writeback: Move clearing of I_SYNC into inode_sync_complete()
[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/export.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/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * 4MB minimal write chunk size
34  */
35 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
36
37 /*
38  * Passed into wb_writeback(), essentially a subset of writeback_control
39  */
40 struct wb_writeback_work {
41         long nr_pages;
42         struct super_block *sb;
43         unsigned long *older_than_this;
44         enum writeback_sync_modes sync_mode;
45         unsigned int tagged_writepages:1;
46         unsigned int for_kupdate:1;
47         unsigned int range_cyclic:1;
48         unsigned int for_background:1;
49         enum wb_reason reason;          /* why was writeback initiated? */
50
51         struct list_head list;          /* pending work list */
52         struct completion *done;        /* set if the caller waits */
53 };
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 /*
88  * Include the creation of the trace points after defining the
89  * wb_writeback_work structure and inline functions so that the definition
90  * remains local to this file.
91  */
92 #define CREATE_TRACE_POINTS
93 #include <trace/events/writeback.h>
94
95 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
96 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
97 {
98         if (bdi->wb.task) {
99                 wake_up_process(bdi->wb.task);
100         } else {
101                 /*
102                  * The bdi thread isn't there, wake up the forker thread which
103                  * will create and run it.
104                  */
105                 wake_up_process(default_backing_dev_info.wb.task);
106         }
107 }
108
109 static void bdi_queue_work(struct backing_dev_info *bdi,
110                            struct wb_writeback_work *work)
111 {
112         trace_writeback_queue(bdi, work);
113
114         spin_lock_bh(&bdi->wb_lock);
115         list_add_tail(&work->list, &bdi->work_list);
116         if (!bdi->wb.task)
117                 trace_writeback_nothread(bdi, work);
118         bdi_wakeup_flusher(bdi);
119         spin_unlock_bh(&bdi->wb_lock);
120 }
121
122 static void
123 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
124                       bool range_cyclic, enum wb_reason reason)
125 {
126         struct wb_writeback_work *work;
127
128         /*
129          * This is WB_SYNC_NONE writeback, so if allocation fails just
130          * wakeup the thread for old dirty data writeback
131          */
132         work = kzalloc(sizeof(*work), GFP_ATOMIC);
133         if (!work) {
134                 if (bdi->wb.task) {
135                         trace_writeback_nowork(bdi);
136                         wake_up_process(bdi->wb.task);
137                 }
138                 return;
139         }
140
141         work->sync_mode = WB_SYNC_NONE;
142         work->nr_pages  = nr_pages;
143         work->range_cyclic = range_cyclic;
144         work->reason    = reason;
145
146         bdi_queue_work(bdi, work);
147 }
148
149 /**
150  * bdi_start_writeback - start writeback
151  * @bdi: the backing device to write from
152  * @nr_pages: the number of pages to write
153  * @reason: reason why some writeback work was initiated
154  *
155  * Description:
156  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
157  *   started when this function returns, we make no guarantees on
158  *   completion. Caller need not hold sb s_umount semaphore.
159  *
160  */
161 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
162                         enum wb_reason reason)
163 {
164         __bdi_start_writeback(bdi, nr_pages, true, reason);
165 }
166
167 /**
168  * bdi_start_background_writeback - start background writeback
169  * @bdi: the backing device to write from
170  *
171  * Description:
172  *   This makes sure WB_SYNC_NONE background writeback happens. When
173  *   this function returns, it is only guaranteed that for given BDI
174  *   some IO is happening if we are over background dirty threshold.
175  *   Caller need not hold sb s_umount semaphore.
176  */
177 void bdi_start_background_writeback(struct backing_dev_info *bdi)
178 {
179         /*
180          * We just wake up the flusher thread. It will perform background
181          * writeback as soon as there is no other work to do.
182          */
183         trace_writeback_wake_background(bdi);
184         spin_lock_bh(&bdi->wb_lock);
185         bdi_wakeup_flusher(bdi);
186         spin_unlock_bh(&bdi->wb_lock);
187 }
188
189 /*
190  * Remove the inode from the writeback list it is on.
191  */
192 void inode_wb_list_del(struct inode *inode)
193 {
194         struct backing_dev_info *bdi = inode_to_bdi(inode);
195
196         spin_lock(&bdi->wb.list_lock);
197         list_del_init(&inode->i_wb_list);
198         spin_unlock(&bdi->wb.list_lock);
199 }
200
201 /*
202  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
203  * furthest end of its superblock's dirty-inode list.
204  *
205  * Before stamping the inode's ->dirtied_when, we check to see whether it is
206  * already the most-recently-dirtied inode on the b_dirty list.  If that is
207  * the case then the inode must have been redirtied while it was being written
208  * out and we don't reset its dirtied_when.
209  */
210 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
211 {
212         assert_spin_locked(&wb->list_lock);
213         if (!list_empty(&wb->b_dirty)) {
214                 struct inode *tail;
215
216                 tail = wb_inode(wb->b_dirty.next);
217                 if (time_before(inode->dirtied_when, tail->dirtied_when))
218                         inode->dirtied_when = jiffies;
219         }
220         list_move(&inode->i_wb_list, &wb->b_dirty);
221 }
222
223 /*
224  * requeue inode for re-scanning after bdi->b_io list is exhausted.
225  */
226 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
227 {
228         assert_spin_locked(&wb->list_lock);
229         list_move(&inode->i_wb_list, &wb->b_more_io);
230 }
231
232 static void inode_sync_complete(struct inode *inode)
233 {
234         inode->i_state &= ~I_SYNC;
235         /* Waiters must see I_SYNC cleared before being woken up */
236         smp_mb();
237         wake_up_bit(&inode->i_state, __I_SYNC);
238 }
239
240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
241 {
242         bool ret = time_after(inode->dirtied_when, t);
243 #ifndef CONFIG_64BIT
244         /*
245          * For inodes being constantly redirtied, dirtied_when can get stuck.
246          * It _appears_ to be in the future, but is actually in distant past.
247          * This test is necessary to prevent such wrapped-around relative times
248          * from permanently stopping the whole bdi writeback.
249          */
250         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
251 #endif
252         return ret;
253 }
254
255 /*
256  * Move expired (dirtied after work->older_than_this) dirty inodes from
257  * @delaying_queue to @dispatch_queue.
258  */
259 static int move_expired_inodes(struct list_head *delaying_queue,
260                                struct list_head *dispatch_queue,
261                                struct wb_writeback_work *work)
262 {
263         LIST_HEAD(tmp);
264         struct list_head *pos, *node;
265         struct super_block *sb = NULL;
266         struct inode *inode;
267         int do_sb_sort = 0;
268         int moved = 0;
269
270         while (!list_empty(delaying_queue)) {
271                 inode = wb_inode(delaying_queue->prev);
272                 if (work->older_than_this &&
273                     inode_dirtied_after(inode, *work->older_than_this))
274                         break;
275                 if (sb && sb != inode->i_sb)
276                         do_sb_sort = 1;
277                 sb = inode->i_sb;
278                 list_move(&inode->i_wb_list, &tmp);
279                 moved++;
280         }
281
282         /* just one sb in list, splice to dispatch_queue and we're done */
283         if (!do_sb_sort) {
284                 list_splice(&tmp, dispatch_queue);
285                 goto out;
286         }
287
288         /* Move inodes from one superblock together */
289         while (!list_empty(&tmp)) {
290                 sb = wb_inode(tmp.prev)->i_sb;
291                 list_for_each_prev_safe(pos, node, &tmp) {
292                         inode = wb_inode(pos);
293                         if (inode->i_sb == sb)
294                                 list_move(&inode->i_wb_list, dispatch_queue);
295                 }
296         }
297 out:
298         return moved;
299 }
300
301 /*
302  * Queue all expired dirty inodes for io, eldest first.
303  * Before
304  *         newly dirtied     b_dirty    b_io    b_more_io
305  *         =============>    gf         edc     BA
306  * After
307  *         newly dirtied     b_dirty    b_io    b_more_io
308  *         =============>    g          fBAedc
309  *                                           |
310  *                                           +--> dequeue for IO
311  */
312 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
313 {
314         int moved;
315         assert_spin_locked(&wb->list_lock);
316         list_splice_init(&wb->b_more_io, &wb->b_io);
317         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
318         trace_writeback_queue_io(wb, work, moved);
319 }
320
321 static int write_inode(struct inode *inode, struct writeback_control *wbc)
322 {
323         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
324                 return inode->i_sb->s_op->write_inode(inode, wbc);
325         return 0;
326 }
327
328 /*
329  * Wait for writeback on an inode to complete.
330  */
331 static void inode_wait_for_writeback(struct inode *inode,
332                                      struct bdi_writeback *wb)
333 {
334         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
335         wait_queue_head_t *wqh;
336
337         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
338         while (inode->i_state & I_SYNC) {
339                 spin_unlock(&inode->i_lock);
340                 spin_unlock(&wb->list_lock);
341                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
342                 spin_lock(&wb->list_lock);
343                 spin_lock(&inode->i_lock);
344         }
345 }
346
347 /*
348  * Write out an inode's dirty pages.  Called under wb->list_lock and
349  * inode->i_lock.  Either the caller has an active reference on the inode or
350  * the inode has I_WILL_FREE set.
351  *
352  * If `wait' is set, wait on the writeout.
353  *
354  * The whole writeout design is quite complex and fragile.  We want to avoid
355  * starvation of particular inodes when others are being redirtied, prevent
356  * livelocks, etc.
357  */
358 static int
359 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
360                        struct writeback_control *wbc)
361 {
362         struct address_space *mapping = inode->i_mapping;
363         long nr_to_write = wbc->nr_to_write;
364         unsigned dirty;
365         int ret;
366
367         assert_spin_locked(&wb->list_lock);
368         assert_spin_locked(&inode->i_lock);
369
370         if (!atomic_read(&inode->i_count))
371                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
372         else
373                 WARN_ON(inode->i_state & I_WILL_FREE);
374
375         if (inode->i_state & I_SYNC) {
376                 /*
377                  * If this inode is locked for writeback and we are not doing
378                  * writeback-for-data-integrity, move it to b_more_io so that
379                  * writeback can proceed with the other inodes on s_io.
380                  *
381                  * We'll have another go at writing back this inode when we
382                  * completed a full scan of b_io.
383                  */
384                 if (wbc->sync_mode != WB_SYNC_ALL) {
385                         requeue_io(inode, wb);
386                         trace_writeback_single_inode_requeue(inode, wbc,
387                                                              nr_to_write);
388                         return 0;
389                 }
390
391                 /*
392                  * It's a data-integrity sync.  We must wait.
393                  */
394                 inode_wait_for_writeback(inode, wb);
395         }
396
397         BUG_ON(inode->i_state & I_SYNC);
398
399         /* Set I_SYNC, reset I_DIRTY_PAGES */
400         inode->i_state |= I_SYNC;
401         inode->i_state &= ~I_DIRTY_PAGES;
402         spin_unlock(&inode->i_lock);
403         spin_unlock(&wb->list_lock);
404
405         ret = do_writepages(mapping, wbc);
406
407         /*
408          * Make sure to wait on the data before writing out the metadata.
409          * This is important for filesystems that modify metadata on data
410          * I/O completion.
411          */
412         if (wbc->sync_mode == WB_SYNC_ALL) {
413                 int err = filemap_fdatawait(mapping);
414                 if (ret == 0)
415                         ret = err;
416         }
417
418         /*
419          * Some filesystems may redirty the inode during the writeback
420          * due to delalloc, clear dirty metadata flags right before
421          * write_inode()
422          */
423         spin_lock(&inode->i_lock);
424         dirty = inode->i_state & I_DIRTY;
425         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
426         spin_unlock(&inode->i_lock);
427         /* Don't write the inode if only I_DIRTY_PAGES was set */
428         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
429                 int err = write_inode(inode, wbc);
430                 if (ret == 0)
431                         ret = err;
432         }
433
434         spin_lock(&wb->list_lock);
435         spin_lock(&inode->i_lock);
436         if (!(inode->i_state & I_FREEING)) {
437                 /*
438                  * Sync livelock prevention. Each inode is tagged and synced in
439                  * one shot. If still dirty, it will be redirty_tail()'ed below.
440                  * Update the dirty time to prevent enqueue and sync it again.
441                  */
442                 if ((inode->i_state & I_DIRTY) &&
443                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
444                         inode->dirtied_when = jiffies;
445
446                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
447                         /*
448                          * We didn't write back all the pages.  nfs_writepages()
449                          * sometimes bales out without doing anything.
450                          */
451                         inode->i_state |= I_DIRTY_PAGES;
452                         if (wbc->nr_to_write <= 0) {
453                                 /*
454                                  * slice used up: queue for next turn
455                                  */
456                                 requeue_io(inode, wb);
457                         } else {
458                                 /*
459                                  * Writeback blocked by something other than
460                                  * congestion. Delay the inode for some time to
461                                  * avoid spinning on the CPU (100% iowait)
462                                  * retrying writeback of the dirty page/inode
463                                  * that cannot be performed immediately.
464                                  */
465                                 redirty_tail(inode, wb);
466                         }
467                 } else if (inode->i_state & I_DIRTY) {
468                         /*
469                          * Filesystems can dirty the inode during writeback
470                          * operations, such as delayed allocation during
471                          * submission or metadata updates after data IO
472                          * completion.
473                          */
474                         redirty_tail(inode, wb);
475                 } else {
476                         /*
477                          * The inode is clean.  At this point we either have
478                          * a reference to the inode or it's on it's way out.
479                          * No need to add it back to the LRU.
480                          */
481                         list_del_init(&inode->i_wb_list);
482                 }
483         }
484         inode_sync_complete(inode);
485         trace_writeback_single_inode(inode, wbc, nr_to_write);
486         return ret;
487 }
488
489 static long writeback_chunk_size(struct backing_dev_info *bdi,
490                                  struct wb_writeback_work *work)
491 {
492         long pages;
493
494         /*
495          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
496          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
497          * here avoids calling into writeback_inodes_wb() more than once.
498          *
499          * The intended call sequence for WB_SYNC_ALL writeback is:
500          *
501          *      wb_writeback()
502          *          writeback_sb_inodes()       <== called only once
503          *              write_cache_pages()     <== called once for each inode
504          *                   (quickly) tag currently dirty pages
505          *                   (maybe slowly) sync all tagged pages
506          */
507         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
508                 pages = LONG_MAX;
509         else {
510                 pages = min(bdi->avg_write_bandwidth / 2,
511                             global_dirty_limit / DIRTY_SCOPE);
512                 pages = min(pages, work->nr_pages);
513                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
514                                    MIN_WRITEBACK_PAGES);
515         }
516
517         return pages;
518 }
519
520 /*
521  * Write a portion of b_io inodes which belong to @sb.
522  *
523  * If @only_this_sb is true, then find and write all such
524  * inodes. Otherwise write only ones which go sequentially
525  * in reverse order.
526  *
527  * Return the number of pages and/or inodes written.
528  */
529 static long writeback_sb_inodes(struct super_block *sb,
530                                 struct bdi_writeback *wb,
531                                 struct wb_writeback_work *work)
532 {
533         struct writeback_control wbc = {
534                 .sync_mode              = work->sync_mode,
535                 .tagged_writepages      = work->tagged_writepages,
536                 .for_kupdate            = work->for_kupdate,
537                 .for_background         = work->for_background,
538                 .range_cyclic           = work->range_cyclic,
539                 .range_start            = 0,
540                 .range_end              = LLONG_MAX,
541         };
542         unsigned long start_time = jiffies;
543         long write_chunk;
544         long wrote = 0;  /* count both pages and inodes */
545
546         while (!list_empty(&wb->b_io)) {
547                 struct inode *inode = wb_inode(wb->b_io.prev);
548
549                 if (inode->i_sb != sb) {
550                         if (work->sb) {
551                                 /*
552                                  * We only want to write back data for this
553                                  * superblock, move all inodes not belonging
554                                  * to it back onto the dirty list.
555                                  */
556                                 redirty_tail(inode, wb);
557                                 continue;
558                         }
559
560                         /*
561                          * The inode belongs to a different superblock.
562                          * Bounce back to the caller to unpin this and
563                          * pin the next superblock.
564                          */
565                         break;
566                 }
567
568                 /*
569                  * Don't bother with new inodes or inodes beeing freed, first
570                  * kind does not need peridic writeout yet, and for the latter
571                  * kind writeout is handled by the freer.
572                  */
573                 spin_lock(&inode->i_lock);
574                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
575                         spin_unlock(&inode->i_lock);
576                         redirty_tail(inode, wb);
577                         continue;
578                 }
579                 __iget(inode);
580                 write_chunk = writeback_chunk_size(wb->bdi, work);
581                 wbc.nr_to_write = write_chunk;
582                 wbc.pages_skipped = 0;
583
584                 writeback_single_inode(inode, wb, &wbc);
585
586                 work->nr_pages -= write_chunk - wbc.nr_to_write;
587                 wrote += write_chunk - wbc.nr_to_write;
588                 if (!(inode->i_state & I_DIRTY))
589                         wrote++;
590                 if (wbc.pages_skipped) {
591                         /*
592                          * writeback is not making progress due to locked
593                          * buffers.  Skip this inode for now.
594                          */
595                         redirty_tail(inode, wb);
596                 }
597                 spin_unlock(&inode->i_lock);
598                 spin_unlock(&wb->list_lock);
599                 iput(inode);
600                 cond_resched();
601                 spin_lock(&wb->list_lock);
602                 /*
603                  * bail out to wb_writeback() often enough to check
604                  * background threshold and other termination conditions.
605                  */
606                 if (wrote) {
607                         if (time_is_before_jiffies(start_time + HZ / 10UL))
608                                 break;
609                         if (work->nr_pages <= 0)
610                                 break;
611                 }
612         }
613         return wrote;
614 }
615
616 static long __writeback_inodes_wb(struct bdi_writeback *wb,
617                                   struct wb_writeback_work *work)
618 {
619         unsigned long start_time = jiffies;
620         long wrote = 0;
621
622         while (!list_empty(&wb->b_io)) {
623                 struct inode *inode = wb_inode(wb->b_io.prev);
624                 struct super_block *sb = inode->i_sb;
625
626                 if (!grab_super_passive(sb)) {
627                         /*
628                          * grab_super_passive() may fail consistently due to
629                          * s_umount being grabbed by someone else. Don't use
630                          * requeue_io() to avoid busy retrying the inode/sb.
631                          */
632                         redirty_tail(inode, wb);
633                         continue;
634                 }
635                 wrote += writeback_sb_inodes(sb, wb, work);
636                 drop_super(sb);
637
638                 /* refer to the same tests at the end of writeback_sb_inodes */
639                 if (wrote) {
640                         if (time_is_before_jiffies(start_time + HZ / 10UL))
641                                 break;
642                         if (work->nr_pages <= 0)
643                                 break;
644                 }
645         }
646         /* Leave any unwritten inodes on b_io */
647         return wrote;
648 }
649
650 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
651                                 enum wb_reason reason)
652 {
653         struct wb_writeback_work work = {
654                 .nr_pages       = nr_pages,
655                 .sync_mode      = WB_SYNC_NONE,
656                 .range_cyclic   = 1,
657                 .reason         = reason,
658         };
659
660         spin_lock(&wb->list_lock);
661         if (list_empty(&wb->b_io))
662                 queue_io(wb, &work);
663         __writeback_inodes_wb(wb, &work);
664         spin_unlock(&wb->list_lock);
665
666         return nr_pages - work.nr_pages;
667 }
668
669 static bool over_bground_thresh(struct backing_dev_info *bdi)
670 {
671         unsigned long background_thresh, dirty_thresh;
672
673         global_dirty_limits(&background_thresh, &dirty_thresh);
674
675         if (global_page_state(NR_FILE_DIRTY) +
676             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
677                 return true;
678
679         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
680                                 bdi_dirty_limit(bdi, background_thresh))
681                 return true;
682
683         return false;
684 }
685
686 /*
687  * Called under wb->list_lock. If there are multiple wb per bdi,
688  * only the flusher working on the first wb should do it.
689  */
690 static void wb_update_bandwidth(struct bdi_writeback *wb,
691                                 unsigned long start_time)
692 {
693         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
694 }
695
696 /*
697  * Explicit flushing or periodic writeback of "old" data.
698  *
699  * Define "old": the first time one of an inode's pages is dirtied, we mark the
700  * dirtying-time in the inode's address_space.  So this periodic writeback code
701  * just walks the superblock inode list, writing back any inodes which are
702  * older than a specific point in time.
703  *
704  * Try to run once per dirty_writeback_interval.  But if a writeback event
705  * takes longer than a dirty_writeback_interval interval, then leave a
706  * one-second gap.
707  *
708  * older_than_this takes precedence over nr_to_write.  So we'll only write back
709  * all dirty pages if they are all attached to "old" mappings.
710  */
711 static long wb_writeback(struct bdi_writeback *wb,
712                          struct wb_writeback_work *work)
713 {
714         unsigned long wb_start = jiffies;
715         long nr_pages = work->nr_pages;
716         unsigned long oldest_jif;
717         struct inode *inode;
718         long progress;
719
720         oldest_jif = jiffies;
721         work->older_than_this = &oldest_jif;
722
723         spin_lock(&wb->list_lock);
724         for (;;) {
725                 /*
726                  * Stop writeback when nr_pages has been consumed
727                  */
728                 if (work->nr_pages <= 0)
729                         break;
730
731                 /*
732                  * Background writeout and kupdate-style writeback may
733                  * run forever. Stop them if there is other work to do
734                  * so that e.g. sync can proceed. They'll be restarted
735                  * after the other works are all done.
736                  */
737                 if ((work->for_background || work->for_kupdate) &&
738                     !list_empty(&wb->bdi->work_list))
739                         break;
740
741                 /*
742                  * For background writeout, stop when we are below the
743                  * background dirty threshold
744                  */
745                 if (work->for_background && !over_bground_thresh(wb->bdi))
746                         break;
747
748                 /*
749                  * Kupdate and background works are special and we want to
750                  * include all inodes that need writing. Livelock avoidance is
751                  * handled by these works yielding to any other work so we are
752                  * safe.
753                  */
754                 if (work->for_kupdate) {
755                         oldest_jif = jiffies -
756                                 msecs_to_jiffies(dirty_expire_interval * 10);
757                 } else if (work->for_background)
758                         oldest_jif = jiffies;
759
760                 trace_writeback_start(wb->bdi, work);
761                 if (list_empty(&wb->b_io))
762                         queue_io(wb, work);
763                 if (work->sb)
764                         progress = writeback_sb_inodes(work->sb, wb, work);
765                 else
766                         progress = __writeback_inodes_wb(wb, work);
767                 trace_writeback_written(wb->bdi, work);
768
769                 wb_update_bandwidth(wb, wb_start);
770
771                 /*
772                  * Did we write something? Try for more
773                  *
774                  * Dirty inodes are moved to b_io for writeback in batches.
775                  * The completion of the current batch does not necessarily
776                  * mean the overall work is done. So we keep looping as long
777                  * as made some progress on cleaning pages or inodes.
778                  */
779                 if (progress)
780                         continue;
781                 /*
782                  * No more inodes for IO, bail
783                  */
784                 if (list_empty(&wb->b_more_io))
785                         break;
786                 /*
787                  * Nothing written. Wait for some inode to
788                  * become available for writeback. Otherwise
789                  * we'll just busyloop.
790                  */
791                 if (!list_empty(&wb->b_more_io))  {
792                         trace_writeback_wait(wb->bdi, work);
793                         inode = wb_inode(wb->b_more_io.prev);
794                         spin_lock(&inode->i_lock);
795                         inode_wait_for_writeback(inode, wb);
796                         spin_unlock(&inode->i_lock);
797                 }
798         }
799         spin_unlock(&wb->list_lock);
800
801         return nr_pages - work->nr_pages;
802 }
803
804 /*
805  * Return the next wb_writeback_work struct that hasn't been processed yet.
806  */
807 static struct wb_writeback_work *
808 get_next_work_item(struct backing_dev_info *bdi)
809 {
810         struct wb_writeback_work *work = NULL;
811
812         spin_lock_bh(&bdi->wb_lock);
813         if (!list_empty(&bdi->work_list)) {
814                 work = list_entry(bdi->work_list.next,
815                                   struct wb_writeback_work, list);
816                 list_del_init(&work->list);
817         }
818         spin_unlock_bh(&bdi->wb_lock);
819         return work;
820 }
821
822 /*
823  * Add in the number of potentially dirty inodes, because each inode
824  * write can dirty pagecache in the underlying blockdev.
825  */
826 static unsigned long get_nr_dirty_pages(void)
827 {
828         return global_page_state(NR_FILE_DIRTY) +
829                 global_page_state(NR_UNSTABLE_NFS) +
830                 get_nr_dirty_inodes();
831 }
832
833 static long wb_check_background_flush(struct bdi_writeback *wb)
834 {
835         if (over_bground_thresh(wb->bdi)) {
836
837                 struct wb_writeback_work work = {
838                         .nr_pages       = LONG_MAX,
839                         .sync_mode      = WB_SYNC_NONE,
840                         .for_background = 1,
841                         .range_cyclic   = 1,
842                         .reason         = WB_REASON_BACKGROUND,
843                 };
844
845                 return wb_writeback(wb, &work);
846         }
847
848         return 0;
849 }
850
851 static long wb_check_old_data_flush(struct bdi_writeback *wb)
852 {
853         unsigned long expired;
854         long nr_pages;
855
856         /*
857          * When set to zero, disable periodic writeback
858          */
859         if (!dirty_writeback_interval)
860                 return 0;
861
862         expired = wb->last_old_flush +
863                         msecs_to_jiffies(dirty_writeback_interval * 10);
864         if (time_before(jiffies, expired))
865                 return 0;
866
867         wb->last_old_flush = jiffies;
868         nr_pages = get_nr_dirty_pages();
869
870         if (nr_pages) {
871                 struct wb_writeback_work work = {
872                         .nr_pages       = nr_pages,
873                         .sync_mode      = WB_SYNC_NONE,
874                         .for_kupdate    = 1,
875                         .range_cyclic   = 1,
876                         .reason         = WB_REASON_PERIODIC,
877                 };
878
879                 return wb_writeback(wb, &work);
880         }
881
882         return 0;
883 }
884
885 /*
886  * Retrieve work items and do the writeback they describe
887  */
888 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
889 {
890         struct backing_dev_info *bdi = wb->bdi;
891         struct wb_writeback_work *work;
892         long wrote = 0;
893
894         set_bit(BDI_writeback_running, &wb->bdi->state);
895         while ((work = get_next_work_item(bdi)) != NULL) {
896                 /*
897                  * Override sync mode, in case we must wait for completion
898                  * because this thread is exiting now.
899                  */
900                 if (force_wait)
901                         work->sync_mode = WB_SYNC_ALL;
902
903                 trace_writeback_exec(bdi, work);
904
905                 wrote += wb_writeback(wb, work);
906
907                 /*
908                  * Notify the caller of completion if this is a synchronous
909                  * work item, otherwise just free it.
910                  */
911                 if (work->done)
912                         complete(work->done);
913                 else
914                         kfree(work);
915         }
916
917         /*
918          * Check for periodic writeback, kupdated() style
919          */
920         wrote += wb_check_old_data_flush(wb);
921         wrote += wb_check_background_flush(wb);
922         clear_bit(BDI_writeback_running, &wb->bdi->state);
923
924         return wrote;
925 }
926
927 /*
928  * Handle writeback of dirty data for the device backed by this bdi. Also
929  * wakes up periodically and does kupdated style flushing.
930  */
931 int bdi_writeback_thread(void *data)
932 {
933         struct bdi_writeback *wb = data;
934         struct backing_dev_info *bdi = wb->bdi;
935         long pages_written;
936
937         current->flags |= PF_SWAPWRITE;
938         set_freezable();
939         wb->last_active = jiffies;
940
941         /*
942          * Our parent may run at a different priority, just set us to normal
943          */
944         set_user_nice(current, 0);
945
946         trace_writeback_thread_start(bdi);
947
948         while (!kthread_freezable_should_stop(NULL)) {
949                 /*
950                  * Remove own delayed wake-up timer, since we are already awake
951                  * and we'll take care of the preriodic write-back.
952                  */
953                 del_timer(&wb->wakeup_timer);
954
955                 pages_written = wb_do_writeback(wb, 0);
956
957                 trace_writeback_pages_written(pages_written);
958
959                 if (pages_written)
960                         wb->last_active = jiffies;
961
962                 set_current_state(TASK_INTERRUPTIBLE);
963                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
964                         __set_current_state(TASK_RUNNING);
965                         continue;
966                 }
967
968                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
969                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
970                 else {
971                         /*
972                          * We have nothing to do, so can go sleep without any
973                          * timeout and save power. When a work is queued or
974                          * something is made dirty - we will be woken up.
975                          */
976                         schedule();
977                 }
978         }
979
980         /* Flush any work that raced with us exiting */
981         if (!list_empty(&bdi->work_list))
982                 wb_do_writeback(wb, 1);
983
984         trace_writeback_thread_stop(bdi);
985         return 0;
986 }
987
988
989 /*
990  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
991  * the whole world.
992  */
993 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
994 {
995         struct backing_dev_info *bdi;
996
997         if (!nr_pages) {
998                 nr_pages = global_page_state(NR_FILE_DIRTY) +
999                                 global_page_state(NR_UNSTABLE_NFS);
1000         }
1001
1002         rcu_read_lock();
1003         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1004                 if (!bdi_has_dirty_io(bdi))
1005                         continue;
1006                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1007         }
1008         rcu_read_unlock();
1009 }
1010
1011 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1012 {
1013         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1014                 struct dentry *dentry;
1015                 const char *name = "?";
1016
1017                 dentry = d_find_alias(inode);
1018                 if (dentry) {
1019                         spin_lock(&dentry->d_lock);
1020                         name = (const char *) dentry->d_name.name;
1021                 }
1022                 printk(KERN_DEBUG
1023                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1024                        current->comm, task_pid_nr(current), inode->i_ino,
1025                        name, inode->i_sb->s_id);
1026                 if (dentry) {
1027                         spin_unlock(&dentry->d_lock);
1028                         dput(dentry);
1029                 }
1030         }
1031 }
1032
1033 /**
1034  *      __mark_inode_dirty -    internal function
1035  *      @inode: inode to mark
1036  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1037  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1038  *      mark_inode_dirty_sync.
1039  *
1040  * Put the inode on the super block's dirty list.
1041  *
1042  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1043  * dirty list only if it is hashed or if it refers to a blockdev.
1044  * If it was not hashed, it will never be added to the dirty list
1045  * even if it is later hashed, as it will have been marked dirty already.
1046  *
1047  * In short, make sure you hash any inodes _before_ you start marking
1048  * them dirty.
1049  *
1050  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1051  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1052  * the kernel-internal blockdev inode represents the dirtying time of the
1053  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1054  * page->mapping->host, so the page-dirtying time is recorded in the internal
1055  * blockdev inode.
1056  */
1057 void __mark_inode_dirty(struct inode *inode, int flags)
1058 {
1059         struct super_block *sb = inode->i_sb;
1060         struct backing_dev_info *bdi = NULL;
1061
1062         /*
1063          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1064          * dirty the inode itself
1065          */
1066         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1067                 if (sb->s_op->dirty_inode)
1068                         sb->s_op->dirty_inode(inode, flags);
1069         }
1070
1071         /*
1072          * make sure that changes are seen by all cpus before we test i_state
1073          * -- mikulas
1074          */
1075         smp_mb();
1076
1077         /* avoid the locking if we can */
1078         if ((inode->i_state & flags) == flags)
1079                 return;
1080
1081         if (unlikely(block_dump))
1082                 block_dump___mark_inode_dirty(inode);
1083
1084         spin_lock(&inode->i_lock);
1085         if ((inode->i_state & flags) != flags) {
1086                 const int was_dirty = inode->i_state & I_DIRTY;
1087
1088                 inode->i_state |= flags;
1089
1090                 /*
1091                  * If the inode is being synced, just update its dirty state.
1092                  * The unlocker will place the inode on the appropriate
1093                  * superblock list, based upon its state.
1094                  */
1095                 if (inode->i_state & I_SYNC)
1096                         goto out_unlock_inode;
1097
1098                 /*
1099                  * Only add valid (hashed) inodes to the superblock's
1100                  * dirty list.  Add blockdev inodes as well.
1101                  */
1102                 if (!S_ISBLK(inode->i_mode)) {
1103                         if (inode_unhashed(inode))
1104                                 goto out_unlock_inode;
1105                 }
1106                 if (inode->i_state & I_FREEING)
1107                         goto out_unlock_inode;
1108
1109                 /*
1110                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1111                  * reposition it (that would break b_dirty time-ordering).
1112                  */
1113                 if (!was_dirty) {
1114                         bool wakeup_bdi = false;
1115                         bdi = inode_to_bdi(inode);
1116
1117                         if (bdi_cap_writeback_dirty(bdi)) {
1118                                 WARN(!test_bit(BDI_registered, &bdi->state),
1119                                      "bdi-%s not registered\n", bdi->name);
1120
1121                                 /*
1122                                  * If this is the first dirty inode for this
1123                                  * bdi, we have to wake-up the corresponding
1124                                  * bdi thread to make sure background
1125                                  * write-back happens later.
1126                                  */
1127                                 if (!wb_has_dirty_io(&bdi->wb))
1128                                         wakeup_bdi = true;
1129                         }
1130
1131                         spin_unlock(&inode->i_lock);
1132                         spin_lock(&bdi->wb.list_lock);
1133                         inode->dirtied_when = jiffies;
1134                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1135                         spin_unlock(&bdi->wb.list_lock);
1136
1137                         if (wakeup_bdi)
1138                                 bdi_wakeup_thread_delayed(bdi);
1139                         return;
1140                 }
1141         }
1142 out_unlock_inode:
1143         spin_unlock(&inode->i_lock);
1144
1145 }
1146 EXPORT_SYMBOL(__mark_inode_dirty);
1147
1148 static void wait_sb_inodes(struct super_block *sb)
1149 {
1150         struct inode *inode, *old_inode = NULL;
1151
1152         /*
1153          * We need to be protected against the filesystem going from
1154          * r/o to r/w or vice versa.
1155          */
1156         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1157
1158         spin_lock(&inode_sb_list_lock);
1159
1160         /*
1161          * Data integrity sync. Must wait for all pages under writeback,
1162          * because there may have been pages dirtied before our sync
1163          * call, but which had writeout started before we write it out.
1164          * In which case, the inode may not be on the dirty list, but
1165          * we still have to wait for that writeout.
1166          */
1167         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1168                 struct address_space *mapping = inode->i_mapping;
1169
1170                 spin_lock(&inode->i_lock);
1171                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1172                     (mapping->nrpages == 0)) {
1173                         spin_unlock(&inode->i_lock);
1174                         continue;
1175                 }
1176                 __iget(inode);
1177                 spin_unlock(&inode->i_lock);
1178                 spin_unlock(&inode_sb_list_lock);
1179
1180                 /*
1181                  * We hold a reference to 'inode' so it couldn't have been
1182                  * removed from s_inodes list while we dropped the
1183                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1184                  * be holding the last reference and we cannot iput it under
1185                  * inode_sb_list_lock. So we keep the reference and iput it
1186                  * later.
1187                  */
1188                 iput(old_inode);
1189                 old_inode = inode;
1190
1191                 filemap_fdatawait(mapping);
1192
1193                 cond_resched();
1194
1195                 spin_lock(&inode_sb_list_lock);
1196         }
1197         spin_unlock(&inode_sb_list_lock);
1198         iput(old_inode);
1199 }
1200
1201 /**
1202  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1203  * @sb: the superblock
1204  * @nr: the number of pages to write
1205  * @reason: reason why some writeback work initiated
1206  *
1207  * Start writeback on some inodes on this super_block. No guarantees are made
1208  * on how many (if any) will be written, and this function does not wait
1209  * for IO completion of submitted IO.
1210  */
1211 void writeback_inodes_sb_nr(struct super_block *sb,
1212                             unsigned long nr,
1213                             enum wb_reason reason)
1214 {
1215         DECLARE_COMPLETION_ONSTACK(done);
1216         struct wb_writeback_work work = {
1217                 .sb                     = sb,
1218                 .sync_mode              = WB_SYNC_NONE,
1219                 .tagged_writepages      = 1,
1220                 .done                   = &done,
1221                 .nr_pages               = nr,
1222                 .reason                 = reason,
1223         };
1224
1225         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1226         bdi_queue_work(sb->s_bdi, &work);
1227         wait_for_completion(&done);
1228 }
1229 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1230
1231 /**
1232  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1233  * @sb: the superblock
1234  * @reason: reason why some writeback work was initiated
1235  *
1236  * Start writeback on some inodes on this super_block. No guarantees are made
1237  * on how many (if any) will be written, and this function does not wait
1238  * for IO completion of submitted IO.
1239  */
1240 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1241 {
1242         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1243 }
1244 EXPORT_SYMBOL(writeback_inodes_sb);
1245
1246 /**
1247  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1248  * @sb: the superblock
1249  * @reason: reason why some writeback work was initiated
1250  *
1251  * Invoke writeback_inodes_sb if no writeback is currently underway.
1252  * Returns 1 if writeback was started, 0 if not.
1253  */
1254 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1255 {
1256         if (!writeback_in_progress(sb->s_bdi)) {
1257                 down_read(&sb->s_umount);
1258                 writeback_inodes_sb(sb, reason);
1259                 up_read(&sb->s_umount);
1260                 return 1;
1261         } else
1262                 return 0;
1263 }
1264 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1265
1266 /**
1267  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1268  * @sb: the superblock
1269  * @nr: the number of pages to write
1270  * @reason: reason why some writeback work was initiated
1271  *
1272  * Invoke writeback_inodes_sb if no writeback is currently underway.
1273  * Returns 1 if writeback was started, 0 if not.
1274  */
1275 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1276                                    unsigned long nr,
1277                                    enum wb_reason reason)
1278 {
1279         if (!writeback_in_progress(sb->s_bdi)) {
1280                 down_read(&sb->s_umount);
1281                 writeback_inodes_sb_nr(sb, nr, reason);
1282                 up_read(&sb->s_umount);
1283                 return 1;
1284         } else
1285                 return 0;
1286 }
1287 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1288
1289 /**
1290  * sync_inodes_sb       -       sync sb inode pages
1291  * @sb: the superblock
1292  *
1293  * This function writes and waits on any dirty inode belonging to this
1294  * super_block.
1295  */
1296 void sync_inodes_sb(struct super_block *sb)
1297 {
1298         DECLARE_COMPLETION_ONSTACK(done);
1299         struct wb_writeback_work work = {
1300                 .sb             = sb,
1301                 .sync_mode      = WB_SYNC_ALL,
1302                 .nr_pages       = LONG_MAX,
1303                 .range_cyclic   = 0,
1304                 .done           = &done,
1305                 .reason         = WB_REASON_SYNC,
1306         };
1307
1308         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1309
1310         bdi_queue_work(sb->s_bdi, &work);
1311         wait_for_completion(&done);
1312
1313         wait_sb_inodes(sb);
1314 }
1315 EXPORT_SYMBOL(sync_inodes_sb);
1316
1317 /**
1318  * write_inode_now      -       write an inode to disk
1319  * @inode: inode to write to disk
1320  * @sync: whether the write should be synchronous or not
1321  *
1322  * This function commits an inode to disk immediately if it is dirty. This is
1323  * primarily needed by knfsd.
1324  *
1325  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1326  */
1327 int write_inode_now(struct inode *inode, int sync)
1328 {
1329         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1330         int ret;
1331         struct writeback_control wbc = {
1332                 .nr_to_write = LONG_MAX,
1333                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1334                 .range_start = 0,
1335                 .range_end = LLONG_MAX,
1336         };
1337
1338         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1339                 wbc.nr_to_write = 0;
1340
1341         might_sleep();
1342         spin_lock(&wb->list_lock);
1343         spin_lock(&inode->i_lock);
1344         ret = writeback_single_inode(inode, wb, &wbc);
1345         spin_unlock(&inode->i_lock);
1346         spin_unlock(&wb->list_lock);
1347         return ret;
1348 }
1349 EXPORT_SYMBOL(write_inode_now);
1350
1351 /**
1352  * sync_inode - write an inode and its pages to disk.
1353  * @inode: the inode to sync
1354  * @wbc: controls the writeback mode
1355  *
1356  * sync_inode() will write an inode and its pages to disk.  It will also
1357  * correctly update the inode on its superblock's dirty inode lists and will
1358  * update inode->i_state.
1359  *
1360  * The caller must have a ref on the inode.
1361  */
1362 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1363 {
1364         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1365         int ret;
1366
1367         spin_lock(&wb->list_lock);
1368         spin_lock(&inode->i_lock);
1369         ret = writeback_single_inode(inode, wb, wbc);
1370         spin_unlock(&inode->i_lock);
1371         spin_unlock(&wb->list_lock);
1372         return ret;
1373 }
1374 EXPORT_SYMBOL(sync_inode);
1375
1376 /**
1377  * sync_inode_metadata - write an inode to disk
1378  * @inode: the inode to sync
1379  * @wait: wait for I/O to complete.
1380  *
1381  * Write an inode to disk and adjust its dirty state after completion.
1382  *
1383  * Note: only writes the actual inode, no associated data or other metadata.
1384  */
1385 int sync_inode_metadata(struct inode *inode, int wait)
1386 {
1387         struct writeback_control wbc = {
1388                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1389                 .nr_to_write = 0, /* metadata-only */
1390         };
1391
1392         return sync_inode(inode, &wbc);
1393 }
1394 EXPORT_SYMBOL(sync_inode_metadata);