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