]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/fs-writeback.c
Merge tag 'iommu-fixes-v4.3-rc7' of git://git.kernel.org/pub/scm/linux/kernel/git...
[karo-tx-linux.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
30 #include <linux/memcontrol.h>
31 #include "internal.h"
32
33 /*
34  * 4MB minimal write chunk size
35  */
36 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_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}
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 wb_queue_work(struct bdi_writeback *wb,
177                           struct wb_writeback_work *work)
178 {
179         trace_writeback_queue(wb, work);
180
181         spin_lock_bh(&wb->work_lock);
182         if (!test_bit(WB_registered, &wb->state))
183                 goto out_unlock;
184         if (work->done)
185                 atomic_inc(&work->done->cnt);
186         list_add_tail(&work->list, &wb->work_list);
187         mod_delayed_work(bdi_wq, &wb->dwork, 0);
188 out_unlock:
189         spin_unlock_bh(&wb->work_lock);
190 }
191
192 /**
193  * wb_wait_for_completion - wait for completion of bdi_writeback_works
194  * @bdi: bdi work items were issued to
195  * @done: target wb_completion
196  *
197  * Wait for one or more work items issued to @bdi with their ->done field
198  * set to @done, which should have been defined with
199  * DEFINE_WB_COMPLETION_ONSTACK().  This function returns after all such
200  * work items are completed.  Work items which are waited upon aren't freed
201  * automatically on completion.
202  */
203 static void wb_wait_for_completion(struct backing_dev_info *bdi,
204                                    struct wb_completion *done)
205 {
206         atomic_dec(&done->cnt);         /* put down the initial count */
207         wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
208 }
209
210 #ifdef CONFIG_CGROUP_WRITEBACK
211
212 /* parameters for foreign inode detection, see wb_detach_inode() */
213 #define WB_FRN_TIME_SHIFT       13      /* 1s = 2^13, upto 8 secs w/ 16bit */
214 #define WB_FRN_TIME_AVG_SHIFT   3       /* avg = avg * 7/8 + new * 1/8 */
215 #define WB_FRN_TIME_CUT_DIV     2       /* ignore rounds < avg / 2 */
216 #define WB_FRN_TIME_PERIOD      (2 * (1 << WB_FRN_TIME_SHIFT))  /* 2s */
217
218 #define WB_FRN_HIST_SLOTS       16      /* inode->i_wb_frn_history is 16bit */
219 #define WB_FRN_HIST_UNIT        (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
220                                         /* each slot's duration is 2s / 16 */
221 #define WB_FRN_HIST_THR_SLOTS   (WB_FRN_HIST_SLOTS / 2)
222                                         /* if foreign slots >= 8, switch */
223 #define WB_FRN_HIST_MAX_SLOTS   (WB_FRN_HIST_THR_SLOTS / 2 + 1)
224                                         /* one round can affect upto 5 slots */
225
226 void __inode_attach_wb(struct inode *inode, struct page *page)
227 {
228         struct backing_dev_info *bdi = inode_to_bdi(inode);
229         struct bdi_writeback *wb = NULL;
230
231         if (inode_cgwb_enabled(inode)) {
232                 struct cgroup_subsys_state *memcg_css;
233
234                 if (page) {
235                         memcg_css = mem_cgroup_css_from_page(page);
236                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
237                 } else {
238                         /* must pin memcg_css, see wb_get_create() */
239                         memcg_css = task_get_css(current, memory_cgrp_id);
240                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
241                         css_put(memcg_css);
242                 }
243         }
244
245         if (!wb)
246                 wb = &bdi->wb;
247
248         /*
249          * There may be multiple instances of this function racing to
250          * update the same inode.  Use cmpxchg() to tell the winner.
251          */
252         if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
253                 wb_put(wb);
254 }
255
256 /**
257  * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
258  * @inode: inode of interest with i_lock held
259  *
260  * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
261  * held on entry and is released on return.  The returned wb is guaranteed
262  * to stay @inode's associated wb until its list_lock is released.
263  */
264 static struct bdi_writeback *
265 locked_inode_to_wb_and_lock_list(struct inode *inode)
266         __releases(&inode->i_lock)
267         __acquires(&wb->list_lock)
268 {
269         while (true) {
270                 struct bdi_writeback *wb = inode_to_wb(inode);
271
272                 /*
273                  * inode_to_wb() association is protected by both
274                  * @inode->i_lock and @wb->list_lock but list_lock nests
275                  * outside i_lock.  Drop i_lock and verify that the
276                  * association hasn't changed after acquiring list_lock.
277                  */
278                 wb_get(wb);
279                 spin_unlock(&inode->i_lock);
280                 spin_lock(&wb->list_lock);
281                 wb_put(wb);             /* not gonna deref it anymore */
282
283                 /* i_wb may have changed inbetween, can't use inode_to_wb() */
284                 if (likely(wb == inode->i_wb))
285                         return wb;      /* @inode already has ref */
286
287                 spin_unlock(&wb->list_lock);
288                 cpu_relax();
289                 spin_lock(&inode->i_lock);
290         }
291 }
292
293 /**
294  * inode_to_wb_and_lock_list - determine an inode's wb and lock it
295  * @inode: inode of interest
296  *
297  * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
298  * on entry.
299  */
300 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
301         __acquires(&wb->list_lock)
302 {
303         spin_lock(&inode->i_lock);
304         return locked_inode_to_wb_and_lock_list(inode);
305 }
306
307 struct inode_switch_wbs_context {
308         struct inode            *inode;
309         struct bdi_writeback    *new_wb;
310
311         struct rcu_head         rcu_head;
312         struct work_struct      work;
313 };
314
315 static void inode_switch_wbs_work_fn(struct work_struct *work)
316 {
317         struct inode_switch_wbs_context *isw =
318                 container_of(work, struct inode_switch_wbs_context, work);
319         struct inode *inode = isw->inode;
320         struct address_space *mapping = inode->i_mapping;
321         struct bdi_writeback *old_wb = inode->i_wb;
322         struct bdi_writeback *new_wb = isw->new_wb;
323         struct radix_tree_iter iter;
324         bool switched = false;
325         void **slot;
326
327         /*
328          * By the time control reaches here, RCU grace period has passed
329          * since I_WB_SWITCH assertion and all wb stat update transactions
330          * between unlocked_inode_to_wb_begin/end() are guaranteed to be
331          * synchronizing against mapping->tree_lock.
332          *
333          * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
334          * gives us exclusion against all wb related operations on @inode
335          * including IO list manipulations and stat updates.
336          */
337         if (old_wb < new_wb) {
338                 spin_lock(&old_wb->list_lock);
339                 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
340         } else {
341                 spin_lock(&new_wb->list_lock);
342                 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
343         }
344         spin_lock(&inode->i_lock);
345         spin_lock_irq(&mapping->tree_lock);
346
347         /*
348          * Once I_FREEING is visible under i_lock, the eviction path owns
349          * the inode and we shouldn't modify ->i_io_list.
350          */
351         if (unlikely(inode->i_state & I_FREEING))
352                 goto skip_switch;
353
354         /*
355          * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
356          * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
357          * pages actually under underwriteback.
358          */
359         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
360                                    PAGECACHE_TAG_DIRTY) {
361                 struct page *page = radix_tree_deref_slot_protected(slot,
362                                                         &mapping->tree_lock);
363                 if (likely(page) && PageDirty(page)) {
364                         __dec_wb_stat(old_wb, WB_RECLAIMABLE);
365                         __inc_wb_stat(new_wb, WB_RECLAIMABLE);
366                 }
367         }
368
369         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
370                                    PAGECACHE_TAG_WRITEBACK) {
371                 struct page *page = radix_tree_deref_slot_protected(slot,
372                                                         &mapping->tree_lock);
373                 if (likely(page)) {
374                         WARN_ON_ONCE(!PageWriteback(page));
375                         __dec_wb_stat(old_wb, WB_WRITEBACK);
376                         __inc_wb_stat(new_wb, WB_WRITEBACK);
377                 }
378         }
379
380         wb_get(new_wb);
381
382         /*
383          * Transfer to @new_wb's IO list if necessary.  The specific list
384          * @inode was on is ignored and the inode is put on ->b_dirty which
385          * is always correct including from ->b_dirty_time.  The transfer
386          * preserves @inode->dirtied_when ordering.
387          */
388         if (!list_empty(&inode->i_io_list)) {
389                 struct inode *pos;
390
391                 inode_io_list_del_locked(inode, old_wb);
392                 inode->i_wb = new_wb;
393                 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
394                         if (time_after_eq(inode->dirtied_when,
395                                           pos->dirtied_when))
396                                 break;
397                 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
398         } else {
399                 inode->i_wb = new_wb;
400         }
401
402         /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
403         inode->i_wb_frn_winner = 0;
404         inode->i_wb_frn_avg_time = 0;
405         inode->i_wb_frn_history = 0;
406         switched = true;
407 skip_switch:
408         /*
409          * Paired with load_acquire in unlocked_inode_to_wb_begin() and
410          * ensures that the new wb is visible if they see !I_WB_SWITCH.
411          */
412         smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
413
414         spin_unlock_irq(&mapping->tree_lock);
415         spin_unlock(&inode->i_lock);
416         spin_unlock(&new_wb->list_lock);
417         spin_unlock(&old_wb->list_lock);
418
419         if (switched) {
420                 wb_wakeup(new_wb);
421                 wb_put(old_wb);
422         }
423         wb_put(new_wb);
424
425         iput(inode);
426         kfree(isw);
427 }
428
429 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
430 {
431         struct inode_switch_wbs_context *isw = container_of(rcu_head,
432                                 struct inode_switch_wbs_context, rcu_head);
433
434         /* needs to grab bh-unsafe locks, bounce to work item */
435         INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
436         schedule_work(&isw->work);
437 }
438
439 /**
440  * inode_switch_wbs - change the wb association of an inode
441  * @inode: target inode
442  * @new_wb_id: ID of the new wb
443  *
444  * Switch @inode's wb association to the wb identified by @new_wb_id.  The
445  * switching is performed asynchronously and may fail silently.
446  */
447 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
448 {
449         struct backing_dev_info *bdi = inode_to_bdi(inode);
450         struct cgroup_subsys_state *memcg_css;
451         struct inode_switch_wbs_context *isw;
452
453         /* noop if seems to be already in progress */
454         if (inode->i_state & I_WB_SWITCH)
455                 return;
456
457         isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
458         if (!isw)
459                 return;
460
461         /* find and pin the new wb */
462         rcu_read_lock();
463         memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
464         if (memcg_css)
465                 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
466         rcu_read_unlock();
467         if (!isw->new_wb)
468                 goto out_free;
469
470         /* while holding I_WB_SWITCH, no one else can update the association */
471         spin_lock(&inode->i_lock);
472         if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
473             inode_to_wb(inode) == isw->new_wb) {
474                 spin_unlock(&inode->i_lock);
475                 goto out_free;
476         }
477         inode->i_state |= I_WB_SWITCH;
478         spin_unlock(&inode->i_lock);
479
480         ihold(inode);
481         isw->inode = inode;
482
483         /*
484          * In addition to synchronizing among switchers, I_WB_SWITCH tells
485          * the RCU protected stat update paths to grab the mapping's
486          * tree_lock so that stat transfer can synchronize against them.
487          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
488          */
489         call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
490         return;
491
492 out_free:
493         if (isw->new_wb)
494                 wb_put(isw->new_wb);
495         kfree(isw);
496 }
497
498 /**
499  * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
500  * @wbc: writeback_control of interest
501  * @inode: target inode
502  *
503  * @inode is locked and about to be written back under the control of @wbc.
504  * Record @inode's writeback context into @wbc and unlock the i_lock.  On
505  * writeback completion, wbc_detach_inode() should be called.  This is used
506  * to track the cgroup writeback context.
507  */
508 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
509                                  struct inode *inode)
510 {
511         if (!inode_cgwb_enabled(inode)) {
512                 spin_unlock(&inode->i_lock);
513                 return;
514         }
515
516         wbc->wb = inode_to_wb(inode);
517         wbc->inode = inode;
518
519         wbc->wb_id = wbc->wb->memcg_css->id;
520         wbc->wb_lcand_id = inode->i_wb_frn_winner;
521         wbc->wb_tcand_id = 0;
522         wbc->wb_bytes = 0;
523         wbc->wb_lcand_bytes = 0;
524         wbc->wb_tcand_bytes = 0;
525
526         wb_get(wbc->wb);
527         spin_unlock(&inode->i_lock);
528
529         /*
530          * A dying wb indicates that the memcg-blkcg mapping has changed
531          * and a new wb is already serving the memcg.  Switch immediately.
532          */
533         if (unlikely(wb_dying(wbc->wb)))
534                 inode_switch_wbs(inode, wbc->wb_id);
535 }
536
537 /**
538  * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
539  * @wbc: writeback_control of the just finished writeback
540  *
541  * To be called after a writeback attempt of an inode finishes and undoes
542  * wbc_attach_and_unlock_inode().  Can be called under any context.
543  *
544  * As concurrent write sharing of an inode is expected to be very rare and
545  * memcg only tracks page ownership on first-use basis severely confining
546  * the usefulness of such sharing, cgroup writeback tracks ownership
547  * per-inode.  While the support for concurrent write sharing of an inode
548  * is deemed unnecessary, an inode being written to by different cgroups at
549  * different points in time is a lot more common, and, more importantly,
550  * charging only by first-use can too readily lead to grossly incorrect
551  * behaviors (single foreign page can lead to gigabytes of writeback to be
552  * incorrectly attributed).
553  *
554  * To resolve this issue, cgroup writeback detects the majority dirtier of
555  * an inode and transfers the ownership to it.  To avoid unnnecessary
556  * oscillation, the detection mechanism keeps track of history and gives
557  * out the switch verdict only if the foreign usage pattern is stable over
558  * a certain amount of time and/or writeback attempts.
559  *
560  * On each writeback attempt, @wbc tries to detect the majority writer
561  * using Boyer-Moore majority vote algorithm.  In addition to the byte
562  * count from the majority voting, it also counts the bytes written for the
563  * current wb and the last round's winner wb (max of last round's current
564  * wb, the winner from two rounds ago, and the last round's majority
565  * candidate).  Keeping track of the historical winner helps the algorithm
566  * to semi-reliably detect the most active writer even when it's not the
567  * absolute majority.
568  *
569  * Once the winner of the round is determined, whether the winner is
570  * foreign or not and how much IO time the round consumed is recorded in
571  * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
572  * over a certain threshold, the switch verdict is given.
573  */
574 void wbc_detach_inode(struct writeback_control *wbc)
575 {
576         struct bdi_writeback *wb = wbc->wb;
577         struct inode *inode = wbc->inode;
578         unsigned long avg_time, max_bytes, max_time;
579         u16 history;
580         int max_id;
581
582         if (!wb)
583                 return;
584
585         history = inode->i_wb_frn_history;
586         avg_time = inode->i_wb_frn_avg_time;
587
588         /* pick the winner of this round */
589         if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
590             wbc->wb_bytes >= wbc->wb_tcand_bytes) {
591                 max_id = wbc->wb_id;
592                 max_bytes = wbc->wb_bytes;
593         } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
594                 max_id = wbc->wb_lcand_id;
595                 max_bytes = wbc->wb_lcand_bytes;
596         } else {
597                 max_id = wbc->wb_tcand_id;
598                 max_bytes = wbc->wb_tcand_bytes;
599         }
600
601         /*
602          * Calculate the amount of IO time the winner consumed and fold it
603          * into the running average kept per inode.  If the consumed IO
604          * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
605          * deciding whether to switch or not.  This is to prevent one-off
606          * small dirtiers from skewing the verdict.
607          */
608         max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
609                                 wb->avg_write_bandwidth);
610         if (avg_time)
611                 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
612                             (avg_time >> WB_FRN_TIME_AVG_SHIFT);
613         else
614                 avg_time = max_time;    /* immediate catch up on first run */
615
616         if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
617                 int slots;
618
619                 /*
620                  * The switch verdict is reached if foreign wb's consume
621                  * more than a certain proportion of IO time in a
622                  * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
623                  * history mask where each bit represents one sixteenth of
624                  * the period.  Determine the number of slots to shift into
625                  * history from @max_time.
626                  */
627                 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
628                             (unsigned long)WB_FRN_HIST_MAX_SLOTS);
629                 history <<= slots;
630                 if (wbc->wb_id != max_id)
631                         history |= (1U << slots) - 1;
632
633                 /*
634                  * Switch if the current wb isn't the consistent winner.
635                  * If there are multiple closely competing dirtiers, the
636                  * inode may switch across them repeatedly over time, which
637                  * is okay.  The main goal is avoiding keeping an inode on
638                  * the wrong wb for an extended period of time.
639                  */
640                 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
641                         inode_switch_wbs(inode, max_id);
642         }
643
644         /*
645          * Multiple instances of this function may race to update the
646          * following fields but we don't mind occassional inaccuracies.
647          */
648         inode->i_wb_frn_winner = max_id;
649         inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
650         inode->i_wb_frn_history = history;
651
652         wb_put(wbc->wb);
653         wbc->wb = NULL;
654 }
655
656 /**
657  * wbc_account_io - account IO issued during writeback
658  * @wbc: writeback_control of the writeback in progress
659  * @page: page being written out
660  * @bytes: number of bytes being written out
661  *
662  * @bytes from @page are about to written out during the writeback
663  * controlled by @wbc.  Keep the book for foreign inode detection.  See
664  * wbc_detach_inode().
665  */
666 void wbc_account_io(struct writeback_control *wbc, struct page *page,
667                     size_t bytes)
668 {
669         int id;
670
671         /*
672          * pageout() path doesn't attach @wbc to the inode being written
673          * out.  This is intentional as we don't want the function to block
674          * behind a slow cgroup.  Ultimately, we want pageout() to kick off
675          * regular writeback instead of writing things out itself.
676          */
677         if (!wbc->wb)
678                 return;
679
680         rcu_read_lock();
681         id = mem_cgroup_css_from_page(page)->id;
682         rcu_read_unlock();
683
684         if (id == wbc->wb_id) {
685                 wbc->wb_bytes += bytes;
686                 return;
687         }
688
689         if (id == wbc->wb_lcand_id)
690                 wbc->wb_lcand_bytes += bytes;
691
692         /* Boyer-Moore majority vote algorithm */
693         if (!wbc->wb_tcand_bytes)
694                 wbc->wb_tcand_id = id;
695         if (id == wbc->wb_tcand_id)
696                 wbc->wb_tcand_bytes += bytes;
697         else
698                 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
699 }
700 EXPORT_SYMBOL_GPL(wbc_account_io);
701
702 /**
703  * inode_congested - test whether an inode is congested
704  * @inode: inode to test for congestion (may be NULL)
705  * @cong_bits: mask of WB_[a]sync_congested bits to test
706  *
707  * Tests whether @inode is congested.  @cong_bits is the mask of congestion
708  * bits to test and the return value is the mask of set bits.
709  *
710  * If cgroup writeback is enabled for @inode, the congestion state is
711  * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
712  * associated with @inode is congested; otherwise, the root wb's congestion
713  * state is used.
714  *
715  * @inode is allowed to be NULL as this function is often called on
716  * mapping->host which is NULL for the swapper space.
717  */
718 int inode_congested(struct inode *inode, int cong_bits)
719 {
720         /*
721          * Once set, ->i_wb never becomes NULL while the inode is alive.
722          * Start transaction iff ->i_wb is visible.
723          */
724         if (inode && inode_to_wb_is_valid(inode)) {
725                 struct bdi_writeback *wb;
726                 bool locked, congested;
727
728                 wb = unlocked_inode_to_wb_begin(inode, &locked);
729                 congested = wb_congested(wb, cong_bits);
730                 unlocked_inode_to_wb_end(inode, locked);
731                 return congested;
732         }
733
734         return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
735 }
736 EXPORT_SYMBOL_GPL(inode_congested);
737
738 /**
739  * wb_split_bdi_pages - split nr_pages to write according to bandwidth
740  * @wb: target bdi_writeback to split @nr_pages to
741  * @nr_pages: number of pages to write for the whole bdi
742  *
743  * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
744  * relation to the total write bandwidth of all wb's w/ dirty inodes on
745  * @wb->bdi.
746  */
747 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
748 {
749         unsigned long this_bw = wb->avg_write_bandwidth;
750         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
751
752         if (nr_pages == LONG_MAX)
753                 return LONG_MAX;
754
755         /*
756          * This may be called on clean wb's and proportional distribution
757          * may not make sense, just use the original @nr_pages in those
758          * cases.  In general, we wanna err on the side of writing more.
759          */
760         if (!tot_bw || this_bw >= tot_bw)
761                 return nr_pages;
762         else
763                 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
764 }
765
766 /**
767  * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
768  * @bdi: target backing_dev_info
769  * @base_work: wb_writeback_work to issue
770  * @skip_if_busy: skip wb's which already have writeback in progress
771  *
772  * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
773  * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
774  * distributed to the busy wbs according to each wb's proportion in the
775  * total active write bandwidth of @bdi.
776  */
777 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
778                                   struct wb_writeback_work *base_work,
779                                   bool skip_if_busy)
780 {
781         struct bdi_writeback *last_wb = NULL;
782         struct bdi_writeback *wb = list_entry_rcu(&bdi->wb_list,
783                                                 struct bdi_writeback, bdi_node);
784
785         might_sleep();
786 restart:
787         rcu_read_lock();
788         list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
789                 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
790                 struct wb_writeback_work fallback_work;
791                 struct wb_writeback_work *work;
792                 long nr_pages;
793
794                 if (last_wb) {
795                         wb_put(last_wb);
796                         last_wb = NULL;
797                 }
798
799                 /* SYNC_ALL writes out I_DIRTY_TIME too */
800                 if (!wb_has_dirty_io(wb) &&
801                     (base_work->sync_mode == WB_SYNC_NONE ||
802                      list_empty(&wb->b_dirty_time)))
803                         continue;
804                 if (skip_if_busy && writeback_in_progress(wb))
805                         continue;
806
807                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
808
809                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
810                 if (work) {
811                         *work = *base_work;
812                         work->nr_pages = nr_pages;
813                         work->auto_free = 1;
814                         wb_queue_work(wb, work);
815                         continue;
816                 }
817
818                 /* alloc failed, execute synchronously using on-stack fallback */
819                 work = &fallback_work;
820                 *work = *base_work;
821                 work->nr_pages = nr_pages;
822                 work->auto_free = 0;
823                 work->done = &fallback_work_done;
824
825                 wb_queue_work(wb, work);
826
827                 /*
828                  * Pin @wb so that it stays on @bdi->wb_list.  This allows
829                  * continuing iteration from @wb after dropping and
830                  * regrabbing rcu read lock.
831                  */
832                 wb_get(wb);
833                 last_wb = wb;
834
835                 rcu_read_unlock();
836                 wb_wait_for_completion(bdi, &fallback_work_done);
837                 goto restart;
838         }
839         rcu_read_unlock();
840
841         if (last_wb)
842                 wb_put(last_wb);
843 }
844
845 #else   /* CONFIG_CGROUP_WRITEBACK */
846
847 static struct bdi_writeback *
848 locked_inode_to_wb_and_lock_list(struct inode *inode)
849         __releases(&inode->i_lock)
850         __acquires(&wb->list_lock)
851 {
852         struct bdi_writeback *wb = inode_to_wb(inode);
853
854         spin_unlock(&inode->i_lock);
855         spin_lock(&wb->list_lock);
856         return wb;
857 }
858
859 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
860         __acquires(&wb->list_lock)
861 {
862         struct bdi_writeback *wb = inode_to_wb(inode);
863
864         spin_lock(&wb->list_lock);
865         return wb;
866 }
867
868 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
869 {
870         return nr_pages;
871 }
872
873 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
874                                   struct wb_writeback_work *base_work,
875                                   bool skip_if_busy)
876 {
877         might_sleep();
878
879         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
880                 base_work->auto_free = 0;
881                 wb_queue_work(&bdi->wb, base_work);
882         }
883 }
884
885 #endif  /* CONFIG_CGROUP_WRITEBACK */
886
887 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
888                         bool range_cyclic, enum wb_reason reason)
889 {
890         struct wb_writeback_work *work;
891
892         if (!wb_has_dirty_io(wb))
893                 return;
894
895         /*
896          * This is WB_SYNC_NONE writeback, so if allocation fails just
897          * wakeup the thread for old dirty data writeback
898          */
899         work = kzalloc(sizeof(*work), GFP_ATOMIC);
900         if (!work) {
901                 trace_writeback_nowork(wb);
902                 wb_wakeup(wb);
903                 return;
904         }
905
906         work->sync_mode = WB_SYNC_NONE;
907         work->nr_pages  = nr_pages;
908         work->range_cyclic = range_cyclic;
909         work->reason    = reason;
910         work->auto_free = 1;
911
912         wb_queue_work(wb, work);
913 }
914
915 /**
916  * wb_start_background_writeback - start background writeback
917  * @wb: bdi_writback to write from
918  *
919  * Description:
920  *   This makes sure WB_SYNC_NONE background writeback happens. When
921  *   this function returns, it is only guaranteed that for given wb
922  *   some IO is happening if we are over background dirty threshold.
923  *   Caller need not hold sb s_umount semaphore.
924  */
925 void wb_start_background_writeback(struct bdi_writeback *wb)
926 {
927         /*
928          * We just wake up the flusher thread. It will perform background
929          * writeback as soon as there is no other work to do.
930          */
931         trace_writeback_wake_background(wb);
932         wb_wakeup(wb);
933 }
934
935 /*
936  * Remove the inode from the writeback list it is on.
937  */
938 void inode_io_list_del(struct inode *inode)
939 {
940         struct bdi_writeback *wb;
941
942         wb = inode_to_wb_and_lock_list(inode);
943         inode_io_list_del_locked(inode, wb);
944         spin_unlock(&wb->list_lock);
945 }
946
947 /*
948  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
949  * furthest end of its superblock's dirty-inode list.
950  *
951  * Before stamping the inode's ->dirtied_when, we check to see whether it is
952  * already the most-recently-dirtied inode on the b_dirty list.  If that is
953  * the case then the inode must have been redirtied while it was being written
954  * out and we don't reset its dirtied_when.
955  */
956 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
957 {
958         if (!list_empty(&wb->b_dirty)) {
959                 struct inode *tail;
960
961                 tail = wb_inode(wb->b_dirty.next);
962                 if (time_before(inode->dirtied_when, tail->dirtied_when))
963                         inode->dirtied_when = jiffies;
964         }
965         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
966 }
967
968 /*
969  * requeue inode for re-scanning after bdi->b_io list is exhausted.
970  */
971 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
972 {
973         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
974 }
975
976 static void inode_sync_complete(struct inode *inode)
977 {
978         inode->i_state &= ~I_SYNC;
979         /* If inode is clean an unused, put it into LRU now... */
980         inode_add_lru(inode);
981         /* Waiters must see I_SYNC cleared before being woken up */
982         smp_mb();
983         wake_up_bit(&inode->i_state, __I_SYNC);
984 }
985
986 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
987 {
988         bool ret = time_after(inode->dirtied_when, t);
989 #ifndef CONFIG_64BIT
990         /*
991          * For inodes being constantly redirtied, dirtied_when can get stuck.
992          * It _appears_ to be in the future, but is actually in distant past.
993          * This test is necessary to prevent such wrapped-around relative times
994          * from permanently stopping the whole bdi writeback.
995          */
996         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
997 #endif
998         return ret;
999 }
1000
1001 #define EXPIRE_DIRTY_ATIME 0x0001
1002
1003 /*
1004  * Move expired (dirtied before work->older_than_this) dirty inodes from
1005  * @delaying_queue to @dispatch_queue.
1006  */
1007 static int move_expired_inodes(struct list_head *delaying_queue,
1008                                struct list_head *dispatch_queue,
1009                                int flags,
1010                                struct wb_writeback_work *work)
1011 {
1012         unsigned long *older_than_this = NULL;
1013         unsigned long expire_time;
1014         LIST_HEAD(tmp);
1015         struct list_head *pos, *node;
1016         struct super_block *sb = NULL;
1017         struct inode *inode;
1018         int do_sb_sort = 0;
1019         int moved = 0;
1020
1021         if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1022                 older_than_this = work->older_than_this;
1023         else if (!work->for_sync) {
1024                 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1025                 older_than_this = &expire_time;
1026         }
1027         while (!list_empty(delaying_queue)) {
1028                 inode = wb_inode(delaying_queue->prev);
1029                 if (older_than_this &&
1030                     inode_dirtied_after(inode, *older_than_this))
1031                         break;
1032                 list_move(&inode->i_io_list, &tmp);
1033                 moved++;
1034                 if (flags & EXPIRE_DIRTY_ATIME)
1035                         set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1036                 if (sb_is_blkdev_sb(inode->i_sb))
1037                         continue;
1038                 if (sb && sb != inode->i_sb)
1039                         do_sb_sort = 1;
1040                 sb = inode->i_sb;
1041         }
1042
1043         /* just one sb in list, splice to dispatch_queue and we're done */
1044         if (!do_sb_sort) {
1045                 list_splice(&tmp, dispatch_queue);
1046                 goto out;
1047         }
1048
1049         /* Move inodes from one superblock together */
1050         while (!list_empty(&tmp)) {
1051                 sb = wb_inode(tmp.prev)->i_sb;
1052                 list_for_each_prev_safe(pos, node, &tmp) {
1053                         inode = wb_inode(pos);
1054                         if (inode->i_sb == sb)
1055                                 list_move(&inode->i_io_list, dispatch_queue);
1056                 }
1057         }
1058 out:
1059         return moved;
1060 }
1061
1062 /*
1063  * Queue all expired dirty inodes for io, eldest first.
1064  * Before
1065  *         newly dirtied     b_dirty    b_io    b_more_io
1066  *         =============>    gf         edc     BA
1067  * After
1068  *         newly dirtied     b_dirty    b_io    b_more_io
1069  *         =============>    g          fBAedc
1070  *                                           |
1071  *                                           +--> dequeue for IO
1072  */
1073 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1074 {
1075         int moved;
1076
1077         assert_spin_locked(&wb->list_lock);
1078         list_splice_init(&wb->b_more_io, &wb->b_io);
1079         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1080         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1081                                      EXPIRE_DIRTY_ATIME, work);
1082         if (moved)
1083                 wb_io_lists_populated(wb);
1084         trace_writeback_queue_io(wb, work, moved);
1085 }
1086
1087 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1088 {
1089         int ret;
1090
1091         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1092                 trace_writeback_write_inode_start(inode, wbc);
1093                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1094                 trace_writeback_write_inode(inode, wbc);
1095                 return ret;
1096         }
1097         return 0;
1098 }
1099
1100 /*
1101  * Wait for writeback on an inode to complete. Called with i_lock held.
1102  * Caller must make sure inode cannot go away when we drop i_lock.
1103  */
1104 static void __inode_wait_for_writeback(struct inode *inode)
1105         __releases(inode->i_lock)
1106         __acquires(inode->i_lock)
1107 {
1108         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1109         wait_queue_head_t *wqh;
1110
1111         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1112         while (inode->i_state & I_SYNC) {
1113                 spin_unlock(&inode->i_lock);
1114                 __wait_on_bit(wqh, &wq, bit_wait,
1115                               TASK_UNINTERRUPTIBLE);
1116                 spin_lock(&inode->i_lock);
1117         }
1118 }
1119
1120 /*
1121  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1122  */
1123 void inode_wait_for_writeback(struct inode *inode)
1124 {
1125         spin_lock(&inode->i_lock);
1126         __inode_wait_for_writeback(inode);
1127         spin_unlock(&inode->i_lock);
1128 }
1129
1130 /*
1131  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1132  * held and drops it. It is aimed for callers not holding any inode reference
1133  * so once i_lock is dropped, inode can go away.
1134  */
1135 static void inode_sleep_on_writeback(struct inode *inode)
1136         __releases(inode->i_lock)
1137 {
1138         DEFINE_WAIT(wait);
1139         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1140         int sleep;
1141
1142         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1143         sleep = inode->i_state & I_SYNC;
1144         spin_unlock(&inode->i_lock);
1145         if (sleep)
1146                 schedule();
1147         finish_wait(wqh, &wait);
1148 }
1149
1150 /*
1151  * Find proper writeback list for the inode depending on its current state and
1152  * possibly also change of its state while we were doing writeback.  Here we
1153  * handle things such as livelock prevention or fairness of writeback among
1154  * inodes. This function can be called only by flusher thread - noone else
1155  * processes all inodes in writeback lists and requeueing inodes behind flusher
1156  * thread's back can have unexpected consequences.
1157  */
1158 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1159                           struct writeback_control *wbc)
1160 {
1161         if (inode->i_state & I_FREEING)
1162                 return;
1163
1164         /*
1165          * Sync livelock prevention. Each inode is tagged and synced in one
1166          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1167          * the dirty time to prevent enqueue and sync it again.
1168          */
1169         if ((inode->i_state & I_DIRTY) &&
1170             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1171                 inode->dirtied_when = jiffies;
1172
1173         if (wbc->pages_skipped) {
1174                 /*
1175                  * writeback is not making progress due to locked
1176                  * buffers. Skip this inode for now.
1177                  */
1178                 redirty_tail(inode, wb);
1179                 return;
1180         }
1181
1182         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1183                 /*
1184                  * We didn't write back all the pages.  nfs_writepages()
1185                  * sometimes bales out without doing anything.
1186                  */
1187                 if (wbc->nr_to_write <= 0) {
1188                         /* Slice used up. Queue for next turn. */
1189                         requeue_io(inode, wb);
1190                 } else {
1191                         /*
1192                          * Writeback blocked by something other than
1193                          * congestion. Delay the inode for some time to
1194                          * avoid spinning on the CPU (100% iowait)
1195                          * retrying writeback of the dirty page/inode
1196                          * that cannot be performed immediately.
1197                          */
1198                         redirty_tail(inode, wb);
1199                 }
1200         } else if (inode->i_state & I_DIRTY) {
1201                 /*
1202                  * Filesystems can dirty the inode during writeback operations,
1203                  * such as delayed allocation during submission or metadata
1204                  * updates after data IO completion.
1205                  */
1206                 redirty_tail(inode, wb);
1207         } else if (inode->i_state & I_DIRTY_TIME) {
1208                 inode->dirtied_when = jiffies;
1209                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1210         } else {
1211                 /* The inode is clean. Remove from writeback lists. */
1212                 inode_io_list_del_locked(inode, wb);
1213         }
1214 }
1215
1216 /*
1217  * Write out an inode and its dirty pages. Do not update the writeback list
1218  * linkage. That is left to the caller. The caller is also responsible for
1219  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1220  */
1221 static int
1222 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1223 {
1224         struct address_space *mapping = inode->i_mapping;
1225         long nr_to_write = wbc->nr_to_write;
1226         unsigned dirty;
1227         int ret;
1228
1229         WARN_ON(!(inode->i_state & I_SYNC));
1230
1231         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1232
1233         ret = do_writepages(mapping, wbc);
1234
1235         /*
1236          * Make sure to wait on the data before writing out the metadata.
1237          * This is important for filesystems that modify metadata on data
1238          * I/O completion. We don't do it for sync(2) writeback because it has a
1239          * separate, external IO completion path and ->sync_fs for guaranteeing
1240          * inode metadata is written back correctly.
1241          */
1242         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1243                 int err = filemap_fdatawait(mapping);
1244                 if (ret == 0)
1245                         ret = err;
1246         }
1247
1248         /*
1249          * Some filesystems may redirty the inode during the writeback
1250          * due to delalloc, clear dirty metadata flags right before
1251          * write_inode()
1252          */
1253         spin_lock(&inode->i_lock);
1254
1255         dirty = inode->i_state & I_DIRTY;
1256         if (inode->i_state & I_DIRTY_TIME) {
1257                 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1258                     unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1259                     unlikely(time_after(jiffies,
1260                                         (inode->dirtied_time_when +
1261                                          dirtytime_expire_interval * HZ)))) {
1262                         dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1263                         trace_writeback_lazytime(inode);
1264                 }
1265         } else
1266                 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1267         inode->i_state &= ~dirty;
1268
1269         /*
1270          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1271          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1272          * either they see the I_DIRTY bits cleared or we see the dirtied
1273          * inode.
1274          *
1275          * I_DIRTY_PAGES is always cleared together above even if @mapping
1276          * still has dirty pages.  The flag is reinstated after smp_mb() if
1277          * necessary.  This guarantees that either __mark_inode_dirty()
1278          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1279          */
1280         smp_mb();
1281
1282         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1283                 inode->i_state |= I_DIRTY_PAGES;
1284
1285         spin_unlock(&inode->i_lock);
1286
1287         if (dirty & I_DIRTY_TIME)
1288                 mark_inode_dirty_sync(inode);
1289         /* Don't write the inode if only I_DIRTY_PAGES was set */
1290         if (dirty & ~I_DIRTY_PAGES) {
1291                 int err = write_inode(inode, wbc);
1292                 if (ret == 0)
1293                         ret = err;
1294         }
1295         trace_writeback_single_inode(inode, wbc, nr_to_write);
1296         return ret;
1297 }
1298
1299 /*
1300  * Write out an inode's dirty pages. Either the caller has an active reference
1301  * on the inode or the inode has I_WILL_FREE set.
1302  *
1303  * This function is designed to be called for writing back one inode which
1304  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1305  * and does more profound writeback list handling in writeback_sb_inodes().
1306  */
1307 static int
1308 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1309                        struct writeback_control *wbc)
1310 {
1311         int ret = 0;
1312
1313         spin_lock(&inode->i_lock);
1314         if (!atomic_read(&inode->i_count))
1315                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1316         else
1317                 WARN_ON(inode->i_state & I_WILL_FREE);
1318
1319         if (inode->i_state & I_SYNC) {
1320                 if (wbc->sync_mode != WB_SYNC_ALL)
1321                         goto out;
1322                 /*
1323                  * It's a data-integrity sync. We must wait. Since callers hold
1324                  * inode reference or inode has I_WILL_FREE set, it cannot go
1325                  * away under us.
1326                  */
1327                 __inode_wait_for_writeback(inode);
1328         }
1329         WARN_ON(inode->i_state & I_SYNC);
1330         /*
1331          * Skip inode if it is clean and we have no outstanding writeback in
1332          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1333          * function since flusher thread may be doing for example sync in
1334          * parallel and if we move the inode, it could get skipped. So here we
1335          * make sure inode is on some writeback list and leave it there unless
1336          * we have completely cleaned the inode.
1337          */
1338         if (!(inode->i_state & I_DIRTY_ALL) &&
1339             (wbc->sync_mode != WB_SYNC_ALL ||
1340              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1341                 goto out;
1342         inode->i_state |= I_SYNC;
1343         wbc_attach_and_unlock_inode(wbc, inode);
1344
1345         ret = __writeback_single_inode(inode, wbc);
1346
1347         wbc_detach_inode(wbc);
1348         spin_lock(&wb->list_lock);
1349         spin_lock(&inode->i_lock);
1350         /*
1351          * If inode is clean, remove it from writeback lists. Otherwise don't
1352          * touch it. See comment above for explanation.
1353          */
1354         if (!(inode->i_state & I_DIRTY_ALL))
1355                 inode_io_list_del_locked(inode, wb);
1356         spin_unlock(&wb->list_lock);
1357         inode_sync_complete(inode);
1358 out:
1359         spin_unlock(&inode->i_lock);
1360         return ret;
1361 }
1362
1363 static long writeback_chunk_size(struct bdi_writeback *wb,
1364                                  struct wb_writeback_work *work)
1365 {
1366         long pages;
1367
1368         /*
1369          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1370          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1371          * here avoids calling into writeback_inodes_wb() more than once.
1372          *
1373          * The intended call sequence for WB_SYNC_ALL writeback is:
1374          *
1375          *      wb_writeback()
1376          *          writeback_sb_inodes()       <== called only once
1377          *              write_cache_pages()     <== called once for each inode
1378          *                   (quickly) tag currently dirty pages
1379          *                   (maybe slowly) sync all tagged pages
1380          */
1381         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1382                 pages = LONG_MAX;
1383         else {
1384                 pages = min(wb->avg_write_bandwidth / 2,
1385                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1386                 pages = min(pages, work->nr_pages);
1387                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1388                                    MIN_WRITEBACK_PAGES);
1389         }
1390
1391         return pages;
1392 }
1393
1394 /*
1395  * Write a portion of b_io inodes which belong to @sb.
1396  *
1397  * Return the number of pages and/or inodes written.
1398  *
1399  * NOTE! This is called with wb->list_lock held, and will
1400  * unlock and relock that for each inode it ends up doing
1401  * IO for.
1402  */
1403 static long writeback_sb_inodes(struct super_block *sb,
1404                                 struct bdi_writeback *wb,
1405                                 struct wb_writeback_work *work)
1406 {
1407         struct writeback_control wbc = {
1408                 .sync_mode              = work->sync_mode,
1409                 .tagged_writepages      = work->tagged_writepages,
1410                 .for_kupdate            = work->for_kupdate,
1411                 .for_background         = work->for_background,
1412                 .for_sync               = work->for_sync,
1413                 .range_cyclic           = work->range_cyclic,
1414                 .range_start            = 0,
1415                 .range_end              = LLONG_MAX,
1416         };
1417         unsigned long start_time = jiffies;
1418         long write_chunk;
1419         long wrote = 0;  /* count both pages and inodes */
1420
1421         while (!list_empty(&wb->b_io)) {
1422                 struct inode *inode = wb_inode(wb->b_io.prev);
1423
1424                 if (inode->i_sb != sb) {
1425                         if (work->sb) {
1426                                 /*
1427                                  * We only want to write back data for this
1428                                  * superblock, move all inodes not belonging
1429                                  * to it back onto the dirty list.
1430                                  */
1431                                 redirty_tail(inode, wb);
1432                                 continue;
1433                         }
1434
1435                         /*
1436                          * The inode belongs to a different superblock.
1437                          * Bounce back to the caller to unpin this and
1438                          * pin the next superblock.
1439                          */
1440                         break;
1441                 }
1442
1443                 /*
1444                  * Don't bother with new inodes or inodes being freed, first
1445                  * kind does not need periodic writeout yet, and for the latter
1446                  * kind writeout is handled by the freer.
1447                  */
1448                 spin_lock(&inode->i_lock);
1449                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1450                         spin_unlock(&inode->i_lock);
1451                         redirty_tail(inode, wb);
1452                         continue;
1453                 }
1454                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1455                         /*
1456                          * If this inode is locked for writeback and we are not
1457                          * doing writeback-for-data-integrity, move it to
1458                          * b_more_io so that writeback can proceed with the
1459                          * other inodes on s_io.
1460                          *
1461                          * We'll have another go at writing back this inode
1462                          * when we completed a full scan of b_io.
1463                          */
1464                         spin_unlock(&inode->i_lock);
1465                         requeue_io(inode, wb);
1466                         trace_writeback_sb_inodes_requeue(inode);
1467                         continue;
1468                 }
1469                 spin_unlock(&wb->list_lock);
1470
1471                 /*
1472                  * We already requeued the inode if it had I_SYNC set and we
1473                  * are doing WB_SYNC_NONE writeback. So this catches only the
1474                  * WB_SYNC_ALL case.
1475                  */
1476                 if (inode->i_state & I_SYNC) {
1477                         /* Wait for I_SYNC. This function drops i_lock... */
1478                         inode_sleep_on_writeback(inode);
1479                         /* Inode may be gone, start again */
1480                         spin_lock(&wb->list_lock);
1481                         continue;
1482                 }
1483                 inode->i_state |= I_SYNC;
1484                 wbc_attach_and_unlock_inode(&wbc, inode);
1485
1486                 write_chunk = writeback_chunk_size(wb, work);
1487                 wbc.nr_to_write = write_chunk;
1488                 wbc.pages_skipped = 0;
1489
1490                 /*
1491                  * We use I_SYNC to pin the inode in memory. While it is set
1492                  * evict_inode() will wait so the inode cannot be freed.
1493                  */
1494                 __writeback_single_inode(inode, &wbc);
1495
1496                 wbc_detach_inode(&wbc);
1497                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1498                 wrote += write_chunk - wbc.nr_to_write;
1499
1500                 if (need_resched()) {
1501                         /*
1502                          * We're trying to balance between building up a nice
1503                          * long list of IOs to improve our merge rate, and
1504                          * getting those IOs out quickly for anyone throttling
1505                          * in balance_dirty_pages().  cond_resched() doesn't
1506                          * unplug, so get our IOs out the door before we
1507                          * give up the CPU.
1508                          */
1509                         blk_flush_plug(current);
1510                         cond_resched();
1511                 }
1512
1513
1514                 spin_lock(&wb->list_lock);
1515                 spin_lock(&inode->i_lock);
1516                 if (!(inode->i_state & I_DIRTY_ALL))
1517                         wrote++;
1518                 requeue_inode(inode, wb, &wbc);
1519                 inode_sync_complete(inode);
1520                 spin_unlock(&inode->i_lock);
1521
1522                 /*
1523                  * bail out to wb_writeback() often enough to check
1524                  * background threshold and other termination conditions.
1525                  */
1526                 if (wrote) {
1527                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1528                                 break;
1529                         if (work->nr_pages <= 0)
1530                                 break;
1531                 }
1532         }
1533         return wrote;
1534 }
1535
1536 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1537                                   struct wb_writeback_work *work)
1538 {
1539         unsigned long start_time = jiffies;
1540         long wrote = 0;
1541
1542         while (!list_empty(&wb->b_io)) {
1543                 struct inode *inode = wb_inode(wb->b_io.prev);
1544                 struct super_block *sb = inode->i_sb;
1545
1546                 if (!trylock_super(sb)) {
1547                         /*
1548                          * trylock_super() may fail consistently due to
1549                          * s_umount being grabbed by someone else. Don't use
1550                          * requeue_io() to avoid busy retrying the inode/sb.
1551                          */
1552                         redirty_tail(inode, wb);
1553                         continue;
1554                 }
1555                 wrote += writeback_sb_inodes(sb, wb, work);
1556                 up_read(&sb->s_umount);
1557
1558                 /* refer to the same tests at the end of writeback_sb_inodes */
1559                 if (wrote) {
1560                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1561                                 break;
1562                         if (work->nr_pages <= 0)
1563                                 break;
1564                 }
1565         }
1566         /* Leave any unwritten inodes on b_io */
1567         return wrote;
1568 }
1569
1570 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1571                                 enum wb_reason reason)
1572 {
1573         struct wb_writeback_work work = {
1574                 .nr_pages       = nr_pages,
1575                 .sync_mode      = WB_SYNC_NONE,
1576                 .range_cyclic   = 1,
1577                 .reason         = reason,
1578         };
1579         struct blk_plug plug;
1580
1581         blk_start_plug(&plug);
1582         spin_lock(&wb->list_lock);
1583         if (list_empty(&wb->b_io))
1584                 queue_io(wb, &work);
1585         __writeback_inodes_wb(wb, &work);
1586         spin_unlock(&wb->list_lock);
1587         blk_finish_plug(&plug);
1588
1589         return nr_pages - work.nr_pages;
1590 }
1591
1592 /*
1593  * Explicit flushing or periodic writeback of "old" data.
1594  *
1595  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1596  * dirtying-time in the inode's address_space.  So this periodic writeback code
1597  * just walks the superblock inode list, writing back any inodes which are
1598  * older than a specific point in time.
1599  *
1600  * Try to run once per dirty_writeback_interval.  But if a writeback event
1601  * takes longer than a dirty_writeback_interval interval, then leave a
1602  * one-second gap.
1603  *
1604  * older_than_this takes precedence over nr_to_write.  So we'll only write back
1605  * all dirty pages if they are all attached to "old" mappings.
1606  */
1607 static long wb_writeback(struct bdi_writeback *wb,
1608                          struct wb_writeback_work *work)
1609 {
1610         unsigned long wb_start = jiffies;
1611         long nr_pages = work->nr_pages;
1612         unsigned long oldest_jif;
1613         struct inode *inode;
1614         long progress;
1615         struct blk_plug plug;
1616
1617         oldest_jif = jiffies;
1618         work->older_than_this = &oldest_jif;
1619
1620         blk_start_plug(&plug);
1621         spin_lock(&wb->list_lock);
1622         for (;;) {
1623                 /*
1624                  * Stop writeback when nr_pages has been consumed
1625                  */
1626                 if (work->nr_pages <= 0)
1627                         break;
1628
1629                 /*
1630                  * Background writeout and kupdate-style writeback may
1631                  * run forever. Stop them if there is other work to do
1632                  * so that e.g. sync can proceed. They'll be restarted
1633                  * after the other works are all done.
1634                  */
1635                 if ((work->for_background || work->for_kupdate) &&
1636                     !list_empty(&wb->work_list))
1637                         break;
1638
1639                 /*
1640                  * For background writeout, stop when we are below the
1641                  * background dirty threshold
1642                  */
1643                 if (work->for_background && !wb_over_bg_thresh(wb))
1644                         break;
1645
1646                 /*
1647                  * Kupdate and background works are special and we want to
1648                  * include all inodes that need writing. Livelock avoidance is
1649                  * handled by these works yielding to any other work so we are
1650                  * safe.
1651                  */
1652                 if (work->for_kupdate) {
1653                         oldest_jif = jiffies -
1654                                 msecs_to_jiffies(dirty_expire_interval * 10);
1655                 } else if (work->for_background)
1656                         oldest_jif = jiffies;
1657
1658                 trace_writeback_start(wb, work);
1659                 if (list_empty(&wb->b_io))
1660                         queue_io(wb, work);
1661                 if (work->sb)
1662                         progress = writeback_sb_inodes(work->sb, wb, work);
1663                 else
1664                         progress = __writeback_inodes_wb(wb, work);
1665                 trace_writeback_written(wb, work);
1666
1667                 wb_update_bandwidth(wb, wb_start);
1668
1669                 /*
1670                  * Did we write something? Try for more
1671                  *
1672                  * Dirty inodes are moved to b_io for writeback in batches.
1673                  * The completion of the current batch does not necessarily
1674                  * mean the overall work is done. So we keep looping as long
1675                  * as made some progress on cleaning pages or inodes.
1676                  */
1677                 if (progress)
1678                         continue;
1679                 /*
1680                  * No more inodes for IO, bail
1681                  */
1682                 if (list_empty(&wb->b_more_io))
1683                         break;
1684                 /*
1685                  * Nothing written. Wait for some inode to
1686                  * become available for writeback. Otherwise
1687                  * we'll just busyloop.
1688                  */
1689                 if (!list_empty(&wb->b_more_io))  {
1690                         trace_writeback_wait(wb, work);
1691                         inode = wb_inode(wb->b_more_io.prev);
1692                         spin_lock(&inode->i_lock);
1693                         spin_unlock(&wb->list_lock);
1694                         /* This function drops i_lock... */
1695                         inode_sleep_on_writeback(inode);
1696                         spin_lock(&wb->list_lock);
1697                 }
1698         }
1699         spin_unlock(&wb->list_lock);
1700         blk_finish_plug(&plug);
1701
1702         return nr_pages - work->nr_pages;
1703 }
1704
1705 /*
1706  * Return the next wb_writeback_work struct that hasn't been processed yet.
1707  */
1708 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1709 {
1710         struct wb_writeback_work *work = NULL;
1711
1712         spin_lock_bh(&wb->work_lock);
1713         if (!list_empty(&wb->work_list)) {
1714                 work = list_entry(wb->work_list.next,
1715                                   struct wb_writeback_work, list);
1716                 list_del_init(&work->list);
1717         }
1718         spin_unlock_bh(&wb->work_lock);
1719         return work;
1720 }
1721
1722 /*
1723  * Add in the number of potentially dirty inodes, because each inode
1724  * write can dirty pagecache in the underlying blockdev.
1725  */
1726 static unsigned long get_nr_dirty_pages(void)
1727 {
1728         return global_page_state(NR_FILE_DIRTY) +
1729                 global_page_state(NR_UNSTABLE_NFS) +
1730                 get_nr_dirty_inodes();
1731 }
1732
1733 static long wb_check_background_flush(struct bdi_writeback *wb)
1734 {
1735         if (wb_over_bg_thresh(wb)) {
1736
1737                 struct wb_writeback_work work = {
1738                         .nr_pages       = LONG_MAX,
1739                         .sync_mode      = WB_SYNC_NONE,
1740                         .for_background = 1,
1741                         .range_cyclic   = 1,
1742                         .reason         = WB_REASON_BACKGROUND,
1743                 };
1744
1745                 return wb_writeback(wb, &work);
1746         }
1747
1748         return 0;
1749 }
1750
1751 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1752 {
1753         unsigned long expired;
1754         long nr_pages;
1755
1756         /*
1757          * When set to zero, disable periodic writeback
1758          */
1759         if (!dirty_writeback_interval)
1760                 return 0;
1761
1762         expired = wb->last_old_flush +
1763                         msecs_to_jiffies(dirty_writeback_interval * 10);
1764         if (time_before(jiffies, expired))
1765                 return 0;
1766
1767         wb->last_old_flush = jiffies;
1768         nr_pages = get_nr_dirty_pages();
1769
1770         if (nr_pages) {
1771                 struct wb_writeback_work work = {
1772                         .nr_pages       = nr_pages,
1773                         .sync_mode      = WB_SYNC_NONE,
1774                         .for_kupdate    = 1,
1775                         .range_cyclic   = 1,
1776                         .reason         = WB_REASON_PERIODIC,
1777                 };
1778
1779                 return wb_writeback(wb, &work);
1780         }
1781
1782         return 0;
1783 }
1784
1785 /*
1786  * Retrieve work items and do the writeback they describe
1787  */
1788 static long wb_do_writeback(struct bdi_writeback *wb)
1789 {
1790         struct wb_writeback_work *work;
1791         long wrote = 0;
1792
1793         set_bit(WB_writeback_running, &wb->state);
1794         while ((work = get_next_work_item(wb)) != NULL) {
1795                 struct wb_completion *done = work->done;
1796
1797                 trace_writeback_exec(wb, work);
1798
1799                 wrote += wb_writeback(wb, work);
1800
1801                 if (work->auto_free)
1802                         kfree(work);
1803                 if (done && atomic_dec_and_test(&done->cnt))
1804                         wake_up_all(&wb->bdi->wb_waitq);
1805         }
1806
1807         /*
1808          * Check for periodic writeback, kupdated() style
1809          */
1810         wrote += wb_check_old_data_flush(wb);
1811         wrote += wb_check_background_flush(wb);
1812         clear_bit(WB_writeback_running, &wb->state);
1813
1814         return wrote;
1815 }
1816
1817 /*
1818  * Handle writeback of dirty data for the device backed by this bdi. Also
1819  * reschedules periodically and does kupdated style flushing.
1820  */
1821 void wb_workfn(struct work_struct *work)
1822 {
1823         struct bdi_writeback *wb = container_of(to_delayed_work(work),
1824                                                 struct bdi_writeback, dwork);
1825         long pages_written;
1826
1827         set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1828         current->flags |= PF_SWAPWRITE;
1829
1830         if (likely(!current_is_workqueue_rescuer() ||
1831                    !test_bit(WB_registered, &wb->state))) {
1832                 /*
1833                  * The normal path.  Keep writing back @wb until its
1834                  * work_list is empty.  Note that this path is also taken
1835                  * if @wb is shutting down even when we're running off the
1836                  * rescuer as work_list needs to be drained.
1837                  */
1838                 do {
1839                         pages_written = wb_do_writeback(wb);
1840                         trace_writeback_pages_written(pages_written);
1841                 } while (!list_empty(&wb->work_list));
1842         } else {
1843                 /*
1844                  * bdi_wq can't get enough workers and we're running off
1845                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
1846                  * enough for efficient IO.
1847                  */
1848                 pages_written = writeback_inodes_wb(wb, 1024,
1849                                                     WB_REASON_FORKER_THREAD);
1850                 trace_writeback_pages_written(pages_written);
1851         }
1852
1853         if (!list_empty(&wb->work_list))
1854                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1855         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1856                 wb_wakeup_delayed(wb);
1857
1858         current->flags &= ~PF_SWAPWRITE;
1859 }
1860
1861 /*
1862  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1863  * the whole world.
1864  */
1865 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1866 {
1867         struct backing_dev_info *bdi;
1868
1869         if (!nr_pages)
1870                 nr_pages = get_nr_dirty_pages();
1871
1872         rcu_read_lock();
1873         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1874                 struct bdi_writeback *wb;
1875
1876                 if (!bdi_has_dirty_io(bdi))
1877                         continue;
1878
1879                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
1880                         wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1881                                            false, reason);
1882         }
1883         rcu_read_unlock();
1884 }
1885
1886 /*
1887  * Wake up bdi's periodically to make sure dirtytime inodes gets
1888  * written back periodically.  We deliberately do *not* check the
1889  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1890  * kernel to be constantly waking up once there are any dirtytime
1891  * inodes on the system.  So instead we define a separate delayed work
1892  * function which gets called much more rarely.  (By default, only
1893  * once every 12 hours.)
1894  *
1895  * If there is any other write activity going on in the file system,
1896  * this function won't be necessary.  But if the only thing that has
1897  * happened on the file system is a dirtytime inode caused by an atime
1898  * update, we need this infrastructure below to make sure that inode
1899  * eventually gets pushed out to disk.
1900  */
1901 static void wakeup_dirtytime_writeback(struct work_struct *w);
1902 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1903
1904 static void wakeup_dirtytime_writeback(struct work_struct *w)
1905 {
1906         struct backing_dev_info *bdi;
1907
1908         rcu_read_lock();
1909         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1910                 struct bdi_writeback *wb;
1911
1912                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
1913                         if (!list_empty(&wb->b_dirty_time))
1914                                 wb_wakeup(wb);
1915         }
1916         rcu_read_unlock();
1917         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1918 }
1919
1920 static int __init start_dirtytime_writeback(void)
1921 {
1922         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1923         return 0;
1924 }
1925 __initcall(start_dirtytime_writeback);
1926
1927 int dirtytime_interval_handler(struct ctl_table *table, int write,
1928                                void __user *buffer, size_t *lenp, loff_t *ppos)
1929 {
1930         int ret;
1931
1932         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1933         if (ret == 0 && write)
1934                 mod_delayed_work(system_wq, &dirtytime_work, 0);
1935         return ret;
1936 }
1937
1938 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1939 {
1940         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1941                 struct dentry *dentry;
1942                 const char *name = "?";
1943
1944                 dentry = d_find_alias(inode);
1945                 if (dentry) {
1946                         spin_lock(&dentry->d_lock);
1947                         name = (const char *) dentry->d_name.name;
1948                 }
1949                 printk(KERN_DEBUG
1950                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1951                        current->comm, task_pid_nr(current), inode->i_ino,
1952                        name, inode->i_sb->s_id);
1953                 if (dentry) {
1954                         spin_unlock(&dentry->d_lock);
1955                         dput(dentry);
1956                 }
1957         }
1958 }
1959
1960 /**
1961  *      __mark_inode_dirty -    internal function
1962  *      @inode: inode to mark
1963  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1964  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1965  *      mark_inode_dirty_sync.
1966  *
1967  * Put the inode on the super block's dirty list.
1968  *
1969  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1970  * dirty list only if it is hashed or if it refers to a blockdev.
1971  * If it was not hashed, it will never be added to the dirty list
1972  * even if it is later hashed, as it will have been marked dirty already.
1973  *
1974  * In short, make sure you hash any inodes _before_ you start marking
1975  * them dirty.
1976  *
1977  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1978  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1979  * the kernel-internal blockdev inode represents the dirtying time of the
1980  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1981  * page->mapping->host, so the page-dirtying time is recorded in the internal
1982  * blockdev inode.
1983  */
1984 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1985 void __mark_inode_dirty(struct inode *inode, int flags)
1986 {
1987         struct super_block *sb = inode->i_sb;
1988         int dirtytime;
1989
1990         trace_writeback_mark_inode_dirty(inode, flags);
1991
1992         /*
1993          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1994          * dirty the inode itself
1995          */
1996         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
1997                 trace_writeback_dirty_inode_start(inode, flags);
1998
1999                 if (sb->s_op->dirty_inode)
2000                         sb->s_op->dirty_inode(inode, flags);
2001
2002                 trace_writeback_dirty_inode(inode, flags);
2003         }
2004         if (flags & I_DIRTY_INODE)
2005                 flags &= ~I_DIRTY_TIME;
2006         dirtytime = flags & I_DIRTY_TIME;
2007
2008         /*
2009          * Paired with smp_mb() in __writeback_single_inode() for the
2010          * following lockless i_state test.  See there for details.
2011          */
2012         smp_mb();
2013
2014         if (((inode->i_state & flags) == flags) ||
2015             (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2016                 return;
2017
2018         if (unlikely(block_dump))
2019                 block_dump___mark_inode_dirty(inode);
2020
2021         spin_lock(&inode->i_lock);
2022         if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2023                 goto out_unlock_inode;
2024         if ((inode->i_state & flags) != flags) {
2025                 const int was_dirty = inode->i_state & I_DIRTY;
2026
2027                 inode_attach_wb(inode, NULL);
2028
2029                 if (flags & I_DIRTY_INODE)
2030                         inode->i_state &= ~I_DIRTY_TIME;
2031                 inode->i_state |= flags;
2032
2033                 /*
2034                  * If the inode is being synced, just update its dirty state.
2035                  * The unlocker will place the inode on the appropriate
2036                  * superblock list, based upon its state.
2037                  */
2038                 if (inode->i_state & I_SYNC)
2039                         goto out_unlock_inode;
2040
2041                 /*
2042                  * Only add valid (hashed) inodes to the superblock's
2043                  * dirty list.  Add blockdev inodes as well.
2044                  */
2045                 if (!S_ISBLK(inode->i_mode)) {
2046                         if (inode_unhashed(inode))
2047                                 goto out_unlock_inode;
2048                 }
2049                 if (inode->i_state & I_FREEING)
2050                         goto out_unlock_inode;
2051
2052                 /*
2053                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2054                  * reposition it (that would break b_dirty time-ordering).
2055                  */
2056                 if (!was_dirty) {
2057                         struct bdi_writeback *wb;
2058                         struct list_head *dirty_list;
2059                         bool wakeup_bdi = false;
2060
2061                         wb = locked_inode_to_wb_and_lock_list(inode);
2062
2063                         WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2064                              !test_bit(WB_registered, &wb->state),
2065                              "bdi-%s not registered\n", wb->bdi->name);
2066
2067                         inode->dirtied_when = jiffies;
2068                         if (dirtytime)
2069                                 inode->dirtied_time_when = jiffies;
2070
2071                         if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2072                                 dirty_list = &wb->b_dirty;
2073                         else
2074                                 dirty_list = &wb->b_dirty_time;
2075
2076                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2077                                                                dirty_list);
2078
2079                         spin_unlock(&wb->list_lock);
2080                         trace_writeback_dirty_inode_enqueue(inode);
2081
2082                         /*
2083                          * If this is the first dirty inode for this bdi,
2084                          * we have to wake-up the corresponding bdi thread
2085                          * to make sure background write-back happens
2086                          * later.
2087                          */
2088                         if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2089                                 wb_wakeup_delayed(wb);
2090                         return;
2091                 }
2092         }
2093 out_unlock_inode:
2094         spin_unlock(&inode->i_lock);
2095
2096 }
2097 EXPORT_SYMBOL(__mark_inode_dirty);
2098
2099 /*
2100  * The @s_sync_lock is used to serialise concurrent sync operations
2101  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2102  * Concurrent callers will block on the s_sync_lock rather than doing contending
2103  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2104  * has been issued up to the time this function is enter is guaranteed to be
2105  * completed by the time we have gained the lock and waited for all IO that is
2106  * in progress regardless of the order callers are granted the lock.
2107  */
2108 static void wait_sb_inodes(struct super_block *sb)
2109 {
2110         struct inode *inode, *old_inode = NULL;
2111
2112         /*
2113          * We need to be protected against the filesystem going from
2114          * r/o to r/w or vice versa.
2115          */
2116         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2117
2118         mutex_lock(&sb->s_sync_lock);
2119         spin_lock(&sb->s_inode_list_lock);
2120
2121         /*
2122          * Data integrity sync. Must wait for all pages under writeback,
2123          * because there may have been pages dirtied before our sync
2124          * call, but which had writeout started before we write it out.
2125          * In which case, the inode may not be on the dirty list, but
2126          * we still have to wait for that writeout.
2127          */
2128         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
2129                 struct address_space *mapping = inode->i_mapping;
2130
2131                 spin_lock(&inode->i_lock);
2132                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2133                     (mapping->nrpages == 0)) {
2134                         spin_unlock(&inode->i_lock);
2135                         continue;
2136                 }
2137                 __iget(inode);
2138                 spin_unlock(&inode->i_lock);
2139                 spin_unlock(&sb->s_inode_list_lock);
2140
2141                 /*
2142                  * We hold a reference to 'inode' so it couldn't have been
2143                  * removed from s_inodes list while we dropped the
2144                  * s_inode_list_lock.  We cannot iput the inode now as we can
2145                  * be holding the last reference and we cannot iput it under
2146                  * s_inode_list_lock. So we keep the reference and iput it
2147                  * later.
2148                  */
2149                 iput(old_inode);
2150                 old_inode = inode;
2151
2152                 filemap_fdatawait(mapping);
2153
2154                 cond_resched();
2155
2156                 spin_lock(&sb->s_inode_list_lock);
2157         }
2158         spin_unlock(&sb->s_inode_list_lock);
2159         iput(old_inode);
2160         mutex_unlock(&sb->s_sync_lock);
2161 }
2162
2163 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2164                                      enum wb_reason reason, bool skip_if_busy)
2165 {
2166         DEFINE_WB_COMPLETION_ONSTACK(done);
2167         struct wb_writeback_work work = {
2168                 .sb                     = sb,
2169                 .sync_mode              = WB_SYNC_NONE,
2170                 .tagged_writepages      = 1,
2171                 .done                   = &done,
2172                 .nr_pages               = nr,
2173                 .reason                 = reason,
2174         };
2175         struct backing_dev_info *bdi = sb->s_bdi;
2176
2177         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2178                 return;
2179         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2180
2181         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2182         wb_wait_for_completion(bdi, &done);
2183 }
2184
2185 /**
2186  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2187  * @sb: the superblock
2188  * @nr: the number of pages to write
2189  * @reason: reason why some writeback work initiated
2190  *
2191  * Start writeback on some inodes on this super_block. No guarantees are made
2192  * on how many (if any) will be written, and this function does not wait
2193  * for IO completion of submitted IO.
2194  */
2195 void writeback_inodes_sb_nr(struct super_block *sb,
2196                             unsigned long nr,
2197                             enum wb_reason reason)
2198 {
2199         __writeback_inodes_sb_nr(sb, nr, reason, false);
2200 }
2201 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2202
2203 /**
2204  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2205  * @sb: the superblock
2206  * @reason: reason why some writeback work was initiated
2207  *
2208  * Start writeback on some inodes on this super_block. No guarantees are made
2209  * on how many (if any) will be written, and this function does not wait
2210  * for IO completion of submitted IO.
2211  */
2212 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2213 {
2214         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2215 }
2216 EXPORT_SYMBOL(writeback_inodes_sb);
2217
2218 /**
2219  * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2220  * @sb: the superblock
2221  * @nr: the number of pages to write
2222  * @reason: the reason of writeback
2223  *
2224  * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2225  * Returns 1 if writeback was started, 0 if not.
2226  */
2227 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2228                                    enum wb_reason reason)
2229 {
2230         if (!down_read_trylock(&sb->s_umount))
2231                 return false;
2232
2233         __writeback_inodes_sb_nr(sb, nr, reason, true);
2234         up_read(&sb->s_umount);
2235         return true;
2236 }
2237 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2238
2239 /**
2240  * try_to_writeback_inodes_sb - try to start writeback if none underway
2241  * @sb: the superblock
2242  * @reason: reason why some writeback work was initiated
2243  *
2244  * Implement by try_to_writeback_inodes_sb_nr()
2245  * Returns 1 if writeback was started, 0 if not.
2246  */
2247 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2248 {
2249         return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2250 }
2251 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2252
2253 /**
2254  * sync_inodes_sb       -       sync sb inode pages
2255  * @sb: the superblock
2256  *
2257  * This function writes and waits on any dirty inode belonging to this
2258  * super_block.
2259  */
2260 void sync_inodes_sb(struct super_block *sb)
2261 {
2262         DEFINE_WB_COMPLETION_ONSTACK(done);
2263         struct wb_writeback_work work = {
2264                 .sb             = sb,
2265                 .sync_mode      = WB_SYNC_ALL,
2266                 .nr_pages       = LONG_MAX,
2267                 .range_cyclic   = 0,
2268                 .done           = &done,
2269                 .reason         = WB_REASON_SYNC,
2270                 .for_sync       = 1,
2271         };
2272         struct backing_dev_info *bdi = sb->s_bdi;
2273
2274         /*
2275          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2276          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2277          * bdi_has_dirty() need to be written out too.
2278          */
2279         if (bdi == &noop_backing_dev_info)
2280                 return;
2281         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2282
2283         bdi_split_work_to_wbs(bdi, &work, false);
2284         wb_wait_for_completion(bdi, &done);
2285
2286         wait_sb_inodes(sb);
2287 }
2288 EXPORT_SYMBOL(sync_inodes_sb);
2289
2290 /**
2291  * write_inode_now      -       write an inode to disk
2292  * @inode: inode to write to disk
2293  * @sync: whether the write should be synchronous or not
2294  *
2295  * This function commits an inode to disk immediately if it is dirty. This is
2296  * primarily needed by knfsd.
2297  *
2298  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2299  */
2300 int write_inode_now(struct inode *inode, int sync)
2301 {
2302         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
2303         struct writeback_control wbc = {
2304                 .nr_to_write = LONG_MAX,
2305                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2306                 .range_start = 0,
2307                 .range_end = LLONG_MAX,
2308         };
2309
2310         if (!mapping_cap_writeback_dirty(inode->i_mapping))
2311                 wbc.nr_to_write = 0;
2312
2313         might_sleep();
2314         return writeback_single_inode(inode, wb, &wbc);
2315 }
2316 EXPORT_SYMBOL(write_inode_now);
2317
2318 /**
2319  * sync_inode - write an inode and its pages to disk.
2320  * @inode: the inode to sync
2321  * @wbc: controls the writeback mode
2322  *
2323  * sync_inode() will write an inode and its pages to disk.  It will also
2324  * correctly update the inode on its superblock's dirty inode lists and will
2325  * update inode->i_state.
2326  *
2327  * The caller must have a ref on the inode.
2328  */
2329 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2330 {
2331         return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
2332 }
2333 EXPORT_SYMBOL(sync_inode);
2334
2335 /**
2336  * sync_inode_metadata - write an inode to disk
2337  * @inode: the inode to sync
2338  * @wait: wait for I/O to complete.
2339  *
2340  * Write an inode to disk and adjust its dirty state after completion.
2341  *
2342  * Note: only writes the actual inode, no associated data or other metadata.
2343  */
2344 int sync_inode_metadata(struct inode *inode, int wait)
2345 {
2346         struct writeback_control wbc = {
2347                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2348                 .nr_to_write = 0, /* metadata-only */
2349         };
2350
2351         return sync_inode(inode, &wbc);
2352 }
2353 EXPORT_SYMBOL(sync_inode_metadata);