arm64: alternative: Provide if/else/endif assembler macros
[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/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include "internal.h"
30
31 /*
32  * 4MB minimal write chunk size
33  */
34 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
35
36 /*
37  * Passed into wb_writeback(), essentially a subset of writeback_control
38  */
39 struct wb_writeback_work {
40         long nr_pages;
41         struct super_block *sb;
42         unsigned long *older_than_this;
43         enum writeback_sync_modes sync_mode;
44         unsigned int tagged_writepages:1;
45         unsigned int for_kupdate:1;
46         unsigned int range_cyclic:1;
47         unsigned int for_background:1;
48         enum wb_reason reason;          /* why was writeback initiated? */
49
50         struct list_head list;          /* pending work list */
51         struct completion *done;        /* set if the caller waits */
52 };
53
54 /**
55  * writeback_in_progress - determine whether there is writeback in progress
56  * @bdi: the device's backing_dev_info structure.
57  *
58  * Determine whether there is writeback waiting to be handled against a
59  * backing device.
60  */
61 int writeback_in_progress(struct backing_dev_info *bdi)
62 {
63         return test_bit(BDI_writeback_running, &bdi->state);
64 }
65 EXPORT_SYMBOL(writeback_in_progress);
66
67 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
68 {
69         struct super_block *sb = inode->i_sb;
70
71         if (strcmp(sb->s_type->name, "bdev") == 0)
72                 return inode->i_mapping->backing_dev_info;
73
74         return sb->s_bdi;
75 }
76
77 static inline struct inode *wb_inode(struct list_head *head)
78 {
79         return list_entry(head, struct inode, i_wb_list);
80 }
81
82 /*
83  * Include the creation of the trace points after defining the
84  * wb_writeback_work structure and inline functions so that the definition
85  * remains local to this file.
86  */
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/writeback.h>
89
90 static void bdi_wakeup_thread(struct backing_dev_info *bdi)
91 {
92         spin_lock_bh(&bdi->wb_lock);
93         if (test_bit(BDI_registered, &bdi->state))
94                 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
95         spin_unlock_bh(&bdi->wb_lock);
96 }
97
98 static void bdi_queue_work(struct backing_dev_info *bdi,
99                            struct wb_writeback_work *work)
100 {
101         trace_writeback_queue(bdi, work);
102
103         spin_lock_bh(&bdi->wb_lock);
104         if (!test_bit(BDI_registered, &bdi->state)) {
105                 if (work->done)
106                         complete(work->done);
107                 goto out_unlock;
108         }
109         list_add_tail(&work->list, &bdi->work_list);
110         mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
111 out_unlock:
112         spin_unlock_bh(&bdi->wb_lock);
113 }
114
115 static void
116 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
117                       bool range_cyclic, enum wb_reason reason)
118 {
119         struct wb_writeback_work *work;
120
121         /*
122          * This is WB_SYNC_NONE writeback, so if allocation fails just
123          * wakeup the thread for old dirty data writeback
124          */
125         work = kzalloc(sizeof(*work), GFP_ATOMIC);
126         if (!work) {
127                 trace_writeback_nowork(bdi);
128                 bdi_wakeup_thread(bdi);
129                 return;
130         }
131
132         work->sync_mode = WB_SYNC_NONE;
133         work->nr_pages  = nr_pages;
134         work->range_cyclic = range_cyclic;
135         work->reason    = reason;
136
137         bdi_queue_work(bdi, work);
138 }
139
140 /**
141  * bdi_start_writeback - start writeback
142  * @bdi: the backing device to write from
143  * @nr_pages: the number of pages to write
144  * @reason: reason why some writeback work was initiated
145  *
146  * Description:
147  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
148  *   started when this function returns, we make no guarantees on
149  *   completion. Caller need not hold sb s_umount semaphore.
150  *
151  */
152 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
153                         enum wb_reason reason)
154 {
155         __bdi_start_writeback(bdi, nr_pages, true, reason);
156 }
157
158 /**
159  * bdi_start_background_writeback - start background writeback
160  * @bdi: the backing device to write from
161  *
162  * Description:
163  *   This makes sure WB_SYNC_NONE background writeback happens. When
164  *   this function returns, it is only guaranteed that for given BDI
165  *   some IO is happening if we are over background dirty threshold.
166  *   Caller need not hold sb s_umount semaphore.
167  */
168 void bdi_start_background_writeback(struct backing_dev_info *bdi)
169 {
170         /*
171          * We just wake up the flusher thread. It will perform background
172          * writeback as soon as there is no other work to do.
173          */
174         trace_writeback_wake_background(bdi);
175         bdi_wakeup_thread(bdi);
176 }
177
178 /*
179  * Remove the inode from the writeback list it is on.
180  */
181 void inode_wb_list_del(struct inode *inode)
182 {
183         struct backing_dev_info *bdi = inode_to_bdi(inode);
184
185         spin_lock(&bdi->wb.list_lock);
186         list_del_init(&inode->i_wb_list);
187         spin_unlock(&bdi->wb.list_lock);
188 }
189
190 /*
191  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
192  * furthest end of its superblock's dirty-inode list.
193  *
194  * Before stamping the inode's ->dirtied_when, we check to see whether it is
195  * already the most-recently-dirtied inode on the b_dirty list.  If that is
196  * the case then the inode must have been redirtied while it was being written
197  * out and we don't reset its dirtied_when.
198  */
199 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
200 {
201         assert_spin_locked(&wb->list_lock);
202         if (!list_empty(&wb->b_dirty)) {
203                 struct inode *tail;
204
205                 tail = wb_inode(wb->b_dirty.next);
206                 if (time_before(inode->dirtied_when, tail->dirtied_when))
207                         inode->dirtied_when = jiffies;
208         }
209         list_move(&inode->i_wb_list, &wb->b_dirty);
210 }
211
212 /*
213  * requeue inode for re-scanning after bdi->b_io list is exhausted.
214  */
215 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
216 {
217         assert_spin_locked(&wb->list_lock);
218         list_move(&inode->i_wb_list, &wb->b_more_io);
219 }
220
221 static void inode_sync_complete(struct inode *inode)
222 {
223         inode->i_state &= ~I_SYNC;
224         /* If inode is clean an unused, put it into LRU now... */
225         inode_add_lru(inode);
226         /* Waiters must see I_SYNC cleared before being woken up */
227         smp_mb();
228         wake_up_bit(&inode->i_state, __I_SYNC);
229 }
230
231 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
232 {
233         bool ret = time_after(inode->dirtied_when, t);
234 #ifndef CONFIG_64BIT
235         /*
236          * For inodes being constantly redirtied, dirtied_when can get stuck.
237          * It _appears_ to be in the future, but is actually in distant past.
238          * This test is necessary to prevent such wrapped-around relative times
239          * from permanently stopping the whole bdi writeback.
240          */
241         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
242 #endif
243         return ret;
244 }
245
246 /*
247  * Move expired (dirtied before work->older_than_this) dirty inodes from
248  * @delaying_queue to @dispatch_queue.
249  */
250 static int move_expired_inodes(struct list_head *delaying_queue,
251                                struct list_head *dispatch_queue,
252                                struct wb_writeback_work *work)
253 {
254         LIST_HEAD(tmp);
255         struct list_head *pos, *node;
256         struct super_block *sb = NULL;
257         struct inode *inode;
258         int do_sb_sort = 0;
259         int moved = 0;
260
261         while (!list_empty(delaying_queue)) {
262                 inode = wb_inode(delaying_queue->prev);
263                 if (work->older_than_this &&
264                     inode_dirtied_after(inode, *work->older_than_this))
265                         break;
266                 if (sb && sb != inode->i_sb)
267                         do_sb_sort = 1;
268                 sb = inode->i_sb;
269                 list_move(&inode->i_wb_list, &tmp);
270                 moved++;
271         }
272
273         /* just one sb in list, splice to dispatch_queue and we're done */
274         if (!do_sb_sort) {
275                 list_splice(&tmp, dispatch_queue);
276                 goto out;
277         }
278
279         /* Move inodes from one superblock together */
280         while (!list_empty(&tmp)) {
281                 sb = wb_inode(tmp.prev)->i_sb;
282                 list_for_each_prev_safe(pos, node, &tmp) {
283                         inode = wb_inode(pos);
284                         if (inode->i_sb == sb)
285                                 list_move(&inode->i_wb_list, dispatch_queue);
286                 }
287         }
288 out:
289         return moved;
290 }
291
292 /*
293  * Queue all expired dirty inodes for io, eldest first.
294  * Before
295  *         newly dirtied     b_dirty    b_io    b_more_io
296  *         =============>    gf         edc     BA
297  * After
298  *         newly dirtied     b_dirty    b_io    b_more_io
299  *         =============>    g          fBAedc
300  *                                           |
301  *                                           +--> dequeue for IO
302  */
303 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
304 {
305         int moved;
306         assert_spin_locked(&wb->list_lock);
307         list_splice_init(&wb->b_more_io, &wb->b_io);
308         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
309         trace_writeback_queue_io(wb, work, moved);
310 }
311
312 static int write_inode(struct inode *inode, struct writeback_control *wbc)
313 {
314         int ret;
315
316         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
317                 trace_writeback_write_inode_start(inode, wbc);
318                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
319                 trace_writeback_write_inode(inode, wbc);
320                 return ret;
321         }
322         return 0;
323 }
324
325 /*
326  * Wait for writeback on an inode to complete. Called with i_lock held.
327  * Caller must make sure inode cannot go away when we drop i_lock.
328  */
329 static void __inode_wait_for_writeback(struct inode *inode)
330         __releases(inode->i_lock)
331         __acquires(inode->i_lock)
332 {
333         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
334         wait_queue_head_t *wqh;
335
336         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
337         while (inode->i_state & I_SYNC) {
338                 spin_unlock(&inode->i_lock);
339                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
340                 spin_lock(&inode->i_lock);
341         }
342 }
343
344 /*
345  * Wait for writeback on an inode to complete. Caller must have inode pinned.
346  */
347 void inode_wait_for_writeback(struct inode *inode)
348 {
349         spin_lock(&inode->i_lock);
350         __inode_wait_for_writeback(inode);
351         spin_unlock(&inode->i_lock);
352 }
353
354 /*
355  * Sleep until I_SYNC is cleared. This function must be called with i_lock
356  * held and drops it. It is aimed for callers not holding any inode reference
357  * so once i_lock is dropped, inode can go away.
358  */
359 static void inode_sleep_on_writeback(struct inode *inode)
360         __releases(inode->i_lock)
361 {
362         DEFINE_WAIT(wait);
363         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
364         int sleep;
365
366         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
367         sleep = inode->i_state & I_SYNC;
368         spin_unlock(&inode->i_lock);
369         if (sleep)
370                 schedule();
371         finish_wait(wqh, &wait);
372 }
373
374 /*
375  * Find proper writeback list for the inode depending on its current state and
376  * possibly also change of its state while we were doing writeback.  Here we
377  * handle things such as livelock prevention or fairness of writeback among
378  * inodes. This function can be called only by flusher thread - noone else
379  * processes all inodes in writeback lists and requeueing inodes behind flusher
380  * thread's back can have unexpected consequences.
381  */
382 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
383                           struct writeback_control *wbc)
384 {
385         if (inode->i_state & I_FREEING)
386                 return;
387
388         /*
389          * Sync livelock prevention. Each inode is tagged and synced in one
390          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
391          * the dirty time to prevent enqueue and sync it again.
392          */
393         if ((inode->i_state & I_DIRTY) &&
394             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
395                 inode->dirtied_when = jiffies;
396
397         if (wbc->pages_skipped) {
398                 /*
399                  * writeback is not making progress due to locked
400                  * buffers. Skip this inode for now.
401                  */
402                 redirty_tail(inode, wb);
403                 return;
404         }
405
406         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
407                 /*
408                  * We didn't write back all the pages.  nfs_writepages()
409                  * sometimes bales out without doing anything.
410                  */
411                 if (wbc->nr_to_write <= 0) {
412                         /* Slice used up. Queue for next turn. */
413                         requeue_io(inode, wb);
414                 } else {
415                         /*
416                          * Writeback blocked by something other than
417                          * congestion. Delay the inode for some time to
418                          * avoid spinning on the CPU (100% iowait)
419                          * retrying writeback of the dirty page/inode
420                          * that cannot be performed immediately.
421                          */
422                         redirty_tail(inode, wb);
423                 }
424         } else if (inode->i_state & I_DIRTY) {
425                 /*
426                  * Filesystems can dirty the inode during writeback operations,
427                  * such as delayed allocation during submission or metadata
428                  * updates after data IO completion.
429                  */
430                 redirty_tail(inode, wb);
431         } else {
432                 /* The inode is clean. Remove from writeback lists. */
433                 list_del_init(&inode->i_wb_list);
434         }
435 }
436
437 /*
438  * Write out an inode and its dirty pages. Do not update the writeback list
439  * linkage. That is left to the caller. The caller is also responsible for
440  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
441  */
442 static int
443 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
444 {
445         struct address_space *mapping = inode->i_mapping;
446         long nr_to_write = wbc->nr_to_write;
447         unsigned dirty;
448         int ret;
449
450         WARN_ON(!(inode->i_state & I_SYNC));
451
452         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
453
454         ret = do_writepages(mapping, wbc);
455
456         /*
457          * Make sure to wait on the data before writing out the metadata.
458          * This is important for filesystems that modify metadata on data
459          * I/O completion.
460          */
461         if (wbc->sync_mode == WB_SYNC_ALL) {
462                 int err = filemap_fdatawait(mapping);
463                 if (ret == 0)
464                         ret = err;
465         }
466
467         /*
468          * Some filesystems may redirty the inode during the writeback
469          * due to delalloc, clear dirty metadata flags right before
470          * write_inode()
471          */
472         spin_lock(&inode->i_lock);
473
474         dirty = inode->i_state & I_DIRTY;
475         inode->i_state &= ~I_DIRTY;
476
477         /*
478          * Paired with smp_mb() in __mark_inode_dirty().  This allows
479          * __mark_inode_dirty() to test i_state without grabbing i_lock -
480          * either they see the I_DIRTY bits cleared or we see the dirtied
481          * inode.
482          *
483          * I_DIRTY_PAGES is always cleared together above even if @mapping
484          * still has dirty pages.  The flag is reinstated after smp_mb() if
485          * necessary.  This guarantees that either __mark_inode_dirty()
486          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
487          */
488         smp_mb();
489
490         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
491                 inode->i_state |= I_DIRTY_PAGES;
492
493         spin_unlock(&inode->i_lock);
494
495         /* Don't write the inode if only I_DIRTY_PAGES was set */
496         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
497                 int err = write_inode(inode, wbc);
498                 if (ret == 0)
499                         ret = err;
500         }
501         trace_writeback_single_inode(inode, wbc, nr_to_write);
502         return ret;
503 }
504
505 /*
506  * Write out an inode's dirty pages. Either the caller has an active reference
507  * on the inode or the inode has I_WILL_FREE set.
508  *
509  * This function is designed to be called for writing back one inode which
510  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
511  * and does more profound writeback list handling in writeback_sb_inodes().
512  */
513 static int
514 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
515                        struct writeback_control *wbc)
516 {
517         int ret = 0;
518
519         spin_lock(&inode->i_lock);
520         if (!atomic_read(&inode->i_count))
521                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
522         else
523                 WARN_ON(inode->i_state & I_WILL_FREE);
524
525         if (inode->i_state & I_SYNC) {
526                 if (wbc->sync_mode != WB_SYNC_ALL)
527                         goto out;
528                 /*
529                  * It's a data-integrity sync. We must wait. Since callers hold
530                  * inode reference or inode has I_WILL_FREE set, it cannot go
531                  * away under us.
532                  */
533                 __inode_wait_for_writeback(inode);
534         }
535         WARN_ON(inode->i_state & I_SYNC);
536         /*
537          * Skip inode if it is clean and we have no outstanding writeback in
538          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
539          * function since flusher thread may be doing for example sync in
540          * parallel and if we move the inode, it could get skipped. So here we
541          * make sure inode is on some writeback list and leave it there unless
542          * we have completely cleaned the inode.
543          */
544         if (!(inode->i_state & I_DIRTY) &&
545             (wbc->sync_mode != WB_SYNC_ALL ||
546              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
547                 goto out;
548         inode->i_state |= I_SYNC;
549         spin_unlock(&inode->i_lock);
550
551         ret = __writeback_single_inode(inode, wbc);
552
553         spin_lock(&wb->list_lock);
554         spin_lock(&inode->i_lock);
555         /*
556          * If inode is clean, remove it from writeback lists. Otherwise don't
557          * touch it. See comment above for explanation.
558          */
559         if (!(inode->i_state & I_DIRTY))
560                 list_del_init(&inode->i_wb_list);
561         spin_unlock(&wb->list_lock);
562         inode_sync_complete(inode);
563 out:
564         spin_unlock(&inode->i_lock);
565         return ret;
566 }
567
568 static long writeback_chunk_size(struct backing_dev_info *bdi,
569                                  struct wb_writeback_work *work)
570 {
571         long pages;
572
573         /*
574          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
575          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
576          * here avoids calling into writeback_inodes_wb() more than once.
577          *
578          * The intended call sequence for WB_SYNC_ALL writeback is:
579          *
580          *      wb_writeback()
581          *          writeback_sb_inodes()       <== called only once
582          *              write_cache_pages()     <== called once for each inode
583          *                   (quickly) tag currently dirty pages
584          *                   (maybe slowly) sync all tagged pages
585          */
586         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
587                 pages = LONG_MAX;
588         else {
589                 pages = min(bdi->avg_write_bandwidth / 2,
590                             global_dirty_limit / DIRTY_SCOPE);
591                 pages = min(pages, work->nr_pages);
592                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
593                                    MIN_WRITEBACK_PAGES);
594         }
595
596         return pages;
597 }
598
599 /*
600  * Write a portion of b_io inodes which belong to @sb.
601  *
602  * Return the number of pages and/or inodes written.
603  */
604 static long writeback_sb_inodes(struct super_block *sb,
605                                 struct bdi_writeback *wb,
606                                 struct wb_writeback_work *work)
607 {
608         struct writeback_control wbc = {
609                 .sync_mode              = work->sync_mode,
610                 .tagged_writepages      = work->tagged_writepages,
611                 .for_kupdate            = work->for_kupdate,
612                 .for_background         = work->for_background,
613                 .range_cyclic           = work->range_cyclic,
614                 .range_start            = 0,
615                 .range_end              = LLONG_MAX,
616         };
617         unsigned long start_time = jiffies;
618         long write_chunk;
619         long wrote = 0;  /* count both pages and inodes */
620
621         while (!list_empty(&wb->b_io)) {
622                 struct inode *inode = wb_inode(wb->b_io.prev);
623
624                 if (inode->i_sb != sb) {
625                         if (work->sb) {
626                                 /*
627                                  * We only want to write back data for this
628                                  * superblock, move all inodes not belonging
629                                  * to it back onto the dirty list.
630                                  */
631                                 redirty_tail(inode, wb);
632                                 continue;
633                         }
634
635                         /*
636                          * The inode belongs to a different superblock.
637                          * Bounce back to the caller to unpin this and
638                          * pin the next superblock.
639                          */
640                         break;
641                 }
642
643                 /*
644                  * Don't bother with new inodes or inodes being freed, first
645                  * kind does not need periodic writeout yet, and for the latter
646                  * kind writeout is handled by the freer.
647                  */
648                 spin_lock(&inode->i_lock);
649                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
650                         spin_unlock(&inode->i_lock);
651                         redirty_tail(inode, wb);
652                         continue;
653                 }
654                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
655                         /*
656                          * If this inode is locked for writeback and we are not
657                          * doing writeback-for-data-integrity, move it to
658                          * b_more_io so that writeback can proceed with the
659                          * other inodes on s_io.
660                          *
661                          * We'll have another go at writing back this inode
662                          * when we completed a full scan of b_io.
663                          */
664                         spin_unlock(&inode->i_lock);
665                         requeue_io(inode, wb);
666                         trace_writeback_sb_inodes_requeue(inode);
667                         continue;
668                 }
669                 spin_unlock(&wb->list_lock);
670
671                 /*
672                  * We already requeued the inode if it had I_SYNC set and we
673                  * are doing WB_SYNC_NONE writeback. So this catches only the
674                  * WB_SYNC_ALL case.
675                  */
676                 if (inode->i_state & I_SYNC) {
677                         /* Wait for I_SYNC. This function drops i_lock... */
678                         inode_sleep_on_writeback(inode);
679                         /* Inode may be gone, start again */
680                         spin_lock(&wb->list_lock);
681                         continue;
682                 }
683                 inode->i_state |= I_SYNC;
684                 spin_unlock(&inode->i_lock);
685
686                 write_chunk = writeback_chunk_size(wb->bdi, work);
687                 wbc.nr_to_write = write_chunk;
688                 wbc.pages_skipped = 0;
689
690                 /*
691                  * We use I_SYNC to pin the inode in memory. While it is set
692                  * evict_inode() will wait so the inode cannot be freed.
693                  */
694                 __writeback_single_inode(inode, &wbc);
695
696                 work->nr_pages -= write_chunk - wbc.nr_to_write;
697                 wrote += write_chunk - wbc.nr_to_write;
698                 spin_lock(&wb->list_lock);
699                 spin_lock(&inode->i_lock);
700                 if (!(inode->i_state & I_DIRTY))
701                         wrote++;
702                 requeue_inode(inode, wb, &wbc);
703                 inode_sync_complete(inode);
704                 spin_unlock(&inode->i_lock);
705                 cond_resched_lock(&wb->list_lock);
706                 /*
707                  * bail out to wb_writeback() often enough to check
708                  * background threshold and other termination conditions.
709                  */
710                 if (wrote) {
711                         if (time_is_before_jiffies(start_time + HZ / 10UL))
712                                 break;
713                         if (work->nr_pages <= 0)
714                                 break;
715                 }
716         }
717         return wrote;
718 }
719
720 static long __writeback_inodes_wb(struct bdi_writeback *wb,
721                                   struct wb_writeback_work *work)
722 {
723         unsigned long start_time = jiffies;
724         long wrote = 0;
725
726         while (!list_empty(&wb->b_io)) {
727                 struct inode *inode = wb_inode(wb->b_io.prev);
728                 struct super_block *sb = inode->i_sb;
729
730                 if (!grab_super_passive(sb)) {
731                         /*
732                          * grab_super_passive() may fail consistently due to
733                          * s_umount being grabbed by someone else. Don't use
734                          * requeue_io() to avoid busy retrying the inode/sb.
735                          */
736                         redirty_tail(inode, wb);
737                         continue;
738                 }
739                 wrote += writeback_sb_inodes(sb, wb, work);
740                 drop_super(sb);
741
742                 /* refer to the same tests at the end of writeback_sb_inodes */
743                 if (wrote) {
744                         if (time_is_before_jiffies(start_time + HZ / 10UL))
745                                 break;
746                         if (work->nr_pages <= 0)
747                                 break;
748                 }
749         }
750         /* Leave any unwritten inodes on b_io */
751         return wrote;
752 }
753
754 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
755                                 enum wb_reason reason)
756 {
757         struct wb_writeback_work work = {
758                 .nr_pages       = nr_pages,
759                 .sync_mode      = WB_SYNC_NONE,
760                 .range_cyclic   = 1,
761                 .reason         = reason,
762         };
763
764         spin_lock(&wb->list_lock);
765         if (list_empty(&wb->b_io))
766                 queue_io(wb, &work);
767         __writeback_inodes_wb(wb, &work);
768         spin_unlock(&wb->list_lock);
769
770         return nr_pages - work.nr_pages;
771 }
772
773 static bool over_bground_thresh(struct backing_dev_info *bdi)
774 {
775         unsigned long background_thresh, dirty_thresh;
776
777         global_dirty_limits(&background_thresh, &dirty_thresh);
778
779         if (global_page_state(NR_FILE_DIRTY) +
780             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
781                 return true;
782
783         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
784                                 bdi_dirty_limit(bdi, background_thresh))
785                 return true;
786
787         return false;
788 }
789
790 /*
791  * Called under wb->list_lock. If there are multiple wb per bdi,
792  * only the flusher working on the first wb should do it.
793  */
794 static void wb_update_bandwidth(struct bdi_writeback *wb,
795                                 unsigned long start_time)
796 {
797         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
798 }
799
800 /*
801  * Explicit flushing or periodic writeback of "old" data.
802  *
803  * Define "old": the first time one of an inode's pages is dirtied, we mark the
804  * dirtying-time in the inode's address_space.  So this periodic writeback code
805  * just walks the superblock inode list, writing back any inodes which are
806  * older than a specific point in time.
807  *
808  * Try to run once per dirty_writeback_interval.  But if a writeback event
809  * takes longer than a dirty_writeback_interval interval, then leave a
810  * one-second gap.
811  *
812  * older_than_this takes precedence over nr_to_write.  So we'll only write back
813  * all dirty pages if they are all attached to "old" mappings.
814  */
815 static long wb_writeback(struct bdi_writeback *wb,
816                          struct wb_writeback_work *work)
817 {
818         unsigned long wb_start = jiffies;
819         long nr_pages = work->nr_pages;
820         unsigned long oldest_jif;
821         struct inode *inode;
822         long progress;
823
824         oldest_jif = jiffies;
825         work->older_than_this = &oldest_jif;
826
827         spin_lock(&wb->list_lock);
828         for (;;) {
829                 /*
830                  * Stop writeback when nr_pages has been consumed
831                  */
832                 if (work->nr_pages <= 0)
833                         break;
834
835                 /*
836                  * Background writeout and kupdate-style writeback may
837                  * run forever. Stop them if there is other work to do
838                  * so that e.g. sync can proceed. They'll be restarted
839                  * after the other works are all done.
840                  */
841                 if ((work->for_background || work->for_kupdate) &&
842                     !list_empty(&wb->bdi->work_list))
843                         break;
844
845                 /*
846                  * For background writeout, stop when we are below the
847                  * background dirty threshold
848                  */
849                 if (work->for_background && !over_bground_thresh(wb->bdi))
850                         break;
851
852                 /*
853                  * Kupdate and background works are special and we want to
854                  * include all inodes that need writing. Livelock avoidance is
855                  * handled by these works yielding to any other work so we are
856                  * safe.
857                  */
858                 if (work->for_kupdate) {
859                         oldest_jif = jiffies -
860                                 msecs_to_jiffies(dirty_expire_interval * 10);
861                 } else if (work->for_background)
862                         oldest_jif = jiffies;
863
864                 trace_writeback_start(wb->bdi, work);
865                 if (list_empty(&wb->b_io))
866                         queue_io(wb, work);
867                 if (work->sb)
868                         progress = writeback_sb_inodes(work->sb, wb, work);
869                 else
870                         progress = __writeback_inodes_wb(wb, work);
871                 trace_writeback_written(wb->bdi, work);
872
873                 wb_update_bandwidth(wb, wb_start);
874
875                 /*
876                  * Did we write something? Try for more
877                  *
878                  * Dirty inodes are moved to b_io for writeback in batches.
879                  * The completion of the current batch does not necessarily
880                  * mean the overall work is done. So we keep looping as long
881                  * as made some progress on cleaning pages or inodes.
882                  */
883                 if (progress)
884                         continue;
885                 /*
886                  * No more inodes for IO, bail
887                  */
888                 if (list_empty(&wb->b_more_io))
889                         break;
890                 /*
891                  * Nothing written. Wait for some inode to
892                  * become available for writeback. Otherwise
893                  * we'll just busyloop.
894                  */
895                 if (!list_empty(&wb->b_more_io))  {
896                         trace_writeback_wait(wb->bdi, work);
897                         inode = wb_inode(wb->b_more_io.prev);
898                         spin_lock(&inode->i_lock);
899                         spin_unlock(&wb->list_lock);
900                         /* This function drops i_lock... */
901                         inode_sleep_on_writeback(inode);
902                         spin_lock(&wb->list_lock);
903                 }
904         }
905         spin_unlock(&wb->list_lock);
906
907         return nr_pages - work->nr_pages;
908 }
909
910 /*
911  * Return the next wb_writeback_work struct that hasn't been processed yet.
912  */
913 static struct wb_writeback_work *
914 get_next_work_item(struct backing_dev_info *bdi)
915 {
916         struct wb_writeback_work *work = NULL;
917
918         spin_lock_bh(&bdi->wb_lock);
919         if (!list_empty(&bdi->work_list)) {
920                 work = list_entry(bdi->work_list.next,
921                                   struct wb_writeback_work, list);
922                 list_del_init(&work->list);
923         }
924         spin_unlock_bh(&bdi->wb_lock);
925         return work;
926 }
927
928 /*
929  * Add in the number of potentially dirty inodes, because each inode
930  * write can dirty pagecache in the underlying blockdev.
931  */
932 static unsigned long get_nr_dirty_pages(void)
933 {
934         return global_page_state(NR_FILE_DIRTY) +
935                 global_page_state(NR_UNSTABLE_NFS) +
936                 get_nr_dirty_inodes();
937 }
938
939 static long wb_check_background_flush(struct bdi_writeback *wb)
940 {
941         if (over_bground_thresh(wb->bdi)) {
942
943                 struct wb_writeback_work work = {
944                         .nr_pages       = LONG_MAX,
945                         .sync_mode      = WB_SYNC_NONE,
946                         .for_background = 1,
947                         .range_cyclic   = 1,
948                         .reason         = WB_REASON_BACKGROUND,
949                 };
950
951                 return wb_writeback(wb, &work);
952         }
953
954         return 0;
955 }
956
957 static long wb_check_old_data_flush(struct bdi_writeback *wb)
958 {
959         unsigned long expired;
960         long nr_pages;
961
962         /*
963          * When set to zero, disable periodic writeback
964          */
965         if (!dirty_writeback_interval)
966                 return 0;
967
968         expired = wb->last_old_flush +
969                         msecs_to_jiffies(dirty_writeback_interval * 10);
970         if (time_before(jiffies, expired))
971                 return 0;
972
973         wb->last_old_flush = jiffies;
974         nr_pages = get_nr_dirty_pages();
975
976         if (nr_pages) {
977                 struct wb_writeback_work work = {
978                         .nr_pages       = nr_pages,
979                         .sync_mode      = WB_SYNC_NONE,
980                         .for_kupdate    = 1,
981                         .range_cyclic   = 1,
982                         .reason         = WB_REASON_PERIODIC,
983                 };
984
985                 return wb_writeback(wb, &work);
986         }
987
988         return 0;
989 }
990
991 /*
992  * Retrieve work items and do the writeback they describe
993  */
994 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
995 {
996         struct backing_dev_info *bdi = wb->bdi;
997         struct wb_writeback_work *work;
998         long wrote = 0;
999
1000         set_bit(BDI_writeback_running, &wb->bdi->state);
1001         while ((work = get_next_work_item(bdi)) != NULL) {
1002                 /*
1003                  * Override sync mode, in case we must wait for completion
1004                  * because this thread is exiting now.
1005                  */
1006                 if (force_wait)
1007                         work->sync_mode = WB_SYNC_ALL;
1008
1009                 trace_writeback_exec(bdi, work);
1010
1011                 wrote += wb_writeback(wb, work);
1012
1013                 /*
1014                  * Notify the caller of completion if this is a synchronous
1015                  * work item, otherwise just free it.
1016                  */
1017                 if (work->done)
1018                         complete(work->done);
1019                 else
1020                         kfree(work);
1021         }
1022
1023         /*
1024          * Check for periodic writeback, kupdated() style
1025          */
1026         wrote += wb_check_old_data_flush(wb);
1027         wrote += wb_check_background_flush(wb);
1028         clear_bit(BDI_writeback_running, &wb->bdi->state);
1029
1030         return wrote;
1031 }
1032
1033 /*
1034  * Handle writeback of dirty data for the device backed by this bdi. Also
1035  * reschedules periodically and does kupdated style flushing.
1036  */
1037 void bdi_writeback_workfn(struct work_struct *work)
1038 {
1039         struct bdi_writeback *wb = container_of(to_delayed_work(work),
1040                                                 struct bdi_writeback, dwork);
1041         struct backing_dev_info *bdi = wb->bdi;
1042         long pages_written;
1043
1044         set_worker_desc("flush-%s", dev_name(bdi->dev));
1045         current->flags |= PF_SWAPWRITE;
1046
1047         if (likely(!current_is_workqueue_rescuer() ||
1048                    !test_bit(BDI_registered, &bdi->state))) {
1049                 /*
1050                  * The normal path.  Keep writing back @bdi until its
1051                  * work_list is empty.  Note that this path is also taken
1052                  * if @bdi is shutting down even when we're running off the
1053                  * rescuer as work_list needs to be drained.
1054                  */
1055                 do {
1056                         pages_written = wb_do_writeback(wb, 0);
1057                         trace_writeback_pages_written(pages_written);
1058                 } while (!list_empty(&bdi->work_list));
1059         } else {
1060                 /*
1061                  * bdi_wq can't get enough workers and we're running off
1062                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
1063                  * enough for efficient IO.
1064                  */
1065                 pages_written = writeback_inodes_wb(&bdi->wb, 1024,
1066                                                     WB_REASON_FORKER_THREAD);
1067                 trace_writeback_pages_written(pages_written);
1068         }
1069
1070         if (!list_empty(&bdi->work_list))
1071                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1072         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1073                 bdi_wakeup_thread_delayed(bdi);
1074
1075         current->flags &= ~PF_SWAPWRITE;
1076 }
1077
1078 /*
1079  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1080  * the whole world.
1081  */
1082 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1083 {
1084         struct backing_dev_info *bdi;
1085
1086         if (!nr_pages) {
1087                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1088                                 global_page_state(NR_UNSTABLE_NFS);
1089         }
1090
1091         rcu_read_lock();
1092         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1093                 if (!bdi_has_dirty_io(bdi))
1094                         continue;
1095                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1096         }
1097         rcu_read_unlock();
1098 }
1099
1100 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1101 {
1102         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1103                 struct dentry *dentry;
1104                 const char *name = "?";
1105
1106                 dentry = d_find_alias(inode);
1107                 if (dentry) {
1108                         spin_lock(&dentry->d_lock);
1109                         name = (const char *) dentry->d_name.name;
1110                 }
1111                 printk(KERN_DEBUG
1112                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1113                        current->comm, task_pid_nr(current), inode->i_ino,
1114                        name, inode->i_sb->s_id);
1115                 if (dentry) {
1116                         spin_unlock(&dentry->d_lock);
1117                         dput(dentry);
1118                 }
1119         }
1120 }
1121
1122 /**
1123  *      __mark_inode_dirty -    internal function
1124  *      @inode: inode to mark
1125  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1126  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1127  *      mark_inode_dirty_sync.
1128  *
1129  * Put the inode on the super block's dirty list.
1130  *
1131  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1132  * dirty list only if it is hashed or if it refers to a blockdev.
1133  * If it was not hashed, it will never be added to the dirty list
1134  * even if it is later hashed, as it will have been marked dirty already.
1135  *
1136  * In short, make sure you hash any inodes _before_ you start marking
1137  * them dirty.
1138  *
1139  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1140  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1141  * the kernel-internal blockdev inode represents the dirtying time of the
1142  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1143  * page->mapping->host, so the page-dirtying time is recorded in the internal
1144  * blockdev inode.
1145  */
1146 void __mark_inode_dirty(struct inode *inode, int flags)
1147 {
1148         struct super_block *sb = inode->i_sb;
1149         struct backing_dev_info *bdi = NULL;
1150
1151         /*
1152          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1153          * dirty the inode itself
1154          */
1155         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1156                 trace_writeback_dirty_inode_start(inode, flags);
1157
1158                 if (sb->s_op->dirty_inode)
1159                         sb->s_op->dirty_inode(inode, flags);
1160
1161                 trace_writeback_dirty_inode(inode, flags);
1162         }
1163
1164         /*
1165          * Paired with smp_mb() in __writeback_single_inode() for the
1166          * following lockless i_state test.  See there for details.
1167          */
1168         smp_mb();
1169
1170         if ((inode->i_state & flags) == flags)
1171                 return;
1172
1173         if (unlikely(block_dump > 1))
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         if (sb->s_bdi == &noop_backing_dev_info)
1318                 return;
1319         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1320         bdi_queue_work(sb->s_bdi, &work);
1321         wait_for_completion(&done);
1322 }
1323 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1324
1325 /**
1326  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1327  * @sb: the superblock
1328  * @reason: reason why some writeback work was initiated
1329  *
1330  * Start writeback on some inodes on this super_block. No guarantees are made
1331  * on how many (if any) will be written, and this function does not wait
1332  * for IO completion of submitted IO.
1333  */
1334 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1335 {
1336         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1337 }
1338 EXPORT_SYMBOL(writeback_inodes_sb);
1339
1340 /**
1341  * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1342  * @sb: the superblock
1343  * @nr: the number of pages to write
1344  * @reason: the reason of writeback
1345  *
1346  * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1347  * Returns 1 if writeback was started, 0 if not.
1348  */
1349 int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1350                                   unsigned long nr,
1351                                   enum wb_reason reason)
1352 {
1353         if (writeback_in_progress(sb->s_bdi))
1354                 return 1;
1355
1356         if (!down_read_trylock(&sb->s_umount))
1357                 return 0;
1358
1359         writeback_inodes_sb_nr(sb, nr, reason);
1360         up_read(&sb->s_umount);
1361         return 1;
1362 }
1363 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1364
1365 /**
1366  * try_to_writeback_inodes_sb - try to start writeback if none underway
1367  * @sb: the superblock
1368  * @reason: reason why some writeback work was initiated
1369  *
1370  * Implement by try_to_writeback_inodes_sb_nr()
1371  * Returns 1 if writeback was started, 0 if not.
1372  */
1373 int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1374 {
1375         return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1376 }
1377 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
1378
1379 /**
1380  * sync_inodes_sb       -       sync sb inode pages
1381  * @sb: the superblock
1382  *
1383  * This function writes and waits on any dirty inode belonging to this
1384  * super_block.
1385  */
1386 void sync_inodes_sb(struct super_block *sb)
1387 {
1388         DECLARE_COMPLETION_ONSTACK(done);
1389         struct wb_writeback_work work = {
1390                 .sb             = sb,
1391                 .sync_mode      = WB_SYNC_ALL,
1392                 .nr_pages       = LONG_MAX,
1393                 .range_cyclic   = 0,
1394                 .done           = &done,
1395                 .reason         = WB_REASON_SYNC,
1396         };
1397
1398         /* Nothing to do? */
1399         if (sb->s_bdi == &noop_backing_dev_info)
1400                 return;
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);