Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6
[linux-3.10.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 /*
56  * We don't actually have pdflush, but this one is exported though /proc...
57  */
58 int nr_pdflush_threads;
59
60 /**
61  * writeback_in_progress - determine whether there is writeback in progress
62  * @bdi: the device's backing_dev_info structure.
63  *
64  * Determine whether there is writeback waiting to be handled against a
65  * backing device.
66  */
67 int writeback_in_progress(struct backing_dev_info *bdi)
68 {
69         return test_bit(BDI_writeback_running, &bdi->state);
70 }
71
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
73 {
74         struct super_block *sb = inode->i_sb;
75
76         if (strcmp(sb->s_type->name, "bdev") == 0)
77                 return inode->i_mapping->backing_dev_info;
78
79         return sb->s_bdi;
80 }
81
82 static inline struct inode *wb_inode(struct list_head *head)
83 {
84         return list_entry(head, struct inode, i_wb_list);
85 }
86
87 static void bdi_queue_work(struct backing_dev_info *bdi,
88                 struct wb_writeback_work *work)
89 {
90         trace_writeback_queue(bdi, work);
91
92         spin_lock_bh(&bdi->wb_lock);
93         list_add_tail(&work->list, &bdi->work_list);
94         if (bdi->wb.task) {
95                 wake_up_process(bdi->wb.task);
96         } else {
97                 /*
98                  * The bdi thread isn't there, wake up the forker thread which
99                  * will create and run it.
100                  */
101                 trace_writeback_nothread(bdi, work);
102                 wake_up_process(default_backing_dev_info.wb.task);
103         }
104         spin_unlock_bh(&bdi->wb_lock);
105 }
106
107 static void
108 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
109                 bool range_cyclic, bool for_background)
110 {
111         struct wb_writeback_work *work;
112
113         /*
114          * This is WB_SYNC_NONE writeback, so if allocation fails just
115          * wakeup the thread for old dirty data writeback
116          */
117         work = kzalloc(sizeof(*work), GFP_ATOMIC);
118         if (!work) {
119                 if (bdi->wb.task) {
120                         trace_writeback_nowork(bdi);
121                         wake_up_process(bdi->wb.task);
122                 }
123                 return;
124         }
125
126         work->sync_mode = WB_SYNC_NONE;
127         work->nr_pages  = nr_pages;
128         work->range_cyclic = range_cyclic;
129         work->for_background = for_background;
130
131         bdi_queue_work(bdi, work);
132 }
133
134 /**
135  * bdi_start_writeback - start writeback
136  * @bdi: the backing device to write from
137  * @nr_pages: the number of pages to write
138  *
139  * Description:
140  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
141  *   started when this function returns, we make no guarentees on
142  *   completion. Caller need not hold sb s_umount semaphore.
143  *
144  */
145 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
146 {
147         __bdi_start_writeback(bdi, nr_pages, true, false);
148 }
149
150 /**
151  * bdi_start_background_writeback - start background writeback
152  * @bdi: the backing device to write from
153  *
154  * Description:
155  *   This does WB_SYNC_NONE background writeback. The IO is only
156  *   started when this function returns, we make no guarentees on
157  *   completion. Caller need not hold sb s_umount semaphore.
158  */
159 void bdi_start_background_writeback(struct backing_dev_info *bdi)
160 {
161         __bdi_start_writeback(bdi, LONG_MAX, true, true);
162 }
163
164 /*
165  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
166  * furthest end of its superblock's dirty-inode list.
167  *
168  * Before stamping the inode's ->dirtied_when, we check to see whether it is
169  * already the most-recently-dirtied inode on the b_dirty list.  If that is
170  * the case then the inode must have been redirtied while it was being written
171  * out and we don't reset its dirtied_when.
172  */
173 static void redirty_tail(struct inode *inode)
174 {
175         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
176
177         if (!list_empty(&wb->b_dirty)) {
178                 struct inode *tail;
179
180                 tail = wb_inode(wb->b_dirty.next);
181                 if (time_before(inode->dirtied_when, tail->dirtied_when))
182                         inode->dirtied_when = jiffies;
183         }
184         list_move(&inode->i_wb_list, &wb->b_dirty);
185 }
186
187 /*
188  * requeue inode for re-scanning after bdi->b_io list is exhausted.
189  */
190 static void requeue_io(struct inode *inode)
191 {
192         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
193
194         list_move(&inode->i_wb_list, &wb->b_more_io);
195 }
196
197 static void inode_sync_complete(struct inode *inode)
198 {
199         /*
200          * Prevent speculative execution through spin_unlock(&inode_lock);
201          */
202         smp_mb();
203         wake_up_bit(&inode->i_state, __I_SYNC);
204 }
205
206 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
207 {
208         bool ret = time_after(inode->dirtied_when, t);
209 #ifndef CONFIG_64BIT
210         /*
211          * For inodes being constantly redirtied, dirtied_when can get stuck.
212          * It _appears_ to be in the future, but is actually in distant past.
213          * This test is necessary to prevent such wrapped-around relative times
214          * from permanently stopping the whole bdi writeback.
215          */
216         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
217 #endif
218         return ret;
219 }
220
221 /*
222  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
223  */
224 static void move_expired_inodes(struct list_head *delaying_queue,
225                                struct list_head *dispatch_queue,
226                                 unsigned long *older_than_this)
227 {
228         LIST_HEAD(tmp);
229         struct list_head *pos, *node;
230         struct super_block *sb = NULL;
231         struct inode *inode;
232         int do_sb_sort = 0;
233
234         while (!list_empty(delaying_queue)) {
235                 inode = wb_inode(delaying_queue->prev);
236                 if (older_than_this &&
237                     inode_dirtied_after(inode, *older_than_this))
238                         break;
239                 if (sb && sb != inode->i_sb)
240                         do_sb_sort = 1;
241                 sb = inode->i_sb;
242                 list_move(&inode->i_wb_list, &tmp);
243         }
244
245         /* just one sb in list, splice to dispatch_queue and we're done */
246         if (!do_sb_sort) {
247                 list_splice(&tmp, dispatch_queue);
248                 return;
249         }
250
251         /* Move inodes from one superblock together */
252         while (!list_empty(&tmp)) {
253                 sb = wb_inode(tmp.prev)->i_sb;
254                 list_for_each_prev_safe(pos, node, &tmp) {
255                         inode = wb_inode(pos);
256                         if (inode->i_sb == sb)
257                                 list_move(&inode->i_wb_list, dispatch_queue);
258                 }
259         }
260 }
261
262 /*
263  * Queue all expired dirty inodes for io, eldest first.
264  * Before
265  *         newly dirtied     b_dirty    b_io    b_more_io
266  *         =============>    gf         edc     BA
267  * After
268  *         newly dirtied     b_dirty    b_io    b_more_io
269  *         =============>    g          fBAedc
270  *                                           |
271  *                                           +--> dequeue for IO
272  */
273 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
274 {
275         list_splice_init(&wb->b_more_io, &wb->b_io);
276         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
277 }
278
279 static int write_inode(struct inode *inode, struct writeback_control *wbc)
280 {
281         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
282                 return inode->i_sb->s_op->write_inode(inode, wbc);
283         return 0;
284 }
285
286 /*
287  * Wait for writeback on an inode to complete.
288  */
289 static void inode_wait_for_writeback(struct inode *inode)
290 {
291         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
292         wait_queue_head_t *wqh;
293
294         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
295          while (inode->i_state & I_SYNC) {
296                 spin_unlock(&inode_lock);
297                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
298                 spin_lock(&inode_lock);
299         }
300 }
301
302 /*
303  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
304  * caller has ref on the inode (either via __iget or via syscall against an fd)
305  * or the inode has I_WILL_FREE set (via generic_forget_inode)
306  *
307  * If `wait' is set, wait on the writeout.
308  *
309  * The whole writeout design is quite complex and fragile.  We want to avoid
310  * starvation of particular inodes when others are being redirtied, prevent
311  * livelocks, etc.
312  *
313  * Called under inode_lock.
314  */
315 static int
316 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
317 {
318         struct address_space *mapping = inode->i_mapping;
319         unsigned dirty;
320         int ret;
321
322         if (!atomic_read(&inode->i_count))
323                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
324         else
325                 WARN_ON(inode->i_state & I_WILL_FREE);
326
327         if (inode->i_state & I_SYNC) {
328                 /*
329                  * If this inode is locked for writeback and we are not doing
330                  * writeback-for-data-integrity, move it to b_more_io so that
331                  * writeback can proceed with the other inodes on s_io.
332                  *
333                  * We'll have another go at writing back this inode when we
334                  * completed a full scan of b_io.
335                  */
336                 if (wbc->sync_mode != WB_SYNC_ALL) {
337                         requeue_io(inode);
338                         return 0;
339                 }
340
341                 /*
342                  * It's a data-integrity sync.  We must wait.
343                  */
344                 inode_wait_for_writeback(inode);
345         }
346
347         BUG_ON(inode->i_state & I_SYNC);
348
349         /* Set I_SYNC, reset I_DIRTY_PAGES */
350         inode->i_state |= I_SYNC;
351         inode->i_state &= ~I_DIRTY_PAGES;
352         spin_unlock(&inode_lock);
353
354         ret = do_writepages(mapping, wbc);
355
356         /*
357          * Make sure to wait on the data before writing out the metadata.
358          * This is important for filesystems that modify metadata on data
359          * I/O completion.
360          */
361         if (wbc->sync_mode == WB_SYNC_ALL) {
362                 int err = filemap_fdatawait(mapping);
363                 if (ret == 0)
364                         ret = err;
365         }
366
367         /*
368          * Some filesystems may redirty the inode during the writeback
369          * due to delalloc, clear dirty metadata flags right before
370          * write_inode()
371          */
372         spin_lock(&inode_lock);
373         dirty = inode->i_state & I_DIRTY;
374         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
375         spin_unlock(&inode_lock);
376         /* Don't write the inode if only I_DIRTY_PAGES was set */
377         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
378                 int err = write_inode(inode, wbc);
379                 if (ret == 0)
380                         ret = err;
381         }
382
383         spin_lock(&inode_lock);
384         inode->i_state &= ~I_SYNC;
385         if (!(inode->i_state & I_FREEING)) {
386                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
387                         /*
388                          * We didn't write back all the pages.  nfs_writepages()
389                          * sometimes bales out without doing anything.
390                          */
391                         inode->i_state |= I_DIRTY_PAGES;
392                         if (wbc->nr_to_write <= 0) {
393                                 /*
394                                  * slice used up: queue for next turn
395                                  */
396                                 requeue_io(inode);
397                         } else {
398                                 /*
399                                  * Writeback blocked by something other than
400                                  * congestion. Delay the inode for some time to
401                                  * avoid spinning on the CPU (100% iowait)
402                                  * retrying writeback of the dirty page/inode
403                                  * that cannot be performed immediately.
404                                  */
405                                 redirty_tail(inode);
406                         }
407                 } else if (inode->i_state & I_DIRTY) {
408                         /*
409                          * Filesystems can dirty the inode during writeback
410                          * operations, such as delayed allocation during
411                          * submission or metadata updates after data IO
412                          * completion.
413                          */
414                         redirty_tail(inode);
415                 } else {
416                         /*
417                          * The inode is clean.  At this point we either have
418                          * a reference to the inode or it's on it's way out.
419                          * No need to add it back to the LRU.
420                          */
421                         list_del_init(&inode->i_wb_list);
422                 }
423         }
424         inode_sync_complete(inode);
425         return ret;
426 }
427
428 /*
429  * For background writeback the caller does not have the sb pinned
430  * before calling writeback. So make sure that we do pin it, so it doesn't
431  * go away while we are writing inodes from it.
432  */
433 static bool pin_sb_for_writeback(struct super_block *sb)
434 {
435         spin_lock(&sb_lock);
436         if (list_empty(&sb->s_instances)) {
437                 spin_unlock(&sb_lock);
438                 return false;
439         }
440
441         sb->s_count++;
442         spin_unlock(&sb_lock);
443
444         if (down_read_trylock(&sb->s_umount)) {
445                 if (sb->s_root)
446                         return true;
447                 up_read(&sb->s_umount);
448         }
449
450         put_super(sb);
451         return false;
452 }
453
454 /*
455  * Write a portion of b_io inodes which belong to @sb.
456  *
457  * If @only_this_sb is true, then find and write all such
458  * inodes. Otherwise write only ones which go sequentially
459  * in reverse order.
460  *
461  * Return 1, if the caller writeback routine should be
462  * interrupted. Otherwise return 0.
463  */
464 static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
465                 struct writeback_control *wbc, bool only_this_sb)
466 {
467         while (!list_empty(&wb->b_io)) {
468                 long pages_skipped;
469                 struct inode *inode = wb_inode(wb->b_io.prev);
470
471                 if (inode->i_sb != sb) {
472                         if (only_this_sb) {
473                                 /*
474                                  * We only want to write back data for this
475                                  * superblock, move all inodes not belonging
476                                  * to it back onto the dirty list.
477                                  */
478                                 redirty_tail(inode);
479                                 continue;
480                         }
481
482                         /*
483                          * The inode belongs to a different superblock.
484                          * Bounce back to the caller to unpin this and
485                          * pin the next superblock.
486                          */
487                         return 0;
488                 }
489
490                 /*
491                  * Don't bother with new inodes or inodes beeing freed, first
492                  * kind does not need peridic writeout yet, and for the latter
493                  * kind writeout is handled by the freer.
494                  */
495                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
496                         requeue_io(inode);
497                         continue;
498                 }
499
500                 /*
501                  * Was this inode dirtied after sync_sb_inodes was called?
502                  * This keeps sync from extra jobs and livelock.
503                  */
504                 if (inode_dirtied_after(inode, wbc->wb_start))
505                         return 1;
506
507                 __iget(inode);
508                 pages_skipped = wbc->pages_skipped;
509                 writeback_single_inode(inode, wbc);
510                 if (wbc->pages_skipped != pages_skipped) {
511                         /*
512                          * writeback is not making progress due to locked
513                          * buffers.  Skip this inode for now.
514                          */
515                         redirty_tail(inode);
516                 }
517                 spin_unlock(&inode_lock);
518                 iput(inode);
519                 cond_resched();
520                 spin_lock(&inode_lock);
521                 if (wbc->nr_to_write <= 0) {
522                         wbc->more_io = 1;
523                         return 1;
524                 }
525                 if (!list_empty(&wb->b_more_io))
526                         wbc->more_io = 1;
527         }
528         /* b_io is empty */
529         return 1;
530 }
531
532 void writeback_inodes_wb(struct bdi_writeback *wb,
533                 struct writeback_control *wbc)
534 {
535         int ret = 0;
536
537         if (!wbc->wb_start)
538                 wbc->wb_start = jiffies; /* livelock avoidance */
539         spin_lock(&inode_lock);
540         if (!wbc->for_kupdate || list_empty(&wb->b_io))
541                 queue_io(wb, wbc->older_than_this);
542
543         while (!list_empty(&wb->b_io)) {
544                 struct inode *inode = wb_inode(wb->b_io.prev);
545                 struct super_block *sb = inode->i_sb;
546
547                 if (!pin_sb_for_writeback(sb)) {
548                         requeue_io(inode);
549                         continue;
550                 }
551                 ret = writeback_sb_inodes(sb, wb, wbc, false);
552                 drop_super(sb);
553
554                 if (ret)
555                         break;
556         }
557         spin_unlock(&inode_lock);
558         /* Leave any unwritten inodes on b_io */
559 }
560
561 static void __writeback_inodes_sb(struct super_block *sb,
562                 struct bdi_writeback *wb, struct writeback_control *wbc)
563 {
564         WARN_ON(!rwsem_is_locked(&sb->s_umount));
565
566         spin_lock(&inode_lock);
567         if (!wbc->for_kupdate || list_empty(&wb->b_io))
568                 queue_io(wb, wbc->older_than_this);
569         writeback_sb_inodes(sb, wb, wbc, true);
570         spin_unlock(&inode_lock);
571 }
572
573 /*
574  * The maximum number of pages to writeout in a single bdi flush/kupdate
575  * operation.  We do this so we don't hold I_SYNC against an inode for
576  * enormous amounts of time, which would block a userspace task which has
577  * been forced to throttle against that inode.  Also, the code reevaluates
578  * the dirty each time it has written this many pages.
579  */
580 #define MAX_WRITEBACK_PAGES     1024
581
582 static inline bool over_bground_thresh(void)
583 {
584         unsigned long background_thresh, dirty_thresh;
585
586         global_dirty_limits(&background_thresh, &dirty_thresh);
587
588         return (global_page_state(NR_FILE_DIRTY) +
589                 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
590 }
591
592 /*
593  * Explicit flushing or periodic writeback of "old" data.
594  *
595  * Define "old": the first time one of an inode's pages is dirtied, we mark the
596  * dirtying-time in the inode's address_space.  So this periodic writeback code
597  * just walks the superblock inode list, writing back any inodes which are
598  * older than a specific point in time.
599  *
600  * Try to run once per dirty_writeback_interval.  But if a writeback event
601  * takes longer than a dirty_writeback_interval interval, then leave a
602  * one-second gap.
603  *
604  * older_than_this takes precedence over nr_to_write.  So we'll only write back
605  * all dirty pages if they are all attached to "old" mappings.
606  */
607 static long wb_writeback(struct bdi_writeback *wb,
608                          struct wb_writeback_work *work)
609 {
610         struct writeback_control wbc = {
611                 .sync_mode              = work->sync_mode,
612                 .older_than_this        = NULL,
613                 .for_kupdate            = work->for_kupdate,
614                 .for_background         = work->for_background,
615                 .range_cyclic           = work->range_cyclic,
616         };
617         unsigned long oldest_jif;
618         long wrote = 0;
619         struct inode *inode;
620
621         if (wbc.for_kupdate) {
622                 wbc.older_than_this = &oldest_jif;
623                 oldest_jif = jiffies -
624                                 msecs_to_jiffies(dirty_expire_interval * 10);
625         }
626         if (!wbc.range_cyclic) {
627                 wbc.range_start = 0;
628                 wbc.range_end = LLONG_MAX;
629         }
630
631         wbc.wb_start = jiffies; /* livelock avoidance */
632         for (;;) {
633                 /*
634                  * Stop writeback when nr_pages has been consumed
635                  */
636                 if (work->nr_pages <= 0)
637                         break;
638
639                 /*
640                  * For background writeout, stop when we are below the
641                  * background dirty threshold
642                  */
643                 if (work->for_background && !over_bground_thresh())
644                         break;
645
646                 wbc.more_io = 0;
647                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
648                 wbc.pages_skipped = 0;
649
650                 trace_wbc_writeback_start(&wbc, wb->bdi);
651                 if (work->sb)
652                         __writeback_inodes_sb(work->sb, wb, &wbc);
653                 else
654                         writeback_inodes_wb(wb, &wbc);
655                 trace_wbc_writeback_written(&wbc, wb->bdi);
656
657                 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
658                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
659
660                 /*
661                  * If we consumed everything, see if we have more
662                  */
663                 if (wbc.nr_to_write <= 0)
664                         continue;
665                 /*
666                  * Didn't write everything and we don't have more IO, bail
667                  */
668                 if (!wbc.more_io)
669                         break;
670                 /*
671                  * Did we write something? Try for more
672                  */
673                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
674                         continue;
675                 /*
676                  * Nothing written. Wait for some inode to
677                  * become available for writeback. Otherwise
678                  * we'll just busyloop.
679                  */
680                 spin_lock(&inode_lock);
681                 if (!list_empty(&wb->b_more_io))  {
682                         inode = wb_inode(wb->b_more_io.prev);
683                         trace_wbc_writeback_wait(&wbc, wb->bdi);
684                         inode_wait_for_writeback(inode);
685                 }
686                 spin_unlock(&inode_lock);
687         }
688
689         return wrote;
690 }
691
692 /*
693  * Return the next wb_writeback_work struct that hasn't been processed yet.
694  */
695 static struct wb_writeback_work *
696 get_next_work_item(struct backing_dev_info *bdi)
697 {
698         struct wb_writeback_work *work = NULL;
699
700         spin_lock_bh(&bdi->wb_lock);
701         if (!list_empty(&bdi->work_list)) {
702                 work = list_entry(bdi->work_list.next,
703                                   struct wb_writeback_work, list);
704                 list_del_init(&work->list);
705         }
706         spin_unlock_bh(&bdi->wb_lock);
707         return work;
708 }
709
710 static long wb_check_old_data_flush(struct bdi_writeback *wb)
711 {
712         unsigned long expired;
713         long nr_pages;
714
715         /*
716          * When set to zero, disable periodic writeback
717          */
718         if (!dirty_writeback_interval)
719                 return 0;
720
721         expired = wb->last_old_flush +
722                         msecs_to_jiffies(dirty_writeback_interval * 10);
723         if (time_before(jiffies, expired))
724                 return 0;
725
726         wb->last_old_flush = jiffies;
727         /*
728          * Add in the number of potentially dirty inodes, because each inode
729          * write can dirty pagecache in the underlying blockdev.
730          */
731         nr_pages = global_page_state(NR_FILE_DIRTY) +
732                         global_page_state(NR_UNSTABLE_NFS) +
733                         get_nr_dirty_inodes();
734
735         if (nr_pages) {
736                 struct wb_writeback_work work = {
737                         .nr_pages       = nr_pages,
738                         .sync_mode      = WB_SYNC_NONE,
739                         .for_kupdate    = 1,
740                         .range_cyclic   = 1,
741                 };
742
743                 return wb_writeback(wb, &work);
744         }
745
746         return 0;
747 }
748
749 /*
750  * Retrieve work items and do the writeback they describe
751  */
752 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
753 {
754         struct backing_dev_info *bdi = wb->bdi;
755         struct wb_writeback_work *work;
756         long wrote = 0;
757
758         set_bit(BDI_writeback_running, &wb->bdi->state);
759         while ((work = get_next_work_item(bdi)) != NULL) {
760                 /*
761                  * Override sync mode, in case we must wait for completion
762                  * because this thread is exiting now.
763                  */
764                 if (force_wait)
765                         work->sync_mode = WB_SYNC_ALL;
766
767                 trace_writeback_exec(bdi, work);
768
769                 wrote += wb_writeback(wb, work);
770
771                 /*
772                  * Notify the caller of completion if this is a synchronous
773                  * work item, otherwise just free it.
774                  */
775                 if (work->done)
776                         complete(work->done);
777                 else
778                         kfree(work);
779         }
780
781         /*
782          * Check for periodic writeback, kupdated() style
783          */
784         wrote += wb_check_old_data_flush(wb);
785         clear_bit(BDI_writeback_running, &wb->bdi->state);
786
787         return wrote;
788 }
789
790 /*
791  * Handle writeback of dirty data for the device backed by this bdi. Also
792  * wakes up periodically and does kupdated style flushing.
793  */
794 int bdi_writeback_thread(void *data)
795 {
796         struct bdi_writeback *wb = data;
797         struct backing_dev_info *bdi = wb->bdi;
798         long pages_written;
799
800         current->flags |= PF_SWAPWRITE;
801         set_freezable();
802         wb->last_active = jiffies;
803
804         /*
805          * Our parent may run at a different priority, just set us to normal
806          */
807         set_user_nice(current, 0);
808
809         trace_writeback_thread_start(bdi);
810
811         while (!kthread_should_stop()) {
812                 /*
813                  * Remove own delayed wake-up timer, since we are already awake
814                  * and we'll take care of the preriodic write-back.
815                  */
816                 del_timer(&wb->wakeup_timer);
817
818                 pages_written = wb_do_writeback(wb, 0);
819
820                 trace_writeback_pages_written(pages_written);
821
822                 if (pages_written)
823                         wb->last_active = jiffies;
824
825                 set_current_state(TASK_INTERRUPTIBLE);
826                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
827                         __set_current_state(TASK_RUNNING);
828                         continue;
829                 }
830
831                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
832                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
833                 else {
834                         /*
835                          * We have nothing to do, so can go sleep without any
836                          * timeout and save power. When a work is queued or
837                          * something is made dirty - we will be woken up.
838                          */
839                         schedule();
840                 }
841
842                 try_to_freeze();
843         }
844
845         /* Flush any work that raced with us exiting */
846         if (!list_empty(&bdi->work_list))
847                 wb_do_writeback(wb, 1);
848
849         trace_writeback_thread_stop(bdi);
850         return 0;
851 }
852
853
854 /*
855  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
856  * the whole world.
857  */
858 void wakeup_flusher_threads(long nr_pages)
859 {
860         struct backing_dev_info *bdi;
861
862         if (!nr_pages) {
863                 nr_pages = global_page_state(NR_FILE_DIRTY) +
864                                 global_page_state(NR_UNSTABLE_NFS);
865         }
866
867         rcu_read_lock();
868         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
869                 if (!bdi_has_dirty_io(bdi))
870                         continue;
871                 __bdi_start_writeback(bdi, nr_pages, false, false);
872         }
873         rcu_read_unlock();
874 }
875
876 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
877 {
878         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
879                 struct dentry *dentry;
880                 const char *name = "?";
881
882                 dentry = d_find_alias(inode);
883                 if (dentry) {
884                         spin_lock(&dentry->d_lock);
885                         name = (const char *) dentry->d_name.name;
886                 }
887                 printk(KERN_DEBUG
888                        "%s(%d): dirtied inode %lu (%s) on %s\n",
889                        current->comm, task_pid_nr(current), inode->i_ino,
890                        name, inode->i_sb->s_id);
891                 if (dentry) {
892                         spin_unlock(&dentry->d_lock);
893                         dput(dentry);
894                 }
895         }
896 }
897
898 /**
899  *      __mark_inode_dirty -    internal function
900  *      @inode: inode to mark
901  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
902  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
903  *      mark_inode_dirty_sync.
904  *
905  * Put the inode on the super block's dirty list.
906  *
907  * CAREFUL! We mark it dirty unconditionally, but move it onto the
908  * dirty list only if it is hashed or if it refers to a blockdev.
909  * If it was not hashed, it will never be added to the dirty list
910  * even if it is later hashed, as it will have been marked dirty already.
911  *
912  * In short, make sure you hash any inodes _before_ you start marking
913  * them dirty.
914  *
915  * This function *must* be atomic for the I_DIRTY_PAGES case -
916  * set_page_dirty() is called under spinlock in several places.
917  *
918  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
919  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
920  * the kernel-internal blockdev inode represents the dirtying time of the
921  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
922  * page->mapping->host, so the page-dirtying time is recorded in the internal
923  * blockdev inode.
924  */
925 void __mark_inode_dirty(struct inode *inode, int flags)
926 {
927         struct super_block *sb = inode->i_sb;
928         struct backing_dev_info *bdi = NULL;
929         bool wakeup_bdi = false;
930
931         /*
932          * Don't do this for I_DIRTY_PAGES - that doesn't actually
933          * dirty the inode itself
934          */
935         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
936                 if (sb->s_op->dirty_inode)
937                         sb->s_op->dirty_inode(inode);
938         }
939
940         /*
941          * make sure that changes are seen by all cpus before we test i_state
942          * -- mikulas
943          */
944         smp_mb();
945
946         /* avoid the locking if we can */
947         if ((inode->i_state & flags) == flags)
948                 return;
949
950         if (unlikely(block_dump))
951                 block_dump___mark_inode_dirty(inode);
952
953         spin_lock(&inode_lock);
954         if ((inode->i_state & flags) != flags) {
955                 const int was_dirty = inode->i_state & I_DIRTY;
956
957                 inode->i_state |= flags;
958
959                 /*
960                  * If the inode is being synced, just update its dirty state.
961                  * The unlocker will place the inode on the appropriate
962                  * superblock list, based upon its state.
963                  */
964                 if (inode->i_state & I_SYNC)
965                         goto out;
966
967                 /*
968                  * Only add valid (hashed) inodes to the superblock's
969                  * dirty list.  Add blockdev inodes as well.
970                  */
971                 if (!S_ISBLK(inode->i_mode)) {
972                         if (inode_unhashed(inode))
973                                 goto out;
974                 }
975                 if (inode->i_state & I_FREEING)
976                         goto out;
977
978                 /*
979                  * If the inode was already on b_dirty/b_io/b_more_io, don't
980                  * reposition it (that would break b_dirty time-ordering).
981                  */
982                 if (!was_dirty) {
983                         bdi = inode_to_bdi(inode);
984
985                         if (bdi_cap_writeback_dirty(bdi)) {
986                                 WARN(!test_bit(BDI_registered, &bdi->state),
987                                      "bdi-%s not registered\n", bdi->name);
988
989                                 /*
990                                  * If this is the first dirty inode for this
991                                  * bdi, we have to wake-up the corresponding
992                                  * bdi thread to make sure background
993                                  * write-back happens later.
994                                  */
995                                 if (!wb_has_dirty_io(&bdi->wb))
996                                         wakeup_bdi = true;
997                         }
998
999                         inode->dirtied_when = jiffies;
1000                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1001                 }
1002         }
1003 out:
1004         spin_unlock(&inode_lock);
1005
1006         if (wakeup_bdi)
1007                 bdi_wakeup_thread_delayed(bdi);
1008 }
1009 EXPORT_SYMBOL(__mark_inode_dirty);
1010
1011 /*
1012  * Write out a superblock's list of dirty inodes.  A wait will be performed
1013  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1014  *
1015  * If older_than_this is non-NULL, then only write out inodes which
1016  * had their first dirtying at a time earlier than *older_than_this.
1017  *
1018  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1019  * This function assumes that the blockdev superblock's inodes are backed by
1020  * a variety of queues, so all inodes are searched.  For other superblocks,
1021  * assume that all inodes are backed by the same queue.
1022  *
1023  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1024  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1025  * on the writer throttling path, and we get decent balancing between many
1026  * throttled threads: we don't want them all piling up on inode_sync_wait.
1027  */
1028 static void wait_sb_inodes(struct super_block *sb)
1029 {
1030         struct inode *inode, *old_inode = NULL;
1031
1032         /*
1033          * We need to be protected against the filesystem going from
1034          * r/o to r/w or vice versa.
1035          */
1036         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1037
1038         spin_lock(&inode_lock);
1039
1040         /*
1041          * Data integrity sync. Must wait for all pages under writeback,
1042          * because there may have been pages dirtied before our sync
1043          * call, but which had writeout started before we write it out.
1044          * In which case, the inode may not be on the dirty list, but
1045          * we still have to wait for that writeout.
1046          */
1047         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1048                 struct address_space *mapping;
1049
1050                 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW))
1051                         continue;
1052                 mapping = inode->i_mapping;
1053                 if (mapping->nrpages == 0)
1054                         continue;
1055                 __iget(inode);
1056                 spin_unlock(&inode_lock);
1057                 /*
1058                  * We hold a reference to 'inode' so it couldn't have
1059                  * been removed from s_inodes list while we dropped the
1060                  * inode_lock.  We cannot iput the inode now as we can
1061                  * be holding the last reference and we cannot iput it
1062                  * under inode_lock. So we keep the reference and iput
1063                  * it later.
1064                  */
1065                 iput(old_inode);
1066                 old_inode = inode;
1067
1068                 filemap_fdatawait(mapping);
1069
1070                 cond_resched();
1071
1072                 spin_lock(&inode_lock);
1073         }
1074         spin_unlock(&inode_lock);
1075         iput(old_inode);
1076 }
1077
1078 /**
1079  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1080  * @sb: the superblock
1081  *
1082  * Start writeback on some inodes on this super_block. No guarantees are made
1083  * on how many (if any) will be written, and this function does not wait
1084  * for IO completion of submitted IO. The number of pages submitted is
1085  * returned.
1086  */
1087 void writeback_inodes_sb(struct super_block *sb)
1088 {
1089         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1090         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1091         DECLARE_COMPLETION_ONSTACK(done);
1092         struct wb_writeback_work work = {
1093                 .sb             = sb,
1094                 .sync_mode      = WB_SYNC_NONE,
1095                 .done           = &done,
1096         };
1097
1098         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1099
1100         work.nr_pages = nr_dirty + nr_unstable + get_nr_dirty_inodes();
1101
1102         bdi_queue_work(sb->s_bdi, &work);
1103         wait_for_completion(&done);
1104 }
1105 EXPORT_SYMBOL(writeback_inodes_sb);
1106
1107 /**
1108  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1109  * @sb: the superblock
1110  *
1111  * Invoke writeback_inodes_sb if no writeback is currently underway.
1112  * Returns 1 if writeback was started, 0 if not.
1113  */
1114 int writeback_inodes_sb_if_idle(struct super_block *sb)
1115 {
1116         if (!writeback_in_progress(sb->s_bdi)) {
1117                 down_read(&sb->s_umount);
1118                 writeback_inodes_sb(sb);
1119                 up_read(&sb->s_umount);
1120                 return 1;
1121         } else
1122                 return 0;
1123 }
1124 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1125
1126 /**
1127  * sync_inodes_sb       -       sync sb inode pages
1128  * @sb: the superblock
1129  *
1130  * This function writes and waits on any dirty inode belonging to this
1131  * super_block. The number of pages synced is returned.
1132  */
1133 void sync_inodes_sb(struct super_block *sb)
1134 {
1135         DECLARE_COMPLETION_ONSTACK(done);
1136         struct wb_writeback_work work = {
1137                 .sb             = sb,
1138                 .sync_mode      = WB_SYNC_ALL,
1139                 .nr_pages       = LONG_MAX,
1140                 .range_cyclic   = 0,
1141                 .done           = &done,
1142         };
1143
1144         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1145
1146         bdi_queue_work(sb->s_bdi, &work);
1147         wait_for_completion(&done);
1148
1149         wait_sb_inodes(sb);
1150 }
1151 EXPORT_SYMBOL(sync_inodes_sb);
1152
1153 /**
1154  * write_inode_now      -       write an inode to disk
1155  * @inode: inode to write to disk
1156  * @sync: whether the write should be synchronous or not
1157  *
1158  * This function commits an inode to disk immediately if it is dirty. This is
1159  * primarily needed by knfsd.
1160  *
1161  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1162  */
1163 int write_inode_now(struct inode *inode, int sync)
1164 {
1165         int ret;
1166         struct writeback_control wbc = {
1167                 .nr_to_write = LONG_MAX,
1168                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1169                 .range_start = 0,
1170                 .range_end = LLONG_MAX,
1171         };
1172
1173         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1174                 wbc.nr_to_write = 0;
1175
1176         might_sleep();
1177         spin_lock(&inode_lock);
1178         ret = writeback_single_inode(inode, &wbc);
1179         spin_unlock(&inode_lock);
1180         if (sync)
1181                 inode_sync_wait(inode);
1182         return ret;
1183 }
1184 EXPORT_SYMBOL(write_inode_now);
1185
1186 /**
1187  * sync_inode - write an inode and its pages to disk.
1188  * @inode: the inode to sync
1189  * @wbc: controls the writeback mode
1190  *
1191  * sync_inode() will write an inode and its pages to disk.  It will also
1192  * correctly update the inode on its superblock's dirty inode lists and will
1193  * update inode->i_state.
1194  *
1195  * The caller must have a ref on the inode.
1196  */
1197 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1198 {
1199         int ret;
1200
1201         spin_lock(&inode_lock);
1202         ret = writeback_single_inode(inode, wbc);
1203         spin_unlock(&inode_lock);
1204         return ret;
1205 }
1206 EXPORT_SYMBOL(sync_inode);
1207
1208 /**
1209  * sync_inode - write an inode to disk
1210  * @inode: the inode to sync
1211  * @wait: wait for I/O to complete.
1212  *
1213  * Write an inode to disk and adjust it's dirty state after completion.
1214  *
1215  * Note: only writes the actual inode, no associated data or other metadata.
1216  */
1217 int sync_inode_metadata(struct inode *inode, int wait)
1218 {
1219         struct writeback_control wbc = {
1220                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1221                 .nr_to_write = 0, /* metadata-only */
1222         };
1223
1224         return sync_inode(inode, &wbc);
1225 }
1226 EXPORT_SYMBOL(sync_inode_metadata);