docs: hwmon: (pmbus/ltc2978) Update datasheet URLs to analog.com.
[linux-2.6-microblaze.git] / fs / fs-writeback.c
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         unsigned long *older_than_this;
46         enum writeback_sync_modes sync_mode;
47         unsigned int tagged_writepages:1;
48         unsigned int for_kupdate:1;
49         unsigned int range_cyclic:1;
50         unsigned int for_background:1;
51         unsigned int for_sync:1;        /* sync(2) WB_SYNC_ALL writeback */
52         unsigned int auto_free:1;       /* free on completion */
53         enum wb_reason reason;          /* why was writeback initiated? */
54
55         struct list_head list;          /* pending work list */
56         struct wb_completion *done;     /* set if the caller waits */
57 };
58
59 /*
60  * If an inode is constantly having its pages dirtied, but then the
61  * updates stop dirtytime_expire_interval seconds in the past, it's
62  * possible for the worst case time between when an inode has its
63  * timestamps updated and when they finally get written out to be two
64  * dirtytime_expire_intervals.  We set the default to 12 hours (in
65  * seconds), which means most of the time inodes will have their
66  * timestamps written to disk after 12 hours, but in the worst case a
67  * few inodes might not their timestamps updated for 24 hours.
68  */
69 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
70
71 static inline struct inode *wb_inode(struct list_head *head)
72 {
73         return list_entry(head, struct inode, i_io_list);
74 }
75
76 /*
77  * Include the creation of the trace points after defining the
78  * wb_writeback_work structure and inline functions so that the definition
79  * remains local to this file.
80  */
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/writeback.h>
83
84 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
85
86 static bool wb_io_lists_populated(struct bdi_writeback *wb)
87 {
88         if (wb_has_dirty_io(wb)) {
89                 return false;
90         } else {
91                 set_bit(WB_has_dirty_io, &wb->state);
92                 WARN_ON_ONCE(!wb->avg_write_bandwidth);
93                 atomic_long_add(wb->avg_write_bandwidth,
94                                 &wb->bdi->tot_write_bandwidth);
95                 return true;
96         }
97 }
98
99 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
100 {
101         if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
102             list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
103                 clear_bit(WB_has_dirty_io, &wb->state);
104                 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
105                                         &wb->bdi->tot_write_bandwidth) < 0);
106         }
107 }
108
109 /**
110  * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
111  * @inode: inode to be moved
112  * @wb: target bdi_writeback
113  * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
114  *
115  * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
116  * Returns %true if @inode is the first occupant of the !dirty_time IO
117  * lists; otherwise, %false.
118  */
119 static bool inode_io_list_move_locked(struct inode *inode,
120                                       struct bdi_writeback *wb,
121                                       struct list_head *head)
122 {
123         assert_spin_locked(&wb->list_lock);
124
125         list_move(&inode->i_io_list, head);
126
127         /* dirty_time doesn't count as dirty_io until expiration */
128         if (head != &wb->b_dirty_time)
129                 return wb_io_lists_populated(wb);
130
131         wb_io_lists_depopulated(wb);
132         return false;
133 }
134
135 /**
136  * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
137  * @inode: inode to be removed
138  * @wb: bdi_writeback @inode is being removed from
139  *
140  * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
141  * clear %WB_has_dirty_io if all are empty afterwards.
142  */
143 static void inode_io_list_del_locked(struct inode *inode,
144                                      struct bdi_writeback *wb)
145 {
146         assert_spin_locked(&wb->list_lock);
147
148         list_del_init(&inode->i_io_list);
149         wb_io_lists_depopulated(wb);
150 }
151
152 static void wb_wakeup(struct bdi_writeback *wb)
153 {
154         spin_lock_bh(&wb->work_lock);
155         if (test_bit(WB_registered, &wb->state))
156                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
157         spin_unlock_bh(&wb->work_lock);
158 }
159
160 static void finish_writeback_work(struct bdi_writeback *wb,
161                                   struct wb_writeback_work *work)
162 {
163         struct wb_completion *done = work->done;
164
165         if (work->auto_free)
166                 kfree(work);
167         if (done) {
168                 wait_queue_head_t *waitq = done->waitq;
169
170                 /* @done can't be accessed after the following dec */
171                 if (atomic_dec_and_test(&done->cnt))
172                         wake_up_all(waitq);
173         }
174 }
175
176 static void wb_queue_work(struct bdi_writeback *wb,
177                           struct wb_writeback_work *work)
178 {
179         trace_writeback_queue(wb, work);
180
181         if (work->done)
182                 atomic_inc(&work->done->cnt);
183
184         spin_lock_bh(&wb->work_lock);
185
186         if (test_bit(WB_registered, &wb->state)) {
187                 list_add_tail(&work->list, &wb->work_list);
188                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
189         } else
190                 finish_writeback_work(wb, work);
191
192         spin_unlock_bh(&wb->work_lock);
193 }
194
195 /**
196  * wb_wait_for_completion - wait for completion of bdi_writeback_works
197  * @done: target wb_completion
198  *
199  * Wait for one or more work items issued to @bdi with their ->done field
200  * set to @done, which should have been initialized with
201  * DEFINE_WB_COMPLETION().  This function returns after all such work items
202  * are completed.  Work items which are waited upon aren't freed
203  * automatically on completion.
204  */
205 void wb_wait_for_completion(struct wb_completion *done)
206 {
207         atomic_dec(&done->cnt);         /* put down the initial count */
208         wait_event(*done->waitq, !atomic_read(&done->cnt));
209 }
210
211 #ifdef CONFIG_CGROUP_WRITEBACK
212
213 /*
214  * Parameters for foreign inode detection, see wbc_detach_inode() to see
215  * how they're used.
216  *
217  * These paramters are inherently heuristical as the detection target
218  * itself is fuzzy.  All we want to do is detaching an inode from the
219  * current owner if it's being written to by some other cgroups too much.
220  *
221  * The current cgroup writeback is built on the assumption that multiple
222  * cgroups writing to the same inode concurrently is very rare and a mode
223  * of operation which isn't well supported.  As such, the goal is not
224  * taking too long when a different cgroup takes over an inode while
225  * avoiding too aggressive flip-flops from occasional foreign writes.
226  *
227  * We record, very roughly, 2s worth of IO time history and if more than
228  * half of that is foreign, trigger the switch.  The recording is quantized
229  * to 16 slots.  To avoid tiny writes from swinging the decision too much,
230  * writes smaller than 1/8 of avg size are ignored.
231  */
232 #define WB_FRN_TIME_SHIFT       13      /* 1s = 2^13, upto 8 secs w/ 16bit */
233 #define WB_FRN_TIME_AVG_SHIFT   3       /* avg = avg * 7/8 + new * 1/8 */
234 #define WB_FRN_TIME_CUT_DIV     8       /* ignore rounds < avg / 8 */
235 #define WB_FRN_TIME_PERIOD      (2 * (1 << WB_FRN_TIME_SHIFT))  /* 2s */
236
237 #define WB_FRN_HIST_SLOTS       16      /* inode->i_wb_frn_history is 16bit */
238 #define WB_FRN_HIST_UNIT        (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
239                                         /* each slot's duration is 2s / 16 */
240 #define WB_FRN_HIST_THR_SLOTS   (WB_FRN_HIST_SLOTS / 2)
241                                         /* if foreign slots >= 8, switch */
242 #define WB_FRN_HIST_MAX_SLOTS   (WB_FRN_HIST_THR_SLOTS / 2 + 1)
243                                         /* one round can affect upto 5 slots */
244 #define WB_FRN_MAX_IN_FLIGHT    1024    /* don't queue too many concurrently */
245
246 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
247 static struct workqueue_struct *isw_wq;
248
249 void __inode_attach_wb(struct inode *inode, struct page *page)
250 {
251         struct backing_dev_info *bdi = inode_to_bdi(inode);
252         struct bdi_writeback *wb = NULL;
253
254         if (inode_cgwb_enabled(inode)) {
255                 struct cgroup_subsys_state *memcg_css;
256
257                 if (page) {
258                         memcg_css = mem_cgroup_css_from_page(page);
259                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
260                 } else {
261                         /* must pin memcg_css, see wb_get_create() */
262                         memcg_css = task_get_css(current, memory_cgrp_id);
263                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
264                         css_put(memcg_css);
265                 }
266         }
267
268         if (!wb)
269                 wb = &bdi->wb;
270
271         /*
272          * There may be multiple instances of this function racing to
273          * update the same inode.  Use cmpxchg() to tell the winner.
274          */
275         if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
276                 wb_put(wb);
277 }
278 EXPORT_SYMBOL_GPL(__inode_attach_wb);
279
280 /**
281  * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
282  * @inode: inode of interest with i_lock held
283  *
284  * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
285  * held on entry and is released on return.  The returned wb is guaranteed
286  * to stay @inode's associated wb until its list_lock is released.
287  */
288 static struct bdi_writeback *
289 locked_inode_to_wb_and_lock_list(struct inode *inode)
290         __releases(&inode->i_lock)
291         __acquires(&wb->list_lock)
292 {
293         while (true) {
294                 struct bdi_writeback *wb = inode_to_wb(inode);
295
296                 /*
297                  * inode_to_wb() association is protected by both
298                  * @inode->i_lock and @wb->list_lock but list_lock nests
299                  * outside i_lock.  Drop i_lock and verify that the
300                  * association hasn't changed after acquiring list_lock.
301                  */
302                 wb_get(wb);
303                 spin_unlock(&inode->i_lock);
304                 spin_lock(&wb->list_lock);
305
306                 /* i_wb may have changed inbetween, can't use inode_to_wb() */
307                 if (likely(wb == inode->i_wb)) {
308                         wb_put(wb);     /* @inode already has ref */
309                         return wb;
310                 }
311
312                 spin_unlock(&wb->list_lock);
313                 wb_put(wb);
314                 cpu_relax();
315                 spin_lock(&inode->i_lock);
316         }
317 }
318
319 /**
320  * inode_to_wb_and_lock_list - determine an inode's wb and lock it
321  * @inode: inode of interest
322  *
323  * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
324  * on entry.
325  */
326 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
327         __acquires(&wb->list_lock)
328 {
329         spin_lock(&inode->i_lock);
330         return locked_inode_to_wb_and_lock_list(inode);
331 }
332
333 struct inode_switch_wbs_context {
334         struct inode            *inode;
335         struct bdi_writeback    *new_wb;
336
337         struct rcu_head         rcu_head;
338         struct work_struct      work;
339 };
340
341 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
342 {
343         down_write(&bdi->wb_switch_rwsem);
344 }
345
346 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
347 {
348         up_write(&bdi->wb_switch_rwsem);
349 }
350
351 static void inode_switch_wbs_work_fn(struct work_struct *work)
352 {
353         struct inode_switch_wbs_context *isw =
354                 container_of(work, struct inode_switch_wbs_context, work);
355         struct inode *inode = isw->inode;
356         struct backing_dev_info *bdi = inode_to_bdi(inode);
357         struct address_space *mapping = inode->i_mapping;
358         struct bdi_writeback *old_wb = inode->i_wb;
359         struct bdi_writeback *new_wb = isw->new_wb;
360         XA_STATE(xas, &mapping->i_pages, 0);
361         struct page *page;
362         bool switched = false;
363
364         /*
365          * If @inode switches cgwb membership while sync_inodes_sb() is
366          * being issued, sync_inodes_sb() might miss it.  Synchronize.
367          */
368         down_read(&bdi->wb_switch_rwsem);
369
370         /*
371          * By the time control reaches here, RCU grace period has passed
372          * since I_WB_SWITCH assertion and all wb stat update transactions
373          * between unlocked_inode_to_wb_begin/end() are guaranteed to be
374          * synchronizing against the i_pages lock.
375          *
376          * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
377          * gives us exclusion against all wb related operations on @inode
378          * including IO list manipulations and stat updates.
379          */
380         if (old_wb < new_wb) {
381                 spin_lock(&old_wb->list_lock);
382                 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
383         } else {
384                 spin_lock(&new_wb->list_lock);
385                 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
386         }
387         spin_lock(&inode->i_lock);
388         xa_lock_irq(&mapping->i_pages);
389
390         /*
391          * Once I_FREEING is visible under i_lock, the eviction path owns
392          * the inode and we shouldn't modify ->i_io_list.
393          */
394         if (unlikely(inode->i_state & I_FREEING))
395                 goto skip_switch;
396
397         trace_inode_switch_wbs(inode, old_wb, new_wb);
398
399         /*
400          * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
401          * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
402          * pages actually under writeback.
403          */
404         xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
405                 if (PageDirty(page)) {
406                         dec_wb_stat(old_wb, WB_RECLAIMABLE);
407                         inc_wb_stat(new_wb, WB_RECLAIMABLE);
408                 }
409         }
410
411         xas_set(&xas, 0);
412         xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
413                 WARN_ON_ONCE(!PageWriteback(page));
414                 dec_wb_stat(old_wb, WB_WRITEBACK);
415                 inc_wb_stat(new_wb, WB_WRITEBACK);
416         }
417
418         wb_get(new_wb);
419
420         /*
421          * Transfer to @new_wb's IO list if necessary.  The specific list
422          * @inode was on is ignored and the inode is put on ->b_dirty which
423          * is always correct including from ->b_dirty_time.  The transfer
424          * preserves @inode->dirtied_when ordering.
425          */
426         if (!list_empty(&inode->i_io_list)) {
427                 struct inode *pos;
428
429                 inode_io_list_del_locked(inode, old_wb);
430                 inode->i_wb = new_wb;
431                 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
432                         if (time_after_eq(inode->dirtied_when,
433                                           pos->dirtied_when))
434                                 break;
435                 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
436         } else {
437                 inode->i_wb = new_wb;
438         }
439
440         /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
441         inode->i_wb_frn_winner = 0;
442         inode->i_wb_frn_avg_time = 0;
443         inode->i_wb_frn_history = 0;
444         switched = true;
445 skip_switch:
446         /*
447          * Paired with load_acquire in unlocked_inode_to_wb_begin() and
448          * ensures that the new wb is visible if they see !I_WB_SWITCH.
449          */
450         smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
451
452         xa_unlock_irq(&mapping->i_pages);
453         spin_unlock(&inode->i_lock);
454         spin_unlock(&new_wb->list_lock);
455         spin_unlock(&old_wb->list_lock);
456
457         up_read(&bdi->wb_switch_rwsem);
458
459         if (switched) {
460                 wb_wakeup(new_wb);
461                 wb_put(old_wb);
462         }
463         wb_put(new_wb);
464
465         iput(inode);
466         kfree(isw);
467
468         atomic_dec(&isw_nr_in_flight);
469 }
470
471 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
472 {
473         struct inode_switch_wbs_context *isw = container_of(rcu_head,
474                                 struct inode_switch_wbs_context, rcu_head);
475
476         /* needs to grab bh-unsafe locks, bounce to work item */
477         INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
478         queue_work(isw_wq, &isw->work);
479 }
480
481 /**
482  * inode_switch_wbs - change the wb association of an inode
483  * @inode: target inode
484  * @new_wb_id: ID of the new wb
485  *
486  * Switch @inode's wb association to the wb identified by @new_wb_id.  The
487  * switching is performed asynchronously and may fail silently.
488  */
489 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
490 {
491         struct backing_dev_info *bdi = inode_to_bdi(inode);
492         struct cgroup_subsys_state *memcg_css;
493         struct inode_switch_wbs_context *isw;
494
495         /* noop if seems to be already in progress */
496         if (inode->i_state & I_WB_SWITCH)
497                 return;
498
499         /* avoid queueing a new switch if too many are already in flight */
500         if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
501                 return;
502
503         isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
504         if (!isw)
505                 return;
506
507         /* find and pin the new wb */
508         rcu_read_lock();
509         memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
510         if (memcg_css)
511                 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
512         rcu_read_unlock();
513         if (!isw->new_wb)
514                 goto out_free;
515
516         /* while holding I_WB_SWITCH, no one else can update the association */
517         spin_lock(&inode->i_lock);
518         if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
519             inode->i_state & (I_WB_SWITCH | I_FREEING) ||
520             inode_to_wb(inode) == isw->new_wb) {
521                 spin_unlock(&inode->i_lock);
522                 goto out_free;
523         }
524         inode->i_state |= I_WB_SWITCH;
525         __iget(inode);
526         spin_unlock(&inode->i_lock);
527
528         isw->inode = inode;
529
530         /*
531          * In addition to synchronizing among switchers, I_WB_SWITCH tells
532          * the RCU protected stat update paths to grab the i_page
533          * lock so that stat transfer can synchronize against them.
534          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
535          */
536         call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
537
538         atomic_inc(&isw_nr_in_flight);
539         return;
540
541 out_free:
542         if (isw->new_wb)
543                 wb_put(isw->new_wb);
544         kfree(isw);
545 }
546
547 /**
548  * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
549  * @wbc: writeback_control of interest
550  * @inode: target inode
551  *
552  * @inode is locked and about to be written back under the control of @wbc.
553  * Record @inode's writeback context into @wbc and unlock the i_lock.  On
554  * writeback completion, wbc_detach_inode() should be called.  This is used
555  * to track the cgroup writeback context.
556  */
557 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
558                                  struct inode *inode)
559 {
560         if (!inode_cgwb_enabled(inode)) {
561                 spin_unlock(&inode->i_lock);
562                 return;
563         }
564
565         wbc->wb = inode_to_wb(inode);
566         wbc->inode = inode;
567
568         wbc->wb_id = wbc->wb->memcg_css->id;
569         wbc->wb_lcand_id = inode->i_wb_frn_winner;
570         wbc->wb_tcand_id = 0;
571         wbc->wb_bytes = 0;
572         wbc->wb_lcand_bytes = 0;
573         wbc->wb_tcand_bytes = 0;
574
575         wb_get(wbc->wb);
576         spin_unlock(&inode->i_lock);
577
578         /*
579          * A dying wb indicates that either the blkcg associated with the
580          * memcg changed or the associated memcg is dying.  In the first
581          * case, a replacement wb should already be available and we should
582          * refresh the wb immediately.  In the second case, trying to
583          * refresh will keep failing.
584          */
585         if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
586                 inode_switch_wbs(inode, wbc->wb_id);
587 }
588 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
589
590 /**
591  * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
592  * @wbc: writeback_control of the just finished writeback
593  *
594  * To be called after a writeback attempt of an inode finishes and undoes
595  * wbc_attach_and_unlock_inode().  Can be called under any context.
596  *
597  * As concurrent write sharing of an inode is expected to be very rare and
598  * memcg only tracks page ownership on first-use basis severely confining
599  * the usefulness of such sharing, cgroup writeback tracks ownership
600  * per-inode.  While the support for concurrent write sharing of an inode
601  * is deemed unnecessary, an inode being written to by different cgroups at
602  * different points in time is a lot more common, and, more importantly,
603  * charging only by first-use can too readily lead to grossly incorrect
604  * behaviors (single foreign page can lead to gigabytes of writeback to be
605  * incorrectly attributed).
606  *
607  * To resolve this issue, cgroup writeback detects the majority dirtier of
608  * an inode and transfers the ownership to it.  To avoid unnnecessary
609  * oscillation, the detection mechanism keeps track of history and gives
610  * out the switch verdict only if the foreign usage pattern is stable over
611  * a certain amount of time and/or writeback attempts.
612  *
613  * On each writeback attempt, @wbc tries to detect the majority writer
614  * using Boyer-Moore majority vote algorithm.  In addition to the byte
615  * count from the majority voting, it also counts the bytes written for the
616  * current wb and the last round's winner wb (max of last round's current
617  * wb, the winner from two rounds ago, and the last round's majority
618  * candidate).  Keeping track of the historical winner helps the algorithm
619  * to semi-reliably detect the most active writer even when it's not the
620  * absolute majority.
621  *
622  * Once the winner of the round is determined, whether the winner is
623  * foreign or not and how much IO time the round consumed is recorded in
624  * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
625  * over a certain threshold, the switch verdict is given.
626  */
627 void wbc_detach_inode(struct writeback_control *wbc)
628 {
629         struct bdi_writeback *wb = wbc->wb;
630         struct inode *inode = wbc->inode;
631         unsigned long avg_time, max_bytes, max_time;
632         u16 history;
633         int max_id;
634
635         if (!wb)
636                 return;
637
638         history = inode->i_wb_frn_history;
639         avg_time = inode->i_wb_frn_avg_time;
640
641         /* pick the winner of this round */
642         if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
643             wbc->wb_bytes >= wbc->wb_tcand_bytes) {
644                 max_id = wbc->wb_id;
645                 max_bytes = wbc->wb_bytes;
646         } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
647                 max_id = wbc->wb_lcand_id;
648                 max_bytes = wbc->wb_lcand_bytes;
649         } else {
650                 max_id = wbc->wb_tcand_id;
651                 max_bytes = wbc->wb_tcand_bytes;
652         }
653
654         /*
655          * Calculate the amount of IO time the winner consumed and fold it
656          * into the running average kept per inode.  If the consumed IO
657          * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
658          * deciding whether to switch or not.  This is to prevent one-off
659          * small dirtiers from skewing the verdict.
660          */
661         max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
662                                 wb->avg_write_bandwidth);
663         if (avg_time)
664                 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
665                             (avg_time >> WB_FRN_TIME_AVG_SHIFT);
666         else
667                 avg_time = max_time;    /* immediate catch up on first run */
668
669         if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
670                 int slots;
671
672                 /*
673                  * The switch verdict is reached if foreign wb's consume
674                  * more than a certain proportion of IO time in a
675                  * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
676                  * history mask where each bit represents one sixteenth of
677                  * the period.  Determine the number of slots to shift into
678                  * history from @max_time.
679                  */
680                 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
681                             (unsigned long)WB_FRN_HIST_MAX_SLOTS);
682                 history <<= slots;
683                 if (wbc->wb_id != max_id)
684                         history |= (1U << slots) - 1;
685
686                 if (history)
687                         trace_inode_foreign_history(inode, wbc, history);
688
689                 /*
690                  * Switch if the current wb isn't the consistent winner.
691                  * If there are multiple closely competing dirtiers, the
692                  * inode may switch across them repeatedly over time, which
693                  * is okay.  The main goal is avoiding keeping an inode on
694                  * the wrong wb for an extended period of time.
695                  */
696                 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
697                         inode_switch_wbs(inode, max_id);
698         }
699
700         /*
701          * Multiple instances of this function may race to update the
702          * following fields but we don't mind occassional inaccuracies.
703          */
704         inode->i_wb_frn_winner = max_id;
705         inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
706         inode->i_wb_frn_history = history;
707
708         wb_put(wbc->wb);
709         wbc->wb = NULL;
710 }
711 EXPORT_SYMBOL_GPL(wbc_detach_inode);
712
713 /**
714  * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
715  * @wbc: writeback_control of the writeback in progress
716  * @page: page being written out
717  * @bytes: number of bytes being written out
718  *
719  * @bytes from @page are about to written out during the writeback
720  * controlled by @wbc.  Keep the book for foreign inode detection.  See
721  * wbc_detach_inode().
722  */
723 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
724                               size_t bytes)
725 {
726         struct cgroup_subsys_state *css;
727         int id;
728
729         /*
730          * pageout() path doesn't attach @wbc to the inode being written
731          * out.  This is intentional as we don't want the function to block
732          * behind a slow cgroup.  Ultimately, we want pageout() to kick off
733          * regular writeback instead of writing things out itself.
734          */
735         if (!wbc->wb || wbc->no_cgroup_owner)
736                 return;
737
738         css = mem_cgroup_css_from_page(page);
739         /* dead cgroups shouldn't contribute to inode ownership arbitration */
740         if (!(css->flags & CSS_ONLINE))
741                 return;
742
743         id = css->id;
744
745         if (id == wbc->wb_id) {
746                 wbc->wb_bytes += bytes;
747                 return;
748         }
749
750         if (id == wbc->wb_lcand_id)
751                 wbc->wb_lcand_bytes += bytes;
752
753         /* Boyer-Moore majority vote algorithm */
754         if (!wbc->wb_tcand_bytes)
755                 wbc->wb_tcand_id = id;
756         if (id == wbc->wb_tcand_id)
757                 wbc->wb_tcand_bytes += bytes;
758         else
759                 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
760 }
761 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
762
763 /**
764  * inode_congested - test whether an inode is congested
765  * @inode: inode to test for congestion (may be NULL)
766  * @cong_bits: mask of WB_[a]sync_congested bits to test
767  *
768  * Tests whether @inode is congested.  @cong_bits is the mask of congestion
769  * bits to test and the return value is the mask of set bits.
770  *
771  * If cgroup writeback is enabled for @inode, the congestion state is
772  * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
773  * associated with @inode is congested; otherwise, the root wb's congestion
774  * state is used.
775  *
776  * @inode is allowed to be NULL as this function is often called on
777  * mapping->host which is NULL for the swapper space.
778  */
779 int inode_congested(struct inode *inode, int cong_bits)
780 {
781         /*
782          * Once set, ->i_wb never becomes NULL while the inode is alive.
783          * Start transaction iff ->i_wb is visible.
784          */
785         if (inode && inode_to_wb_is_valid(inode)) {
786                 struct bdi_writeback *wb;
787                 struct wb_lock_cookie lock_cookie = {};
788                 bool congested;
789
790                 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
791                 congested = wb_congested(wb, cong_bits);
792                 unlocked_inode_to_wb_end(inode, &lock_cookie);
793                 return congested;
794         }
795
796         return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
797 }
798 EXPORT_SYMBOL_GPL(inode_congested);
799
800 /**
801  * wb_split_bdi_pages - split nr_pages to write according to bandwidth
802  * @wb: target bdi_writeback to split @nr_pages to
803  * @nr_pages: number of pages to write for the whole bdi
804  *
805  * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
806  * relation to the total write bandwidth of all wb's w/ dirty inodes on
807  * @wb->bdi.
808  */
809 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
810 {
811         unsigned long this_bw = wb->avg_write_bandwidth;
812         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
813
814         if (nr_pages == LONG_MAX)
815                 return LONG_MAX;
816
817         /*
818          * This may be called on clean wb's and proportional distribution
819          * may not make sense, just use the original @nr_pages in those
820          * cases.  In general, we wanna err on the side of writing more.
821          */
822         if (!tot_bw || this_bw >= tot_bw)
823                 return nr_pages;
824         else
825                 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
826 }
827
828 /**
829  * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
830  * @bdi: target backing_dev_info
831  * @base_work: wb_writeback_work to issue
832  * @skip_if_busy: skip wb's which already have writeback in progress
833  *
834  * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
835  * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
836  * distributed to the busy wbs according to each wb's proportion in the
837  * total active write bandwidth of @bdi.
838  */
839 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
840                                   struct wb_writeback_work *base_work,
841                                   bool skip_if_busy)
842 {
843         struct bdi_writeback *last_wb = NULL;
844         struct bdi_writeback *wb = list_entry(&bdi->wb_list,
845                                               struct bdi_writeback, bdi_node);
846
847         might_sleep();
848 restart:
849         rcu_read_lock();
850         list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
851                 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
852                 struct wb_writeback_work fallback_work;
853                 struct wb_writeback_work *work;
854                 long nr_pages;
855
856                 if (last_wb) {
857                         wb_put(last_wb);
858                         last_wb = NULL;
859                 }
860
861                 /* SYNC_ALL writes out I_DIRTY_TIME too */
862                 if (!wb_has_dirty_io(wb) &&
863                     (base_work->sync_mode == WB_SYNC_NONE ||
864                      list_empty(&wb->b_dirty_time)))
865                         continue;
866                 if (skip_if_busy && writeback_in_progress(wb))
867                         continue;
868
869                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
870
871                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
872                 if (work) {
873                         *work = *base_work;
874                         work->nr_pages = nr_pages;
875                         work->auto_free = 1;
876                         wb_queue_work(wb, work);
877                         continue;
878                 }
879
880                 /* alloc failed, execute synchronously using on-stack fallback */
881                 work = &fallback_work;
882                 *work = *base_work;
883                 work->nr_pages = nr_pages;
884                 work->auto_free = 0;
885                 work->done = &fallback_work_done;
886
887                 wb_queue_work(wb, work);
888
889                 /*
890                  * Pin @wb so that it stays on @bdi->wb_list.  This allows
891                  * continuing iteration from @wb after dropping and
892                  * regrabbing rcu read lock.
893                  */
894                 wb_get(wb);
895                 last_wb = wb;
896
897                 rcu_read_unlock();
898                 wb_wait_for_completion(&fallback_work_done);
899                 goto restart;
900         }
901         rcu_read_unlock();
902
903         if (last_wb)
904                 wb_put(last_wb);
905 }
906
907 /**
908  * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
909  * @bdi_id: target bdi id
910  * @memcg_id: target memcg css id
911  * @nr: number of pages to write, 0 for best-effort dirty flushing
912  * @reason: reason why some writeback work initiated
913  * @done: target wb_completion
914  *
915  * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
916  * with the specified parameters.
917  */
918 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
919                            enum wb_reason reason, struct wb_completion *done)
920 {
921         struct backing_dev_info *bdi;
922         struct cgroup_subsys_state *memcg_css;
923         struct bdi_writeback *wb;
924         struct wb_writeback_work *work;
925         int ret;
926
927         /* lookup bdi and memcg */
928         bdi = bdi_get_by_id(bdi_id);
929         if (!bdi)
930                 return -ENOENT;
931
932         rcu_read_lock();
933         memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
934         if (memcg_css && !css_tryget(memcg_css))
935                 memcg_css = NULL;
936         rcu_read_unlock();
937         if (!memcg_css) {
938                 ret = -ENOENT;
939                 goto out_bdi_put;
940         }
941
942         /*
943          * And find the associated wb.  If the wb isn't there already
944          * there's nothing to flush, don't create one.
945          */
946         wb = wb_get_lookup(bdi, memcg_css);
947         if (!wb) {
948                 ret = -ENOENT;
949                 goto out_css_put;
950         }
951
952         /*
953          * If @nr is zero, the caller is attempting to write out most of
954          * the currently dirty pages.  Let's take the current dirty page
955          * count and inflate it by 25% which should be large enough to
956          * flush out most dirty pages while avoiding getting livelocked by
957          * concurrent dirtiers.
958          */
959         if (!nr) {
960                 unsigned long filepages, headroom, dirty, writeback;
961
962                 mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
963                                       &writeback);
964                 nr = dirty * 10 / 8;
965         }
966
967         /* issue the writeback work */
968         work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
969         if (work) {
970                 work->nr_pages = nr;
971                 work->sync_mode = WB_SYNC_NONE;
972                 work->range_cyclic = 1;
973                 work->reason = reason;
974                 work->done = done;
975                 work->auto_free = 1;
976                 wb_queue_work(wb, work);
977                 ret = 0;
978         } else {
979                 ret = -ENOMEM;
980         }
981
982         wb_put(wb);
983 out_css_put:
984         css_put(memcg_css);
985 out_bdi_put:
986         bdi_put(bdi);
987         return ret;
988 }
989
990 /**
991  * cgroup_writeback_umount - flush inode wb switches for umount
992  *
993  * This function is called when a super_block is about to be destroyed and
994  * flushes in-flight inode wb switches.  An inode wb switch goes through
995  * RCU and then workqueue, so the two need to be flushed in order to ensure
996  * that all previously scheduled switches are finished.  As wb switches are
997  * rare occurrences and synchronize_rcu() can take a while, perform
998  * flushing iff wb switches are in flight.
999  */
1000 void cgroup_writeback_umount(void)
1001 {
1002         if (atomic_read(&isw_nr_in_flight)) {
1003                 /*
1004                  * Use rcu_barrier() to wait for all pending callbacks to
1005                  * ensure that all in-flight wb switches are in the workqueue.
1006                  */
1007                 rcu_barrier();
1008                 flush_workqueue(isw_wq);
1009         }
1010 }
1011
1012 static int __init cgroup_writeback_init(void)
1013 {
1014         isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1015         if (!isw_wq)
1016                 return -ENOMEM;
1017         return 0;
1018 }
1019 fs_initcall(cgroup_writeback_init);
1020
1021 #else   /* CONFIG_CGROUP_WRITEBACK */
1022
1023 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1024 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1025
1026 static struct bdi_writeback *
1027 locked_inode_to_wb_and_lock_list(struct inode *inode)
1028         __releases(&inode->i_lock)
1029         __acquires(&wb->list_lock)
1030 {
1031         struct bdi_writeback *wb = inode_to_wb(inode);
1032
1033         spin_unlock(&inode->i_lock);
1034         spin_lock(&wb->list_lock);
1035         return wb;
1036 }
1037
1038 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1039         __acquires(&wb->list_lock)
1040 {
1041         struct bdi_writeback *wb = inode_to_wb(inode);
1042
1043         spin_lock(&wb->list_lock);
1044         return wb;
1045 }
1046
1047 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1048 {
1049         return nr_pages;
1050 }
1051
1052 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1053                                   struct wb_writeback_work *base_work,
1054                                   bool skip_if_busy)
1055 {
1056         might_sleep();
1057
1058         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1059                 base_work->auto_free = 0;
1060                 wb_queue_work(&bdi->wb, base_work);
1061         }
1062 }
1063
1064 #endif  /* CONFIG_CGROUP_WRITEBACK */
1065
1066 /*
1067  * Add in the number of potentially dirty inodes, because each inode
1068  * write can dirty pagecache in the underlying blockdev.
1069  */
1070 static unsigned long get_nr_dirty_pages(void)
1071 {
1072         return global_node_page_state(NR_FILE_DIRTY) +
1073                 global_node_page_state(NR_UNSTABLE_NFS) +
1074                 get_nr_dirty_inodes();
1075 }
1076
1077 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1078 {
1079         if (!wb_has_dirty_io(wb))
1080                 return;
1081
1082         /*
1083          * All callers of this function want to start writeback of all
1084          * dirty pages. Places like vmscan can call this at a very
1085          * high frequency, causing pointless allocations of tons of
1086          * work items and keeping the flusher threads busy retrieving
1087          * that work. Ensure that we only allow one of them pending and
1088          * inflight at the time.
1089          */
1090         if (test_bit(WB_start_all, &wb->state) ||
1091             test_and_set_bit(WB_start_all, &wb->state))
1092                 return;
1093
1094         wb->start_all_reason = reason;
1095         wb_wakeup(wb);
1096 }
1097
1098 /**
1099  * wb_start_background_writeback - start background writeback
1100  * @wb: bdi_writback to write from
1101  *
1102  * Description:
1103  *   This makes sure WB_SYNC_NONE background writeback happens. When
1104  *   this function returns, it is only guaranteed that for given wb
1105  *   some IO is happening if we are over background dirty threshold.
1106  *   Caller need not hold sb s_umount semaphore.
1107  */
1108 void wb_start_background_writeback(struct bdi_writeback *wb)
1109 {
1110         /*
1111          * We just wake up the flusher thread. It will perform background
1112          * writeback as soon as there is no other work to do.
1113          */
1114         trace_writeback_wake_background(wb);
1115         wb_wakeup(wb);
1116 }
1117
1118 /*
1119  * Remove the inode from the writeback list it is on.
1120  */
1121 void inode_io_list_del(struct inode *inode)
1122 {
1123         struct bdi_writeback *wb;
1124
1125         wb = inode_to_wb_and_lock_list(inode);
1126         inode_io_list_del_locked(inode, wb);
1127         spin_unlock(&wb->list_lock);
1128 }
1129
1130 /*
1131  * mark an inode as under writeback on the sb
1132  */
1133 void sb_mark_inode_writeback(struct inode *inode)
1134 {
1135         struct super_block *sb = inode->i_sb;
1136         unsigned long flags;
1137
1138         if (list_empty(&inode->i_wb_list)) {
1139                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1140                 if (list_empty(&inode->i_wb_list)) {
1141                         list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1142                         trace_sb_mark_inode_writeback(inode);
1143                 }
1144                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1145         }
1146 }
1147
1148 /*
1149  * clear an inode as under writeback on the sb
1150  */
1151 void sb_clear_inode_writeback(struct inode *inode)
1152 {
1153         struct super_block *sb = inode->i_sb;
1154         unsigned long flags;
1155
1156         if (!list_empty(&inode->i_wb_list)) {
1157                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1158                 if (!list_empty(&inode->i_wb_list)) {
1159                         list_del_init(&inode->i_wb_list);
1160                         trace_sb_clear_inode_writeback(inode);
1161                 }
1162                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1163         }
1164 }
1165
1166 /*
1167  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1168  * furthest end of its superblock's dirty-inode list.
1169  *
1170  * Before stamping the inode's ->dirtied_when, we check to see whether it is
1171  * already the most-recently-dirtied inode on the b_dirty list.  If that is
1172  * the case then the inode must have been redirtied while it was being written
1173  * out and we don't reset its dirtied_when.
1174  */
1175 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1176 {
1177         if (!list_empty(&wb->b_dirty)) {
1178                 struct inode *tail;
1179
1180                 tail = wb_inode(wb->b_dirty.next);
1181                 if (time_before(inode->dirtied_when, tail->dirtied_when))
1182                         inode->dirtied_when = jiffies;
1183         }
1184         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1185 }
1186
1187 /*
1188  * requeue inode for re-scanning after bdi->b_io list is exhausted.
1189  */
1190 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1191 {
1192         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1193 }
1194
1195 static void inode_sync_complete(struct inode *inode)
1196 {
1197         inode->i_state &= ~I_SYNC;
1198         /* If inode is clean an unused, put it into LRU now... */
1199         inode_add_lru(inode);
1200         /* Waiters must see I_SYNC cleared before being woken up */
1201         smp_mb();
1202         wake_up_bit(&inode->i_state, __I_SYNC);
1203 }
1204
1205 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1206 {
1207         bool ret = time_after(inode->dirtied_when, t);
1208 #ifndef CONFIG_64BIT
1209         /*
1210          * For inodes being constantly redirtied, dirtied_when can get stuck.
1211          * It _appears_ to be in the future, but is actually in distant past.
1212          * This test is necessary to prevent such wrapped-around relative times
1213          * from permanently stopping the whole bdi writeback.
1214          */
1215         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1216 #endif
1217         return ret;
1218 }
1219
1220 #define EXPIRE_DIRTY_ATIME 0x0001
1221
1222 /*
1223  * Move expired (dirtied before work->older_than_this) dirty inodes from
1224  * @delaying_queue to @dispatch_queue.
1225  */
1226 static int move_expired_inodes(struct list_head *delaying_queue,
1227                                struct list_head *dispatch_queue,
1228                                int flags,
1229                                struct wb_writeback_work *work)
1230 {
1231         unsigned long *older_than_this = NULL;
1232         unsigned long expire_time;
1233         LIST_HEAD(tmp);
1234         struct list_head *pos, *node;
1235         struct super_block *sb = NULL;
1236         struct inode *inode;
1237         int do_sb_sort = 0;
1238         int moved = 0;
1239
1240         if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1241                 older_than_this = work->older_than_this;
1242         else if (!work->for_sync) {
1243                 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1244                 older_than_this = &expire_time;
1245         }
1246         while (!list_empty(delaying_queue)) {
1247                 inode = wb_inode(delaying_queue->prev);
1248                 if (older_than_this &&
1249                     inode_dirtied_after(inode, *older_than_this))
1250                         break;
1251                 list_move(&inode->i_io_list, &tmp);
1252                 moved++;
1253                 if (flags & EXPIRE_DIRTY_ATIME)
1254                         set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1255                 if (sb_is_blkdev_sb(inode->i_sb))
1256                         continue;
1257                 if (sb && sb != inode->i_sb)
1258                         do_sb_sort = 1;
1259                 sb = inode->i_sb;
1260         }
1261
1262         /* just one sb in list, splice to dispatch_queue and we're done */
1263         if (!do_sb_sort) {
1264                 list_splice(&tmp, dispatch_queue);
1265                 goto out;
1266         }
1267
1268         /* Move inodes from one superblock together */
1269         while (!list_empty(&tmp)) {
1270                 sb = wb_inode(tmp.prev)->i_sb;
1271                 list_for_each_prev_safe(pos, node, &tmp) {
1272                         inode = wb_inode(pos);
1273                         if (inode->i_sb == sb)
1274                                 list_move(&inode->i_io_list, dispatch_queue);
1275                 }
1276         }
1277 out:
1278         return moved;
1279 }
1280
1281 /*
1282  * Queue all expired dirty inodes for io, eldest first.
1283  * Before
1284  *         newly dirtied     b_dirty    b_io    b_more_io
1285  *         =============>    gf         edc     BA
1286  * After
1287  *         newly dirtied     b_dirty    b_io    b_more_io
1288  *         =============>    g          fBAedc
1289  *                                           |
1290  *                                           +--> dequeue for IO
1291  */
1292 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1293 {
1294         int moved;
1295
1296         assert_spin_locked(&wb->list_lock);
1297         list_splice_init(&wb->b_more_io, &wb->b_io);
1298         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1299         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1300                                      EXPIRE_DIRTY_ATIME, work);
1301         if (moved)
1302                 wb_io_lists_populated(wb);
1303         trace_writeback_queue_io(wb, work, moved);
1304 }
1305
1306 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1307 {
1308         int ret;
1309
1310         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1311                 trace_writeback_write_inode_start(inode, wbc);
1312                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1313                 trace_writeback_write_inode(inode, wbc);
1314                 return ret;
1315         }
1316         return 0;
1317 }
1318
1319 /*
1320  * Wait for writeback on an inode to complete. Called with i_lock held.
1321  * Caller must make sure inode cannot go away when we drop i_lock.
1322  */
1323 static void __inode_wait_for_writeback(struct inode *inode)
1324         __releases(inode->i_lock)
1325         __acquires(inode->i_lock)
1326 {
1327         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1328         wait_queue_head_t *wqh;
1329
1330         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1331         while (inode->i_state & I_SYNC) {
1332                 spin_unlock(&inode->i_lock);
1333                 __wait_on_bit(wqh, &wq, bit_wait,
1334                               TASK_UNINTERRUPTIBLE);
1335                 spin_lock(&inode->i_lock);
1336         }
1337 }
1338
1339 /*
1340  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1341  */
1342 void inode_wait_for_writeback(struct inode *inode)
1343 {
1344         spin_lock(&inode->i_lock);
1345         __inode_wait_for_writeback(inode);
1346         spin_unlock(&inode->i_lock);
1347 }
1348
1349 /*
1350  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1351  * held and drops it. It is aimed for callers not holding any inode reference
1352  * so once i_lock is dropped, inode can go away.
1353  */
1354 static void inode_sleep_on_writeback(struct inode *inode)
1355         __releases(inode->i_lock)
1356 {
1357         DEFINE_WAIT(wait);
1358         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1359         int sleep;
1360
1361         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1362         sleep = inode->i_state & I_SYNC;
1363         spin_unlock(&inode->i_lock);
1364         if (sleep)
1365                 schedule();
1366         finish_wait(wqh, &wait);
1367 }
1368
1369 /*
1370  * Find proper writeback list for the inode depending on its current state and
1371  * possibly also change of its state while we were doing writeback.  Here we
1372  * handle things such as livelock prevention or fairness of writeback among
1373  * inodes. This function can be called only by flusher thread - noone else
1374  * processes all inodes in writeback lists and requeueing inodes behind flusher
1375  * thread's back can have unexpected consequences.
1376  */
1377 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1378                           struct writeback_control *wbc)
1379 {
1380         if (inode->i_state & I_FREEING)
1381                 return;
1382
1383         /*
1384          * Sync livelock prevention. Each inode is tagged and synced in one
1385          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1386          * the dirty time to prevent enqueue and sync it again.
1387          */
1388         if ((inode->i_state & I_DIRTY) &&
1389             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1390                 inode->dirtied_when = jiffies;
1391
1392         if (wbc->pages_skipped) {
1393                 /*
1394                  * writeback is not making progress due to locked
1395                  * buffers. Skip this inode for now.
1396                  */
1397                 redirty_tail(inode, wb);
1398                 return;
1399         }
1400
1401         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1402                 /*
1403                  * We didn't write back all the pages.  nfs_writepages()
1404                  * sometimes bales out without doing anything.
1405                  */
1406                 if (wbc->nr_to_write <= 0) {
1407                         /* Slice used up. Queue for next turn. */
1408                         requeue_io(inode, wb);
1409                 } else {
1410                         /*
1411                          * Writeback blocked by something other than
1412                          * congestion. Delay the inode for some time to
1413                          * avoid spinning on the CPU (100% iowait)
1414                          * retrying writeback of the dirty page/inode
1415                          * that cannot be performed immediately.
1416                          */
1417                         redirty_tail(inode, wb);
1418                 }
1419         } else if (inode->i_state & I_DIRTY) {
1420                 /*
1421                  * Filesystems can dirty the inode during writeback operations,
1422                  * such as delayed allocation during submission or metadata
1423                  * updates after data IO completion.
1424                  */
1425                 redirty_tail(inode, wb);
1426         } else if (inode->i_state & I_DIRTY_TIME) {
1427                 inode->dirtied_when = jiffies;
1428                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1429         } else {
1430                 /* The inode is clean. Remove from writeback lists. */
1431                 inode_io_list_del_locked(inode, wb);
1432         }
1433 }
1434
1435 /*
1436  * Write out an inode and its dirty pages. Do not update the writeback list
1437  * linkage. That is left to the caller. The caller is also responsible for
1438  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1439  */
1440 static int
1441 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1442 {
1443         struct address_space *mapping = inode->i_mapping;
1444         long nr_to_write = wbc->nr_to_write;
1445         unsigned dirty;
1446         int ret;
1447
1448         WARN_ON(!(inode->i_state & I_SYNC));
1449
1450         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1451
1452         ret = do_writepages(mapping, wbc);
1453
1454         /*
1455          * Make sure to wait on the data before writing out the metadata.
1456          * This is important for filesystems that modify metadata on data
1457          * I/O completion. We don't do it for sync(2) writeback because it has a
1458          * separate, external IO completion path and ->sync_fs for guaranteeing
1459          * inode metadata is written back correctly.
1460          */
1461         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1462                 int err = filemap_fdatawait(mapping);
1463                 if (ret == 0)
1464                         ret = err;
1465         }
1466
1467         /*
1468          * Some filesystems may redirty the inode during the writeback
1469          * due to delalloc, clear dirty metadata flags right before
1470          * write_inode()
1471          */
1472         spin_lock(&inode->i_lock);
1473
1474         dirty = inode->i_state & I_DIRTY;
1475         if (inode->i_state & I_DIRTY_TIME) {
1476                 if ((dirty & I_DIRTY_INODE) ||
1477                     wbc->sync_mode == WB_SYNC_ALL ||
1478                     unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1479                     unlikely(time_after(jiffies,
1480                                         (inode->dirtied_time_when +
1481                                          dirtytime_expire_interval * HZ)))) {
1482                         dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1483                         trace_writeback_lazytime(inode);
1484                 }
1485         } else
1486                 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1487         inode->i_state &= ~dirty;
1488
1489         /*
1490          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1491          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1492          * either they see the I_DIRTY bits cleared or we see the dirtied
1493          * inode.
1494          *
1495          * I_DIRTY_PAGES is always cleared together above even if @mapping
1496          * still has dirty pages.  The flag is reinstated after smp_mb() if
1497          * necessary.  This guarantees that either __mark_inode_dirty()
1498          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1499          */
1500         smp_mb();
1501
1502         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1503                 inode->i_state |= I_DIRTY_PAGES;
1504
1505         spin_unlock(&inode->i_lock);
1506
1507         if (dirty & I_DIRTY_TIME)
1508                 mark_inode_dirty_sync(inode);
1509         /* Don't write the inode if only I_DIRTY_PAGES was set */
1510         if (dirty & ~I_DIRTY_PAGES) {
1511                 int err = write_inode(inode, wbc);
1512                 if (ret == 0)
1513                         ret = err;
1514         }
1515         trace_writeback_single_inode(inode, wbc, nr_to_write);
1516         return ret;
1517 }
1518
1519 /*
1520  * Write out an inode's dirty pages. Either the caller has an active reference
1521  * on the inode or the inode has I_WILL_FREE set.
1522  *
1523  * This function is designed to be called for writing back one inode which
1524  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1525  * and does more profound writeback list handling in writeback_sb_inodes().
1526  */
1527 static int writeback_single_inode(struct inode *inode,
1528                                   struct writeback_control *wbc)
1529 {
1530         struct bdi_writeback *wb;
1531         int ret = 0;
1532
1533         spin_lock(&inode->i_lock);
1534         if (!atomic_read(&inode->i_count))
1535                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1536         else
1537                 WARN_ON(inode->i_state & I_WILL_FREE);
1538
1539         if (inode->i_state & I_SYNC) {
1540                 if (wbc->sync_mode != WB_SYNC_ALL)
1541                         goto out;
1542                 /*
1543                  * It's a data-integrity sync. We must wait. Since callers hold
1544                  * inode reference or inode has I_WILL_FREE set, it cannot go
1545                  * away under us.
1546                  */
1547                 __inode_wait_for_writeback(inode);
1548         }
1549         WARN_ON(inode->i_state & I_SYNC);
1550         /*
1551          * Skip inode if it is clean and we have no outstanding writeback in
1552          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1553          * function since flusher thread may be doing for example sync in
1554          * parallel and if we move the inode, it could get skipped. So here we
1555          * make sure inode is on some writeback list and leave it there unless
1556          * we have completely cleaned the inode.
1557          */
1558         if (!(inode->i_state & I_DIRTY_ALL) &&
1559             (wbc->sync_mode != WB_SYNC_ALL ||
1560              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1561                 goto out;
1562         inode->i_state |= I_SYNC;
1563         wbc_attach_and_unlock_inode(wbc, inode);
1564
1565         ret = __writeback_single_inode(inode, wbc);
1566
1567         wbc_detach_inode(wbc);
1568
1569         wb = inode_to_wb_and_lock_list(inode);
1570         spin_lock(&inode->i_lock);
1571         /*
1572          * If inode is clean, remove it from writeback lists. Otherwise don't
1573          * touch it. See comment above for explanation.
1574          */
1575         if (!(inode->i_state & I_DIRTY_ALL))
1576                 inode_io_list_del_locked(inode, wb);
1577         spin_unlock(&wb->list_lock);
1578         inode_sync_complete(inode);
1579 out:
1580         spin_unlock(&inode->i_lock);
1581         return ret;
1582 }
1583
1584 static long writeback_chunk_size(struct bdi_writeback *wb,
1585                                  struct wb_writeback_work *work)
1586 {
1587         long pages;
1588
1589         /*
1590          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1591          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1592          * here avoids calling into writeback_inodes_wb() more than once.
1593          *
1594          * The intended call sequence for WB_SYNC_ALL writeback is:
1595          *
1596          *      wb_writeback()
1597          *          writeback_sb_inodes()       <== called only once
1598          *              write_cache_pages()     <== called once for each inode
1599          *                   (quickly) tag currently dirty pages
1600          *                   (maybe slowly) sync all tagged pages
1601          */
1602         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1603                 pages = LONG_MAX;
1604         else {
1605                 pages = min(wb->avg_write_bandwidth / 2,
1606                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1607                 pages = min(pages, work->nr_pages);
1608                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1609                                    MIN_WRITEBACK_PAGES);
1610         }
1611
1612         return pages;
1613 }
1614
1615 /*
1616  * Write a portion of b_io inodes which belong to @sb.
1617  *
1618  * Return the number of pages and/or inodes written.
1619  *
1620  * NOTE! This is called with wb->list_lock held, and will
1621  * unlock and relock that for each inode it ends up doing
1622  * IO for.
1623  */
1624 static long writeback_sb_inodes(struct super_block *sb,
1625                                 struct bdi_writeback *wb,
1626                                 struct wb_writeback_work *work)
1627 {
1628         struct writeback_control wbc = {
1629                 .sync_mode              = work->sync_mode,
1630                 .tagged_writepages      = work->tagged_writepages,
1631                 .for_kupdate            = work->for_kupdate,
1632                 .for_background         = work->for_background,
1633                 .for_sync               = work->for_sync,
1634                 .range_cyclic           = work->range_cyclic,
1635                 .range_start            = 0,
1636                 .range_end              = LLONG_MAX,
1637         };
1638         unsigned long start_time = jiffies;
1639         long write_chunk;
1640         long wrote = 0;  /* count both pages and inodes */
1641
1642         while (!list_empty(&wb->b_io)) {
1643                 struct inode *inode = wb_inode(wb->b_io.prev);
1644                 struct bdi_writeback *tmp_wb;
1645
1646                 if (inode->i_sb != sb) {
1647                         if (work->sb) {
1648                                 /*
1649                                  * We only want to write back data for this
1650                                  * superblock, move all inodes not belonging
1651                                  * to it back onto the dirty list.
1652                                  */
1653                                 redirty_tail(inode, wb);
1654                                 continue;
1655                         }
1656
1657                         /*
1658                          * The inode belongs to a different superblock.
1659                          * Bounce back to the caller to unpin this and
1660                          * pin the next superblock.
1661                          */
1662                         break;
1663                 }
1664
1665                 /*
1666                  * Don't bother with new inodes or inodes being freed, first
1667                  * kind does not need periodic writeout yet, and for the latter
1668                  * kind writeout is handled by the freer.
1669                  */
1670                 spin_lock(&inode->i_lock);
1671                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1672                         spin_unlock(&inode->i_lock);
1673                         redirty_tail(inode, wb);
1674                         continue;
1675                 }
1676                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1677                         /*
1678                          * If this inode is locked for writeback and we are not
1679                          * doing writeback-for-data-integrity, move it to
1680                          * b_more_io so that writeback can proceed with the
1681                          * other inodes on s_io.
1682                          *
1683                          * We'll have another go at writing back this inode
1684                          * when we completed a full scan of b_io.
1685                          */
1686                         spin_unlock(&inode->i_lock);
1687                         requeue_io(inode, wb);
1688                         trace_writeback_sb_inodes_requeue(inode);
1689                         continue;
1690                 }
1691                 spin_unlock(&wb->list_lock);
1692
1693                 /*
1694                  * We already requeued the inode if it had I_SYNC set and we
1695                  * are doing WB_SYNC_NONE writeback. So this catches only the
1696                  * WB_SYNC_ALL case.
1697                  */
1698                 if (inode->i_state & I_SYNC) {
1699                         /* Wait for I_SYNC. This function drops i_lock... */
1700                         inode_sleep_on_writeback(inode);
1701                         /* Inode may be gone, start again */
1702                         spin_lock(&wb->list_lock);
1703                         continue;
1704                 }
1705                 inode->i_state |= I_SYNC;
1706                 wbc_attach_and_unlock_inode(&wbc, inode);
1707
1708                 write_chunk = writeback_chunk_size(wb, work);
1709                 wbc.nr_to_write = write_chunk;
1710                 wbc.pages_skipped = 0;
1711
1712                 /*
1713                  * We use I_SYNC to pin the inode in memory. While it is set
1714                  * evict_inode() will wait so the inode cannot be freed.
1715                  */
1716                 __writeback_single_inode(inode, &wbc);
1717
1718                 wbc_detach_inode(&wbc);
1719                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1720                 wrote += write_chunk - wbc.nr_to_write;
1721
1722                 if (need_resched()) {
1723                         /*
1724                          * We're trying to balance between building up a nice
1725                          * long list of IOs to improve our merge rate, and
1726                          * getting those IOs out quickly for anyone throttling
1727                          * in balance_dirty_pages().  cond_resched() doesn't
1728                          * unplug, so get our IOs out the door before we
1729                          * give up the CPU.
1730                          */
1731                         blk_flush_plug(current);
1732                         cond_resched();
1733                 }
1734
1735                 /*
1736                  * Requeue @inode if still dirty.  Be careful as @inode may
1737                  * have been switched to another wb in the meantime.
1738                  */
1739                 tmp_wb = inode_to_wb_and_lock_list(inode);
1740                 spin_lock(&inode->i_lock);
1741                 if (!(inode->i_state & I_DIRTY_ALL))
1742                         wrote++;
1743                 requeue_inode(inode, tmp_wb, &wbc);
1744                 inode_sync_complete(inode);
1745                 spin_unlock(&inode->i_lock);
1746
1747                 if (unlikely(tmp_wb != wb)) {
1748                         spin_unlock(&tmp_wb->list_lock);
1749                         spin_lock(&wb->list_lock);
1750                 }
1751
1752                 /*
1753                  * bail out to wb_writeback() often enough to check
1754                  * background threshold and other termination conditions.
1755                  */
1756                 if (wrote) {
1757                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1758                                 break;
1759                         if (work->nr_pages <= 0)
1760                                 break;
1761                 }
1762         }
1763         return wrote;
1764 }
1765
1766 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1767                                   struct wb_writeback_work *work)
1768 {
1769         unsigned long start_time = jiffies;
1770         long wrote = 0;
1771
1772         while (!list_empty(&wb->b_io)) {
1773                 struct inode *inode = wb_inode(wb->b_io.prev);
1774                 struct super_block *sb = inode->i_sb;
1775
1776                 if (!trylock_super(sb)) {
1777                         /*
1778                          * trylock_super() may fail consistently due to
1779                          * s_umount being grabbed by someone else. Don't use
1780                          * requeue_io() to avoid busy retrying the inode/sb.
1781                          */
1782                         redirty_tail(inode, wb);
1783                         continue;
1784                 }
1785                 wrote += writeback_sb_inodes(sb, wb, work);
1786                 up_read(&sb->s_umount);
1787
1788                 /* refer to the same tests at the end of writeback_sb_inodes */
1789                 if (wrote) {
1790                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1791                                 break;
1792                         if (work->nr_pages <= 0)
1793                                 break;
1794                 }
1795         }
1796         /* Leave any unwritten inodes on b_io */
1797         return wrote;
1798 }
1799
1800 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1801                                 enum wb_reason reason)
1802 {
1803         struct wb_writeback_work work = {
1804                 .nr_pages       = nr_pages,
1805                 .sync_mode      = WB_SYNC_NONE,
1806                 .range_cyclic   = 1,
1807                 .reason         = reason,
1808         };
1809         struct blk_plug plug;
1810
1811         blk_start_plug(&plug);
1812         spin_lock(&wb->list_lock);
1813         if (list_empty(&wb->b_io))
1814                 queue_io(wb, &work);
1815         __writeback_inodes_wb(wb, &work);
1816         spin_unlock(&wb->list_lock);
1817         blk_finish_plug(&plug);
1818
1819         return nr_pages - work.nr_pages;
1820 }
1821
1822 /*
1823  * Explicit flushing or periodic writeback of "old" data.
1824  *
1825  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1826  * dirtying-time in the inode's address_space.  So this periodic writeback code
1827  * just walks the superblock inode list, writing back any inodes which are
1828  * older than a specific point in time.
1829  *
1830  * Try to run once per dirty_writeback_interval.  But if a writeback event
1831  * takes longer than a dirty_writeback_interval interval, then leave a
1832  * one-second gap.
1833  *
1834  * older_than_this takes precedence over nr_to_write.  So we'll only write back
1835  * all dirty pages if they are all attached to "old" mappings.
1836  */
1837 static long wb_writeback(struct bdi_writeback *wb,
1838                          struct wb_writeback_work *work)
1839 {
1840         unsigned long wb_start = jiffies;
1841         long nr_pages = work->nr_pages;
1842         unsigned long oldest_jif;
1843         struct inode *inode;
1844         long progress;
1845         struct blk_plug plug;
1846
1847         oldest_jif = jiffies;
1848         work->older_than_this = &oldest_jif;
1849
1850         blk_start_plug(&plug);
1851         spin_lock(&wb->list_lock);
1852         for (;;) {
1853                 /*
1854                  * Stop writeback when nr_pages has been consumed
1855                  */
1856                 if (work->nr_pages <= 0)
1857                         break;
1858
1859                 /*
1860                  * Background writeout and kupdate-style writeback may
1861                  * run forever. Stop them if there is other work to do
1862                  * so that e.g. sync can proceed. They'll be restarted
1863                  * after the other works are all done.
1864                  */
1865                 if ((work->for_background || work->for_kupdate) &&
1866                     !list_empty(&wb->work_list))
1867                         break;
1868
1869                 /*
1870                  * For background writeout, stop when we are below the
1871                  * background dirty threshold
1872                  */
1873                 if (work->for_background && !wb_over_bg_thresh(wb))
1874                         break;
1875
1876                 /*
1877                  * Kupdate and background works are special and we want to
1878                  * include all inodes that need writing. Livelock avoidance is
1879                  * handled by these works yielding to any other work so we are
1880                  * safe.
1881                  */
1882                 if (work->for_kupdate) {
1883                         oldest_jif = jiffies -
1884                                 msecs_to_jiffies(dirty_expire_interval * 10);
1885                 } else if (work->for_background)
1886                         oldest_jif = jiffies;
1887
1888                 trace_writeback_start(wb, work);
1889                 if (list_empty(&wb->b_io))
1890                         queue_io(wb, work);
1891                 if (work->sb)
1892                         progress = writeback_sb_inodes(work->sb, wb, work);
1893                 else
1894                         progress = __writeback_inodes_wb(wb, work);
1895                 trace_writeback_written(wb, work);
1896
1897                 wb_update_bandwidth(wb, wb_start);
1898
1899                 /*
1900                  * Did we write something? Try for more
1901                  *
1902                  * Dirty inodes are moved to b_io for writeback in batches.
1903                  * The completion of the current batch does not necessarily
1904                  * mean the overall work is done. So we keep looping as long
1905                  * as made some progress on cleaning pages or inodes.
1906                  */
1907                 if (progress)
1908                         continue;
1909                 /*
1910                  * No more inodes for IO, bail
1911                  */
1912                 if (list_empty(&wb->b_more_io))
1913                         break;
1914                 /*
1915                  * Nothing written. Wait for some inode to
1916                  * become available for writeback. Otherwise
1917                  * we'll just busyloop.
1918                  */
1919                 trace_writeback_wait(wb, work);
1920                 inode = wb_inode(wb->b_more_io.prev);
1921                 spin_lock(&inode->i_lock);
1922                 spin_unlock(&wb->list_lock);
1923                 /* This function drops i_lock... */
1924                 inode_sleep_on_writeback(inode);
1925                 spin_lock(&wb->list_lock);
1926         }
1927         spin_unlock(&wb->list_lock);
1928         blk_finish_plug(&plug);
1929
1930         return nr_pages - work->nr_pages;
1931 }
1932
1933 /*
1934  * Return the next wb_writeback_work struct that hasn't been processed yet.
1935  */
1936 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1937 {
1938         struct wb_writeback_work *work = NULL;
1939
1940         spin_lock_bh(&wb->work_lock);
1941         if (!list_empty(&wb->work_list)) {
1942                 work = list_entry(wb->work_list.next,
1943                                   struct wb_writeback_work, list);
1944                 list_del_init(&work->list);
1945         }
1946         spin_unlock_bh(&wb->work_lock);
1947         return work;
1948 }
1949
1950 static long wb_check_background_flush(struct bdi_writeback *wb)
1951 {
1952         if (wb_over_bg_thresh(wb)) {
1953
1954                 struct wb_writeback_work work = {
1955                         .nr_pages       = LONG_MAX,
1956                         .sync_mode      = WB_SYNC_NONE,
1957                         .for_background = 1,
1958                         .range_cyclic   = 1,
1959                         .reason         = WB_REASON_BACKGROUND,
1960                 };
1961
1962                 return wb_writeback(wb, &work);
1963         }
1964
1965         return 0;
1966 }
1967
1968 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1969 {
1970         unsigned long expired;
1971         long nr_pages;
1972
1973         /*
1974          * When set to zero, disable periodic writeback
1975          */
1976         if (!dirty_writeback_interval)
1977                 return 0;
1978
1979         expired = wb->last_old_flush +
1980                         msecs_to_jiffies(dirty_writeback_interval * 10);
1981         if (time_before(jiffies, expired))
1982                 return 0;
1983
1984         wb->last_old_flush = jiffies;
1985         nr_pages = get_nr_dirty_pages();
1986
1987         if (nr_pages) {
1988                 struct wb_writeback_work work = {
1989                         .nr_pages       = nr_pages,
1990                         .sync_mode      = WB_SYNC_NONE,
1991                         .for_kupdate    = 1,
1992                         .range_cyclic   = 1,
1993                         .reason         = WB_REASON_PERIODIC,
1994                 };
1995
1996                 return wb_writeback(wb, &work);
1997         }
1998
1999         return 0;
2000 }
2001
2002 static long wb_check_start_all(struct bdi_writeback *wb)
2003 {
2004         long nr_pages;
2005
2006         if (!test_bit(WB_start_all, &wb->state))
2007                 return 0;
2008
2009         nr_pages = get_nr_dirty_pages();
2010         if (nr_pages) {
2011                 struct wb_writeback_work work = {
2012                         .nr_pages       = wb_split_bdi_pages(wb, nr_pages),
2013                         .sync_mode      = WB_SYNC_NONE,
2014                         .range_cyclic   = 1,
2015                         .reason         = wb->start_all_reason,
2016                 };
2017
2018                 nr_pages = wb_writeback(wb, &work);
2019         }
2020
2021         clear_bit(WB_start_all, &wb->state);
2022         return nr_pages;
2023 }
2024
2025
2026 /*
2027  * Retrieve work items and do the writeback they describe
2028  */
2029 static long wb_do_writeback(struct bdi_writeback *wb)
2030 {
2031         struct wb_writeback_work *work;
2032         long wrote = 0;
2033
2034         set_bit(WB_writeback_running, &wb->state);
2035         while ((work = get_next_work_item(wb)) != NULL) {
2036                 trace_writeback_exec(wb, work);
2037                 wrote += wb_writeback(wb, work);
2038                 finish_writeback_work(wb, work);
2039         }
2040
2041         /*
2042          * Check for a flush-everything request
2043          */
2044         wrote += wb_check_start_all(wb);
2045
2046         /*
2047          * Check for periodic writeback, kupdated() style
2048          */
2049         wrote += wb_check_old_data_flush(wb);
2050         wrote += wb_check_background_flush(wb);
2051         clear_bit(WB_writeback_running, &wb->state);
2052
2053         return wrote;
2054 }
2055
2056 /*
2057  * Handle writeback of dirty data for the device backed by this bdi. Also
2058  * reschedules periodically and does kupdated style flushing.
2059  */
2060 void wb_workfn(struct work_struct *work)
2061 {
2062         struct bdi_writeback *wb = container_of(to_delayed_work(work),
2063                                                 struct bdi_writeback, dwork);
2064         long pages_written;
2065
2066         set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2067         current->flags |= PF_SWAPWRITE;
2068
2069         if (likely(!current_is_workqueue_rescuer() ||
2070                    !test_bit(WB_registered, &wb->state))) {
2071                 /*
2072                  * The normal path.  Keep writing back @wb until its
2073                  * work_list is empty.  Note that this path is also taken
2074                  * if @wb is shutting down even when we're running off the
2075                  * rescuer as work_list needs to be drained.
2076                  */
2077                 do {
2078                         pages_written = wb_do_writeback(wb);
2079                         trace_writeback_pages_written(pages_written);
2080                 } while (!list_empty(&wb->work_list));
2081         } else {
2082                 /*
2083                  * bdi_wq can't get enough workers and we're running off
2084                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
2085                  * enough for efficient IO.
2086                  */
2087                 pages_written = writeback_inodes_wb(wb, 1024,
2088                                                     WB_REASON_FORKER_THREAD);
2089                 trace_writeback_pages_written(pages_written);
2090         }
2091
2092         if (!list_empty(&wb->work_list))
2093                 wb_wakeup(wb);
2094         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2095                 wb_wakeup_delayed(wb);
2096
2097         current->flags &= ~PF_SWAPWRITE;
2098 }
2099
2100 /*
2101  * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2102  * write back the whole world.
2103  */
2104 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2105                                          enum wb_reason reason)
2106 {
2107         struct bdi_writeback *wb;
2108
2109         if (!bdi_has_dirty_io(bdi))
2110                 return;
2111
2112         list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2113                 wb_start_writeback(wb, reason);
2114 }
2115
2116 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2117                                 enum wb_reason reason)
2118 {
2119         rcu_read_lock();
2120         __wakeup_flusher_threads_bdi(bdi, reason);
2121         rcu_read_unlock();
2122 }
2123
2124 /*
2125  * Wakeup the flusher threads to start writeback of all currently dirty pages
2126  */
2127 void wakeup_flusher_threads(enum wb_reason reason)
2128 {
2129         struct backing_dev_info *bdi;
2130
2131         /*
2132          * If we are expecting writeback progress we must submit plugged IO.
2133          */
2134         if (blk_needs_flush_plug(current))
2135                 blk_schedule_flush_plug(current);
2136
2137         rcu_read_lock();
2138         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2139                 __wakeup_flusher_threads_bdi(bdi, reason);
2140         rcu_read_unlock();
2141 }
2142
2143 /*
2144  * Wake up bdi's periodically to make sure dirtytime inodes gets
2145  * written back periodically.  We deliberately do *not* check the
2146  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2147  * kernel to be constantly waking up once there are any dirtytime
2148  * inodes on the system.  So instead we define a separate delayed work
2149  * function which gets called much more rarely.  (By default, only
2150  * once every 12 hours.)
2151  *
2152  * If there is any other write activity going on in the file system,
2153  * this function won't be necessary.  But if the only thing that has
2154  * happened on the file system is a dirtytime inode caused by an atime
2155  * update, we need this infrastructure below to make sure that inode
2156  * eventually gets pushed out to disk.
2157  */
2158 static void wakeup_dirtytime_writeback(struct work_struct *w);
2159 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2160
2161 static void wakeup_dirtytime_writeback(struct work_struct *w)
2162 {
2163         struct backing_dev_info *bdi;
2164
2165         rcu_read_lock();
2166         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2167                 struct bdi_writeback *wb;
2168
2169                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2170                         if (!list_empty(&wb->b_dirty_time))
2171                                 wb_wakeup(wb);
2172         }
2173         rcu_read_unlock();
2174         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2175 }
2176
2177 static int __init start_dirtytime_writeback(void)
2178 {
2179         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2180         return 0;
2181 }
2182 __initcall(start_dirtytime_writeback);
2183
2184 int dirtytime_interval_handler(struct ctl_table *table, int write,
2185                                void __user *buffer, size_t *lenp, loff_t *ppos)
2186 {
2187         int ret;
2188
2189         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2190         if (ret == 0 && write)
2191                 mod_delayed_work(system_wq, &dirtytime_work, 0);
2192         return ret;
2193 }
2194
2195 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2196 {
2197         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2198                 struct dentry *dentry;
2199                 const char *name = "?";
2200
2201                 dentry = d_find_alias(inode);
2202                 if (dentry) {
2203                         spin_lock(&dentry->d_lock);
2204                         name = (const char *) dentry->d_name.name;
2205                 }
2206                 printk(KERN_DEBUG
2207                        "%s(%d): dirtied inode %lu (%s) on %s\n",
2208                        current->comm, task_pid_nr(current), inode->i_ino,
2209                        name, inode->i_sb->s_id);
2210                 if (dentry) {
2211                         spin_unlock(&dentry->d_lock);
2212                         dput(dentry);
2213                 }
2214         }
2215 }
2216
2217 /**
2218  * __mark_inode_dirty - internal function
2219  *
2220  * @inode: inode to mark
2221  * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2222  *
2223  * Mark an inode as dirty. Callers should use mark_inode_dirty or
2224  * mark_inode_dirty_sync.
2225  *
2226  * Put the inode on the super block's dirty list.
2227  *
2228  * CAREFUL! We mark it dirty unconditionally, but move it onto the
2229  * dirty list only if it is hashed or if it refers to a blockdev.
2230  * If it was not hashed, it will never be added to the dirty list
2231  * even if it is later hashed, as it will have been marked dirty already.
2232  *
2233  * In short, make sure you hash any inodes _before_ you start marking
2234  * them dirty.
2235  *
2236  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2237  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2238  * the kernel-internal blockdev inode represents the dirtying time of the
2239  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2240  * page->mapping->host, so the page-dirtying time is recorded in the internal
2241  * blockdev inode.
2242  */
2243 void __mark_inode_dirty(struct inode *inode, int flags)
2244 {
2245         struct super_block *sb = inode->i_sb;
2246         int dirtytime;
2247
2248         trace_writeback_mark_inode_dirty(inode, flags);
2249
2250         /*
2251          * Don't do this for I_DIRTY_PAGES - that doesn't actually
2252          * dirty the inode itself
2253          */
2254         if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) {
2255                 trace_writeback_dirty_inode_start(inode, flags);
2256
2257                 if (sb->s_op->dirty_inode)
2258                         sb->s_op->dirty_inode(inode, flags);
2259
2260                 trace_writeback_dirty_inode(inode, flags);
2261         }
2262         if (flags & I_DIRTY_INODE)
2263                 flags &= ~I_DIRTY_TIME;
2264         dirtytime = flags & I_DIRTY_TIME;
2265
2266         /*
2267          * Paired with smp_mb() in __writeback_single_inode() for the
2268          * following lockless i_state test.  See there for details.
2269          */
2270         smp_mb();
2271
2272         if (((inode->i_state & flags) == flags) ||
2273             (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2274                 return;
2275
2276         if (unlikely(block_dump))
2277                 block_dump___mark_inode_dirty(inode);
2278
2279         spin_lock(&inode->i_lock);
2280         if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2281                 goto out_unlock_inode;
2282         if ((inode->i_state & flags) != flags) {
2283                 const int was_dirty = inode->i_state & I_DIRTY;
2284
2285                 inode_attach_wb(inode, NULL);
2286
2287                 if (flags & I_DIRTY_INODE)
2288                         inode->i_state &= ~I_DIRTY_TIME;
2289                 inode->i_state |= flags;
2290
2291                 /*
2292                  * If the inode is being synced, just update its dirty state.
2293                  * The unlocker will place the inode on the appropriate
2294                  * superblock list, based upon its state.
2295                  */
2296                 if (inode->i_state & I_SYNC)
2297                         goto out_unlock_inode;
2298
2299                 /*
2300                  * Only add valid (hashed) inodes to the superblock's
2301                  * dirty list.  Add blockdev inodes as well.
2302                  */
2303                 if (!S_ISBLK(inode->i_mode)) {
2304                         if (inode_unhashed(inode))
2305                                 goto out_unlock_inode;
2306                 }
2307                 if (inode->i_state & I_FREEING)
2308                         goto out_unlock_inode;
2309
2310                 /*
2311                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2312                  * reposition it (that would break b_dirty time-ordering).
2313                  */
2314                 if (!was_dirty) {
2315                         struct bdi_writeback *wb;
2316                         struct list_head *dirty_list;
2317                         bool wakeup_bdi = false;
2318
2319                         wb = locked_inode_to_wb_and_lock_list(inode);
2320
2321                         WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2322                              !test_bit(WB_registered, &wb->state),
2323                              "bdi-%s not registered\n", wb->bdi->name);
2324
2325                         inode->dirtied_when = jiffies;
2326                         if (dirtytime)
2327                                 inode->dirtied_time_when = jiffies;
2328
2329                         if (inode->i_state & I_DIRTY)
2330                                 dirty_list = &wb->b_dirty;
2331                         else
2332                                 dirty_list = &wb->b_dirty_time;
2333
2334                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2335                                                                dirty_list);
2336
2337                         spin_unlock(&wb->list_lock);
2338                         trace_writeback_dirty_inode_enqueue(inode);
2339
2340                         /*
2341                          * If this is the first dirty inode for this bdi,
2342                          * we have to wake-up the corresponding bdi thread
2343                          * to make sure background write-back happens
2344                          * later.
2345                          */
2346                         if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2347                                 wb_wakeup_delayed(wb);
2348                         return;
2349                 }
2350         }
2351 out_unlock_inode:
2352         spin_unlock(&inode->i_lock);
2353 }
2354 EXPORT_SYMBOL(__mark_inode_dirty);
2355
2356 /*
2357  * The @s_sync_lock is used to serialise concurrent sync operations
2358  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2359  * Concurrent callers will block on the s_sync_lock rather than doing contending
2360  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2361  * has been issued up to the time this function is enter is guaranteed to be
2362  * completed by the time we have gained the lock and waited for all IO that is
2363  * in progress regardless of the order callers are granted the lock.
2364  */
2365 static void wait_sb_inodes(struct super_block *sb)
2366 {
2367         LIST_HEAD(sync_list);
2368
2369         /*
2370          * We need to be protected against the filesystem going from
2371          * r/o to r/w or vice versa.
2372          */
2373         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2374
2375         mutex_lock(&sb->s_sync_lock);
2376
2377         /*
2378          * Splice the writeback list onto a temporary list to avoid waiting on
2379          * inodes that have started writeback after this point.
2380          *
2381          * Use rcu_read_lock() to keep the inodes around until we have a
2382          * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2383          * the local list because inodes can be dropped from either by writeback
2384          * completion.
2385          */
2386         rcu_read_lock();
2387         spin_lock_irq(&sb->s_inode_wblist_lock);
2388         list_splice_init(&sb->s_inodes_wb, &sync_list);
2389
2390         /*
2391          * Data integrity sync. Must wait for all pages under writeback, because
2392          * there may have been pages dirtied before our sync call, but which had
2393          * writeout started before we write it out.  In which case, the inode
2394          * may not be on the dirty list, but we still have to wait for that
2395          * writeout.
2396          */
2397         while (!list_empty(&sync_list)) {
2398                 struct inode *inode = list_first_entry(&sync_list, struct inode,
2399                                                        i_wb_list);
2400                 struct address_space *mapping = inode->i_mapping;
2401
2402                 /*
2403                  * Move each inode back to the wb list before we drop the lock
2404                  * to preserve consistency between i_wb_list and the mapping
2405                  * writeback tag. Writeback completion is responsible to remove
2406                  * the inode from either list once the writeback tag is cleared.
2407                  */
2408                 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2409
2410                 /*
2411                  * The mapping can appear untagged while still on-list since we
2412                  * do not have the mapping lock. Skip it here, wb completion
2413                  * will remove it.
2414                  */
2415                 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2416                         continue;
2417
2418                 spin_unlock_irq(&sb->s_inode_wblist_lock);
2419
2420                 spin_lock(&inode->i_lock);
2421                 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2422                         spin_unlock(&inode->i_lock);
2423
2424                         spin_lock_irq(&sb->s_inode_wblist_lock);
2425                         continue;
2426                 }
2427                 __iget(inode);
2428                 spin_unlock(&inode->i_lock);
2429                 rcu_read_unlock();
2430
2431                 /*
2432                  * We keep the error status of individual mapping so that
2433                  * applications can catch the writeback error using fsync(2).
2434                  * See filemap_fdatawait_keep_errors() for details.
2435                  */
2436                 filemap_fdatawait_keep_errors(mapping);
2437
2438                 cond_resched();
2439
2440                 iput(inode);
2441
2442                 rcu_read_lock();
2443                 spin_lock_irq(&sb->s_inode_wblist_lock);
2444         }
2445         spin_unlock_irq(&sb->s_inode_wblist_lock);
2446         rcu_read_unlock();
2447         mutex_unlock(&sb->s_sync_lock);
2448 }
2449
2450 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2451                                      enum wb_reason reason, bool skip_if_busy)
2452 {
2453         struct backing_dev_info *bdi = sb->s_bdi;
2454         DEFINE_WB_COMPLETION(done, bdi);
2455         struct wb_writeback_work work = {
2456                 .sb                     = sb,
2457                 .sync_mode              = WB_SYNC_NONE,
2458                 .tagged_writepages      = 1,
2459                 .done                   = &done,
2460                 .nr_pages               = nr,
2461                 .reason                 = reason,
2462         };
2463
2464         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2465                 return;
2466         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2467
2468         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2469         wb_wait_for_completion(&done);
2470 }
2471
2472 /**
2473  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2474  * @sb: the superblock
2475  * @nr: the number of pages to write
2476  * @reason: reason why some writeback work initiated
2477  *
2478  * Start writeback on some inodes on this super_block. No guarantees are made
2479  * on how many (if any) will be written, and this function does not wait
2480  * for IO completion of submitted IO.
2481  */
2482 void writeback_inodes_sb_nr(struct super_block *sb,
2483                             unsigned long nr,
2484                             enum wb_reason reason)
2485 {
2486         __writeback_inodes_sb_nr(sb, nr, reason, false);
2487 }
2488 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2489
2490 /**
2491  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2492  * @sb: the superblock
2493  * @reason: reason why some writeback work was initiated
2494  *
2495  * Start writeback on some inodes on this super_block. No guarantees are made
2496  * on how many (if any) will be written, and this function does not wait
2497  * for IO completion of submitted IO.
2498  */
2499 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2500 {
2501         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2502 }
2503 EXPORT_SYMBOL(writeback_inodes_sb);
2504
2505 /**
2506  * try_to_writeback_inodes_sb - try to start writeback if none underway
2507  * @sb: the superblock
2508  * @reason: reason why some writeback work was initiated
2509  *
2510  * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2511  */
2512 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2513 {
2514         if (!down_read_trylock(&sb->s_umount))
2515                 return;
2516
2517         __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2518         up_read(&sb->s_umount);
2519 }
2520 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2521
2522 /**
2523  * sync_inodes_sb       -       sync sb inode pages
2524  * @sb: the superblock
2525  *
2526  * This function writes and waits on any dirty inode belonging to this
2527  * super_block.
2528  */
2529 void sync_inodes_sb(struct super_block *sb)
2530 {
2531         struct backing_dev_info *bdi = sb->s_bdi;
2532         DEFINE_WB_COMPLETION(done, bdi);
2533         struct wb_writeback_work work = {
2534                 .sb             = sb,
2535                 .sync_mode      = WB_SYNC_ALL,
2536                 .nr_pages       = LONG_MAX,
2537                 .range_cyclic   = 0,
2538                 .done           = &done,
2539                 .reason         = WB_REASON_SYNC,
2540                 .for_sync       = 1,
2541         };
2542
2543         /*
2544          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2545          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2546          * bdi_has_dirty() need to be written out too.
2547          */
2548         if (bdi == &noop_backing_dev_info)
2549                 return;
2550         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2551
2552         /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2553         bdi_down_write_wb_switch_rwsem(bdi);
2554         bdi_split_work_to_wbs(bdi, &work, false);
2555         wb_wait_for_completion(&done);
2556         bdi_up_write_wb_switch_rwsem(bdi);
2557
2558         wait_sb_inodes(sb);
2559 }
2560 EXPORT_SYMBOL(sync_inodes_sb);
2561
2562 /**
2563  * write_inode_now      -       write an inode to disk
2564  * @inode: inode to write to disk
2565  * @sync: whether the write should be synchronous or not
2566  *
2567  * This function commits an inode to disk immediately if it is dirty. This is
2568  * primarily needed by knfsd.
2569  *
2570  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2571  */
2572 int write_inode_now(struct inode *inode, int sync)
2573 {
2574         struct writeback_control wbc = {
2575                 .nr_to_write = LONG_MAX,
2576                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2577                 .range_start = 0,
2578                 .range_end = LLONG_MAX,
2579         };
2580
2581         if (!mapping_cap_writeback_dirty(inode->i_mapping))
2582                 wbc.nr_to_write = 0;
2583
2584         might_sleep();
2585         return writeback_single_inode(inode, &wbc);
2586 }
2587 EXPORT_SYMBOL(write_inode_now);
2588
2589 /**
2590  * sync_inode - write an inode and its pages to disk.
2591  * @inode: the inode to sync
2592  * @wbc: controls the writeback mode
2593  *
2594  * sync_inode() will write an inode and its pages to disk.  It will also
2595  * correctly update the inode on its superblock's dirty inode lists and will
2596  * update inode->i_state.
2597  *
2598  * The caller must have a ref on the inode.
2599  */
2600 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2601 {
2602         return writeback_single_inode(inode, wbc);
2603 }
2604 EXPORT_SYMBOL(sync_inode);
2605
2606 /**
2607  * sync_inode_metadata - write an inode to disk
2608  * @inode: the inode to sync
2609  * @wait: wait for I/O to complete.
2610  *
2611  * Write an inode to disk and adjust its dirty state after completion.
2612  *
2613  * Note: only writes the actual inode, no associated data or other metadata.
2614  */
2615 int sync_inode_metadata(struct inode *inode, int wait)
2616 {
2617         struct writeback_control wbc = {
2618                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2619                 .nr_to_write = 0, /* metadata-only */
2620         };
2621
2622         return sync_inode(inode, &wbc);
2623 }
2624 EXPORT_SYMBOL(sync_inode_metadata);