]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/fs-writeback.c
Merge tag 'efi-next' of git://git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi...
[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         int next_memcg_id = 0;
782         struct bdi_writeback *wb;
783         struct wb_iter iter;
784
785         might_sleep();
786 restart:
787         rcu_read_lock();
788         bdi_for_each_wb(wb, bdi, &iter, next_memcg_id) {
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                 /* SYNC_ALL writes out I_DIRTY_TIME too */
795                 if (!wb_has_dirty_io(wb) &&
796                     (base_work->sync_mode == WB_SYNC_NONE ||
797                      list_empty(&wb->b_dirty_time)))
798                         continue;
799                 if (skip_if_busy && writeback_in_progress(wb))
800                         continue;
801
802                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
803
804                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
805                 if (work) {
806                         *work = *base_work;
807                         work->nr_pages = nr_pages;
808                         work->auto_free = 1;
809                         wb_queue_work(wb, work);
810                         continue;
811                 }
812
813                 /* alloc failed, execute synchronously using on-stack fallback */
814                 work = &fallback_work;
815                 *work = *base_work;
816                 work->nr_pages = nr_pages;
817                 work->auto_free = 0;
818                 work->done = &fallback_work_done;
819
820                 wb_queue_work(wb, work);
821
822                 next_memcg_id = wb->memcg_css->id + 1;
823                 rcu_read_unlock();
824                 wb_wait_for_completion(bdi, &fallback_work_done);
825                 goto restart;
826         }
827         rcu_read_unlock();
828 }
829
830 #else   /* CONFIG_CGROUP_WRITEBACK */
831
832 static struct bdi_writeback *
833 locked_inode_to_wb_and_lock_list(struct inode *inode)
834         __releases(&inode->i_lock)
835         __acquires(&wb->list_lock)
836 {
837         struct bdi_writeback *wb = inode_to_wb(inode);
838
839         spin_unlock(&inode->i_lock);
840         spin_lock(&wb->list_lock);
841         return wb;
842 }
843
844 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
845         __acquires(&wb->list_lock)
846 {
847         struct bdi_writeback *wb = inode_to_wb(inode);
848
849         spin_lock(&wb->list_lock);
850         return wb;
851 }
852
853 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
854 {
855         return nr_pages;
856 }
857
858 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
859                                   struct wb_writeback_work *base_work,
860                                   bool skip_if_busy)
861 {
862         might_sleep();
863
864         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
865                 base_work->auto_free = 0;
866                 wb_queue_work(&bdi->wb, base_work);
867         }
868 }
869
870 #endif  /* CONFIG_CGROUP_WRITEBACK */
871
872 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
873                         bool range_cyclic, enum wb_reason reason)
874 {
875         struct wb_writeback_work *work;
876
877         if (!wb_has_dirty_io(wb))
878                 return;
879
880         /*
881          * This is WB_SYNC_NONE writeback, so if allocation fails just
882          * wakeup the thread for old dirty data writeback
883          */
884         work = kzalloc(sizeof(*work), GFP_ATOMIC);
885         if (!work) {
886                 trace_writeback_nowork(wb);
887                 wb_wakeup(wb);
888                 return;
889         }
890
891         work->sync_mode = WB_SYNC_NONE;
892         work->nr_pages  = nr_pages;
893         work->range_cyclic = range_cyclic;
894         work->reason    = reason;
895         work->auto_free = 1;
896
897         wb_queue_work(wb, work);
898 }
899
900 /**
901  * wb_start_background_writeback - start background writeback
902  * @wb: bdi_writback to write from
903  *
904  * Description:
905  *   This makes sure WB_SYNC_NONE background writeback happens. When
906  *   this function returns, it is only guaranteed that for given wb
907  *   some IO is happening if we are over background dirty threshold.
908  *   Caller need not hold sb s_umount semaphore.
909  */
910 void wb_start_background_writeback(struct bdi_writeback *wb)
911 {
912         /*
913          * We just wake up the flusher thread. It will perform background
914          * writeback as soon as there is no other work to do.
915          */
916         trace_writeback_wake_background(wb);
917         wb_wakeup(wb);
918 }
919
920 /*
921  * Remove the inode from the writeback list it is on.
922  */
923 void inode_io_list_del(struct inode *inode)
924 {
925         struct bdi_writeback *wb;
926
927         wb = inode_to_wb_and_lock_list(inode);
928         inode_io_list_del_locked(inode, wb);
929         spin_unlock(&wb->list_lock);
930 }
931
932 /*
933  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
934  * furthest end of its superblock's dirty-inode list.
935  *
936  * Before stamping the inode's ->dirtied_when, we check to see whether it is
937  * already the most-recently-dirtied inode on the b_dirty list.  If that is
938  * the case then the inode must have been redirtied while it was being written
939  * out and we don't reset its dirtied_when.
940  */
941 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
942 {
943         if (!list_empty(&wb->b_dirty)) {
944                 struct inode *tail;
945
946                 tail = wb_inode(wb->b_dirty.next);
947                 if (time_before(inode->dirtied_when, tail->dirtied_when))
948                         inode->dirtied_when = jiffies;
949         }
950         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
951 }
952
953 /*
954  * requeue inode for re-scanning after bdi->b_io list is exhausted.
955  */
956 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
957 {
958         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
959 }
960
961 static void inode_sync_complete(struct inode *inode)
962 {
963         inode->i_state &= ~I_SYNC;
964         /* If inode is clean an unused, put it into LRU now... */
965         inode_add_lru(inode);
966         /* Waiters must see I_SYNC cleared before being woken up */
967         smp_mb();
968         wake_up_bit(&inode->i_state, __I_SYNC);
969 }
970
971 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
972 {
973         bool ret = time_after(inode->dirtied_when, t);
974 #ifndef CONFIG_64BIT
975         /*
976          * For inodes being constantly redirtied, dirtied_when can get stuck.
977          * It _appears_ to be in the future, but is actually in distant past.
978          * This test is necessary to prevent such wrapped-around relative times
979          * from permanently stopping the whole bdi writeback.
980          */
981         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
982 #endif
983         return ret;
984 }
985
986 #define EXPIRE_DIRTY_ATIME 0x0001
987
988 /*
989  * Move expired (dirtied before work->older_than_this) dirty inodes from
990  * @delaying_queue to @dispatch_queue.
991  */
992 static int move_expired_inodes(struct list_head *delaying_queue,
993                                struct list_head *dispatch_queue,
994                                int flags,
995                                struct wb_writeback_work *work)
996 {
997         unsigned long *older_than_this = NULL;
998         unsigned long expire_time;
999         LIST_HEAD(tmp);
1000         struct list_head *pos, *node;
1001         struct super_block *sb = NULL;
1002         struct inode *inode;
1003         int do_sb_sort = 0;
1004         int moved = 0;
1005
1006         if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1007                 older_than_this = work->older_than_this;
1008         else if (!work->for_sync) {
1009                 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1010                 older_than_this = &expire_time;
1011         }
1012         while (!list_empty(delaying_queue)) {
1013                 inode = wb_inode(delaying_queue->prev);
1014                 if (older_than_this &&
1015                     inode_dirtied_after(inode, *older_than_this))
1016                         break;
1017                 list_move(&inode->i_io_list, &tmp);
1018                 moved++;
1019                 if (flags & EXPIRE_DIRTY_ATIME)
1020                         set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1021                 if (sb_is_blkdev_sb(inode->i_sb))
1022                         continue;
1023                 if (sb && sb != inode->i_sb)
1024                         do_sb_sort = 1;
1025                 sb = inode->i_sb;
1026         }
1027
1028         /* just one sb in list, splice to dispatch_queue and we're done */
1029         if (!do_sb_sort) {
1030                 list_splice(&tmp, dispatch_queue);
1031                 goto out;
1032         }
1033
1034         /* Move inodes from one superblock together */
1035         while (!list_empty(&tmp)) {
1036                 sb = wb_inode(tmp.prev)->i_sb;
1037                 list_for_each_prev_safe(pos, node, &tmp) {
1038                         inode = wb_inode(pos);
1039                         if (inode->i_sb == sb)
1040                                 list_move(&inode->i_io_list, dispatch_queue);
1041                 }
1042         }
1043 out:
1044         return moved;
1045 }
1046
1047 /*
1048  * Queue all expired dirty inodes for io, eldest first.
1049  * Before
1050  *         newly dirtied     b_dirty    b_io    b_more_io
1051  *         =============>    gf         edc     BA
1052  * After
1053  *         newly dirtied     b_dirty    b_io    b_more_io
1054  *         =============>    g          fBAedc
1055  *                                           |
1056  *                                           +--> dequeue for IO
1057  */
1058 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1059 {
1060         int moved;
1061
1062         assert_spin_locked(&wb->list_lock);
1063         list_splice_init(&wb->b_more_io, &wb->b_io);
1064         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1065         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1066                                      EXPIRE_DIRTY_ATIME, work);
1067         if (moved)
1068                 wb_io_lists_populated(wb);
1069         trace_writeback_queue_io(wb, work, moved);
1070 }
1071
1072 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1073 {
1074         int ret;
1075
1076         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1077                 trace_writeback_write_inode_start(inode, wbc);
1078                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1079                 trace_writeback_write_inode(inode, wbc);
1080                 return ret;
1081         }
1082         return 0;
1083 }
1084
1085 /*
1086  * Wait for writeback on an inode to complete. Called with i_lock held.
1087  * Caller must make sure inode cannot go away when we drop i_lock.
1088  */
1089 static void __inode_wait_for_writeback(struct inode *inode)
1090         __releases(inode->i_lock)
1091         __acquires(inode->i_lock)
1092 {
1093         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1094         wait_queue_head_t *wqh;
1095
1096         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1097         while (inode->i_state & I_SYNC) {
1098                 spin_unlock(&inode->i_lock);
1099                 __wait_on_bit(wqh, &wq, bit_wait,
1100                               TASK_UNINTERRUPTIBLE);
1101                 spin_lock(&inode->i_lock);
1102         }
1103 }
1104
1105 /*
1106  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1107  */
1108 void inode_wait_for_writeback(struct inode *inode)
1109 {
1110         spin_lock(&inode->i_lock);
1111         __inode_wait_for_writeback(inode);
1112         spin_unlock(&inode->i_lock);
1113 }
1114
1115 /*
1116  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1117  * held and drops it. It is aimed for callers not holding any inode reference
1118  * so once i_lock is dropped, inode can go away.
1119  */
1120 static void inode_sleep_on_writeback(struct inode *inode)
1121         __releases(inode->i_lock)
1122 {
1123         DEFINE_WAIT(wait);
1124         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1125         int sleep;
1126
1127         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1128         sleep = inode->i_state & I_SYNC;
1129         spin_unlock(&inode->i_lock);
1130         if (sleep)
1131                 schedule();
1132         finish_wait(wqh, &wait);
1133 }
1134
1135 /*
1136  * Find proper writeback list for the inode depending on its current state and
1137  * possibly also change of its state while we were doing writeback.  Here we
1138  * handle things such as livelock prevention or fairness of writeback among
1139  * inodes. This function can be called only by flusher thread - noone else
1140  * processes all inodes in writeback lists and requeueing inodes behind flusher
1141  * thread's back can have unexpected consequences.
1142  */
1143 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1144                           struct writeback_control *wbc)
1145 {
1146         if (inode->i_state & I_FREEING)
1147                 return;
1148
1149         /*
1150          * Sync livelock prevention. Each inode is tagged and synced in one
1151          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1152          * the dirty time to prevent enqueue and sync it again.
1153          */
1154         if ((inode->i_state & I_DIRTY) &&
1155             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1156                 inode->dirtied_when = jiffies;
1157
1158         if (wbc->pages_skipped) {
1159                 /*
1160                  * writeback is not making progress due to locked
1161                  * buffers. Skip this inode for now.
1162                  */
1163                 redirty_tail(inode, wb);
1164                 return;
1165         }
1166
1167         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1168                 /*
1169                  * We didn't write back all the pages.  nfs_writepages()
1170                  * sometimes bales out without doing anything.
1171                  */
1172                 if (wbc->nr_to_write <= 0) {
1173                         /* Slice used up. Queue for next turn. */
1174                         requeue_io(inode, wb);
1175                 } else {
1176                         /*
1177                          * Writeback blocked by something other than
1178                          * congestion. Delay the inode for some time to
1179                          * avoid spinning on the CPU (100% iowait)
1180                          * retrying writeback of the dirty page/inode
1181                          * that cannot be performed immediately.
1182                          */
1183                         redirty_tail(inode, wb);
1184                 }
1185         } else if (inode->i_state & I_DIRTY) {
1186                 /*
1187                  * Filesystems can dirty the inode during writeback operations,
1188                  * such as delayed allocation during submission or metadata
1189                  * updates after data IO completion.
1190                  */
1191                 redirty_tail(inode, wb);
1192         } else if (inode->i_state & I_DIRTY_TIME) {
1193                 inode->dirtied_when = jiffies;
1194                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1195         } else {
1196                 /* The inode is clean. Remove from writeback lists. */
1197                 inode_io_list_del_locked(inode, wb);
1198         }
1199 }
1200
1201 /*
1202  * Write out an inode and its dirty pages. Do not update the writeback list
1203  * linkage. That is left to the caller. The caller is also responsible for
1204  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1205  */
1206 static int
1207 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1208 {
1209         struct address_space *mapping = inode->i_mapping;
1210         long nr_to_write = wbc->nr_to_write;
1211         unsigned dirty;
1212         int ret;
1213
1214         WARN_ON(!(inode->i_state & I_SYNC));
1215
1216         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1217
1218         ret = do_writepages(mapping, wbc);
1219
1220         /*
1221          * Make sure to wait on the data before writing out the metadata.
1222          * This is important for filesystems that modify metadata on data
1223          * I/O completion. We don't do it for sync(2) writeback because it has a
1224          * separate, external IO completion path and ->sync_fs for guaranteeing
1225          * inode metadata is written back correctly.
1226          */
1227         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1228                 int err = filemap_fdatawait(mapping);
1229                 if (ret == 0)
1230                         ret = err;
1231         }
1232
1233         /*
1234          * Some filesystems may redirty the inode during the writeback
1235          * due to delalloc, clear dirty metadata flags right before
1236          * write_inode()
1237          */
1238         spin_lock(&inode->i_lock);
1239
1240         dirty = inode->i_state & I_DIRTY;
1241         if (inode->i_state & I_DIRTY_TIME) {
1242                 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1243                     unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1244                     unlikely(time_after(jiffies,
1245                                         (inode->dirtied_time_when +
1246                                          dirtytime_expire_interval * HZ)))) {
1247                         dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1248                         trace_writeback_lazytime(inode);
1249                 }
1250         } else
1251                 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1252         inode->i_state &= ~dirty;
1253
1254         /*
1255          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1256          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1257          * either they see the I_DIRTY bits cleared or we see the dirtied
1258          * inode.
1259          *
1260          * I_DIRTY_PAGES is always cleared together above even if @mapping
1261          * still has dirty pages.  The flag is reinstated after smp_mb() if
1262          * necessary.  This guarantees that either __mark_inode_dirty()
1263          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1264          */
1265         smp_mb();
1266
1267         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1268                 inode->i_state |= I_DIRTY_PAGES;
1269
1270         spin_unlock(&inode->i_lock);
1271
1272         if (dirty & I_DIRTY_TIME)
1273                 mark_inode_dirty_sync(inode);
1274         /* Don't write the inode if only I_DIRTY_PAGES was set */
1275         if (dirty & ~I_DIRTY_PAGES) {
1276                 int err = write_inode(inode, wbc);
1277                 if (ret == 0)
1278                         ret = err;
1279         }
1280         trace_writeback_single_inode(inode, wbc, nr_to_write);
1281         return ret;
1282 }
1283
1284 /*
1285  * Write out an inode's dirty pages. Either the caller has an active reference
1286  * on the inode or the inode has I_WILL_FREE set.
1287  *
1288  * This function is designed to be called for writing back one inode which
1289  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1290  * and does more profound writeback list handling in writeback_sb_inodes().
1291  */
1292 static int
1293 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1294                        struct writeback_control *wbc)
1295 {
1296         int ret = 0;
1297
1298         spin_lock(&inode->i_lock);
1299         if (!atomic_read(&inode->i_count))
1300                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1301         else
1302                 WARN_ON(inode->i_state & I_WILL_FREE);
1303
1304         if (inode->i_state & I_SYNC) {
1305                 if (wbc->sync_mode != WB_SYNC_ALL)
1306                         goto out;
1307                 /*
1308                  * It's a data-integrity sync. We must wait. Since callers hold
1309                  * inode reference or inode has I_WILL_FREE set, it cannot go
1310                  * away under us.
1311                  */
1312                 __inode_wait_for_writeback(inode);
1313         }
1314         WARN_ON(inode->i_state & I_SYNC);
1315         /*
1316          * Skip inode if it is clean and we have no outstanding writeback in
1317          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1318          * function since flusher thread may be doing for example sync in
1319          * parallel and if we move the inode, it could get skipped. So here we
1320          * make sure inode is on some writeback list and leave it there unless
1321          * we have completely cleaned the inode.
1322          */
1323         if (!(inode->i_state & I_DIRTY_ALL) &&
1324             (wbc->sync_mode != WB_SYNC_ALL ||
1325              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1326                 goto out;
1327         inode->i_state |= I_SYNC;
1328         wbc_attach_and_unlock_inode(wbc, inode);
1329
1330         ret = __writeback_single_inode(inode, wbc);
1331
1332         wbc_detach_inode(wbc);
1333         spin_lock(&wb->list_lock);
1334         spin_lock(&inode->i_lock);
1335         /*
1336          * If inode is clean, remove it from writeback lists. Otherwise don't
1337          * touch it. See comment above for explanation.
1338          */
1339         if (!(inode->i_state & I_DIRTY_ALL))
1340                 inode_io_list_del_locked(inode, wb);
1341         spin_unlock(&wb->list_lock);
1342         inode_sync_complete(inode);
1343 out:
1344         spin_unlock(&inode->i_lock);
1345         return ret;
1346 }
1347
1348 static long writeback_chunk_size(struct bdi_writeback *wb,
1349                                  struct wb_writeback_work *work)
1350 {
1351         long pages;
1352
1353         /*
1354          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1355          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1356          * here avoids calling into writeback_inodes_wb() more than once.
1357          *
1358          * The intended call sequence for WB_SYNC_ALL writeback is:
1359          *
1360          *      wb_writeback()
1361          *          writeback_sb_inodes()       <== called only once
1362          *              write_cache_pages()     <== called once for each inode
1363          *                   (quickly) tag currently dirty pages
1364          *                   (maybe slowly) sync all tagged pages
1365          */
1366         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1367                 pages = LONG_MAX;
1368         else {
1369                 pages = min(wb->avg_write_bandwidth / 2,
1370                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1371                 pages = min(pages, work->nr_pages);
1372                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1373                                    MIN_WRITEBACK_PAGES);
1374         }
1375
1376         return pages;
1377 }
1378
1379 /*
1380  * Write a portion of b_io inodes which belong to @sb.
1381  *
1382  * Return the number of pages and/or inodes written.
1383  *
1384  * NOTE! This is called with wb->list_lock held, and will
1385  * unlock and relock that for each inode it ends up doing
1386  * IO for.
1387  */
1388 static long writeback_sb_inodes(struct super_block *sb,
1389                                 struct bdi_writeback *wb,
1390                                 struct wb_writeback_work *work)
1391 {
1392         struct writeback_control wbc = {
1393                 .sync_mode              = work->sync_mode,
1394                 .tagged_writepages      = work->tagged_writepages,
1395                 .for_kupdate            = work->for_kupdate,
1396                 .for_background         = work->for_background,
1397                 .for_sync               = work->for_sync,
1398                 .range_cyclic           = work->range_cyclic,
1399                 .range_start            = 0,
1400                 .range_end              = LLONG_MAX,
1401         };
1402         unsigned long start_time = jiffies;
1403         long write_chunk;
1404         long wrote = 0;  /* count both pages and inodes */
1405
1406         while (!list_empty(&wb->b_io)) {
1407                 struct inode *inode = wb_inode(wb->b_io.prev);
1408
1409                 if (inode->i_sb != sb) {
1410                         if (work->sb) {
1411                                 /*
1412                                  * We only want to write back data for this
1413                                  * superblock, move all inodes not belonging
1414                                  * to it back onto the dirty list.
1415                                  */
1416                                 redirty_tail(inode, wb);
1417                                 continue;
1418                         }
1419
1420                         /*
1421                          * The inode belongs to a different superblock.
1422                          * Bounce back to the caller to unpin this and
1423                          * pin the next superblock.
1424                          */
1425                         break;
1426                 }
1427
1428                 /*
1429                  * Don't bother with new inodes or inodes being freed, first
1430                  * kind does not need periodic writeout yet, and for the latter
1431                  * kind writeout is handled by the freer.
1432                  */
1433                 spin_lock(&inode->i_lock);
1434                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1435                         spin_unlock(&inode->i_lock);
1436                         redirty_tail(inode, wb);
1437                         continue;
1438                 }
1439                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1440                         /*
1441                          * If this inode is locked for writeback and we are not
1442                          * doing writeback-for-data-integrity, move it to
1443                          * b_more_io so that writeback can proceed with the
1444                          * other inodes on s_io.
1445                          *
1446                          * We'll have another go at writing back this inode
1447                          * when we completed a full scan of b_io.
1448                          */
1449                         spin_unlock(&inode->i_lock);
1450                         requeue_io(inode, wb);
1451                         trace_writeback_sb_inodes_requeue(inode);
1452                         continue;
1453                 }
1454                 spin_unlock(&wb->list_lock);
1455
1456                 /*
1457                  * We already requeued the inode if it had I_SYNC set and we
1458                  * are doing WB_SYNC_NONE writeback. So this catches only the
1459                  * WB_SYNC_ALL case.
1460                  */
1461                 if (inode->i_state & I_SYNC) {
1462                         /* Wait for I_SYNC. This function drops i_lock... */
1463                         inode_sleep_on_writeback(inode);
1464                         /* Inode may be gone, start again */
1465                         spin_lock(&wb->list_lock);
1466                         continue;
1467                 }
1468                 inode->i_state |= I_SYNC;
1469                 wbc_attach_and_unlock_inode(&wbc, inode);
1470
1471                 write_chunk = writeback_chunk_size(wb, work);
1472                 wbc.nr_to_write = write_chunk;
1473                 wbc.pages_skipped = 0;
1474
1475                 /*
1476                  * We use I_SYNC to pin the inode in memory. While it is set
1477                  * evict_inode() will wait so the inode cannot be freed.
1478                  */
1479                 __writeback_single_inode(inode, &wbc);
1480
1481                 wbc_detach_inode(&wbc);
1482                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1483                 wrote += write_chunk - wbc.nr_to_write;
1484
1485                 if (need_resched()) {
1486                         /*
1487                          * We're trying to balance between building up a nice
1488                          * long list of IOs to improve our merge rate, and
1489                          * getting those IOs out quickly for anyone throttling
1490                          * in balance_dirty_pages().  cond_resched() doesn't
1491                          * unplug, so get our IOs out the door before we
1492                          * give up the CPU.
1493                          */
1494                         blk_flush_plug(current);
1495                         cond_resched();
1496                 }
1497
1498
1499                 spin_lock(&wb->list_lock);
1500                 spin_lock(&inode->i_lock);
1501                 if (!(inode->i_state & I_DIRTY_ALL))
1502                         wrote++;
1503                 requeue_inode(inode, wb, &wbc);
1504                 inode_sync_complete(inode);
1505                 spin_unlock(&inode->i_lock);
1506
1507                 /*
1508                  * bail out to wb_writeback() often enough to check
1509                  * background threshold and other termination conditions.
1510                  */
1511                 if (wrote) {
1512                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1513                                 break;
1514                         if (work->nr_pages <= 0)
1515                                 break;
1516                 }
1517         }
1518         return wrote;
1519 }
1520
1521 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1522                                   struct wb_writeback_work *work)
1523 {
1524         unsigned long start_time = jiffies;
1525         long wrote = 0;
1526
1527         while (!list_empty(&wb->b_io)) {
1528                 struct inode *inode = wb_inode(wb->b_io.prev);
1529                 struct super_block *sb = inode->i_sb;
1530
1531                 if (!trylock_super(sb)) {
1532                         /*
1533                          * trylock_super() may fail consistently due to
1534                          * s_umount being grabbed by someone else. Don't use
1535                          * requeue_io() to avoid busy retrying the inode/sb.
1536                          */
1537                         redirty_tail(inode, wb);
1538                         continue;
1539                 }
1540                 wrote += writeback_sb_inodes(sb, wb, work);
1541                 up_read(&sb->s_umount);
1542
1543                 /* refer to the same tests at the end of writeback_sb_inodes */
1544                 if (wrote) {
1545                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1546                                 break;
1547                         if (work->nr_pages <= 0)
1548                                 break;
1549                 }
1550         }
1551         /* Leave any unwritten inodes on b_io */
1552         return wrote;
1553 }
1554
1555 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1556                                 enum wb_reason reason)
1557 {
1558         struct wb_writeback_work work = {
1559                 .nr_pages       = nr_pages,
1560                 .sync_mode      = WB_SYNC_NONE,
1561                 .range_cyclic   = 1,
1562                 .reason         = reason,
1563         };
1564         struct blk_plug plug;
1565
1566         blk_start_plug(&plug);
1567         spin_lock(&wb->list_lock);
1568         if (list_empty(&wb->b_io))
1569                 queue_io(wb, &work);
1570         __writeback_inodes_wb(wb, &work);
1571         spin_unlock(&wb->list_lock);
1572         blk_finish_plug(&plug);
1573
1574         return nr_pages - work.nr_pages;
1575 }
1576
1577 /*
1578  * Explicit flushing or periodic writeback of "old" data.
1579  *
1580  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1581  * dirtying-time in the inode's address_space.  So this periodic writeback code
1582  * just walks the superblock inode list, writing back any inodes which are
1583  * older than a specific point in time.
1584  *
1585  * Try to run once per dirty_writeback_interval.  But if a writeback event
1586  * takes longer than a dirty_writeback_interval interval, then leave a
1587  * one-second gap.
1588  *
1589  * older_than_this takes precedence over nr_to_write.  So we'll only write back
1590  * all dirty pages if they are all attached to "old" mappings.
1591  */
1592 static long wb_writeback(struct bdi_writeback *wb,
1593                          struct wb_writeback_work *work)
1594 {
1595         unsigned long wb_start = jiffies;
1596         long nr_pages = work->nr_pages;
1597         unsigned long oldest_jif;
1598         struct inode *inode;
1599         long progress;
1600         struct blk_plug plug;
1601
1602         oldest_jif = jiffies;
1603         work->older_than_this = &oldest_jif;
1604
1605         blk_start_plug(&plug);
1606         spin_lock(&wb->list_lock);
1607         for (;;) {
1608                 /*
1609                  * Stop writeback when nr_pages has been consumed
1610                  */
1611                 if (work->nr_pages <= 0)
1612                         break;
1613
1614                 /*
1615                  * Background writeout and kupdate-style writeback may
1616                  * run forever. Stop them if there is other work to do
1617                  * so that e.g. sync can proceed. They'll be restarted
1618                  * after the other works are all done.
1619                  */
1620                 if ((work->for_background || work->for_kupdate) &&
1621                     !list_empty(&wb->work_list))
1622                         break;
1623
1624                 /*
1625                  * For background writeout, stop when we are below the
1626                  * background dirty threshold
1627                  */
1628                 if (work->for_background && !wb_over_bg_thresh(wb))
1629                         break;
1630
1631                 /*
1632                  * Kupdate and background works are special and we want to
1633                  * include all inodes that need writing. Livelock avoidance is
1634                  * handled by these works yielding to any other work so we are
1635                  * safe.
1636                  */
1637                 if (work->for_kupdate) {
1638                         oldest_jif = jiffies -
1639                                 msecs_to_jiffies(dirty_expire_interval * 10);
1640                 } else if (work->for_background)
1641                         oldest_jif = jiffies;
1642
1643                 trace_writeback_start(wb, work);
1644                 if (list_empty(&wb->b_io))
1645                         queue_io(wb, work);
1646                 if (work->sb)
1647                         progress = writeback_sb_inodes(work->sb, wb, work);
1648                 else
1649                         progress = __writeback_inodes_wb(wb, work);
1650                 trace_writeback_written(wb, work);
1651
1652                 wb_update_bandwidth(wb, wb_start);
1653
1654                 /*
1655                  * Did we write something? Try for more
1656                  *
1657                  * Dirty inodes are moved to b_io for writeback in batches.
1658                  * The completion of the current batch does not necessarily
1659                  * mean the overall work is done. So we keep looping as long
1660                  * as made some progress on cleaning pages or inodes.
1661                  */
1662                 if (progress)
1663                         continue;
1664                 /*
1665                  * No more inodes for IO, bail
1666                  */
1667                 if (list_empty(&wb->b_more_io))
1668                         break;
1669                 /*
1670                  * Nothing written. Wait for some inode to
1671                  * become available for writeback. Otherwise
1672                  * we'll just busyloop.
1673                  */
1674                 if (!list_empty(&wb->b_more_io))  {
1675                         trace_writeback_wait(wb, work);
1676                         inode = wb_inode(wb->b_more_io.prev);
1677                         spin_lock(&inode->i_lock);
1678                         spin_unlock(&wb->list_lock);
1679                         /* This function drops i_lock... */
1680                         inode_sleep_on_writeback(inode);
1681                         spin_lock(&wb->list_lock);
1682                 }
1683         }
1684         spin_unlock(&wb->list_lock);
1685         blk_finish_plug(&plug);
1686
1687         return nr_pages - work->nr_pages;
1688 }
1689
1690 /*
1691  * Return the next wb_writeback_work struct that hasn't been processed yet.
1692  */
1693 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1694 {
1695         struct wb_writeback_work *work = NULL;
1696
1697         spin_lock_bh(&wb->work_lock);
1698         if (!list_empty(&wb->work_list)) {
1699                 work = list_entry(wb->work_list.next,
1700                                   struct wb_writeback_work, list);
1701                 list_del_init(&work->list);
1702         }
1703         spin_unlock_bh(&wb->work_lock);
1704         return work;
1705 }
1706
1707 /*
1708  * Add in the number of potentially dirty inodes, because each inode
1709  * write can dirty pagecache in the underlying blockdev.
1710  */
1711 static unsigned long get_nr_dirty_pages(void)
1712 {
1713         return global_page_state(NR_FILE_DIRTY) +
1714                 global_page_state(NR_UNSTABLE_NFS) +
1715                 get_nr_dirty_inodes();
1716 }
1717
1718 static long wb_check_background_flush(struct bdi_writeback *wb)
1719 {
1720         if (wb_over_bg_thresh(wb)) {
1721
1722                 struct wb_writeback_work work = {
1723                         .nr_pages       = LONG_MAX,
1724                         .sync_mode      = WB_SYNC_NONE,
1725                         .for_background = 1,
1726                         .range_cyclic   = 1,
1727                         .reason         = WB_REASON_BACKGROUND,
1728                 };
1729
1730                 return wb_writeback(wb, &work);
1731         }
1732
1733         return 0;
1734 }
1735
1736 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1737 {
1738         unsigned long expired;
1739         long nr_pages;
1740
1741         /*
1742          * When set to zero, disable periodic writeback
1743          */
1744         if (!dirty_writeback_interval)
1745                 return 0;
1746
1747         expired = wb->last_old_flush +
1748                         msecs_to_jiffies(dirty_writeback_interval * 10);
1749         if (time_before(jiffies, expired))
1750                 return 0;
1751
1752         wb->last_old_flush = jiffies;
1753         nr_pages = get_nr_dirty_pages();
1754
1755         if (nr_pages) {
1756                 struct wb_writeback_work work = {
1757                         .nr_pages       = nr_pages,
1758                         .sync_mode      = WB_SYNC_NONE,
1759                         .for_kupdate    = 1,
1760                         .range_cyclic   = 1,
1761                         .reason         = WB_REASON_PERIODIC,
1762                 };
1763
1764                 return wb_writeback(wb, &work);
1765         }
1766
1767         return 0;
1768 }
1769
1770 /*
1771  * Retrieve work items and do the writeback they describe
1772  */
1773 static long wb_do_writeback(struct bdi_writeback *wb)
1774 {
1775         struct wb_writeback_work *work;
1776         long wrote = 0;
1777
1778         set_bit(WB_writeback_running, &wb->state);
1779         while ((work = get_next_work_item(wb)) != NULL) {
1780                 struct wb_completion *done = work->done;
1781
1782                 trace_writeback_exec(wb, work);
1783
1784                 wrote += wb_writeback(wb, work);
1785
1786                 if (work->auto_free)
1787                         kfree(work);
1788                 if (done && atomic_dec_and_test(&done->cnt))
1789                         wake_up_all(&wb->bdi->wb_waitq);
1790         }
1791
1792         /*
1793          * Check for periodic writeback, kupdated() style
1794          */
1795         wrote += wb_check_old_data_flush(wb);
1796         wrote += wb_check_background_flush(wb);
1797         clear_bit(WB_writeback_running, &wb->state);
1798
1799         return wrote;
1800 }
1801
1802 /*
1803  * Handle writeback of dirty data for the device backed by this bdi. Also
1804  * reschedules periodically and does kupdated style flushing.
1805  */
1806 void wb_workfn(struct work_struct *work)
1807 {
1808         struct bdi_writeback *wb = container_of(to_delayed_work(work),
1809                                                 struct bdi_writeback, dwork);
1810         long pages_written;
1811
1812         set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1813         current->flags |= PF_SWAPWRITE;
1814
1815         if (likely(!current_is_workqueue_rescuer() ||
1816                    !test_bit(WB_registered, &wb->state))) {
1817                 /*
1818                  * The normal path.  Keep writing back @wb until its
1819                  * work_list is empty.  Note that this path is also taken
1820                  * if @wb is shutting down even when we're running off the
1821                  * rescuer as work_list needs to be drained.
1822                  */
1823                 do {
1824                         pages_written = wb_do_writeback(wb);
1825                         trace_writeback_pages_written(pages_written);
1826                 } while (!list_empty(&wb->work_list));
1827         } else {
1828                 /*
1829                  * bdi_wq can't get enough workers and we're running off
1830                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
1831                  * enough for efficient IO.
1832                  */
1833                 pages_written = writeback_inodes_wb(wb, 1024,
1834                                                     WB_REASON_FORKER_THREAD);
1835                 trace_writeback_pages_written(pages_written);
1836         }
1837
1838         if (!list_empty(&wb->work_list))
1839                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1840         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1841                 wb_wakeup_delayed(wb);
1842
1843         current->flags &= ~PF_SWAPWRITE;
1844 }
1845
1846 /*
1847  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1848  * the whole world.
1849  */
1850 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1851 {
1852         struct backing_dev_info *bdi;
1853
1854         if (!nr_pages)
1855                 nr_pages = get_nr_dirty_pages();
1856
1857         rcu_read_lock();
1858         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1859                 struct bdi_writeback *wb;
1860                 struct wb_iter iter;
1861
1862                 if (!bdi_has_dirty_io(bdi))
1863                         continue;
1864
1865                 bdi_for_each_wb(wb, bdi, &iter, 0)
1866                         wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1867                                            false, reason);
1868         }
1869         rcu_read_unlock();
1870 }
1871
1872 /*
1873  * Wake up bdi's periodically to make sure dirtytime inodes gets
1874  * written back periodically.  We deliberately do *not* check the
1875  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1876  * kernel to be constantly waking up once there are any dirtytime
1877  * inodes on the system.  So instead we define a separate delayed work
1878  * function which gets called much more rarely.  (By default, only
1879  * once every 12 hours.)
1880  *
1881  * If there is any other write activity going on in the file system,
1882  * this function won't be necessary.  But if the only thing that has
1883  * happened on the file system is a dirtytime inode caused by an atime
1884  * update, we need this infrastructure below to make sure that inode
1885  * eventually gets pushed out to disk.
1886  */
1887 static void wakeup_dirtytime_writeback(struct work_struct *w);
1888 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1889
1890 static void wakeup_dirtytime_writeback(struct work_struct *w)
1891 {
1892         struct backing_dev_info *bdi;
1893
1894         rcu_read_lock();
1895         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1896                 struct bdi_writeback *wb;
1897                 struct wb_iter iter;
1898
1899                 bdi_for_each_wb(wb, bdi, &iter, 0)
1900                         if (!list_empty(&bdi->wb.b_dirty_time))
1901                                 wb_wakeup(&bdi->wb);
1902         }
1903         rcu_read_unlock();
1904         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1905 }
1906
1907 static int __init start_dirtytime_writeback(void)
1908 {
1909         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1910         return 0;
1911 }
1912 __initcall(start_dirtytime_writeback);
1913
1914 int dirtytime_interval_handler(struct ctl_table *table, int write,
1915                                void __user *buffer, size_t *lenp, loff_t *ppos)
1916 {
1917         int ret;
1918
1919         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1920         if (ret == 0 && write)
1921                 mod_delayed_work(system_wq, &dirtytime_work, 0);
1922         return ret;
1923 }
1924
1925 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1926 {
1927         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1928                 struct dentry *dentry;
1929                 const char *name = "?";
1930
1931                 dentry = d_find_alias(inode);
1932                 if (dentry) {
1933                         spin_lock(&dentry->d_lock);
1934                         name = (const char *) dentry->d_name.name;
1935                 }
1936                 printk(KERN_DEBUG
1937                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1938                        current->comm, task_pid_nr(current), inode->i_ino,
1939                        name, inode->i_sb->s_id);
1940                 if (dentry) {
1941                         spin_unlock(&dentry->d_lock);
1942                         dput(dentry);
1943                 }
1944         }
1945 }
1946
1947 /**
1948  *      __mark_inode_dirty -    internal function
1949  *      @inode: inode to mark
1950  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1951  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1952  *      mark_inode_dirty_sync.
1953  *
1954  * Put the inode on the super block's dirty list.
1955  *
1956  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1957  * dirty list only if it is hashed or if it refers to a blockdev.
1958  * If it was not hashed, it will never be added to the dirty list
1959  * even if it is later hashed, as it will have been marked dirty already.
1960  *
1961  * In short, make sure you hash any inodes _before_ you start marking
1962  * them dirty.
1963  *
1964  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1965  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1966  * the kernel-internal blockdev inode represents the dirtying time of the
1967  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1968  * page->mapping->host, so the page-dirtying time is recorded in the internal
1969  * blockdev inode.
1970  */
1971 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1972 void __mark_inode_dirty(struct inode *inode, int flags)
1973 {
1974         struct super_block *sb = inode->i_sb;
1975         int dirtytime;
1976
1977         trace_writeback_mark_inode_dirty(inode, flags);
1978
1979         /*
1980          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1981          * dirty the inode itself
1982          */
1983         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
1984                 trace_writeback_dirty_inode_start(inode, flags);
1985
1986                 if (sb->s_op->dirty_inode)
1987                         sb->s_op->dirty_inode(inode, flags);
1988
1989                 trace_writeback_dirty_inode(inode, flags);
1990         }
1991         if (flags & I_DIRTY_INODE)
1992                 flags &= ~I_DIRTY_TIME;
1993         dirtytime = flags & I_DIRTY_TIME;
1994
1995         /*
1996          * Paired with smp_mb() in __writeback_single_inode() for the
1997          * following lockless i_state test.  See there for details.
1998          */
1999         smp_mb();
2000
2001         if (((inode->i_state & flags) == flags) ||
2002             (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2003                 return;
2004
2005         if (unlikely(block_dump))
2006                 block_dump___mark_inode_dirty(inode);
2007
2008         spin_lock(&inode->i_lock);
2009         if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2010                 goto out_unlock_inode;
2011         if ((inode->i_state & flags) != flags) {
2012                 const int was_dirty = inode->i_state & I_DIRTY;
2013
2014                 inode_attach_wb(inode, NULL);
2015
2016                 if (flags & I_DIRTY_INODE)
2017                         inode->i_state &= ~I_DIRTY_TIME;
2018                 inode->i_state |= flags;
2019
2020                 /*
2021                  * If the inode is being synced, just update its dirty state.
2022                  * The unlocker will place the inode on the appropriate
2023                  * superblock list, based upon its state.
2024                  */
2025                 if (inode->i_state & I_SYNC)
2026                         goto out_unlock_inode;
2027
2028                 /*
2029                  * Only add valid (hashed) inodes to the superblock's
2030                  * dirty list.  Add blockdev inodes as well.
2031                  */
2032                 if (!S_ISBLK(inode->i_mode)) {
2033                         if (inode_unhashed(inode))
2034                                 goto out_unlock_inode;
2035                 }
2036                 if (inode->i_state & I_FREEING)
2037                         goto out_unlock_inode;
2038
2039                 /*
2040                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2041                  * reposition it (that would break b_dirty time-ordering).
2042                  */
2043                 if (!was_dirty) {
2044                         struct bdi_writeback *wb;
2045                         struct list_head *dirty_list;
2046                         bool wakeup_bdi = false;
2047
2048                         wb = locked_inode_to_wb_and_lock_list(inode);
2049
2050                         WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2051                              !test_bit(WB_registered, &wb->state),
2052                              "bdi-%s not registered\n", wb->bdi->name);
2053
2054                         inode->dirtied_when = jiffies;
2055                         if (dirtytime)
2056                                 inode->dirtied_time_when = jiffies;
2057
2058                         if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2059                                 dirty_list = &wb->b_dirty;
2060                         else
2061                                 dirty_list = &wb->b_dirty_time;
2062
2063                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2064                                                                dirty_list);
2065
2066                         spin_unlock(&wb->list_lock);
2067                         trace_writeback_dirty_inode_enqueue(inode);
2068
2069                         /*
2070                          * If this is the first dirty inode for this bdi,
2071                          * we have to wake-up the corresponding bdi thread
2072                          * to make sure background write-back happens
2073                          * later.
2074                          */
2075                         if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2076                                 wb_wakeup_delayed(wb);
2077                         return;
2078                 }
2079         }
2080 out_unlock_inode:
2081         spin_unlock(&inode->i_lock);
2082
2083 }
2084 EXPORT_SYMBOL(__mark_inode_dirty);
2085
2086 /*
2087  * The @s_sync_lock is used to serialise concurrent sync operations
2088  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2089  * Concurrent callers will block on the s_sync_lock rather than doing contending
2090  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2091  * has been issued up to the time this function is enter is guaranteed to be
2092  * completed by the time we have gained the lock and waited for all IO that is
2093  * in progress regardless of the order callers are granted the lock.
2094  */
2095 static void wait_sb_inodes(struct super_block *sb)
2096 {
2097         struct inode *inode, *old_inode = NULL;
2098
2099         /*
2100          * We need to be protected against the filesystem going from
2101          * r/o to r/w or vice versa.
2102          */
2103         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2104
2105         mutex_lock(&sb->s_sync_lock);
2106         spin_lock(&sb->s_inode_list_lock);
2107
2108         /*
2109          * Data integrity sync. Must wait for all pages under writeback,
2110          * because there may have been pages dirtied before our sync
2111          * call, but which had writeout started before we write it out.
2112          * In which case, the inode may not be on the dirty list, but
2113          * we still have to wait for that writeout.
2114          */
2115         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
2116                 struct address_space *mapping = inode->i_mapping;
2117
2118                 spin_lock(&inode->i_lock);
2119                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2120                     (mapping->nrpages == 0)) {
2121                         spin_unlock(&inode->i_lock);
2122                         continue;
2123                 }
2124                 __iget(inode);
2125                 spin_unlock(&inode->i_lock);
2126                 spin_unlock(&sb->s_inode_list_lock);
2127
2128                 /*
2129                  * We hold a reference to 'inode' so it couldn't have been
2130                  * removed from s_inodes list while we dropped the
2131                  * s_inode_list_lock.  We cannot iput the inode now as we can
2132                  * be holding the last reference and we cannot iput it under
2133                  * s_inode_list_lock. So we keep the reference and iput it
2134                  * later.
2135                  */
2136                 iput(old_inode);
2137                 old_inode = inode;
2138
2139                 filemap_fdatawait(mapping);
2140
2141                 cond_resched();
2142
2143                 spin_lock(&sb->s_inode_list_lock);
2144         }
2145         spin_unlock(&sb->s_inode_list_lock);
2146         iput(old_inode);
2147         mutex_unlock(&sb->s_sync_lock);
2148 }
2149
2150 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2151                                      enum wb_reason reason, bool skip_if_busy)
2152 {
2153         DEFINE_WB_COMPLETION_ONSTACK(done);
2154         struct wb_writeback_work work = {
2155                 .sb                     = sb,
2156                 .sync_mode              = WB_SYNC_NONE,
2157                 .tagged_writepages      = 1,
2158                 .done                   = &done,
2159                 .nr_pages               = nr,
2160                 .reason                 = reason,
2161         };
2162         struct backing_dev_info *bdi = sb->s_bdi;
2163
2164         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2165                 return;
2166         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2167
2168         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2169         wb_wait_for_completion(bdi, &done);
2170 }
2171
2172 /**
2173  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2174  * @sb: the superblock
2175  * @nr: the number of pages to write
2176  * @reason: reason why some writeback work initiated
2177  *
2178  * Start writeback on some inodes on this super_block. No guarantees are made
2179  * on how many (if any) will be written, and this function does not wait
2180  * for IO completion of submitted IO.
2181  */
2182 void writeback_inodes_sb_nr(struct super_block *sb,
2183                             unsigned long nr,
2184                             enum wb_reason reason)
2185 {
2186         __writeback_inodes_sb_nr(sb, nr, reason, false);
2187 }
2188 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2189
2190 /**
2191  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2192  * @sb: the superblock
2193  * @reason: reason why some writeback work was initiated
2194  *
2195  * Start writeback on some inodes on this super_block. No guarantees are made
2196  * on how many (if any) will be written, and this function does not wait
2197  * for IO completion of submitted IO.
2198  */
2199 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2200 {
2201         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2202 }
2203 EXPORT_SYMBOL(writeback_inodes_sb);
2204
2205 /**
2206  * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2207  * @sb: the superblock
2208  * @nr: the number of pages to write
2209  * @reason: the reason of writeback
2210  *
2211  * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2212  * Returns 1 if writeback was started, 0 if not.
2213  */
2214 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2215                                    enum wb_reason reason)
2216 {
2217         if (!down_read_trylock(&sb->s_umount))
2218                 return false;
2219
2220         __writeback_inodes_sb_nr(sb, nr, reason, true);
2221         up_read(&sb->s_umount);
2222         return true;
2223 }
2224 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2225
2226 /**
2227  * try_to_writeback_inodes_sb - try to start writeback if none underway
2228  * @sb: the superblock
2229  * @reason: reason why some writeback work was initiated
2230  *
2231  * Implement by try_to_writeback_inodes_sb_nr()
2232  * Returns 1 if writeback was started, 0 if not.
2233  */
2234 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2235 {
2236         return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2237 }
2238 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2239
2240 /**
2241  * sync_inodes_sb       -       sync sb inode pages
2242  * @sb: the superblock
2243  *
2244  * This function writes and waits on any dirty inode belonging to this
2245  * super_block.
2246  */
2247 void sync_inodes_sb(struct super_block *sb)
2248 {
2249         DEFINE_WB_COMPLETION_ONSTACK(done);
2250         struct wb_writeback_work work = {
2251                 .sb             = sb,
2252                 .sync_mode      = WB_SYNC_ALL,
2253                 .nr_pages       = LONG_MAX,
2254                 .range_cyclic   = 0,
2255                 .done           = &done,
2256                 .reason         = WB_REASON_SYNC,
2257                 .for_sync       = 1,
2258         };
2259         struct backing_dev_info *bdi = sb->s_bdi;
2260
2261         /*
2262          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2263          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2264          * bdi_has_dirty() need to be written out too.
2265          */
2266         if (bdi == &noop_backing_dev_info)
2267                 return;
2268         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2269
2270         bdi_split_work_to_wbs(bdi, &work, false);
2271         wb_wait_for_completion(bdi, &done);
2272
2273         wait_sb_inodes(sb);
2274 }
2275 EXPORT_SYMBOL(sync_inodes_sb);
2276
2277 /**
2278  * write_inode_now      -       write an inode to disk
2279  * @inode: inode to write to disk
2280  * @sync: whether the write should be synchronous or not
2281  *
2282  * This function commits an inode to disk immediately if it is dirty. This is
2283  * primarily needed by knfsd.
2284  *
2285  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2286  */
2287 int write_inode_now(struct inode *inode, int sync)
2288 {
2289         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
2290         struct writeback_control wbc = {
2291                 .nr_to_write = LONG_MAX,
2292                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2293                 .range_start = 0,
2294                 .range_end = LLONG_MAX,
2295         };
2296
2297         if (!mapping_cap_writeback_dirty(inode->i_mapping))
2298                 wbc.nr_to_write = 0;
2299
2300         might_sleep();
2301         return writeback_single_inode(inode, wb, &wbc);
2302 }
2303 EXPORT_SYMBOL(write_inode_now);
2304
2305 /**
2306  * sync_inode - write an inode and its pages to disk.
2307  * @inode: the inode to sync
2308  * @wbc: controls the writeback mode
2309  *
2310  * sync_inode() will write an inode and its pages to disk.  It will also
2311  * correctly update the inode on its superblock's dirty inode lists and will
2312  * update inode->i_state.
2313  *
2314  * The caller must have a ref on the inode.
2315  */
2316 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2317 {
2318         return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
2319 }
2320 EXPORT_SYMBOL(sync_inode);
2321
2322 /**
2323  * sync_inode_metadata - write an inode to disk
2324  * @inode: the inode to sync
2325  * @wait: wait for I/O to complete.
2326  *
2327  * Write an inode to disk and adjust its dirty state after completion.
2328  *
2329  * Note: only writes the actual inode, no associated data or other metadata.
2330  */
2331 int sync_inode_metadata(struct inode *inode, int wait)
2332 {
2333         struct writeback_control wbc = {
2334                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2335                 .nr_to_write = 0, /* metadata-only */
2336         };
2337
2338         return sync_inode(inode, &wbc);
2339 }
2340 EXPORT_SYMBOL(sync_inode_metadata);