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