writeback: make writeback_control.nr_to_write straight
[linux-3.10.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * The maximum number of pages to writeout in a single bdi flush/kupdate
34  * operation.  We do this so we don't hold I_SYNC against an inode for
35  * enormous amounts of time, which would block a userspace task which has
36  * been forced to throttle against that inode.  Also, the code reevaluates
37  * the dirty each time it has written this many pages.
38  */
39 #define MAX_WRITEBACK_PAGES     1024L
40
41 /*
42  * Passed into wb_writeback(), essentially a subset of writeback_control
43  */
44 struct wb_writeback_work {
45         long nr_pages;
46         struct super_block *sb;
47         unsigned long *older_than_this;
48         enum writeback_sync_modes sync_mode;
49         unsigned int tagged_writepages:1;
50         unsigned int for_kupdate:1;
51         unsigned int range_cyclic:1;
52         unsigned int for_background:1;
53
54         struct list_head list;          /* pending work list */
55         struct completion *done;        /* set if the caller waits */
56 };
57
58 /*
59  * Include the creation of the trace points after defining the
60  * wb_writeback_work structure so that the definition remains local to this
61  * file.
62  */
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/writeback.h>
65
66 /*
67  * We don't actually have pdflush, but this one is exported though /proc...
68  */
69 int nr_pdflush_threads;
70
71 /**
72  * writeback_in_progress - determine whether there is writeback in progress
73  * @bdi: the device's backing_dev_info structure.
74  *
75  * Determine whether there is writeback waiting to be handled against a
76  * backing device.
77  */
78 int writeback_in_progress(struct backing_dev_info *bdi)
79 {
80         return test_bit(BDI_writeback_running, &bdi->state);
81 }
82
83 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
84 {
85         struct super_block *sb = inode->i_sb;
86
87         if (strcmp(sb->s_type->name, "bdev") == 0)
88                 return inode->i_mapping->backing_dev_info;
89
90         return sb->s_bdi;
91 }
92
93 static inline struct inode *wb_inode(struct list_head *head)
94 {
95         return list_entry(head, struct inode, i_wb_list);
96 }
97
98 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
99 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
100 {
101         if (bdi->wb.task) {
102                 wake_up_process(bdi->wb.task);
103         } else {
104                 /*
105                  * The bdi thread isn't there, wake up the forker thread which
106                  * will create and run it.
107                  */
108                 wake_up_process(default_backing_dev_info.wb.task);
109         }
110 }
111
112 static void bdi_queue_work(struct backing_dev_info *bdi,
113                            struct wb_writeback_work *work)
114 {
115         trace_writeback_queue(bdi, work);
116
117         spin_lock_bh(&bdi->wb_lock);
118         list_add_tail(&work->list, &bdi->work_list);
119         if (!bdi->wb.task)
120                 trace_writeback_nothread(bdi, work);
121         bdi_wakeup_flusher(bdi);
122         spin_unlock_bh(&bdi->wb_lock);
123 }
124
125 static void
126 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
127                       bool range_cyclic)
128 {
129         struct wb_writeback_work *work;
130
131         /*
132          * This is WB_SYNC_NONE writeback, so if allocation fails just
133          * wakeup the thread for old dirty data writeback
134          */
135         work = kzalloc(sizeof(*work), GFP_ATOMIC);
136         if (!work) {
137                 if (bdi->wb.task) {
138                         trace_writeback_nowork(bdi);
139                         wake_up_process(bdi->wb.task);
140                 }
141                 return;
142         }
143
144         work->sync_mode = WB_SYNC_NONE;
145         work->nr_pages  = nr_pages;
146         work->range_cyclic = range_cyclic;
147
148         bdi_queue_work(bdi, work);
149 }
150
151 /**
152  * bdi_start_writeback - start writeback
153  * @bdi: the backing device to write from
154  * @nr_pages: the number of pages to write
155  *
156  * Description:
157  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
158  *   started when this function returns, we make no guarantees on
159  *   completion. Caller need not hold sb s_umount semaphore.
160  *
161  */
162 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
163 {
164         __bdi_start_writeback(bdi, nr_pages, true);
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         /*
235          * Prevent speculative execution through
236          * spin_unlock(&wb->list_lock);
237          */
238
239         smp_mb();
240         wake_up_bit(&inode->i_state, __I_SYNC);
241 }
242
243 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
244 {
245         bool ret = time_after(inode->dirtied_when, t);
246 #ifndef CONFIG_64BIT
247         /*
248          * For inodes being constantly redirtied, dirtied_when can get stuck.
249          * It _appears_ to be in the future, but is actually in distant past.
250          * This test is necessary to prevent such wrapped-around relative times
251          * from permanently stopping the whole bdi writeback.
252          */
253         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
254 #endif
255         return ret;
256 }
257
258 /*
259  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
260  */
261 static int move_expired_inodes(struct list_head *delaying_queue,
262                                struct list_head *dispatch_queue,
263                                unsigned long *older_than_this)
264 {
265         LIST_HEAD(tmp);
266         struct list_head *pos, *node;
267         struct super_block *sb = NULL;
268         struct inode *inode;
269         int do_sb_sort = 0;
270         int moved = 0;
271
272         while (!list_empty(delaying_queue)) {
273                 inode = wb_inode(delaying_queue->prev);
274                 if (older_than_this &&
275                     inode_dirtied_after(inode, *older_than_this))
276                         break;
277                 if (sb && sb != inode->i_sb)
278                         do_sb_sort = 1;
279                 sb = inode->i_sb;
280                 list_move(&inode->i_wb_list, &tmp);
281                 moved++;
282         }
283
284         /* just one sb in list, splice to dispatch_queue and we're done */
285         if (!do_sb_sort) {
286                 list_splice(&tmp, dispatch_queue);
287                 goto out;
288         }
289
290         /* Move inodes from one superblock together */
291         while (!list_empty(&tmp)) {
292                 sb = wb_inode(tmp.prev)->i_sb;
293                 list_for_each_prev_safe(pos, node, &tmp) {
294                         inode = wb_inode(pos);
295                         if (inode->i_sb == sb)
296                                 list_move(&inode->i_wb_list, dispatch_queue);
297                 }
298         }
299 out:
300         return moved;
301 }
302
303 /*
304  * Queue all expired dirty inodes for io, eldest first.
305  * Before
306  *         newly dirtied     b_dirty    b_io    b_more_io
307  *         =============>    gf         edc     BA
308  * After
309  *         newly dirtied     b_dirty    b_io    b_more_io
310  *         =============>    g          fBAedc
311  *                                           |
312  *                                           +--> dequeue for IO
313  */
314 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
315 {
316         int moved;
317         assert_spin_locked(&wb->list_lock);
318         list_splice_init(&wb->b_more_io, &wb->b_io);
319         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
320         trace_writeback_queue_io(wb, older_than_this, moved);
321 }
322
323 static int write_inode(struct inode *inode, struct writeback_control *wbc)
324 {
325         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
326                 return inode->i_sb->s_op->write_inode(inode, wbc);
327         return 0;
328 }
329
330 /*
331  * Wait for writeback on an inode to complete.
332  */
333 static void inode_wait_for_writeback(struct inode *inode,
334                                      struct bdi_writeback *wb)
335 {
336         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
337         wait_queue_head_t *wqh;
338
339         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
340         while (inode->i_state & I_SYNC) {
341                 spin_unlock(&inode->i_lock);
342                 spin_unlock(&wb->list_lock);
343                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
344                 spin_lock(&wb->list_lock);
345                 spin_lock(&inode->i_lock);
346         }
347 }
348
349 /*
350  * Write out an inode's dirty pages.  Called under wb->list_lock and
351  * inode->i_lock.  Either the caller has an active reference on the inode or
352  * the inode has I_WILL_FREE set.
353  *
354  * If `wait' is set, wait on the writeout.
355  *
356  * The whole writeout design is quite complex and fragile.  We want to avoid
357  * starvation of particular inodes when others are being redirtied, prevent
358  * livelocks, etc.
359  */
360 static int
361 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
362                        struct writeback_control *wbc)
363 {
364         struct address_space *mapping = inode->i_mapping;
365         long nr_to_write = wbc->nr_to_write;
366         unsigned dirty;
367         int ret;
368
369         assert_spin_locked(&wb->list_lock);
370         assert_spin_locked(&inode->i_lock);
371
372         if (!atomic_read(&inode->i_count))
373                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
374         else
375                 WARN_ON(inode->i_state & I_WILL_FREE);
376
377         if (inode->i_state & I_SYNC) {
378                 /*
379                  * If this inode is locked for writeback and we are not doing
380                  * writeback-for-data-integrity, move it to b_more_io so that
381                  * writeback can proceed with the other inodes on s_io.
382                  *
383                  * We'll have another go at writing back this inode when we
384                  * completed a full scan of b_io.
385                  */
386                 if (wbc->sync_mode != WB_SYNC_ALL) {
387                         requeue_io(inode, wb);
388                         trace_writeback_single_inode_requeue(inode, wbc,
389                                                              nr_to_write);
390                         return 0;
391                 }
392
393                 /*
394                  * It's a data-integrity sync.  We must wait.
395                  */
396                 inode_wait_for_writeback(inode, wb);
397         }
398
399         BUG_ON(inode->i_state & I_SYNC);
400
401         /* Set I_SYNC, reset I_DIRTY_PAGES */
402         inode->i_state |= I_SYNC;
403         inode->i_state &= ~I_DIRTY_PAGES;
404         spin_unlock(&inode->i_lock);
405         spin_unlock(&wb->list_lock);
406
407         ret = do_writepages(mapping, wbc);
408
409         /*
410          * Make sure to wait on the data before writing out the metadata.
411          * This is important for filesystems that modify metadata on data
412          * I/O completion.
413          */
414         if (wbc->sync_mode == WB_SYNC_ALL) {
415                 int err = filemap_fdatawait(mapping);
416                 if (ret == 0)
417                         ret = err;
418         }
419
420         /*
421          * Some filesystems may redirty the inode during the writeback
422          * due to delalloc, clear dirty metadata flags right before
423          * write_inode()
424          */
425         spin_lock(&inode->i_lock);
426         dirty = inode->i_state & I_DIRTY;
427         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
428         spin_unlock(&inode->i_lock);
429         /* Don't write the inode if only I_DIRTY_PAGES was set */
430         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
431                 int err = write_inode(inode, wbc);
432                 if (ret == 0)
433                         ret = err;
434         }
435
436         spin_lock(&wb->list_lock);
437         spin_lock(&inode->i_lock);
438         inode->i_state &= ~I_SYNC;
439         if (!(inode->i_state & I_FREEING)) {
440                 /*
441                  * Sync livelock prevention. Each inode is tagged and synced in
442                  * one shot. If still dirty, it will be redirty_tail()'ed below.
443                  * Update the dirty time to prevent enqueue and sync it again.
444                  */
445                 if ((inode->i_state & I_DIRTY) &&
446                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
447                         inode->dirtied_when = jiffies;
448
449                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
450                         /*
451                          * We didn't write back all the pages.  nfs_writepages()
452                          * sometimes bales out without doing anything.
453                          */
454                         inode->i_state |= I_DIRTY_PAGES;
455                         if (wbc->nr_to_write <= 0) {
456                                 /*
457                                  * slice used up: queue for next turn
458                                  */
459                                 requeue_io(inode, wb);
460                         } else {
461                                 /*
462                                  * Writeback blocked by something other than
463                                  * congestion. Delay the inode for some time to
464                                  * avoid spinning on the CPU (100% iowait)
465                                  * retrying writeback of the dirty page/inode
466                                  * that cannot be performed immediately.
467                                  */
468                                 redirty_tail(inode, wb);
469                         }
470                 } else if (inode->i_state & I_DIRTY) {
471                         /*
472                          * Filesystems can dirty the inode during writeback
473                          * operations, such as delayed allocation during
474                          * submission or metadata updates after data IO
475                          * completion.
476                          */
477                         redirty_tail(inode, wb);
478                 } else {
479                         /*
480                          * The inode is clean.  At this point we either have
481                          * a reference to the inode or it's on it's way out.
482                          * No need to add it back to the LRU.
483                          */
484                         list_del_init(&inode->i_wb_list);
485                 }
486         }
487         inode_sync_complete(inode);
488         trace_writeback_single_inode(inode, wbc, nr_to_write);
489         return ret;
490 }
491
492 /*
493  * For background writeback the caller does not have the sb pinned
494  * before calling writeback. So make sure that we do pin it, so it doesn't
495  * go away while we are writing inodes from it.
496  */
497 static bool pin_sb_for_writeback(struct super_block *sb)
498 {
499         spin_lock(&sb_lock);
500         if (list_empty(&sb->s_instances)) {
501                 spin_unlock(&sb_lock);
502                 return false;
503         }
504
505         sb->s_count++;
506         spin_unlock(&sb_lock);
507
508         if (down_read_trylock(&sb->s_umount)) {
509                 if (sb->s_root)
510                         return true;
511                 up_read(&sb->s_umount);
512         }
513
514         put_super(sb);
515         return false;
516 }
517
518 static long writeback_chunk_size(struct wb_writeback_work *work)
519 {
520         long pages;
521
522         /*
523          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
524          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
525          * here avoids calling into writeback_inodes_wb() more than once.
526          *
527          * The intended call sequence for WB_SYNC_ALL writeback is:
528          *
529          *      wb_writeback()
530          *          writeback_sb_inodes()       <== called only once
531          *              write_cache_pages()     <== called once for each inode
532          *                   (quickly) tag currently dirty pages
533          *                   (maybe slowly) sync all tagged pages
534          */
535         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
536                 pages = LONG_MAX;
537         else
538                 pages = min(MAX_WRITEBACK_PAGES, work->nr_pages);
539
540         return pages;
541 }
542
543 /*
544  * Write a portion of b_io inodes which belong to @sb.
545  *
546  * If @only_this_sb is true, then find and write all such
547  * inodes. Otherwise write only ones which go sequentially
548  * in reverse order.
549  *
550  * Return the number of pages and/or inodes written.
551  */
552 static long writeback_sb_inodes(struct super_block *sb,
553                                 struct bdi_writeback *wb,
554                                 struct wb_writeback_work *work)
555 {
556         struct writeback_control wbc = {
557                 .sync_mode              = work->sync_mode,
558                 .tagged_writepages      = work->tagged_writepages,
559                 .for_kupdate            = work->for_kupdate,
560                 .for_background         = work->for_background,
561                 .range_cyclic           = work->range_cyclic,
562                 .range_start            = 0,
563                 .range_end              = LLONG_MAX,
564         };
565         unsigned long start_time = jiffies;
566         long write_chunk;
567         long wrote = 0;  /* count both pages and inodes */
568
569         while (!list_empty(&wb->b_io)) {
570                 struct inode *inode = wb_inode(wb->b_io.prev);
571
572                 if (inode->i_sb != sb) {
573                         if (work->sb) {
574                                 /*
575                                  * We only want to write back data for this
576                                  * superblock, move all inodes not belonging
577                                  * to it back onto the dirty list.
578                                  */
579                                 redirty_tail(inode, wb);
580                                 continue;
581                         }
582
583                         /*
584                          * The inode belongs to a different superblock.
585                          * Bounce back to the caller to unpin this and
586                          * pin the next superblock.
587                          */
588                         break;
589                 }
590
591                 /*
592                  * Don't bother with new inodes or inodes beeing freed, first
593                  * kind does not need peridic writeout yet, and for the latter
594                  * kind writeout is handled by the freer.
595                  */
596                 spin_lock(&inode->i_lock);
597                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
598                         spin_unlock(&inode->i_lock);
599                         requeue_io(inode, wb);
600                         continue;
601                 }
602                 __iget(inode);
603                 write_chunk = writeback_chunk_size(work);
604                 wbc.nr_to_write = write_chunk;
605                 wbc.pages_skipped = 0;
606
607                 writeback_single_inode(inode, wb, &wbc);
608
609                 work->nr_pages -= write_chunk - wbc.nr_to_write;
610                 wrote += write_chunk - wbc.nr_to_write;
611                 if (!(inode->i_state & I_DIRTY))
612                         wrote++;
613                 if (wbc.pages_skipped) {
614                         /*
615                          * writeback is not making progress due to locked
616                          * buffers.  Skip this inode for now.
617                          */
618                         redirty_tail(inode, wb);
619                 }
620                 spin_unlock(&inode->i_lock);
621                 spin_unlock(&wb->list_lock);
622                 iput(inode);
623                 cond_resched();
624                 spin_lock(&wb->list_lock);
625                 /*
626                  * bail out to wb_writeback() often enough to check
627                  * background threshold and other termination conditions.
628                  */
629                 if (wrote) {
630                         if (time_is_before_jiffies(start_time + HZ / 10UL))
631                                 break;
632                         if (work->nr_pages <= 0)
633                                 break;
634                 }
635         }
636         return wrote;
637 }
638
639 static long __writeback_inodes_wb(struct bdi_writeback *wb,
640                                   struct wb_writeback_work *work)
641 {
642         unsigned long start_time = jiffies;
643         long wrote = 0;
644
645         while (!list_empty(&wb->b_io)) {
646                 struct inode *inode = wb_inode(wb->b_io.prev);
647                 struct super_block *sb = inode->i_sb;
648
649                 if (!pin_sb_for_writeback(sb)) {
650                         requeue_io(inode, wb);
651                         continue;
652                 }
653                 wrote += writeback_sb_inodes(sb, wb, work);
654                 drop_super(sb);
655
656                 /* refer to the same tests at the end of writeback_sb_inodes */
657                 if (wrote) {
658                         if (time_is_before_jiffies(start_time + HZ / 10UL))
659                                 break;
660                         if (work->nr_pages <= 0)
661                                 break;
662                 }
663         }
664         /* Leave any unwritten inodes on b_io */
665         return wrote;
666 }
667
668 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages)
669 {
670         struct wb_writeback_work work = {
671                 .nr_pages       = nr_pages,
672                 .sync_mode      = WB_SYNC_NONE,
673                 .range_cyclic   = 1,
674         };
675
676         spin_lock(&wb->list_lock);
677         if (list_empty(&wb->b_io))
678                 queue_io(wb, NULL);
679         __writeback_inodes_wb(wb, &work);
680         spin_unlock(&wb->list_lock);
681
682         return nr_pages - work.nr_pages;
683 }
684
685 static inline bool over_bground_thresh(void)
686 {
687         unsigned long background_thresh, dirty_thresh;
688
689         global_dirty_limits(&background_thresh, &dirty_thresh);
690
691         return (global_page_state(NR_FILE_DIRTY) +
692                 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
693 }
694
695 /*
696  * Explicit flushing or periodic writeback of "old" data.
697  *
698  * Define "old": the first time one of an inode's pages is dirtied, we mark the
699  * dirtying-time in the inode's address_space.  So this periodic writeback code
700  * just walks the superblock inode list, writing back any inodes which are
701  * older than a specific point in time.
702  *
703  * Try to run once per dirty_writeback_interval.  But if a writeback event
704  * takes longer than a dirty_writeback_interval interval, then leave a
705  * one-second gap.
706  *
707  * older_than_this takes precedence over nr_to_write.  So we'll only write back
708  * all dirty pages if they are all attached to "old" mappings.
709  */
710 static long wb_writeback(struct bdi_writeback *wb,
711                          struct wb_writeback_work *work)
712 {
713         long nr_pages = work->nr_pages;
714         unsigned long oldest_jif;
715         struct inode *inode;
716         long progress;
717
718         oldest_jif = jiffies;
719         work->older_than_this = &oldest_jif;
720
721         spin_lock(&wb->list_lock);
722         for (;;) {
723                 /*
724                  * Stop writeback when nr_pages has been consumed
725                  */
726                 if (work->nr_pages <= 0)
727                         break;
728
729                 /*
730                  * Background writeout and kupdate-style writeback may
731                  * run forever. Stop them if there is other work to do
732                  * so that e.g. sync can proceed. They'll be restarted
733                  * after the other works are all done.
734                  */
735                 if ((work->for_background || work->for_kupdate) &&
736                     !list_empty(&wb->bdi->work_list))
737                         break;
738
739                 /*
740                  * For background writeout, stop when we are below the
741                  * background dirty threshold
742                  */
743                 if (work->for_background && !over_bground_thresh())
744                         break;
745
746                 if (work->for_kupdate) {
747                         oldest_jif = jiffies -
748                                 msecs_to_jiffies(dirty_expire_interval * 10);
749                         work->older_than_this = &oldest_jif;
750                 }
751
752                 trace_writeback_start(wb->bdi, work);
753                 if (list_empty(&wb->b_io))
754                         queue_io(wb, work->older_than_this);
755                 if (work->sb)
756                         progress = writeback_sb_inodes(work->sb, wb, work);
757                 else
758                         progress = __writeback_inodes_wb(wb, work);
759                 trace_writeback_written(wb->bdi, work);
760
761                 /*
762                  * Did we write something? Try for more
763                  *
764                  * Dirty inodes are moved to b_io for writeback in batches.
765                  * The completion of the current batch does not necessarily
766                  * mean the overall work is done. So we keep looping as long
767                  * as made some progress on cleaning pages or inodes.
768                  */
769                 if (progress)
770                         continue;
771                 /*
772                  * No more inodes for IO, bail
773                  */
774                 if (list_empty(&wb->b_more_io))
775                         break;
776                 /*
777                  * Nothing written. Wait for some inode to
778                  * become available for writeback. Otherwise
779                  * we'll just busyloop.
780                  */
781                 if (!list_empty(&wb->b_more_io))  {
782                         trace_writeback_wait(wb->bdi, work);
783                         inode = wb_inode(wb->b_more_io.prev);
784                         spin_lock(&inode->i_lock);
785                         inode_wait_for_writeback(inode, wb);
786                         spin_unlock(&inode->i_lock);
787                 }
788         }
789         spin_unlock(&wb->list_lock);
790
791         return nr_pages - work->nr_pages;
792 }
793
794 /*
795  * Return the next wb_writeback_work struct that hasn't been processed yet.
796  */
797 static struct wb_writeback_work *
798 get_next_work_item(struct backing_dev_info *bdi)
799 {
800         struct wb_writeback_work *work = NULL;
801
802         spin_lock_bh(&bdi->wb_lock);
803         if (!list_empty(&bdi->work_list)) {
804                 work = list_entry(bdi->work_list.next,
805                                   struct wb_writeback_work, list);
806                 list_del_init(&work->list);
807         }
808         spin_unlock_bh(&bdi->wb_lock);
809         return work;
810 }
811
812 /*
813  * Add in the number of potentially dirty inodes, because each inode
814  * write can dirty pagecache in the underlying blockdev.
815  */
816 static unsigned long get_nr_dirty_pages(void)
817 {
818         return global_page_state(NR_FILE_DIRTY) +
819                 global_page_state(NR_UNSTABLE_NFS) +
820                 get_nr_dirty_inodes();
821 }
822
823 static long wb_check_background_flush(struct bdi_writeback *wb)
824 {
825         if (over_bground_thresh()) {
826
827                 struct wb_writeback_work work = {
828                         .nr_pages       = LONG_MAX,
829                         .sync_mode      = WB_SYNC_NONE,
830                         .for_background = 1,
831                         .range_cyclic   = 1,
832                 };
833
834                 return wb_writeback(wb, &work);
835         }
836
837         return 0;
838 }
839
840 static long wb_check_old_data_flush(struct bdi_writeback *wb)
841 {
842         unsigned long expired;
843         long nr_pages;
844
845         /*
846          * When set to zero, disable periodic writeback
847          */
848         if (!dirty_writeback_interval)
849                 return 0;
850
851         expired = wb->last_old_flush +
852                         msecs_to_jiffies(dirty_writeback_interval * 10);
853         if (time_before(jiffies, expired))
854                 return 0;
855
856         wb->last_old_flush = jiffies;
857         nr_pages = get_nr_dirty_pages();
858
859         if (nr_pages) {
860                 struct wb_writeback_work work = {
861                         .nr_pages       = nr_pages,
862                         .sync_mode      = WB_SYNC_NONE,
863                         .for_kupdate    = 1,
864                         .range_cyclic   = 1,
865                 };
866
867                 return wb_writeback(wb, &work);
868         }
869
870         return 0;
871 }
872
873 /*
874  * Retrieve work items and do the writeback they describe
875  */
876 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
877 {
878         struct backing_dev_info *bdi = wb->bdi;
879         struct wb_writeback_work *work;
880         long wrote = 0;
881
882         set_bit(BDI_writeback_running, &wb->bdi->state);
883         while ((work = get_next_work_item(bdi)) != NULL) {
884                 /*
885                  * Override sync mode, in case we must wait for completion
886                  * because this thread is exiting now.
887                  */
888                 if (force_wait)
889                         work->sync_mode = WB_SYNC_ALL;
890
891                 trace_writeback_exec(bdi, work);
892
893                 wrote += wb_writeback(wb, work);
894
895                 /*
896                  * Notify the caller of completion if this is a synchronous
897                  * work item, otherwise just free it.
898                  */
899                 if (work->done)
900                         complete(work->done);
901                 else
902                         kfree(work);
903         }
904
905         /*
906          * Check for periodic writeback, kupdated() style
907          */
908         wrote += wb_check_old_data_flush(wb);
909         wrote += wb_check_background_flush(wb);
910         clear_bit(BDI_writeback_running, &wb->bdi->state);
911
912         return wrote;
913 }
914
915 /*
916  * Handle writeback of dirty data for the device backed by this bdi. Also
917  * wakes up periodically and does kupdated style flushing.
918  */
919 int bdi_writeback_thread(void *data)
920 {
921         struct bdi_writeback *wb = data;
922         struct backing_dev_info *bdi = wb->bdi;
923         long pages_written;
924
925         current->flags |= PF_SWAPWRITE;
926         set_freezable();
927         wb->last_active = jiffies;
928
929         /*
930          * Our parent may run at a different priority, just set us to normal
931          */
932         set_user_nice(current, 0);
933
934         trace_writeback_thread_start(bdi);
935
936         while (!kthread_should_stop()) {
937                 /*
938                  * Remove own delayed wake-up timer, since we are already awake
939                  * and we'll take care of the preriodic write-back.
940                  */
941                 del_timer(&wb->wakeup_timer);
942
943                 pages_written = wb_do_writeback(wb, 0);
944
945                 trace_writeback_pages_written(pages_written);
946
947                 if (pages_written)
948                         wb->last_active = jiffies;
949
950                 set_current_state(TASK_INTERRUPTIBLE);
951                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
952                         __set_current_state(TASK_RUNNING);
953                         continue;
954                 }
955
956                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
957                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
958                 else {
959                         /*
960                          * We have nothing to do, so can go sleep without any
961                          * timeout and save power. When a work is queued or
962                          * something is made dirty - we will be woken up.
963                          */
964                         schedule();
965                 }
966
967                 try_to_freeze();
968         }
969
970         /* Flush any work that raced with us exiting */
971         if (!list_empty(&bdi->work_list))
972                 wb_do_writeback(wb, 1);
973
974         trace_writeback_thread_stop(bdi);
975         return 0;
976 }
977
978
979 /*
980  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
981  * the whole world.
982  */
983 void wakeup_flusher_threads(long nr_pages)
984 {
985         struct backing_dev_info *bdi;
986
987         if (!nr_pages) {
988                 nr_pages = global_page_state(NR_FILE_DIRTY) +
989                                 global_page_state(NR_UNSTABLE_NFS);
990         }
991
992         rcu_read_lock();
993         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
994                 if (!bdi_has_dirty_io(bdi))
995                         continue;
996                 __bdi_start_writeback(bdi, nr_pages, false);
997         }
998         rcu_read_unlock();
999 }
1000
1001 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1002 {
1003         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1004                 struct dentry *dentry;
1005                 const char *name = "?";
1006
1007                 dentry = d_find_alias(inode);
1008                 if (dentry) {
1009                         spin_lock(&dentry->d_lock);
1010                         name = (const char *) dentry->d_name.name;
1011                 }
1012                 printk(KERN_DEBUG
1013                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1014                        current->comm, task_pid_nr(current), inode->i_ino,
1015                        name, inode->i_sb->s_id);
1016                 if (dentry) {
1017                         spin_unlock(&dentry->d_lock);
1018                         dput(dentry);
1019                 }
1020         }
1021 }
1022
1023 /**
1024  *      __mark_inode_dirty -    internal function
1025  *      @inode: inode to mark
1026  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1027  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1028  *      mark_inode_dirty_sync.
1029  *
1030  * Put the inode on the super block's dirty list.
1031  *
1032  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1033  * dirty list only if it is hashed or if it refers to a blockdev.
1034  * If it was not hashed, it will never be added to the dirty list
1035  * even if it is later hashed, as it will have been marked dirty already.
1036  *
1037  * In short, make sure you hash any inodes _before_ you start marking
1038  * them dirty.
1039  *
1040  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1041  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1042  * the kernel-internal blockdev inode represents the dirtying time of the
1043  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1044  * page->mapping->host, so the page-dirtying time is recorded in the internal
1045  * blockdev inode.
1046  */
1047 void __mark_inode_dirty(struct inode *inode, int flags)
1048 {
1049         struct super_block *sb = inode->i_sb;
1050         struct backing_dev_info *bdi = NULL;
1051
1052         /*
1053          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1054          * dirty the inode itself
1055          */
1056         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1057                 if (sb->s_op->dirty_inode)
1058                         sb->s_op->dirty_inode(inode, flags);
1059         }
1060
1061         /*
1062          * make sure that changes are seen by all cpus before we test i_state
1063          * -- mikulas
1064          */
1065         smp_mb();
1066
1067         /* avoid the locking if we can */
1068         if ((inode->i_state & flags) == flags)
1069                 return;
1070
1071         if (unlikely(block_dump))
1072                 block_dump___mark_inode_dirty(inode);
1073
1074         spin_lock(&inode->i_lock);
1075         if ((inode->i_state & flags) != flags) {
1076                 const int was_dirty = inode->i_state & I_DIRTY;
1077
1078                 inode->i_state |= flags;
1079
1080                 /*
1081                  * If the inode is being synced, just update its dirty state.
1082                  * The unlocker will place the inode on the appropriate
1083                  * superblock list, based upon its state.
1084                  */
1085                 if (inode->i_state & I_SYNC)
1086                         goto out_unlock_inode;
1087
1088                 /*
1089                  * Only add valid (hashed) inodes to the superblock's
1090                  * dirty list.  Add blockdev inodes as well.
1091                  */
1092                 if (!S_ISBLK(inode->i_mode)) {
1093                         if (inode_unhashed(inode))
1094                                 goto out_unlock_inode;
1095                 }
1096                 if (inode->i_state & I_FREEING)
1097                         goto out_unlock_inode;
1098
1099                 /*
1100                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1101                  * reposition it (that would break b_dirty time-ordering).
1102                  */
1103                 if (!was_dirty) {
1104                         bool wakeup_bdi = false;
1105                         bdi = inode_to_bdi(inode);
1106
1107                         if (bdi_cap_writeback_dirty(bdi)) {
1108                                 WARN(!test_bit(BDI_registered, &bdi->state),
1109                                      "bdi-%s not registered\n", bdi->name);
1110
1111                                 /*
1112                                  * If this is the first dirty inode for this
1113                                  * bdi, we have to wake-up the corresponding
1114                                  * bdi thread to make sure background
1115                                  * write-back happens later.
1116                                  */
1117                                 if (!wb_has_dirty_io(&bdi->wb))
1118                                         wakeup_bdi = true;
1119                         }
1120
1121                         spin_unlock(&inode->i_lock);
1122                         spin_lock(&bdi->wb.list_lock);
1123                         inode->dirtied_when = jiffies;
1124                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1125                         spin_unlock(&bdi->wb.list_lock);
1126
1127                         if (wakeup_bdi)
1128                                 bdi_wakeup_thread_delayed(bdi);
1129                         return;
1130                 }
1131         }
1132 out_unlock_inode:
1133         spin_unlock(&inode->i_lock);
1134
1135 }
1136 EXPORT_SYMBOL(__mark_inode_dirty);
1137
1138 /*
1139  * Write out a superblock's list of dirty inodes.  A wait will be performed
1140  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1141  *
1142  * If older_than_this is non-NULL, then only write out inodes which
1143  * had their first dirtying at a time earlier than *older_than_this.
1144  *
1145  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1146  * This function assumes that the blockdev superblock's inodes are backed by
1147  * a variety of queues, so all inodes are searched.  For other superblocks,
1148  * assume that all inodes are backed by the same queue.
1149  *
1150  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1151  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1152  * on the writer throttling path, and we get decent balancing between many
1153  * throttled threads: we don't want them all piling up on inode_sync_wait.
1154  */
1155 static void wait_sb_inodes(struct super_block *sb)
1156 {
1157         struct inode *inode, *old_inode = NULL;
1158
1159         /*
1160          * We need to be protected against the filesystem going from
1161          * r/o to r/w or vice versa.
1162          */
1163         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1164
1165         spin_lock(&inode_sb_list_lock);
1166
1167         /*
1168          * Data integrity sync. Must wait for all pages under writeback,
1169          * because there may have been pages dirtied before our sync
1170          * call, but which had writeout started before we write it out.
1171          * In which case, the inode may not be on the dirty list, but
1172          * we still have to wait for that writeout.
1173          */
1174         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1175                 struct address_space *mapping = inode->i_mapping;
1176
1177                 spin_lock(&inode->i_lock);
1178                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1179                     (mapping->nrpages == 0)) {
1180                         spin_unlock(&inode->i_lock);
1181                         continue;
1182                 }
1183                 __iget(inode);
1184                 spin_unlock(&inode->i_lock);
1185                 spin_unlock(&inode_sb_list_lock);
1186
1187                 /*
1188                  * We hold a reference to 'inode' so it couldn't have been
1189                  * removed from s_inodes list while we dropped the
1190                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1191                  * be holding the last reference and we cannot iput it under
1192                  * inode_sb_list_lock. So we keep the reference and iput it
1193                  * later.
1194                  */
1195                 iput(old_inode);
1196                 old_inode = inode;
1197
1198                 filemap_fdatawait(mapping);
1199
1200                 cond_resched();
1201
1202                 spin_lock(&inode_sb_list_lock);
1203         }
1204         spin_unlock(&inode_sb_list_lock);
1205         iput(old_inode);
1206 }
1207
1208 /**
1209  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1210  * @sb: the superblock
1211  * @nr: the number of pages to write
1212  *
1213  * Start writeback on some inodes on this super_block. No guarantees are made
1214  * on how many (if any) will be written, and this function does not wait
1215  * for IO completion of submitted IO.
1216  */
1217 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1218 {
1219         DECLARE_COMPLETION_ONSTACK(done);
1220         struct wb_writeback_work work = {
1221                 .sb                     = sb,
1222                 .sync_mode              = WB_SYNC_NONE,
1223                 .tagged_writepages      = 1,
1224                 .done                   = &done,
1225                 .nr_pages               = nr,
1226         };
1227
1228         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1229         bdi_queue_work(sb->s_bdi, &work);
1230         wait_for_completion(&done);
1231 }
1232 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1233
1234 /**
1235  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1236  * @sb: the superblock
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)
1243 {
1244         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
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  *
1252  * Invoke writeback_inodes_sb if no writeback is currently underway.
1253  * Returns 1 if writeback was started, 0 if not.
1254  */
1255 int writeback_inodes_sb_if_idle(struct super_block *sb)
1256 {
1257         if (!writeback_in_progress(sb->s_bdi)) {
1258                 down_read(&sb->s_umount);
1259                 writeback_inodes_sb(sb);
1260                 up_read(&sb->s_umount);
1261                 return 1;
1262         } else
1263                 return 0;
1264 }
1265 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1266
1267 /**
1268  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1269  * @sb: the superblock
1270  * @nr: the number of pages to write
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 {
1278         if (!writeback_in_progress(sb->s_bdi)) {
1279                 down_read(&sb->s_umount);
1280                 writeback_inodes_sb_nr(sb, nr);
1281                 up_read(&sb->s_umount);
1282                 return 1;
1283         } else
1284                 return 0;
1285 }
1286 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1287
1288 /**
1289  * sync_inodes_sb       -       sync sb inode pages
1290  * @sb: the superblock
1291  *
1292  * This function writes and waits on any dirty inode belonging to this
1293  * super_block.
1294  */
1295 void sync_inodes_sb(struct super_block *sb)
1296 {
1297         DECLARE_COMPLETION_ONSTACK(done);
1298         struct wb_writeback_work work = {
1299                 .sb             = sb,
1300                 .sync_mode      = WB_SYNC_ALL,
1301                 .nr_pages       = LONG_MAX,
1302                 .range_cyclic   = 0,
1303                 .done           = &done,
1304         };
1305
1306         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1307
1308         bdi_queue_work(sb->s_bdi, &work);
1309         wait_for_completion(&done);
1310
1311         wait_sb_inodes(sb);
1312 }
1313 EXPORT_SYMBOL(sync_inodes_sb);
1314
1315 /**
1316  * write_inode_now      -       write an inode to disk
1317  * @inode: inode to write to disk
1318  * @sync: whether the write should be synchronous or not
1319  *
1320  * This function commits an inode to disk immediately if it is dirty. This is
1321  * primarily needed by knfsd.
1322  *
1323  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1324  */
1325 int write_inode_now(struct inode *inode, int sync)
1326 {
1327         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1328         int ret;
1329         struct writeback_control wbc = {
1330                 .nr_to_write = LONG_MAX,
1331                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1332                 .range_start = 0,
1333                 .range_end = LLONG_MAX,
1334         };
1335
1336         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1337                 wbc.nr_to_write = 0;
1338
1339         might_sleep();
1340         spin_lock(&wb->list_lock);
1341         spin_lock(&inode->i_lock);
1342         ret = writeback_single_inode(inode, wb, &wbc);
1343         spin_unlock(&inode->i_lock);
1344         spin_unlock(&wb->list_lock);
1345         if (sync)
1346                 inode_sync_wait(inode);
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);