Merge tag 'writeback' of git://git.kernel.org/pub/scm/linux/kernel/git/wfg/linux
[linux-3.10.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * 4MB minimal write chunk size
34  */
35 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
36
37 /*
38  * Passed into wb_writeback(), essentially a subset of writeback_control
39  */
40 struct wb_writeback_work {
41         long nr_pages;
42         struct super_block *sb;
43         unsigned long *older_than_this;
44         enum writeback_sync_modes sync_mode;
45         unsigned int tagged_writepages:1;
46         unsigned int for_kupdate:1;
47         unsigned int range_cyclic:1;
48         unsigned int for_background:1;
49         enum wb_reason reason;          /* why was writeback initiated? */
50
51         struct list_head list;          /* pending work list */
52         struct completion *done;        /* set if the caller waits */
53 };
54
55 /*
56  * We don't actually have pdflush, but this one is exported though /proc...
57  */
58 int nr_pdflush_threads;
59
60 /**
61  * writeback_in_progress - determine whether there is writeback in progress
62  * @bdi: the device's backing_dev_info structure.
63  *
64  * Determine whether there is writeback waiting to be handled against a
65  * backing device.
66  */
67 int writeback_in_progress(struct backing_dev_info *bdi)
68 {
69         return test_bit(BDI_writeback_running, &bdi->state);
70 }
71
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
73 {
74         struct super_block *sb = inode->i_sb;
75
76         if (strcmp(sb->s_type->name, "bdev") == 0)
77                 return inode->i_mapping->backing_dev_info;
78
79         return sb->s_bdi;
80 }
81
82 static inline struct inode *wb_inode(struct list_head *head)
83 {
84         return list_entry(head, struct inode, i_wb_list);
85 }
86
87 /*
88  * Include the creation of the trace points after defining the
89  * wb_writeback_work structure and inline functions so that the definition
90  * remains local to this file.
91  */
92 #define CREATE_TRACE_POINTS
93 #include <trace/events/writeback.h>
94
95 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
96 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
97 {
98         if (bdi->wb.task) {
99                 wake_up_process(bdi->wb.task);
100         } else {
101                 /*
102                  * The bdi thread isn't there, wake up the forker thread which
103                  * will create and run it.
104                  */
105                 wake_up_process(default_backing_dev_info.wb.task);
106         }
107 }
108
109 static void bdi_queue_work(struct backing_dev_info *bdi,
110                            struct wb_writeback_work *work)
111 {
112         trace_writeback_queue(bdi, work);
113
114         spin_lock_bh(&bdi->wb_lock);
115         list_add_tail(&work->list, &bdi->work_list);
116         if (!bdi->wb.task)
117                 trace_writeback_nothread(bdi, work);
118         bdi_wakeup_flusher(bdi);
119         spin_unlock_bh(&bdi->wb_lock);
120 }
121
122 static void
123 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
124                       bool range_cyclic, enum wb_reason reason)
125 {
126         struct wb_writeback_work *work;
127
128         /*
129          * This is WB_SYNC_NONE writeback, so if allocation fails just
130          * wakeup the thread for old dirty data writeback
131          */
132         work = kzalloc(sizeof(*work), GFP_ATOMIC);
133         if (!work) {
134                 if (bdi->wb.task) {
135                         trace_writeback_nowork(bdi);
136                         wake_up_process(bdi->wb.task);
137                 }
138                 return;
139         }
140
141         work->sync_mode = WB_SYNC_NONE;
142         work->nr_pages  = nr_pages;
143         work->range_cyclic = range_cyclic;
144         work->reason    = reason;
145
146         bdi_queue_work(bdi, work);
147 }
148
149 /**
150  * bdi_start_writeback - start writeback
151  * @bdi: the backing device to write from
152  * @nr_pages: the number of pages to write
153  * @reason: reason why some writeback work was initiated
154  *
155  * Description:
156  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
157  *   started when this function returns, we make no guarantees on
158  *   completion. Caller need not hold sb s_umount semaphore.
159  *
160  */
161 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
162                         enum wb_reason reason)
163 {
164         __bdi_start_writeback(bdi, nr_pages, true, reason);
165 }
166
167 /**
168  * bdi_start_background_writeback - start background writeback
169  * @bdi: the backing device to write from
170  *
171  * Description:
172  *   This makes sure WB_SYNC_NONE background writeback happens. When
173  *   this function returns, it is only guaranteed that for given BDI
174  *   some IO is happening if we are over background dirty threshold.
175  *   Caller need not hold sb s_umount semaphore.
176  */
177 void bdi_start_background_writeback(struct backing_dev_info *bdi)
178 {
179         /*
180          * We just wake up the flusher thread. It will perform background
181          * writeback as soon as there is no other work to do.
182          */
183         trace_writeback_wake_background(bdi);
184         spin_lock_bh(&bdi->wb_lock);
185         bdi_wakeup_flusher(bdi);
186         spin_unlock_bh(&bdi->wb_lock);
187 }
188
189 /*
190  * Remove the inode from the writeback list it is on.
191  */
192 void inode_wb_list_del(struct inode *inode)
193 {
194         struct backing_dev_info *bdi = inode_to_bdi(inode);
195
196         spin_lock(&bdi->wb.list_lock);
197         list_del_init(&inode->i_wb_list);
198         spin_unlock(&bdi->wb.list_lock);
199 }
200
201 /*
202  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
203  * furthest end of its superblock's dirty-inode list.
204  *
205  * Before stamping the inode's ->dirtied_when, we check to see whether it is
206  * already the most-recently-dirtied inode on the b_dirty list.  If that is
207  * the case then the inode must have been redirtied while it was being written
208  * out and we don't reset its dirtied_when.
209  */
210 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
211 {
212         assert_spin_locked(&wb->list_lock);
213         if (!list_empty(&wb->b_dirty)) {
214                 struct inode *tail;
215
216                 tail = wb_inode(wb->b_dirty.next);
217                 if (time_before(inode->dirtied_when, tail->dirtied_when))
218                         inode->dirtied_when = jiffies;
219         }
220         list_move(&inode->i_wb_list, &wb->b_dirty);
221 }
222
223 /*
224  * requeue inode for re-scanning after bdi->b_io list is exhausted.
225  */
226 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
227 {
228         assert_spin_locked(&wb->list_lock);
229         list_move(&inode->i_wb_list, &wb->b_more_io);
230 }
231
232 static void inode_sync_complete(struct inode *inode)
233 {
234         inode->i_state &= ~I_SYNC;
235         /* Waiters must see I_SYNC cleared before being woken up */
236         smp_mb();
237         wake_up_bit(&inode->i_state, __I_SYNC);
238 }
239
240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
241 {
242         bool ret = time_after(inode->dirtied_when, t);
243 #ifndef CONFIG_64BIT
244         /*
245          * For inodes being constantly redirtied, dirtied_when can get stuck.
246          * It _appears_ to be in the future, but is actually in distant past.
247          * This test is necessary to prevent such wrapped-around relative times
248          * from permanently stopping the whole bdi writeback.
249          */
250         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
251 #endif
252         return ret;
253 }
254
255 /*
256  * Move expired (dirtied after work->older_than_this) dirty inodes from
257  * @delaying_queue to @dispatch_queue.
258  */
259 static int move_expired_inodes(struct list_head *delaying_queue,
260                                struct list_head *dispatch_queue,
261                                struct wb_writeback_work *work)
262 {
263         LIST_HEAD(tmp);
264         struct list_head *pos, *node;
265         struct super_block *sb = NULL;
266         struct inode *inode;
267         int do_sb_sort = 0;
268         int moved = 0;
269
270         while (!list_empty(delaying_queue)) {
271                 inode = wb_inode(delaying_queue->prev);
272                 if (work->older_than_this &&
273                     inode_dirtied_after(inode, *work->older_than_this))
274                         break;
275                 if (sb && sb != inode->i_sb)
276                         do_sb_sort = 1;
277                 sb = inode->i_sb;
278                 list_move(&inode->i_wb_list, &tmp);
279                 moved++;
280         }
281
282         /* just one sb in list, splice to dispatch_queue and we're done */
283         if (!do_sb_sort) {
284                 list_splice(&tmp, dispatch_queue);
285                 goto out;
286         }
287
288         /* Move inodes from one superblock together */
289         while (!list_empty(&tmp)) {
290                 sb = wb_inode(tmp.prev)->i_sb;
291                 list_for_each_prev_safe(pos, node, &tmp) {
292                         inode = wb_inode(pos);
293                         if (inode->i_sb == sb)
294                                 list_move(&inode->i_wb_list, dispatch_queue);
295                 }
296         }
297 out:
298         return moved;
299 }
300
301 /*
302  * Queue all expired dirty inodes for io, eldest first.
303  * Before
304  *         newly dirtied     b_dirty    b_io    b_more_io
305  *         =============>    gf         edc     BA
306  * After
307  *         newly dirtied     b_dirty    b_io    b_more_io
308  *         =============>    g          fBAedc
309  *                                           |
310  *                                           +--> dequeue for IO
311  */
312 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
313 {
314         int moved;
315         assert_spin_locked(&wb->list_lock);
316         list_splice_init(&wb->b_more_io, &wb->b_io);
317         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
318         trace_writeback_queue_io(wb, work, moved);
319 }
320
321 static int write_inode(struct inode *inode, struct writeback_control *wbc)
322 {
323         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
324                 return inode->i_sb->s_op->write_inode(inode, wbc);
325         return 0;
326 }
327
328 /*
329  * Wait for writeback on an inode to complete. Called with i_lock held.
330  * Caller must make sure inode cannot go away when we drop i_lock.
331  */
332 static void __inode_wait_for_writeback(struct inode *inode)
333         __releases(inode->i_lock)
334         __acquires(inode->i_lock)
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                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
343                 spin_lock(&inode->i_lock);
344         }
345 }
346
347 /*
348  * Wait for writeback on an inode to complete. Caller must have inode pinned.
349  */
350 void inode_wait_for_writeback(struct inode *inode)
351 {
352         spin_lock(&inode->i_lock);
353         __inode_wait_for_writeback(inode);
354         spin_unlock(&inode->i_lock);
355 }
356
357 /*
358  * Sleep until I_SYNC is cleared. This function must be called with i_lock
359  * held and drops it. It is aimed for callers not holding any inode reference
360  * so once i_lock is dropped, inode can go away.
361  */
362 static void inode_sleep_on_writeback(struct inode *inode)
363         __releases(inode->i_lock)
364 {
365         DEFINE_WAIT(wait);
366         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
367         int sleep;
368
369         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
370         sleep = inode->i_state & I_SYNC;
371         spin_unlock(&inode->i_lock);
372         if (sleep)
373                 schedule();
374         finish_wait(wqh, &wait);
375 }
376
377 /*
378  * Find proper writeback list for the inode depending on its current state and
379  * possibly also change of its state while we were doing writeback.  Here we
380  * handle things such as livelock prevention or fairness of writeback among
381  * inodes. This function can be called only by flusher thread - noone else
382  * processes all inodes in writeback lists and requeueing inodes behind flusher
383  * thread's back can have unexpected consequences.
384  */
385 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
386                           struct writeback_control *wbc)
387 {
388         if (inode->i_state & I_FREEING)
389                 return;
390
391         /*
392          * Sync livelock prevention. Each inode is tagged and synced in one
393          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
394          * the dirty time to prevent enqueue and sync it again.
395          */
396         if ((inode->i_state & I_DIRTY) &&
397             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
398                 inode->dirtied_when = jiffies;
399
400         if (wbc->pages_skipped) {
401                 /*
402                  * writeback is not making progress due to locked
403                  * buffers. Skip this inode for now.
404                  */
405                 redirty_tail(inode, wb);
406                 return;
407         }
408
409         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
410                 /*
411                  * We didn't write back all the pages.  nfs_writepages()
412                  * sometimes bales out without doing anything.
413                  */
414                 if (wbc->nr_to_write <= 0) {
415                         /* Slice used up. Queue for next turn. */
416                         requeue_io(inode, wb);
417                 } else {
418                         /*
419                          * Writeback blocked by something other than
420                          * congestion. Delay the inode for some time to
421                          * avoid spinning on the CPU (100% iowait)
422                          * retrying writeback of the dirty page/inode
423                          * that cannot be performed immediately.
424                          */
425                         redirty_tail(inode, wb);
426                 }
427         } else if (inode->i_state & I_DIRTY) {
428                 /*
429                  * Filesystems can dirty the inode during writeback operations,
430                  * such as delayed allocation during submission or metadata
431                  * updates after data IO completion.
432                  */
433                 redirty_tail(inode, wb);
434         } else {
435                 /* The inode is clean. Remove from writeback lists. */
436                 list_del_init(&inode->i_wb_list);
437         }
438 }
439
440 /*
441  * Write out an inode and its dirty pages. Do not update the writeback list
442  * linkage. That is left to the caller. The caller is also responsible for
443  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
444  */
445 static int
446 __writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
447                          struct writeback_control *wbc)
448 {
449         struct address_space *mapping = inode->i_mapping;
450         long nr_to_write = wbc->nr_to_write;
451         unsigned dirty;
452         int ret;
453
454         WARN_ON(!(inode->i_state & I_SYNC));
455
456         ret = do_writepages(mapping, wbc);
457
458         /*
459          * Make sure to wait on the data before writing out the metadata.
460          * This is important for filesystems that modify metadata on data
461          * I/O completion.
462          */
463         if (wbc->sync_mode == WB_SYNC_ALL) {
464                 int err = filemap_fdatawait(mapping);
465                 if (ret == 0)
466                         ret = err;
467         }
468
469         /*
470          * Some filesystems may redirty the inode during the writeback
471          * due to delalloc, clear dirty metadata flags right before
472          * write_inode()
473          */
474         spin_lock(&inode->i_lock);
475         /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
476         if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
477                 inode->i_state &= ~I_DIRTY_PAGES;
478         dirty = inode->i_state & I_DIRTY;
479         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
480         spin_unlock(&inode->i_lock);
481         /* Don't write the inode if only I_DIRTY_PAGES was set */
482         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
483                 int err = write_inode(inode, wbc);
484                 if (ret == 0)
485                         ret = err;
486         }
487         trace_writeback_single_inode(inode, wbc, nr_to_write);
488         return ret;
489 }
490
491 /*
492  * Write out an inode's dirty pages. Either the caller has an active reference
493  * on the inode or the inode has I_WILL_FREE set.
494  *
495  * This function is designed to be called for writing back one inode which
496  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
497  * and does more profound writeback list handling in writeback_sb_inodes().
498  */
499 static int
500 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
501                        struct writeback_control *wbc)
502 {
503         int ret = 0;
504
505         spin_lock(&inode->i_lock);
506         if (!atomic_read(&inode->i_count))
507                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
508         else
509                 WARN_ON(inode->i_state & I_WILL_FREE);
510
511         if (inode->i_state & I_SYNC) {
512                 if (wbc->sync_mode != WB_SYNC_ALL)
513                         goto out;
514                 /*
515                  * It's a data-integrity sync. We must wait. Since callers hold
516                  * inode reference or inode has I_WILL_FREE set, it cannot go
517                  * away under us.
518                  */
519                 __inode_wait_for_writeback(inode);
520         }
521         WARN_ON(inode->i_state & I_SYNC);
522         /*
523          * Skip inode if it is clean. We don't want to mess with writeback
524          * lists in this function since flusher thread may be doing for example
525          * sync in parallel and if we move the inode, it could get skipped. So
526          * here we make sure inode is on some writeback list and leave it there
527          * unless we have completely cleaned the inode.
528          */
529         if (!(inode->i_state & I_DIRTY))
530                 goto out;
531         inode->i_state |= I_SYNC;
532         spin_unlock(&inode->i_lock);
533
534         ret = __writeback_single_inode(inode, wb, wbc);
535
536         spin_lock(&wb->list_lock);
537         spin_lock(&inode->i_lock);
538         /*
539          * If inode is clean, remove it from writeback lists. Otherwise don't
540          * touch it. See comment above for explanation.
541          */
542         if (!(inode->i_state & I_DIRTY))
543                 list_del_init(&inode->i_wb_list);
544         spin_unlock(&wb->list_lock);
545         inode_sync_complete(inode);
546 out:
547         spin_unlock(&inode->i_lock);
548         return ret;
549 }
550
551 static long writeback_chunk_size(struct backing_dev_info *bdi,
552                                  struct wb_writeback_work *work)
553 {
554         long pages;
555
556         /*
557          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
558          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
559          * here avoids calling into writeback_inodes_wb() more than once.
560          *
561          * The intended call sequence for WB_SYNC_ALL writeback is:
562          *
563          *      wb_writeback()
564          *          writeback_sb_inodes()       <== called only once
565          *              write_cache_pages()     <== called once for each inode
566          *                   (quickly) tag currently dirty pages
567          *                   (maybe slowly) sync all tagged pages
568          */
569         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
570                 pages = LONG_MAX;
571         else {
572                 pages = min(bdi->avg_write_bandwidth / 2,
573                             global_dirty_limit / DIRTY_SCOPE);
574                 pages = min(pages, work->nr_pages);
575                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
576                                    MIN_WRITEBACK_PAGES);
577         }
578
579         return pages;
580 }
581
582 /*
583  * Write a portion of b_io inodes which belong to @sb.
584  *
585  * If @only_this_sb is true, then find and write all such
586  * inodes. Otherwise write only ones which go sequentially
587  * in reverse order.
588  *
589  * Return the number of pages and/or inodes written.
590  */
591 static long writeback_sb_inodes(struct super_block *sb,
592                                 struct bdi_writeback *wb,
593                                 struct wb_writeback_work *work)
594 {
595         struct writeback_control wbc = {
596                 .sync_mode              = work->sync_mode,
597                 .tagged_writepages      = work->tagged_writepages,
598                 .for_kupdate            = work->for_kupdate,
599                 .for_background         = work->for_background,
600                 .range_cyclic           = work->range_cyclic,
601                 .range_start            = 0,
602                 .range_end              = LLONG_MAX,
603         };
604         unsigned long start_time = jiffies;
605         long write_chunk;
606         long wrote = 0;  /* count both pages and inodes */
607
608         while (!list_empty(&wb->b_io)) {
609                 struct inode *inode = wb_inode(wb->b_io.prev);
610
611                 if (inode->i_sb != sb) {
612                         if (work->sb) {
613                                 /*
614                                  * We only want to write back data for this
615                                  * superblock, move all inodes not belonging
616                                  * to it back onto the dirty list.
617                                  */
618                                 redirty_tail(inode, wb);
619                                 continue;
620                         }
621
622                         /*
623                          * The inode belongs to a different superblock.
624                          * Bounce back to the caller to unpin this and
625                          * pin the next superblock.
626                          */
627                         break;
628                 }
629
630                 /*
631                  * Don't bother with new inodes or inodes beeing freed, first
632                  * kind does not need peridic writeout yet, and for the latter
633                  * kind writeout is handled by the freer.
634                  */
635                 spin_lock(&inode->i_lock);
636                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
637                         spin_unlock(&inode->i_lock);
638                         redirty_tail(inode, wb);
639                         continue;
640                 }
641                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
642                         /*
643                          * If this inode is locked for writeback and we are not
644                          * doing writeback-for-data-integrity, move it to
645                          * b_more_io so that writeback can proceed with the
646                          * other inodes on s_io.
647                          *
648                          * We'll have another go at writing back this inode
649                          * when we completed a full scan of b_io.
650                          */
651                         spin_unlock(&inode->i_lock);
652                         requeue_io(inode, wb);
653                         trace_writeback_sb_inodes_requeue(inode);
654                         continue;
655                 }
656                 spin_unlock(&wb->list_lock);
657
658                 /*
659                  * We already requeued the inode if it had I_SYNC set and we
660                  * are doing WB_SYNC_NONE writeback. So this catches only the
661                  * WB_SYNC_ALL case.
662                  */
663                 if (inode->i_state & I_SYNC) {
664                         /* Wait for I_SYNC. This function drops i_lock... */
665                         inode_sleep_on_writeback(inode);
666                         /* Inode may be gone, start again */
667                         continue;
668                 }
669                 inode->i_state |= I_SYNC;
670                 spin_unlock(&inode->i_lock);
671
672                 write_chunk = writeback_chunk_size(wb->bdi, work);
673                 wbc.nr_to_write = write_chunk;
674                 wbc.pages_skipped = 0;
675
676                 /*
677                  * We use I_SYNC to pin the inode in memory. While it is set
678                  * evict_inode() will wait so the inode cannot be freed.
679                  */
680                 __writeback_single_inode(inode, wb, &wbc);
681
682                 work->nr_pages -= write_chunk - wbc.nr_to_write;
683                 wrote += write_chunk - wbc.nr_to_write;
684                 spin_lock(&wb->list_lock);
685                 spin_lock(&inode->i_lock);
686                 if (!(inode->i_state & I_DIRTY))
687                         wrote++;
688                 requeue_inode(inode, wb, &wbc);
689                 inode_sync_complete(inode);
690                 spin_unlock(&inode->i_lock);
691                 cond_resched_lock(&wb->list_lock);
692                 /*
693                  * bail out to wb_writeback() often enough to check
694                  * background threshold and other termination conditions.
695                  */
696                 if (wrote) {
697                         if (time_is_before_jiffies(start_time + HZ / 10UL))
698                                 break;
699                         if (work->nr_pages <= 0)
700                                 break;
701                 }
702         }
703         return wrote;
704 }
705
706 static long __writeback_inodes_wb(struct bdi_writeback *wb,
707                                   struct wb_writeback_work *work)
708 {
709         unsigned long start_time = jiffies;
710         long wrote = 0;
711
712         while (!list_empty(&wb->b_io)) {
713                 struct inode *inode = wb_inode(wb->b_io.prev);
714                 struct super_block *sb = inode->i_sb;
715
716                 if (!grab_super_passive(sb)) {
717                         /*
718                          * grab_super_passive() may fail consistently due to
719                          * s_umount being grabbed by someone else. Don't use
720                          * requeue_io() to avoid busy retrying the inode/sb.
721                          */
722                         redirty_tail(inode, wb);
723                         continue;
724                 }
725                 wrote += writeback_sb_inodes(sb, wb, work);
726                 drop_super(sb);
727
728                 /* refer to the same tests at the end of writeback_sb_inodes */
729                 if (wrote) {
730                         if (time_is_before_jiffies(start_time + HZ / 10UL))
731                                 break;
732                         if (work->nr_pages <= 0)
733                                 break;
734                 }
735         }
736         /* Leave any unwritten inodes on b_io */
737         return wrote;
738 }
739
740 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
741                                 enum wb_reason reason)
742 {
743         struct wb_writeback_work work = {
744                 .nr_pages       = nr_pages,
745                 .sync_mode      = WB_SYNC_NONE,
746                 .range_cyclic   = 1,
747                 .reason         = reason,
748         };
749
750         spin_lock(&wb->list_lock);
751         if (list_empty(&wb->b_io))
752                 queue_io(wb, &work);
753         __writeback_inodes_wb(wb, &work);
754         spin_unlock(&wb->list_lock);
755
756         return nr_pages - work.nr_pages;
757 }
758
759 static bool over_bground_thresh(struct backing_dev_info *bdi)
760 {
761         unsigned long background_thresh, dirty_thresh;
762
763         global_dirty_limits(&background_thresh, &dirty_thresh);
764
765         if (global_page_state(NR_FILE_DIRTY) +
766             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
767                 return true;
768
769         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
770                                 bdi_dirty_limit(bdi, background_thresh))
771                 return true;
772
773         return false;
774 }
775
776 /*
777  * Called under wb->list_lock. If there are multiple wb per bdi,
778  * only the flusher working on the first wb should do it.
779  */
780 static void wb_update_bandwidth(struct bdi_writeback *wb,
781                                 unsigned long start_time)
782 {
783         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
784 }
785
786 /*
787  * Explicit flushing or periodic writeback of "old" data.
788  *
789  * Define "old": the first time one of an inode's pages is dirtied, we mark the
790  * dirtying-time in the inode's address_space.  So this periodic writeback code
791  * just walks the superblock inode list, writing back any inodes which are
792  * older than a specific point in time.
793  *
794  * Try to run once per dirty_writeback_interval.  But if a writeback event
795  * takes longer than a dirty_writeback_interval interval, then leave a
796  * one-second gap.
797  *
798  * older_than_this takes precedence over nr_to_write.  So we'll only write back
799  * all dirty pages if they are all attached to "old" mappings.
800  */
801 static long wb_writeback(struct bdi_writeback *wb,
802                          struct wb_writeback_work *work)
803 {
804         unsigned long wb_start = jiffies;
805         long nr_pages = work->nr_pages;
806         unsigned long oldest_jif;
807         struct inode *inode;
808         long progress;
809
810         oldest_jif = jiffies;
811         work->older_than_this = &oldest_jif;
812
813         spin_lock(&wb->list_lock);
814         for (;;) {
815                 /*
816                  * Stop writeback when nr_pages has been consumed
817                  */
818                 if (work->nr_pages <= 0)
819                         break;
820
821                 /*
822                  * Background writeout and kupdate-style writeback may
823                  * run forever. Stop them if there is other work to do
824                  * so that e.g. sync can proceed. They'll be restarted
825                  * after the other works are all done.
826                  */
827                 if ((work->for_background || work->for_kupdate) &&
828                     !list_empty(&wb->bdi->work_list))
829                         break;
830
831                 /*
832                  * For background writeout, stop when we are below the
833                  * background dirty threshold
834                  */
835                 if (work->for_background && !over_bground_thresh(wb->bdi))
836                         break;
837
838                 /*
839                  * Kupdate and background works are special and we want to
840                  * include all inodes that need writing. Livelock avoidance is
841                  * handled by these works yielding to any other work so we are
842                  * safe.
843                  */
844                 if (work->for_kupdate) {
845                         oldest_jif = jiffies -
846                                 msecs_to_jiffies(dirty_expire_interval * 10);
847                 } else if (work->for_background)
848                         oldest_jif = jiffies;
849
850                 trace_writeback_start(wb->bdi, work);
851                 if (list_empty(&wb->b_io))
852                         queue_io(wb, work);
853                 if (work->sb)
854                         progress = writeback_sb_inodes(work->sb, wb, work);
855                 else
856                         progress = __writeback_inodes_wb(wb, work);
857                 trace_writeback_written(wb->bdi, work);
858
859                 wb_update_bandwidth(wb, wb_start);
860
861                 /*
862                  * Did we write something? Try for more
863                  *
864                  * Dirty inodes are moved to b_io for writeback in batches.
865                  * The completion of the current batch does not necessarily
866                  * mean the overall work is done. So we keep looping as long
867                  * as made some progress on cleaning pages or inodes.
868                  */
869                 if (progress)
870                         continue;
871                 /*
872                  * No more inodes for IO, bail
873                  */
874                 if (list_empty(&wb->b_more_io))
875                         break;
876                 /*
877                  * Nothing written. Wait for some inode to
878                  * become available for writeback. Otherwise
879                  * we'll just busyloop.
880                  */
881                 if (!list_empty(&wb->b_more_io))  {
882                         trace_writeback_wait(wb->bdi, work);
883                         inode = wb_inode(wb->b_more_io.prev);
884                         spin_lock(&inode->i_lock);
885                         spin_unlock(&wb->list_lock);
886                         /* This function drops i_lock... */
887                         inode_sleep_on_writeback(inode);
888                         spin_lock(&wb->list_lock);
889                 }
890         }
891         spin_unlock(&wb->list_lock);
892
893         return nr_pages - work->nr_pages;
894 }
895
896 /*
897  * Return the next wb_writeback_work struct that hasn't been processed yet.
898  */
899 static struct wb_writeback_work *
900 get_next_work_item(struct backing_dev_info *bdi)
901 {
902         struct wb_writeback_work *work = NULL;
903
904         spin_lock_bh(&bdi->wb_lock);
905         if (!list_empty(&bdi->work_list)) {
906                 work = list_entry(bdi->work_list.next,
907                                   struct wb_writeback_work, list);
908                 list_del_init(&work->list);
909         }
910         spin_unlock_bh(&bdi->wb_lock);
911         return work;
912 }
913
914 /*
915  * Add in the number of potentially dirty inodes, because each inode
916  * write can dirty pagecache in the underlying blockdev.
917  */
918 static unsigned long get_nr_dirty_pages(void)
919 {
920         return global_page_state(NR_FILE_DIRTY) +
921                 global_page_state(NR_UNSTABLE_NFS) +
922                 get_nr_dirty_inodes();
923 }
924
925 static long wb_check_background_flush(struct bdi_writeback *wb)
926 {
927         if (over_bground_thresh(wb->bdi)) {
928
929                 struct wb_writeback_work work = {
930                         .nr_pages       = LONG_MAX,
931                         .sync_mode      = WB_SYNC_NONE,
932                         .for_background = 1,
933                         .range_cyclic   = 1,
934                         .reason         = WB_REASON_BACKGROUND,
935                 };
936
937                 return wb_writeback(wb, &work);
938         }
939
940         return 0;
941 }
942
943 static long wb_check_old_data_flush(struct bdi_writeback *wb)
944 {
945         unsigned long expired;
946         long nr_pages;
947
948         /*
949          * When set to zero, disable periodic writeback
950          */
951         if (!dirty_writeback_interval)
952                 return 0;
953
954         expired = wb->last_old_flush +
955                         msecs_to_jiffies(dirty_writeback_interval * 10);
956         if (time_before(jiffies, expired))
957                 return 0;
958
959         wb->last_old_flush = jiffies;
960         nr_pages = get_nr_dirty_pages();
961
962         if (nr_pages) {
963                 struct wb_writeback_work work = {
964                         .nr_pages       = nr_pages,
965                         .sync_mode      = WB_SYNC_NONE,
966                         .for_kupdate    = 1,
967                         .range_cyclic   = 1,
968                         .reason         = WB_REASON_PERIODIC,
969                 };
970
971                 return wb_writeback(wb, &work);
972         }
973
974         return 0;
975 }
976
977 /*
978  * Retrieve work items and do the writeback they describe
979  */
980 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
981 {
982         struct backing_dev_info *bdi = wb->bdi;
983         struct wb_writeback_work *work;
984         long wrote = 0;
985
986         set_bit(BDI_writeback_running, &wb->bdi->state);
987         while ((work = get_next_work_item(bdi)) != NULL) {
988                 /*
989                  * Override sync mode, in case we must wait for completion
990                  * because this thread is exiting now.
991                  */
992                 if (force_wait)
993                         work->sync_mode = WB_SYNC_ALL;
994
995                 trace_writeback_exec(bdi, work);
996
997                 wrote += wb_writeback(wb, work);
998
999                 /*
1000                  * Notify the caller of completion if this is a synchronous
1001                  * work item, otherwise just free it.
1002                  */
1003                 if (work->done)
1004                         complete(work->done);
1005                 else
1006                         kfree(work);
1007         }
1008
1009         /*
1010          * Check for periodic writeback, kupdated() style
1011          */
1012         wrote += wb_check_old_data_flush(wb);
1013         wrote += wb_check_background_flush(wb);
1014         clear_bit(BDI_writeback_running, &wb->bdi->state);
1015
1016         return wrote;
1017 }
1018
1019 /*
1020  * Handle writeback of dirty data for the device backed by this bdi. Also
1021  * wakes up periodically and does kupdated style flushing.
1022  */
1023 int bdi_writeback_thread(void *data)
1024 {
1025         struct bdi_writeback *wb = data;
1026         struct backing_dev_info *bdi = wb->bdi;
1027         long pages_written;
1028
1029         current->flags |= PF_SWAPWRITE;
1030         set_freezable();
1031         wb->last_active = jiffies;
1032
1033         /*
1034          * Our parent may run at a different priority, just set us to normal
1035          */
1036         set_user_nice(current, 0);
1037
1038         trace_writeback_thread_start(bdi);
1039
1040         while (!kthread_freezable_should_stop(NULL)) {
1041                 /*
1042                  * Remove own delayed wake-up timer, since we are already awake
1043                  * and we'll take care of the preriodic write-back.
1044                  */
1045                 del_timer(&wb->wakeup_timer);
1046
1047                 pages_written = wb_do_writeback(wb, 0);
1048
1049                 trace_writeback_pages_written(pages_written);
1050
1051                 if (pages_written)
1052                         wb->last_active = jiffies;
1053
1054                 set_current_state(TASK_INTERRUPTIBLE);
1055                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
1056                         __set_current_state(TASK_RUNNING);
1057                         continue;
1058                 }
1059
1060                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1061                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
1062                 else {
1063                         /*
1064                          * We have nothing to do, so can go sleep without any
1065                          * timeout and save power. When a work is queued or
1066                          * something is made dirty - we will be woken up.
1067                          */
1068                         schedule();
1069                 }
1070         }
1071
1072         /* Flush any work that raced with us exiting */
1073         if (!list_empty(&bdi->work_list))
1074                 wb_do_writeback(wb, 1);
1075
1076         trace_writeback_thread_stop(bdi);
1077         return 0;
1078 }
1079
1080
1081 /*
1082  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1083  * the whole world.
1084  */
1085 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1086 {
1087         struct backing_dev_info *bdi;
1088
1089         if (!nr_pages) {
1090                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1091                                 global_page_state(NR_UNSTABLE_NFS);
1092         }
1093
1094         rcu_read_lock();
1095         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1096                 if (!bdi_has_dirty_io(bdi))
1097                         continue;
1098                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1099         }
1100         rcu_read_unlock();
1101 }
1102
1103 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1104 {
1105         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1106                 struct dentry *dentry;
1107                 const char *name = "?";
1108
1109                 dentry = d_find_alias(inode);
1110                 if (dentry) {
1111                         spin_lock(&dentry->d_lock);
1112                         name = (const char *) dentry->d_name.name;
1113                 }
1114                 printk(KERN_DEBUG
1115                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1116                        current->comm, task_pid_nr(current), inode->i_ino,
1117                        name, inode->i_sb->s_id);
1118                 if (dentry) {
1119                         spin_unlock(&dentry->d_lock);
1120                         dput(dentry);
1121                 }
1122         }
1123 }
1124
1125 /**
1126  *      __mark_inode_dirty -    internal function
1127  *      @inode: inode to mark
1128  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1129  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1130  *      mark_inode_dirty_sync.
1131  *
1132  * Put the inode on the super block's dirty list.
1133  *
1134  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1135  * dirty list only if it is hashed or if it refers to a blockdev.
1136  * If it was not hashed, it will never be added to the dirty list
1137  * even if it is later hashed, as it will have been marked dirty already.
1138  *
1139  * In short, make sure you hash any inodes _before_ you start marking
1140  * them dirty.
1141  *
1142  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1143  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1144  * the kernel-internal blockdev inode represents the dirtying time of the
1145  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1146  * page->mapping->host, so the page-dirtying time is recorded in the internal
1147  * blockdev inode.
1148  */
1149 void __mark_inode_dirty(struct inode *inode, int flags)
1150 {
1151         struct super_block *sb = inode->i_sb;
1152         struct backing_dev_info *bdi = NULL;
1153
1154         /*
1155          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1156          * dirty the inode itself
1157          */
1158         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1159                 if (sb->s_op->dirty_inode)
1160                         sb->s_op->dirty_inode(inode, flags);
1161         }
1162
1163         /*
1164          * make sure that changes are seen by all cpus before we test i_state
1165          * -- mikulas
1166          */
1167         smp_mb();
1168
1169         /* avoid the locking if we can */
1170         if ((inode->i_state & flags) == flags)
1171                 return;
1172
1173         if (unlikely(block_dump))
1174                 block_dump___mark_inode_dirty(inode);
1175
1176         spin_lock(&inode->i_lock);
1177         if ((inode->i_state & flags) != flags) {
1178                 const int was_dirty = inode->i_state & I_DIRTY;
1179
1180                 inode->i_state |= flags;
1181
1182                 /*
1183                  * If the inode is being synced, just update its dirty state.
1184                  * The unlocker will place the inode on the appropriate
1185                  * superblock list, based upon its state.
1186                  */
1187                 if (inode->i_state & I_SYNC)
1188                         goto out_unlock_inode;
1189
1190                 /*
1191                  * Only add valid (hashed) inodes to the superblock's
1192                  * dirty list.  Add blockdev inodes as well.
1193                  */
1194                 if (!S_ISBLK(inode->i_mode)) {
1195                         if (inode_unhashed(inode))
1196                                 goto out_unlock_inode;
1197                 }
1198                 if (inode->i_state & I_FREEING)
1199                         goto out_unlock_inode;
1200
1201                 /*
1202                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1203                  * reposition it (that would break b_dirty time-ordering).
1204                  */
1205                 if (!was_dirty) {
1206                         bool wakeup_bdi = false;
1207                         bdi = inode_to_bdi(inode);
1208
1209                         if (bdi_cap_writeback_dirty(bdi)) {
1210                                 WARN(!test_bit(BDI_registered, &bdi->state),
1211                                      "bdi-%s not registered\n", bdi->name);
1212
1213                                 /*
1214                                  * If this is the first dirty inode for this
1215                                  * bdi, we have to wake-up the corresponding
1216                                  * bdi thread to make sure background
1217                                  * write-back happens later.
1218                                  */
1219                                 if (!wb_has_dirty_io(&bdi->wb))
1220                                         wakeup_bdi = true;
1221                         }
1222
1223                         spin_unlock(&inode->i_lock);
1224                         spin_lock(&bdi->wb.list_lock);
1225                         inode->dirtied_when = jiffies;
1226                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1227                         spin_unlock(&bdi->wb.list_lock);
1228
1229                         if (wakeup_bdi)
1230                                 bdi_wakeup_thread_delayed(bdi);
1231                         return;
1232                 }
1233         }
1234 out_unlock_inode:
1235         spin_unlock(&inode->i_lock);
1236
1237 }
1238 EXPORT_SYMBOL(__mark_inode_dirty);
1239
1240 static void wait_sb_inodes(struct super_block *sb)
1241 {
1242         struct inode *inode, *old_inode = NULL;
1243
1244         /*
1245          * We need to be protected against the filesystem going from
1246          * r/o to r/w or vice versa.
1247          */
1248         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1249
1250         spin_lock(&inode_sb_list_lock);
1251
1252         /*
1253          * Data integrity sync. Must wait for all pages under writeback,
1254          * because there may have been pages dirtied before our sync
1255          * call, but which had writeout started before we write it out.
1256          * In which case, the inode may not be on the dirty list, but
1257          * we still have to wait for that writeout.
1258          */
1259         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1260                 struct address_space *mapping = inode->i_mapping;
1261
1262                 spin_lock(&inode->i_lock);
1263                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1264                     (mapping->nrpages == 0)) {
1265                         spin_unlock(&inode->i_lock);
1266                         continue;
1267                 }
1268                 __iget(inode);
1269                 spin_unlock(&inode->i_lock);
1270                 spin_unlock(&inode_sb_list_lock);
1271
1272                 /*
1273                  * We hold a reference to 'inode' so it couldn't have been
1274                  * removed from s_inodes list while we dropped the
1275                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1276                  * be holding the last reference and we cannot iput it under
1277                  * inode_sb_list_lock. So we keep the reference and iput it
1278                  * later.
1279                  */
1280                 iput(old_inode);
1281                 old_inode = inode;
1282
1283                 filemap_fdatawait(mapping);
1284
1285                 cond_resched();
1286
1287                 spin_lock(&inode_sb_list_lock);
1288         }
1289         spin_unlock(&inode_sb_list_lock);
1290         iput(old_inode);
1291 }
1292
1293 /**
1294  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1295  * @sb: the superblock
1296  * @nr: the number of pages to write
1297  * @reason: reason why some writeback work initiated
1298  *
1299  * Start writeback on some inodes on this super_block. No guarantees are made
1300  * on how many (if any) will be written, and this function does not wait
1301  * for IO completion of submitted IO.
1302  */
1303 void writeback_inodes_sb_nr(struct super_block *sb,
1304                             unsigned long nr,
1305                             enum wb_reason reason)
1306 {
1307         DECLARE_COMPLETION_ONSTACK(done);
1308         struct wb_writeback_work work = {
1309                 .sb                     = sb,
1310                 .sync_mode              = WB_SYNC_NONE,
1311                 .tagged_writepages      = 1,
1312                 .done                   = &done,
1313                 .nr_pages               = nr,
1314                 .reason                 = reason,
1315         };
1316
1317         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1318         bdi_queue_work(sb->s_bdi, &work);
1319         wait_for_completion(&done);
1320 }
1321 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1322
1323 /**
1324  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1325  * @sb: the superblock
1326  * @reason: reason why some writeback work was initiated
1327  *
1328  * Start writeback on some inodes on this super_block. No guarantees are made
1329  * on how many (if any) will be written, and this function does not wait
1330  * for IO completion of submitted IO.
1331  */
1332 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1333 {
1334         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1335 }
1336 EXPORT_SYMBOL(writeback_inodes_sb);
1337
1338 /**
1339  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1340  * @sb: the superblock
1341  * @reason: reason why some writeback work was initiated
1342  *
1343  * Invoke writeback_inodes_sb if no writeback is currently underway.
1344  * Returns 1 if writeback was started, 0 if not.
1345  */
1346 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1347 {
1348         if (!writeback_in_progress(sb->s_bdi)) {
1349                 down_read(&sb->s_umount);
1350                 writeback_inodes_sb(sb, reason);
1351                 up_read(&sb->s_umount);
1352                 return 1;
1353         } else
1354                 return 0;
1355 }
1356 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1357
1358 /**
1359  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1360  * @sb: the superblock
1361  * @nr: the number of pages to write
1362  * @reason: reason why some writeback work was initiated
1363  *
1364  * Invoke writeback_inodes_sb if no writeback is currently underway.
1365  * Returns 1 if writeback was started, 0 if not.
1366  */
1367 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1368                                    unsigned long nr,
1369                                    enum wb_reason reason)
1370 {
1371         if (!writeback_in_progress(sb->s_bdi)) {
1372                 down_read(&sb->s_umount);
1373                 writeback_inodes_sb_nr(sb, nr, reason);
1374                 up_read(&sb->s_umount);
1375                 return 1;
1376         } else
1377                 return 0;
1378 }
1379 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1380
1381 /**
1382  * sync_inodes_sb       -       sync sb inode pages
1383  * @sb: the superblock
1384  *
1385  * This function writes and waits on any dirty inode belonging to this
1386  * super_block.
1387  */
1388 void sync_inodes_sb(struct super_block *sb)
1389 {
1390         DECLARE_COMPLETION_ONSTACK(done);
1391         struct wb_writeback_work work = {
1392                 .sb             = sb,
1393                 .sync_mode      = WB_SYNC_ALL,
1394                 .nr_pages       = LONG_MAX,
1395                 .range_cyclic   = 0,
1396                 .done           = &done,
1397                 .reason         = WB_REASON_SYNC,
1398         };
1399
1400         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1401
1402         bdi_queue_work(sb->s_bdi, &work);
1403         wait_for_completion(&done);
1404
1405         wait_sb_inodes(sb);
1406 }
1407 EXPORT_SYMBOL(sync_inodes_sb);
1408
1409 /**
1410  * write_inode_now      -       write an inode to disk
1411  * @inode: inode to write to disk
1412  * @sync: whether the write should be synchronous or not
1413  *
1414  * This function commits an inode to disk immediately if it is dirty. This is
1415  * primarily needed by knfsd.
1416  *
1417  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1418  */
1419 int write_inode_now(struct inode *inode, int sync)
1420 {
1421         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1422         struct writeback_control wbc = {
1423                 .nr_to_write = LONG_MAX,
1424                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1425                 .range_start = 0,
1426                 .range_end = LLONG_MAX,
1427         };
1428
1429         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1430                 wbc.nr_to_write = 0;
1431
1432         might_sleep();
1433         return writeback_single_inode(inode, wb, &wbc);
1434 }
1435 EXPORT_SYMBOL(write_inode_now);
1436
1437 /**
1438  * sync_inode - write an inode and its pages to disk.
1439  * @inode: the inode to sync
1440  * @wbc: controls the writeback mode
1441  *
1442  * sync_inode() will write an inode and its pages to disk.  It will also
1443  * correctly update the inode on its superblock's dirty inode lists and will
1444  * update inode->i_state.
1445  *
1446  * The caller must have a ref on the inode.
1447  */
1448 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1449 {
1450         return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1451 }
1452 EXPORT_SYMBOL(sync_inode);
1453
1454 /**
1455  * sync_inode_metadata - write an inode to disk
1456  * @inode: the inode to sync
1457  * @wait: wait for I/O to complete.
1458  *
1459  * Write an inode to disk and adjust its dirty state after completion.
1460  *
1461  * Note: only writes the actual inode, no associated data or other metadata.
1462  */
1463 int sync_inode_metadata(struct inode *inode, int wait)
1464 {
1465         struct writeback_control wbc = {
1466                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1467                 .nr_to_write = 0, /* metadata-only */
1468         };
1469
1470         return sync_inode(inode, &wbc);
1471 }
1472 EXPORT_SYMBOL(sync_inode_metadata);