fs: protect inode->i_state with inode->i_lock
[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 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
88 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
89 {
90         if (bdi->wb.task) {
91                 wake_up_process(bdi->wb.task);
92         } else {
93                 /*
94                  * The bdi thread isn't there, wake up the forker thread which
95                  * will create and run it.
96                  */
97                 wake_up_process(default_backing_dev_info.wb.task);
98         }
99 }
100
101 static void bdi_queue_work(struct backing_dev_info *bdi,
102                            struct wb_writeback_work *work)
103 {
104         trace_writeback_queue(bdi, work);
105
106         spin_lock_bh(&bdi->wb_lock);
107         list_add_tail(&work->list, &bdi->work_list);
108         if (!bdi->wb.task)
109                 trace_writeback_nothread(bdi, work);
110         bdi_wakeup_flusher(bdi);
111         spin_unlock_bh(&bdi->wb_lock);
112 }
113
114 static void
115 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
116                       bool range_cyclic)
117 {
118         struct wb_writeback_work *work;
119
120         /*
121          * This is WB_SYNC_NONE writeback, so if allocation fails just
122          * wakeup the thread for old dirty data writeback
123          */
124         work = kzalloc(sizeof(*work), GFP_ATOMIC);
125         if (!work) {
126                 if (bdi->wb.task) {
127                         trace_writeback_nowork(bdi);
128                         wake_up_process(bdi->wb.task);
129                 }
130                 return;
131         }
132
133         work->sync_mode = WB_SYNC_NONE;
134         work->nr_pages  = nr_pages;
135         work->range_cyclic = range_cyclic;
136
137         bdi_queue_work(bdi, work);
138 }
139
140 /**
141  * bdi_start_writeback - start writeback
142  * @bdi: the backing device to write from
143  * @nr_pages: the number of pages to write
144  *
145  * Description:
146  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
147  *   started when this function returns, we make no guarentees on
148  *   completion. Caller need not hold sb s_umount semaphore.
149  *
150  */
151 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
152 {
153         __bdi_start_writeback(bdi, nr_pages, true);
154 }
155
156 /**
157  * bdi_start_background_writeback - start background writeback
158  * @bdi: the backing device to write from
159  *
160  * Description:
161  *   This makes sure WB_SYNC_NONE background writeback happens. When
162  *   this function returns, it is only guaranteed that for given BDI
163  *   some IO is happening if we are over background dirty threshold.
164  *   Caller need not hold sb s_umount semaphore.
165  */
166 void bdi_start_background_writeback(struct backing_dev_info *bdi)
167 {
168         /*
169          * We just wake up the flusher thread. It will perform background
170          * writeback as soon as there is no other work to do.
171          */
172         trace_writeback_wake_background(bdi);
173         spin_lock_bh(&bdi->wb_lock);
174         bdi_wakeup_flusher(bdi);
175         spin_unlock_bh(&bdi->wb_lock);
176 }
177
178 /*
179  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
180  * furthest end of its superblock's dirty-inode list.
181  *
182  * Before stamping the inode's ->dirtied_when, we check to see whether it is
183  * already the most-recently-dirtied inode on the b_dirty list.  If that is
184  * the case then the inode must have been redirtied while it was being written
185  * out and we don't reset its dirtied_when.
186  */
187 static void redirty_tail(struct inode *inode)
188 {
189         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
190
191         if (!list_empty(&wb->b_dirty)) {
192                 struct inode *tail;
193
194                 tail = wb_inode(wb->b_dirty.next);
195                 if (time_before(inode->dirtied_when, tail->dirtied_when))
196                         inode->dirtied_when = jiffies;
197         }
198         list_move(&inode->i_wb_list, &wb->b_dirty);
199 }
200
201 /*
202  * requeue inode for re-scanning after bdi->b_io list is exhausted.
203  */
204 static void requeue_io(struct inode *inode)
205 {
206         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
207
208         list_move(&inode->i_wb_list, &wb->b_more_io);
209 }
210
211 static void inode_sync_complete(struct inode *inode)
212 {
213         /*
214          * Prevent speculative execution through spin_unlock(&inode_lock);
215          */
216         smp_mb();
217         wake_up_bit(&inode->i_state, __I_SYNC);
218 }
219
220 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
221 {
222         bool ret = time_after(inode->dirtied_when, t);
223 #ifndef CONFIG_64BIT
224         /*
225          * For inodes being constantly redirtied, dirtied_when can get stuck.
226          * It _appears_ to be in the future, but is actually in distant past.
227          * This test is necessary to prevent such wrapped-around relative times
228          * from permanently stopping the whole bdi writeback.
229          */
230         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
231 #endif
232         return ret;
233 }
234
235 /*
236  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
237  */
238 static void move_expired_inodes(struct list_head *delaying_queue,
239                                struct list_head *dispatch_queue,
240                                 unsigned long *older_than_this)
241 {
242         LIST_HEAD(tmp);
243         struct list_head *pos, *node;
244         struct super_block *sb = NULL;
245         struct inode *inode;
246         int do_sb_sort = 0;
247
248         while (!list_empty(delaying_queue)) {
249                 inode = wb_inode(delaying_queue->prev);
250                 if (older_than_this &&
251                     inode_dirtied_after(inode, *older_than_this))
252                         break;
253                 if (sb && sb != inode->i_sb)
254                         do_sb_sort = 1;
255                 sb = inode->i_sb;
256                 list_move(&inode->i_wb_list, &tmp);
257         }
258
259         /* just one sb in list, splice to dispatch_queue and we're done */
260         if (!do_sb_sort) {
261                 list_splice(&tmp, dispatch_queue);
262                 return;
263         }
264
265         /* Move inodes from one superblock together */
266         while (!list_empty(&tmp)) {
267                 sb = wb_inode(tmp.prev)->i_sb;
268                 list_for_each_prev_safe(pos, node, &tmp) {
269                         inode = wb_inode(pos);
270                         if (inode->i_sb == sb)
271                                 list_move(&inode->i_wb_list, dispatch_queue);
272                 }
273         }
274 }
275
276 /*
277  * Queue all expired dirty inodes for io, eldest first.
278  * Before
279  *         newly dirtied     b_dirty    b_io    b_more_io
280  *         =============>    gf         edc     BA
281  * After
282  *         newly dirtied     b_dirty    b_io    b_more_io
283  *         =============>    g          fBAedc
284  *                                           |
285  *                                           +--> dequeue for IO
286  */
287 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
288 {
289         list_splice_init(&wb->b_more_io, &wb->b_io);
290         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
291 }
292
293 static int write_inode(struct inode *inode, struct writeback_control *wbc)
294 {
295         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
296                 return inode->i_sb->s_op->write_inode(inode, wbc);
297         return 0;
298 }
299
300 /*
301  * Wait for writeback on an inode to complete.
302  */
303 static void inode_wait_for_writeback(struct inode *inode)
304 {
305         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
306         wait_queue_head_t *wqh;
307
308         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
309         while (inode->i_state & I_SYNC) {
310                 spin_unlock(&inode->i_lock);
311                 spin_unlock(&inode_lock);
312                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
313                 spin_lock(&inode_lock);
314                 spin_lock(&inode->i_lock);
315         }
316 }
317
318 /*
319  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
320  * caller has ref on the inode (either via __iget or via syscall against an fd)
321  * or the inode has I_WILL_FREE set (via generic_forget_inode)
322  *
323  * If `wait' is set, wait on the writeout.
324  *
325  * The whole writeout design is quite complex and fragile.  We want to avoid
326  * starvation of particular inodes when others are being redirtied, prevent
327  * livelocks, etc.
328  *
329  * Called under inode_lock.
330  */
331 static int
332 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
333 {
334         struct address_space *mapping = inode->i_mapping;
335         unsigned dirty;
336         int ret;
337
338         spin_lock(&inode->i_lock);
339         if (!atomic_read(&inode->i_count))
340                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
341         else
342                 WARN_ON(inode->i_state & I_WILL_FREE);
343
344         if (inode->i_state & I_SYNC) {
345                 /*
346                  * If this inode is locked for writeback and we are not doing
347                  * writeback-for-data-integrity, move it to b_more_io so that
348                  * writeback can proceed with the other inodes on s_io.
349                  *
350                  * We'll have another go at writing back this inode when we
351                  * completed a full scan of b_io.
352                  */
353                 if (wbc->sync_mode != WB_SYNC_ALL) {
354                         spin_unlock(&inode->i_lock);
355                         requeue_io(inode);
356                         return 0;
357                 }
358
359                 /*
360                  * It's a data-integrity sync.  We must wait.
361                  */
362                 inode_wait_for_writeback(inode);
363         }
364
365         BUG_ON(inode->i_state & I_SYNC);
366
367         /* Set I_SYNC, reset I_DIRTY_PAGES */
368         inode->i_state |= I_SYNC;
369         inode->i_state &= ~I_DIRTY_PAGES;
370         spin_unlock(&inode->i_lock);
371         spin_unlock(&inode_lock);
372
373         ret = do_writepages(mapping, wbc);
374
375         /*
376          * Make sure to wait on the data before writing out the metadata.
377          * This is important for filesystems that modify metadata on data
378          * I/O completion.
379          */
380         if (wbc->sync_mode == WB_SYNC_ALL) {
381                 int err = filemap_fdatawait(mapping);
382                 if (ret == 0)
383                         ret = err;
384         }
385
386         /*
387          * Some filesystems may redirty the inode during the writeback
388          * due to delalloc, clear dirty metadata flags right before
389          * write_inode()
390          */
391         spin_lock(&inode_lock);
392         spin_lock(&inode->i_lock);
393         dirty = inode->i_state & I_DIRTY;
394         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
395         spin_unlock(&inode->i_lock);
396         spin_unlock(&inode_lock);
397         /* Don't write the inode if only I_DIRTY_PAGES was set */
398         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
399                 int err = write_inode(inode, wbc);
400                 if (ret == 0)
401                         ret = err;
402         }
403
404         spin_lock(&inode_lock);
405         spin_lock(&inode->i_lock);
406         inode->i_state &= ~I_SYNC;
407         if (!(inode->i_state & I_FREEING)) {
408                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
409                         /*
410                          * We didn't write back all the pages.  nfs_writepages()
411                          * sometimes bales out without doing anything.
412                          */
413                         inode->i_state |= I_DIRTY_PAGES;
414                         if (wbc->nr_to_write <= 0) {
415                                 /*
416                                  * slice used up: queue for next turn
417                                  */
418                                 requeue_io(inode);
419                         } else {
420                                 /*
421                                  * Writeback blocked by something other than
422                                  * congestion. Delay the inode for some time to
423                                  * avoid spinning on the CPU (100% iowait)
424                                  * retrying writeback of the dirty page/inode
425                                  * that cannot be performed immediately.
426                                  */
427                                 redirty_tail(inode);
428                         }
429                 } else if (inode->i_state & I_DIRTY) {
430                         /*
431                          * Filesystems can dirty the inode during writeback
432                          * operations, such as delayed allocation during
433                          * submission or metadata updates after data IO
434                          * completion.
435                          */
436                         redirty_tail(inode);
437                 } else {
438                         /*
439                          * The inode is clean.  At this point we either have
440                          * a reference to the inode or it's on it's way out.
441                          * No need to add it back to the LRU.
442                          */
443                         list_del_init(&inode->i_wb_list);
444                 }
445         }
446         inode_sync_complete(inode);
447         spin_unlock(&inode->i_lock);
448         return ret;
449 }
450
451 /*
452  * For background writeback the caller does not have the sb pinned
453  * before calling writeback. So make sure that we do pin it, so it doesn't
454  * go away while we are writing inodes from it.
455  */
456 static bool pin_sb_for_writeback(struct super_block *sb)
457 {
458         spin_lock(&sb_lock);
459         if (list_empty(&sb->s_instances)) {
460                 spin_unlock(&sb_lock);
461                 return false;
462         }
463
464         sb->s_count++;
465         spin_unlock(&sb_lock);
466
467         if (down_read_trylock(&sb->s_umount)) {
468                 if (sb->s_root)
469                         return true;
470                 up_read(&sb->s_umount);
471         }
472
473         put_super(sb);
474         return false;
475 }
476
477 /*
478  * Write a portion of b_io inodes which belong to @sb.
479  *
480  * If @only_this_sb is true, then find and write all such
481  * inodes. Otherwise write only ones which go sequentially
482  * in reverse order.
483  *
484  * Return 1, if the caller writeback routine should be
485  * interrupted. Otherwise return 0.
486  */
487 static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
488                 struct writeback_control *wbc, bool only_this_sb)
489 {
490         while (!list_empty(&wb->b_io)) {
491                 long pages_skipped;
492                 struct inode *inode = wb_inode(wb->b_io.prev);
493
494                 if (inode->i_sb != sb) {
495                         if (only_this_sb) {
496                                 /*
497                                  * We only want to write back data for this
498                                  * superblock, move all inodes not belonging
499                                  * to it back onto the dirty list.
500                                  */
501                                 redirty_tail(inode);
502                                 continue;
503                         }
504
505                         /*
506                          * The inode belongs to a different superblock.
507                          * Bounce back to the caller to unpin this and
508                          * pin the next superblock.
509                          */
510                         return 0;
511                 }
512
513                 /*
514                  * Don't bother with new inodes or inodes beeing freed, first
515                  * kind does not need peridic writeout yet, and for the latter
516                  * kind writeout is handled by the freer.
517                  */
518                 spin_lock(&inode->i_lock);
519                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
520                         spin_unlock(&inode->i_lock);
521                         requeue_io(inode);
522                         continue;
523                 }
524
525                 /*
526                  * Was this inode dirtied after sync_sb_inodes was called?
527                  * This keeps sync from extra jobs and livelock.
528                  */
529                 if (inode_dirtied_after(inode, wbc->wb_start)) {
530                         spin_unlock(&inode->i_lock);
531                         return 1;
532                 }
533
534                 __iget(inode);
535                 spin_unlock(&inode->i_lock);
536
537                 pages_skipped = wbc->pages_skipped;
538                 writeback_single_inode(inode, wbc);
539                 if (wbc->pages_skipped != pages_skipped) {
540                         /*
541                          * writeback is not making progress due to locked
542                          * buffers.  Skip this inode for now.
543                          */
544                         redirty_tail(inode);
545                 }
546                 spin_unlock(&inode_lock);
547                 iput(inode);
548                 cond_resched();
549                 spin_lock(&inode_lock);
550                 if (wbc->nr_to_write <= 0) {
551                         wbc->more_io = 1;
552                         return 1;
553                 }
554                 if (!list_empty(&wb->b_more_io))
555                         wbc->more_io = 1;
556         }
557         /* b_io is empty */
558         return 1;
559 }
560
561 void writeback_inodes_wb(struct bdi_writeback *wb,
562                 struct writeback_control *wbc)
563 {
564         int ret = 0;
565
566         if (!wbc->wb_start)
567                 wbc->wb_start = jiffies; /* livelock avoidance */
568         spin_lock(&inode_lock);
569         if (!wbc->for_kupdate || list_empty(&wb->b_io))
570                 queue_io(wb, wbc->older_than_this);
571
572         while (!list_empty(&wb->b_io)) {
573                 struct inode *inode = wb_inode(wb->b_io.prev);
574                 struct super_block *sb = inode->i_sb;
575
576                 if (!pin_sb_for_writeback(sb)) {
577                         requeue_io(inode);
578                         continue;
579                 }
580                 ret = writeback_sb_inodes(sb, wb, wbc, false);
581                 drop_super(sb);
582
583                 if (ret)
584                         break;
585         }
586         spin_unlock(&inode_lock);
587         /* Leave any unwritten inodes on b_io */
588 }
589
590 static void __writeback_inodes_sb(struct super_block *sb,
591                 struct bdi_writeback *wb, struct writeback_control *wbc)
592 {
593         WARN_ON(!rwsem_is_locked(&sb->s_umount));
594
595         spin_lock(&inode_lock);
596         if (!wbc->for_kupdate || list_empty(&wb->b_io))
597                 queue_io(wb, wbc->older_than_this);
598         writeback_sb_inodes(sb, wb, wbc, true);
599         spin_unlock(&inode_lock);
600 }
601
602 /*
603  * The maximum number of pages to writeout in a single bdi flush/kupdate
604  * operation.  We do this so we don't hold I_SYNC against an inode for
605  * enormous amounts of time, which would block a userspace task which has
606  * been forced to throttle against that inode.  Also, the code reevaluates
607  * the dirty each time it has written this many pages.
608  */
609 #define MAX_WRITEBACK_PAGES     1024
610
611 static inline bool over_bground_thresh(void)
612 {
613         unsigned long background_thresh, dirty_thresh;
614
615         global_dirty_limits(&background_thresh, &dirty_thresh);
616
617         return (global_page_state(NR_FILE_DIRTY) +
618                 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
619 }
620
621 /*
622  * Explicit flushing or periodic writeback of "old" data.
623  *
624  * Define "old": the first time one of an inode's pages is dirtied, we mark the
625  * dirtying-time in the inode's address_space.  So this periodic writeback code
626  * just walks the superblock inode list, writing back any inodes which are
627  * older than a specific point in time.
628  *
629  * Try to run once per dirty_writeback_interval.  But if a writeback event
630  * takes longer than a dirty_writeback_interval interval, then leave a
631  * one-second gap.
632  *
633  * older_than_this takes precedence over nr_to_write.  So we'll only write back
634  * all dirty pages if they are all attached to "old" mappings.
635  */
636 static long wb_writeback(struct bdi_writeback *wb,
637                          struct wb_writeback_work *work)
638 {
639         struct writeback_control wbc = {
640                 .sync_mode              = work->sync_mode,
641                 .older_than_this        = NULL,
642                 .for_kupdate            = work->for_kupdate,
643                 .for_background         = work->for_background,
644                 .range_cyclic           = work->range_cyclic,
645         };
646         unsigned long oldest_jif;
647         long wrote = 0;
648         long write_chunk;
649         struct inode *inode;
650
651         if (wbc.for_kupdate) {
652                 wbc.older_than_this = &oldest_jif;
653                 oldest_jif = jiffies -
654                                 msecs_to_jiffies(dirty_expire_interval * 10);
655         }
656         if (!wbc.range_cyclic) {
657                 wbc.range_start = 0;
658                 wbc.range_end = LLONG_MAX;
659         }
660
661         /*
662          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
663          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
664          * here avoids calling into writeback_inodes_wb() more than once.
665          *
666          * The intended call sequence for WB_SYNC_ALL writeback is:
667          *
668          *      wb_writeback()
669          *          __writeback_inodes_sb()     <== called only once
670          *              write_cache_pages()     <== called once for each inode
671          *                   (quickly) tag currently dirty pages
672          *                   (maybe slowly) sync all tagged pages
673          */
674         if (wbc.sync_mode == WB_SYNC_NONE)
675                 write_chunk = MAX_WRITEBACK_PAGES;
676         else
677                 write_chunk = LONG_MAX;
678
679         wbc.wb_start = jiffies; /* livelock avoidance */
680         for (;;) {
681                 /*
682                  * Stop writeback when nr_pages has been consumed
683                  */
684                 if (work->nr_pages <= 0)
685                         break;
686
687                 /*
688                  * Background writeout and kupdate-style writeback may
689                  * run forever. Stop them if there is other work to do
690                  * so that e.g. sync can proceed. They'll be restarted
691                  * after the other works are all done.
692                  */
693                 if ((work->for_background || work->for_kupdate) &&
694                     !list_empty(&wb->bdi->work_list))
695                         break;
696
697                 /*
698                  * For background writeout, stop when we are below the
699                  * background dirty threshold
700                  */
701                 if (work->for_background && !over_bground_thresh())
702                         break;
703
704                 wbc.more_io = 0;
705                 wbc.nr_to_write = write_chunk;
706                 wbc.pages_skipped = 0;
707
708                 trace_wbc_writeback_start(&wbc, wb->bdi);
709                 if (work->sb)
710                         __writeback_inodes_sb(work->sb, wb, &wbc);
711                 else
712                         writeback_inodes_wb(wb, &wbc);
713                 trace_wbc_writeback_written(&wbc, wb->bdi);
714
715                 work->nr_pages -= write_chunk - wbc.nr_to_write;
716                 wrote += write_chunk - wbc.nr_to_write;
717
718                 /*
719                  * If we consumed everything, see if we have more
720                  */
721                 if (wbc.nr_to_write <= 0)
722                         continue;
723                 /*
724                  * Didn't write everything and we don't have more IO, bail
725                  */
726                 if (!wbc.more_io)
727                         break;
728                 /*
729                  * Did we write something? Try for more
730                  */
731                 if (wbc.nr_to_write < write_chunk)
732                         continue;
733                 /*
734                  * Nothing written. Wait for some inode to
735                  * become available for writeback. Otherwise
736                  * we'll just busyloop.
737                  */
738                 spin_lock(&inode_lock);
739                 if (!list_empty(&wb->b_more_io))  {
740                         inode = wb_inode(wb->b_more_io.prev);
741                         trace_wbc_writeback_wait(&wbc, wb->bdi);
742                         spin_lock(&inode->i_lock);
743                         inode_wait_for_writeback(inode);
744                         spin_unlock(&inode->i_lock);
745                 }
746                 spin_unlock(&inode_lock);
747         }
748
749         return wrote;
750 }
751
752 /*
753  * Return the next wb_writeback_work struct that hasn't been processed yet.
754  */
755 static struct wb_writeback_work *
756 get_next_work_item(struct backing_dev_info *bdi)
757 {
758         struct wb_writeback_work *work = NULL;
759
760         spin_lock_bh(&bdi->wb_lock);
761         if (!list_empty(&bdi->work_list)) {
762                 work = list_entry(bdi->work_list.next,
763                                   struct wb_writeback_work, list);
764                 list_del_init(&work->list);
765         }
766         spin_unlock_bh(&bdi->wb_lock);
767         return work;
768 }
769
770 /*
771  * Add in the number of potentially dirty inodes, because each inode
772  * write can dirty pagecache in the underlying blockdev.
773  */
774 static unsigned long get_nr_dirty_pages(void)
775 {
776         return global_page_state(NR_FILE_DIRTY) +
777                 global_page_state(NR_UNSTABLE_NFS) +
778                 get_nr_dirty_inodes();
779 }
780
781 static long wb_check_background_flush(struct bdi_writeback *wb)
782 {
783         if (over_bground_thresh()) {
784
785                 struct wb_writeback_work work = {
786                         .nr_pages       = LONG_MAX,
787                         .sync_mode      = WB_SYNC_NONE,
788                         .for_background = 1,
789                         .range_cyclic   = 1,
790                 };
791
792                 return wb_writeback(wb, &work);
793         }
794
795         return 0;
796 }
797
798 static long wb_check_old_data_flush(struct bdi_writeback *wb)
799 {
800         unsigned long expired;
801         long nr_pages;
802
803         /*
804          * When set to zero, disable periodic writeback
805          */
806         if (!dirty_writeback_interval)
807                 return 0;
808
809         expired = wb->last_old_flush +
810                         msecs_to_jiffies(dirty_writeback_interval * 10);
811         if (time_before(jiffies, expired))
812                 return 0;
813
814         wb->last_old_flush = jiffies;
815         nr_pages = get_nr_dirty_pages();
816
817         if (nr_pages) {
818                 struct wb_writeback_work work = {
819                         .nr_pages       = nr_pages,
820                         .sync_mode      = WB_SYNC_NONE,
821                         .for_kupdate    = 1,
822                         .range_cyclic   = 1,
823                 };
824
825                 return wb_writeback(wb, &work);
826         }
827
828         return 0;
829 }
830
831 /*
832  * Retrieve work items and do the writeback they describe
833  */
834 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
835 {
836         struct backing_dev_info *bdi = wb->bdi;
837         struct wb_writeback_work *work;
838         long wrote = 0;
839
840         set_bit(BDI_writeback_running, &wb->bdi->state);
841         while ((work = get_next_work_item(bdi)) != NULL) {
842                 /*
843                  * Override sync mode, in case we must wait for completion
844                  * because this thread is exiting now.
845                  */
846                 if (force_wait)
847                         work->sync_mode = WB_SYNC_ALL;
848
849                 trace_writeback_exec(bdi, work);
850
851                 wrote += wb_writeback(wb, work);
852
853                 /*
854                  * Notify the caller of completion if this is a synchronous
855                  * work item, otherwise just free it.
856                  */
857                 if (work->done)
858                         complete(work->done);
859                 else
860                         kfree(work);
861         }
862
863         /*
864          * Check for periodic writeback, kupdated() style
865          */
866         wrote += wb_check_old_data_flush(wb);
867         wrote += wb_check_background_flush(wb);
868         clear_bit(BDI_writeback_running, &wb->bdi->state);
869
870         return wrote;
871 }
872
873 /*
874  * Handle writeback of dirty data for the device backed by this bdi. Also
875  * wakes up periodically and does kupdated style flushing.
876  */
877 int bdi_writeback_thread(void *data)
878 {
879         struct bdi_writeback *wb = data;
880         struct backing_dev_info *bdi = wb->bdi;
881         long pages_written;
882
883         current->flags |= PF_SWAPWRITE;
884         set_freezable();
885         wb->last_active = jiffies;
886
887         /*
888          * Our parent may run at a different priority, just set us to normal
889          */
890         set_user_nice(current, 0);
891
892         trace_writeback_thread_start(bdi);
893
894         while (!kthread_should_stop()) {
895                 /*
896                  * Remove own delayed wake-up timer, since we are already awake
897                  * and we'll take care of the preriodic write-back.
898                  */
899                 del_timer(&wb->wakeup_timer);
900
901                 pages_written = wb_do_writeback(wb, 0);
902
903                 trace_writeback_pages_written(pages_written);
904
905                 if (pages_written)
906                         wb->last_active = jiffies;
907
908                 set_current_state(TASK_INTERRUPTIBLE);
909                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
910                         __set_current_state(TASK_RUNNING);
911                         continue;
912                 }
913
914                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
915                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
916                 else {
917                         /*
918                          * We have nothing to do, so can go sleep without any
919                          * timeout and save power. When a work is queued or
920                          * something is made dirty - we will be woken up.
921                          */
922                         schedule();
923                 }
924
925                 try_to_freeze();
926         }
927
928         /* Flush any work that raced with us exiting */
929         if (!list_empty(&bdi->work_list))
930                 wb_do_writeback(wb, 1);
931
932         trace_writeback_thread_stop(bdi);
933         return 0;
934 }
935
936
937 /*
938  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
939  * the whole world.
940  */
941 void wakeup_flusher_threads(long nr_pages)
942 {
943         struct backing_dev_info *bdi;
944
945         if (!nr_pages) {
946                 nr_pages = global_page_state(NR_FILE_DIRTY) +
947                                 global_page_state(NR_UNSTABLE_NFS);
948         }
949
950         rcu_read_lock();
951         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
952                 if (!bdi_has_dirty_io(bdi))
953                         continue;
954                 __bdi_start_writeback(bdi, nr_pages, false);
955         }
956         rcu_read_unlock();
957 }
958
959 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
960 {
961         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
962                 struct dentry *dentry;
963                 const char *name = "?";
964
965                 dentry = d_find_alias(inode);
966                 if (dentry) {
967                         spin_lock(&dentry->d_lock);
968                         name = (const char *) dentry->d_name.name;
969                 }
970                 printk(KERN_DEBUG
971                        "%s(%d): dirtied inode %lu (%s) on %s\n",
972                        current->comm, task_pid_nr(current), inode->i_ino,
973                        name, inode->i_sb->s_id);
974                 if (dentry) {
975                         spin_unlock(&dentry->d_lock);
976                         dput(dentry);
977                 }
978         }
979 }
980
981 /**
982  *      __mark_inode_dirty -    internal function
983  *      @inode: inode to mark
984  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
985  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
986  *      mark_inode_dirty_sync.
987  *
988  * Put the inode on the super block's dirty list.
989  *
990  * CAREFUL! We mark it dirty unconditionally, but move it onto the
991  * dirty list only if it is hashed or if it refers to a blockdev.
992  * If it was not hashed, it will never be added to the dirty list
993  * even if it is later hashed, as it will have been marked dirty already.
994  *
995  * In short, make sure you hash any inodes _before_ you start marking
996  * them dirty.
997  *
998  * This function *must* be atomic for the I_DIRTY_PAGES case -
999  * set_page_dirty() is called under spinlock in several places.
1000  *
1001  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1002  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1003  * the kernel-internal blockdev inode represents the dirtying time of the
1004  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1005  * page->mapping->host, so the page-dirtying time is recorded in the internal
1006  * blockdev inode.
1007  */
1008 void __mark_inode_dirty(struct inode *inode, int flags)
1009 {
1010         struct super_block *sb = inode->i_sb;
1011         struct backing_dev_info *bdi = NULL;
1012         bool wakeup_bdi = false;
1013
1014         /*
1015          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1016          * dirty the inode itself
1017          */
1018         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1019                 if (sb->s_op->dirty_inode)
1020                         sb->s_op->dirty_inode(inode);
1021         }
1022
1023         /*
1024          * make sure that changes are seen by all cpus before we test i_state
1025          * -- mikulas
1026          */
1027         smp_mb();
1028
1029         /* avoid the locking if we can */
1030         if ((inode->i_state & flags) == flags)
1031                 return;
1032
1033         if (unlikely(block_dump))
1034                 block_dump___mark_inode_dirty(inode);
1035
1036         spin_lock(&inode_lock);
1037         spin_lock(&inode->i_lock);
1038         if ((inode->i_state & flags) != flags) {
1039                 const int was_dirty = inode->i_state & I_DIRTY;
1040
1041                 inode->i_state |= flags;
1042
1043                 /*
1044                  * If the inode is being synced, just update its dirty state.
1045                  * The unlocker will place the inode on the appropriate
1046                  * superblock list, based upon its state.
1047                  */
1048                 if (inode->i_state & I_SYNC)
1049                         goto out_unlock_inode;
1050
1051                 /*
1052                  * Only add valid (hashed) inodes to the superblock's
1053                  * dirty list.  Add blockdev inodes as well.
1054                  */
1055                 if (!S_ISBLK(inode->i_mode)) {
1056                         if (inode_unhashed(inode))
1057                                 goto out_unlock_inode;
1058                 }
1059                 if (inode->i_state & I_FREEING)
1060                         goto out_unlock_inode;
1061
1062                 spin_unlock(&inode->i_lock);
1063                 /*
1064                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1065                  * reposition it (that would break b_dirty time-ordering).
1066                  */
1067                 if (!was_dirty) {
1068                         bdi = inode_to_bdi(inode);
1069
1070                         if (bdi_cap_writeback_dirty(bdi)) {
1071                                 WARN(!test_bit(BDI_registered, &bdi->state),
1072                                      "bdi-%s not registered\n", bdi->name);
1073
1074                                 /*
1075                                  * If this is the first dirty inode for this
1076                                  * bdi, we have to wake-up the corresponding
1077                                  * bdi thread to make sure background
1078                                  * write-back happens later.
1079                                  */
1080                                 if (!wb_has_dirty_io(&bdi->wb))
1081                                         wakeup_bdi = true;
1082                         }
1083
1084                         inode->dirtied_when = jiffies;
1085                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1086                 }
1087                 goto out;
1088         }
1089 out_unlock_inode:
1090         spin_unlock(&inode->i_lock);
1091 out:
1092         spin_unlock(&inode_lock);
1093
1094         if (wakeup_bdi)
1095                 bdi_wakeup_thread_delayed(bdi);
1096 }
1097 EXPORT_SYMBOL(__mark_inode_dirty);
1098
1099 /*
1100  * Write out a superblock's list of dirty inodes.  A wait will be performed
1101  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1102  *
1103  * If older_than_this is non-NULL, then only write out inodes which
1104  * had their first dirtying at a time earlier than *older_than_this.
1105  *
1106  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1107  * This function assumes that the blockdev superblock's inodes are backed by
1108  * a variety of queues, so all inodes are searched.  For other superblocks,
1109  * assume that all inodes are backed by the same queue.
1110  *
1111  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1112  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1113  * on the writer throttling path, and we get decent balancing between many
1114  * throttled threads: we don't want them all piling up on inode_sync_wait.
1115  */
1116 static void wait_sb_inodes(struct super_block *sb)
1117 {
1118         struct inode *inode, *old_inode = NULL;
1119
1120         /*
1121          * We need to be protected against the filesystem going from
1122          * r/o to r/w or vice versa.
1123          */
1124         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1125
1126         spin_lock(&inode_lock);
1127
1128         /*
1129          * Data integrity sync. Must wait for all pages under writeback,
1130          * because there may have been pages dirtied before our sync
1131          * call, but which had writeout started before we write it out.
1132          * In which case, the inode may not be on the dirty list, but
1133          * we still have to wait for that writeout.
1134          */
1135         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1136                 struct address_space *mapping = inode->i_mapping;
1137
1138                 spin_lock(&inode->i_lock);
1139                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1140                     (mapping->nrpages == 0)) {
1141                         spin_unlock(&inode->i_lock);
1142                         continue;
1143                 }
1144                 __iget(inode);
1145                 spin_unlock(&inode->i_lock);
1146                 spin_unlock(&inode_lock);
1147                 /*
1148                  * We hold a reference to 'inode' so it couldn't have
1149                  * been removed from s_inodes list while we dropped the
1150                  * inode_lock.  We cannot iput the inode now as we can
1151                  * be holding the last reference and we cannot iput it
1152                  * under inode_lock. So we keep the reference and iput
1153                  * it later.
1154                  */
1155                 iput(old_inode);
1156                 old_inode = inode;
1157
1158                 filemap_fdatawait(mapping);
1159
1160                 cond_resched();
1161
1162                 spin_lock(&inode_lock);
1163         }
1164         spin_unlock(&inode_lock);
1165         iput(old_inode);
1166 }
1167
1168 /**
1169  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1170  * @sb: the superblock
1171  * @nr: the number of pages to write
1172  *
1173  * Start writeback on some inodes on this super_block. No guarantees are made
1174  * on how many (if any) will be written, and this function does not wait
1175  * for IO completion of submitted IO.
1176  */
1177 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1178 {
1179         DECLARE_COMPLETION_ONSTACK(done);
1180         struct wb_writeback_work work = {
1181                 .sb             = sb,
1182                 .sync_mode      = WB_SYNC_NONE,
1183                 .done           = &done,
1184                 .nr_pages       = nr,
1185         };
1186
1187         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1188         bdi_queue_work(sb->s_bdi, &work);
1189         wait_for_completion(&done);
1190 }
1191 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1192
1193 /**
1194  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1195  * @sb: the superblock
1196  *
1197  * Start writeback on some inodes on this super_block. No guarantees are made
1198  * on how many (if any) will be written, and this function does not wait
1199  * for IO completion of submitted IO.
1200  */
1201 void writeback_inodes_sb(struct super_block *sb)
1202 {
1203         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1204 }
1205 EXPORT_SYMBOL(writeback_inodes_sb);
1206
1207 /**
1208  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1209  * @sb: the superblock
1210  *
1211  * Invoke writeback_inodes_sb if no writeback is currently underway.
1212  * Returns 1 if writeback was started, 0 if not.
1213  */
1214 int writeback_inodes_sb_if_idle(struct super_block *sb)
1215 {
1216         if (!writeback_in_progress(sb->s_bdi)) {
1217                 down_read(&sb->s_umount);
1218                 writeback_inodes_sb(sb);
1219                 up_read(&sb->s_umount);
1220                 return 1;
1221         } else
1222                 return 0;
1223 }
1224 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1225
1226 /**
1227  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1228  * @sb: the superblock
1229  * @nr: the number of pages to write
1230  *
1231  * Invoke writeback_inodes_sb if no writeback is currently underway.
1232  * Returns 1 if writeback was started, 0 if not.
1233  */
1234 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1235                                    unsigned long nr)
1236 {
1237         if (!writeback_in_progress(sb->s_bdi)) {
1238                 down_read(&sb->s_umount);
1239                 writeback_inodes_sb_nr(sb, nr);
1240                 up_read(&sb->s_umount);
1241                 return 1;
1242         } else
1243                 return 0;
1244 }
1245 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1246
1247 /**
1248  * sync_inodes_sb       -       sync sb inode pages
1249  * @sb: the superblock
1250  *
1251  * This function writes and waits on any dirty inode belonging to this
1252  * super_block.
1253  */
1254 void sync_inodes_sb(struct super_block *sb)
1255 {
1256         DECLARE_COMPLETION_ONSTACK(done);
1257         struct wb_writeback_work work = {
1258                 .sb             = sb,
1259                 .sync_mode      = WB_SYNC_ALL,
1260                 .nr_pages       = LONG_MAX,
1261                 .range_cyclic   = 0,
1262                 .done           = &done,
1263         };
1264
1265         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1266
1267         bdi_queue_work(sb->s_bdi, &work);
1268         wait_for_completion(&done);
1269
1270         wait_sb_inodes(sb);
1271 }
1272 EXPORT_SYMBOL(sync_inodes_sb);
1273
1274 /**
1275  * write_inode_now      -       write an inode to disk
1276  * @inode: inode to write to disk
1277  * @sync: whether the write should be synchronous or not
1278  *
1279  * This function commits an inode to disk immediately if it is dirty. This is
1280  * primarily needed by knfsd.
1281  *
1282  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1283  */
1284 int write_inode_now(struct inode *inode, int sync)
1285 {
1286         int ret;
1287         struct writeback_control wbc = {
1288                 .nr_to_write = LONG_MAX,
1289                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1290                 .range_start = 0,
1291                 .range_end = LLONG_MAX,
1292         };
1293
1294         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1295                 wbc.nr_to_write = 0;
1296
1297         might_sleep();
1298         spin_lock(&inode_lock);
1299         ret = writeback_single_inode(inode, &wbc);
1300         spin_unlock(&inode_lock);
1301         if (sync)
1302                 inode_sync_wait(inode);
1303         return ret;
1304 }
1305 EXPORT_SYMBOL(write_inode_now);
1306
1307 /**
1308  * sync_inode - write an inode and its pages to disk.
1309  * @inode: the inode to sync
1310  * @wbc: controls the writeback mode
1311  *
1312  * sync_inode() will write an inode and its pages to disk.  It will also
1313  * correctly update the inode on its superblock's dirty inode lists and will
1314  * update inode->i_state.
1315  *
1316  * The caller must have a ref on the inode.
1317  */
1318 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1319 {
1320         int ret;
1321
1322         spin_lock(&inode_lock);
1323         ret = writeback_single_inode(inode, wbc);
1324         spin_unlock(&inode_lock);
1325         return ret;
1326 }
1327 EXPORT_SYMBOL(sync_inode);
1328
1329 /**
1330  * sync_inode_metadata - write an inode to disk
1331  * @inode: the inode to sync
1332  * @wait: wait for I/O to complete.
1333  *
1334  * Write an inode to disk and adjust its dirty state after completion.
1335  *
1336  * Note: only writes the actual inode, no associated data or other metadata.
1337  */
1338 int sync_inode_metadata(struct inode *inode, int wait)
1339 {
1340         struct writeback_control wbc = {
1341                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1342                 .nr_to_write = 0, /* metadata-only */
1343         };
1344
1345         return sync_inode(inode, &wbc);
1346 }
1347 EXPORT_SYMBOL(sync_inode_metadata);