writeback: Fix lock imbalance in writeback_sb_inodes()
[linux-3.10.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33  * 4MB minimal write chunk size
34  */
35 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
36
37 /*
38  * Passed into wb_writeback(), essentially a subset of writeback_control
39  */
40 struct wb_writeback_work {
41         long nr_pages;
42         struct super_block *sb;
43         unsigned long *older_than_this;
44         enum writeback_sync_modes sync_mode;
45         unsigned int tagged_writepages:1;
46         unsigned int for_kupdate:1;
47         unsigned int range_cyclic:1;
48         unsigned int for_background:1;
49         enum wb_reason reason;          /* why was writeback initiated? */
50
51         struct list_head list;          /* pending work list */
52         struct completion *done;        /* set if the caller waits */
53 };
54
55 /*
56  * We don't actually have pdflush, but this one is exported though /proc...
57  */
58 int nr_pdflush_threads;
59
60 /**
61  * writeback_in_progress - determine whether there is writeback in progress
62  * @bdi: the device's backing_dev_info structure.
63  *
64  * Determine whether there is writeback waiting to be handled against a
65  * backing device.
66  */
67 int writeback_in_progress(struct backing_dev_info *bdi)
68 {
69         return test_bit(BDI_writeback_running, &bdi->state);
70 }
71
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
73 {
74         struct super_block *sb = inode->i_sb;
75
76         if (strcmp(sb->s_type->name, "bdev") == 0)
77                 return inode->i_mapping->backing_dev_info;
78
79         return sb->s_bdi;
80 }
81
82 static inline struct inode *wb_inode(struct list_head *head)
83 {
84         return list_entry(head, struct inode, i_wb_list);
85 }
86
87 /*
88  * Include the creation of the trace points after defining the
89  * wb_writeback_work structure and inline functions so that the definition
90  * remains local to this file.
91  */
92 #define CREATE_TRACE_POINTS
93 #include <trace/events/writeback.h>
94
95 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
96 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
97 {
98         if (bdi->wb.task) {
99                 wake_up_process(bdi->wb.task);
100         } else {
101                 /*
102                  * The bdi thread isn't there, wake up the forker thread which
103                  * will create and run it.
104                  */
105                 wake_up_process(default_backing_dev_info.wb.task);
106         }
107 }
108
109 static void bdi_queue_work(struct backing_dev_info *bdi,
110                            struct wb_writeback_work *work)
111 {
112         trace_writeback_queue(bdi, work);
113
114         spin_lock_bh(&bdi->wb_lock);
115         list_add_tail(&work->list, &bdi->work_list);
116         if (!bdi->wb.task)
117                 trace_writeback_nothread(bdi, work);
118         bdi_wakeup_flusher(bdi);
119         spin_unlock_bh(&bdi->wb_lock);
120 }
121
122 static void
123 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
124                       bool range_cyclic, enum wb_reason reason)
125 {
126         struct wb_writeback_work *work;
127
128         /*
129          * This is WB_SYNC_NONE writeback, so if allocation fails just
130          * wakeup the thread for old dirty data writeback
131          */
132         work = kzalloc(sizeof(*work), GFP_ATOMIC);
133         if (!work) {
134                 if (bdi->wb.task) {
135                         trace_writeback_nowork(bdi);
136                         wake_up_process(bdi->wb.task);
137                 }
138                 return;
139         }
140
141         work->sync_mode = WB_SYNC_NONE;
142         work->nr_pages  = nr_pages;
143         work->range_cyclic = range_cyclic;
144         work->reason    = reason;
145
146         bdi_queue_work(bdi, work);
147 }
148
149 /**
150  * bdi_start_writeback - start writeback
151  * @bdi: the backing device to write from
152  * @nr_pages: the number of pages to write
153  * @reason: reason why some writeback work was initiated
154  *
155  * Description:
156  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
157  *   started when this function returns, we make no guarantees on
158  *   completion. Caller need not hold sb s_umount semaphore.
159  *
160  */
161 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
162                         enum wb_reason reason)
163 {
164         __bdi_start_writeback(bdi, nr_pages, true, reason);
165 }
166
167 /**
168  * bdi_start_background_writeback - start background writeback
169  * @bdi: the backing device to write from
170  *
171  * Description:
172  *   This makes sure WB_SYNC_NONE background writeback happens. When
173  *   this function returns, it is only guaranteed that for given BDI
174  *   some IO is happening if we are over background dirty threshold.
175  *   Caller need not hold sb s_umount semaphore.
176  */
177 void bdi_start_background_writeback(struct backing_dev_info *bdi)
178 {
179         /*
180          * We just wake up the flusher thread. It will perform background
181          * writeback as soon as there is no other work to do.
182          */
183         trace_writeback_wake_background(bdi);
184         spin_lock_bh(&bdi->wb_lock);
185         bdi_wakeup_flusher(bdi);
186         spin_unlock_bh(&bdi->wb_lock);
187 }
188
189 /*
190  * Remove the inode from the writeback list it is on.
191  */
192 void inode_wb_list_del(struct inode *inode)
193 {
194         struct backing_dev_info *bdi = inode_to_bdi(inode);
195
196         spin_lock(&bdi->wb.list_lock);
197         list_del_init(&inode->i_wb_list);
198         spin_unlock(&bdi->wb.list_lock);
199 }
200
201 /*
202  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
203  * furthest end of its superblock's dirty-inode list.
204  *
205  * Before stamping the inode's ->dirtied_when, we check to see whether it is
206  * already the most-recently-dirtied inode on the b_dirty list.  If that is
207  * the case then the inode must have been redirtied while it was being written
208  * out and we don't reset its dirtied_when.
209  */
210 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
211 {
212         assert_spin_locked(&wb->list_lock);
213         if (!list_empty(&wb->b_dirty)) {
214                 struct inode *tail;
215
216                 tail = wb_inode(wb->b_dirty.next);
217                 if (time_before(inode->dirtied_when, tail->dirtied_when))
218                         inode->dirtied_when = jiffies;
219         }
220         list_move(&inode->i_wb_list, &wb->b_dirty);
221 }
222
223 /*
224  * requeue inode for re-scanning after bdi->b_io list is exhausted.
225  */
226 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
227 {
228         assert_spin_locked(&wb->list_lock);
229         list_move(&inode->i_wb_list, &wb->b_more_io);
230 }
231
232 static void inode_sync_complete(struct inode *inode)
233 {
234         inode->i_state &= ~I_SYNC;
235         /* Waiters must see I_SYNC cleared before being woken up */
236         smp_mb();
237         wake_up_bit(&inode->i_state, __I_SYNC);
238 }
239
240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
241 {
242         bool ret = time_after(inode->dirtied_when, t);
243 #ifndef CONFIG_64BIT
244         /*
245          * For inodes being constantly redirtied, dirtied_when can get stuck.
246          * It _appears_ to be in the future, but is actually in distant past.
247          * This test is necessary to prevent such wrapped-around relative times
248          * from permanently stopping the whole bdi writeback.
249          */
250         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
251 #endif
252         return ret;
253 }
254
255 /*
256  * Move expired (dirtied after work->older_than_this) dirty inodes from
257  * @delaying_queue to @dispatch_queue.
258  */
259 static int move_expired_inodes(struct list_head *delaying_queue,
260                                struct list_head *dispatch_queue,
261                                struct wb_writeback_work *work)
262 {
263         LIST_HEAD(tmp);
264         struct list_head *pos, *node;
265         struct super_block *sb = NULL;
266         struct inode *inode;
267         int do_sb_sort = 0;
268         int moved = 0;
269
270         while (!list_empty(delaying_queue)) {
271                 inode = wb_inode(delaying_queue->prev);
272                 if (work->older_than_this &&
273                     inode_dirtied_after(inode, *work->older_than_this))
274                         break;
275                 if (sb && sb != inode->i_sb)
276                         do_sb_sort = 1;
277                 sb = inode->i_sb;
278                 list_move(&inode->i_wb_list, &tmp);
279                 moved++;
280         }
281
282         /* just one sb in list, splice to dispatch_queue and we're done */
283         if (!do_sb_sort) {
284                 list_splice(&tmp, dispatch_queue);
285                 goto out;
286         }
287
288         /* Move inodes from one superblock together */
289         while (!list_empty(&tmp)) {
290                 sb = wb_inode(tmp.prev)->i_sb;
291                 list_for_each_prev_safe(pos, node, &tmp) {
292                         inode = wb_inode(pos);
293                         if (inode->i_sb == sb)
294                                 list_move(&inode->i_wb_list, dispatch_queue);
295                 }
296         }
297 out:
298         return moved;
299 }
300
301 /*
302  * Queue all expired dirty inodes for io, eldest first.
303  * Before
304  *         newly dirtied     b_dirty    b_io    b_more_io
305  *         =============>    gf         edc     BA
306  * After
307  *         newly dirtied     b_dirty    b_io    b_more_io
308  *         =============>    g          fBAedc
309  *                                           |
310  *                                           +--> dequeue for IO
311  */
312 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
313 {
314         int moved;
315         assert_spin_locked(&wb->list_lock);
316         list_splice_init(&wb->b_more_io, &wb->b_io);
317         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
318         trace_writeback_queue_io(wb, work, moved);
319 }
320
321 static int write_inode(struct inode *inode, struct writeback_control *wbc)
322 {
323         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
324                 return inode->i_sb->s_op->write_inode(inode, wbc);
325         return 0;
326 }
327
328 /*
329  * Wait for writeback on an inode to complete. 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                         spin_lock(&wb->list_lock);
668                         continue;
669                 }
670                 inode->i_state |= I_SYNC;
671                 spin_unlock(&inode->i_lock);
672
673                 write_chunk = writeback_chunk_size(wb->bdi, work);
674                 wbc.nr_to_write = write_chunk;
675                 wbc.pages_skipped = 0;
676
677                 /*
678                  * We use I_SYNC to pin the inode in memory. While it is set
679                  * evict_inode() will wait so the inode cannot be freed.
680                  */
681                 __writeback_single_inode(inode, wb, &wbc);
682
683                 work->nr_pages -= write_chunk - wbc.nr_to_write;
684                 wrote += write_chunk - wbc.nr_to_write;
685                 spin_lock(&wb->list_lock);
686                 spin_lock(&inode->i_lock);
687                 if (!(inode->i_state & I_DIRTY))
688                         wrote++;
689                 requeue_inode(inode, wb, &wbc);
690                 inode_sync_complete(inode);
691                 spin_unlock(&inode->i_lock);
692                 cond_resched_lock(&wb->list_lock);
693                 /*
694                  * bail out to wb_writeback() often enough to check
695                  * background threshold and other termination conditions.
696                  */
697                 if (wrote) {
698                         if (time_is_before_jiffies(start_time + HZ / 10UL))
699                                 break;
700                         if (work->nr_pages <= 0)
701                                 break;
702                 }
703         }
704         return wrote;
705 }
706
707 static long __writeback_inodes_wb(struct bdi_writeback *wb,
708                                   struct wb_writeback_work *work)
709 {
710         unsigned long start_time = jiffies;
711         long wrote = 0;
712
713         while (!list_empty(&wb->b_io)) {
714                 struct inode *inode = wb_inode(wb->b_io.prev);
715                 struct super_block *sb = inode->i_sb;
716
717                 if (!grab_super_passive(sb)) {
718                         /*
719                          * grab_super_passive() may fail consistently due to
720                          * s_umount being grabbed by someone else. Don't use
721                          * requeue_io() to avoid busy retrying the inode/sb.
722                          */
723                         redirty_tail(inode, wb);
724                         continue;
725                 }
726                 wrote += writeback_sb_inodes(sb, wb, work);
727                 drop_super(sb);
728
729                 /* refer to the same tests at the end of writeback_sb_inodes */
730                 if (wrote) {
731                         if (time_is_before_jiffies(start_time + HZ / 10UL))
732                                 break;
733                         if (work->nr_pages <= 0)
734                                 break;
735                 }
736         }
737         /* Leave any unwritten inodes on b_io */
738         return wrote;
739 }
740
741 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
742                                 enum wb_reason reason)
743 {
744         struct wb_writeback_work work = {
745                 .nr_pages       = nr_pages,
746                 .sync_mode      = WB_SYNC_NONE,
747                 .range_cyclic   = 1,
748                 .reason         = reason,
749         };
750
751         spin_lock(&wb->list_lock);
752         if (list_empty(&wb->b_io))
753                 queue_io(wb, &work);
754         __writeback_inodes_wb(wb, &work);
755         spin_unlock(&wb->list_lock);
756
757         return nr_pages - work.nr_pages;
758 }
759
760 static bool over_bground_thresh(struct backing_dev_info *bdi)
761 {
762         unsigned long background_thresh, dirty_thresh;
763
764         global_dirty_limits(&background_thresh, &dirty_thresh);
765
766         if (global_page_state(NR_FILE_DIRTY) +
767             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
768                 return true;
769
770         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
771                                 bdi_dirty_limit(bdi, background_thresh))
772                 return true;
773
774         return false;
775 }
776
777 /*
778  * Called under wb->list_lock. If there are multiple wb per bdi,
779  * only the flusher working on the first wb should do it.
780  */
781 static void wb_update_bandwidth(struct bdi_writeback *wb,
782                                 unsigned long start_time)
783 {
784         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
785 }
786
787 /*
788  * Explicit flushing or periodic writeback of "old" data.
789  *
790  * Define "old": the first time one of an inode's pages is dirtied, we mark the
791  * dirtying-time in the inode's address_space.  So this periodic writeback code
792  * just walks the superblock inode list, writing back any inodes which are
793  * older than a specific point in time.
794  *
795  * Try to run once per dirty_writeback_interval.  But if a writeback event
796  * takes longer than a dirty_writeback_interval interval, then leave a
797  * one-second gap.
798  *
799  * older_than_this takes precedence over nr_to_write.  So we'll only write back
800  * all dirty pages if they are all attached to "old" mappings.
801  */
802 static long wb_writeback(struct bdi_writeback *wb,
803                          struct wb_writeback_work *work)
804 {
805         unsigned long wb_start = jiffies;
806         long nr_pages = work->nr_pages;
807         unsigned long oldest_jif;
808         struct inode *inode;
809         long progress;
810
811         oldest_jif = jiffies;
812         work->older_than_this = &oldest_jif;
813
814         spin_lock(&wb->list_lock);
815         for (;;) {
816                 /*
817                  * Stop writeback when nr_pages has been consumed
818                  */
819                 if (work->nr_pages <= 0)
820                         break;
821
822                 /*
823                  * Background writeout and kupdate-style writeback may
824                  * run forever. Stop them if there is other work to do
825                  * so that e.g. sync can proceed. They'll be restarted
826                  * after the other works are all done.
827                  */
828                 if ((work->for_background || work->for_kupdate) &&
829                     !list_empty(&wb->bdi->work_list))
830                         break;
831
832                 /*
833                  * For background writeout, stop when we are below the
834                  * background dirty threshold
835                  */
836                 if (work->for_background && !over_bground_thresh(wb->bdi))
837                         break;
838
839                 /*
840                  * Kupdate and background works are special and we want to
841                  * include all inodes that need writing. Livelock avoidance is
842                  * handled by these works yielding to any other work so we are
843                  * safe.
844                  */
845                 if (work->for_kupdate) {
846                         oldest_jif = jiffies -
847                                 msecs_to_jiffies(dirty_expire_interval * 10);
848                 } else if (work->for_background)
849                         oldest_jif = jiffies;
850
851                 trace_writeback_start(wb->bdi, work);
852                 if (list_empty(&wb->b_io))
853                         queue_io(wb, work);
854                 if (work->sb)
855                         progress = writeback_sb_inodes(work->sb, wb, work);
856                 else
857                         progress = __writeback_inodes_wb(wb, work);
858                 trace_writeback_written(wb->bdi, work);
859
860                 wb_update_bandwidth(wb, wb_start);
861
862                 /*
863                  * Did we write something? Try for more
864                  *
865                  * Dirty inodes are moved to b_io for writeback in batches.
866                  * The completion of the current batch does not necessarily
867                  * mean the overall work is done. So we keep looping as long
868                  * as made some progress on cleaning pages or inodes.
869                  */
870                 if (progress)
871                         continue;
872                 /*
873                  * No more inodes for IO, bail
874                  */
875                 if (list_empty(&wb->b_more_io))
876                         break;
877                 /*
878                  * Nothing written. Wait for some inode to
879                  * become available for writeback. Otherwise
880                  * we'll just busyloop.
881                  */
882                 if (!list_empty(&wb->b_more_io))  {
883                         trace_writeback_wait(wb->bdi, work);
884                         inode = wb_inode(wb->b_more_io.prev);
885                         spin_lock(&inode->i_lock);
886                         spin_unlock(&wb->list_lock);
887                         /* This function drops i_lock... */
888                         inode_sleep_on_writeback(inode);
889                         spin_lock(&wb->list_lock);
890                 }
891         }
892         spin_unlock(&wb->list_lock);
893
894         return nr_pages - work->nr_pages;
895 }
896
897 /*
898  * Return the next wb_writeback_work struct that hasn't been processed yet.
899  */
900 static struct wb_writeback_work *
901 get_next_work_item(struct backing_dev_info *bdi)
902 {
903         struct wb_writeback_work *work = NULL;
904
905         spin_lock_bh(&bdi->wb_lock);
906         if (!list_empty(&bdi->work_list)) {
907                 work = list_entry(bdi->work_list.next,
908                                   struct wb_writeback_work, list);
909                 list_del_init(&work->list);
910         }
911         spin_unlock_bh(&bdi->wb_lock);
912         return work;
913 }
914
915 /*
916  * Add in the number of potentially dirty inodes, because each inode
917  * write can dirty pagecache in the underlying blockdev.
918  */
919 static unsigned long get_nr_dirty_pages(void)
920 {
921         return global_page_state(NR_FILE_DIRTY) +
922                 global_page_state(NR_UNSTABLE_NFS) +
923                 get_nr_dirty_inodes();
924 }
925
926 static long wb_check_background_flush(struct bdi_writeback *wb)
927 {
928         if (over_bground_thresh(wb->bdi)) {
929
930                 struct wb_writeback_work work = {
931                         .nr_pages       = LONG_MAX,
932                         .sync_mode      = WB_SYNC_NONE,
933                         .for_background = 1,
934                         .range_cyclic   = 1,
935                         .reason         = WB_REASON_BACKGROUND,
936                 };
937
938                 return wb_writeback(wb, &work);
939         }
940
941         return 0;
942 }
943
944 static long wb_check_old_data_flush(struct bdi_writeback *wb)
945 {
946         unsigned long expired;
947         long nr_pages;
948
949         /*
950          * When set to zero, disable periodic writeback
951          */
952         if (!dirty_writeback_interval)
953                 return 0;
954
955         expired = wb->last_old_flush +
956                         msecs_to_jiffies(dirty_writeback_interval * 10);
957         if (time_before(jiffies, expired))
958                 return 0;
959
960         wb->last_old_flush = jiffies;
961         nr_pages = get_nr_dirty_pages();
962
963         if (nr_pages) {
964                 struct wb_writeback_work work = {
965                         .nr_pages       = nr_pages,
966                         .sync_mode      = WB_SYNC_NONE,
967                         .for_kupdate    = 1,
968                         .range_cyclic   = 1,
969                         .reason         = WB_REASON_PERIODIC,
970                 };
971
972                 return wb_writeback(wb, &work);
973         }
974
975         return 0;
976 }
977
978 /*
979  * Retrieve work items and do the writeback they describe
980  */
981 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
982 {
983         struct backing_dev_info *bdi = wb->bdi;
984         struct wb_writeback_work *work;
985         long wrote = 0;
986
987         set_bit(BDI_writeback_running, &wb->bdi->state);
988         while ((work = get_next_work_item(bdi)) != NULL) {
989                 /*
990                  * Override sync mode, in case we must wait for completion
991                  * because this thread is exiting now.
992                  */
993                 if (force_wait)
994                         work->sync_mode = WB_SYNC_ALL;
995
996                 trace_writeback_exec(bdi, work);
997
998                 wrote += wb_writeback(wb, work);
999
1000                 /*
1001                  * Notify the caller of completion if this is a synchronous
1002                  * work item, otherwise just free it.
1003                  */
1004                 if (work->done)
1005                         complete(work->done);
1006                 else
1007                         kfree(work);
1008         }
1009
1010         /*
1011          * Check for periodic writeback, kupdated() style
1012          */
1013         wrote += wb_check_old_data_flush(wb);
1014         wrote += wb_check_background_flush(wb);
1015         clear_bit(BDI_writeback_running, &wb->bdi->state);
1016
1017         return wrote;
1018 }
1019
1020 /*
1021  * Handle writeback of dirty data for the device backed by this bdi. Also
1022  * wakes up periodically and does kupdated style flushing.
1023  */
1024 int bdi_writeback_thread(void *data)
1025 {
1026         struct bdi_writeback *wb = data;
1027         struct backing_dev_info *bdi = wb->bdi;
1028         long pages_written;
1029
1030         current->flags |= PF_SWAPWRITE;
1031         set_freezable();
1032         wb->last_active = jiffies;
1033
1034         /*
1035          * Our parent may run at a different priority, just set us to normal
1036          */
1037         set_user_nice(current, 0);
1038
1039         trace_writeback_thread_start(bdi);
1040
1041         while (!kthread_freezable_should_stop(NULL)) {
1042                 /*
1043                  * Remove own delayed wake-up timer, since we are already awake
1044                  * and we'll take care of the preriodic write-back.
1045                  */
1046                 del_timer(&wb->wakeup_timer);
1047
1048                 pages_written = wb_do_writeback(wb, 0);
1049
1050                 trace_writeback_pages_written(pages_written);
1051
1052                 if (pages_written)
1053                         wb->last_active = jiffies;
1054
1055                 set_current_state(TASK_INTERRUPTIBLE);
1056                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
1057                         __set_current_state(TASK_RUNNING);
1058                         continue;
1059                 }
1060
1061                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1062                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
1063                 else {
1064                         /*
1065                          * We have nothing to do, so can go sleep without any
1066                          * timeout and save power. When a work is queued or
1067                          * something is made dirty - we will be woken up.
1068                          */
1069                         schedule();
1070                 }
1071         }
1072
1073         /* Flush any work that raced with us exiting */
1074         if (!list_empty(&bdi->work_list))
1075                 wb_do_writeback(wb, 1);
1076
1077         trace_writeback_thread_stop(bdi);
1078         return 0;
1079 }
1080
1081
1082 /*
1083  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1084  * the whole world.
1085  */
1086 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1087 {
1088         struct backing_dev_info *bdi;
1089
1090         if (!nr_pages) {
1091                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1092                                 global_page_state(NR_UNSTABLE_NFS);
1093         }
1094
1095         rcu_read_lock();
1096         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1097                 if (!bdi_has_dirty_io(bdi))
1098                         continue;
1099                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1100         }
1101         rcu_read_unlock();
1102 }
1103
1104 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1105 {
1106         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1107                 struct dentry *dentry;
1108                 const char *name = "?";
1109
1110                 dentry = d_find_alias(inode);
1111                 if (dentry) {
1112                         spin_lock(&dentry->d_lock);
1113                         name = (const char *) dentry->d_name.name;
1114                 }
1115                 printk(KERN_DEBUG
1116                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1117                        current->comm, task_pid_nr(current), inode->i_ino,
1118                        name, inode->i_sb->s_id);
1119                 if (dentry) {
1120                         spin_unlock(&dentry->d_lock);
1121                         dput(dentry);
1122                 }
1123         }
1124 }
1125
1126 /**
1127  *      __mark_inode_dirty -    internal function
1128  *      @inode: inode to mark
1129  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1130  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1131  *      mark_inode_dirty_sync.
1132  *
1133  * Put the inode on the super block's dirty list.
1134  *
1135  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1136  * dirty list only if it is hashed or if it refers to a blockdev.
1137  * If it was not hashed, it will never be added to the dirty list
1138  * even if it is later hashed, as it will have been marked dirty already.
1139  *
1140  * In short, make sure you hash any inodes _before_ you start marking
1141  * them dirty.
1142  *
1143  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1144  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1145  * the kernel-internal blockdev inode represents the dirtying time of the
1146  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1147  * page->mapping->host, so the page-dirtying time is recorded in the internal
1148  * blockdev inode.
1149  */
1150 void __mark_inode_dirty(struct inode *inode, int flags)
1151 {
1152         struct super_block *sb = inode->i_sb;
1153         struct backing_dev_info *bdi = NULL;
1154
1155         /*
1156          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1157          * dirty the inode itself
1158          */
1159         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1160                 if (sb->s_op->dirty_inode)
1161                         sb->s_op->dirty_inode(inode, flags);
1162         }
1163
1164         /*
1165          * make sure that changes are seen by all cpus before we test i_state
1166          * -- mikulas
1167          */
1168         smp_mb();
1169
1170         /* avoid the locking if we can */
1171         if ((inode->i_state & flags) == flags)
1172                 return;
1173
1174         if (unlikely(block_dump))
1175                 block_dump___mark_inode_dirty(inode);
1176
1177         spin_lock(&inode->i_lock);
1178         if ((inode->i_state & flags) != flags) {
1179                 const int was_dirty = inode->i_state & I_DIRTY;
1180
1181                 inode->i_state |= flags;
1182
1183                 /*
1184                  * If the inode is being synced, just update its dirty state.
1185                  * The unlocker will place the inode on the appropriate
1186                  * superblock list, based upon its state.
1187                  */
1188                 if (inode->i_state & I_SYNC)
1189                         goto out_unlock_inode;
1190
1191                 /*
1192                  * Only add valid (hashed) inodes to the superblock's
1193                  * dirty list.  Add blockdev inodes as well.
1194                  */
1195                 if (!S_ISBLK(inode->i_mode)) {
1196                         if (inode_unhashed(inode))
1197                                 goto out_unlock_inode;
1198                 }
1199                 if (inode->i_state & I_FREEING)
1200                         goto out_unlock_inode;
1201
1202                 /*
1203                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1204                  * reposition it (that would break b_dirty time-ordering).
1205                  */
1206                 if (!was_dirty) {
1207                         bool wakeup_bdi = false;
1208                         bdi = inode_to_bdi(inode);
1209
1210                         if (bdi_cap_writeback_dirty(bdi)) {
1211                                 WARN(!test_bit(BDI_registered, &bdi->state),
1212                                      "bdi-%s not registered\n", bdi->name);
1213
1214                                 /*
1215                                  * If this is the first dirty inode for this
1216                                  * bdi, we have to wake-up the corresponding
1217                                  * bdi thread to make sure background
1218                                  * write-back happens later.
1219                                  */
1220                                 if (!wb_has_dirty_io(&bdi->wb))
1221                                         wakeup_bdi = true;
1222                         }
1223
1224                         spin_unlock(&inode->i_lock);
1225                         spin_lock(&bdi->wb.list_lock);
1226                         inode->dirtied_when = jiffies;
1227                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1228                         spin_unlock(&bdi->wb.list_lock);
1229
1230                         if (wakeup_bdi)
1231                                 bdi_wakeup_thread_delayed(bdi);
1232                         return;
1233                 }
1234         }
1235 out_unlock_inode:
1236         spin_unlock(&inode->i_lock);
1237
1238 }
1239 EXPORT_SYMBOL(__mark_inode_dirty);
1240
1241 static void wait_sb_inodes(struct super_block *sb)
1242 {
1243         struct inode *inode, *old_inode = NULL;
1244
1245         /*
1246          * We need to be protected against the filesystem going from
1247          * r/o to r/w or vice versa.
1248          */
1249         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1250
1251         spin_lock(&inode_sb_list_lock);
1252
1253         /*
1254          * Data integrity sync. Must wait for all pages under writeback,
1255          * because there may have been pages dirtied before our sync
1256          * call, but which had writeout started before we write it out.
1257          * In which case, the inode may not be on the dirty list, but
1258          * we still have to wait for that writeout.
1259          */
1260         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1261                 struct address_space *mapping = inode->i_mapping;
1262
1263                 spin_lock(&inode->i_lock);
1264                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1265                     (mapping->nrpages == 0)) {
1266                         spin_unlock(&inode->i_lock);
1267                         continue;
1268                 }
1269                 __iget(inode);
1270                 spin_unlock(&inode->i_lock);
1271                 spin_unlock(&inode_sb_list_lock);
1272
1273                 /*
1274                  * We hold a reference to 'inode' so it couldn't have been
1275                  * removed from s_inodes list while we dropped the
1276                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1277                  * be holding the last reference and we cannot iput it under
1278                  * inode_sb_list_lock. So we keep the reference and iput it
1279                  * later.
1280                  */
1281                 iput(old_inode);
1282                 old_inode = inode;
1283
1284                 filemap_fdatawait(mapping);
1285
1286                 cond_resched();
1287
1288                 spin_lock(&inode_sb_list_lock);
1289         }
1290         spin_unlock(&inode_sb_list_lock);
1291         iput(old_inode);
1292 }
1293
1294 /**
1295  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1296  * @sb: the superblock
1297  * @nr: the number of pages to write
1298  * @reason: reason why some writeback work initiated
1299  *
1300  * Start writeback on some inodes on this super_block. No guarantees are made
1301  * on how many (if any) will be written, and this function does not wait
1302  * for IO completion of submitted IO.
1303  */
1304 void writeback_inodes_sb_nr(struct super_block *sb,
1305                             unsigned long nr,
1306                             enum wb_reason reason)
1307 {
1308         DECLARE_COMPLETION_ONSTACK(done);
1309         struct wb_writeback_work work = {
1310                 .sb                     = sb,
1311                 .sync_mode              = WB_SYNC_NONE,
1312                 .tagged_writepages      = 1,
1313                 .done                   = &done,
1314                 .nr_pages               = nr,
1315                 .reason                 = reason,
1316         };
1317
1318         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1319         bdi_queue_work(sb->s_bdi, &work);
1320         wait_for_completion(&done);
1321 }
1322 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1323
1324 /**
1325  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1326  * @sb: the superblock
1327  * @reason: reason why some writeback work was initiated
1328  *
1329  * Start writeback on some inodes on this super_block. No guarantees are made
1330  * on how many (if any) will be written, and this function does not wait
1331  * for IO completion of submitted IO.
1332  */
1333 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1334 {
1335         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1336 }
1337 EXPORT_SYMBOL(writeback_inodes_sb);
1338
1339 /**
1340  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1341  * @sb: the superblock
1342  * @reason: reason why some writeback work was initiated
1343  *
1344  * Invoke writeback_inodes_sb if no writeback is currently underway.
1345  * Returns 1 if writeback was started, 0 if not.
1346  */
1347 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1348 {
1349         if (!writeback_in_progress(sb->s_bdi)) {
1350                 down_read(&sb->s_umount);
1351                 writeback_inodes_sb(sb, reason);
1352                 up_read(&sb->s_umount);
1353                 return 1;
1354         } else
1355                 return 0;
1356 }
1357 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1358
1359 /**
1360  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1361  * @sb: the superblock
1362  * @nr: the number of pages to write
1363  * @reason: reason why some writeback work was initiated
1364  *
1365  * Invoke writeback_inodes_sb if no writeback is currently underway.
1366  * Returns 1 if writeback was started, 0 if not.
1367  */
1368 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1369                                    unsigned long nr,
1370                                    enum wb_reason reason)
1371 {
1372         if (!writeback_in_progress(sb->s_bdi)) {
1373                 down_read(&sb->s_umount);
1374                 writeback_inodes_sb_nr(sb, nr, reason);
1375                 up_read(&sb->s_umount);
1376                 return 1;
1377         } else
1378                 return 0;
1379 }
1380 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1381
1382 /**
1383  * sync_inodes_sb       -       sync sb inode pages
1384  * @sb: the superblock
1385  *
1386  * This function writes and waits on any dirty inode belonging to this
1387  * super_block.
1388  */
1389 void sync_inodes_sb(struct super_block *sb)
1390 {
1391         DECLARE_COMPLETION_ONSTACK(done);
1392         struct wb_writeback_work work = {
1393                 .sb             = sb,
1394                 .sync_mode      = WB_SYNC_ALL,
1395                 .nr_pages       = LONG_MAX,
1396                 .range_cyclic   = 0,
1397                 .done           = &done,
1398                 .reason         = WB_REASON_SYNC,
1399         };
1400
1401         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1402
1403         bdi_queue_work(sb->s_bdi, &work);
1404         wait_for_completion(&done);
1405
1406         wait_sb_inodes(sb);
1407 }
1408 EXPORT_SYMBOL(sync_inodes_sb);
1409
1410 /**
1411  * write_inode_now      -       write an inode to disk
1412  * @inode: inode to write to disk
1413  * @sync: whether the write should be synchronous or not
1414  *
1415  * This function commits an inode to disk immediately if it is dirty. This is
1416  * primarily needed by knfsd.
1417  *
1418  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1419  */
1420 int write_inode_now(struct inode *inode, int sync)
1421 {
1422         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1423         struct writeback_control wbc = {
1424                 .nr_to_write = LONG_MAX,
1425                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1426                 .range_start = 0,
1427                 .range_end = LLONG_MAX,
1428         };
1429
1430         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1431                 wbc.nr_to_write = 0;
1432
1433         might_sleep();
1434         return writeback_single_inode(inode, wb, &wbc);
1435 }
1436 EXPORT_SYMBOL(write_inode_now);
1437
1438 /**
1439  * sync_inode - write an inode and its pages to disk.
1440  * @inode: the inode to sync
1441  * @wbc: controls the writeback mode
1442  *
1443  * sync_inode() will write an inode and its pages to disk.  It will also
1444  * correctly update the inode on its superblock's dirty inode lists and will
1445  * update inode->i_state.
1446  *
1447  * The caller must have a ref on the inode.
1448  */
1449 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1450 {
1451         return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1452 }
1453 EXPORT_SYMBOL(sync_inode);
1454
1455 /**
1456  * sync_inode_metadata - write an inode to disk
1457  * @inode: the inode to sync
1458  * @wait: wait for I/O to complete.
1459  *
1460  * Write an inode to disk and adjust its dirty state after completion.
1461  *
1462  * Note: only writes the actual inode, no associated data or other metadata.
1463  */
1464 int sync_inode_metadata(struct inode *inode, int wait)
1465 {
1466         struct writeback_control wbc = {
1467                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1468                 .nr_to_write = 0, /* metadata-only */
1469         };
1470
1471         return sync_inode(inode, &wbc);
1472 }
1473 EXPORT_SYMBOL(sync_inode_metadata);