mm: avoid resetting wb_start after each writeback round
[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 "internal.h"
30
31 #define inode_to_bdi(inode)     ((inode)->i_mapping->backing_dev_info)
32
33 /*
34  * We don't actually have pdflush, but this one is exported though /proc...
35  */
36 int nr_pdflush_threads;
37
38 /*
39  * Passed into wb_writeback(), essentially a subset of writeback_control
40  */
41 struct wb_writeback_work {
42         long nr_pages;
43         struct super_block *sb;
44         enum writeback_sync_modes sync_mode;
45         unsigned int for_kupdate:1;
46         unsigned int range_cyclic:1;
47         unsigned int for_background:1;
48
49         struct list_head list;          /* pending work list */
50         struct completion *done;        /* set if the caller waits */
51 };
52
53 /**
54  * writeback_in_progress - determine whether there is writeback in progress
55  * @bdi: the device's backing_dev_info structure.
56  *
57  * Determine whether there is writeback waiting to be handled against a
58  * backing device.
59  */
60 int writeback_in_progress(struct backing_dev_info *bdi)
61 {
62         return !list_empty(&bdi->work_list);
63 }
64
65 static void bdi_queue_work(struct backing_dev_info *bdi,
66                 struct wb_writeback_work *work)
67 {
68         spin_lock(&bdi->wb_lock);
69         list_add_tail(&work->list, &bdi->work_list);
70         spin_unlock(&bdi->wb_lock);
71
72         /*
73          * If the default thread isn't there, make sure we add it. When
74          * it gets created and wakes up, we'll run this work.
75          */
76         if (unlikely(list_empty_careful(&bdi->wb_list)))
77                 wake_up_process(default_backing_dev_info.wb.task);
78         else {
79                 struct bdi_writeback *wb = &bdi->wb;
80
81                 if (wb->task)
82                         wake_up_process(wb->task);
83         }
84 }
85
86 static void
87 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
88                 bool range_cyclic, bool for_background)
89 {
90         struct wb_writeback_work *work;
91
92         /*
93          * This is WB_SYNC_NONE writeback, so if allocation fails just
94          * wakeup the thread for old dirty data writeback
95          */
96         work = kzalloc(sizeof(*work), GFP_ATOMIC);
97         if (!work) {
98                 if (bdi->wb.task)
99                         wake_up_process(bdi->wb.task);
100                 return;
101         }
102
103         work->sync_mode = WB_SYNC_NONE;
104         work->nr_pages  = nr_pages;
105         work->range_cyclic = range_cyclic;
106         work->for_background = for_background;
107
108         bdi_queue_work(bdi, work);
109 }
110
111 /**
112  * bdi_start_writeback - start writeback
113  * @bdi: the backing device to write from
114  * @nr_pages: the number of pages to write
115  *
116  * Description:
117  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
118  *   started when this function returns, we make no guarentees on
119  *   completion. Caller need not hold sb s_umount semaphore.
120  *
121  */
122 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
123 {
124         __bdi_start_writeback(bdi, nr_pages, true, false);
125 }
126
127 /**
128  * bdi_start_background_writeback - start background writeback
129  * @bdi: the backing device to write from
130  *
131  * Description:
132  *   This does WB_SYNC_NONE background writeback. The IO is only
133  *   started when this function returns, we make no guarentees on
134  *   completion. Caller need not hold sb s_umount semaphore.
135  */
136 void bdi_start_background_writeback(struct backing_dev_info *bdi)
137 {
138         __bdi_start_writeback(bdi, LONG_MAX, true, true);
139 }
140
141 /*
142  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
143  * furthest end of its superblock's dirty-inode list.
144  *
145  * Before stamping the inode's ->dirtied_when, we check to see whether it is
146  * already the most-recently-dirtied inode on the b_dirty list.  If that is
147  * the case then the inode must have been redirtied while it was being written
148  * out and we don't reset its dirtied_when.
149  */
150 static void redirty_tail(struct inode *inode)
151 {
152         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
153
154         if (!list_empty(&wb->b_dirty)) {
155                 struct inode *tail;
156
157                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
158                 if (time_before(inode->dirtied_when, tail->dirtied_when))
159                         inode->dirtied_when = jiffies;
160         }
161         list_move(&inode->i_list, &wb->b_dirty);
162 }
163
164 /*
165  * requeue inode for re-scanning after bdi->b_io list is exhausted.
166  */
167 static void requeue_io(struct inode *inode)
168 {
169         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
170
171         list_move(&inode->i_list, &wb->b_more_io);
172 }
173
174 static void inode_sync_complete(struct inode *inode)
175 {
176         /*
177          * Prevent speculative execution through spin_unlock(&inode_lock);
178          */
179         smp_mb();
180         wake_up_bit(&inode->i_state, __I_SYNC);
181 }
182
183 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
184 {
185         bool ret = time_after(inode->dirtied_when, t);
186 #ifndef CONFIG_64BIT
187         /*
188          * For inodes being constantly redirtied, dirtied_when can get stuck.
189          * It _appears_ to be in the future, but is actually in distant past.
190          * This test is necessary to prevent such wrapped-around relative times
191          * from permanently stopping the whole bdi writeback.
192          */
193         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
194 #endif
195         return ret;
196 }
197
198 /*
199  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
200  */
201 static void move_expired_inodes(struct list_head *delaying_queue,
202                                struct list_head *dispatch_queue,
203                                 unsigned long *older_than_this)
204 {
205         LIST_HEAD(tmp);
206         struct list_head *pos, *node;
207         struct super_block *sb = NULL;
208         struct inode *inode;
209         int do_sb_sort = 0;
210
211         while (!list_empty(delaying_queue)) {
212                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
213                 if (older_than_this &&
214                     inode_dirtied_after(inode, *older_than_this))
215                         break;
216                 if (sb && sb != inode->i_sb)
217                         do_sb_sort = 1;
218                 sb = inode->i_sb;
219                 list_move(&inode->i_list, &tmp);
220         }
221
222         /* just one sb in list, splice to dispatch_queue and we're done */
223         if (!do_sb_sort) {
224                 list_splice(&tmp, dispatch_queue);
225                 return;
226         }
227
228         /* Move inodes from one superblock together */
229         while (!list_empty(&tmp)) {
230                 inode = list_entry(tmp.prev, struct inode, i_list);
231                 sb = inode->i_sb;
232                 list_for_each_prev_safe(pos, node, &tmp) {
233                         inode = list_entry(pos, struct inode, i_list);
234                         if (inode->i_sb == sb)
235                                 list_move(&inode->i_list, dispatch_queue);
236                 }
237         }
238 }
239
240 /*
241  * Queue all expired dirty inodes for io, eldest first.
242  */
243 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
244 {
245         list_splice_init(&wb->b_more_io, wb->b_io.prev);
246         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
247 }
248
249 static int write_inode(struct inode *inode, struct writeback_control *wbc)
250 {
251         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
252                 return inode->i_sb->s_op->write_inode(inode, wbc);
253         return 0;
254 }
255
256 /*
257  * Wait for writeback on an inode to complete.
258  */
259 static void inode_wait_for_writeback(struct inode *inode)
260 {
261         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
262         wait_queue_head_t *wqh;
263
264         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
265          while (inode->i_state & I_SYNC) {
266                 spin_unlock(&inode_lock);
267                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
268                 spin_lock(&inode_lock);
269         }
270 }
271
272 /*
273  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
274  * caller has ref on the inode (either via __iget or via syscall against an fd)
275  * or the inode has I_WILL_FREE set (via generic_forget_inode)
276  *
277  * If `wait' is set, wait on the writeout.
278  *
279  * The whole writeout design is quite complex and fragile.  We want to avoid
280  * starvation of particular inodes when others are being redirtied, prevent
281  * livelocks, etc.
282  *
283  * Called under inode_lock.
284  */
285 static int
286 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
287 {
288         struct address_space *mapping = inode->i_mapping;
289         unsigned dirty;
290         int ret;
291
292         if (!atomic_read(&inode->i_count))
293                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
294         else
295                 WARN_ON(inode->i_state & I_WILL_FREE);
296
297         if (inode->i_state & I_SYNC) {
298                 /*
299                  * If this inode is locked for writeback and we are not doing
300                  * writeback-for-data-integrity, move it to b_more_io so that
301                  * writeback can proceed with the other inodes on s_io.
302                  *
303                  * We'll have another go at writing back this inode when we
304                  * completed a full scan of b_io.
305                  */
306                 if (wbc->sync_mode != WB_SYNC_ALL) {
307                         requeue_io(inode);
308                         return 0;
309                 }
310
311                 /*
312                  * It's a data-integrity sync.  We must wait.
313                  */
314                 inode_wait_for_writeback(inode);
315         }
316
317         BUG_ON(inode->i_state & I_SYNC);
318
319         /* Set I_SYNC, reset I_DIRTY_PAGES */
320         inode->i_state |= I_SYNC;
321         inode->i_state &= ~I_DIRTY_PAGES;
322         spin_unlock(&inode_lock);
323
324         ret = do_writepages(mapping, wbc);
325
326         /*
327          * Make sure to wait on the data before writing out the metadata.
328          * This is important for filesystems that modify metadata on data
329          * I/O completion.
330          */
331         if (wbc->sync_mode == WB_SYNC_ALL) {
332                 int err = filemap_fdatawait(mapping);
333                 if (ret == 0)
334                         ret = err;
335         }
336
337         /*
338          * Some filesystems may redirty the inode during the writeback
339          * due to delalloc, clear dirty metadata flags right before
340          * write_inode()
341          */
342         spin_lock(&inode_lock);
343         dirty = inode->i_state & I_DIRTY;
344         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
345         spin_unlock(&inode_lock);
346         /* Don't write the inode if only I_DIRTY_PAGES was set */
347         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
348                 int err = write_inode(inode, wbc);
349                 if (ret == 0)
350                         ret = err;
351         }
352
353         spin_lock(&inode_lock);
354         inode->i_state &= ~I_SYNC;
355         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
356                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
357                         /*
358                          * More pages get dirtied by a fast dirtier.
359                          */
360                         goto select_queue;
361                 } else if (inode->i_state & I_DIRTY) {
362                         /*
363                          * At least XFS will redirty the inode during the
364                          * writeback (delalloc) and on io completion (isize).
365                          */
366                         redirty_tail(inode);
367                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
368                         /*
369                          * We didn't write back all the pages.  nfs_writepages()
370                          * sometimes bales out without doing anything. Redirty
371                          * the inode; Move it from b_io onto b_more_io/b_dirty.
372                          */
373                         /*
374                          * akpm: if the caller was the kupdate function we put
375                          * this inode at the head of b_dirty so it gets first
376                          * consideration.  Otherwise, move it to the tail, for
377                          * the reasons described there.  I'm not really sure
378                          * how much sense this makes.  Presumably I had a good
379                          * reasons for doing it this way, and I'd rather not
380                          * muck with it at present.
381                          */
382                         if (wbc->for_kupdate) {
383                                 /*
384                                  * For the kupdate function we move the inode
385                                  * to b_more_io so it will get more writeout as
386                                  * soon as the queue becomes uncongested.
387                                  */
388                                 inode->i_state |= I_DIRTY_PAGES;
389 select_queue:
390                                 if (wbc->nr_to_write <= 0) {
391                                         /*
392                                          * slice used up: queue for next turn
393                                          */
394                                         requeue_io(inode);
395                                 } else {
396                                         /*
397                                          * somehow blocked: retry later
398                                          */
399                                         redirty_tail(inode);
400                                 }
401                         } else {
402                                 /*
403                                  * Otherwise fully redirty the inode so that
404                                  * other inodes on this superblock will get some
405                                  * writeout.  Otherwise heavy writing to one
406                                  * file would indefinitely suspend writeout of
407                                  * all the other files.
408                                  */
409                                 inode->i_state |= I_DIRTY_PAGES;
410                                 redirty_tail(inode);
411                         }
412                 } else if (atomic_read(&inode->i_count)) {
413                         /*
414                          * The inode is clean, inuse
415                          */
416                         list_move(&inode->i_list, &inode_in_use);
417                 } else {
418                         /*
419                          * The inode is clean, unused
420                          */
421                         list_move(&inode->i_list, &inode_unused);
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 = list_entry(wb->b_io.prev,
470                                                  struct inode, i_list);
471
472                 if (inode->i_sb != sb) {
473                         if (only_this_sb) {
474                                 /*
475                                  * We only want to write back data for this
476                                  * superblock, move all inodes not belonging
477                                  * to it back onto the dirty list.
478                                  */
479                                 redirty_tail(inode);
480                                 continue;
481                         }
482
483                         /*
484                          * The inode belongs to a different superblock.
485                          * Bounce back to the caller to unpin this and
486                          * pin the next superblock.
487                          */
488                         return 0;
489                 }
490
491                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
492                         requeue_io(inode);
493                         continue;
494                 }
495                 /*
496                  * Was this inode dirtied after sync_sb_inodes was called?
497                  * This keeps sync from extra jobs and livelock.
498                  */
499                 if (inode_dirtied_after(inode, wbc->wb_start))
500                         return 1;
501
502                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
503                 __iget(inode);
504                 pages_skipped = wbc->pages_skipped;
505                 writeback_single_inode(inode, wbc);
506                 if (wbc->pages_skipped != pages_skipped) {
507                         /*
508                          * writeback is not making progress due to locked
509                          * buffers.  Skip this inode for now.
510                          */
511                         redirty_tail(inode);
512                 }
513                 spin_unlock(&inode_lock);
514                 iput(inode);
515                 cond_resched();
516                 spin_lock(&inode_lock);
517                 if (wbc->nr_to_write <= 0) {
518                         wbc->more_io = 1;
519                         return 1;
520                 }
521                 if (!list_empty(&wb->b_more_io))
522                         wbc->more_io = 1;
523         }
524         /* b_io is empty */
525         return 1;
526 }
527
528 void writeback_inodes_wb(struct bdi_writeback *wb,
529                 struct writeback_control *wbc)
530 {
531         int ret = 0;
532
533         if (!wbc->wb_start)
534                 wbc->wb_start = jiffies; /* livelock avoidance */
535         spin_lock(&inode_lock);
536         if (!wbc->for_kupdate || list_empty(&wb->b_io))
537                 queue_io(wb, wbc->older_than_this);
538
539         while (!list_empty(&wb->b_io)) {
540                 struct inode *inode = list_entry(wb->b_io.prev,
541                                                  struct inode, i_list);
542                 struct super_block *sb = inode->i_sb;
543
544                 if (!pin_sb_for_writeback(sb)) {
545                         requeue_io(inode);
546                         continue;
547                 }
548                 ret = writeback_sb_inodes(sb, wb, wbc, false);
549                 drop_super(sb);
550
551                 if (ret)
552                         break;
553         }
554         spin_unlock(&inode_lock);
555         /* Leave any unwritten inodes on b_io */
556 }
557
558 static void __writeback_inodes_sb(struct super_block *sb,
559                 struct bdi_writeback *wb, struct writeback_control *wbc)
560 {
561         WARN_ON(!rwsem_is_locked(&sb->s_umount));
562
563         spin_lock(&inode_lock);
564         if (!wbc->for_kupdate || list_empty(&wb->b_io))
565                 queue_io(wb, wbc->older_than_this);
566         writeback_sb_inodes(sb, wb, wbc, true);
567         spin_unlock(&inode_lock);
568 }
569
570 /*
571  * The maximum number of pages to writeout in a single bdi flush/kupdate
572  * operation.  We do this so we don't hold I_SYNC against an inode for
573  * enormous amounts of time, which would block a userspace task which has
574  * been forced to throttle against that inode.  Also, the code reevaluates
575  * the dirty each time it has written this many pages.
576  */
577 #define MAX_WRITEBACK_PAGES     1024
578
579 static inline bool over_bground_thresh(void)
580 {
581         unsigned long background_thresh, dirty_thresh;
582
583         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
584
585         return (global_page_state(NR_FILE_DIRTY) +
586                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
587 }
588
589 /*
590  * Explicit flushing or periodic writeback of "old" data.
591  *
592  * Define "old": the first time one of an inode's pages is dirtied, we mark the
593  * dirtying-time in the inode's address_space.  So this periodic writeback code
594  * just walks the superblock inode list, writing back any inodes which are
595  * older than a specific point in time.
596  *
597  * Try to run once per dirty_writeback_interval.  But if a writeback event
598  * takes longer than a dirty_writeback_interval interval, then leave a
599  * one-second gap.
600  *
601  * older_than_this takes precedence over nr_to_write.  So we'll only write back
602  * all dirty pages if they are all attached to "old" mappings.
603  */
604 static long wb_writeback(struct bdi_writeback *wb,
605                          struct wb_writeback_work *work)
606 {
607         struct writeback_control wbc = {
608                 .sync_mode              = work->sync_mode,
609                 .older_than_this        = NULL,
610                 .for_kupdate            = work->for_kupdate,
611                 .for_background         = work->for_background,
612                 .range_cyclic           = work->range_cyclic,
613         };
614         unsigned long oldest_jif;
615         long wrote = 0;
616         struct inode *inode;
617
618         if (wbc.for_kupdate) {
619                 wbc.older_than_this = &oldest_jif;
620                 oldest_jif = jiffies -
621                                 msecs_to_jiffies(dirty_expire_interval * 10);
622         }
623         if (!wbc.range_cyclic) {
624                 wbc.range_start = 0;
625                 wbc.range_end = LLONG_MAX;
626         }
627
628         wbc.wb_start = jiffies; /* livelock avoidance */
629         for (;;) {
630                 /*
631                  * Stop writeback when nr_pages has been consumed
632                  */
633                 if (work->nr_pages <= 0)
634                         break;
635
636                 /*
637                  * For background writeout, stop when we are below the
638                  * background dirty threshold
639                  */
640                 if (work->for_background && !over_bground_thresh())
641                         break;
642
643                 wbc.more_io = 0;
644                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
645                 wbc.pages_skipped = 0;
646                 if (work->sb)
647                         __writeback_inodes_sb(work->sb, wb, &wbc);
648                 else
649                         writeback_inodes_wb(wb, &wbc);
650                 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
651                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
652
653                 /*
654                  * If we consumed everything, see if we have more
655                  */
656                 if (wbc.nr_to_write <= 0)
657                         continue;
658                 /*
659                  * Didn't write everything and we don't have more IO, bail
660                  */
661                 if (!wbc.more_io)
662                         break;
663                 /*
664                  * Did we write something? Try for more
665                  */
666                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
667                         continue;
668                 /*
669                  * Nothing written. Wait for some inode to
670                  * become available for writeback. Otherwise
671                  * we'll just busyloop.
672                  */
673                 spin_lock(&inode_lock);
674                 if (!list_empty(&wb->b_more_io))  {
675                         inode = list_entry(wb->b_more_io.prev,
676                                                 struct inode, i_list);
677                         inode_wait_for_writeback(inode);
678                 }
679                 spin_unlock(&inode_lock);
680         }
681
682         return wrote;
683 }
684
685 /*
686  * Return the next wb_writeback_work struct that hasn't been processed yet.
687  */
688 static struct wb_writeback_work *
689 get_next_work_item(struct backing_dev_info *bdi, struct bdi_writeback *wb)
690 {
691         struct wb_writeback_work *work = NULL;
692
693         spin_lock(&bdi->wb_lock);
694         if (!list_empty(&bdi->work_list)) {
695                 work = list_entry(bdi->work_list.next,
696                                   struct wb_writeback_work, list);
697                 list_del_init(&work->list);
698         }
699         spin_unlock(&bdi->wb_lock);
700         return work;
701 }
702
703 static long wb_check_old_data_flush(struct bdi_writeback *wb)
704 {
705         unsigned long expired;
706         long nr_pages;
707
708         /*
709          * When set to zero, disable periodic writeback
710          */
711         if (!dirty_writeback_interval)
712                 return 0;
713
714         expired = wb->last_old_flush +
715                         msecs_to_jiffies(dirty_writeback_interval * 10);
716         if (time_before(jiffies, expired))
717                 return 0;
718
719         wb->last_old_flush = jiffies;
720         nr_pages = global_page_state(NR_FILE_DIRTY) +
721                         global_page_state(NR_UNSTABLE_NFS) +
722                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
723
724         if (nr_pages) {
725                 struct wb_writeback_work work = {
726                         .nr_pages       = nr_pages,
727                         .sync_mode      = WB_SYNC_NONE,
728                         .for_kupdate    = 1,
729                         .range_cyclic   = 1,
730                 };
731
732                 return wb_writeback(wb, &work);
733         }
734
735         return 0;
736 }
737
738 /*
739  * Retrieve work items and do the writeback they describe
740  */
741 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
742 {
743         struct backing_dev_info *bdi = wb->bdi;
744         struct wb_writeback_work *work;
745         long wrote = 0;
746
747         while ((work = get_next_work_item(bdi, wb)) != NULL) {
748                 /*
749                  * Override sync mode, in case we must wait for completion
750                  * because this thread is exiting now.
751                  */
752                 if (force_wait)
753                         work->sync_mode = WB_SYNC_ALL;
754
755                 wrote += wb_writeback(wb, work);
756
757                 /*
758                  * Notify the caller of completion if this is a synchronous
759                  * work item, otherwise just free it.
760                  */
761                 if (work->done)
762                         complete(work->done);
763                 else
764                         kfree(work);
765         }
766
767         /*
768          * Check for periodic writeback, kupdated() style
769          */
770         wrote += wb_check_old_data_flush(wb);
771
772         return wrote;
773 }
774
775 /*
776  * Handle writeback of dirty data for the device backed by this bdi. Also
777  * wakes up periodically and does kupdated style flushing.
778  */
779 int bdi_writeback_task(struct bdi_writeback *wb)
780 {
781         unsigned long last_active = jiffies;
782         unsigned long wait_jiffies = -1UL;
783         long pages_written;
784
785         while (!kthread_should_stop()) {
786                 pages_written = wb_do_writeback(wb, 0);
787
788                 if (pages_written)
789                         last_active = jiffies;
790                 else if (wait_jiffies != -1UL) {
791                         unsigned long max_idle;
792
793                         /*
794                          * Longest period of inactivity that we tolerate. If we
795                          * see dirty data again later, the task will get
796                          * recreated automatically.
797                          */
798                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
799                         if (time_after(jiffies, max_idle + last_active))
800                                 break;
801                 }
802
803                 if (dirty_writeback_interval) {
804                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
805                         schedule_timeout_interruptible(wait_jiffies);
806                 } else {
807                         set_current_state(TASK_INTERRUPTIBLE);
808                         if (list_empty_careful(&wb->bdi->work_list) &&
809                             !kthread_should_stop())
810                                 schedule();
811                         __set_current_state(TASK_RUNNING);
812                 }
813
814                 try_to_freeze();
815         }
816
817         return 0;
818 }
819
820 /*
821  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
822  * the whole world.
823  */
824 void wakeup_flusher_threads(long nr_pages)
825 {
826         struct backing_dev_info *bdi;
827
828         if (!nr_pages) {
829                 nr_pages = global_page_state(NR_FILE_DIRTY) +
830                                 global_page_state(NR_UNSTABLE_NFS);
831         }
832
833         rcu_read_lock();
834         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
835                 if (!bdi_has_dirty_io(bdi))
836                         continue;
837                 __bdi_start_writeback(bdi, nr_pages, false, false);
838         }
839         rcu_read_unlock();
840 }
841
842 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
843 {
844         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
845                 struct dentry *dentry;
846                 const char *name = "?";
847
848                 dentry = d_find_alias(inode);
849                 if (dentry) {
850                         spin_lock(&dentry->d_lock);
851                         name = (const char *) dentry->d_name.name;
852                 }
853                 printk(KERN_DEBUG
854                        "%s(%d): dirtied inode %lu (%s) on %s\n",
855                        current->comm, task_pid_nr(current), inode->i_ino,
856                        name, inode->i_sb->s_id);
857                 if (dentry) {
858                         spin_unlock(&dentry->d_lock);
859                         dput(dentry);
860                 }
861         }
862 }
863
864 /**
865  *      __mark_inode_dirty -    internal function
866  *      @inode: inode to mark
867  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
868  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
869  *      mark_inode_dirty_sync.
870  *
871  * Put the inode on the super block's dirty list.
872  *
873  * CAREFUL! We mark it dirty unconditionally, but move it onto the
874  * dirty list only if it is hashed or if it refers to a blockdev.
875  * If it was not hashed, it will never be added to the dirty list
876  * even if it is later hashed, as it will have been marked dirty already.
877  *
878  * In short, make sure you hash any inodes _before_ you start marking
879  * them dirty.
880  *
881  * This function *must* be atomic for the I_DIRTY_PAGES case -
882  * set_page_dirty() is called under spinlock in several places.
883  *
884  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
885  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
886  * the kernel-internal blockdev inode represents the dirtying time of the
887  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
888  * page->mapping->host, so the page-dirtying time is recorded in the internal
889  * blockdev inode.
890  */
891 void __mark_inode_dirty(struct inode *inode, int flags)
892 {
893         struct super_block *sb = inode->i_sb;
894
895         /*
896          * Don't do this for I_DIRTY_PAGES - that doesn't actually
897          * dirty the inode itself
898          */
899         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
900                 if (sb->s_op->dirty_inode)
901                         sb->s_op->dirty_inode(inode);
902         }
903
904         /*
905          * make sure that changes are seen by all cpus before we test i_state
906          * -- mikulas
907          */
908         smp_mb();
909
910         /* avoid the locking if we can */
911         if ((inode->i_state & flags) == flags)
912                 return;
913
914         if (unlikely(block_dump))
915                 block_dump___mark_inode_dirty(inode);
916
917         spin_lock(&inode_lock);
918         if ((inode->i_state & flags) != flags) {
919                 const int was_dirty = inode->i_state & I_DIRTY;
920
921                 inode->i_state |= flags;
922
923                 /*
924                  * If the inode is being synced, just update its dirty state.
925                  * The unlocker will place the inode on the appropriate
926                  * superblock list, based upon its state.
927                  */
928                 if (inode->i_state & I_SYNC)
929                         goto out;
930
931                 /*
932                  * Only add valid (hashed) inodes to the superblock's
933                  * dirty list.  Add blockdev inodes as well.
934                  */
935                 if (!S_ISBLK(inode->i_mode)) {
936                         if (hlist_unhashed(&inode->i_hash))
937                                 goto out;
938                 }
939                 if (inode->i_state & (I_FREEING|I_CLEAR))
940                         goto out;
941
942                 /*
943                  * If the inode was already on b_dirty/b_io/b_more_io, don't
944                  * reposition it (that would break b_dirty time-ordering).
945                  */
946                 if (!was_dirty) {
947                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
948                         struct backing_dev_info *bdi = wb->bdi;
949
950                         if (bdi_cap_writeback_dirty(bdi) &&
951                             !test_bit(BDI_registered, &bdi->state)) {
952                                 WARN_ON(1);
953                                 printk(KERN_ERR "bdi-%s not registered\n",
954                                                                 bdi->name);
955                         }
956
957                         inode->dirtied_when = jiffies;
958                         list_move(&inode->i_list, &wb->b_dirty);
959                 }
960         }
961 out:
962         spin_unlock(&inode_lock);
963 }
964 EXPORT_SYMBOL(__mark_inode_dirty);
965
966 /*
967  * Write out a superblock's list of dirty inodes.  A wait will be performed
968  * upon no inodes, all inodes or the final one, depending upon sync_mode.
969  *
970  * If older_than_this is non-NULL, then only write out inodes which
971  * had their first dirtying at a time earlier than *older_than_this.
972  *
973  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
974  * This function assumes that the blockdev superblock's inodes are backed by
975  * a variety of queues, so all inodes are searched.  For other superblocks,
976  * assume that all inodes are backed by the same queue.
977  *
978  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
979  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
980  * on the writer throttling path, and we get decent balancing between many
981  * throttled threads: we don't want them all piling up on inode_sync_wait.
982  */
983 static void wait_sb_inodes(struct super_block *sb)
984 {
985         struct inode *inode, *old_inode = NULL;
986
987         /*
988          * We need to be protected against the filesystem going from
989          * r/o to r/w or vice versa.
990          */
991         WARN_ON(!rwsem_is_locked(&sb->s_umount));
992
993         spin_lock(&inode_lock);
994
995         /*
996          * Data integrity sync. Must wait for all pages under writeback,
997          * because there may have been pages dirtied before our sync
998          * call, but which had writeout started before we write it out.
999          * In which case, the inode may not be on the dirty list, but
1000          * we still have to wait for that writeout.
1001          */
1002         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1003                 struct address_space *mapping;
1004
1005                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1006                         continue;
1007                 mapping = inode->i_mapping;
1008                 if (mapping->nrpages == 0)
1009                         continue;
1010                 __iget(inode);
1011                 spin_unlock(&inode_lock);
1012                 /*
1013                  * We hold a reference to 'inode' so it couldn't have
1014                  * been removed from s_inodes list while we dropped the
1015                  * inode_lock.  We cannot iput the inode now as we can
1016                  * be holding the last reference and we cannot iput it
1017                  * under inode_lock. So we keep the reference and iput
1018                  * it later.
1019                  */
1020                 iput(old_inode);
1021                 old_inode = inode;
1022
1023                 filemap_fdatawait(mapping);
1024
1025                 cond_resched();
1026
1027                 spin_lock(&inode_lock);
1028         }
1029         spin_unlock(&inode_lock);
1030         iput(old_inode);
1031 }
1032
1033 /**
1034  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1035  * @sb: the superblock
1036  *
1037  * Start writeback on some inodes on this super_block. No guarantees are made
1038  * on how many (if any) will be written, and this function does not wait
1039  * for IO completion of submitted IO. The number of pages submitted is
1040  * returned.
1041  */
1042 void writeback_inodes_sb(struct super_block *sb)
1043 {
1044         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1045         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1046         DECLARE_COMPLETION_ONSTACK(done);
1047         struct wb_writeback_work work = {
1048                 .sb             = sb,
1049                 .sync_mode      = WB_SYNC_NONE,
1050                 .done           = &done,
1051         };
1052
1053         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1054
1055         work.nr_pages = nr_dirty + nr_unstable +
1056                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1057
1058         bdi_queue_work(sb->s_bdi, &work);
1059         wait_for_completion(&done);
1060 }
1061 EXPORT_SYMBOL(writeback_inodes_sb);
1062
1063 /**
1064  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1065  * @sb: the superblock
1066  *
1067  * Invoke writeback_inodes_sb if no writeback is currently underway.
1068  * Returns 1 if writeback was started, 0 if not.
1069  */
1070 int writeback_inodes_sb_if_idle(struct super_block *sb)
1071 {
1072         if (!writeback_in_progress(sb->s_bdi)) {
1073                 down_read(&sb->s_umount);
1074                 writeback_inodes_sb(sb);
1075                 up_read(&sb->s_umount);
1076                 return 1;
1077         } else
1078                 return 0;
1079 }
1080 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1081
1082 /**
1083  * sync_inodes_sb       -       sync sb inode pages
1084  * @sb: the superblock
1085  *
1086  * This function writes and waits on any dirty inode belonging to this
1087  * super_block. The number of pages synced is returned.
1088  */
1089 void sync_inodes_sb(struct super_block *sb)
1090 {
1091         DECLARE_COMPLETION_ONSTACK(done);
1092         struct wb_writeback_work work = {
1093                 .sb             = sb,
1094                 .sync_mode      = WB_SYNC_ALL,
1095                 .nr_pages       = LONG_MAX,
1096                 .range_cyclic   = 0,
1097                 .done           = &done,
1098         };
1099
1100         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1101
1102         bdi_queue_work(sb->s_bdi, &work);
1103         wait_for_completion(&done);
1104
1105         wait_sb_inodes(sb);
1106 }
1107 EXPORT_SYMBOL(sync_inodes_sb);
1108
1109 /**
1110  * write_inode_now      -       write an inode to disk
1111  * @inode: inode to write to disk
1112  * @sync: whether the write should be synchronous or not
1113  *
1114  * This function commits an inode to disk immediately if it is dirty. This is
1115  * primarily needed by knfsd.
1116  *
1117  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1118  */
1119 int write_inode_now(struct inode *inode, int sync)
1120 {
1121         int ret;
1122         struct writeback_control wbc = {
1123                 .nr_to_write = LONG_MAX,
1124                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1125                 .range_start = 0,
1126                 .range_end = LLONG_MAX,
1127         };
1128
1129         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1130                 wbc.nr_to_write = 0;
1131
1132         might_sleep();
1133         spin_lock(&inode_lock);
1134         ret = writeback_single_inode(inode, &wbc);
1135         spin_unlock(&inode_lock);
1136         if (sync)
1137                 inode_sync_wait(inode);
1138         return ret;
1139 }
1140 EXPORT_SYMBOL(write_inode_now);
1141
1142 /**
1143  * sync_inode - write an inode and its pages to disk.
1144  * @inode: the inode to sync
1145  * @wbc: controls the writeback mode
1146  *
1147  * sync_inode() will write an inode and its pages to disk.  It will also
1148  * correctly update the inode on its superblock's dirty inode lists and will
1149  * update inode->i_state.
1150  *
1151  * The caller must have a ref on the inode.
1152  */
1153 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1154 {
1155         int ret;
1156
1157         spin_lock(&inode_lock);
1158         ret = writeback_single_inode(inode, wbc);
1159         spin_unlock(&inode_lock);
1160         return ret;
1161 }
1162 EXPORT_SYMBOL(sync_inode);