Merge tag 'powerpc-6.1-2' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc...
[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         enum writeback_sync_modes sync_mode;
46         unsigned int tagged_writepages:1;
47         unsigned int for_kupdate:1;
48         unsigned int range_cyclic:1;
49         unsigned int for_background:1;
50         unsigned int for_sync:1;        /* sync(2) WB_SYNC_ALL writeback */
51         unsigned int auto_free:1;       /* free on completion */
52         enum wb_reason reason;          /* why was writeback initiated? */
53
54         struct list_head list;          /* pending work list */
55         struct wb_completion *done;     /* set if the caller waits */
56 };
57
58 /*
59  * If an inode is constantly having its pages dirtied, but then the
60  * updates stop dirtytime_expire_interval seconds in the past, it's
61  * possible for the worst case time between when an inode has its
62  * timestamps updated and when they finally get written out to be two
63  * dirtytime_expire_intervals.  We set the default to 12 hours (in
64  * seconds), which means most of the time inodes will have their
65  * timestamps written to disk after 12 hours, but in the worst case a
66  * few inodes might not their timestamps updated for 24 hours.
67  */
68 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
70 static inline struct inode *wb_inode(struct list_head *head)
71 {
72         return list_entry(head, struct inode, i_io_list);
73 }
74
75 /*
76  * Include the creation of the trace points after defining the
77  * wb_writeback_work structure and inline functions so that the definition
78  * remains local to this file.
79  */
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
82
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
85 static bool wb_io_lists_populated(struct bdi_writeback *wb)
86 {
87         if (wb_has_dirty_io(wb)) {
88                 return false;
89         } else {
90                 set_bit(WB_has_dirty_io, &wb->state);
91                 WARN_ON_ONCE(!wb->avg_write_bandwidth);
92                 atomic_long_add(wb->avg_write_bandwidth,
93                                 &wb->bdi->tot_write_bandwidth);
94                 return true;
95         }
96 }
97
98 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99 {
100         if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101             list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102                 clear_bit(WB_has_dirty_io, &wb->state);
103                 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104                                         &wb->bdi->tot_write_bandwidth) < 0);
105         }
106 }
107
108 /**
109  * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110  * @inode: inode to be moved
111  * @wb: target bdi_writeback
112  * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
113  *
114  * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115  * Returns %true if @inode is the first occupant of the !dirty_time IO
116  * lists; otherwise, %false.
117  */
118 static bool inode_io_list_move_locked(struct inode *inode,
119                                       struct bdi_writeback *wb,
120                                       struct list_head *head)
121 {
122         assert_spin_locked(&wb->list_lock);
123         assert_spin_locked(&inode->i_lock);
124
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 static void wb_wakeup(struct bdi_writeback *wb)
136 {
137         spin_lock_irq(&wb->work_lock);
138         if (test_bit(WB_registered, &wb->state))
139                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
140         spin_unlock_irq(&wb->work_lock);
141 }
142
143 static void finish_writeback_work(struct bdi_writeback *wb,
144                                   struct wb_writeback_work *work)
145 {
146         struct wb_completion *done = work->done;
147
148         if (work->auto_free)
149                 kfree(work);
150         if (done) {
151                 wait_queue_head_t *waitq = done->waitq;
152
153                 /* @done can't be accessed after the following dec */
154                 if (atomic_dec_and_test(&done->cnt))
155                         wake_up_all(waitq);
156         }
157 }
158
159 static void wb_queue_work(struct bdi_writeback *wb,
160                           struct wb_writeback_work *work)
161 {
162         trace_writeback_queue(wb, work);
163
164         if (work->done)
165                 atomic_inc(&work->done->cnt);
166
167         spin_lock_irq(&wb->work_lock);
168
169         if (test_bit(WB_registered, &wb->state)) {
170                 list_add_tail(&work->list, &wb->work_list);
171                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
172         } else
173                 finish_writeback_work(wb, work);
174
175         spin_unlock_irq(&wb->work_lock);
176 }
177
178 /**
179  * wb_wait_for_completion - wait for completion of bdi_writeback_works
180  * @done: target wb_completion
181  *
182  * Wait for one or more work items issued to @bdi with their ->done field
183  * set to @done, which should have been initialized with
184  * DEFINE_WB_COMPLETION().  This function returns after all such work items
185  * are completed.  Work items which are waited upon aren't freed
186  * automatically on completion.
187  */
188 void wb_wait_for_completion(struct wb_completion *done)
189 {
190         atomic_dec(&done->cnt);         /* put down the initial count */
191         wait_event(*done->waitq, !atomic_read(&done->cnt));
192 }
193
194 #ifdef CONFIG_CGROUP_WRITEBACK
195
196 /*
197  * Parameters for foreign inode detection, see wbc_detach_inode() to see
198  * how they're used.
199  *
200  * These paramters are inherently heuristical as the detection target
201  * itself is fuzzy.  All we want to do is detaching an inode from the
202  * current owner if it's being written to by some other cgroups too much.
203  *
204  * The current cgroup writeback is built on the assumption that multiple
205  * cgroups writing to the same inode concurrently is very rare and a mode
206  * of operation which isn't well supported.  As such, the goal is not
207  * taking too long when a different cgroup takes over an inode while
208  * avoiding too aggressive flip-flops from occasional foreign writes.
209  *
210  * We record, very roughly, 2s worth of IO time history and if more than
211  * half of that is foreign, trigger the switch.  The recording is quantized
212  * to 16 slots.  To avoid tiny writes from swinging the decision too much,
213  * writes smaller than 1/8 of avg size are ignored.
214  */
215 #define WB_FRN_TIME_SHIFT       13      /* 1s = 2^13, upto 8 secs w/ 16bit */
216 #define WB_FRN_TIME_AVG_SHIFT   3       /* avg = avg * 7/8 + new * 1/8 */
217 #define WB_FRN_TIME_CUT_DIV     8       /* ignore rounds < avg / 8 */
218 #define WB_FRN_TIME_PERIOD      (2 * (1 << WB_FRN_TIME_SHIFT))  /* 2s */
219
220 #define WB_FRN_HIST_SLOTS       16      /* inode->i_wb_frn_history is 16bit */
221 #define WB_FRN_HIST_UNIT        (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
222                                         /* each slot's duration is 2s / 16 */
223 #define WB_FRN_HIST_THR_SLOTS   (WB_FRN_HIST_SLOTS / 2)
224                                         /* if foreign slots >= 8, switch */
225 #define WB_FRN_HIST_MAX_SLOTS   (WB_FRN_HIST_THR_SLOTS / 2 + 1)
226                                         /* one round can affect upto 5 slots */
227 #define WB_FRN_MAX_IN_FLIGHT    1024    /* don't queue too many concurrently */
228
229 /*
230  * Maximum inodes per isw.  A specific value has been chosen to make
231  * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
232  */
233 #define WB_MAX_INODES_PER_ISW  ((1024UL - sizeof(struct inode_switch_wbs_context)) \
234                                 / sizeof(struct inode *))
235
236 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
237 static struct workqueue_struct *isw_wq;
238
239 void __inode_attach_wb(struct inode *inode, struct page *page)
240 {
241         struct backing_dev_info *bdi = inode_to_bdi(inode);
242         struct bdi_writeback *wb = NULL;
243
244         if (inode_cgwb_enabled(inode)) {
245                 struct cgroup_subsys_state *memcg_css;
246
247                 if (page) {
248                         memcg_css = mem_cgroup_css_from_page(page);
249                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
250                 } else {
251                         /* must pin memcg_css, see wb_get_create() */
252                         memcg_css = task_get_css(current, memory_cgrp_id);
253                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254                         css_put(memcg_css);
255                 }
256         }
257
258         if (!wb)
259                 wb = &bdi->wb;
260
261         /*
262          * There may be multiple instances of this function racing to
263          * update the same inode.  Use cmpxchg() to tell the winner.
264          */
265         if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
266                 wb_put(wb);
267 }
268 EXPORT_SYMBOL_GPL(__inode_attach_wb);
269
270 /**
271  * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
272  * @inode: inode of interest with i_lock held
273  * @wb: target bdi_writeback
274  *
275  * Remove the inode from wb's io lists and if necessarily put onto b_attached
276  * list.  Only inodes attached to cgwb's are kept on this list.
277  */
278 static void inode_cgwb_move_to_attached(struct inode *inode,
279                                         struct bdi_writeback *wb)
280 {
281         assert_spin_locked(&wb->list_lock);
282         assert_spin_locked(&inode->i_lock);
283
284         inode->i_state &= ~I_SYNC_QUEUED;
285         if (wb != &wb->bdi->wb)
286                 list_move(&inode->i_io_list, &wb->b_attached);
287         else
288                 list_del_init(&inode->i_io_list);
289         wb_io_lists_depopulated(wb);
290 }
291
292 /**
293  * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
294  * @inode: inode of interest with i_lock held
295  *
296  * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
297  * held on entry and is released on return.  The returned wb is guaranteed
298  * to stay @inode's associated wb until its list_lock is released.
299  */
300 static struct bdi_writeback *
301 locked_inode_to_wb_and_lock_list(struct inode *inode)
302         __releases(&inode->i_lock)
303         __acquires(&wb->list_lock)
304 {
305         while (true) {
306                 struct bdi_writeback *wb = inode_to_wb(inode);
307
308                 /*
309                  * inode_to_wb() association is protected by both
310                  * @inode->i_lock and @wb->list_lock but list_lock nests
311                  * outside i_lock.  Drop i_lock and verify that the
312                  * association hasn't changed after acquiring list_lock.
313                  */
314                 wb_get(wb);
315                 spin_unlock(&inode->i_lock);
316                 spin_lock(&wb->list_lock);
317
318                 /* i_wb may have changed inbetween, can't use inode_to_wb() */
319                 if (likely(wb == inode->i_wb)) {
320                         wb_put(wb);     /* @inode already has ref */
321                         return wb;
322                 }
323
324                 spin_unlock(&wb->list_lock);
325                 wb_put(wb);
326                 cpu_relax();
327                 spin_lock(&inode->i_lock);
328         }
329 }
330
331 /**
332  * inode_to_wb_and_lock_list - determine an inode's wb and lock it
333  * @inode: inode of interest
334  *
335  * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
336  * on entry.
337  */
338 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
339         __acquires(&wb->list_lock)
340 {
341         spin_lock(&inode->i_lock);
342         return locked_inode_to_wb_and_lock_list(inode);
343 }
344
345 struct inode_switch_wbs_context {
346         struct rcu_work         work;
347
348         /*
349          * Multiple inodes can be switched at once.  The switching procedure
350          * consists of two parts, separated by a RCU grace period.  To make
351          * sure that the second part is executed for each inode gone through
352          * the first part, all inode pointers are placed into a NULL-terminated
353          * array embedded into struct inode_switch_wbs_context.  Otherwise
354          * an inode could be left in a non-consistent state.
355          */
356         struct bdi_writeback    *new_wb;
357         struct inode            *inodes[];
358 };
359
360 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
361 {
362         down_write(&bdi->wb_switch_rwsem);
363 }
364
365 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
366 {
367         up_write(&bdi->wb_switch_rwsem);
368 }
369
370 static bool inode_do_switch_wbs(struct inode *inode,
371                                 struct bdi_writeback *old_wb,
372                                 struct bdi_writeback *new_wb)
373 {
374         struct address_space *mapping = inode->i_mapping;
375         XA_STATE(xas, &mapping->i_pages, 0);
376         struct folio *folio;
377         bool switched = false;
378
379         spin_lock(&inode->i_lock);
380         xa_lock_irq(&mapping->i_pages);
381
382         /*
383          * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
384          * path owns the inode and we shouldn't modify ->i_io_list.
385          */
386         if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
387                 goto skip_switch;
388
389         trace_inode_switch_wbs(inode, old_wb, new_wb);
390
391         /*
392          * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
393          * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
394          * folios actually under writeback.
395          */
396         xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
397                 if (folio_test_dirty(folio)) {
398                         long nr = folio_nr_pages(folio);
399                         wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
400                         wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
401                 }
402         }
403
404         xas_set(&xas, 0);
405         xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
406                 long nr = folio_nr_pages(folio);
407                 WARN_ON_ONCE(!folio_test_writeback(folio));
408                 wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
409                 wb_stat_mod(new_wb, WB_WRITEBACK, nr);
410         }
411
412         if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
413                 atomic_dec(&old_wb->writeback_inodes);
414                 atomic_inc(&new_wb->writeback_inodes);
415         }
416
417         wb_get(new_wb);
418
419         /*
420          * Transfer to @new_wb's IO list if necessary.  If the @inode is dirty,
421          * the specific list @inode was on is ignored and the @inode is put on
422          * ->b_dirty which is always correct including from ->b_dirty_time.
423          * The transfer preserves @inode->dirtied_when ordering.  If the @inode
424          * was clean, it means it was on the b_attached list, so move it onto
425          * the b_attached list of @new_wb.
426          */
427         if (!list_empty(&inode->i_io_list)) {
428                 inode->i_wb = new_wb;
429
430                 if (inode->i_state & I_DIRTY_ALL) {
431                         struct inode *pos;
432
433                         list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
434                                 if (time_after_eq(inode->dirtied_when,
435                                                   pos->dirtied_when))
436                                         break;
437                         inode_io_list_move_locked(inode, new_wb,
438                                                   pos->i_io_list.prev);
439                 } else {
440                         inode_cgwb_move_to_attached(inode, new_wb);
441                 }
442         } else {
443                 inode->i_wb = new_wb;
444         }
445
446         /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
447         inode->i_wb_frn_winner = 0;
448         inode->i_wb_frn_avg_time = 0;
449         inode->i_wb_frn_history = 0;
450         switched = true;
451 skip_switch:
452         /*
453          * Paired with load_acquire in unlocked_inode_to_wb_begin() and
454          * ensures that the new wb is visible if they see !I_WB_SWITCH.
455          */
456         smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
457
458         xa_unlock_irq(&mapping->i_pages);
459         spin_unlock(&inode->i_lock);
460
461         return switched;
462 }
463
464 static void inode_switch_wbs_work_fn(struct work_struct *work)
465 {
466         struct inode_switch_wbs_context *isw =
467                 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
468         struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
469         struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
470         struct bdi_writeback *new_wb = isw->new_wb;
471         unsigned long nr_switched = 0;
472         struct inode **inodep;
473
474         /*
475          * If @inode switches cgwb membership while sync_inodes_sb() is
476          * being issued, sync_inodes_sb() might miss it.  Synchronize.
477          */
478         down_read(&bdi->wb_switch_rwsem);
479
480         /*
481          * By the time control reaches here, RCU grace period has passed
482          * since I_WB_SWITCH assertion and all wb stat update transactions
483          * between unlocked_inode_to_wb_begin/end() are guaranteed to be
484          * synchronizing against the i_pages lock.
485          *
486          * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
487          * gives us exclusion against all wb related operations on @inode
488          * including IO list manipulations and stat updates.
489          */
490         if (old_wb < new_wb) {
491                 spin_lock(&old_wb->list_lock);
492                 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
493         } else {
494                 spin_lock(&new_wb->list_lock);
495                 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
496         }
497
498         for (inodep = isw->inodes; *inodep; inodep++) {
499                 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
500                 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
501                         nr_switched++;
502         }
503
504         spin_unlock(&new_wb->list_lock);
505         spin_unlock(&old_wb->list_lock);
506
507         up_read(&bdi->wb_switch_rwsem);
508
509         if (nr_switched) {
510                 wb_wakeup(new_wb);
511                 wb_put_many(old_wb, nr_switched);
512         }
513
514         for (inodep = isw->inodes; *inodep; inodep++)
515                 iput(*inodep);
516         wb_put(new_wb);
517         kfree(isw);
518         atomic_dec(&isw_nr_in_flight);
519 }
520
521 static bool inode_prepare_wbs_switch(struct inode *inode,
522                                      struct bdi_writeback *new_wb)
523 {
524         /*
525          * Paired with smp_mb() in cgroup_writeback_umount().
526          * isw_nr_in_flight must be increased before checking SB_ACTIVE and
527          * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
528          * in cgroup_writeback_umount() and the isw_wq will be not flushed.
529          */
530         smp_mb();
531
532         if (IS_DAX(inode))
533                 return false;
534
535         /* while holding I_WB_SWITCH, no one else can update the association */
536         spin_lock(&inode->i_lock);
537         if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
538             inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
539             inode_to_wb(inode) == new_wb) {
540                 spin_unlock(&inode->i_lock);
541                 return false;
542         }
543         inode->i_state |= I_WB_SWITCH;
544         __iget(inode);
545         spin_unlock(&inode->i_lock);
546
547         return true;
548 }
549
550 /**
551  * inode_switch_wbs - change the wb association of an inode
552  * @inode: target inode
553  * @new_wb_id: ID of the new wb
554  *
555  * Switch @inode's wb association to the wb identified by @new_wb_id.  The
556  * switching is performed asynchronously and may fail silently.
557  */
558 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
559 {
560         struct backing_dev_info *bdi = inode_to_bdi(inode);
561         struct cgroup_subsys_state *memcg_css;
562         struct inode_switch_wbs_context *isw;
563
564         /* noop if seems to be already in progress */
565         if (inode->i_state & I_WB_SWITCH)
566                 return;
567
568         /* avoid queueing a new switch if too many are already in flight */
569         if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
570                 return;
571
572         isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
573         if (!isw)
574                 return;
575
576         atomic_inc(&isw_nr_in_flight);
577
578         /* find and pin the new wb */
579         rcu_read_lock();
580         memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
581         if (memcg_css && !css_tryget(memcg_css))
582                 memcg_css = NULL;
583         rcu_read_unlock();
584         if (!memcg_css)
585                 goto out_free;
586
587         isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
588         css_put(memcg_css);
589         if (!isw->new_wb)
590                 goto out_free;
591
592         if (!inode_prepare_wbs_switch(inode, isw->new_wb))
593                 goto out_free;
594
595         isw->inodes[0] = inode;
596
597         /*
598          * In addition to synchronizing among switchers, I_WB_SWITCH tells
599          * the RCU protected stat update paths to grab the i_page
600          * lock so that stat transfer can synchronize against them.
601          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
602          */
603         INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
604         queue_rcu_work(isw_wq, &isw->work);
605         return;
606
607 out_free:
608         atomic_dec(&isw_nr_in_flight);
609         if (isw->new_wb)
610                 wb_put(isw->new_wb);
611         kfree(isw);
612 }
613
614 /**
615  * cleanup_offline_cgwb - detach associated inodes
616  * @wb: target wb
617  *
618  * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
619  * to eventually release the dying @wb.  Returns %true if not all inodes were
620  * switched and the function has to be restarted.
621  */
622 bool cleanup_offline_cgwb(struct bdi_writeback *wb)
623 {
624         struct cgroup_subsys_state *memcg_css;
625         struct inode_switch_wbs_context *isw;
626         struct inode *inode;
627         int nr;
628         bool restart = false;
629
630         isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
631                       GFP_KERNEL);
632         if (!isw)
633                 return restart;
634
635         atomic_inc(&isw_nr_in_flight);
636
637         for (memcg_css = wb->memcg_css->parent; memcg_css;
638              memcg_css = memcg_css->parent) {
639                 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
640                 if (isw->new_wb)
641                         break;
642         }
643         if (unlikely(!isw->new_wb))
644                 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
645
646         nr = 0;
647         spin_lock(&wb->list_lock);
648         list_for_each_entry(inode, &wb->b_attached, i_io_list) {
649                 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
650                         continue;
651
652                 isw->inodes[nr++] = inode;
653
654                 if (nr >= WB_MAX_INODES_PER_ISW - 1) {
655                         restart = true;
656                         break;
657                 }
658         }
659         spin_unlock(&wb->list_lock);
660
661         /* no attached inodes? bail out */
662         if (nr == 0) {
663                 atomic_dec(&isw_nr_in_flight);
664                 wb_put(isw->new_wb);
665                 kfree(isw);
666                 return restart;
667         }
668
669         /*
670          * In addition to synchronizing among switchers, I_WB_SWITCH tells
671          * the RCU protected stat update paths to grab the i_page
672          * lock so that stat transfer can synchronize against them.
673          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
674          */
675         INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
676         queue_rcu_work(isw_wq, &isw->work);
677
678         return restart;
679 }
680
681 /**
682  * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
683  * @wbc: writeback_control of interest
684  * @inode: target inode
685  *
686  * @inode is locked and about to be written back under the control of @wbc.
687  * Record @inode's writeback context into @wbc and unlock the i_lock.  On
688  * writeback completion, wbc_detach_inode() should be called.  This is used
689  * to track the cgroup writeback context.
690  */
691 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
692                                  struct inode *inode)
693 {
694         if (!inode_cgwb_enabled(inode)) {
695                 spin_unlock(&inode->i_lock);
696                 return;
697         }
698
699         wbc->wb = inode_to_wb(inode);
700         wbc->inode = inode;
701
702         wbc->wb_id = wbc->wb->memcg_css->id;
703         wbc->wb_lcand_id = inode->i_wb_frn_winner;
704         wbc->wb_tcand_id = 0;
705         wbc->wb_bytes = 0;
706         wbc->wb_lcand_bytes = 0;
707         wbc->wb_tcand_bytes = 0;
708
709         wb_get(wbc->wb);
710         spin_unlock(&inode->i_lock);
711
712         /*
713          * A dying wb indicates that either the blkcg associated with the
714          * memcg changed or the associated memcg is dying.  In the first
715          * case, a replacement wb should already be available and we should
716          * refresh the wb immediately.  In the second case, trying to
717          * refresh will keep failing.
718          */
719         if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
720                 inode_switch_wbs(inode, wbc->wb_id);
721 }
722 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
723
724 /**
725  * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
726  * @wbc: writeback_control of the just finished writeback
727  *
728  * To be called after a writeback attempt of an inode finishes and undoes
729  * wbc_attach_and_unlock_inode().  Can be called under any context.
730  *
731  * As concurrent write sharing of an inode is expected to be very rare and
732  * memcg only tracks page ownership on first-use basis severely confining
733  * the usefulness of such sharing, cgroup writeback tracks ownership
734  * per-inode.  While the support for concurrent write sharing of an inode
735  * is deemed unnecessary, an inode being written to by different cgroups at
736  * different points in time is a lot more common, and, more importantly,
737  * charging only by first-use can too readily lead to grossly incorrect
738  * behaviors (single foreign page can lead to gigabytes of writeback to be
739  * incorrectly attributed).
740  *
741  * To resolve this issue, cgroup writeback detects the majority dirtier of
742  * an inode and transfers the ownership to it.  To avoid unnecessary
743  * oscillation, the detection mechanism keeps track of history and gives
744  * out the switch verdict only if the foreign usage pattern is stable over
745  * a certain amount of time and/or writeback attempts.
746  *
747  * On each writeback attempt, @wbc tries to detect the majority writer
748  * using Boyer-Moore majority vote algorithm.  In addition to the byte
749  * count from the majority voting, it also counts the bytes written for the
750  * current wb and the last round's winner wb (max of last round's current
751  * wb, the winner from two rounds ago, and the last round's majority
752  * candidate).  Keeping track of the historical winner helps the algorithm
753  * to semi-reliably detect the most active writer even when it's not the
754  * absolute majority.
755  *
756  * Once the winner of the round is determined, whether the winner is
757  * foreign or not and how much IO time the round consumed is recorded in
758  * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
759  * over a certain threshold, the switch verdict is given.
760  */
761 void wbc_detach_inode(struct writeback_control *wbc)
762 {
763         struct bdi_writeback *wb = wbc->wb;
764         struct inode *inode = wbc->inode;
765         unsigned long avg_time, max_bytes, max_time;
766         u16 history;
767         int max_id;
768
769         if (!wb)
770                 return;
771
772         history = inode->i_wb_frn_history;
773         avg_time = inode->i_wb_frn_avg_time;
774
775         /* pick the winner of this round */
776         if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
777             wbc->wb_bytes >= wbc->wb_tcand_bytes) {
778                 max_id = wbc->wb_id;
779                 max_bytes = wbc->wb_bytes;
780         } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
781                 max_id = wbc->wb_lcand_id;
782                 max_bytes = wbc->wb_lcand_bytes;
783         } else {
784                 max_id = wbc->wb_tcand_id;
785                 max_bytes = wbc->wb_tcand_bytes;
786         }
787
788         /*
789          * Calculate the amount of IO time the winner consumed and fold it
790          * into the running average kept per inode.  If the consumed IO
791          * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
792          * deciding whether to switch or not.  This is to prevent one-off
793          * small dirtiers from skewing the verdict.
794          */
795         max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
796                                 wb->avg_write_bandwidth);
797         if (avg_time)
798                 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
799                             (avg_time >> WB_FRN_TIME_AVG_SHIFT);
800         else
801                 avg_time = max_time;    /* immediate catch up on first run */
802
803         if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
804                 int slots;
805
806                 /*
807                  * The switch verdict is reached if foreign wb's consume
808                  * more than a certain proportion of IO time in a
809                  * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
810                  * history mask where each bit represents one sixteenth of
811                  * the period.  Determine the number of slots to shift into
812                  * history from @max_time.
813                  */
814                 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
815                             (unsigned long)WB_FRN_HIST_MAX_SLOTS);
816                 history <<= slots;
817                 if (wbc->wb_id != max_id)
818                         history |= (1U << slots) - 1;
819
820                 if (history)
821                         trace_inode_foreign_history(inode, wbc, history);
822
823                 /*
824                  * Switch if the current wb isn't the consistent winner.
825                  * If there are multiple closely competing dirtiers, the
826                  * inode may switch across them repeatedly over time, which
827                  * is okay.  The main goal is avoiding keeping an inode on
828                  * the wrong wb for an extended period of time.
829                  */
830                 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
831                         inode_switch_wbs(inode, max_id);
832         }
833
834         /*
835          * Multiple instances of this function may race to update the
836          * following fields but we don't mind occassional inaccuracies.
837          */
838         inode->i_wb_frn_winner = max_id;
839         inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
840         inode->i_wb_frn_history = history;
841
842         wb_put(wbc->wb);
843         wbc->wb = NULL;
844 }
845 EXPORT_SYMBOL_GPL(wbc_detach_inode);
846
847 /**
848  * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
849  * @wbc: writeback_control of the writeback in progress
850  * @page: page being written out
851  * @bytes: number of bytes being written out
852  *
853  * @bytes from @page are about to written out during the writeback
854  * controlled by @wbc.  Keep the book for foreign inode detection.  See
855  * wbc_detach_inode().
856  */
857 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
858                               size_t bytes)
859 {
860         struct cgroup_subsys_state *css;
861         int id;
862
863         /*
864          * pageout() path doesn't attach @wbc to the inode being written
865          * out.  This is intentional as we don't want the function to block
866          * behind a slow cgroup.  Ultimately, we want pageout() to kick off
867          * regular writeback instead of writing things out itself.
868          */
869         if (!wbc->wb || wbc->no_cgroup_owner)
870                 return;
871
872         css = mem_cgroup_css_from_page(page);
873         /* dead cgroups shouldn't contribute to inode ownership arbitration */
874         if (!(css->flags & CSS_ONLINE))
875                 return;
876
877         id = css->id;
878
879         if (id == wbc->wb_id) {
880                 wbc->wb_bytes += bytes;
881                 return;
882         }
883
884         if (id == wbc->wb_lcand_id)
885                 wbc->wb_lcand_bytes += bytes;
886
887         /* Boyer-Moore majority vote algorithm */
888         if (!wbc->wb_tcand_bytes)
889                 wbc->wb_tcand_id = id;
890         if (id == wbc->wb_tcand_id)
891                 wbc->wb_tcand_bytes += bytes;
892         else
893                 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
894 }
895 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
896
897 /**
898  * wb_split_bdi_pages - split nr_pages to write according to bandwidth
899  * @wb: target bdi_writeback to split @nr_pages to
900  * @nr_pages: number of pages to write for the whole bdi
901  *
902  * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
903  * relation to the total write bandwidth of all wb's w/ dirty inodes on
904  * @wb->bdi.
905  */
906 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
907 {
908         unsigned long this_bw = wb->avg_write_bandwidth;
909         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
910
911         if (nr_pages == LONG_MAX)
912                 return LONG_MAX;
913
914         /*
915          * This may be called on clean wb's and proportional distribution
916          * may not make sense, just use the original @nr_pages in those
917          * cases.  In general, we wanna err on the side of writing more.
918          */
919         if (!tot_bw || this_bw >= tot_bw)
920                 return nr_pages;
921         else
922                 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
923 }
924
925 /**
926  * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
927  * @bdi: target backing_dev_info
928  * @base_work: wb_writeback_work to issue
929  * @skip_if_busy: skip wb's which already have writeback in progress
930  *
931  * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
932  * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
933  * distributed to the busy wbs according to each wb's proportion in the
934  * total active write bandwidth of @bdi.
935  */
936 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
937                                   struct wb_writeback_work *base_work,
938                                   bool skip_if_busy)
939 {
940         struct bdi_writeback *last_wb = NULL;
941         struct bdi_writeback *wb = list_entry(&bdi->wb_list,
942                                               struct bdi_writeback, bdi_node);
943
944         might_sleep();
945 restart:
946         rcu_read_lock();
947         list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
948                 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
949                 struct wb_writeback_work fallback_work;
950                 struct wb_writeback_work *work;
951                 long nr_pages;
952
953                 if (last_wb) {
954                         wb_put(last_wb);
955                         last_wb = NULL;
956                 }
957
958                 /* SYNC_ALL writes out I_DIRTY_TIME too */
959                 if (!wb_has_dirty_io(wb) &&
960                     (base_work->sync_mode == WB_SYNC_NONE ||
961                      list_empty(&wb->b_dirty_time)))
962                         continue;
963                 if (skip_if_busy && writeback_in_progress(wb))
964                         continue;
965
966                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
967
968                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
969                 if (work) {
970                         *work = *base_work;
971                         work->nr_pages = nr_pages;
972                         work->auto_free = 1;
973                         wb_queue_work(wb, work);
974                         continue;
975                 }
976
977                 /* alloc failed, execute synchronously using on-stack fallback */
978                 work = &fallback_work;
979                 *work = *base_work;
980                 work->nr_pages = nr_pages;
981                 work->auto_free = 0;
982                 work->done = &fallback_work_done;
983
984                 wb_queue_work(wb, work);
985
986                 /*
987                  * Pin @wb so that it stays on @bdi->wb_list.  This allows
988                  * continuing iteration from @wb after dropping and
989                  * regrabbing rcu read lock.
990                  */
991                 wb_get(wb);
992                 last_wb = wb;
993
994                 rcu_read_unlock();
995                 wb_wait_for_completion(&fallback_work_done);
996                 goto restart;
997         }
998         rcu_read_unlock();
999
1000         if (last_wb)
1001                 wb_put(last_wb);
1002 }
1003
1004 /**
1005  * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1006  * @bdi_id: target bdi id
1007  * @memcg_id: target memcg css id
1008  * @reason: reason why some writeback work initiated
1009  * @done: target wb_completion
1010  *
1011  * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1012  * with the specified parameters.
1013  */
1014 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1015                            enum wb_reason reason, struct wb_completion *done)
1016 {
1017         struct backing_dev_info *bdi;
1018         struct cgroup_subsys_state *memcg_css;
1019         struct bdi_writeback *wb;
1020         struct wb_writeback_work *work;
1021         unsigned long dirty;
1022         int ret;
1023
1024         /* lookup bdi and memcg */
1025         bdi = bdi_get_by_id(bdi_id);
1026         if (!bdi)
1027                 return -ENOENT;
1028
1029         rcu_read_lock();
1030         memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1031         if (memcg_css && !css_tryget(memcg_css))
1032                 memcg_css = NULL;
1033         rcu_read_unlock();
1034         if (!memcg_css) {
1035                 ret = -ENOENT;
1036                 goto out_bdi_put;
1037         }
1038
1039         /*
1040          * And find the associated wb.  If the wb isn't there already
1041          * there's nothing to flush, don't create one.
1042          */
1043         wb = wb_get_lookup(bdi, memcg_css);
1044         if (!wb) {
1045                 ret = -ENOENT;
1046                 goto out_css_put;
1047         }
1048
1049         /*
1050          * The caller is attempting to write out most of
1051          * the currently dirty pages.  Let's take the current dirty page
1052          * count and inflate it by 25% which should be large enough to
1053          * flush out most dirty pages while avoiding getting livelocked by
1054          * concurrent dirtiers.
1055          *
1056          * BTW the memcg stats are flushed periodically and this is best-effort
1057          * estimation, so some potential error is ok.
1058          */
1059         dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1060         dirty = dirty * 10 / 8;
1061
1062         /* issue the writeback work */
1063         work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1064         if (work) {
1065                 work->nr_pages = dirty;
1066                 work->sync_mode = WB_SYNC_NONE;
1067                 work->range_cyclic = 1;
1068                 work->reason = reason;
1069                 work->done = done;
1070                 work->auto_free = 1;
1071                 wb_queue_work(wb, work);
1072                 ret = 0;
1073         } else {
1074                 ret = -ENOMEM;
1075         }
1076
1077         wb_put(wb);
1078 out_css_put:
1079         css_put(memcg_css);
1080 out_bdi_put:
1081         bdi_put(bdi);
1082         return ret;
1083 }
1084
1085 /**
1086  * cgroup_writeback_umount - flush inode wb switches for umount
1087  *
1088  * This function is called when a super_block is about to be destroyed and
1089  * flushes in-flight inode wb switches.  An inode wb switch goes through
1090  * RCU and then workqueue, so the two need to be flushed in order to ensure
1091  * that all previously scheduled switches are finished.  As wb switches are
1092  * rare occurrences and synchronize_rcu() can take a while, perform
1093  * flushing iff wb switches are in flight.
1094  */
1095 void cgroup_writeback_umount(void)
1096 {
1097         /*
1098          * SB_ACTIVE should be reliably cleared before checking
1099          * isw_nr_in_flight, see generic_shutdown_super().
1100          */
1101         smp_mb();
1102
1103         if (atomic_read(&isw_nr_in_flight)) {
1104                 /*
1105                  * Use rcu_barrier() to wait for all pending callbacks to
1106                  * ensure that all in-flight wb switches are in the workqueue.
1107                  */
1108                 rcu_barrier();
1109                 flush_workqueue(isw_wq);
1110         }
1111 }
1112
1113 static int __init cgroup_writeback_init(void)
1114 {
1115         isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1116         if (!isw_wq)
1117                 return -ENOMEM;
1118         return 0;
1119 }
1120 fs_initcall(cgroup_writeback_init);
1121
1122 #else   /* CONFIG_CGROUP_WRITEBACK */
1123
1124 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1125 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1126
1127 static void inode_cgwb_move_to_attached(struct inode *inode,
1128                                         struct bdi_writeback *wb)
1129 {
1130         assert_spin_locked(&wb->list_lock);
1131         assert_spin_locked(&inode->i_lock);
1132
1133         inode->i_state &= ~I_SYNC_QUEUED;
1134         list_del_init(&inode->i_io_list);
1135         wb_io_lists_depopulated(wb);
1136 }
1137
1138 static struct bdi_writeback *
1139 locked_inode_to_wb_and_lock_list(struct inode *inode)
1140         __releases(&inode->i_lock)
1141         __acquires(&wb->list_lock)
1142 {
1143         struct bdi_writeback *wb = inode_to_wb(inode);
1144
1145         spin_unlock(&inode->i_lock);
1146         spin_lock(&wb->list_lock);
1147         return wb;
1148 }
1149
1150 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1151         __acquires(&wb->list_lock)
1152 {
1153         struct bdi_writeback *wb = inode_to_wb(inode);
1154
1155         spin_lock(&wb->list_lock);
1156         return wb;
1157 }
1158
1159 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1160 {
1161         return nr_pages;
1162 }
1163
1164 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1165                                   struct wb_writeback_work *base_work,
1166                                   bool skip_if_busy)
1167 {
1168         might_sleep();
1169
1170         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1171                 base_work->auto_free = 0;
1172                 wb_queue_work(&bdi->wb, base_work);
1173         }
1174 }
1175
1176 #endif  /* CONFIG_CGROUP_WRITEBACK */
1177
1178 /*
1179  * Add in the number of potentially dirty inodes, because each inode
1180  * write can dirty pagecache in the underlying blockdev.
1181  */
1182 static unsigned long get_nr_dirty_pages(void)
1183 {
1184         return global_node_page_state(NR_FILE_DIRTY) +
1185                 get_nr_dirty_inodes();
1186 }
1187
1188 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1189 {
1190         if (!wb_has_dirty_io(wb))
1191                 return;
1192
1193         /*
1194          * All callers of this function want to start writeback of all
1195          * dirty pages. Places like vmscan can call this at a very
1196          * high frequency, causing pointless allocations of tons of
1197          * work items and keeping the flusher threads busy retrieving
1198          * that work. Ensure that we only allow one of them pending and
1199          * inflight at the time.
1200          */
1201         if (test_bit(WB_start_all, &wb->state) ||
1202             test_and_set_bit(WB_start_all, &wb->state))
1203                 return;
1204
1205         wb->start_all_reason = reason;
1206         wb_wakeup(wb);
1207 }
1208
1209 /**
1210  * wb_start_background_writeback - start background writeback
1211  * @wb: bdi_writback to write from
1212  *
1213  * Description:
1214  *   This makes sure WB_SYNC_NONE background writeback happens. When
1215  *   this function returns, it is only guaranteed that for given wb
1216  *   some IO is happening if we are over background dirty threshold.
1217  *   Caller need not hold sb s_umount semaphore.
1218  */
1219 void wb_start_background_writeback(struct bdi_writeback *wb)
1220 {
1221         /*
1222          * We just wake up the flusher thread. It will perform background
1223          * writeback as soon as there is no other work to do.
1224          */
1225         trace_writeback_wake_background(wb);
1226         wb_wakeup(wb);
1227 }
1228
1229 /*
1230  * Remove the inode from the writeback list it is on.
1231  */
1232 void inode_io_list_del(struct inode *inode)
1233 {
1234         struct bdi_writeback *wb;
1235
1236         wb = inode_to_wb_and_lock_list(inode);
1237         spin_lock(&inode->i_lock);
1238
1239         inode->i_state &= ~I_SYNC_QUEUED;
1240         list_del_init(&inode->i_io_list);
1241         wb_io_lists_depopulated(wb);
1242
1243         spin_unlock(&inode->i_lock);
1244         spin_unlock(&wb->list_lock);
1245 }
1246 EXPORT_SYMBOL(inode_io_list_del);
1247
1248 /*
1249  * mark an inode as under writeback on the sb
1250  */
1251 void sb_mark_inode_writeback(struct inode *inode)
1252 {
1253         struct super_block *sb = inode->i_sb;
1254         unsigned long flags;
1255
1256         if (list_empty(&inode->i_wb_list)) {
1257                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1258                 if (list_empty(&inode->i_wb_list)) {
1259                         list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1260                         trace_sb_mark_inode_writeback(inode);
1261                 }
1262                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1263         }
1264 }
1265
1266 /*
1267  * clear an inode as under writeback on the sb
1268  */
1269 void sb_clear_inode_writeback(struct inode *inode)
1270 {
1271         struct super_block *sb = inode->i_sb;
1272         unsigned long flags;
1273
1274         if (!list_empty(&inode->i_wb_list)) {
1275                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1276                 if (!list_empty(&inode->i_wb_list)) {
1277                         list_del_init(&inode->i_wb_list);
1278                         trace_sb_clear_inode_writeback(inode);
1279                 }
1280                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1281         }
1282 }
1283
1284 /*
1285  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1286  * furthest end of its superblock's dirty-inode list.
1287  *
1288  * Before stamping the inode's ->dirtied_when, we check to see whether it is
1289  * already the most-recently-dirtied inode on the b_dirty list.  If that is
1290  * the case then the inode must have been redirtied while it was being written
1291  * out and we don't reset its dirtied_when.
1292  */
1293 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1294 {
1295         assert_spin_locked(&inode->i_lock);
1296
1297         if (!list_empty(&wb->b_dirty)) {
1298                 struct inode *tail;
1299
1300                 tail = wb_inode(wb->b_dirty.next);
1301                 if (time_before(inode->dirtied_when, tail->dirtied_when))
1302                         inode->dirtied_when = jiffies;
1303         }
1304         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1305         inode->i_state &= ~I_SYNC_QUEUED;
1306 }
1307
1308 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1309 {
1310         spin_lock(&inode->i_lock);
1311         redirty_tail_locked(inode, wb);
1312         spin_unlock(&inode->i_lock);
1313 }
1314
1315 /*
1316  * requeue inode for re-scanning after bdi->b_io list is exhausted.
1317  */
1318 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1319 {
1320         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1321 }
1322
1323 static void inode_sync_complete(struct inode *inode)
1324 {
1325         inode->i_state &= ~I_SYNC;
1326         /* If inode is clean an unused, put it into LRU now... */
1327         inode_add_lru(inode);
1328         /* Waiters must see I_SYNC cleared before being woken up */
1329         smp_mb();
1330         wake_up_bit(&inode->i_state, __I_SYNC);
1331 }
1332
1333 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1334 {
1335         bool ret = time_after(inode->dirtied_when, t);
1336 #ifndef CONFIG_64BIT
1337         /*
1338          * For inodes being constantly redirtied, dirtied_when can get stuck.
1339          * It _appears_ to be in the future, but is actually in distant past.
1340          * This test is necessary to prevent such wrapped-around relative times
1341          * from permanently stopping the whole bdi writeback.
1342          */
1343         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1344 #endif
1345         return ret;
1346 }
1347
1348 #define EXPIRE_DIRTY_ATIME 0x0001
1349
1350 /*
1351  * Move expired (dirtied before dirtied_before) dirty inodes from
1352  * @delaying_queue to @dispatch_queue.
1353  */
1354 static int move_expired_inodes(struct list_head *delaying_queue,
1355                                struct list_head *dispatch_queue,
1356                                unsigned long dirtied_before)
1357 {
1358         LIST_HEAD(tmp);
1359         struct list_head *pos, *node;
1360         struct super_block *sb = NULL;
1361         struct inode *inode;
1362         int do_sb_sort = 0;
1363         int moved = 0;
1364
1365         while (!list_empty(delaying_queue)) {
1366                 inode = wb_inode(delaying_queue->prev);
1367                 if (inode_dirtied_after(inode, dirtied_before))
1368                         break;
1369                 spin_lock(&inode->i_lock);
1370                 list_move(&inode->i_io_list, &tmp);
1371                 moved++;
1372                 inode->i_state |= I_SYNC_QUEUED;
1373                 spin_unlock(&inode->i_lock);
1374                 if (sb_is_blkdev_sb(inode->i_sb))
1375                         continue;
1376                 if (sb && sb != inode->i_sb)
1377                         do_sb_sort = 1;
1378                 sb = inode->i_sb;
1379         }
1380
1381         /* just one sb in list, splice to dispatch_queue and we're done */
1382         if (!do_sb_sort) {
1383                 list_splice(&tmp, dispatch_queue);
1384                 goto out;
1385         }
1386
1387         /*
1388          * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1389          * we don't take inode->i_lock here because it is just a pointless overhead.
1390          * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1391          * fully under our control.
1392          */
1393         while (!list_empty(&tmp)) {
1394                 sb = wb_inode(tmp.prev)->i_sb;
1395                 list_for_each_prev_safe(pos, node, &tmp) {
1396                         inode = wb_inode(pos);
1397                         if (inode->i_sb == sb)
1398                                 list_move(&inode->i_io_list, dispatch_queue);
1399                 }
1400         }
1401 out:
1402         return moved;
1403 }
1404
1405 /*
1406  * Queue all expired dirty inodes for io, eldest first.
1407  * Before
1408  *         newly dirtied     b_dirty    b_io    b_more_io
1409  *         =============>    gf         edc     BA
1410  * After
1411  *         newly dirtied     b_dirty    b_io    b_more_io
1412  *         =============>    g          fBAedc
1413  *                                           |
1414  *                                           +--> dequeue for IO
1415  */
1416 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1417                      unsigned long dirtied_before)
1418 {
1419         int moved;
1420         unsigned long time_expire_jif = dirtied_before;
1421
1422         assert_spin_locked(&wb->list_lock);
1423         list_splice_init(&wb->b_more_io, &wb->b_io);
1424         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1425         if (!work->for_sync)
1426                 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1427         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1428                                      time_expire_jif);
1429         if (moved)
1430                 wb_io_lists_populated(wb);
1431         trace_writeback_queue_io(wb, work, dirtied_before, moved);
1432 }
1433
1434 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1435 {
1436         int ret;
1437
1438         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1439                 trace_writeback_write_inode_start(inode, wbc);
1440                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1441                 trace_writeback_write_inode(inode, wbc);
1442                 return ret;
1443         }
1444         return 0;
1445 }
1446
1447 /*
1448  * Wait for writeback on an inode to complete. Called with i_lock held.
1449  * Caller must make sure inode cannot go away when we drop i_lock.
1450  */
1451 static void __inode_wait_for_writeback(struct inode *inode)
1452         __releases(inode->i_lock)
1453         __acquires(inode->i_lock)
1454 {
1455         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1456         wait_queue_head_t *wqh;
1457
1458         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1459         while (inode->i_state & I_SYNC) {
1460                 spin_unlock(&inode->i_lock);
1461                 __wait_on_bit(wqh, &wq, bit_wait,
1462                               TASK_UNINTERRUPTIBLE);
1463                 spin_lock(&inode->i_lock);
1464         }
1465 }
1466
1467 /*
1468  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1469  */
1470 void inode_wait_for_writeback(struct inode *inode)
1471 {
1472         spin_lock(&inode->i_lock);
1473         __inode_wait_for_writeback(inode);
1474         spin_unlock(&inode->i_lock);
1475 }
1476
1477 /*
1478  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1479  * held and drops it. It is aimed for callers not holding any inode reference
1480  * so once i_lock is dropped, inode can go away.
1481  */
1482 static void inode_sleep_on_writeback(struct inode *inode)
1483         __releases(inode->i_lock)
1484 {
1485         DEFINE_WAIT(wait);
1486         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1487         int sleep;
1488
1489         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1490         sleep = inode->i_state & I_SYNC;
1491         spin_unlock(&inode->i_lock);
1492         if (sleep)
1493                 schedule();
1494         finish_wait(wqh, &wait);
1495 }
1496
1497 /*
1498  * Find proper writeback list for the inode depending on its current state and
1499  * possibly also change of its state while we were doing writeback.  Here we
1500  * handle things such as livelock prevention or fairness of writeback among
1501  * inodes. This function can be called only by flusher thread - noone else
1502  * processes all inodes in writeback lists and requeueing inodes behind flusher
1503  * thread's back can have unexpected consequences.
1504  */
1505 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1506                           struct writeback_control *wbc)
1507 {
1508         if (inode->i_state & I_FREEING)
1509                 return;
1510
1511         /*
1512          * Sync livelock prevention. Each inode is tagged and synced in one
1513          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1514          * the dirty time to prevent enqueue and sync it again.
1515          */
1516         if ((inode->i_state & I_DIRTY) &&
1517             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1518                 inode->dirtied_when = jiffies;
1519
1520         if (wbc->pages_skipped) {
1521                 /*
1522                  * writeback is not making progress due to locked
1523                  * buffers. Skip this inode for now.
1524                  */
1525                 redirty_tail_locked(inode, wb);
1526                 return;
1527         }
1528
1529         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1530                 /*
1531                  * We didn't write back all the pages.  nfs_writepages()
1532                  * sometimes bales out without doing anything.
1533                  */
1534                 if (wbc->nr_to_write <= 0) {
1535                         /* Slice used up. Queue for next turn. */
1536                         requeue_io(inode, wb);
1537                 } else {
1538                         /*
1539                          * Writeback blocked by something other than
1540                          * congestion. Delay the inode for some time to
1541                          * avoid spinning on the CPU (100% iowait)
1542                          * retrying writeback of the dirty page/inode
1543                          * that cannot be performed immediately.
1544                          */
1545                         redirty_tail_locked(inode, wb);
1546                 }
1547         } else if (inode->i_state & I_DIRTY) {
1548                 /*
1549                  * Filesystems can dirty the inode during writeback operations,
1550                  * such as delayed allocation during submission or metadata
1551                  * updates after data IO completion.
1552                  */
1553                 redirty_tail_locked(inode, wb);
1554         } else if (inode->i_state & I_DIRTY_TIME) {
1555                 inode->dirtied_when = jiffies;
1556                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1557                 inode->i_state &= ~I_SYNC_QUEUED;
1558         } else {
1559                 /* The inode is clean. Remove from writeback lists. */
1560                 inode_cgwb_move_to_attached(inode, wb);
1561         }
1562 }
1563
1564 /*
1565  * Write out an inode and its dirty pages (or some of its dirty pages, depending
1566  * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1567  *
1568  * This doesn't remove the inode from the writeback list it is on, except
1569  * potentially to move it from b_dirty_time to b_dirty due to timestamp
1570  * expiration.  The caller is otherwise responsible for writeback list handling.
1571  *
1572  * The caller is also responsible for setting the I_SYNC flag beforehand and
1573  * calling inode_sync_complete() to clear it afterwards.
1574  */
1575 static int
1576 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1577 {
1578         struct address_space *mapping = inode->i_mapping;
1579         long nr_to_write = wbc->nr_to_write;
1580         unsigned dirty;
1581         int ret;
1582
1583         WARN_ON(!(inode->i_state & I_SYNC));
1584
1585         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1586
1587         ret = do_writepages(mapping, wbc);
1588
1589         /*
1590          * Make sure to wait on the data before writing out the metadata.
1591          * This is important for filesystems that modify metadata on data
1592          * I/O completion. We don't do it for sync(2) writeback because it has a
1593          * separate, external IO completion path and ->sync_fs for guaranteeing
1594          * inode metadata is written back correctly.
1595          */
1596         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1597                 int err = filemap_fdatawait(mapping);
1598                 if (ret == 0)
1599                         ret = err;
1600         }
1601
1602         /*
1603          * If the inode has dirty timestamps and we need to write them, call
1604          * mark_inode_dirty_sync() to notify the filesystem about it and to
1605          * change I_DIRTY_TIME into I_DIRTY_SYNC.
1606          */
1607         if ((inode->i_state & I_DIRTY_TIME) &&
1608             (wbc->sync_mode == WB_SYNC_ALL ||
1609              time_after(jiffies, inode->dirtied_time_when +
1610                         dirtytime_expire_interval * HZ))) {
1611                 trace_writeback_lazytime(inode);
1612                 mark_inode_dirty_sync(inode);
1613         }
1614
1615         /*
1616          * Get and clear the dirty flags from i_state.  This needs to be done
1617          * after calling writepages because some filesystems may redirty the
1618          * inode during writepages due to delalloc.  It also needs to be done
1619          * after handling timestamp expiration, as that may dirty the inode too.
1620          */
1621         spin_lock(&inode->i_lock);
1622         dirty = inode->i_state & I_DIRTY;
1623         inode->i_state &= ~dirty;
1624
1625         /*
1626          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1627          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1628          * either they see the I_DIRTY bits cleared or we see the dirtied
1629          * inode.
1630          *
1631          * I_DIRTY_PAGES is always cleared together above even if @mapping
1632          * still has dirty pages.  The flag is reinstated after smp_mb() if
1633          * necessary.  This guarantees that either __mark_inode_dirty()
1634          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1635          */
1636         smp_mb();
1637
1638         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1639                 inode->i_state |= I_DIRTY_PAGES;
1640         else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) {
1641                 if (!(inode->i_state & I_DIRTY_PAGES)) {
1642                         inode->i_state &= ~I_PINNING_FSCACHE_WB;
1643                         wbc->unpinned_fscache_wb = true;
1644                         dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */
1645                 }
1646         }
1647
1648         spin_unlock(&inode->i_lock);
1649
1650         /* Don't write the inode if only I_DIRTY_PAGES was set */
1651         if (dirty & ~I_DIRTY_PAGES) {
1652                 int err = write_inode(inode, wbc);
1653                 if (ret == 0)
1654                         ret = err;
1655         }
1656         wbc->unpinned_fscache_wb = false;
1657         trace_writeback_single_inode(inode, wbc, nr_to_write);
1658         return ret;
1659 }
1660
1661 /*
1662  * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1663  * the regular batched writeback done by the flusher threads in
1664  * writeback_sb_inodes().  @wbc controls various aspects of the write, such as
1665  * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1666  *
1667  * To prevent the inode from going away, either the caller must have a reference
1668  * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1669  */
1670 static int writeback_single_inode(struct inode *inode,
1671                                   struct writeback_control *wbc)
1672 {
1673         struct bdi_writeback *wb;
1674         int ret = 0;
1675
1676         spin_lock(&inode->i_lock);
1677         if (!atomic_read(&inode->i_count))
1678                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1679         else
1680                 WARN_ON(inode->i_state & I_WILL_FREE);
1681
1682         if (inode->i_state & I_SYNC) {
1683                 /*
1684                  * Writeback is already running on the inode.  For WB_SYNC_NONE,
1685                  * that's enough and we can just return.  For WB_SYNC_ALL, we
1686                  * must wait for the existing writeback to complete, then do
1687                  * writeback again if there's anything left.
1688                  */
1689                 if (wbc->sync_mode != WB_SYNC_ALL)
1690                         goto out;
1691                 __inode_wait_for_writeback(inode);
1692         }
1693         WARN_ON(inode->i_state & I_SYNC);
1694         /*
1695          * If the inode is already fully clean, then there's nothing to do.
1696          *
1697          * For data-integrity syncs we also need to check whether any pages are
1698          * still under writeback, e.g. due to prior WB_SYNC_NONE writeback.  If
1699          * there are any such pages, we'll need to wait for them.
1700          */
1701         if (!(inode->i_state & I_DIRTY_ALL) &&
1702             (wbc->sync_mode != WB_SYNC_ALL ||
1703              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1704                 goto out;
1705         inode->i_state |= I_SYNC;
1706         wbc_attach_and_unlock_inode(wbc, inode);
1707
1708         ret = __writeback_single_inode(inode, wbc);
1709
1710         wbc_detach_inode(wbc);
1711
1712         wb = inode_to_wb_and_lock_list(inode);
1713         spin_lock(&inode->i_lock);
1714         /*
1715          * If the inode is now fully clean, then it can be safely removed from
1716          * its writeback list (if any).  Otherwise the flusher threads are
1717          * responsible for the writeback lists.
1718          */
1719         if (!(inode->i_state & I_DIRTY_ALL))
1720                 inode_cgwb_move_to_attached(inode, wb);
1721         else if (!(inode->i_state & I_SYNC_QUEUED)) {
1722                 if ((inode->i_state & I_DIRTY))
1723                         redirty_tail_locked(inode, wb);
1724                 else if (inode->i_state & I_DIRTY_TIME) {
1725                         inode->dirtied_when = jiffies;
1726                         inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1727                 }
1728         }
1729
1730         spin_unlock(&wb->list_lock);
1731         inode_sync_complete(inode);
1732 out:
1733         spin_unlock(&inode->i_lock);
1734         return ret;
1735 }
1736
1737 static long writeback_chunk_size(struct bdi_writeback *wb,
1738                                  struct wb_writeback_work *work)
1739 {
1740         long pages;
1741
1742         /*
1743          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1744          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1745          * here avoids calling into writeback_inodes_wb() more than once.
1746          *
1747          * The intended call sequence for WB_SYNC_ALL writeback is:
1748          *
1749          *      wb_writeback()
1750          *          writeback_sb_inodes()       <== called only once
1751          *              write_cache_pages()     <== called once for each inode
1752          *                   (quickly) tag currently dirty pages
1753          *                   (maybe slowly) sync all tagged pages
1754          */
1755         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1756                 pages = LONG_MAX;
1757         else {
1758                 pages = min(wb->avg_write_bandwidth / 2,
1759                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1760                 pages = min(pages, work->nr_pages);
1761                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1762                                    MIN_WRITEBACK_PAGES);
1763         }
1764
1765         return pages;
1766 }
1767
1768 /*
1769  * Write a portion of b_io inodes which belong to @sb.
1770  *
1771  * Return the number of pages and/or inodes written.
1772  *
1773  * NOTE! This is called with wb->list_lock held, and will
1774  * unlock and relock that for each inode it ends up doing
1775  * IO for.
1776  */
1777 static long writeback_sb_inodes(struct super_block *sb,
1778                                 struct bdi_writeback *wb,
1779                                 struct wb_writeback_work *work)
1780 {
1781         struct writeback_control wbc = {
1782                 .sync_mode              = work->sync_mode,
1783                 .tagged_writepages      = work->tagged_writepages,
1784                 .for_kupdate            = work->for_kupdate,
1785                 .for_background         = work->for_background,
1786                 .for_sync               = work->for_sync,
1787                 .range_cyclic           = work->range_cyclic,
1788                 .range_start            = 0,
1789                 .range_end              = LLONG_MAX,
1790         };
1791         unsigned long start_time = jiffies;
1792         long write_chunk;
1793         long total_wrote = 0;  /* count both pages and inodes */
1794
1795         while (!list_empty(&wb->b_io)) {
1796                 struct inode *inode = wb_inode(wb->b_io.prev);
1797                 struct bdi_writeback *tmp_wb;
1798                 long wrote;
1799
1800                 if (inode->i_sb != sb) {
1801                         if (work->sb) {
1802                                 /*
1803                                  * We only want to write back data for this
1804                                  * superblock, move all inodes not belonging
1805                                  * to it back onto the dirty list.
1806                                  */
1807                                 redirty_tail(inode, wb);
1808                                 continue;
1809                         }
1810
1811                         /*
1812                          * The inode belongs to a different superblock.
1813                          * Bounce back to the caller to unpin this and
1814                          * pin the next superblock.
1815                          */
1816                         break;
1817                 }
1818
1819                 /*
1820                  * Don't bother with new inodes or inodes being freed, first
1821                  * kind does not need periodic writeout yet, and for the latter
1822                  * kind writeout is handled by the freer.
1823                  */
1824                 spin_lock(&inode->i_lock);
1825                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1826                         redirty_tail_locked(inode, wb);
1827                         spin_unlock(&inode->i_lock);
1828                         continue;
1829                 }
1830                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1831                         /*
1832                          * If this inode is locked for writeback and we are not
1833                          * doing writeback-for-data-integrity, move it to
1834                          * b_more_io so that writeback can proceed with the
1835                          * other inodes on s_io.
1836                          *
1837                          * We'll have another go at writing back this inode
1838                          * when we completed a full scan of b_io.
1839                          */
1840                         requeue_io(inode, wb);
1841                         spin_unlock(&inode->i_lock);
1842                         trace_writeback_sb_inodes_requeue(inode);
1843                         continue;
1844                 }
1845                 spin_unlock(&wb->list_lock);
1846
1847                 /*
1848                  * We already requeued the inode if it had I_SYNC set and we
1849                  * are doing WB_SYNC_NONE writeback. So this catches only the
1850                  * WB_SYNC_ALL case.
1851                  */
1852                 if (inode->i_state & I_SYNC) {
1853                         /* Wait for I_SYNC. This function drops i_lock... */
1854                         inode_sleep_on_writeback(inode);
1855                         /* Inode may be gone, start again */
1856                         spin_lock(&wb->list_lock);
1857                         continue;
1858                 }
1859                 inode->i_state |= I_SYNC;
1860                 wbc_attach_and_unlock_inode(&wbc, inode);
1861
1862                 write_chunk = writeback_chunk_size(wb, work);
1863                 wbc.nr_to_write = write_chunk;
1864                 wbc.pages_skipped = 0;
1865
1866                 /*
1867                  * We use I_SYNC to pin the inode in memory. While it is set
1868                  * evict_inode() will wait so the inode cannot be freed.
1869                  */
1870                 __writeback_single_inode(inode, &wbc);
1871
1872                 wbc_detach_inode(&wbc);
1873                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1874                 wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1875                 wrote = wrote < 0 ? 0 : wrote;
1876                 total_wrote += wrote;
1877
1878                 if (need_resched()) {
1879                         /*
1880                          * We're trying to balance between building up a nice
1881                          * long list of IOs to improve our merge rate, and
1882                          * getting those IOs out quickly for anyone throttling
1883                          * in balance_dirty_pages().  cond_resched() doesn't
1884                          * unplug, so get our IOs out the door before we
1885                          * give up the CPU.
1886                          */
1887                         blk_flush_plug(current->plug, false);
1888                         cond_resched();
1889                 }
1890
1891                 /*
1892                  * Requeue @inode if still dirty.  Be careful as @inode may
1893                  * have been switched to another wb in the meantime.
1894                  */
1895                 tmp_wb = inode_to_wb_and_lock_list(inode);
1896                 spin_lock(&inode->i_lock);
1897                 if (!(inode->i_state & I_DIRTY_ALL))
1898                         total_wrote++;
1899                 requeue_inode(inode, tmp_wb, &wbc);
1900                 inode_sync_complete(inode);
1901                 spin_unlock(&inode->i_lock);
1902
1903                 if (unlikely(tmp_wb != wb)) {
1904                         spin_unlock(&tmp_wb->list_lock);
1905                         spin_lock(&wb->list_lock);
1906                 }
1907
1908                 /*
1909                  * bail out to wb_writeback() often enough to check
1910                  * background threshold and other termination conditions.
1911                  */
1912                 if (total_wrote) {
1913                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1914                                 break;
1915                         if (work->nr_pages <= 0)
1916                                 break;
1917                 }
1918         }
1919         return total_wrote;
1920 }
1921
1922 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1923                                   struct wb_writeback_work *work)
1924 {
1925         unsigned long start_time = jiffies;
1926         long wrote = 0;
1927
1928         while (!list_empty(&wb->b_io)) {
1929                 struct inode *inode = wb_inode(wb->b_io.prev);
1930                 struct super_block *sb = inode->i_sb;
1931
1932                 if (!trylock_super(sb)) {
1933                         /*
1934                          * trylock_super() may fail consistently due to
1935                          * s_umount being grabbed by someone else. Don't use
1936                          * requeue_io() to avoid busy retrying the inode/sb.
1937                          */
1938                         redirty_tail(inode, wb);
1939                         continue;
1940                 }
1941                 wrote += writeback_sb_inodes(sb, wb, work);
1942                 up_read(&sb->s_umount);
1943
1944                 /* refer to the same tests at the end of writeback_sb_inodes */
1945                 if (wrote) {
1946                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1947                                 break;
1948                         if (work->nr_pages <= 0)
1949                                 break;
1950                 }
1951         }
1952         /* Leave any unwritten inodes on b_io */
1953         return wrote;
1954 }
1955
1956 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1957                                 enum wb_reason reason)
1958 {
1959         struct wb_writeback_work work = {
1960                 .nr_pages       = nr_pages,
1961                 .sync_mode      = WB_SYNC_NONE,
1962                 .range_cyclic   = 1,
1963                 .reason         = reason,
1964         };
1965         struct blk_plug plug;
1966
1967         blk_start_plug(&plug);
1968         spin_lock(&wb->list_lock);
1969         if (list_empty(&wb->b_io))
1970                 queue_io(wb, &work, jiffies);
1971         __writeback_inodes_wb(wb, &work);
1972         spin_unlock(&wb->list_lock);
1973         blk_finish_plug(&plug);
1974
1975         return nr_pages - work.nr_pages;
1976 }
1977
1978 /*
1979  * Explicit flushing or periodic writeback of "old" data.
1980  *
1981  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1982  * dirtying-time in the inode's address_space.  So this periodic writeback code
1983  * just walks the superblock inode list, writing back any inodes which are
1984  * older than a specific point in time.
1985  *
1986  * Try to run once per dirty_writeback_interval.  But if a writeback event
1987  * takes longer than a dirty_writeback_interval interval, then leave a
1988  * one-second gap.
1989  *
1990  * dirtied_before takes precedence over nr_to_write.  So we'll only write back
1991  * all dirty pages if they are all attached to "old" mappings.
1992  */
1993 static long wb_writeback(struct bdi_writeback *wb,
1994                          struct wb_writeback_work *work)
1995 {
1996         long nr_pages = work->nr_pages;
1997         unsigned long dirtied_before = jiffies;
1998         struct inode *inode;
1999         long progress;
2000         struct blk_plug plug;
2001
2002         blk_start_plug(&plug);
2003         spin_lock(&wb->list_lock);
2004         for (;;) {
2005                 /*
2006                  * Stop writeback when nr_pages has been consumed
2007                  */
2008                 if (work->nr_pages <= 0)
2009                         break;
2010
2011                 /*
2012                  * Background writeout and kupdate-style writeback may
2013                  * run forever. Stop them if there is other work to do
2014                  * so that e.g. sync can proceed. They'll be restarted
2015                  * after the other works are all done.
2016                  */
2017                 if ((work->for_background || work->for_kupdate) &&
2018                     !list_empty(&wb->work_list))
2019                         break;
2020
2021                 /*
2022                  * For background writeout, stop when we are below the
2023                  * background dirty threshold
2024                  */
2025                 if (work->for_background && !wb_over_bg_thresh(wb))
2026                         break;
2027
2028                 /*
2029                  * Kupdate and background works are special and we want to
2030                  * include all inodes that need writing. Livelock avoidance is
2031                  * handled by these works yielding to any other work so we are
2032                  * safe.
2033                  */
2034                 if (work->for_kupdate) {
2035                         dirtied_before = jiffies -
2036                                 msecs_to_jiffies(dirty_expire_interval * 10);
2037                 } else if (work->for_background)
2038                         dirtied_before = jiffies;
2039
2040                 trace_writeback_start(wb, work);
2041                 if (list_empty(&wb->b_io))
2042                         queue_io(wb, work, dirtied_before);
2043                 if (work->sb)
2044                         progress = writeback_sb_inodes(work->sb, wb, work);
2045                 else
2046                         progress = __writeback_inodes_wb(wb, work);
2047                 trace_writeback_written(wb, work);
2048
2049                 /*
2050                  * Did we write something? Try for more
2051                  *
2052                  * Dirty inodes are moved to b_io for writeback in batches.
2053                  * The completion of the current batch does not necessarily
2054                  * mean the overall work is done. So we keep looping as long
2055                  * as made some progress on cleaning pages or inodes.
2056                  */
2057                 if (progress)
2058                         continue;
2059                 /*
2060                  * No more inodes for IO, bail
2061                  */
2062                 if (list_empty(&wb->b_more_io))
2063                         break;
2064                 /*
2065                  * Nothing written. Wait for some inode to
2066                  * become available for writeback. Otherwise
2067                  * we'll just busyloop.
2068                  */
2069                 trace_writeback_wait(wb, work);
2070                 inode = wb_inode(wb->b_more_io.prev);
2071                 spin_lock(&inode->i_lock);
2072                 spin_unlock(&wb->list_lock);
2073                 /* This function drops i_lock... */
2074                 inode_sleep_on_writeback(inode);
2075                 spin_lock(&wb->list_lock);
2076         }
2077         spin_unlock(&wb->list_lock);
2078         blk_finish_plug(&plug);
2079
2080         return nr_pages - work->nr_pages;
2081 }
2082
2083 /*
2084  * Return the next wb_writeback_work struct that hasn't been processed yet.
2085  */
2086 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2087 {
2088         struct wb_writeback_work *work = NULL;
2089
2090         spin_lock_irq(&wb->work_lock);
2091         if (!list_empty(&wb->work_list)) {
2092                 work = list_entry(wb->work_list.next,
2093                                   struct wb_writeback_work, list);
2094                 list_del_init(&work->list);
2095         }
2096         spin_unlock_irq(&wb->work_lock);
2097         return work;
2098 }
2099
2100 static long wb_check_background_flush(struct bdi_writeback *wb)
2101 {
2102         if (wb_over_bg_thresh(wb)) {
2103
2104                 struct wb_writeback_work work = {
2105                         .nr_pages       = LONG_MAX,
2106                         .sync_mode      = WB_SYNC_NONE,
2107                         .for_background = 1,
2108                         .range_cyclic   = 1,
2109                         .reason         = WB_REASON_BACKGROUND,
2110                 };
2111
2112                 return wb_writeback(wb, &work);
2113         }
2114
2115         return 0;
2116 }
2117
2118 static long wb_check_old_data_flush(struct bdi_writeback *wb)
2119 {
2120         unsigned long expired;
2121         long nr_pages;
2122
2123         /*
2124          * When set to zero, disable periodic writeback
2125          */
2126         if (!dirty_writeback_interval)
2127                 return 0;
2128
2129         expired = wb->last_old_flush +
2130                         msecs_to_jiffies(dirty_writeback_interval * 10);
2131         if (time_before(jiffies, expired))
2132                 return 0;
2133
2134         wb->last_old_flush = jiffies;
2135         nr_pages = get_nr_dirty_pages();
2136
2137         if (nr_pages) {
2138                 struct wb_writeback_work work = {
2139                         .nr_pages       = nr_pages,
2140                         .sync_mode      = WB_SYNC_NONE,
2141                         .for_kupdate    = 1,
2142                         .range_cyclic   = 1,
2143                         .reason         = WB_REASON_PERIODIC,
2144                 };
2145
2146                 return wb_writeback(wb, &work);
2147         }
2148
2149         return 0;
2150 }
2151
2152 static long wb_check_start_all(struct bdi_writeback *wb)
2153 {
2154         long nr_pages;
2155
2156         if (!test_bit(WB_start_all, &wb->state))
2157                 return 0;
2158
2159         nr_pages = get_nr_dirty_pages();
2160         if (nr_pages) {
2161                 struct wb_writeback_work work = {
2162                         .nr_pages       = wb_split_bdi_pages(wb, nr_pages),
2163                         .sync_mode      = WB_SYNC_NONE,
2164                         .range_cyclic   = 1,
2165                         .reason         = wb->start_all_reason,
2166                 };
2167
2168                 nr_pages = wb_writeback(wb, &work);
2169         }
2170
2171         clear_bit(WB_start_all, &wb->state);
2172         return nr_pages;
2173 }
2174
2175
2176 /*
2177  * Retrieve work items and do the writeback they describe
2178  */
2179 static long wb_do_writeback(struct bdi_writeback *wb)
2180 {
2181         struct wb_writeback_work *work;
2182         long wrote = 0;
2183
2184         set_bit(WB_writeback_running, &wb->state);
2185         while ((work = get_next_work_item(wb)) != NULL) {
2186                 trace_writeback_exec(wb, work);
2187                 wrote += wb_writeback(wb, work);
2188                 finish_writeback_work(wb, work);
2189         }
2190
2191         /*
2192          * Check for a flush-everything request
2193          */
2194         wrote += wb_check_start_all(wb);
2195
2196         /*
2197          * Check for periodic writeback, kupdated() style
2198          */
2199         wrote += wb_check_old_data_flush(wb);
2200         wrote += wb_check_background_flush(wb);
2201         clear_bit(WB_writeback_running, &wb->state);
2202
2203         return wrote;
2204 }
2205
2206 /*
2207  * Handle writeback of dirty data for the device backed by this bdi. Also
2208  * reschedules periodically and does kupdated style flushing.
2209  */
2210 void wb_workfn(struct work_struct *work)
2211 {
2212         struct bdi_writeback *wb = container_of(to_delayed_work(work),
2213                                                 struct bdi_writeback, dwork);
2214         long pages_written;
2215
2216         set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2217
2218         if (likely(!current_is_workqueue_rescuer() ||
2219                    !test_bit(WB_registered, &wb->state))) {
2220                 /*
2221                  * The normal path.  Keep writing back @wb until its
2222                  * work_list is empty.  Note that this path is also taken
2223                  * if @wb is shutting down even when we're running off the
2224                  * rescuer as work_list needs to be drained.
2225                  */
2226                 do {
2227                         pages_written = wb_do_writeback(wb);
2228                         trace_writeback_pages_written(pages_written);
2229                 } while (!list_empty(&wb->work_list));
2230         } else {
2231                 /*
2232                  * bdi_wq can't get enough workers and we're running off
2233                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
2234                  * enough for efficient IO.
2235                  */
2236                 pages_written = writeback_inodes_wb(wb, 1024,
2237                                                     WB_REASON_FORKER_THREAD);
2238                 trace_writeback_pages_written(pages_written);
2239         }
2240
2241         if (!list_empty(&wb->work_list))
2242                 wb_wakeup(wb);
2243         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2244                 wb_wakeup_delayed(wb);
2245 }
2246
2247 /*
2248  * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2249  * write back the whole world.
2250  */
2251 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2252                                          enum wb_reason reason)
2253 {
2254         struct bdi_writeback *wb;
2255
2256         if (!bdi_has_dirty_io(bdi))
2257                 return;
2258
2259         list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2260                 wb_start_writeback(wb, reason);
2261 }
2262
2263 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2264                                 enum wb_reason reason)
2265 {
2266         rcu_read_lock();
2267         __wakeup_flusher_threads_bdi(bdi, reason);
2268         rcu_read_unlock();
2269 }
2270
2271 /*
2272  * Wakeup the flusher threads to start writeback of all currently dirty pages
2273  */
2274 void wakeup_flusher_threads(enum wb_reason reason)
2275 {
2276         struct backing_dev_info *bdi;
2277
2278         /*
2279          * If we are expecting writeback progress we must submit plugged IO.
2280          */
2281         blk_flush_plug(current->plug, true);
2282
2283         rcu_read_lock();
2284         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2285                 __wakeup_flusher_threads_bdi(bdi, reason);
2286         rcu_read_unlock();
2287 }
2288
2289 /*
2290  * Wake up bdi's periodically to make sure dirtytime inodes gets
2291  * written back periodically.  We deliberately do *not* check the
2292  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2293  * kernel to be constantly waking up once there are any dirtytime
2294  * inodes on the system.  So instead we define a separate delayed work
2295  * function which gets called much more rarely.  (By default, only
2296  * once every 12 hours.)
2297  *
2298  * If there is any other write activity going on in the file system,
2299  * this function won't be necessary.  But if the only thing that has
2300  * happened on the file system is a dirtytime inode caused by an atime
2301  * update, we need this infrastructure below to make sure that inode
2302  * eventually gets pushed out to disk.
2303  */
2304 static void wakeup_dirtytime_writeback(struct work_struct *w);
2305 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2306
2307 static void wakeup_dirtytime_writeback(struct work_struct *w)
2308 {
2309         struct backing_dev_info *bdi;
2310
2311         rcu_read_lock();
2312         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2313                 struct bdi_writeback *wb;
2314
2315                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2316                         if (!list_empty(&wb->b_dirty_time))
2317                                 wb_wakeup(wb);
2318         }
2319         rcu_read_unlock();
2320         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2321 }
2322
2323 static int __init start_dirtytime_writeback(void)
2324 {
2325         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2326         return 0;
2327 }
2328 __initcall(start_dirtytime_writeback);
2329
2330 int dirtytime_interval_handler(struct ctl_table *table, int write,
2331                                void *buffer, size_t *lenp, loff_t *ppos)
2332 {
2333         int ret;
2334
2335         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2336         if (ret == 0 && write)
2337                 mod_delayed_work(system_wq, &dirtytime_work, 0);
2338         return ret;
2339 }
2340
2341 /**
2342  * __mark_inode_dirty - internal function to mark an inode dirty
2343  *
2344  * @inode: inode to mark
2345  * @flags: what kind of dirty, e.g. I_DIRTY_SYNC.  This can be a combination of
2346  *         multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2347  *         with I_DIRTY_PAGES.
2348  *
2349  * Mark an inode as dirty.  We notify the filesystem, then update the inode's
2350  * dirty flags.  Then, if needed we add the inode to the appropriate dirty list.
2351  *
2352  * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2353  * instead of calling this directly.
2354  *
2355  * CAREFUL!  We only add the inode to the dirty list if it is hashed or if it
2356  * refers to a blockdev.  Unhashed inodes will never be added to the dirty list
2357  * even if they are later hashed, as they will have been marked dirty already.
2358  *
2359  * In short, ensure you hash any inodes _before_ you start marking them dirty.
2360  *
2361  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2362  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2363  * the kernel-internal blockdev inode represents the dirtying time of the
2364  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2365  * page->mapping->host, so the page-dirtying time is recorded in the internal
2366  * blockdev inode.
2367  */
2368 void __mark_inode_dirty(struct inode *inode, int flags)
2369 {
2370         struct super_block *sb = inode->i_sb;
2371         int dirtytime = 0;
2372         struct bdi_writeback *wb = NULL;
2373
2374         trace_writeback_mark_inode_dirty(inode, flags);
2375
2376         if (flags & I_DIRTY_INODE) {
2377                 /*
2378                  * Inode timestamp update will piggback on this dirtying.
2379                  * We tell ->dirty_inode callback that timestamps need to
2380                  * be updated by setting I_DIRTY_TIME in flags.
2381                  */
2382                 if (inode->i_state & I_DIRTY_TIME) {
2383                         spin_lock(&inode->i_lock);
2384                         if (inode->i_state & I_DIRTY_TIME) {
2385                                 inode->i_state &= ~I_DIRTY_TIME;
2386                                 flags |= I_DIRTY_TIME;
2387                         }
2388                         spin_unlock(&inode->i_lock);
2389                 }
2390
2391                 /*
2392                  * Notify the filesystem about the inode being dirtied, so that
2393                  * (if needed) it can update on-disk fields and journal the
2394                  * inode.  This is only needed when the inode itself is being
2395                  * dirtied now.  I.e. it's only needed for I_DIRTY_INODE, not
2396                  * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2397                  */
2398                 trace_writeback_dirty_inode_start(inode, flags);
2399                 if (sb->s_op->dirty_inode)
2400                         sb->s_op->dirty_inode(inode,
2401                                 flags & (I_DIRTY_INODE | I_DIRTY_TIME));
2402                 trace_writeback_dirty_inode(inode, flags);
2403
2404                 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2405                 flags &= ~I_DIRTY_TIME;
2406         } else {
2407                 /*
2408                  * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2409                  * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2410                  * in one call to __mark_inode_dirty().)
2411                  */
2412                 dirtytime = flags & I_DIRTY_TIME;
2413                 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2414         }
2415
2416         /*
2417          * Paired with smp_mb() in __writeback_single_inode() for the
2418          * following lockless i_state test.  See there for details.
2419          */
2420         smp_mb();
2421
2422         if ((inode->i_state & flags) == flags)
2423                 return;
2424
2425         spin_lock(&inode->i_lock);
2426         if ((inode->i_state & flags) != flags) {
2427                 const int was_dirty = inode->i_state & I_DIRTY;
2428
2429                 inode_attach_wb(inode, NULL);
2430
2431                 inode->i_state |= flags;
2432
2433                 /*
2434                  * Grab inode's wb early because it requires dropping i_lock and we
2435                  * need to make sure following checks happen atomically with dirty
2436                  * list handling so that we don't move inodes under flush worker's
2437                  * hands.
2438                  */
2439                 if (!was_dirty) {
2440                         wb = locked_inode_to_wb_and_lock_list(inode);
2441                         spin_lock(&inode->i_lock);
2442                 }
2443
2444                 /*
2445                  * If the inode is queued for writeback by flush worker, just
2446                  * update its dirty state. Once the flush worker is done with
2447                  * the inode it will place it on the appropriate superblock
2448                  * list, based upon its state.
2449                  */
2450                 if (inode->i_state & I_SYNC_QUEUED)
2451                         goto out_unlock;
2452
2453                 /*
2454                  * Only add valid (hashed) inodes to the superblock's
2455                  * dirty list.  Add blockdev inodes as well.
2456                  */
2457                 if (!S_ISBLK(inode->i_mode)) {
2458                         if (inode_unhashed(inode))
2459                                 goto out_unlock;
2460                 }
2461                 if (inode->i_state & I_FREEING)
2462                         goto out_unlock;
2463
2464                 /*
2465                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2466                  * reposition it (that would break b_dirty time-ordering).
2467                  */
2468                 if (!was_dirty) {
2469                         struct list_head *dirty_list;
2470                         bool wakeup_bdi = false;
2471
2472                         inode->dirtied_when = jiffies;
2473                         if (dirtytime)
2474                                 inode->dirtied_time_when = jiffies;
2475
2476                         if (inode->i_state & I_DIRTY)
2477                                 dirty_list = &wb->b_dirty;
2478                         else
2479                                 dirty_list = &wb->b_dirty_time;
2480
2481                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2482                                                                dirty_list);
2483
2484                         spin_unlock(&wb->list_lock);
2485                         spin_unlock(&inode->i_lock);
2486                         trace_writeback_dirty_inode_enqueue(inode);
2487
2488                         /*
2489                          * If this is the first dirty inode for this bdi,
2490                          * we have to wake-up the corresponding bdi thread
2491                          * to make sure background write-back happens
2492                          * later.
2493                          */
2494                         if (wakeup_bdi &&
2495                             (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2496                                 wb_wakeup_delayed(wb);
2497                         return;
2498                 }
2499         }
2500 out_unlock:
2501         if (wb)
2502                 spin_unlock(&wb->list_lock);
2503         spin_unlock(&inode->i_lock);
2504 }
2505 EXPORT_SYMBOL(__mark_inode_dirty);
2506
2507 /*
2508  * The @s_sync_lock is used to serialise concurrent sync operations
2509  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2510  * Concurrent callers will block on the s_sync_lock rather than doing contending
2511  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2512  * has been issued up to the time this function is enter is guaranteed to be
2513  * completed by the time we have gained the lock and waited for all IO that is
2514  * in progress regardless of the order callers are granted the lock.
2515  */
2516 static void wait_sb_inodes(struct super_block *sb)
2517 {
2518         LIST_HEAD(sync_list);
2519
2520         /*
2521          * We need to be protected against the filesystem going from
2522          * r/o to r/w or vice versa.
2523          */
2524         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2525
2526         mutex_lock(&sb->s_sync_lock);
2527
2528         /*
2529          * Splice the writeback list onto a temporary list to avoid waiting on
2530          * inodes that have started writeback after this point.
2531          *
2532          * Use rcu_read_lock() to keep the inodes around until we have a
2533          * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2534          * the local list because inodes can be dropped from either by writeback
2535          * completion.
2536          */
2537         rcu_read_lock();
2538         spin_lock_irq(&sb->s_inode_wblist_lock);
2539         list_splice_init(&sb->s_inodes_wb, &sync_list);
2540
2541         /*
2542          * Data integrity sync. Must wait for all pages under writeback, because
2543          * there may have been pages dirtied before our sync call, but which had
2544          * writeout started before we write it out.  In which case, the inode
2545          * may not be on the dirty list, but we still have to wait for that
2546          * writeout.
2547          */
2548         while (!list_empty(&sync_list)) {
2549                 struct inode *inode = list_first_entry(&sync_list, struct inode,
2550                                                        i_wb_list);
2551                 struct address_space *mapping = inode->i_mapping;
2552
2553                 /*
2554                  * Move each inode back to the wb list before we drop the lock
2555                  * to preserve consistency between i_wb_list and the mapping
2556                  * writeback tag. Writeback completion is responsible to remove
2557                  * the inode from either list once the writeback tag is cleared.
2558                  */
2559                 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2560
2561                 /*
2562                  * The mapping can appear untagged while still on-list since we
2563                  * do not have the mapping lock. Skip it here, wb completion
2564                  * will remove it.
2565                  */
2566                 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2567                         continue;
2568
2569                 spin_unlock_irq(&sb->s_inode_wblist_lock);
2570
2571                 spin_lock(&inode->i_lock);
2572                 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2573                         spin_unlock(&inode->i_lock);
2574
2575                         spin_lock_irq(&sb->s_inode_wblist_lock);
2576                         continue;
2577                 }
2578                 __iget(inode);
2579                 spin_unlock(&inode->i_lock);
2580                 rcu_read_unlock();
2581
2582                 /*
2583                  * We keep the error status of individual mapping so that
2584                  * applications can catch the writeback error using fsync(2).
2585                  * See filemap_fdatawait_keep_errors() for details.
2586                  */
2587                 filemap_fdatawait_keep_errors(mapping);
2588
2589                 cond_resched();
2590
2591                 iput(inode);
2592
2593                 rcu_read_lock();
2594                 spin_lock_irq(&sb->s_inode_wblist_lock);
2595         }
2596         spin_unlock_irq(&sb->s_inode_wblist_lock);
2597         rcu_read_unlock();
2598         mutex_unlock(&sb->s_sync_lock);
2599 }
2600
2601 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2602                                      enum wb_reason reason, bool skip_if_busy)
2603 {
2604         struct backing_dev_info *bdi = sb->s_bdi;
2605         DEFINE_WB_COMPLETION(done, bdi);
2606         struct wb_writeback_work work = {
2607                 .sb                     = sb,
2608                 .sync_mode              = WB_SYNC_NONE,
2609                 .tagged_writepages      = 1,
2610                 .done                   = &done,
2611                 .nr_pages               = nr,
2612                 .reason                 = reason,
2613         };
2614
2615         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2616                 return;
2617         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2618
2619         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2620         wb_wait_for_completion(&done);
2621 }
2622
2623 /**
2624  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2625  * @sb: the superblock
2626  * @nr: the number of pages to write
2627  * @reason: reason why some writeback work initiated
2628  *
2629  * Start writeback on some inodes on this super_block. No guarantees are made
2630  * on how many (if any) will be written, and this function does not wait
2631  * for IO completion of submitted IO.
2632  */
2633 void writeback_inodes_sb_nr(struct super_block *sb,
2634                             unsigned long nr,
2635                             enum wb_reason reason)
2636 {
2637         __writeback_inodes_sb_nr(sb, nr, reason, false);
2638 }
2639 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2640
2641 /**
2642  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2643  * @sb: the superblock
2644  * @reason: reason why some writeback work was initiated
2645  *
2646  * Start writeback on some inodes on this super_block. No guarantees are made
2647  * on how many (if any) will be written, and this function does not wait
2648  * for IO completion of submitted IO.
2649  */
2650 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2651 {
2652         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2653 }
2654 EXPORT_SYMBOL(writeback_inodes_sb);
2655
2656 /**
2657  * try_to_writeback_inodes_sb - try to start writeback if none underway
2658  * @sb: the superblock
2659  * @reason: reason why some writeback work was initiated
2660  *
2661  * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2662  */
2663 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2664 {
2665         if (!down_read_trylock(&sb->s_umount))
2666                 return;
2667
2668         __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2669         up_read(&sb->s_umount);
2670 }
2671 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2672
2673 /**
2674  * sync_inodes_sb       -       sync sb inode pages
2675  * @sb: the superblock
2676  *
2677  * This function writes and waits on any dirty inode belonging to this
2678  * super_block.
2679  */
2680 void sync_inodes_sb(struct super_block *sb)
2681 {
2682         struct backing_dev_info *bdi = sb->s_bdi;
2683         DEFINE_WB_COMPLETION(done, bdi);
2684         struct wb_writeback_work work = {
2685                 .sb             = sb,
2686                 .sync_mode      = WB_SYNC_ALL,
2687                 .nr_pages       = LONG_MAX,
2688                 .range_cyclic   = 0,
2689                 .done           = &done,
2690                 .reason         = WB_REASON_SYNC,
2691                 .for_sync       = 1,
2692         };
2693
2694         /*
2695          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2696          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2697          * bdi_has_dirty() need to be written out too.
2698          */
2699         if (bdi == &noop_backing_dev_info)
2700                 return;
2701         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2702
2703         /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2704         bdi_down_write_wb_switch_rwsem(bdi);
2705         bdi_split_work_to_wbs(bdi, &work, false);
2706         wb_wait_for_completion(&done);
2707         bdi_up_write_wb_switch_rwsem(bdi);
2708
2709         wait_sb_inodes(sb);
2710 }
2711 EXPORT_SYMBOL(sync_inodes_sb);
2712
2713 /**
2714  * write_inode_now      -       write an inode to disk
2715  * @inode: inode to write to disk
2716  * @sync: whether the write should be synchronous or not
2717  *
2718  * This function commits an inode to disk immediately if it is dirty. This is
2719  * primarily needed by knfsd.
2720  *
2721  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2722  */
2723 int write_inode_now(struct inode *inode, int sync)
2724 {
2725         struct writeback_control wbc = {
2726                 .nr_to_write = LONG_MAX,
2727                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2728                 .range_start = 0,
2729                 .range_end = LLONG_MAX,
2730         };
2731
2732         if (!mapping_can_writeback(inode->i_mapping))
2733                 wbc.nr_to_write = 0;
2734
2735         might_sleep();
2736         return writeback_single_inode(inode, &wbc);
2737 }
2738 EXPORT_SYMBOL(write_inode_now);
2739
2740 /**
2741  * sync_inode_metadata - write an inode to disk
2742  * @inode: the inode to sync
2743  * @wait: wait for I/O to complete.
2744  *
2745  * Write an inode to disk and adjust its dirty state after completion.
2746  *
2747  * Note: only writes the actual inode, no associated data or other metadata.
2748  */
2749 int sync_inode_metadata(struct inode *inode, int wait)
2750 {
2751         struct writeback_control wbc = {
2752                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2753                 .nr_to_write = 0, /* metadata-only */
2754         };
2755
2756         return writeback_single_inode(inode, &wbc);
2757 }
2758 EXPORT_SYMBOL(sync_inode_metadata);