slub: Fix full list corruption if debugging is on
[linux-2.6.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * Passed into wb_writeback(), essentially a subset of writeback_control
34  */
35 struct wb_writeback_work {
36         long nr_pages;
37         struct super_block *sb;
38         unsigned long *older_than_this;
39         enum writeback_sync_modes sync_mode;
40         unsigned int tagged_writepages:1;
41         unsigned int for_kupdate:1;
42         unsigned int range_cyclic:1;
43         unsigned int for_background:1;
44
45         struct list_head list;          /* pending work list */
46         struct completion *done;        /* set if the caller waits */
47 };
48
49 /*
50  * Include the creation of the trace points after defining the
51  * wb_writeback_work structure so that the definition remains local to this
52  * file.
53  */
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/writeback.h>
56
57 /*
58  * We don't actually have pdflush, but this one is exported though /proc...
59  */
60 int nr_pdflush_threads;
61
62 /**
63  * writeback_in_progress - determine whether there is writeback in progress
64  * @bdi: the device's backing_dev_info structure.
65  *
66  * Determine whether there is writeback waiting to be handled against a
67  * backing device.
68  */
69 int writeback_in_progress(struct backing_dev_info *bdi)
70 {
71         return test_bit(BDI_writeback_running, &bdi->state);
72 }
73
74 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
75 {
76         struct super_block *sb = inode->i_sb;
77
78         if (strcmp(sb->s_type->name, "bdev") == 0)
79                 return inode->i_mapping->backing_dev_info;
80
81         return sb->s_bdi;
82 }
83
84 static inline struct inode *wb_inode(struct list_head *head)
85 {
86         return list_entry(head, struct inode, i_wb_list);
87 }
88
89 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
90 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
91 {
92         if (bdi->wb.task) {
93                 wake_up_process(bdi->wb.task);
94         } else {
95                 /*
96                  * The bdi thread isn't there, wake up the forker thread which
97                  * will create and run it.
98                  */
99                 wake_up_process(default_backing_dev_info.wb.task);
100         }
101 }
102
103 static void bdi_queue_work(struct backing_dev_info *bdi,
104                            struct wb_writeback_work *work)
105 {
106         trace_writeback_queue(bdi, work);
107
108         spin_lock_bh(&bdi->wb_lock);
109         list_add_tail(&work->list, &bdi->work_list);
110         if (!bdi->wb.task)
111                 trace_writeback_nothread(bdi, work);
112         bdi_wakeup_flusher(bdi);
113         spin_unlock_bh(&bdi->wb_lock);
114 }
115
116 static void
117 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
118                       bool range_cyclic)
119 {
120         struct wb_writeback_work *work;
121
122         /*
123          * This is WB_SYNC_NONE writeback, so if allocation fails just
124          * wakeup the thread for old dirty data writeback
125          */
126         work = kzalloc(sizeof(*work), GFP_ATOMIC);
127         if (!work) {
128                 if (bdi->wb.task) {
129                         trace_writeback_nowork(bdi);
130                         wake_up_process(bdi->wb.task);
131                 }
132                 return;
133         }
134
135         work->sync_mode = WB_SYNC_NONE;
136         work->nr_pages  = nr_pages;
137         work->range_cyclic = range_cyclic;
138
139         bdi_queue_work(bdi, work);
140 }
141
142 /**
143  * bdi_start_writeback - start writeback
144  * @bdi: the backing device to write from
145  * @nr_pages: the number of pages to write
146  *
147  * Description:
148  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
149  *   started when this function returns, we make no guarantees on
150  *   completion. Caller need not hold sb s_umount semaphore.
151  *
152  */
153 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
154 {
155         __bdi_start_writeback(bdi, nr_pages, true);
156 }
157
158 /**
159  * bdi_start_background_writeback - start background writeback
160  * @bdi: the backing device to write from
161  *
162  * Description:
163  *   This makes sure WB_SYNC_NONE background writeback happens. When
164  *   this function returns, it is only guaranteed that for given BDI
165  *   some IO is happening if we are over background dirty threshold.
166  *   Caller need not hold sb s_umount semaphore.
167  */
168 void bdi_start_background_writeback(struct backing_dev_info *bdi)
169 {
170         /*
171          * We just wake up the flusher thread. It will perform background
172          * writeback as soon as there is no other work to do.
173          */
174         trace_writeback_wake_background(bdi);
175         spin_lock_bh(&bdi->wb_lock);
176         bdi_wakeup_flusher(bdi);
177         spin_unlock_bh(&bdi->wb_lock);
178 }
179
180 /*
181  * Remove the inode from the writeback list it is on.
182  */
183 void inode_wb_list_del(struct inode *inode)
184 {
185         struct backing_dev_info *bdi = inode_to_bdi(inode);
186
187         spin_lock(&bdi->wb.list_lock);
188         list_del_init(&inode->i_wb_list);
189         spin_unlock(&bdi->wb.list_lock);
190 }
191
192 /*
193  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
194  * furthest end of its superblock's dirty-inode list.
195  *
196  * Before stamping the inode's ->dirtied_when, we check to see whether it is
197  * already the most-recently-dirtied inode on the b_dirty list.  If that is
198  * the case then the inode must have been redirtied while it was being written
199  * out and we don't reset its dirtied_when.
200  */
201 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
202 {
203         assert_spin_locked(&wb->list_lock);
204         if (!list_empty(&wb->b_dirty)) {
205                 struct inode *tail;
206
207                 tail = wb_inode(wb->b_dirty.next);
208                 if (time_before(inode->dirtied_when, tail->dirtied_when))
209                         inode->dirtied_when = jiffies;
210         }
211         list_move(&inode->i_wb_list, &wb->b_dirty);
212 }
213
214 /*
215  * requeue inode for re-scanning after bdi->b_io list is exhausted.
216  */
217 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
218 {
219         assert_spin_locked(&wb->list_lock);
220         list_move(&inode->i_wb_list, &wb->b_more_io);
221 }
222
223 static void inode_sync_complete(struct inode *inode)
224 {
225         /*
226          * Prevent speculative execution through
227          * spin_unlock(&wb->list_lock);
228          */
229
230         smp_mb();
231         wake_up_bit(&inode->i_state, __I_SYNC);
232 }
233
234 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
235 {
236         bool ret = time_after(inode->dirtied_when, t);
237 #ifndef CONFIG_64BIT
238         /*
239          * For inodes being constantly redirtied, dirtied_when can get stuck.
240          * It _appears_ to be in the future, but is actually in distant past.
241          * This test is necessary to prevent such wrapped-around relative times
242          * from permanently stopping the whole bdi writeback.
243          */
244         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
245 #endif
246         return ret;
247 }
248
249 /*
250  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
251  */
252 static int move_expired_inodes(struct list_head *delaying_queue,
253                                struct list_head *dispatch_queue,
254                                unsigned long *older_than_this)
255 {
256         LIST_HEAD(tmp);
257         struct list_head *pos, *node;
258         struct super_block *sb = NULL;
259         struct inode *inode;
260         int do_sb_sort = 0;
261         int moved = 0;
262
263         while (!list_empty(delaying_queue)) {
264                 inode = wb_inode(delaying_queue->prev);
265                 if (older_than_this &&
266                     inode_dirtied_after(inode, *older_than_this))
267                         break;
268                 if (sb && sb != inode->i_sb)
269                         do_sb_sort = 1;
270                 sb = inode->i_sb;
271                 list_move(&inode->i_wb_list, &tmp);
272                 moved++;
273         }
274
275         /* just one sb in list, splice to dispatch_queue and we're done */
276         if (!do_sb_sort) {
277                 list_splice(&tmp, dispatch_queue);
278                 goto out;
279         }
280
281         /* Move inodes from one superblock together */
282         while (!list_empty(&tmp)) {
283                 sb = wb_inode(tmp.prev)->i_sb;
284                 list_for_each_prev_safe(pos, node, &tmp) {
285                         inode = wb_inode(pos);
286                         if (inode->i_sb == sb)
287                                 list_move(&inode->i_wb_list, dispatch_queue);
288                 }
289         }
290 out:
291         return moved;
292 }
293
294 /*
295  * Queue all expired dirty inodes for io, eldest first.
296  * Before
297  *         newly dirtied     b_dirty    b_io    b_more_io
298  *         =============>    gf         edc     BA
299  * After
300  *         newly dirtied     b_dirty    b_io    b_more_io
301  *         =============>    g          fBAedc
302  *                                           |
303  *                                           +--> dequeue for IO
304  */
305 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
306 {
307         int moved;
308         assert_spin_locked(&wb->list_lock);
309         list_splice_init(&wb->b_more_io, &wb->b_io);
310         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
311         trace_writeback_queue_io(wb, older_than_this, moved);
312 }
313
314 static int write_inode(struct inode *inode, struct writeback_control *wbc)
315 {
316         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
317                 return inode->i_sb->s_op->write_inode(inode, wbc);
318         return 0;
319 }
320
321 /*
322  * Wait for writeback on an inode to complete.
323  */
324 static void inode_wait_for_writeback(struct inode *inode,
325                                      struct bdi_writeback *wb)
326 {
327         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
328         wait_queue_head_t *wqh;
329
330         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
331         while (inode->i_state & I_SYNC) {
332                 spin_unlock(&inode->i_lock);
333                 spin_unlock(&wb->list_lock);
334                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
335                 spin_lock(&wb->list_lock);
336                 spin_lock(&inode->i_lock);
337         }
338 }
339
340 /*
341  * Write out an inode's dirty pages.  Called under wb->list_lock and
342  * inode->i_lock.  Either the caller has an active reference on the inode or
343  * the inode has I_WILL_FREE set.
344  *
345  * If `wait' is set, wait on the writeout.
346  *
347  * The whole writeout design is quite complex and fragile.  We want to avoid
348  * starvation of particular inodes when others are being redirtied, prevent
349  * livelocks, etc.
350  */
351 static int
352 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
353                        struct writeback_control *wbc)
354 {
355         struct address_space *mapping = inode->i_mapping;
356         long nr_to_write = wbc->nr_to_write;
357         unsigned dirty;
358         int ret;
359
360         assert_spin_locked(&wb->list_lock);
361         assert_spin_locked(&inode->i_lock);
362
363         if (!atomic_read(&inode->i_count))
364                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
365         else
366                 WARN_ON(inode->i_state & I_WILL_FREE);
367
368         if (inode->i_state & I_SYNC) {
369                 /*
370                  * If this inode is locked for writeback and we are not doing
371                  * writeback-for-data-integrity, move it to b_more_io so that
372                  * writeback can proceed with the other inodes on s_io.
373                  *
374                  * We'll have another go at writing back this inode when we
375                  * completed a full scan of b_io.
376                  */
377                 if (wbc->sync_mode != WB_SYNC_ALL) {
378                         requeue_io(inode, wb);
379                         trace_writeback_single_inode_requeue(inode, wbc,
380                                                              nr_to_write);
381                         return 0;
382                 }
383
384                 /*
385                  * It's a data-integrity sync.  We must wait.
386                  */
387                 inode_wait_for_writeback(inode, wb);
388         }
389
390         BUG_ON(inode->i_state & I_SYNC);
391
392         /* Set I_SYNC, reset I_DIRTY_PAGES */
393         inode->i_state |= I_SYNC;
394         inode->i_state &= ~I_DIRTY_PAGES;
395         spin_unlock(&inode->i_lock);
396         spin_unlock(&wb->list_lock);
397
398         ret = do_writepages(mapping, wbc);
399
400         /*
401          * Make sure to wait on the data before writing out the metadata.
402          * This is important for filesystems that modify metadata on data
403          * I/O completion.
404          */
405         if (wbc->sync_mode == WB_SYNC_ALL) {
406                 int err = filemap_fdatawait(mapping);
407                 if (ret == 0)
408                         ret = err;
409         }
410
411         /*
412          * Some filesystems may redirty the inode during the writeback
413          * due to delalloc, clear dirty metadata flags right before
414          * write_inode()
415          */
416         spin_lock(&inode->i_lock);
417         dirty = inode->i_state & I_DIRTY;
418         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
419         spin_unlock(&inode->i_lock);
420         /* Don't write the inode if only I_DIRTY_PAGES was set */
421         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
422                 int err = write_inode(inode, wbc);
423                 if (ret == 0)
424                         ret = err;
425         }
426
427         spin_lock(&wb->list_lock);
428         spin_lock(&inode->i_lock);
429         inode->i_state &= ~I_SYNC;
430         if (!(inode->i_state & I_FREEING)) {
431                 /*
432                  * Sync livelock prevention. Each inode is tagged and synced in
433                  * one shot. If still dirty, it will be redirty_tail()'ed below.
434                  * Update the dirty time to prevent enqueue and sync it again.
435                  */
436                 if ((inode->i_state & I_DIRTY) &&
437                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
438                         inode->dirtied_when = jiffies;
439
440                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
441                         /*
442                          * We didn't write back all the pages.  nfs_writepages()
443                          * sometimes bales out without doing anything.
444                          */
445                         inode->i_state |= I_DIRTY_PAGES;
446                         if (wbc->nr_to_write <= 0) {
447                                 /*
448                                  * slice used up: queue for next turn
449                                  */
450                                 requeue_io(inode, wb);
451                         } else {
452                                 /*
453                                  * Writeback blocked by something other than
454                                  * congestion. Delay the inode for some time to
455                                  * avoid spinning on the CPU (100% iowait)
456                                  * retrying writeback of the dirty page/inode
457                                  * that cannot be performed immediately.
458                                  */
459                                 redirty_tail(inode, wb);
460                         }
461                 } else if (inode->i_state & I_DIRTY) {
462                         /*
463                          * Filesystems can dirty the inode during writeback
464                          * operations, such as delayed allocation during
465                          * submission or metadata updates after data IO
466                          * completion.
467                          */
468                         redirty_tail(inode, wb);
469                 } else {
470                         /*
471                          * The inode is clean.  At this point we either have
472                          * a reference to the inode or it's on it's way out.
473                          * No need to add it back to the LRU.
474                          */
475                         list_del_init(&inode->i_wb_list);
476                 }
477         }
478         inode_sync_complete(inode);
479         trace_writeback_single_inode(inode, wbc, nr_to_write);
480         return ret;
481 }
482
483 static long writeback_chunk_size(struct backing_dev_info *bdi,
484                                  struct wb_writeback_work *work)
485 {
486         long pages;
487
488         /*
489          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
490          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
491          * here avoids calling into writeback_inodes_wb() more than once.
492          *
493          * The intended call sequence for WB_SYNC_ALL writeback is:
494          *
495          *      wb_writeback()
496          *          writeback_sb_inodes()       <== called only once
497          *              write_cache_pages()     <== called once for each inode
498          *                   (quickly) tag currently dirty pages
499          *                   (maybe slowly) sync all tagged pages
500          */
501         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
502                 pages = LONG_MAX;
503         else {
504                 pages = min(bdi->avg_write_bandwidth / 2,
505                             global_dirty_limit / DIRTY_SCOPE);
506                 pages = min(pages, work->nr_pages);
507                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
508                                    MIN_WRITEBACK_PAGES);
509         }
510
511         return pages;
512 }
513
514 /*
515  * Write a portion of b_io inodes which belong to @sb.
516  *
517  * If @only_this_sb is true, then find and write all such
518  * inodes. Otherwise write only ones which go sequentially
519  * in reverse order.
520  *
521  * Return the number of pages and/or inodes written.
522  */
523 static long writeback_sb_inodes(struct super_block *sb,
524                                 struct bdi_writeback *wb,
525                                 struct wb_writeback_work *work)
526 {
527         struct writeback_control wbc = {
528                 .sync_mode              = work->sync_mode,
529                 .tagged_writepages      = work->tagged_writepages,
530                 .for_kupdate            = work->for_kupdate,
531                 .for_background         = work->for_background,
532                 .range_cyclic           = work->range_cyclic,
533                 .range_start            = 0,
534                 .range_end              = LLONG_MAX,
535         };
536         unsigned long start_time = jiffies;
537         long write_chunk;
538         long wrote = 0;  /* count both pages and inodes */
539
540         while (!list_empty(&wb->b_io)) {
541                 struct inode *inode = wb_inode(wb->b_io.prev);
542
543                 if (inode->i_sb != sb) {
544                         if (work->sb) {
545                                 /*
546                                  * We only want to write back data for this
547                                  * superblock, move all inodes not belonging
548                                  * to it back onto the dirty list.
549                                  */
550                                 redirty_tail(inode, wb);
551                                 continue;
552                         }
553
554                         /*
555                          * The inode belongs to a different superblock.
556                          * Bounce back to the caller to unpin this and
557                          * pin the next superblock.
558                          */
559                         break;
560                 }
561
562                 /*
563                  * Don't bother with new inodes or inodes beeing freed, first
564                  * kind does not need peridic writeout yet, and for the latter
565                  * kind writeout is handled by the freer.
566                  */
567                 spin_lock(&inode->i_lock);
568                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
569                         spin_unlock(&inode->i_lock);
570                         redirty_tail(inode, wb);
571                         continue;
572                 }
573                 __iget(inode);
574                 write_chunk = writeback_chunk_size(wb->bdi, work);
575                 wbc.nr_to_write = write_chunk;
576                 wbc.pages_skipped = 0;
577
578                 writeback_single_inode(inode, wb, &wbc);
579
580                 work->nr_pages -= write_chunk - wbc.nr_to_write;
581                 wrote += write_chunk - wbc.nr_to_write;
582                 if (!(inode->i_state & I_DIRTY))
583                         wrote++;
584                 if (wbc.pages_skipped) {
585                         /*
586                          * writeback is not making progress due to locked
587                          * buffers.  Skip this inode for now.
588                          */
589                         redirty_tail(inode, wb);
590                 }
591                 spin_unlock(&inode->i_lock);
592                 spin_unlock(&wb->list_lock);
593                 iput(inode);
594                 cond_resched();
595                 spin_lock(&wb->list_lock);
596                 /*
597                  * bail out to wb_writeback() often enough to check
598                  * background threshold and other termination conditions.
599                  */
600                 if (wrote) {
601                         if (time_is_before_jiffies(start_time + HZ / 10UL))
602                                 break;
603                         if (work->nr_pages <= 0)
604                                 break;
605                 }
606         }
607         return wrote;
608 }
609
610 static long __writeback_inodes_wb(struct bdi_writeback *wb,
611                                   struct wb_writeback_work *work)
612 {
613         unsigned long start_time = jiffies;
614         long wrote = 0;
615
616         while (!list_empty(&wb->b_io)) {
617                 struct inode *inode = wb_inode(wb->b_io.prev);
618                 struct super_block *sb = inode->i_sb;
619
620                 if (!grab_super_passive(sb)) {
621                         /*
622                          * grab_super_passive() may fail consistently due to
623                          * s_umount being grabbed by someone else. Don't use
624                          * requeue_io() to avoid busy retrying the inode/sb.
625                          */
626                         redirty_tail(inode, wb);
627                         continue;
628                 }
629                 wrote += writeback_sb_inodes(sb, wb, work);
630                 drop_super(sb);
631
632                 /* refer to the same tests at the end of writeback_sb_inodes */
633                 if (wrote) {
634                         if (time_is_before_jiffies(start_time + HZ / 10UL))
635                                 break;
636                         if (work->nr_pages <= 0)
637                                 break;
638                 }
639         }
640         /* Leave any unwritten inodes on b_io */
641         return wrote;
642 }
643
644 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages)
645 {
646         struct wb_writeback_work work = {
647                 .nr_pages       = nr_pages,
648                 .sync_mode      = WB_SYNC_NONE,
649                 .range_cyclic   = 1,
650         };
651
652         spin_lock(&wb->list_lock);
653         if (list_empty(&wb->b_io))
654                 queue_io(wb, NULL);
655         __writeback_inodes_wb(wb, &work);
656         spin_unlock(&wb->list_lock);
657
658         return nr_pages - work.nr_pages;
659 }
660
661 static inline bool over_bground_thresh(void)
662 {
663         unsigned long background_thresh, dirty_thresh;
664
665         global_dirty_limits(&background_thresh, &dirty_thresh);
666
667         return (global_page_state(NR_FILE_DIRTY) +
668                 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
669 }
670
671 /*
672  * Called under wb->list_lock. If there are multiple wb per bdi,
673  * only the flusher working on the first wb should do it.
674  */
675 static void wb_update_bandwidth(struct bdi_writeback *wb,
676                                 unsigned long start_time)
677 {
678         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, start_time);
679 }
680
681 /*
682  * Explicit flushing or periodic writeback of "old" data.
683  *
684  * Define "old": the first time one of an inode's pages is dirtied, we mark the
685  * dirtying-time in the inode's address_space.  So this periodic writeback code
686  * just walks the superblock inode list, writing back any inodes which are
687  * older than a specific point in time.
688  *
689  * Try to run once per dirty_writeback_interval.  But if a writeback event
690  * takes longer than a dirty_writeback_interval interval, then leave a
691  * one-second gap.
692  *
693  * older_than_this takes precedence over nr_to_write.  So we'll only write back
694  * all dirty pages if they are all attached to "old" mappings.
695  */
696 static long wb_writeback(struct bdi_writeback *wb,
697                          struct wb_writeback_work *work)
698 {
699         unsigned long wb_start = jiffies;
700         long nr_pages = work->nr_pages;
701         unsigned long oldest_jif;
702         struct inode *inode;
703         long progress;
704
705         oldest_jif = jiffies;
706         work->older_than_this = &oldest_jif;
707
708         spin_lock(&wb->list_lock);
709         for (;;) {
710                 /*
711                  * Stop writeback when nr_pages has been consumed
712                  */
713                 if (work->nr_pages <= 0)
714                         break;
715
716                 /*
717                  * Background writeout and kupdate-style writeback may
718                  * run forever. Stop them if there is other work to do
719                  * so that e.g. sync can proceed. They'll be restarted
720                  * after the other works are all done.
721                  */
722                 if ((work->for_background || work->for_kupdate) &&
723                     !list_empty(&wb->bdi->work_list))
724                         break;
725
726                 /*
727                  * For background writeout, stop when we are below the
728                  * background dirty threshold
729                  */
730                 if (work->for_background && !over_bground_thresh())
731                         break;
732
733                 if (work->for_kupdate) {
734                         oldest_jif = jiffies -
735                                 msecs_to_jiffies(dirty_expire_interval * 10);
736                         work->older_than_this = &oldest_jif;
737                 }
738
739                 trace_writeback_start(wb->bdi, work);
740                 if (list_empty(&wb->b_io))
741                         queue_io(wb, work->older_than_this);
742                 if (work->sb)
743                         progress = writeback_sb_inodes(work->sb, wb, work);
744                 else
745                         progress = __writeback_inodes_wb(wb, work);
746                 trace_writeback_written(wb->bdi, work);
747
748                 wb_update_bandwidth(wb, wb_start);
749
750                 /*
751                  * Did we write something? Try for more
752                  *
753                  * Dirty inodes are moved to b_io for writeback in batches.
754                  * The completion of the current batch does not necessarily
755                  * mean the overall work is done. So we keep looping as long
756                  * as made some progress on cleaning pages or inodes.
757                  */
758                 if (progress)
759                         continue;
760                 /*
761                  * No more inodes for IO, bail
762                  */
763                 if (list_empty(&wb->b_more_io))
764                         break;
765                 /*
766                  * Nothing written. Wait for some inode to
767                  * become available for writeback. Otherwise
768                  * we'll just busyloop.
769                  */
770                 if (!list_empty(&wb->b_more_io))  {
771                         trace_writeback_wait(wb->bdi, work);
772                         inode = wb_inode(wb->b_more_io.prev);
773                         spin_lock(&inode->i_lock);
774                         inode_wait_for_writeback(inode, wb);
775                         spin_unlock(&inode->i_lock);
776                 }
777         }
778         spin_unlock(&wb->list_lock);
779
780         return nr_pages - work->nr_pages;
781 }
782
783 /*
784  * Return the next wb_writeback_work struct that hasn't been processed yet.
785  */
786 static struct wb_writeback_work *
787 get_next_work_item(struct backing_dev_info *bdi)
788 {
789         struct wb_writeback_work *work = NULL;
790
791         spin_lock_bh(&bdi->wb_lock);
792         if (!list_empty(&bdi->work_list)) {
793                 work = list_entry(bdi->work_list.next,
794                                   struct wb_writeback_work, list);
795                 list_del_init(&work->list);
796         }
797         spin_unlock_bh(&bdi->wb_lock);
798         return work;
799 }
800
801 /*
802  * Add in the number of potentially dirty inodes, because each inode
803  * write can dirty pagecache in the underlying blockdev.
804  */
805 static unsigned long get_nr_dirty_pages(void)
806 {
807         return global_page_state(NR_FILE_DIRTY) +
808                 global_page_state(NR_UNSTABLE_NFS) +
809                 get_nr_dirty_inodes();
810 }
811
812 static long wb_check_background_flush(struct bdi_writeback *wb)
813 {
814         if (over_bground_thresh()) {
815
816                 struct wb_writeback_work work = {
817                         .nr_pages       = LONG_MAX,
818                         .sync_mode      = WB_SYNC_NONE,
819                         .for_background = 1,
820                         .range_cyclic   = 1,
821                 };
822
823                 return wb_writeback(wb, &work);
824         }
825
826         return 0;
827 }
828
829 static long wb_check_old_data_flush(struct bdi_writeback *wb)
830 {
831         unsigned long expired;
832         long nr_pages;
833
834         /*
835          * When set to zero, disable periodic writeback
836          */
837         if (!dirty_writeback_interval)
838                 return 0;
839
840         expired = wb->last_old_flush +
841                         msecs_to_jiffies(dirty_writeback_interval * 10);
842         if (time_before(jiffies, expired))
843                 return 0;
844
845         wb->last_old_flush = jiffies;
846         nr_pages = get_nr_dirty_pages();
847
848         if (nr_pages) {
849                 struct wb_writeback_work work = {
850                         .nr_pages       = nr_pages,
851                         .sync_mode      = WB_SYNC_NONE,
852                         .for_kupdate    = 1,
853                         .range_cyclic   = 1,
854                 };
855
856                 return wb_writeback(wb, &work);
857         }
858
859         return 0;
860 }
861
862 /*
863  * Retrieve work items and do the writeback they describe
864  */
865 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
866 {
867         struct backing_dev_info *bdi = wb->bdi;
868         struct wb_writeback_work *work;
869         long wrote = 0;
870
871         set_bit(BDI_writeback_running, &wb->bdi->state);
872         while ((work = get_next_work_item(bdi)) != NULL) {
873                 /*
874                  * Override sync mode, in case we must wait for completion
875                  * because this thread is exiting now.
876                  */
877                 if (force_wait)
878                         work->sync_mode = WB_SYNC_ALL;
879
880                 trace_writeback_exec(bdi, work);
881
882                 wrote += wb_writeback(wb, work);
883
884                 /*
885                  * Notify the caller of completion if this is a synchronous
886                  * work item, otherwise just free it.
887                  */
888                 if (work->done)
889                         complete(work->done);
890                 else
891                         kfree(work);
892         }
893
894         /*
895          * Check for periodic writeback, kupdated() style
896          */
897         wrote += wb_check_old_data_flush(wb);
898         wrote += wb_check_background_flush(wb);
899         clear_bit(BDI_writeback_running, &wb->bdi->state);
900
901         return wrote;
902 }
903
904 /*
905  * Handle writeback of dirty data for the device backed by this bdi. Also
906  * wakes up periodically and does kupdated style flushing.
907  */
908 int bdi_writeback_thread(void *data)
909 {
910         struct bdi_writeback *wb = data;
911         struct backing_dev_info *bdi = wb->bdi;
912         long pages_written;
913
914         current->flags |= PF_SWAPWRITE;
915         set_freezable();
916         wb->last_active = jiffies;
917
918         /*
919          * Our parent may run at a different priority, just set us to normal
920          */
921         set_user_nice(current, 0);
922
923         trace_writeback_thread_start(bdi);
924
925         while (!kthread_should_stop()) {
926                 /*
927                  * Remove own delayed wake-up timer, since we are already awake
928                  * and we'll take care of the preriodic write-back.
929                  */
930                 del_timer(&wb->wakeup_timer);
931
932                 pages_written = wb_do_writeback(wb, 0);
933
934                 trace_writeback_pages_written(pages_written);
935
936                 if (pages_written)
937                         wb->last_active = jiffies;
938
939                 set_current_state(TASK_INTERRUPTIBLE);
940                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
941                         __set_current_state(TASK_RUNNING);
942                         continue;
943                 }
944
945                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
946                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
947                 else {
948                         /*
949                          * We have nothing to do, so can go sleep without any
950                          * timeout and save power. When a work is queued or
951                          * something is made dirty - we will be woken up.
952                          */
953                         schedule();
954                 }
955
956                 try_to_freeze();
957         }
958
959         /* Flush any work that raced with us exiting */
960         if (!list_empty(&bdi->work_list))
961                 wb_do_writeback(wb, 1);
962
963         trace_writeback_thread_stop(bdi);
964         return 0;
965 }
966
967
968 /*
969  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
970  * the whole world.
971  */
972 void wakeup_flusher_threads(long nr_pages)
973 {
974         struct backing_dev_info *bdi;
975
976         if (!nr_pages) {
977                 nr_pages = global_page_state(NR_FILE_DIRTY) +
978                                 global_page_state(NR_UNSTABLE_NFS);
979         }
980
981         rcu_read_lock();
982         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
983                 if (!bdi_has_dirty_io(bdi))
984                         continue;
985                 __bdi_start_writeback(bdi, nr_pages, false);
986         }
987         rcu_read_unlock();
988 }
989
990 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
991 {
992         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
993                 struct dentry *dentry;
994                 const char *name = "?";
995
996                 dentry = d_find_alias(inode);
997                 if (dentry) {
998                         spin_lock(&dentry->d_lock);
999                         name = (const char *) dentry->d_name.name;
1000                 }
1001                 printk(KERN_DEBUG
1002                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1003                        current->comm, task_pid_nr(current), inode->i_ino,
1004                        name, inode->i_sb->s_id);
1005                 if (dentry) {
1006                         spin_unlock(&dentry->d_lock);
1007                         dput(dentry);
1008                 }
1009         }
1010 }
1011
1012 /**
1013  *      __mark_inode_dirty -    internal function
1014  *      @inode: inode to mark
1015  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1016  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1017  *      mark_inode_dirty_sync.
1018  *
1019  * Put the inode on the super block's dirty list.
1020  *
1021  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1022  * dirty list only if it is hashed or if it refers to a blockdev.
1023  * If it was not hashed, it will never be added to the dirty list
1024  * even if it is later hashed, as it will have been marked dirty already.
1025  *
1026  * In short, make sure you hash any inodes _before_ you start marking
1027  * them dirty.
1028  *
1029  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1030  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1031  * the kernel-internal blockdev inode represents the dirtying time of the
1032  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1033  * page->mapping->host, so the page-dirtying time is recorded in the internal
1034  * blockdev inode.
1035  */
1036 void __mark_inode_dirty(struct inode *inode, int flags)
1037 {
1038         struct super_block *sb = inode->i_sb;
1039         struct backing_dev_info *bdi = NULL;
1040
1041         /*
1042          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1043          * dirty the inode itself
1044          */
1045         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1046                 if (sb->s_op->dirty_inode)
1047                         sb->s_op->dirty_inode(inode, flags);
1048         }
1049
1050         /*
1051          * make sure that changes are seen by all cpus before we test i_state
1052          * -- mikulas
1053          */
1054         smp_mb();
1055
1056         /* avoid the locking if we can */
1057         if ((inode->i_state & flags) == flags)
1058                 return;
1059
1060         if (unlikely(block_dump))
1061                 block_dump___mark_inode_dirty(inode);
1062
1063         spin_lock(&inode->i_lock);
1064         if ((inode->i_state & flags) != flags) {
1065                 const int was_dirty = inode->i_state & I_DIRTY;
1066
1067                 inode->i_state |= flags;
1068
1069                 /*
1070                  * If the inode is being synced, just update its dirty state.
1071                  * The unlocker will place the inode on the appropriate
1072                  * superblock list, based upon its state.
1073                  */
1074                 if (inode->i_state & I_SYNC)
1075                         goto out_unlock_inode;
1076
1077                 /*
1078                  * Only add valid (hashed) inodes to the superblock's
1079                  * dirty list.  Add blockdev inodes as well.
1080                  */
1081                 if (!S_ISBLK(inode->i_mode)) {
1082                         if (inode_unhashed(inode))
1083                                 goto out_unlock_inode;
1084                 }
1085                 if (inode->i_state & I_FREEING)
1086                         goto out_unlock_inode;
1087
1088                 /*
1089                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1090                  * reposition it (that would break b_dirty time-ordering).
1091                  */
1092                 if (!was_dirty) {
1093                         bool wakeup_bdi = false;
1094                         bdi = inode_to_bdi(inode);
1095
1096                         if (bdi_cap_writeback_dirty(bdi)) {
1097                                 WARN(!test_bit(BDI_registered, &bdi->state),
1098                                      "bdi-%s not registered\n", bdi->name);
1099
1100                                 /*
1101                                  * If this is the first dirty inode for this
1102                                  * bdi, we have to wake-up the corresponding
1103                                  * bdi thread to make sure background
1104                                  * write-back happens later.
1105                                  */
1106                                 if (!wb_has_dirty_io(&bdi->wb))
1107                                         wakeup_bdi = true;
1108                         }
1109
1110                         spin_unlock(&inode->i_lock);
1111                         spin_lock(&bdi->wb.list_lock);
1112                         inode->dirtied_when = jiffies;
1113                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1114                         spin_unlock(&bdi->wb.list_lock);
1115
1116                         if (wakeup_bdi)
1117                                 bdi_wakeup_thread_delayed(bdi);
1118                         return;
1119                 }
1120         }
1121 out_unlock_inode:
1122         spin_unlock(&inode->i_lock);
1123
1124 }
1125 EXPORT_SYMBOL(__mark_inode_dirty);
1126
1127 /*
1128  * Write out a superblock's list of dirty inodes.  A wait will be performed
1129  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1130  *
1131  * If older_than_this is non-NULL, then only write out inodes which
1132  * had their first dirtying at a time earlier than *older_than_this.
1133  *
1134  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1135  * This function assumes that the blockdev superblock's inodes are backed by
1136  * a variety of queues, so all inodes are searched.  For other superblocks,
1137  * assume that all inodes are backed by the same queue.
1138  *
1139  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1140  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1141  * on the writer throttling path, and we get decent balancing between many
1142  * throttled threads: we don't want them all piling up on inode_sync_wait.
1143  */
1144 static void wait_sb_inodes(struct super_block *sb)
1145 {
1146         struct inode *inode, *old_inode = NULL;
1147
1148         /*
1149          * We need to be protected against the filesystem going from
1150          * r/o to r/w or vice versa.
1151          */
1152         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1153
1154         spin_lock(&inode_sb_list_lock);
1155
1156         /*
1157          * Data integrity sync. Must wait for all pages under writeback,
1158          * because there may have been pages dirtied before our sync
1159          * call, but which had writeout started before we write it out.
1160          * In which case, the inode may not be on the dirty list, but
1161          * we still have to wait for that writeout.
1162          */
1163         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1164                 struct address_space *mapping = inode->i_mapping;
1165
1166                 spin_lock(&inode->i_lock);
1167                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1168                     (mapping->nrpages == 0)) {
1169                         spin_unlock(&inode->i_lock);
1170                         continue;
1171                 }
1172                 __iget(inode);
1173                 spin_unlock(&inode->i_lock);
1174                 spin_unlock(&inode_sb_list_lock);
1175
1176                 /*
1177                  * We hold a reference to 'inode' so it couldn't have been
1178                  * removed from s_inodes list while we dropped the
1179                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1180                  * be holding the last reference and we cannot iput it under
1181                  * inode_sb_list_lock. So we keep the reference and iput it
1182                  * later.
1183                  */
1184                 iput(old_inode);
1185                 old_inode = inode;
1186
1187                 filemap_fdatawait(mapping);
1188
1189                 cond_resched();
1190
1191                 spin_lock(&inode_sb_list_lock);
1192         }
1193         spin_unlock(&inode_sb_list_lock);
1194         iput(old_inode);
1195 }
1196
1197 /**
1198  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1199  * @sb: the superblock
1200  * @nr: the number of pages to write
1201  *
1202  * Start writeback on some inodes on this super_block. No guarantees are made
1203  * on how many (if any) will be written, and this function does not wait
1204  * for IO completion of submitted IO.
1205  */
1206 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1207 {
1208         DECLARE_COMPLETION_ONSTACK(done);
1209         struct wb_writeback_work work = {
1210                 .sb                     = sb,
1211                 .sync_mode              = WB_SYNC_NONE,
1212                 .tagged_writepages      = 1,
1213                 .done                   = &done,
1214                 .nr_pages               = nr,
1215         };
1216
1217         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1218         bdi_queue_work(sb->s_bdi, &work);
1219         wait_for_completion(&done);
1220 }
1221 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1222
1223 /**
1224  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1225  * @sb: the superblock
1226  *
1227  * Start writeback on some inodes on this super_block. No guarantees are made
1228  * on how many (if any) will be written, and this function does not wait
1229  * for IO completion of submitted IO.
1230  */
1231 void writeback_inodes_sb(struct super_block *sb)
1232 {
1233         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1234 }
1235 EXPORT_SYMBOL(writeback_inodes_sb);
1236
1237 /**
1238  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1239  * @sb: the superblock
1240  *
1241  * Invoke writeback_inodes_sb if no writeback is currently underway.
1242  * Returns 1 if writeback was started, 0 if not.
1243  */
1244 int writeback_inodes_sb_if_idle(struct super_block *sb)
1245 {
1246         if (!writeback_in_progress(sb->s_bdi)) {
1247                 down_read(&sb->s_umount);
1248                 writeback_inodes_sb(sb);
1249                 up_read(&sb->s_umount);
1250                 return 1;
1251         } else
1252                 return 0;
1253 }
1254 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1255
1256 /**
1257  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1258  * @sb: the superblock
1259  * @nr: the number of pages to write
1260  *
1261  * Invoke writeback_inodes_sb if no writeback is currently underway.
1262  * Returns 1 if writeback was started, 0 if not.
1263  */
1264 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1265                                    unsigned long nr)
1266 {
1267         if (!writeback_in_progress(sb->s_bdi)) {
1268                 down_read(&sb->s_umount);
1269                 writeback_inodes_sb_nr(sb, nr);
1270                 up_read(&sb->s_umount);
1271                 return 1;
1272         } else
1273                 return 0;
1274 }
1275 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1276
1277 /**
1278  * sync_inodes_sb       -       sync sb inode pages
1279  * @sb: the superblock
1280  *
1281  * This function writes and waits on any dirty inode belonging to this
1282  * super_block.
1283  */
1284 void sync_inodes_sb(struct super_block *sb)
1285 {
1286         DECLARE_COMPLETION_ONSTACK(done);
1287         struct wb_writeback_work work = {
1288                 .sb             = sb,
1289                 .sync_mode      = WB_SYNC_ALL,
1290                 .nr_pages       = LONG_MAX,
1291                 .range_cyclic   = 0,
1292                 .done           = &done,
1293         };
1294
1295         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1296
1297         bdi_queue_work(sb->s_bdi, &work);
1298         wait_for_completion(&done);
1299
1300         wait_sb_inodes(sb);
1301 }
1302 EXPORT_SYMBOL(sync_inodes_sb);
1303
1304 /**
1305  * write_inode_now      -       write an inode to disk
1306  * @inode: inode to write to disk
1307  * @sync: whether the write should be synchronous or not
1308  *
1309  * This function commits an inode to disk immediately if it is dirty. This is
1310  * primarily needed by knfsd.
1311  *
1312  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1313  */
1314 int write_inode_now(struct inode *inode, int sync)
1315 {
1316         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1317         int ret;
1318         struct writeback_control wbc = {
1319                 .nr_to_write = LONG_MAX,
1320                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1321                 .range_start = 0,
1322                 .range_end = LLONG_MAX,
1323         };
1324
1325         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1326                 wbc.nr_to_write = 0;
1327
1328         might_sleep();
1329         spin_lock(&wb->list_lock);
1330         spin_lock(&inode->i_lock);
1331         ret = writeback_single_inode(inode, wb, &wbc);
1332         spin_unlock(&inode->i_lock);
1333         spin_unlock(&wb->list_lock);
1334         if (sync)
1335                 inode_sync_wait(inode);
1336         return ret;
1337 }
1338 EXPORT_SYMBOL(write_inode_now);
1339
1340 /**
1341  * sync_inode - write an inode and its pages to disk.
1342  * @inode: the inode to sync
1343  * @wbc: controls the writeback mode
1344  *
1345  * sync_inode() will write an inode and its pages to disk.  It will also
1346  * correctly update the inode on its superblock's dirty inode lists and will
1347  * update inode->i_state.
1348  *
1349  * The caller must have a ref on the inode.
1350  */
1351 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1352 {
1353         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1354         int ret;
1355
1356         spin_lock(&wb->list_lock);
1357         spin_lock(&inode->i_lock);
1358         ret = writeback_single_inode(inode, wb, wbc);
1359         spin_unlock(&inode->i_lock);
1360         spin_unlock(&wb->list_lock);
1361         return ret;
1362 }
1363 EXPORT_SYMBOL(sync_inode);
1364
1365 /**
1366  * sync_inode_metadata - write an inode to disk
1367  * @inode: the inode to sync
1368  * @wait: wait for I/O to complete.
1369  *
1370  * Write an inode to disk and adjust its dirty state after completion.
1371  *
1372  * Note: only writes the actual inode, no associated data or other metadata.
1373  */
1374 int sync_inode_metadata(struct inode *inode, int wait)
1375 {
1376         struct writeback_control wbc = {
1377                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1378                 .nr_to_write = 0, /* metadata-only */
1379         };
1380
1381         return sync_inode(inode, &wbc);
1382 }
1383 EXPORT_SYMBOL(sync_inode_metadata);