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