writeback: prevent unnecessary bdi threads wakeups
[pandora-kernel.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/module.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/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * Passed into wb_writeback(), essentially a subset of writeback_control
34  */
35 struct wb_writeback_work {
36         long nr_pages;
37         struct super_block *sb;
38         enum writeback_sync_modes sync_mode;
39         unsigned int for_kupdate:1;
40         unsigned int range_cyclic:1;
41         unsigned int for_background:1;
42
43         struct list_head list;          /* pending work list */
44         struct completion *done;        /* set if the caller waits */
45 };
46
47 /*
48  * Include the creation of the trace points after defining the
49  * wb_writeback_work structure so that the definition remains local to this
50  * file.
51  */
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/writeback.h>
54
55 #define inode_to_bdi(inode)     ((inode)->i_mapping->backing_dev_info)
56
57 /*
58  * We don't actually have pdflush, but this one is exported though /proc...
59  */
60 int nr_pdflush_threads;
61
62 /**
63  * writeback_in_progress - determine whether there is writeback in progress
64  * @bdi: the device's backing_dev_info structure.
65  *
66  * Determine whether there is writeback waiting to be handled against a
67  * backing device.
68  */
69 int writeback_in_progress(struct backing_dev_info *bdi)
70 {
71         return !list_empty(&bdi->work_list);
72 }
73
74 static void bdi_queue_work(struct backing_dev_info *bdi,
75                 struct wb_writeback_work *work)
76 {
77         trace_writeback_queue(bdi, work);
78
79         spin_lock(&bdi->wb_lock);
80         list_add_tail(&work->list, &bdi->work_list);
81         if (bdi->wb.task) {
82                 wake_up_process(bdi->wb.task);
83         } else {
84                 /*
85                  * The bdi thread isn't there, wake up the forker thread which
86                  * will create and run it.
87                  */
88                 trace_writeback_nothread(bdi, work);
89                 wake_up_process(default_backing_dev_info.wb.task);
90         }
91         spin_unlock(&bdi->wb_lock);
92 }
93
94 static void
95 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
96                 bool range_cyclic, bool for_background)
97 {
98         struct wb_writeback_work *work;
99
100         /*
101          * This is WB_SYNC_NONE writeback, so if allocation fails just
102          * wakeup the thread for old dirty data writeback
103          */
104         work = kzalloc(sizeof(*work), GFP_ATOMIC);
105         if (!work) {
106                 if (bdi->wb.task) {
107                         trace_writeback_nowork(bdi);
108                         wake_up_process(bdi->wb.task);
109                 }
110                 return;
111         }
112
113         work->sync_mode = WB_SYNC_NONE;
114         work->nr_pages  = nr_pages;
115         work->range_cyclic = range_cyclic;
116         work->for_background = for_background;
117
118         bdi_queue_work(bdi, work);
119 }
120
121 /**
122  * bdi_start_writeback - start writeback
123  * @bdi: the backing device to write from
124  * @nr_pages: the number of pages to write
125  *
126  * Description:
127  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
128  *   started when this function returns, we make no guarentees on
129  *   completion. Caller need not hold sb s_umount semaphore.
130  *
131  */
132 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
133 {
134         __bdi_start_writeback(bdi, nr_pages, true, false);
135 }
136
137 /**
138  * bdi_start_background_writeback - start background writeback
139  * @bdi: the backing device to write from
140  *
141  * Description:
142  *   This does WB_SYNC_NONE background writeback. The IO is only
143  *   started when this function returns, we make no guarentees on
144  *   completion. Caller need not hold sb s_umount semaphore.
145  */
146 void bdi_start_background_writeback(struct backing_dev_info *bdi)
147 {
148         __bdi_start_writeback(bdi, LONG_MAX, true, true);
149 }
150
151 /*
152  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
153  * furthest end of its superblock's dirty-inode list.
154  *
155  * Before stamping the inode's ->dirtied_when, we check to see whether it is
156  * already the most-recently-dirtied inode on the b_dirty list.  If that is
157  * the case then the inode must have been redirtied while it was being written
158  * out and we don't reset its dirtied_when.
159  */
160 static void redirty_tail(struct inode *inode)
161 {
162         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
163
164         if (!list_empty(&wb->b_dirty)) {
165                 struct inode *tail;
166
167                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
168                 if (time_before(inode->dirtied_when, tail->dirtied_when))
169                         inode->dirtied_when = jiffies;
170         }
171         list_move(&inode->i_list, &wb->b_dirty);
172 }
173
174 /*
175  * requeue inode for re-scanning after bdi->b_io list is exhausted.
176  */
177 static void requeue_io(struct inode *inode)
178 {
179         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
180
181         list_move(&inode->i_list, &wb->b_more_io);
182 }
183
184 static void inode_sync_complete(struct inode *inode)
185 {
186         /*
187          * Prevent speculative execution through spin_unlock(&inode_lock);
188          */
189         smp_mb();
190         wake_up_bit(&inode->i_state, __I_SYNC);
191 }
192
193 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
194 {
195         bool ret = time_after(inode->dirtied_when, t);
196 #ifndef CONFIG_64BIT
197         /*
198          * For inodes being constantly redirtied, dirtied_when can get stuck.
199          * It _appears_ to be in the future, but is actually in distant past.
200          * This test is necessary to prevent such wrapped-around relative times
201          * from permanently stopping the whole bdi writeback.
202          */
203         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
204 #endif
205         return ret;
206 }
207
208 /*
209  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
210  */
211 static void move_expired_inodes(struct list_head *delaying_queue,
212                                struct list_head *dispatch_queue,
213                                 unsigned long *older_than_this)
214 {
215         LIST_HEAD(tmp);
216         struct list_head *pos, *node;
217         struct super_block *sb = NULL;
218         struct inode *inode;
219         int do_sb_sort = 0;
220
221         while (!list_empty(delaying_queue)) {
222                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
223                 if (older_than_this &&
224                     inode_dirtied_after(inode, *older_than_this))
225                         break;
226                 if (sb && sb != inode->i_sb)
227                         do_sb_sort = 1;
228                 sb = inode->i_sb;
229                 list_move(&inode->i_list, &tmp);
230         }
231
232         /* just one sb in list, splice to dispatch_queue and we're done */
233         if (!do_sb_sort) {
234                 list_splice(&tmp, dispatch_queue);
235                 return;
236         }
237
238         /* Move inodes from one superblock together */
239         while (!list_empty(&tmp)) {
240                 inode = list_entry(tmp.prev, struct inode, i_list);
241                 sb = inode->i_sb;
242                 list_for_each_prev_safe(pos, node, &tmp) {
243                         inode = list_entry(pos, struct inode, i_list);
244                         if (inode->i_sb == sb)
245                                 list_move(&inode->i_list, dispatch_queue);
246                 }
247         }
248 }
249
250 /*
251  * Queue all expired dirty inodes for io, eldest first.
252  */
253 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
254 {
255         list_splice_init(&wb->b_more_io, wb->b_io.prev);
256         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
257 }
258
259 static int write_inode(struct inode *inode, struct writeback_control *wbc)
260 {
261         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
262                 return inode->i_sb->s_op->write_inode(inode, wbc);
263         return 0;
264 }
265
266 /*
267  * Wait for writeback on an inode to complete.
268  */
269 static void inode_wait_for_writeback(struct inode *inode)
270 {
271         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
272         wait_queue_head_t *wqh;
273
274         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
275          while (inode->i_state & I_SYNC) {
276                 spin_unlock(&inode_lock);
277                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
278                 spin_lock(&inode_lock);
279         }
280 }
281
282 /*
283  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
284  * caller has ref on the inode (either via __iget or via syscall against an fd)
285  * or the inode has I_WILL_FREE set (via generic_forget_inode)
286  *
287  * If `wait' is set, wait on the writeout.
288  *
289  * The whole writeout design is quite complex and fragile.  We want to avoid
290  * starvation of particular inodes when others are being redirtied, prevent
291  * livelocks, etc.
292  *
293  * Called under inode_lock.
294  */
295 static int
296 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
297 {
298         struct address_space *mapping = inode->i_mapping;
299         unsigned dirty;
300         int ret;
301
302         if (!atomic_read(&inode->i_count))
303                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
304         else
305                 WARN_ON(inode->i_state & I_WILL_FREE);
306
307         if (inode->i_state & I_SYNC) {
308                 /*
309                  * If this inode is locked for writeback and we are not doing
310                  * writeback-for-data-integrity, move it to b_more_io so that
311                  * writeback can proceed with the other inodes on s_io.
312                  *
313                  * We'll have another go at writing back this inode when we
314                  * completed a full scan of b_io.
315                  */
316                 if (wbc->sync_mode != WB_SYNC_ALL) {
317                         requeue_io(inode);
318                         return 0;
319                 }
320
321                 /*
322                  * It's a data-integrity sync.  We must wait.
323                  */
324                 inode_wait_for_writeback(inode);
325         }
326
327         BUG_ON(inode->i_state & I_SYNC);
328
329         /* Set I_SYNC, reset I_DIRTY_PAGES */
330         inode->i_state |= I_SYNC;
331         inode->i_state &= ~I_DIRTY_PAGES;
332         spin_unlock(&inode_lock);
333
334         ret = do_writepages(mapping, wbc);
335
336         /*
337          * Make sure to wait on the data before writing out the metadata.
338          * This is important for filesystems that modify metadata on data
339          * I/O completion.
340          */
341         if (wbc->sync_mode == WB_SYNC_ALL) {
342                 int err = filemap_fdatawait(mapping);
343                 if (ret == 0)
344                         ret = err;
345         }
346
347         /*
348          * Some filesystems may redirty the inode during the writeback
349          * due to delalloc, clear dirty metadata flags right before
350          * write_inode()
351          */
352         spin_lock(&inode_lock);
353         dirty = inode->i_state & I_DIRTY;
354         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
355         spin_unlock(&inode_lock);
356         /* Don't write the inode if only I_DIRTY_PAGES was set */
357         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
358                 int err = write_inode(inode, wbc);
359                 if (ret == 0)
360                         ret = err;
361         }
362
363         spin_lock(&inode_lock);
364         inode->i_state &= ~I_SYNC;
365         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
366                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
367                         /*
368                          * More pages get dirtied by a fast dirtier.
369                          */
370                         goto select_queue;
371                 } else if (inode->i_state & I_DIRTY) {
372                         /*
373                          * At least XFS will redirty the inode during the
374                          * writeback (delalloc) and on io completion (isize).
375                          */
376                         redirty_tail(inode);
377                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
378                         /*
379                          * We didn't write back all the pages.  nfs_writepages()
380                          * sometimes bales out without doing anything. Redirty
381                          * the inode; Move it from b_io onto b_more_io/b_dirty.
382                          */
383                         /*
384                          * akpm: if the caller was the kupdate function we put
385                          * this inode at the head of b_dirty so it gets first
386                          * consideration.  Otherwise, move it to the tail, for
387                          * the reasons described there.  I'm not really sure
388                          * how much sense this makes.  Presumably I had a good
389                          * reasons for doing it this way, and I'd rather not
390                          * muck with it at present.
391                          */
392                         if (wbc->for_kupdate) {
393                                 /*
394                                  * For the kupdate function we move the inode
395                                  * to b_more_io so it will get more writeout as
396                                  * soon as the queue becomes uncongested.
397                                  */
398                                 inode->i_state |= I_DIRTY_PAGES;
399 select_queue:
400                                 if (wbc->nr_to_write <= 0) {
401                                         /*
402                                          * slice used up: queue for next turn
403                                          */
404                                         requeue_io(inode);
405                                 } else {
406                                         /*
407                                          * somehow blocked: retry later
408                                          */
409                                         redirty_tail(inode);
410                                 }
411                         } else {
412                                 /*
413                                  * Otherwise fully redirty the inode so that
414                                  * other inodes on this superblock will get some
415                                  * writeout.  Otherwise heavy writing to one
416                                  * file would indefinitely suspend writeout of
417                                  * all the other files.
418                                  */
419                                 inode->i_state |= I_DIRTY_PAGES;
420                                 redirty_tail(inode);
421                         }
422                 } else if (atomic_read(&inode->i_count)) {
423                         /*
424                          * The inode is clean, inuse
425                          */
426                         list_move(&inode->i_list, &inode_in_use);
427                 } else {
428                         /*
429                          * The inode is clean, unused
430                          */
431                         list_move(&inode->i_list, &inode_unused);
432                 }
433         }
434         inode_sync_complete(inode);
435         return ret;
436 }
437
438 /*
439  * For background writeback the caller does not have the sb pinned
440  * before calling writeback. So make sure that we do pin it, so it doesn't
441  * go away while we are writing inodes from it.
442  */
443 static bool pin_sb_for_writeback(struct super_block *sb)
444 {
445         spin_lock(&sb_lock);
446         if (list_empty(&sb->s_instances)) {
447                 spin_unlock(&sb_lock);
448                 return false;
449         }
450
451         sb->s_count++;
452         spin_unlock(&sb_lock);
453
454         if (down_read_trylock(&sb->s_umount)) {
455                 if (sb->s_root)
456                         return true;
457                 up_read(&sb->s_umount);
458         }
459
460         put_super(sb);
461         return false;
462 }
463
464 /*
465  * Write a portion of b_io inodes which belong to @sb.
466  *
467  * If @only_this_sb is true, then find and write all such
468  * inodes. Otherwise write only ones which go sequentially
469  * in reverse order.
470  *
471  * Return 1, if the caller writeback routine should be
472  * interrupted. Otherwise return 0.
473  */
474 static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
475                 struct writeback_control *wbc, bool only_this_sb)
476 {
477         while (!list_empty(&wb->b_io)) {
478                 long pages_skipped;
479                 struct inode *inode = list_entry(wb->b_io.prev,
480                                                  struct inode, i_list);
481
482                 if (inode->i_sb != sb) {
483                         if (only_this_sb) {
484                                 /*
485                                  * We only want to write back data for this
486                                  * superblock, move all inodes not belonging
487                                  * to it back onto the dirty list.
488                                  */
489                                 redirty_tail(inode);
490                                 continue;
491                         }
492
493                         /*
494                          * The inode belongs to a different superblock.
495                          * Bounce back to the caller to unpin this and
496                          * pin the next superblock.
497                          */
498                         return 0;
499                 }
500
501                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
502                         requeue_io(inode);
503                         continue;
504                 }
505                 /*
506                  * Was this inode dirtied after sync_sb_inodes was called?
507                  * This keeps sync from extra jobs and livelock.
508                  */
509                 if (inode_dirtied_after(inode, wbc->wb_start))
510                         return 1;
511
512                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
513                 __iget(inode);
514                 pages_skipped = wbc->pages_skipped;
515                 writeback_single_inode(inode, wbc);
516                 if (wbc->pages_skipped != pages_skipped) {
517                         /*
518                          * writeback is not making progress due to locked
519                          * buffers.  Skip this inode for now.
520                          */
521                         redirty_tail(inode);
522                 }
523                 spin_unlock(&inode_lock);
524                 iput(inode);
525                 cond_resched();
526                 spin_lock(&inode_lock);
527                 if (wbc->nr_to_write <= 0) {
528                         wbc->more_io = 1;
529                         return 1;
530                 }
531                 if (!list_empty(&wb->b_more_io))
532                         wbc->more_io = 1;
533         }
534         /* b_io is empty */
535         return 1;
536 }
537
538 void writeback_inodes_wb(struct bdi_writeback *wb,
539                 struct writeback_control *wbc)
540 {
541         int ret = 0;
542
543         wbc->wb_start = jiffies; /* livelock avoidance */
544         spin_lock(&inode_lock);
545         if (!wbc->for_kupdate || list_empty(&wb->b_io))
546                 queue_io(wb, wbc->older_than_this);
547
548         while (!list_empty(&wb->b_io)) {
549                 struct inode *inode = list_entry(wb->b_io.prev,
550                                                  struct inode, i_list);
551                 struct super_block *sb = inode->i_sb;
552
553                 if (!pin_sb_for_writeback(sb)) {
554                         requeue_io(inode);
555                         continue;
556                 }
557                 ret = writeback_sb_inodes(sb, wb, wbc, false);
558                 drop_super(sb);
559
560                 if (ret)
561                         break;
562         }
563         spin_unlock(&inode_lock);
564         /* Leave any unwritten inodes on b_io */
565 }
566
567 static void __writeback_inodes_sb(struct super_block *sb,
568                 struct bdi_writeback *wb, struct writeback_control *wbc)
569 {
570         WARN_ON(!rwsem_is_locked(&sb->s_umount));
571
572         wbc->wb_start = jiffies; /* livelock avoidance */
573         spin_lock(&inode_lock);
574         if (!wbc->for_kupdate || list_empty(&wb->b_io))
575                 queue_io(wb, wbc->older_than_this);
576         writeback_sb_inodes(sb, wb, wbc, true);
577         spin_unlock(&inode_lock);
578 }
579
580 /*
581  * The maximum number of pages to writeout in a single bdi flush/kupdate
582  * operation.  We do this so we don't hold I_SYNC against an inode for
583  * enormous amounts of time, which would block a userspace task which has
584  * been forced to throttle against that inode.  Also, the code reevaluates
585  * the dirty each time it has written this many pages.
586  */
587 #define MAX_WRITEBACK_PAGES     1024
588
589 static inline bool over_bground_thresh(void)
590 {
591         unsigned long background_thresh, dirty_thresh;
592
593         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
594
595         return (global_page_state(NR_FILE_DIRTY) +
596                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
597 }
598
599 /*
600  * Explicit flushing or periodic writeback of "old" data.
601  *
602  * Define "old": the first time one of an inode's pages is dirtied, we mark the
603  * dirtying-time in the inode's address_space.  So this periodic writeback code
604  * just walks the superblock inode list, writing back any inodes which are
605  * older than a specific point in time.
606  *
607  * Try to run once per dirty_writeback_interval.  But if a writeback event
608  * takes longer than a dirty_writeback_interval interval, then leave a
609  * one-second gap.
610  *
611  * older_than_this takes precedence over nr_to_write.  So we'll only write back
612  * all dirty pages if they are all attached to "old" mappings.
613  */
614 static long wb_writeback(struct bdi_writeback *wb,
615                          struct wb_writeback_work *work)
616 {
617         struct writeback_control wbc = {
618                 .sync_mode              = work->sync_mode,
619                 .older_than_this        = NULL,
620                 .for_kupdate            = work->for_kupdate,
621                 .for_background         = work->for_background,
622                 .range_cyclic           = work->range_cyclic,
623         };
624         unsigned long oldest_jif;
625         long wrote = 0;
626         struct inode *inode;
627
628         if (wbc.for_kupdate) {
629                 wbc.older_than_this = &oldest_jif;
630                 oldest_jif = jiffies -
631                                 msecs_to_jiffies(dirty_expire_interval * 10);
632         }
633         if (!wbc.range_cyclic) {
634                 wbc.range_start = 0;
635                 wbc.range_end = LLONG_MAX;
636         }
637
638         for (;;) {
639                 /*
640                  * Stop writeback when nr_pages has been consumed
641                  */
642                 if (work->nr_pages <= 0)
643                         break;
644
645                 /*
646                  * For background writeout, stop when we are below the
647                  * background dirty threshold
648                  */
649                 if (work->for_background && !over_bground_thresh())
650                         break;
651
652                 wbc.more_io = 0;
653                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
654                 wbc.pages_skipped = 0;
655
656                 trace_wbc_writeback_start(&wbc, wb->bdi);
657                 if (work->sb)
658                         __writeback_inodes_sb(work->sb, wb, &wbc);
659                 else
660                         writeback_inodes_wb(wb, &wbc);
661                 trace_wbc_writeback_written(&wbc, wb->bdi);
662
663                 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
664                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
665
666                 /*
667                  * If we consumed everything, see if we have more
668                  */
669                 if (wbc.nr_to_write <= 0)
670                         continue;
671                 /*
672                  * Didn't write everything and we don't have more IO, bail
673                  */
674                 if (!wbc.more_io)
675                         break;
676                 /*
677                  * Did we write something? Try for more
678                  */
679                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
680                         continue;
681                 /*
682                  * Nothing written. Wait for some inode to
683                  * become available for writeback. Otherwise
684                  * we'll just busyloop.
685                  */
686                 spin_lock(&inode_lock);
687                 if (!list_empty(&wb->b_more_io))  {
688                         inode = list_entry(wb->b_more_io.prev,
689                                                 struct inode, i_list);
690                         trace_wbc_writeback_wait(&wbc, wb->bdi);
691                         inode_wait_for_writeback(inode);
692                 }
693                 spin_unlock(&inode_lock);
694         }
695
696         return wrote;
697 }
698
699 /*
700  * Return the next wb_writeback_work struct that hasn't been processed yet.
701  */
702 static struct wb_writeback_work *
703 get_next_work_item(struct backing_dev_info *bdi)
704 {
705         struct wb_writeback_work *work = NULL;
706
707         spin_lock(&bdi->wb_lock);
708         if (!list_empty(&bdi->work_list)) {
709                 work = list_entry(bdi->work_list.next,
710                                   struct wb_writeback_work, list);
711                 list_del_init(&work->list);
712         }
713         spin_unlock(&bdi->wb_lock);
714         return work;
715 }
716
717 static long wb_check_old_data_flush(struct bdi_writeback *wb)
718 {
719         unsigned long expired;
720         long nr_pages;
721
722         /*
723          * When set to zero, disable periodic writeback
724          */
725         if (!dirty_writeback_interval)
726                 return 0;
727
728         expired = wb->last_old_flush +
729                         msecs_to_jiffies(dirty_writeback_interval * 10);
730         if (time_before(jiffies, expired))
731                 return 0;
732
733         wb->last_old_flush = jiffies;
734         nr_pages = global_page_state(NR_FILE_DIRTY) +
735                         global_page_state(NR_UNSTABLE_NFS) +
736                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
737
738         if (nr_pages) {
739                 struct wb_writeback_work work = {
740                         .nr_pages       = nr_pages,
741                         .sync_mode      = WB_SYNC_NONE,
742                         .for_kupdate    = 1,
743                         .range_cyclic   = 1,
744                 };
745
746                 return wb_writeback(wb, &work);
747         }
748
749         return 0;
750 }
751
752 /*
753  * Retrieve work items and do the writeback they describe
754  */
755 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
756 {
757         struct backing_dev_info *bdi = wb->bdi;
758         struct wb_writeback_work *work;
759         long wrote = 0;
760
761         while ((work = get_next_work_item(bdi)) != NULL) {
762                 /*
763                  * Override sync mode, in case we must wait for completion
764                  * because this thread is exiting now.
765                  */
766                 if (force_wait)
767                         work->sync_mode = WB_SYNC_ALL;
768
769                 trace_writeback_exec(bdi, work);
770
771                 wrote += wb_writeback(wb, work);
772
773                 /*
774                  * Notify the caller of completion if this is a synchronous
775                  * work item, otherwise just free it.
776                  */
777                 if (work->done)
778                         complete(work->done);
779                 else
780                         kfree(work);
781         }
782
783         /*
784          * Check for periodic writeback, kupdated() style
785          */
786         wrote += wb_check_old_data_flush(wb);
787
788         return wrote;
789 }
790
791 /*
792  * Handle writeback of dirty data for the device backed by this bdi. Also
793  * wakes up periodically and does kupdated style flushing.
794  */
795 int bdi_writeback_thread(void *data)
796 {
797         struct bdi_writeback *wb = data;
798         struct backing_dev_info *bdi = wb->bdi;
799         long pages_written;
800
801         current->flags |= PF_FLUSHER | PF_SWAPWRITE;
802         set_freezable();
803         wb->last_active = jiffies;
804
805         /*
806          * Our parent may run at a different priority, just set us to normal
807          */
808         set_user_nice(current, 0);
809
810         trace_writeback_thread_start(bdi);
811
812         while (!kthread_should_stop()) {
813                 pages_written = wb_do_writeback(wb, 0);
814
815                 trace_writeback_pages_written(pages_written);
816
817                 if (pages_written)
818                         wb->last_active = jiffies;
819
820                 set_current_state(TASK_INTERRUPTIBLE);
821                 if (!list_empty(&bdi->work_list)) {
822                         __set_current_state(TASK_RUNNING);
823                         continue;
824                 }
825
826                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
827                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
828                 else {
829                         /*
830                          * We have nothing to do, so can go sleep without any
831                          * timeout and save power. When a work is queued or
832                          * something is made dirty - we will be woken up.
833                          */
834                         schedule();
835                 }
836
837                 try_to_freeze();
838         }
839
840         /* Flush any work that raced with us exiting */
841         if (!list_empty(&bdi->work_list))
842                 wb_do_writeback(wb, 1);
843
844         trace_writeback_thread_stop(bdi);
845         return 0;
846 }
847
848
849 /*
850  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
851  * the whole world.
852  */
853 void wakeup_flusher_threads(long nr_pages)
854 {
855         struct backing_dev_info *bdi;
856
857         if (!nr_pages) {
858                 nr_pages = global_page_state(NR_FILE_DIRTY) +
859                                 global_page_state(NR_UNSTABLE_NFS);
860         }
861
862         rcu_read_lock();
863         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
864                 if (!bdi_has_dirty_io(bdi))
865                         continue;
866                 __bdi_start_writeback(bdi, nr_pages, false, false);
867         }
868         rcu_read_unlock();
869 }
870
871 /*
872  * This function is used when the first inode for this bdi is marked dirty. It
873  * wakes-up the corresponding bdi thread which should then take care of the
874  * periodic background write-out of dirty inodes.
875  */
876 static void wakeup_bdi_thread(struct backing_dev_info *bdi)
877 {
878         spin_lock(&bdi->wb_lock);
879         if (bdi->wb.task)
880                 wake_up_process(bdi->wb.task);
881         else
882                 /*
883                  * When bdi tasks are inactive for long time, they are killed.
884                  * In this case we have to wake-up the forker thread which
885                  * should create and run the bdi thread.
886                  */
887                 wake_up_process(default_backing_dev_info.wb.task);
888         spin_unlock(&bdi->wb_lock);
889 }
890
891 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
892 {
893         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
894                 struct dentry *dentry;
895                 const char *name = "?";
896
897                 dentry = d_find_alias(inode);
898                 if (dentry) {
899                         spin_lock(&dentry->d_lock);
900                         name = (const char *) dentry->d_name.name;
901                 }
902                 printk(KERN_DEBUG
903                        "%s(%d): dirtied inode %lu (%s) on %s\n",
904                        current->comm, task_pid_nr(current), inode->i_ino,
905                        name, inode->i_sb->s_id);
906                 if (dentry) {
907                         spin_unlock(&dentry->d_lock);
908                         dput(dentry);
909                 }
910         }
911 }
912
913 /**
914  *      __mark_inode_dirty -    internal function
915  *      @inode: inode to mark
916  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
917  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
918  *      mark_inode_dirty_sync.
919  *
920  * Put the inode on the super block's dirty list.
921  *
922  * CAREFUL! We mark it dirty unconditionally, but move it onto the
923  * dirty list only if it is hashed or if it refers to a blockdev.
924  * If it was not hashed, it will never be added to the dirty list
925  * even if it is later hashed, as it will have been marked dirty already.
926  *
927  * In short, make sure you hash any inodes _before_ you start marking
928  * them dirty.
929  *
930  * This function *must* be atomic for the I_DIRTY_PAGES case -
931  * set_page_dirty() is called under spinlock in several places.
932  *
933  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
934  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
935  * the kernel-internal blockdev inode represents the dirtying time of the
936  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
937  * page->mapping->host, so the page-dirtying time is recorded in the internal
938  * blockdev inode.
939  */
940 void __mark_inode_dirty(struct inode *inode, int flags)
941 {
942         struct super_block *sb = inode->i_sb;
943         struct backing_dev_info *bdi = NULL;
944         bool wakeup_bdi = false;
945
946         /*
947          * Don't do this for I_DIRTY_PAGES - that doesn't actually
948          * dirty the inode itself
949          */
950         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
951                 if (sb->s_op->dirty_inode)
952                         sb->s_op->dirty_inode(inode);
953         }
954
955         /*
956          * make sure that changes are seen by all cpus before we test i_state
957          * -- mikulas
958          */
959         smp_mb();
960
961         /* avoid the locking if we can */
962         if ((inode->i_state & flags) == flags)
963                 return;
964
965         if (unlikely(block_dump))
966                 block_dump___mark_inode_dirty(inode);
967
968         spin_lock(&inode_lock);
969         if ((inode->i_state & flags) != flags) {
970                 const int was_dirty = inode->i_state & I_DIRTY;
971
972                 inode->i_state |= flags;
973
974                 /*
975                  * If the inode is being synced, just update its dirty state.
976                  * The unlocker will place the inode on the appropriate
977                  * superblock list, based upon its state.
978                  */
979                 if (inode->i_state & I_SYNC)
980                         goto out;
981
982                 /*
983                  * Only add valid (hashed) inodes to the superblock's
984                  * dirty list.  Add blockdev inodes as well.
985                  */
986                 if (!S_ISBLK(inode->i_mode)) {
987                         if (hlist_unhashed(&inode->i_hash))
988                                 goto out;
989                 }
990                 if (inode->i_state & (I_FREEING|I_CLEAR))
991                         goto out;
992
993                 /*
994                  * If the inode was already on b_dirty/b_io/b_more_io, don't
995                  * reposition it (that would break b_dirty time-ordering).
996                  */
997                 if (!was_dirty) {
998                         bdi = inode_to_bdi(inode);
999
1000                         if (bdi_cap_writeback_dirty(bdi)) {
1001                                 WARN(!test_bit(BDI_registered, &bdi->state),
1002                                      "bdi-%s not registered\n", bdi->name);
1003
1004                                 /*
1005                                  * If this is the first dirty inode for this
1006                                  * bdi, we have to wake-up the corresponding
1007                                  * bdi thread to make sure background
1008                                  * write-back happens later.
1009                                  */
1010                                 if (!wb_has_dirty_io(&bdi->wb))
1011                                         wakeup_bdi = true;
1012                         }
1013
1014                         inode->dirtied_when = jiffies;
1015                         list_move(&inode->i_list, &bdi->wb.b_dirty);
1016                 }
1017         }
1018 out:
1019         spin_unlock(&inode_lock);
1020
1021         if (wakeup_bdi)
1022                 wakeup_bdi_thread(bdi);
1023 }
1024 EXPORT_SYMBOL(__mark_inode_dirty);
1025
1026 /*
1027  * Write out a superblock's list of dirty inodes.  A wait will be performed
1028  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1029  *
1030  * If older_than_this is non-NULL, then only write out inodes which
1031  * had their first dirtying at a time earlier than *older_than_this.
1032  *
1033  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1034  * This function assumes that the blockdev superblock's inodes are backed by
1035  * a variety of queues, so all inodes are searched.  For other superblocks,
1036  * assume that all inodes are backed by the same queue.
1037  *
1038  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1039  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1040  * on the writer throttling path, and we get decent balancing between many
1041  * throttled threads: we don't want them all piling up on inode_sync_wait.
1042  */
1043 static void wait_sb_inodes(struct super_block *sb)
1044 {
1045         struct inode *inode, *old_inode = NULL;
1046
1047         /*
1048          * We need to be protected against the filesystem going from
1049          * r/o to r/w or vice versa.
1050          */
1051         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1052
1053         spin_lock(&inode_lock);
1054
1055         /*
1056          * Data integrity sync. Must wait for all pages under writeback,
1057          * because there may have been pages dirtied before our sync
1058          * call, but which had writeout started before we write it out.
1059          * In which case, the inode may not be on the dirty list, but
1060          * we still have to wait for that writeout.
1061          */
1062         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1063                 struct address_space *mapping;
1064
1065                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1066                         continue;
1067                 mapping = inode->i_mapping;
1068                 if (mapping->nrpages == 0)
1069                         continue;
1070                 __iget(inode);
1071                 spin_unlock(&inode_lock);
1072                 /*
1073                  * We hold a reference to 'inode' so it couldn't have
1074                  * been removed from s_inodes list while we dropped the
1075                  * inode_lock.  We cannot iput the inode now as we can
1076                  * be holding the last reference and we cannot iput it
1077                  * under inode_lock. So we keep the reference and iput
1078                  * it later.
1079                  */
1080                 iput(old_inode);
1081                 old_inode = inode;
1082
1083                 filemap_fdatawait(mapping);
1084
1085                 cond_resched();
1086
1087                 spin_lock(&inode_lock);
1088         }
1089         spin_unlock(&inode_lock);
1090         iput(old_inode);
1091 }
1092
1093 /**
1094  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1095  * @sb: the superblock
1096  *
1097  * Start writeback on some inodes on this super_block. No guarantees are made
1098  * on how many (if any) will be written, and this function does not wait
1099  * for IO completion of submitted IO. The number of pages submitted is
1100  * returned.
1101  */
1102 void writeback_inodes_sb(struct super_block *sb)
1103 {
1104         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1105         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1106         DECLARE_COMPLETION_ONSTACK(done);
1107         struct wb_writeback_work work = {
1108                 .sb             = sb,
1109                 .sync_mode      = WB_SYNC_NONE,
1110                 .done           = &done,
1111         };
1112
1113         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1114
1115         work.nr_pages = nr_dirty + nr_unstable +
1116                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1117
1118         bdi_queue_work(sb->s_bdi, &work);
1119         wait_for_completion(&done);
1120 }
1121 EXPORT_SYMBOL(writeback_inodes_sb);
1122
1123 /**
1124  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1125  * @sb: the superblock
1126  *
1127  * Invoke writeback_inodes_sb if no writeback is currently underway.
1128  * Returns 1 if writeback was started, 0 if not.
1129  */
1130 int writeback_inodes_sb_if_idle(struct super_block *sb)
1131 {
1132         if (!writeback_in_progress(sb->s_bdi)) {
1133                 down_read(&sb->s_umount);
1134                 writeback_inodes_sb(sb);
1135                 up_read(&sb->s_umount);
1136                 return 1;
1137         } else
1138                 return 0;
1139 }
1140 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1141
1142 /**
1143  * sync_inodes_sb       -       sync sb inode pages
1144  * @sb: the superblock
1145  *
1146  * This function writes and waits on any dirty inode belonging to this
1147  * super_block. The number of pages synced is returned.
1148  */
1149 void sync_inodes_sb(struct super_block *sb)
1150 {
1151         DECLARE_COMPLETION_ONSTACK(done);
1152         struct wb_writeback_work work = {
1153                 .sb             = sb,
1154                 .sync_mode      = WB_SYNC_ALL,
1155                 .nr_pages       = LONG_MAX,
1156                 .range_cyclic   = 0,
1157                 .done           = &done,
1158         };
1159
1160         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1161
1162         bdi_queue_work(sb->s_bdi, &work);
1163         wait_for_completion(&done);
1164
1165         wait_sb_inodes(sb);
1166 }
1167 EXPORT_SYMBOL(sync_inodes_sb);
1168
1169 /**
1170  * write_inode_now      -       write an inode to disk
1171  * @inode: inode to write to disk
1172  * @sync: whether the write should be synchronous or not
1173  *
1174  * This function commits an inode to disk immediately if it is dirty. This is
1175  * primarily needed by knfsd.
1176  *
1177  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1178  */
1179 int write_inode_now(struct inode *inode, int sync)
1180 {
1181         int ret;
1182         struct writeback_control wbc = {
1183                 .nr_to_write = LONG_MAX,
1184                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1185                 .range_start = 0,
1186                 .range_end = LLONG_MAX,
1187         };
1188
1189         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1190                 wbc.nr_to_write = 0;
1191
1192         might_sleep();
1193         spin_lock(&inode_lock);
1194         ret = writeback_single_inode(inode, &wbc);
1195         spin_unlock(&inode_lock);
1196         if (sync)
1197                 inode_sync_wait(inode);
1198         return ret;
1199 }
1200 EXPORT_SYMBOL(write_inode_now);
1201
1202 /**
1203  * sync_inode - write an inode and its pages to disk.
1204  * @inode: the inode to sync
1205  * @wbc: controls the writeback mode
1206  *
1207  * sync_inode() will write an inode and its pages to disk.  It will also
1208  * correctly update the inode on its superblock's dirty inode lists and will
1209  * update inode->i_state.
1210  *
1211  * The caller must have a ref on the inode.
1212  */
1213 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1214 {
1215         int ret;
1216
1217         spin_lock(&inode_lock);
1218         ret = writeback_single_inode(inode, wbc);
1219         spin_unlock(&inode_lock);
1220         return ret;
1221 }
1222 EXPORT_SYMBOL(sync_inode);