writeback: Move requeueing when I_SYNC set to writeback_sb_inodes()
[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                 if (wbc->sync_mode != WB_SYNC_ALL)
377                         return 0;
378                 /*
379                  * It's a data-integrity sync.  We must wait.
380                  */
381                 inode_wait_for_writeback(inode, wb);
382         }
383
384         BUG_ON(inode->i_state & I_SYNC);
385
386         /* Set I_SYNC, reset I_DIRTY_PAGES */
387         inode->i_state |= I_SYNC;
388         inode->i_state &= ~I_DIRTY_PAGES;
389         spin_unlock(&inode->i_lock);
390         spin_unlock(&wb->list_lock);
391
392         ret = do_writepages(mapping, wbc);
393
394         /*
395          * Make sure to wait on the data before writing out the metadata.
396          * This is important for filesystems that modify metadata on data
397          * I/O completion.
398          */
399         if (wbc->sync_mode == WB_SYNC_ALL) {
400                 int err = filemap_fdatawait(mapping);
401                 if (ret == 0)
402                         ret = err;
403         }
404
405         /*
406          * Some filesystems may redirty the inode during the writeback
407          * due to delalloc, clear dirty metadata flags right before
408          * write_inode()
409          */
410         spin_lock(&inode->i_lock);
411         dirty = inode->i_state & I_DIRTY;
412         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
413         spin_unlock(&inode->i_lock);
414         /* Don't write the inode if only I_DIRTY_PAGES was set */
415         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
416                 int err = write_inode(inode, wbc);
417                 if (ret == 0)
418                         ret = err;
419         }
420
421         spin_lock(&wb->list_lock);
422         spin_lock(&inode->i_lock);
423         if (!(inode->i_state & I_FREEING)) {
424                 /*
425                  * Sync livelock prevention. Each inode is tagged and synced in
426                  * one shot. If still dirty, it will be redirty_tail()'ed below.
427                  * Update the dirty time to prevent enqueue and sync it again.
428                  */
429                 if ((inode->i_state & I_DIRTY) &&
430                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
431                         inode->dirtied_when = jiffies;
432
433                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
434                         /*
435                          * We didn't write back all the pages.  nfs_writepages()
436                          * sometimes bales out without doing anything.
437                          */
438                         inode->i_state |= I_DIRTY_PAGES;
439                         if (wbc->nr_to_write <= 0) {
440                                 /*
441                                  * slice used up: queue for next turn
442                                  */
443                                 requeue_io(inode, wb);
444                         } else {
445                                 /*
446                                  * Writeback blocked by something other than
447                                  * congestion. Delay the inode for some time to
448                                  * avoid spinning on the CPU (100% iowait)
449                                  * retrying writeback of the dirty page/inode
450                                  * that cannot be performed immediately.
451                                  */
452                                 redirty_tail(inode, wb);
453                         }
454                 } else if (inode->i_state & I_DIRTY) {
455                         /*
456                          * Filesystems can dirty the inode during writeback
457                          * operations, such as delayed allocation during
458                          * submission or metadata updates after data IO
459                          * completion.
460                          */
461                         redirty_tail(inode, wb);
462                 } else {
463                         /*
464                          * The inode is clean.  At this point we either have
465                          * a reference to the inode or it's on it's way out.
466                          * No need to add it back to the LRU.
467                          */
468                         list_del_init(&inode->i_wb_list);
469                 }
470         }
471         inode_sync_complete(inode);
472         trace_writeback_single_inode(inode, wbc, nr_to_write);
473         return ret;
474 }
475
476 static long writeback_chunk_size(struct backing_dev_info *bdi,
477                                  struct wb_writeback_work *work)
478 {
479         long pages;
480
481         /*
482          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
483          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
484          * here avoids calling into writeback_inodes_wb() more than once.
485          *
486          * The intended call sequence for WB_SYNC_ALL writeback is:
487          *
488          *      wb_writeback()
489          *          writeback_sb_inodes()       <== called only once
490          *              write_cache_pages()     <== called once for each inode
491          *                   (quickly) tag currently dirty pages
492          *                   (maybe slowly) sync all tagged pages
493          */
494         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
495                 pages = LONG_MAX;
496         else {
497                 pages = min(bdi->avg_write_bandwidth / 2,
498                             global_dirty_limit / DIRTY_SCOPE);
499                 pages = min(pages, work->nr_pages);
500                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
501                                    MIN_WRITEBACK_PAGES);
502         }
503
504         return pages;
505 }
506
507 /*
508  * Write a portion of b_io inodes which belong to @sb.
509  *
510  * If @only_this_sb is true, then find and write all such
511  * inodes. Otherwise write only ones which go sequentially
512  * in reverse order.
513  *
514  * Return the number of pages and/or inodes written.
515  */
516 static long writeback_sb_inodes(struct super_block *sb,
517                                 struct bdi_writeback *wb,
518                                 struct wb_writeback_work *work)
519 {
520         struct writeback_control wbc = {
521                 .sync_mode              = work->sync_mode,
522                 .tagged_writepages      = work->tagged_writepages,
523                 .for_kupdate            = work->for_kupdate,
524                 .for_background         = work->for_background,
525                 .range_cyclic           = work->range_cyclic,
526                 .range_start            = 0,
527                 .range_end              = LLONG_MAX,
528         };
529         unsigned long start_time = jiffies;
530         long write_chunk;
531         long wrote = 0;  /* count both pages and inodes */
532
533         while (!list_empty(&wb->b_io)) {
534                 struct inode *inode = wb_inode(wb->b_io.prev);
535
536                 if (inode->i_sb != sb) {
537                         if (work->sb) {
538                                 /*
539                                  * We only want to write back data for this
540                                  * superblock, move all inodes not belonging
541                                  * to it back onto the dirty list.
542                                  */
543                                 redirty_tail(inode, wb);
544                                 continue;
545                         }
546
547                         /*
548                          * The inode belongs to a different superblock.
549                          * Bounce back to the caller to unpin this and
550                          * pin the next superblock.
551                          */
552                         break;
553                 }
554
555                 /*
556                  * Don't bother with new inodes or inodes beeing freed, first
557                  * kind does not need peridic writeout yet, and for the latter
558                  * kind writeout is handled by the freer.
559                  */
560                 spin_lock(&inode->i_lock);
561                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
562                         spin_unlock(&inode->i_lock);
563                         redirty_tail(inode, wb);
564                         continue;
565                 }
566                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
567                         /*
568                          * If this inode is locked for writeback and we are not
569                          * doing writeback-for-data-integrity, move it to
570                          * b_more_io so that writeback can proceed with the
571                          * other inodes on s_io.
572                          *
573                          * We'll have another go at writing back this inode
574                          * when we completed a full scan of b_io.
575                          */
576                         spin_unlock(&inode->i_lock);
577                         requeue_io(inode, wb);
578                         trace_writeback_sb_inodes_requeue(inode);
579                         continue;
580                 }
581                 __iget(inode);
582                 write_chunk = writeback_chunk_size(wb->bdi, work);
583                 wbc.nr_to_write = write_chunk;
584                 wbc.pages_skipped = 0;
585
586                 writeback_single_inode(inode, wb, &wbc);
587
588                 work->nr_pages -= write_chunk - wbc.nr_to_write;
589                 wrote += write_chunk - wbc.nr_to_write;
590                 if (!(inode->i_state & I_DIRTY))
591                         wrote++;
592                 if (wbc.pages_skipped) {
593                         /*
594                          * writeback is not making progress due to locked
595                          * buffers.  Skip this inode for now.
596                          */
597                         redirty_tail(inode, wb);
598                 }
599                 spin_unlock(&inode->i_lock);
600                 spin_unlock(&wb->list_lock);
601                 iput(inode);
602                 cond_resched();
603                 spin_lock(&wb->list_lock);
604                 /*
605                  * bail out to wb_writeback() often enough to check
606                  * background threshold and other termination conditions.
607                  */
608                 if (wrote) {
609                         if (time_is_before_jiffies(start_time + HZ / 10UL))
610                                 break;
611                         if (work->nr_pages <= 0)
612                                 break;
613                 }
614         }
615         return wrote;
616 }
617
618 static long __writeback_inodes_wb(struct bdi_writeback *wb,
619                                   struct wb_writeback_work *work)
620 {
621         unsigned long start_time = jiffies;
622         long wrote = 0;
623
624         while (!list_empty(&wb->b_io)) {
625                 struct inode *inode = wb_inode(wb->b_io.prev);
626                 struct super_block *sb = inode->i_sb;
627
628                 if (!grab_super_passive(sb)) {
629                         /*
630                          * grab_super_passive() may fail consistently due to
631                          * s_umount being grabbed by someone else. Don't use
632                          * requeue_io() to avoid busy retrying the inode/sb.
633                          */
634                         redirty_tail(inode, wb);
635                         continue;
636                 }
637                 wrote += writeback_sb_inodes(sb, wb, work);
638                 drop_super(sb);
639
640                 /* refer to the same tests at the end of writeback_sb_inodes */
641                 if (wrote) {
642                         if (time_is_before_jiffies(start_time + HZ / 10UL))
643                                 break;
644                         if (work->nr_pages <= 0)
645                                 break;
646                 }
647         }
648         /* Leave any unwritten inodes on b_io */
649         return wrote;
650 }
651
652 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
653                                 enum wb_reason reason)
654 {
655         struct wb_writeback_work work = {
656                 .nr_pages       = nr_pages,
657                 .sync_mode      = WB_SYNC_NONE,
658                 .range_cyclic   = 1,
659                 .reason         = reason,
660         };
661
662         spin_lock(&wb->list_lock);
663         if (list_empty(&wb->b_io))
664                 queue_io(wb, &work);
665         __writeback_inodes_wb(wb, &work);
666         spin_unlock(&wb->list_lock);
667
668         return nr_pages - work.nr_pages;
669 }
670
671 static bool over_bground_thresh(struct backing_dev_info *bdi)
672 {
673         unsigned long background_thresh, dirty_thresh;
674
675         global_dirty_limits(&background_thresh, &dirty_thresh);
676
677         if (global_page_state(NR_FILE_DIRTY) +
678             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
679                 return true;
680
681         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
682                                 bdi_dirty_limit(bdi, background_thresh))
683                 return true;
684
685         return false;
686 }
687
688 /*
689  * Called under wb->list_lock. If there are multiple wb per bdi,
690  * only the flusher working on the first wb should do it.
691  */
692 static void wb_update_bandwidth(struct bdi_writeback *wb,
693                                 unsigned long start_time)
694 {
695         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
696 }
697
698 /*
699  * Explicit flushing or periodic writeback of "old" data.
700  *
701  * Define "old": the first time one of an inode's pages is dirtied, we mark the
702  * dirtying-time in the inode's address_space.  So this periodic writeback code
703  * just walks the superblock inode list, writing back any inodes which are
704  * older than a specific point in time.
705  *
706  * Try to run once per dirty_writeback_interval.  But if a writeback event
707  * takes longer than a dirty_writeback_interval interval, then leave a
708  * one-second gap.
709  *
710  * older_than_this takes precedence over nr_to_write.  So we'll only write back
711  * all dirty pages if they are all attached to "old" mappings.
712  */
713 static long wb_writeback(struct bdi_writeback *wb,
714                          struct wb_writeback_work *work)
715 {
716         unsigned long wb_start = jiffies;
717         long nr_pages = work->nr_pages;
718         unsigned long oldest_jif;
719         struct inode *inode;
720         long progress;
721
722         oldest_jif = jiffies;
723         work->older_than_this = &oldest_jif;
724
725         spin_lock(&wb->list_lock);
726         for (;;) {
727                 /*
728                  * Stop writeback when nr_pages has been consumed
729                  */
730                 if (work->nr_pages <= 0)
731                         break;
732
733                 /*
734                  * Background writeout and kupdate-style writeback may
735                  * run forever. Stop them if there is other work to do
736                  * so that e.g. sync can proceed. They'll be restarted
737                  * after the other works are all done.
738                  */
739                 if ((work->for_background || work->for_kupdate) &&
740                     !list_empty(&wb->bdi->work_list))
741                         break;
742
743                 /*
744                  * For background writeout, stop when we are below the
745                  * background dirty threshold
746                  */
747                 if (work->for_background && !over_bground_thresh(wb->bdi))
748                         break;
749
750                 /*
751                  * Kupdate and background works are special and we want to
752                  * include all inodes that need writing. Livelock avoidance is
753                  * handled by these works yielding to any other work so we are
754                  * safe.
755                  */
756                 if (work->for_kupdate) {
757                         oldest_jif = jiffies -
758                                 msecs_to_jiffies(dirty_expire_interval * 10);
759                 } else if (work->for_background)
760                         oldest_jif = jiffies;
761
762                 trace_writeback_start(wb->bdi, work);
763                 if (list_empty(&wb->b_io))
764                         queue_io(wb, work);
765                 if (work->sb)
766                         progress = writeback_sb_inodes(work->sb, wb, work);
767                 else
768                         progress = __writeback_inodes_wb(wb, work);
769                 trace_writeback_written(wb->bdi, work);
770
771                 wb_update_bandwidth(wb, wb_start);
772
773                 /*
774                  * Did we write something? Try for more
775                  *
776                  * Dirty inodes are moved to b_io for writeback in batches.
777                  * The completion of the current batch does not necessarily
778                  * mean the overall work is done. So we keep looping as long
779                  * as made some progress on cleaning pages or inodes.
780                  */
781                 if (progress)
782                         continue;
783                 /*
784                  * No more inodes for IO, bail
785                  */
786                 if (list_empty(&wb->b_more_io))
787                         break;
788                 /*
789                  * Nothing written. Wait for some inode to
790                  * become available for writeback. Otherwise
791                  * we'll just busyloop.
792                  */
793                 if (!list_empty(&wb->b_more_io))  {
794                         trace_writeback_wait(wb->bdi, work);
795                         inode = wb_inode(wb->b_more_io.prev);
796                         spin_lock(&inode->i_lock);
797                         inode_wait_for_writeback(inode, wb);
798                         spin_unlock(&inode->i_lock);
799                 }
800         }
801         spin_unlock(&wb->list_lock);
802
803         return nr_pages - work->nr_pages;
804 }
805
806 /*
807  * Return the next wb_writeback_work struct that hasn't been processed yet.
808  */
809 static struct wb_writeback_work *
810 get_next_work_item(struct backing_dev_info *bdi)
811 {
812         struct wb_writeback_work *work = NULL;
813
814         spin_lock_bh(&bdi->wb_lock);
815         if (!list_empty(&bdi->work_list)) {
816                 work = list_entry(bdi->work_list.next,
817                                   struct wb_writeback_work, list);
818                 list_del_init(&work->list);
819         }
820         spin_unlock_bh(&bdi->wb_lock);
821         return work;
822 }
823
824 /*
825  * Add in the number of potentially dirty inodes, because each inode
826  * write can dirty pagecache in the underlying blockdev.
827  */
828 static unsigned long get_nr_dirty_pages(void)
829 {
830         return global_page_state(NR_FILE_DIRTY) +
831                 global_page_state(NR_UNSTABLE_NFS) +
832                 get_nr_dirty_inodes();
833 }
834
835 static long wb_check_background_flush(struct bdi_writeback *wb)
836 {
837         if (over_bground_thresh(wb->bdi)) {
838
839                 struct wb_writeback_work work = {
840                         .nr_pages       = LONG_MAX,
841                         .sync_mode      = WB_SYNC_NONE,
842                         .for_background = 1,
843                         .range_cyclic   = 1,
844                         .reason         = WB_REASON_BACKGROUND,
845                 };
846
847                 return wb_writeback(wb, &work);
848         }
849
850         return 0;
851 }
852
853 static long wb_check_old_data_flush(struct bdi_writeback *wb)
854 {
855         unsigned long expired;
856         long nr_pages;
857
858         /*
859          * When set to zero, disable periodic writeback
860          */
861         if (!dirty_writeback_interval)
862                 return 0;
863
864         expired = wb->last_old_flush +
865                         msecs_to_jiffies(dirty_writeback_interval * 10);
866         if (time_before(jiffies, expired))
867                 return 0;
868
869         wb->last_old_flush = jiffies;
870         nr_pages = get_nr_dirty_pages();
871
872         if (nr_pages) {
873                 struct wb_writeback_work work = {
874                         .nr_pages       = nr_pages,
875                         .sync_mode      = WB_SYNC_NONE,
876                         .for_kupdate    = 1,
877                         .range_cyclic   = 1,
878                         .reason         = WB_REASON_PERIODIC,
879                 };
880
881                 return wb_writeback(wb, &work);
882         }
883
884         return 0;
885 }
886
887 /*
888  * Retrieve work items and do the writeback they describe
889  */
890 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
891 {
892         struct backing_dev_info *bdi = wb->bdi;
893         struct wb_writeback_work *work;
894         long wrote = 0;
895
896         set_bit(BDI_writeback_running, &wb->bdi->state);
897         while ((work = get_next_work_item(bdi)) != NULL) {
898                 /*
899                  * Override sync mode, in case we must wait for completion
900                  * because this thread is exiting now.
901                  */
902                 if (force_wait)
903                         work->sync_mode = WB_SYNC_ALL;
904
905                 trace_writeback_exec(bdi, work);
906
907                 wrote += wb_writeback(wb, work);
908
909                 /*
910                  * Notify the caller of completion if this is a synchronous
911                  * work item, otherwise just free it.
912                  */
913                 if (work->done)
914                         complete(work->done);
915                 else
916                         kfree(work);
917         }
918
919         /*
920          * Check for periodic writeback, kupdated() style
921          */
922         wrote += wb_check_old_data_flush(wb);
923         wrote += wb_check_background_flush(wb);
924         clear_bit(BDI_writeback_running, &wb->bdi->state);
925
926         return wrote;
927 }
928
929 /*
930  * Handle writeback of dirty data for the device backed by this bdi. Also
931  * wakes up periodically and does kupdated style flushing.
932  */
933 int bdi_writeback_thread(void *data)
934 {
935         struct bdi_writeback *wb = data;
936         struct backing_dev_info *bdi = wb->bdi;
937         long pages_written;
938
939         current->flags |= PF_SWAPWRITE;
940         set_freezable();
941         wb->last_active = jiffies;
942
943         /*
944          * Our parent may run at a different priority, just set us to normal
945          */
946         set_user_nice(current, 0);
947
948         trace_writeback_thread_start(bdi);
949
950         while (!kthread_freezable_should_stop(NULL)) {
951                 /*
952                  * Remove own delayed wake-up timer, since we are already awake
953                  * and we'll take care of the preriodic write-back.
954                  */
955                 del_timer(&wb->wakeup_timer);
956
957                 pages_written = wb_do_writeback(wb, 0);
958
959                 trace_writeback_pages_written(pages_written);
960
961                 if (pages_written)
962                         wb->last_active = jiffies;
963
964                 set_current_state(TASK_INTERRUPTIBLE);
965                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
966                         __set_current_state(TASK_RUNNING);
967                         continue;
968                 }
969
970                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
971                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
972                 else {
973                         /*
974                          * We have nothing to do, so can go sleep without any
975                          * timeout and save power. When a work is queued or
976                          * something is made dirty - we will be woken up.
977                          */
978                         schedule();
979                 }
980         }
981
982         /* Flush any work that raced with us exiting */
983         if (!list_empty(&bdi->work_list))
984                 wb_do_writeback(wb, 1);
985
986         trace_writeback_thread_stop(bdi);
987         return 0;
988 }
989
990
991 /*
992  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
993  * the whole world.
994  */
995 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
996 {
997         struct backing_dev_info *bdi;
998
999         if (!nr_pages) {
1000                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1001                                 global_page_state(NR_UNSTABLE_NFS);
1002         }
1003
1004         rcu_read_lock();
1005         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1006                 if (!bdi_has_dirty_io(bdi))
1007                         continue;
1008                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1009         }
1010         rcu_read_unlock();
1011 }
1012
1013 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1014 {
1015         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1016                 struct dentry *dentry;
1017                 const char *name = "?";
1018
1019                 dentry = d_find_alias(inode);
1020                 if (dentry) {
1021                         spin_lock(&dentry->d_lock);
1022                         name = (const char *) dentry->d_name.name;
1023                 }
1024                 printk(KERN_DEBUG
1025                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1026                        current->comm, task_pid_nr(current), inode->i_ino,
1027                        name, inode->i_sb->s_id);
1028                 if (dentry) {
1029                         spin_unlock(&dentry->d_lock);
1030                         dput(dentry);
1031                 }
1032         }
1033 }
1034
1035 /**
1036  *      __mark_inode_dirty -    internal function
1037  *      @inode: inode to mark
1038  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1039  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1040  *      mark_inode_dirty_sync.
1041  *
1042  * Put the inode on the super block's dirty list.
1043  *
1044  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1045  * dirty list only if it is hashed or if it refers to a blockdev.
1046  * If it was not hashed, it will never be added to the dirty list
1047  * even if it is later hashed, as it will have been marked dirty already.
1048  *
1049  * In short, make sure you hash any inodes _before_ you start marking
1050  * them dirty.
1051  *
1052  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1053  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1054  * the kernel-internal blockdev inode represents the dirtying time of the
1055  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1056  * page->mapping->host, so the page-dirtying time is recorded in the internal
1057  * blockdev inode.
1058  */
1059 void __mark_inode_dirty(struct inode *inode, int flags)
1060 {
1061         struct super_block *sb = inode->i_sb;
1062         struct backing_dev_info *bdi = NULL;
1063
1064         /*
1065          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1066          * dirty the inode itself
1067          */
1068         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1069                 if (sb->s_op->dirty_inode)
1070                         sb->s_op->dirty_inode(inode, flags);
1071         }
1072
1073         /*
1074          * make sure that changes are seen by all cpus before we test i_state
1075          * -- mikulas
1076          */
1077         smp_mb();
1078
1079         /* avoid the locking if we can */
1080         if ((inode->i_state & flags) == flags)
1081                 return;
1082
1083         if (unlikely(block_dump))
1084                 block_dump___mark_inode_dirty(inode);
1085
1086         spin_lock(&inode->i_lock);
1087         if ((inode->i_state & flags) != flags) {
1088                 const int was_dirty = inode->i_state & I_DIRTY;
1089
1090                 inode->i_state |= flags;
1091
1092                 /*
1093                  * If the inode is being synced, just update its dirty state.
1094                  * The unlocker will place the inode on the appropriate
1095                  * superblock list, based upon its state.
1096                  */
1097                 if (inode->i_state & I_SYNC)
1098                         goto out_unlock_inode;
1099
1100                 /*
1101                  * Only add valid (hashed) inodes to the superblock's
1102                  * dirty list.  Add blockdev inodes as well.
1103                  */
1104                 if (!S_ISBLK(inode->i_mode)) {
1105                         if (inode_unhashed(inode))
1106                                 goto out_unlock_inode;
1107                 }
1108                 if (inode->i_state & I_FREEING)
1109                         goto out_unlock_inode;
1110
1111                 /*
1112                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1113                  * reposition it (that would break b_dirty time-ordering).
1114                  */
1115                 if (!was_dirty) {
1116                         bool wakeup_bdi = false;
1117                         bdi = inode_to_bdi(inode);
1118
1119                         if (bdi_cap_writeback_dirty(bdi)) {
1120                                 WARN(!test_bit(BDI_registered, &bdi->state),
1121                                      "bdi-%s not registered\n", bdi->name);
1122
1123                                 /*
1124                                  * If this is the first dirty inode for this
1125                                  * bdi, we have to wake-up the corresponding
1126                                  * bdi thread to make sure background
1127                                  * write-back happens later.
1128                                  */
1129                                 if (!wb_has_dirty_io(&bdi->wb))
1130                                         wakeup_bdi = true;
1131                         }
1132
1133                         spin_unlock(&inode->i_lock);
1134                         spin_lock(&bdi->wb.list_lock);
1135                         inode->dirtied_when = jiffies;
1136                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1137                         spin_unlock(&bdi->wb.list_lock);
1138
1139                         if (wakeup_bdi)
1140                                 bdi_wakeup_thread_delayed(bdi);
1141                         return;
1142                 }
1143         }
1144 out_unlock_inode:
1145         spin_unlock(&inode->i_lock);
1146
1147 }
1148 EXPORT_SYMBOL(__mark_inode_dirty);
1149
1150 static void wait_sb_inodes(struct super_block *sb)
1151 {
1152         struct inode *inode, *old_inode = NULL;
1153
1154         /*
1155          * We need to be protected against the filesystem going from
1156          * r/o to r/w or vice versa.
1157          */
1158         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1159
1160         spin_lock(&inode_sb_list_lock);
1161
1162         /*
1163          * Data integrity sync. Must wait for all pages under writeback,
1164          * because there may have been pages dirtied before our sync
1165          * call, but which had writeout started before we write it out.
1166          * In which case, the inode may not be on the dirty list, but
1167          * we still have to wait for that writeout.
1168          */
1169         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1170                 struct address_space *mapping = inode->i_mapping;
1171
1172                 spin_lock(&inode->i_lock);
1173                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1174                     (mapping->nrpages == 0)) {
1175                         spin_unlock(&inode->i_lock);
1176                         continue;
1177                 }
1178                 __iget(inode);
1179                 spin_unlock(&inode->i_lock);
1180                 spin_unlock(&inode_sb_list_lock);
1181
1182                 /*
1183                  * We hold a reference to 'inode' so it couldn't have been
1184                  * removed from s_inodes list while we dropped the
1185                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1186                  * be holding the last reference and we cannot iput it under
1187                  * inode_sb_list_lock. So we keep the reference and iput it
1188                  * later.
1189                  */
1190                 iput(old_inode);
1191                 old_inode = inode;
1192
1193                 filemap_fdatawait(mapping);
1194
1195                 cond_resched();
1196
1197                 spin_lock(&inode_sb_list_lock);
1198         }
1199         spin_unlock(&inode_sb_list_lock);
1200         iput(old_inode);
1201 }
1202
1203 /**
1204  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1205  * @sb: the superblock
1206  * @nr: the number of pages to write
1207  * @reason: reason why some writeback work initiated
1208  *
1209  * Start writeback on some inodes on this super_block. No guarantees are made
1210  * on how many (if any) will be written, and this function does not wait
1211  * for IO completion of submitted IO.
1212  */
1213 void writeback_inodes_sb_nr(struct super_block *sb,
1214                             unsigned long nr,
1215                             enum wb_reason reason)
1216 {
1217         DECLARE_COMPLETION_ONSTACK(done);
1218         struct wb_writeback_work work = {
1219                 .sb                     = sb,
1220                 .sync_mode              = WB_SYNC_NONE,
1221                 .tagged_writepages      = 1,
1222                 .done                   = &done,
1223                 .nr_pages               = nr,
1224                 .reason                 = reason,
1225         };
1226
1227         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1228         bdi_queue_work(sb->s_bdi, &work);
1229         wait_for_completion(&done);
1230 }
1231 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1232
1233 /**
1234  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1235  * @sb: the superblock
1236  * @reason: reason why some writeback work was initiated
1237  *
1238  * Start writeback on some inodes on this super_block. No guarantees are made
1239  * on how many (if any) will be written, and this function does not wait
1240  * for IO completion of submitted IO.
1241  */
1242 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1243 {
1244         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1245 }
1246 EXPORT_SYMBOL(writeback_inodes_sb);
1247
1248 /**
1249  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1250  * @sb: the superblock
1251  * @reason: reason why some writeback work was initiated
1252  *
1253  * Invoke writeback_inodes_sb if no writeback is currently underway.
1254  * Returns 1 if writeback was started, 0 if not.
1255  */
1256 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1257 {
1258         if (!writeback_in_progress(sb->s_bdi)) {
1259                 down_read(&sb->s_umount);
1260                 writeback_inodes_sb(sb, reason);
1261                 up_read(&sb->s_umount);
1262                 return 1;
1263         } else
1264                 return 0;
1265 }
1266 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1267
1268 /**
1269  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1270  * @sb: the superblock
1271  * @nr: the number of pages to write
1272  * @reason: reason why some writeback work was initiated
1273  *
1274  * Invoke writeback_inodes_sb if no writeback is currently underway.
1275  * Returns 1 if writeback was started, 0 if not.
1276  */
1277 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1278                                    unsigned long nr,
1279                                    enum wb_reason reason)
1280 {
1281         if (!writeback_in_progress(sb->s_bdi)) {
1282                 down_read(&sb->s_umount);
1283                 writeback_inodes_sb_nr(sb, nr, reason);
1284                 up_read(&sb->s_umount);
1285                 return 1;
1286         } else
1287                 return 0;
1288 }
1289 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1290
1291 /**
1292  * sync_inodes_sb       -       sync sb inode pages
1293  * @sb: the superblock
1294  *
1295  * This function writes and waits on any dirty inode belonging to this
1296  * super_block.
1297  */
1298 void sync_inodes_sb(struct super_block *sb)
1299 {
1300         DECLARE_COMPLETION_ONSTACK(done);
1301         struct wb_writeback_work work = {
1302                 .sb             = sb,
1303                 .sync_mode      = WB_SYNC_ALL,
1304                 .nr_pages       = LONG_MAX,
1305                 .range_cyclic   = 0,
1306                 .done           = &done,
1307                 .reason         = WB_REASON_SYNC,
1308         };
1309
1310         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1311
1312         bdi_queue_work(sb->s_bdi, &work);
1313         wait_for_completion(&done);
1314
1315         wait_sb_inodes(sb);
1316 }
1317 EXPORT_SYMBOL(sync_inodes_sb);
1318
1319 /**
1320  * write_inode_now      -       write an inode to disk
1321  * @inode: inode to write to disk
1322  * @sync: whether the write should be synchronous or not
1323  *
1324  * This function commits an inode to disk immediately if it is dirty. This is
1325  * primarily needed by knfsd.
1326  *
1327  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1328  */
1329 int write_inode_now(struct inode *inode, int sync)
1330 {
1331         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1332         int ret;
1333         struct writeback_control wbc = {
1334                 .nr_to_write = LONG_MAX,
1335                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1336                 .range_start = 0,
1337                 .range_end = LLONG_MAX,
1338         };
1339
1340         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1341                 wbc.nr_to_write = 0;
1342
1343         might_sleep();
1344         spin_lock(&wb->list_lock);
1345         spin_lock(&inode->i_lock);
1346         ret = writeback_single_inode(inode, wb, &wbc);
1347         spin_unlock(&inode->i_lock);
1348         spin_unlock(&wb->list_lock);
1349         return ret;
1350 }
1351 EXPORT_SYMBOL(write_inode_now);
1352
1353 /**
1354  * sync_inode - write an inode and its pages to disk.
1355  * @inode: the inode to sync
1356  * @wbc: controls the writeback mode
1357  *
1358  * sync_inode() will write an inode and its pages to disk.  It will also
1359  * correctly update the inode on its superblock's dirty inode lists and will
1360  * update inode->i_state.
1361  *
1362  * The caller must have a ref on the inode.
1363  */
1364 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1365 {
1366         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1367         int ret;
1368
1369         spin_lock(&wb->list_lock);
1370         spin_lock(&inode->i_lock);
1371         ret = writeback_single_inode(inode, wb, wbc);
1372         spin_unlock(&inode->i_lock);
1373         spin_unlock(&wb->list_lock);
1374         return ret;
1375 }
1376 EXPORT_SYMBOL(sync_inode);
1377
1378 /**
1379  * sync_inode_metadata - write an inode to disk
1380  * @inode: the inode to sync
1381  * @wait: wait for I/O to complete.
1382  *
1383  * Write an inode to disk and adjust its dirty state after completion.
1384  *
1385  * Note: only writes the actual inode, no associated data or other metadata.
1386  */
1387 int sync_inode_metadata(struct inode *inode, int wait)
1388 {
1389         struct writeback_control wbc = {
1390                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1391                 .nr_to_write = 0, /* metadata-only */
1392         };
1393
1394         return sync_inode(inode, &wbc);
1395 }
1396 EXPORT_SYMBOL(sync_inode_metadata);