @sb.
1784 *
1785 * Return the number of pages and/or inodes written.
1786 *
1787 * NOTE! This is called with wb->list_lock held, and will
1788 * unlock and relock that for each inode it ends up doing
1789 * IO for.
1790 */
1791static long writeback_sb_inodes(struct super_block *sb,
1792				struct bdi_writeback *wb,
1793				struct wb_writeback_work *work)
1794{
1795	struct writeback_control wbc = {
1796		.sync_mode		= work->sync_mode,
1797		.tagged_writepages	= work->tagged_writepages,
1798		.for_kupdate		= work->for_kupdate,
1799		.for_background		= work->for_background,
1800		.for_sync		= work->for_sync,
1801		.range_cyclic		= work->range_cyclic,
1802		.range_start		= 0,
1803		.range_end		= LLONG_MAX,
1804	};
1805	unsigned long start_time = jiffies;
1806	long write_chunk;
1807	long wrote = 0;  /* count both pages and inodes */
1808
1809	while (!list_empty(&wb->b_io)) {
1810		struct inode *inode = wb_inode(wb->b_io.prev);
1811		struct bdi_writeback *tmp_wb;
1812
1813		if (inode->i_sb != sb) {
1814			if (work->sb) {
1815				/*
1816				 * We only want to write back data for this
1817				 * superblock, move all inodes not belonging
1818				 * to it back onto the dirty list.
1819				 */
1820				redirty_tail(inode, wb);
1821				continue;
1822			}
1823
1824			/*
1825			 * The inode belongs to a different superblock.
1826			 * Bounce back to the caller to unpin this and
1827			 * pin the next superblock.
1828			 */
1829			break;
1830		}
1831
1832		/*
1833		 * Don't bother with new inodes or inodes being freed, first
1834		 * kind does not need periodic writeout yet, and for the latter
1835		 * kind writeout is handled by the freer.
1836		 */
1837		spin_lock(&inode->i_lock);
1838		if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1839			redirty_tail_locked(inode, wb);
1840			spin_unlock(&inode->i_lock);
1841			continue;
1842		}
1843		if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1844			/*
1845			 * If this inode is locked for writeback and we are not
1846			 * doing writeback-for-data-integrity, move it to
1847			 * b_more_io so that writeback can proceed with the
1848			 * other inodes on s_io.
1849			 *
1850			 * We'll have another go at writing back this inode
1851			 * when we completed a full scan of b_io.
1852			 */
1853			spin_unlock(&inode->i_lock);
1854			requeue_io(inode, wb);
1855			trace_writeback_sb_inodes_requeue(inode);
1856			continue;
1857		}
1858		spin_unlock(&wb->list_lock);
1859
1860		/*
1861		 * We already requeued the inode if it had I_SYNC set and we
1862		 * are doing WB_SYNC_NONE writeback. So this catches only the
1863		 * WB_SYNC_ALL case.
1864		 */
1865		if (inode->i_state & I_SYNC) {
1866			/* Wait for I_SYNC. This function drops i_lock... */
1867			inode_sleep_on_writeback(inode);
1868			/* Inode may be gone, start again */
1869			spin_lock(&wb->list_lock);
1870			continue;
1871		}
1872		inode->i_state |= I_SYNC;
1873		wbc_attach_and_unlock_inode(&wbc, inode);
1874
1875		write_chunk = writeback_chunk_size(wb, work);
1876		wbc.nr_to_write = write_chunk;
1877		wbc.pages_skipped = 0;
1878
1879		/*
1880		 * We use I_SYNC to pin the inode in memory. While it is set
1881		 * evict_inode() will wait so the inode cannot be freed.
1882		 */
1883		__writeback_single_inode(inode, &wbc);
1884
1885		wbc_detach_inode(&wbc);
1886		work->nr_pages -= write_chunk - wbc.nr_to_write;
1887		wrote += write_chunk - wbc.nr_to_write;
1888
1889		if (need_resched()) {
1890			/*
1891			 * We're trying to balance between building up a nice
1892			 * long list of IOs to improve our merge rate, and
1893			 * getting those IOs out quickly for anyone throttling
1894			 * in balance_dirty_pages().  cond_resched() doesn't
1895			 * unplug, so get our IOs out the door before we
1896			 * give up the CPU.
1897			 */
1898			blk_flush_plug(current);
1899			cond_resched();
1900		}
1901
1902		/*
1903		 * Requeue @inode if still dirty.  Be careful as @inode may
1904		 * have been switched to another wb in the meantime.
1905		 */
1906		tmp_wb = inode_to_wb_and_lock_list(inode);
1907		spin_lock(&inode->i_lock);
1908		if (!(inode->i_state & I_DIRTY_ALL))
1909			wrote++;
1910		requeue_inode(inode, tmp_wb, &wbc);
1911		inode_sync_complete(inode);
1912		spin_unlock(&inode->i_lock);
1913
1914		if (unlikely(tmp_wb != wb)) {
1915			spin_unlock(&tmp_wb->list_lock);
1916			spin_lock(&wb->list_lock);
1917		}
1918
1919		/*
1920		 * bail out to wb_writeback() often enough to check
1921		 * background threshold and other termination conditions.
1922		 */
1923		if (wrote) {
1924			if (time_is_before_jiffies(start_time + HZ / 10UL))
1925				break;
1926			if (work->nr_pages <= 0)
1927				break;
1928		}
1929	}
1930	return wrote;
1931}
1932
1933static long __writeback_inodes_wb(struct bdi_writeback *wb,
1934				  struct wb_writeback_work *work)
1935{
1936	unsigned long start_time = jiffies;
1937	long wrote = 0;
1938
1939	while (!list_empty(&wb->b_io)) {
1940		struct inode *inode = wb_inode(wb->b_io.prev);
1941		struct super_block *sb = inode->i_sb;
1942
1943		if (!trylock_super(sb)) {
1944			/*
1945			 * trylock_super() may fail consistently due to
1946			 * s_umount being grabbed by someone else. Don't use
1947			 * requeue_io() to avoid busy retrying the inode/sb.
1948			 */
1949			redirty_tail(inode, wb);
1950			continue;
1951		}
1952		wrote += writeback_sb_inodes(sb, wb, work);
1953		up_read(&sb->s_umount);
1954
1955		/* refer to the same tests at the end of writeback_sb_inodes */
1956		if (wrote) {
1957			if (time_is_before_jiffies(start_time + HZ / 10UL))
1958				break;
1959			if (work->nr_pages <= 0)
1960				break;
1961		}
1962	}
1963	/* Leave any unwritten inodes on b_io */
1964	return wrote;
1965}
1966
1967static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1968				enum wb_reason reason)
1969{
1970	struct wb_writeback_work work = {
1971		.nr_pages	= nr_pages,
1972		.sync_mode	= WB_SYNC_NONE,
1973		.range_cyclic	= 1,
1974		.reason		= reason,
1975	};
1976	struct blk_plug plug;
1977
1978	blk_start_plug(&plug);
1979	spin_lock(&wb->list_lock);
1980	if (list_empty(&wb->b_io))
1981		queue_io(wb, &work, jiffies);
1982	__writeback_inodes_wb(wb, &work);
1983	spin_unlock(&wb->list_lock);
1984	blk_finish_plug(&plug);
1985
1986	return nr_pages - work.nr_pages;
1987}
1988
1989/*
1990 * Explicit flushing or periodic writeback of "old" data.
1991 *
1992 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1993 * dirtying-time in the inode's address_space.  So this periodic writeback code
1994 * just walks the superblock inode list, writing back any inodes which are
1995 * older than a specific point in time.
1996 *
1997 * Try to run once per dirty_writeback_interval.  But if a writeback event
1998 * takes longer than a dirty_writeback_interval interval, then leave a
1999 * one-second gap.
2000 *
2001 * dirtied_before takes precedence over nr_to_write.  So we'll only write back
2002 * all dirty pages if they are all attached to "old" mappings.
2003 */
2004static long wb_writeback(struct bdi_writeback *wb,
2005			 struct wb_writeback_work *work)
2006{
2007	unsigned long wb_start = jiffies;
2008	long nr_pages = work->nr_pages;
2009	unsigned long dirtied_before = jiffies;
2010	struct inode *inode;
2011	long progress;
2012	struct blk_plug plug;
2013
2014	blk_start_plug(&plug);
2015	spin_lock(&wb->list_lock);
2016	for (;;) {
2017		/*
2018		 * Stop writeback when nr_pages has been consumed
2019		 */
2020		if (work->nr_pages <= 0)
2021			break;
2022
2023		/*
2024		 * Background writeout and kupdate-style writeback may
2025		 * run forever. Stop them if there is other work to do
2026		 * so that e.g. sync can proceed. They'll be restarted
2027		 * after the other works are all done.
2028		 */
2029		if ((work->for_background || work->for_kupdate) &&
2030		    !list_empty(&wb->work_list))
2031			break;
2032
2033		/*
2034		 * For background writeout, stop when we are below the
2035		 * background dirty threshold
2036		 */
2037		if (work->for_background && !wb_over_bg_thresh(wb))
2038			break;
2039
2040		/*
2041		 * Kupdate and background works are special and we want to
2042		 * include all inodes that need writing. Livelock avoidance is
2043		 * handled by these works yielding to any other work so we are
2044		 * safe.
2045		 */
2046		if (work->for_kupdate) {
2047			dirtied_before = jiffies -
2048				msecs_to_jiffies(dirty_expire_interval * 10);
2049		} else if (work->for_background)
2050			dirtied_before = jiffies;
2051
2052		trace_writeback_start(wb, work);
2053		if (list_empty(&wb->b_io))
2054			queue_io(wb, work, dirtied_before);
2055		if (work->sb)
2056			progress = writeback_sb_inodes(work->sb, wb, work);
2057		else
2058			progress = __writeback_inodes_wb(wb, work);
2059		trace_writeback_written(wb, work);
2060
2061		wb_update_bandwidth(wb, wb_start);
2062
2063		/*
2064		 * Did we write something? Try for more
2065		 *
2066		 * Dirty inodes are moved to b_io for writeback in batches.
2067		 * The completion of the current batch does not necessarily
2068		 * mean the overall work is done. So we keep looping as long
2069		 * as made some progress on cleaning pages or inodes.
2070		 */
2071		if (progress)
2072			continue;
2073		/*
2074		 * No more inodes for IO, bail
2075		 */
2076		if (list_empty(&wb->b_more_io))
2077			break;
2078		/*
2079		 * Nothing written. Wait for some inode to
2080		 * become available for writeback. Otherwise
2081		 * we'll just busyloop.
2082		 */
2083		trace_writeback_wait(wb, work);
2084		inode = wb_inode(wb->b_more_io.prev);
2085		spin_lock(&inode->i_lock);
2086		spin_unlock(&wb->list_lock);
2087		/* This function drops i_lock... */
2088		inode_sleep_on_writeback(inode);
2089		spin_lock(&wb->list_lock);
2090	}
2091	spin_unlock(&wb->list_lock);
2092	blk_finish_plug(&plug);
2093
2094	return nr_pages - work->nr_pages;
2095}
2096
2097/*
2098 * Return the next wb_writeback_work struct that hasn't been processed yet.
2099 */
2100static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2101{
2102	struct wb_writeback_work *work = NULL;
2103
2104	spin_lock_bh(&wb->work_lock);
2105	if (!list_empty(&wb->work_list)) {
2106		work = list_entry(wb->work_list.next,
2107				  struct wb_writeback_work, list);
2108		list_del_init(&work->list);
2109	}
2110	spin_unlock_bh(&wb->work_lock);
2111	return work;
2112}
2113
2114static long wb_check_background_flush(struct bdi_writeback *wb)
2115{
2116	if (wb_over_bg_thresh(wb)) {
2117
2118		struct wb_writeback_work work = {
2119			.nr_pages	= LONG_MAX,
2120			.sync_mode	= WB_SYNC_NONE,
2121			.for_background	= 1,
2122			.range_cyclic	= 1,
2123			.reason		= WB_REASON_BACKGROUND,
2124		};
2125
2126		return wb_writeback(wb, &work);
2127	}
2128
2129	return 0;
2130}
2131
2132static long wb_check_old_data_flush(struct bdi_writeback *wb)
2133{
2134	unsigned long expired;
2135	long nr_pages;
2136
2137	/*
2138	 * When set to zero, disable periodic writeback
2139	 */
2140	if (!dirty_writeback_interval)
2141		return 0;
2142
2143	expired = wb->last_old_flush +
2144			msecs_to_jiffies(dirty_writeback_interval * 10);
2145	if (time_before(jiffies, expired))
2146		return 0;
2147
2148	wb->last_old_flush = jiffies;
2149	nr_pages = get_nr_dirty_pages();
2150
2151	if (nr_pages) {
2152		struct wb_writeback_work work = {
2153			.nr_pages	= nr_pages,
2154			.sync_mode	= WB_SYNC_NONE,
2155			.for_kupdate	= 1,
2156			.range_cyclic	= 1,
2157			.reason		= WB_REASON_PERIODIC,
2158		};
2159
2160		return wb_writeback(wb, &work);
2161	}
2162
2163	return 0;
2164}
2165
2166static long wb_check_start_all(struct bdi_writeback *wb)
2167{
2168	long nr_pages;
2169
2170	if (!test_bit(WB_start_all, &wb->state))
2171		return 0;
2172
2173	nr_pages = get_nr_dirty_pages();
2174	if (nr_pages) {
2175		struct wb_writeback_work work = {
2176			.nr_pages	= wb_split_bdi_pages(wb, nr_pages),
2177			.sync_mode	= WB_SYNC_NONE,
2178			.range_cyclic	= 1,
2179			.reason		= wb->start_all_reason,
2180		};
2181
2182		nr_pages = wb_writeback(wb, &work);
2183	}
2184
2185	clear_bit(WB_start_all, &wb->state);
2186	return nr_pages;
2187}
2188
2189
2190/*
2191 * Retrieve work items and do the writeback they describe
2192 */
2193static long wb_do_writeback(struct bdi_writeback *wb)
2194{
2195	struct wb_writeback_work *work;
2196	long wrote = 0;
2197
2198	set_bit(WB_writeback_running, &wb->state);
2199	while ((work = get_next_work_item(wb)) != NULL) {
2200		trace_writeback_exec(wb, work);
2201		wrote += wb_writeback(wb, work);
2202		finish_writeback_work(wb, work);
2203	}
2204
2205	/*
2206	 * Check for a flush-everything request
2207	 */
2208	wrote += wb_check_start_all(wb);
2209
2210	/*
2211	 * Check for periodic writeback, kupdated() style
2212	 */
2213	wrote += wb_check_old_data_flush(wb);
2214	wrote += wb_check_background_flush(wb);
2215	clear_bit(WB_writeback_running, &wb->state);
2216
2217	return wrote;
2218}
2219
2220/*
2221 * Handle writeback of dirty data for the device backed by this bdi. Also
2222 * reschedules periodically and does kupdated style flushing.
2223 */
2224void wb_workfn(struct work_struct *work)
2225{
2226	struct bdi_writeback *wb = container_of(to_delayed_work(work),
2227						struct bdi_writeback, dwork);
2228	long pages_written;
2229
2230	set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2231	current->flags |= PF_SWAPWRITE;
2232
2233	if (likely(!current_is_workqueue_rescuer() ||
2234		   !test_bit(WB_registered, &wb->state))) {
2235		/*
2236		 * The normal path.  Keep writing back @wb until its
2237		 * work_list is empty.  Note that this path is also taken
2238		 * if @wb is shutting down even when we're running off the
2239		 * rescuer as work_list needs to be drained.
2240		 */
2241		do {
2242			pages_written = wb_do_writeback(wb);
2243			trace_writeback_pages_written(pages_written);
2244		} while (!list_empty(&wb->work_list));
2245	} else {
2246		/*
2247		 * bdi_wq can't get enough workers and we're running off
2248		 * the emergency worker.  Don't hog it.  Hopefully, 1024 is
2249		 * enough for efficient IO.
2250		 */
2251		pages_written = writeback_inodes_wb(wb, 1024,
2252						    WB_REASON_FORKER_THREAD);
2253		trace_writeback_pages_written(pages_written);
2254	}
2255
2256	if (!list_empty(&wb->work_list))
2257		wb_wakeup(wb);
2258	else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2259		wb_wakeup_delayed(wb);
2260
2261	current->flags &= ~PF_SWAPWRITE;
2262}
2263
2264/*
2265 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2266 * write back the whole world.
2267 */
2268static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2269					 enum wb_reason reason)
2270{
2271	struct bdi_writeback *wb;
2272
2273	if (!bdi_has_dirty_io(bdi))
2274		return;
2275
2276	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2277		wb_start_writeback(wb, reason);
2278}
2279
2280void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2281				enum wb_reason reason)
2282{
2283	rcu_read_lock();
2284	__wakeup_flusher_threads_bdi(bdi, reason);
2285	rcu_read_unlock();
2286}
2287
2288/*
2289 * Wakeup the flusher threads to start writeback of all currently dirty pages
2290 */
2291void wakeup_flusher_threads(enum wb_reason reason)
2292{
2293	struct backing_dev_info *bdi;
2294
2295	/*
2296	 * If we are expecting writeback progress we must submit plugged IO.
2297	 */
2298	if (blk_needs_flush_plug(current))
2299		blk_schedule_flush_plug(current);
2300
2301	rcu_read_lock();
2302	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2303		__wakeup_flusher_threads_bdi(bdi, reason);
2304	rcu_read_unlock();
2305}
2306
2307/*
2308 * Wake up bdi's periodically to make sure dirtytime inodes gets
2309 * written back periodically.  We deliberately do *not* check the
2310 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2311 * kernel to be constantly waking up once there are any dirtytime
2312 * inodes on the system.  So instead we define a separate delayed work
2313 * function which gets called much more rarely.  (By default, only
2314 * once every 12 hours.)
2315 *
2316 * If there is any other write activity going on in the file system,
2317 * this function won't be necessary.  But if the only thing that has
2318 * happened on the file system is a dirtytime inode caused by an atime
2319 * update, we need this infrastructure below to make sure that inode
2320 * eventually gets pushed out to disk.
2321 */
2322static void wakeup_dirtytime_writeback(struct work_struct *w);
2323static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2324
2325static void wakeup_dirtytime_writeback(struct work_struct *w)
2326{
2327	struct backing_dev_info *bdi;
2328
2329	rcu_read_lock();
2330	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2331		struct bdi_writeback *wb;
2332
2333		list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2334			if (!list_empty(&wb->b_dirty_time))
2335				wb_wakeup(wb);
2336	}
2337	rcu_read_unlock();
2338	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2339}
2340
2341static int __init start_dirtytime_writeback(void)
2342{
2343	schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2344	return 0;
2345}
2346__initcall(start_dirtytime_writeback);
2347
2348int dirtytime_interval_handler(struct ctl_table *table, int write,
2349			       void *buffer, size_t *lenp, loff_t *ppos)
2350{
2351	int ret;
2352
2353	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2354	if (ret == 0 && write)
2355		mod_delayed_work(system_wq, &dirtytime_work, 0);
2356	return ret;
2357}
2358
2359/**
2360 * __mark_inode_dirty -	internal function to mark an inode dirty
2361 *
2362 * @inode: inode to mark
2363 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC.  This can be a combination of
2364 *	   multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2365 *	   with I_DIRTY_PAGES.
2366 *
2367 * Mark an inode as dirty.  We notify the filesystem, then update the inode's
2368 * dirty flags.  Then, if needed we add the inode to the appropriate dirty list.
2369 *
2370 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2371 * instead of calling this directly.
2372 *
2373 * CAREFUL!  We only add the inode to the dirty list if it is hashed or if it
2374 * refers to a blockdev.  Unhashed inodes will never be added to the dirty list
2375 * even if they are later hashed, as they will have been marked dirty already.
2376 *
2377 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2378 *
2379 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2380 * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2381 * the kernel-internal blockdev inode represents the dirtying time of the
2382 * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2383 * page->mapping->host, so the page-dirtying time is recorded in the internal
2384 * blockdev inode.
2385 */
2386void __mark_inode_dirty(struct inode *inode, int flags)
2387{
2388	struct super_block *sb = inode->i_sb;
2389	int dirtytime = 0;
2390
2391	trace_writeback_mark_inode_dirty(inode, flags);
2392
2393	if (flags & I_DIRTY_INODE) {
2394		/*
2395		 * Notify the filesystem about the inode being dirtied, so that
2396		 * (if needed) it can update on-disk fields and journal the
2397		 * inode.  This is only needed when the inode itself is being
2398		 * dirtied now.  I.e. it's only needed for I_DIRTY_INODE, not
2399		 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2400		 */
2401		trace_writeback_dirty_inode_start(inode, flags);
2402		if (sb->s_op->dirty_inode)
2403			sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
2404		trace_writeback_dirty_inode(inode, flags);
2405
2406		/* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2407		flags &= ~I_DIRTY_TIME;
2408	} else {
2409		/*
2410		 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2411		 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2412		 * in one call to __mark_inode_dirty().)
2413		 */
2414		dirtytime = flags & I_DIRTY_TIME;
2415		WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2416	}
2417
2418	/*
2419	 * Paired with smp_mb() in __writeback_single_inode() for the
2420	 * following lockless i_state test.  See there for details.
2421	 */
2422	smp_mb();
2423
2424	if (((inode->i_state & flags) == flags) ||
2425	    (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2426		return;
2427
2428	spin_lock(&inode->i_lock);
2429	if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2430		goto out_unlock_inode;
2431	if ((inode->i_state & flags) != flags) {
2432		const int was_dirty = inode->i_state & I_DIRTY;
2433
2434		inode_attach_wb(inode, NULL);
2435
2436		/* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2437		if (flags & I_DIRTY_INODE)
2438			inode->i_state &= ~I_DIRTY_TIME;
2439		inode->i_state |= flags;
2440
2441		/*
2442		 * If the inode is queued for writeback by flush worker, just
2443		 * update its dirty state. Once the flush worker is done with
2444		 * the inode it will place it on the appropriate superblock
2445		 * list, based upon its state.
2446		 */
2447		if (inode->i_state & I_SYNC_QUEUED)
2448			goto out_unlock_inode;
2449
2450		/*
2451		 * Only add valid (hashed) inodes to the superblock's
2452		 * dirty list.  Add blockdev inodes as well.
2453		 */
2454		if (!S_ISBLK(inode->i_mode)) {
2455			if (inode_unhashed(inode))
2456				goto out_unlock_inode;
2457		}
2458		if (inode->i_state & I_FREEING)
2459			goto out_unlock_inode;
2460
2461		/*
2462		 * If the inode was already on b_dirty/b_io/b_more_io, don't
2463		 * reposition it (that would break b_dirty time-ordering).
2464		 */
2465		if (!was_dirty) {
2466			struct bdi_writeback *wb;
2467			struct list_head *dirty_list;
2468			bool wakeup_bdi = false;
2469
2470			wb = locked_inode_to_wb_and_lock_list(inode);
2471
2472			inode->dirtied_when = jiffies;
2473			if (dirtytime)
2474				inode->dirtied_time_when = jiffies;
2475
2476			if (inode->i_state & I_DIRTY)
2477				dirty_list = &wb->b_dirty;
2478			else
2479				dirty_list = &wb->b_dirty_time;
2480
2481			wakeup_bdi = inode_io_list_move_locked(inode, wb,
2482							       dirty_list);
2483
2484			spin_unlock(&wb->list_lock);
2485			trace_writeback_dirty_inode_enqueue(inode);
2486
2487			/*
2488			 * If this is the first dirty inode for this bdi,
2489			 * we have to wake-up the corresponding bdi thread
2490			 * to make sure background write-back happens
2491			 * later.
2492			 */
2493			if (wakeup_bdi &&
2494			    (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2495				wb_wakeup_delayed(wb);
2496			return;
2497		}
2498	}
2499out_unlock_inode:
2500	spin_unlock(&inode->i_lock);
2501}
2502EXPORT_SYMBOL(__mark_inode_dirty);
2503
2504/*
2505 * The @s_sync_lock is used to serialise concurrent sync operations
2506 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2507 * Concurrent callers will block on the s_sync_lock rather than doing contending
2508 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2509 * has been issued up to the time this function is enter is guaranteed to be
2510 * completed by the time we have gained the lock and waited for all IO that is
2511 * in progress regardless of the order callers are granted the lock.
2512 */
2513static void wait_sb_inodes(struct super_block *sb)
2514{
2515	LIST_HEAD(sync_list);
2516
2517	/*
2518	 * We need to be protected against the filesystem going from
2519	 * r/o to r/w or vice versa.
2520	 */
2521	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2522
2523	mutex_lock(&sb->s_sync_lock);
2524
2525	/*
2526	 * Splice the writeback list onto a temporary list to avoid waiting on
2527	 * inodes that have started writeback after this point.
2528	 *
2529	 * Use rcu_read_lock() to keep the inodes around until we have a
2530	 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2531	 * the local list because inodes can be dropped from either by writeback
2532	 * completion.
2533	 */
2534	rcu_read_lock();
2535	spin_lock_irq(&sb->s_inode_wblist_lock);
2536	list_splice_init(&sb->s_inodes_wb, &sync_list);
2537
2538	/*
2539	 * Data integrity sync. Must wait for all pages under writeback, because
2540	 * there may have been pages dirtied before our sync call, but which had
2541	 * writeout started before we write it out.  In which case, the inode
2542	 * may not be on the dirty list, but we still have to wait for that
2543	 * writeout.
2544	 */
2545	while (!list_empty(&sync_list)) {
2546		struct inode *inode = list_first_entry(&sync_list, struct inode,
2547						       i_wb_list);
2548		struct address_space *mapping = inode->i_mapping;
2549
2550		/*
2551		 * Move each inode back to the wb list before we drop the lock
2552		 * to preserve consistency between i_wb_list and the mapping
2553		 * writeback tag. Writeback completion is responsible to remove
2554		 * the inode from either list once the writeback tag is cleared.
2555		 */
2556		list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2557
2558		/*
2559		 * The mapping can appear untagged while still on-list since we
2560		 * do not have the mapping lock. Skip it here, wb completion
2561		 * will remove it.
2562		 */
2563		if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2564			continue;
2565
2566		spin_unlock_irq(&sb->s_inode_wblist_lock);
2567
2568		spin_lock(&inode->i_lock);
2569		if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2570			spin_unlock(&inode->i_lock);
2571
2572			spin_lock_irq(&sb->s_inode_wblist_lock);
2573			continue;
2574		}
2575		__iget(inode);
2576		spin_unlock(&inode->i_lock);
2577		rcu_read_unlock();
2578
2579		/*
2580		 * We keep the error status of individual mapping so that
2581		 * applications can catch the writeback error using fsync(2).
2582		 * See filemap_fdatawait_keep_errors() for details.
2583		 */
2584		filemap_fdatawait_keep_errors(mapping);
2585
2586		cond_resched();
2587
2588		iput(inode);
2589
2590		rcu_read_lock();
2591		spin_lock_irq(&sb->s_inode_wblist_lock);
2592	}
2593	spin_unlock_irq(&sb->s_inode_wblist_lock);
2594	rcu_read_unlock();
2595	mutex_unlock(&sb->s_sync_lock);
2596}
2597
2598static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2599				     enum wb_reason reason, bool skip_if_busy)
2600{
2601	struct backing_dev_info *bdi = sb->s_bdi;
2602	DEFINE_WB_COMPLETION(done, bdi);
2603	struct wb_writeback_work work = {
2604		.sb			= sb,
2605		.sync_mode		= WB_SYNC_NONE,
2606		.tagged_writepages	= 1,
2607		.done			= &done,
2608		.nr_pages		= nr,
2609		.reason			= reason,
2610	};
2611
2612	if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2613		return;
2614	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2615
2616	bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2617	wb_wait_for_completion(&done);
2618}
2619
2620/**
2621 * writeback_inodes_sb_nr -	writeback dirty inodes from given super_block
2622 * @sb: the superblock
2623 * @nr: the number of pages to write
2624 * @reason: reason why some writeback work initiated
2625 *
2626 * Start writeback on some inodes on this super_block. No guarantees are made
2627 * on how many (if any) will be written, and this function does not wait
2628 * for IO completion of submitted IO.
2629 */
2630void writeback_inodes_sb_nr(struct super_block *sb,
2631			    unsigned long nr,
2632			    enum wb_reason reason)
2633{
2634	__writeback_inodes_sb_nr(sb, nr, reason, false);
2635}
2636EXPORT_SYMBOL(writeback_inodes_sb_nr);
2637
2638/**
2639 * writeback_inodes_sb	-	writeback dirty inodes from given super_block
2640 * @sb: the superblock
2641 * @reason: reason why some writeback work was initiated
2642 *
2643 * Start writeback on some inodes on this super_block. No guarantees are made
2644 * on how many (if any) will be written, and this function does not wait
2645 * for IO completion of submitted IO.
2646 */
2647void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2648{
2649	return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2650}
2651EXPORT_SYMBOL(writeback_inodes_sb);
2652
2653/**
2654 * try_to_writeback_inodes_sb - try to start writeback if none underway
2655 * @sb: the superblock
2656 * @reason: reason why some writeback work was initiated
2657 *
2658 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2659 */
2660void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2661{
2662	if (!down_read_trylock(&sb->s_umount))
2663		return;
2664
2665	__writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2666	up_read(&sb->s_umount);
2667}
2668EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2669
2670/**
2671 * sync_inodes_sb	-	sync sb inode pages
2672 * @sb: the superblock
2673 *
2674 * This function writes and waits on any dirty inode belonging to this
2675 * super_block.
2676 */
2677void sync_inodes_sb(struct super_block *sb)
2678{
2679	struct backing_dev_info *bdi = sb->s_bdi;
2680	DEFINE_WB_COMPLETION(done, bdi);
2681	struct wb_writeback_work work = {
2682		.sb		= sb,
2683		.sync_mode	= WB_SYNC_ALL,
2684		.nr_pages	= LONG_MAX,
2685		.range_cyclic	= 0,
2686		.done		= &done,
2687		.reason		= WB_REASON_SYNC,
2688		.for_sync	= 1,
2689	};
2690
2691	/*
2692	 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2693	 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2694	 * bdi_has_dirty() need to be written out too.
2695	 */
2696	if (bdi == &noop_backing_dev_info)
2697		return;
2698	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2699
2700	/* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2701	bdi_down_write_wb_switch_rwsem(bdi);
2702	bdi_split_work_to_wbs(bdi, &work, false);
2703	wb_wait_for_completion(&done);
2704	bdi_up_write_wb_switch_rwsem(bdi);
2705
2706	wait_sb_inodes(sb);
2707}
2708EXPORT_SYMBOL(sync_inodes_sb);
2709
2710/**
2711 * write_inode_now	-	write an inode to disk
2712 * @inode: inode to write to disk
2713 * @sync: whether the write should be synchronous or not
2714 *
2715 * This function commits an inode to disk immediately if it is dirty. This is
2716 * primarily needed by knfsd.
2717 *
2718 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2719 */
2720int write_inode_now(struct inode *inode, int sync)
2721{
2722	struct writeback_control wbc = {
2723		.nr_to_write = LONG_MAX,
2724		.sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2725		.range_start = 0,
2726		.range_end = LLONG_MAX,
2727	};
2728
2729	if (!mapping_can_writeback(inode->i_mapping))
2730		wbc.nr_to_write = 0;
2731
2732	might_sleep();
2733	return writeback_single_inode(inode, &wbc);
2734}
2735EXPORT_SYMBOL(write_inode_now);
2736
2737/**
2738 * sync_inode - write an inode and its pages to disk.
2739 * @inode: the inode to sync
2740 * @wbc: controls the writeback mode
2741 *
2742 * sync_inode() will write an inode and its pages to disk.  It will also
2743 * correctly update the inode on its superblock's dirty inode lists and will
2744 * update inode->i_state.
2745 *
2746 * The caller must have a ref on the inode.
2747 */
2748int sync_inode(struct inode *inode, struct writeback_control *wbc)
2749{
2750	return writeback_single_inode(inode, wbc);
2751}
2752EXPORT_SYMBOL(sync_inode);
2753
2754/**
2755 * sync_inode_metadata - write an inode to disk
2756 * @inode: the inode to sync
2757 * @wait: wait for I/O to complete.
2758 *
2759 * Write an inode to disk and adjust its dirty state after completion.
2760 *
2761 * Note: only writes the actual inode, no associated data or other metadata.
2762 */
2763int sync_inode_metadata(struct inode *inode, int wait)
2764{
2765	struct writeback_control wbc = {
2766		.sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2767		.nr_to_write = 0, /* metadata-only */
2768	};
2769
2770	return sync_inode(inode, &wbc);
2771}
2772EXPORT_SYMBOL(sync_inode_metadata);
2773