ARM: tegra11: clock: Update EMC clock mux
[linux-2.6.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/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         /*
235          * Prevent speculative execution through
236          * spin_unlock(&wb->list_lock);
237          */
238
239         smp_mb();
240         wake_up_bit(&inode->i_state, __I_SYNC);
241 }
242
243 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
244 {
245         bool ret = time_after(inode->dirtied_when, t);
246 #ifndef CONFIG_64BIT
247         /*
248          * For inodes being constantly redirtied, dirtied_when can get stuck.
249          * It _appears_ to be in the future, but is actually in distant past.
250          * This test is necessary to prevent such wrapped-around relative times
251          * from permanently stopping the whole bdi writeback.
252          */
253         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
254 #endif
255         return ret;
256 }
257
258 /*
259  * Move expired (dirtied after work->older_than_this) dirty inodes from
260  * @delaying_queue to @dispatch_queue.
261  */
262 static int move_expired_inodes(struct list_head *delaying_queue,
263                                struct list_head *dispatch_queue,
264                                struct wb_writeback_work *work)
265 {
266         LIST_HEAD(tmp);
267         struct list_head *pos, *node;
268         struct super_block *sb = NULL;
269         struct inode *inode;
270         int do_sb_sort = 0;
271         int moved = 0;
272
273         while (!list_empty(delaying_queue)) {
274                 inode = wb_inode(delaying_queue->prev);
275                 if (work->older_than_this &&
276                     inode_dirtied_after(inode, *work->older_than_this))
277                         break;
278                 if (sb && sb != inode->i_sb)
279                         do_sb_sort = 1;
280                 sb = inode->i_sb;
281                 list_move(&inode->i_wb_list, &tmp);
282                 moved++;
283         }
284
285         /* just one sb in list, splice to dispatch_queue and we're done */
286         if (!do_sb_sort) {
287                 list_splice(&tmp, dispatch_queue);
288                 goto out;
289         }
290
291         /* Move inodes from one superblock together */
292         while (!list_empty(&tmp)) {
293                 sb = wb_inode(tmp.prev)->i_sb;
294                 list_for_each_prev_safe(pos, node, &tmp) {
295                         inode = wb_inode(pos);
296                         if (inode->i_sb == sb)
297                                 list_move(&inode->i_wb_list, dispatch_queue);
298                 }
299         }
300 out:
301         return moved;
302 }
303
304 /*
305  * Queue all expired dirty inodes for io, eldest first.
306  * Before
307  *         newly dirtied     b_dirty    b_io    b_more_io
308  *         =============>    gf         edc     BA
309  * After
310  *         newly dirtied     b_dirty    b_io    b_more_io
311  *         =============>    g          fBAedc
312  *                                           |
313  *                                           +--> dequeue for IO
314  */
315 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
316 {
317         int moved;
318         assert_spin_locked(&wb->list_lock);
319         list_splice_init(&wb->b_more_io, &wb->b_io);
320         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
321         trace_writeback_queue_io(wb, work, moved);
322 }
323
324 static int write_inode(struct inode *inode, struct writeback_control *wbc)
325 {
326         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
327                 return inode->i_sb->s_op->write_inode(inode, wbc);
328         return 0;
329 }
330
331 /*
332  * Wait for writeback on an inode to complete.
333  */
334 static void inode_wait_for_writeback(struct inode *inode,
335                                      struct bdi_writeback *wb)
336 {
337         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
338         wait_queue_head_t *wqh;
339
340         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
341         while (inode->i_state & I_SYNC) {
342                 spin_unlock(&inode->i_lock);
343                 spin_unlock(&wb->list_lock);
344                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
345                 spin_lock(&wb->list_lock);
346                 spin_lock(&inode->i_lock);
347         }
348 }
349
350 /*
351  * Write out an inode's dirty pages.  Called under wb->list_lock and
352  * inode->i_lock.  Either the caller has an active reference on the inode or
353  * the inode has I_WILL_FREE set.
354  *
355  * If `wait' is set, wait on the writeout.
356  *
357  * The whole writeout design is quite complex and fragile.  We want to avoid
358  * starvation of particular inodes when others are being redirtied, prevent
359  * livelocks, etc.
360  */
361 static int
362 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
363                        struct writeback_control *wbc)
364 {
365         struct address_space *mapping = inode->i_mapping;
366         long nr_to_write = wbc->nr_to_write;
367         unsigned dirty;
368         int ret;
369
370         assert_spin_locked(&wb->list_lock);
371         assert_spin_locked(&inode->i_lock);
372
373         if (!atomic_read(&inode->i_count))
374                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
375         else
376                 WARN_ON(inode->i_state & I_WILL_FREE);
377
378         if (inode->i_state & I_SYNC) {
379                 /*
380                  * If this inode is locked for writeback and we are not doing
381                  * writeback-for-data-integrity, move it to b_more_io so that
382                  * writeback can proceed with the other inodes on s_io.
383                  *
384                  * We'll have another go at writing back this inode when we
385                  * completed a full scan of b_io.
386                  */
387                 if (wbc->sync_mode != WB_SYNC_ALL) {
388                         requeue_io(inode, wb);
389                         trace_writeback_single_inode_requeue(inode, wbc,
390                                                              nr_to_write);
391                         return 0;
392                 }
393
394                 /*
395                  * It's a data-integrity sync.  We must wait.
396                  */
397                 inode_wait_for_writeback(inode, wb);
398         }
399
400         BUG_ON(inode->i_state & I_SYNC);
401
402         /* Set I_SYNC, reset I_DIRTY_PAGES */
403         inode->i_state |= I_SYNC;
404         inode->i_state &= ~I_DIRTY_PAGES;
405         spin_unlock(&inode->i_lock);
406         spin_unlock(&wb->list_lock);
407
408         ret = do_writepages(mapping, wbc);
409
410         /*
411          * Make sure to wait on the data before writing out the metadata.
412          * This is important for filesystems that modify metadata on data
413          * I/O completion.
414          */
415         if (wbc->sync_mode == WB_SYNC_ALL) {
416                 int err = filemap_fdatawait(mapping);
417                 if (ret == 0)
418                         ret = err;
419         }
420
421         /*
422          * Some filesystems may redirty the inode during the writeback
423          * due to delalloc, clear dirty metadata flags right before
424          * write_inode()
425          */
426         spin_lock(&inode->i_lock);
427         dirty = inode->i_state & I_DIRTY;
428         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
429         spin_unlock(&inode->i_lock);
430         /* Don't write the inode if only I_DIRTY_PAGES was set */
431         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
432                 int err = write_inode(inode, wbc);
433                 if (ret == 0)
434                         ret = err;
435         }
436
437         spin_lock(&wb->list_lock);
438         spin_lock(&inode->i_lock);
439         inode->i_state &= ~I_SYNC;
440         if (!(inode->i_state & I_FREEING)) {
441                 /*
442                  * Sync livelock prevention. Each inode is tagged and synced in
443                  * one shot. If still dirty, it will be redirty_tail()'ed below.
444                  * Update the dirty time to prevent enqueue and sync it again.
445                  */
446                 if ((inode->i_state & I_DIRTY) &&
447                     (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
448                         inode->dirtied_when = jiffies;
449
450                 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
451                         /*
452                          * We didn't write back all the pages.  nfs_writepages()
453                          * sometimes bales out without doing anything.
454                          */
455                         inode->i_state |= I_DIRTY_PAGES;
456                         if (wbc->nr_to_write <= 0) {
457                                 /*
458                                  * slice used up: queue for next turn
459                                  */
460                                 requeue_io(inode, wb);
461                         } else {
462                                 /*
463                                  * Writeback blocked by something other than
464                                  * congestion. Delay the inode for some time to
465                                  * avoid spinning on the CPU (100% iowait)
466                                  * retrying writeback of the dirty page/inode
467                                  * that cannot be performed immediately.
468                                  */
469                                 redirty_tail(inode, wb);
470                         }
471                 } else if (inode->i_state & I_DIRTY) {
472                         /*
473                          * Filesystems can dirty the inode during writeback
474                          * operations, such as delayed allocation during
475                          * submission or metadata updates after data IO
476                          * completion.
477                          */
478                         redirty_tail(inode, wb);
479                 } else {
480                         /*
481                          * The inode is clean.  At this point we either have
482                          * a reference to the inode or it's on it's way out.
483                          * No need to add it back to the LRU.
484                          */
485                         list_del_init(&inode->i_wb_list);
486                 }
487         }
488         inode_sync_complete(inode);
489         trace_writeback_single_inode(inode, wbc, nr_to_write);
490         return ret;
491 }
492
493 static long writeback_chunk_size(struct backing_dev_info *bdi,
494                                  struct wb_writeback_work *work)
495 {
496         long pages;
497
498         /*
499          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
500          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
501          * here avoids calling into writeback_inodes_wb() more than once.
502          *
503          * The intended call sequence for WB_SYNC_ALL writeback is:
504          *
505          *      wb_writeback()
506          *          writeback_sb_inodes()       <== called only once
507          *              write_cache_pages()     <== called once for each inode
508          *                   (quickly) tag currently dirty pages
509          *                   (maybe slowly) sync all tagged pages
510          */
511         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
512                 pages = LONG_MAX;
513         else {
514                 pages = min(bdi->avg_write_bandwidth / 2,
515                             global_dirty_limit / DIRTY_SCOPE);
516                 pages = min(pages, work->nr_pages);
517                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
518                                    MIN_WRITEBACK_PAGES);
519         }
520
521         return pages;
522 }
523
524 /*
525  * Write a portion of b_io inodes which belong to @sb.
526  *
527  * If @only_this_sb is true, then find and write all such
528  * inodes. Otherwise write only ones which go sequentially
529  * in reverse order.
530  *
531  * Return the number of pages and/or inodes written.
532  */
533 static long writeback_sb_inodes(struct super_block *sb,
534                                 struct bdi_writeback *wb,
535                                 struct wb_writeback_work *work)
536 {
537         struct writeback_control wbc = {
538                 .sync_mode              = work->sync_mode,
539                 .tagged_writepages      = work->tagged_writepages,
540                 .for_kupdate            = work->for_kupdate,
541                 .for_background         = work->for_background,
542                 .range_cyclic           = work->range_cyclic,
543                 .range_start            = 0,
544                 .range_end              = LLONG_MAX,
545         };
546         unsigned long start_time = jiffies;
547         long write_chunk;
548         long wrote = 0;  /* count both pages and inodes */
549
550         while (!list_empty(&wb->b_io)) {
551                 struct inode *inode = wb_inode(wb->b_io.prev);
552
553                 if (inode->i_sb != sb) {
554                         if (work->sb) {
555                                 /*
556                                  * We only want to write back data for this
557                                  * superblock, move all inodes not belonging
558                                  * to it back onto the dirty list.
559                                  */
560                                 redirty_tail(inode, wb);
561                                 continue;
562                         }
563
564                         /*
565                          * The inode belongs to a different superblock.
566                          * Bounce back to the caller to unpin this and
567                          * pin the next superblock.
568                          */
569                         break;
570                 }
571
572                 /*
573                  * Don't bother with new inodes or inodes beeing freed, first
574                  * kind does not need peridic writeout yet, and for the latter
575                  * kind writeout is handled by the freer.
576                  */
577                 spin_lock(&inode->i_lock);
578                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
579                         spin_unlock(&inode->i_lock);
580                         redirty_tail(inode, wb);
581                         continue;
582                 }
583                 __iget(inode);
584                 write_chunk = writeback_chunk_size(wb->bdi, work);
585                 wbc.nr_to_write = write_chunk;
586                 wbc.pages_skipped = 0;
587
588                 writeback_single_inode(inode, wb, &wbc);
589
590                 work->nr_pages -= write_chunk - wbc.nr_to_write;
591                 wrote += write_chunk - wbc.nr_to_write;
592                 if (!(inode->i_state & I_DIRTY))
593                         wrote++;
594                 if (wbc.pages_skipped) {
595                         /*
596                          * writeback is not making progress due to locked
597                          * buffers.  Skip this inode for now.
598                          */
599                         redirty_tail(inode, wb);
600                 }
601                 spin_unlock(&inode->i_lock);
602                 spin_unlock(&wb->list_lock);
603                 iput(inode);
604                 cond_resched();
605                 spin_lock(&wb->list_lock);
606                 /*
607                  * bail out to wb_writeback() often enough to check
608                  * background threshold and other termination conditions.
609                  */
610                 if (wrote) {
611                         if (time_is_before_jiffies(start_time + HZ / 10UL))
612                                 break;
613                         if (work->nr_pages <= 0)
614                                 break;
615                 }
616         }
617         return wrote;
618 }
619
620 static long __writeback_inodes_wb(struct bdi_writeback *wb,
621                                   struct wb_writeback_work *work)
622 {
623         unsigned long start_time = jiffies;
624         long wrote = 0;
625
626         while (!list_empty(&wb->b_io)) {
627                 struct inode *inode = wb_inode(wb->b_io.prev);
628                 struct super_block *sb = inode->i_sb;
629
630                 if (!grab_super_passive(sb)) {
631                         /*
632                          * grab_super_passive() may fail consistently due to
633                          * s_umount being grabbed by someone else. Don't use
634                          * requeue_io() to avoid busy retrying the inode/sb.
635                          */
636                         redirty_tail(inode, wb);
637                         continue;
638                 }
639                 wrote += writeback_sb_inodes(sb, wb, work);
640                 drop_super(sb);
641
642                 /* refer to the same tests at the end of writeback_sb_inodes */
643                 if (wrote) {
644                         if (time_is_before_jiffies(start_time + HZ / 10UL))
645                                 break;
646                         if (work->nr_pages <= 0)
647                                 break;
648                 }
649         }
650         /* Leave any unwritten inodes on b_io */
651         return wrote;
652 }
653
654 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
655                                 enum wb_reason reason)
656 {
657         struct wb_writeback_work work = {
658                 .nr_pages       = nr_pages,
659                 .sync_mode      = WB_SYNC_NONE,
660                 .range_cyclic   = 1,
661                 .reason         = reason,
662         };
663
664         spin_lock(&wb->list_lock);
665         if (list_empty(&wb->b_io))
666                 queue_io(wb, &work);
667         __writeback_inodes_wb(wb, &work);
668         spin_unlock(&wb->list_lock);
669
670         return nr_pages - work.nr_pages;
671 }
672
673 static bool over_bground_thresh(struct backing_dev_info *bdi)
674 {
675         unsigned long background_thresh, dirty_thresh;
676
677         global_dirty_limits(&background_thresh, &dirty_thresh);
678
679         if (global_page_state(NR_FILE_DIRTY) +
680             global_page_state(NR_UNSTABLE_NFS) > background_thresh)
681                 return true;
682
683         if (bdi_stat(bdi, BDI_RECLAIMABLE) >
684                                 bdi_dirty_limit(bdi, background_thresh))
685                 return true;
686
687         return false;
688 }
689
690 /*
691  * Called under wb->list_lock. If there are multiple wb per bdi,
692  * only the flusher working on the first wb should do it.
693  */
694 static void wb_update_bandwidth(struct bdi_writeback *wb,
695                                 unsigned long start_time)
696 {
697         __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
698 }
699
700 /*
701  * Explicit flushing or periodic writeback of "old" data.
702  *
703  * Define "old": the first time one of an inode's pages is dirtied, we mark the
704  * dirtying-time in the inode's address_space.  So this periodic writeback code
705  * just walks the superblock inode list, writing back any inodes which are
706  * older than a specific point in time.
707  *
708  * Try to run once per dirty_writeback_interval.  But if a writeback event
709  * takes longer than a dirty_writeback_interval interval, then leave a
710  * one-second gap.
711  *
712  * older_than_this takes precedence over nr_to_write.  So we'll only write back
713  * all dirty pages if they are all attached to "old" mappings.
714  */
715 static long wb_writeback(struct bdi_writeback *wb,
716                          struct wb_writeback_work *work)
717 {
718         unsigned long wb_start = jiffies;
719         long nr_pages = work->nr_pages;
720         unsigned long oldest_jif;
721         struct inode *inode;
722         long progress;
723
724         oldest_jif = jiffies;
725         work->older_than_this = &oldest_jif;
726
727         spin_lock(&wb->list_lock);
728         for (;;) {
729                 /*
730                  * Stop writeback when nr_pages has been consumed
731                  */
732                 if (work->nr_pages <= 0)
733                         break;
734
735                 /*
736                  * Background writeout and kupdate-style writeback may
737                  * run forever. Stop them if there is other work to do
738                  * so that e.g. sync can proceed. They'll be restarted
739                  * after the other works are all done.
740                  */
741                 if ((work->for_background || work->for_kupdate) &&
742                     !list_empty(&wb->bdi->work_list))
743                         break;
744
745                 /*
746                  * For background writeout, stop when we are below the
747                  * background dirty threshold
748                  */
749                 if (work->for_background && !over_bground_thresh(wb->bdi))
750                         break;
751
752                 /*
753                  * Kupdate and background works are special and we want to
754                  * include all inodes that need writing. Livelock avoidance is
755                  * handled by these works yielding to any other work so we are
756                  * safe.
757                  */
758                 if (work->for_kupdate) {
759                         oldest_jif = jiffies -
760                                 msecs_to_jiffies(dirty_expire_interval * 10);
761                 } else if (work->for_background)
762                         oldest_jif = jiffies;
763
764                 trace_writeback_start(wb->bdi, work);
765                 if (list_empty(&wb->b_io))
766                         queue_io(wb, work);
767                 if (work->sb)
768                         progress = writeback_sb_inodes(work->sb, wb, work);
769                 else
770                         progress = __writeback_inodes_wb(wb, work);
771                 trace_writeback_written(wb->bdi, work);
772
773                 wb_update_bandwidth(wb, wb_start);
774
775                 /*
776                  * Did we write something? Try for more
777                  *
778                  * Dirty inodes are moved to b_io for writeback in batches.
779                  * The completion of the current batch does not necessarily
780                  * mean the overall work is done. So we keep looping as long
781                  * as made some progress on cleaning pages or inodes.
782                  */
783                 if (progress)
784                         continue;
785                 /*
786                  * No more inodes for IO, bail
787                  */
788                 if (list_empty(&wb->b_more_io))
789                         break;
790                 /*
791                  * Nothing written. Wait for some inode to
792                  * become available for writeback. Otherwise
793                  * we'll just busyloop.
794                  */
795                 if (!list_empty(&wb->b_more_io))  {
796                         trace_writeback_wait(wb->bdi, work);
797                         inode = wb_inode(wb->b_more_io.prev);
798                         spin_lock(&inode->i_lock);
799                         inode_wait_for_writeback(inode, wb);
800                         spin_unlock(&inode->i_lock);
801                 }
802         }
803         spin_unlock(&wb->list_lock);
804
805         return nr_pages - work->nr_pages;
806 }
807
808 /*
809  * Return the next wb_writeback_work struct that hasn't been processed yet.
810  */
811 static struct wb_writeback_work *
812 get_next_work_item(struct backing_dev_info *bdi)
813 {
814         struct wb_writeback_work *work = NULL;
815
816         spin_lock_bh(&bdi->wb_lock);
817         if (!list_empty(&bdi->work_list)) {
818                 work = list_entry(bdi->work_list.next,
819                                   struct wb_writeback_work, list);
820                 list_del_init(&work->list);
821         }
822         spin_unlock_bh(&bdi->wb_lock);
823         return work;
824 }
825
826 /*
827  * Add in the number of potentially dirty inodes, because each inode
828  * write can dirty pagecache in the underlying blockdev.
829  */
830 static unsigned long get_nr_dirty_pages(void)
831 {
832         return global_page_state(NR_FILE_DIRTY) +
833                 global_page_state(NR_UNSTABLE_NFS) +
834                 get_nr_dirty_inodes();
835 }
836
837 static long wb_check_background_flush(struct bdi_writeback *wb)
838 {
839         if (over_bground_thresh(wb->bdi)) {
840
841                 struct wb_writeback_work work = {
842                         .nr_pages       = LONG_MAX,
843                         .sync_mode      = WB_SYNC_NONE,
844                         .for_background = 1,
845                         .range_cyclic   = 1,
846                         .reason         = WB_REASON_BACKGROUND,
847                 };
848
849                 return wb_writeback(wb, &work);
850         }
851
852         return 0;
853 }
854
855 static long wb_check_old_data_flush(struct bdi_writeback *wb)
856 {
857         unsigned long expired;
858         long nr_pages;
859
860         /*
861          * When set to zero, disable periodic writeback
862          */
863         if (!dirty_writeback_interval)
864                 return 0;
865
866         expired = wb->last_old_flush +
867                         msecs_to_jiffies(dirty_writeback_interval * 10);
868         if (time_before(jiffies, expired))
869                 return 0;
870
871         wb->last_old_flush = jiffies;
872         nr_pages = get_nr_dirty_pages();
873
874         if (nr_pages) {
875                 struct wb_writeback_work work = {
876                         .nr_pages       = nr_pages,
877                         .sync_mode      = WB_SYNC_NONE,
878                         .for_kupdate    = 1,
879                         .range_cyclic   = 1,
880                         .reason         = WB_REASON_PERIODIC,
881                 };
882
883                 return wb_writeback(wb, &work);
884         }
885
886         return 0;
887 }
888
889 /*
890  * Retrieve work items and do the writeback they describe
891  */
892 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
893 {
894         struct backing_dev_info *bdi = wb->bdi;
895         struct wb_writeback_work *work;
896         long wrote = 0;
897
898         set_bit(BDI_writeback_running, &wb->bdi->state);
899         while ((work = get_next_work_item(bdi)) != NULL) {
900                 /*
901                  * Override sync mode, in case we must wait for completion
902                  * because this thread is exiting now.
903                  */
904                 if (force_wait)
905                         work->sync_mode = WB_SYNC_ALL;
906
907                 trace_writeback_exec(bdi, work);
908
909                 wrote += wb_writeback(wb, work);
910
911                 /*
912                  * Notify the caller of completion if this is a synchronous
913                  * work item, otherwise just free it.
914                  */
915                 if (work->done)
916                         complete(work->done);
917                 else
918                         kfree(work);
919         }
920
921         /*
922          * Check for periodic writeback, kupdated() style
923          */
924         wrote += wb_check_old_data_flush(wb);
925         wrote += wb_check_background_flush(wb);
926         clear_bit(BDI_writeback_running, &wb->bdi->state);
927
928         return wrote;
929 }
930
931 /*
932  * Handle writeback of dirty data for the device backed by this bdi. Also
933  * wakes up periodically and does kupdated style flushing.
934  */
935 int bdi_writeback_thread(void *data)
936 {
937         struct bdi_writeback *wb = data;
938         struct backing_dev_info *bdi = wb->bdi;
939         long pages_written;
940
941         current->flags |= PF_SWAPWRITE;
942         set_freezable();
943         wb->last_active = jiffies;
944
945         /*
946          * Our parent may run at a different priority, just set us to normal
947          */
948         set_user_nice(current, 0);
949
950         trace_writeback_thread_start(bdi);
951
952         while (!kthread_freezable_should_stop(NULL)) {
953                 /*
954                  * Remove own delayed wake-up timer, since we are already awake
955                  * and we'll take care of the preriodic write-back.
956                  */
957                 del_timer(&wb->wakeup_timer);
958
959                 pages_written = wb_do_writeback(wb, 0);
960
961                 trace_writeback_pages_written(pages_written);
962
963                 if (pages_written)
964                         wb->last_active = jiffies;
965
966                 set_current_state(TASK_INTERRUPTIBLE);
967                 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
968                         __set_current_state(TASK_RUNNING);
969                         continue;
970                 }
971
972                 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
973                         schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
974                 else {
975                         /*
976                          * We have nothing to do, so can go sleep without any
977                          * timeout and save power. When a work is queued or
978                          * something is made dirty - we will be woken up.
979                          */
980                         schedule();
981                 }
982         }
983
984         /* Flush any work that raced with us exiting */
985         if (!list_empty(&bdi->work_list))
986                 wb_do_writeback(wb, 1);
987
988         trace_writeback_thread_stop(bdi);
989         return 0;
990 }
991
992
993 /*
994  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
995  * the whole world.
996  */
997 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
998 {
999         struct backing_dev_info *bdi;
1000
1001         if (!nr_pages) {
1002                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1003                                 global_page_state(NR_UNSTABLE_NFS);
1004         }
1005
1006         rcu_read_lock();
1007         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1008                 if (!bdi_has_dirty_io(bdi))
1009                         continue;
1010                 __bdi_start_writeback(bdi, nr_pages, false, reason);
1011         }
1012         rcu_read_unlock();
1013 }
1014
1015 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1016 {
1017         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1018                 struct dentry *dentry;
1019                 const char *name = "?";
1020
1021                 dentry = d_find_alias(inode);
1022                 if (dentry) {
1023                         spin_lock(&dentry->d_lock);
1024                         name = (const char *) dentry->d_name.name;
1025                 }
1026                 printk(KERN_DEBUG
1027                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1028                        current->comm, task_pid_nr(current), inode->i_ino,
1029                        name, inode->i_sb->s_id);
1030                 if (dentry) {
1031                         spin_unlock(&dentry->d_lock);
1032                         dput(dentry);
1033                 }
1034         }
1035 }
1036
1037 /**
1038  *      __mark_inode_dirty -    internal function
1039  *      @inode: inode to mark
1040  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1041  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1042  *      mark_inode_dirty_sync.
1043  *
1044  * Put the inode on the super block's dirty list.
1045  *
1046  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1047  * dirty list only if it is hashed or if it refers to a blockdev.
1048  * If it was not hashed, it will never be added to the dirty list
1049  * even if it is later hashed, as it will have been marked dirty already.
1050  *
1051  * In short, make sure you hash any inodes _before_ you start marking
1052  * them dirty.
1053  *
1054  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1055  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1056  * the kernel-internal blockdev inode represents the dirtying time of the
1057  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1058  * page->mapping->host, so the page-dirtying time is recorded in the internal
1059  * blockdev inode.
1060  */
1061 void __mark_inode_dirty(struct inode *inode, int flags)
1062 {
1063         struct super_block *sb = inode->i_sb;
1064         struct backing_dev_info *bdi = NULL;
1065
1066         /*
1067          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1068          * dirty the inode itself
1069          */
1070         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1071                 if (sb->s_op->dirty_inode)
1072                         sb->s_op->dirty_inode(inode, flags);
1073         }
1074
1075         /*
1076          * make sure that changes are seen by all cpus before we test i_state
1077          * -- mikulas
1078          */
1079         smp_mb();
1080
1081         /* avoid the locking if we can */
1082         if ((inode->i_state & flags) == flags)
1083                 return;
1084
1085         if (unlikely(block_dump > 1))
1086                 block_dump___mark_inode_dirty(inode);
1087
1088         spin_lock(&inode->i_lock);
1089         if ((inode->i_state & flags) != flags) {
1090                 const int was_dirty = inode->i_state & I_DIRTY;
1091
1092                 inode->i_state |= flags;
1093
1094                 /*
1095                  * If the inode is being synced, just update its dirty state.
1096                  * The unlocker will place the inode on the appropriate
1097                  * superblock list, based upon its state.
1098                  */
1099                 if (inode->i_state & I_SYNC)
1100                         goto out_unlock_inode;
1101
1102                 /*
1103                  * Only add valid (hashed) inodes to the superblock's
1104                  * dirty list.  Add blockdev inodes as well.
1105                  */
1106                 if (!S_ISBLK(inode->i_mode)) {
1107                         if (inode_unhashed(inode))
1108                                 goto out_unlock_inode;
1109                 }
1110                 if (inode->i_state & I_FREEING)
1111                         goto out_unlock_inode;
1112
1113                 /*
1114                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1115                  * reposition it (that would break b_dirty time-ordering).
1116                  */
1117                 if (!was_dirty) {
1118                         bool wakeup_bdi = false;
1119                         bdi = inode_to_bdi(inode);
1120
1121                         if (bdi_cap_writeback_dirty(bdi)) {
1122                                 WARN(!test_bit(BDI_registered, &bdi->state),
1123                                      "bdi-%s not registered\n", bdi->name);
1124
1125                                 /*
1126                                  * If this is the first dirty inode for this
1127                                  * bdi, we have to wake-up the corresponding
1128                                  * bdi thread to make sure background
1129                                  * write-back happens later.
1130                                  */
1131                                 if (!wb_has_dirty_io(&bdi->wb))
1132                                         wakeup_bdi = true;
1133                         }
1134
1135                         spin_unlock(&inode->i_lock);
1136                         spin_lock(&bdi->wb.list_lock);
1137                         inode->dirtied_when = jiffies;
1138                         list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1139                         spin_unlock(&bdi->wb.list_lock);
1140
1141                         if (wakeup_bdi)
1142                                 bdi_wakeup_thread_delayed(bdi);
1143                         return;
1144                 }
1145         }
1146 out_unlock_inode:
1147         spin_unlock(&inode->i_lock);
1148
1149 }
1150 EXPORT_SYMBOL(__mark_inode_dirty);
1151
1152 static void wait_sb_inodes(struct super_block *sb)
1153 {
1154         struct inode *inode, *old_inode = NULL;
1155
1156         /*
1157          * We need to be protected against the filesystem going from
1158          * r/o to r/w or vice versa.
1159          */
1160         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1161
1162         spin_lock(&inode_sb_list_lock);
1163
1164         /*
1165          * Data integrity sync. Must wait for all pages under writeback,
1166          * because there may have been pages dirtied before our sync
1167          * call, but which had writeout started before we write it out.
1168          * In which case, the inode may not be on the dirty list, but
1169          * we still have to wait for that writeout.
1170          */
1171         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1172                 struct address_space *mapping = inode->i_mapping;
1173
1174                 spin_lock(&inode->i_lock);
1175                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1176                     (mapping->nrpages == 0)) {
1177                         spin_unlock(&inode->i_lock);
1178                         continue;
1179                 }
1180                 __iget(inode);
1181                 spin_unlock(&inode->i_lock);
1182                 spin_unlock(&inode_sb_list_lock);
1183
1184                 /*
1185                  * We hold a reference to 'inode' so it couldn't have been
1186                  * removed from s_inodes list while we dropped the
1187                  * inode_sb_list_lock.  We cannot iput the inode now as we can
1188                  * be holding the last reference and we cannot iput it under
1189                  * inode_sb_list_lock. So we keep the reference and iput it
1190                  * later.
1191                  */
1192                 iput(old_inode);
1193                 old_inode = inode;
1194
1195                 filemap_fdatawait(mapping);
1196
1197                 cond_resched();
1198
1199                 spin_lock(&inode_sb_list_lock);
1200         }
1201         spin_unlock(&inode_sb_list_lock);
1202         iput(old_inode);
1203 }
1204
1205 /**
1206  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1207  * @sb: the superblock
1208  * @nr: the number of pages to write
1209  * @reason: reason why some writeback work initiated
1210  *
1211  * Start writeback on some inodes on this super_block. No guarantees are made
1212  * on how many (if any) will be written, and this function does not wait
1213  * for IO completion of submitted IO.
1214  */
1215 void writeback_inodes_sb_nr(struct super_block *sb,
1216                             unsigned long nr,
1217                             enum wb_reason reason)
1218 {
1219         DECLARE_COMPLETION_ONSTACK(done);
1220         struct wb_writeback_work work = {
1221                 .sb                     = sb,
1222                 .sync_mode              = WB_SYNC_NONE,
1223                 .tagged_writepages      = 1,
1224                 .done                   = &done,
1225                 .nr_pages               = nr,
1226                 .reason                 = reason,
1227         };
1228
1229         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1230         bdi_queue_work(sb->s_bdi, &work);
1231         wait_for_completion(&done);
1232 }
1233 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1234
1235 /**
1236  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1237  * @sb: the superblock
1238  * @reason: reason why some writeback work was initiated
1239  *
1240  * Start writeback on some inodes on this super_block. No guarantees are made
1241  * on how many (if any) will be written, and this function does not wait
1242  * for IO completion of submitted IO.
1243  */
1244 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1245 {
1246         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1247 }
1248 EXPORT_SYMBOL(writeback_inodes_sb);
1249
1250 /**
1251  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1252  * @sb: the superblock
1253  * @reason: reason why some writeback work was initiated
1254  *
1255  * Invoke writeback_inodes_sb if no writeback is currently underway.
1256  * Returns 1 if writeback was started, 0 if not.
1257  */
1258 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1259 {
1260         if (!writeback_in_progress(sb->s_bdi)) {
1261                 down_read(&sb->s_umount);
1262                 writeback_inodes_sb(sb, reason);
1263                 up_read(&sb->s_umount);
1264                 return 1;
1265         } else
1266                 return 0;
1267 }
1268 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1269
1270 /**
1271  * writeback_inodes_sb_nr_if_idle       -       start writeback if none underway
1272  * @sb: the superblock
1273  * @nr: the number of pages to write
1274  * @reason: reason why some writeback work was initiated
1275  *
1276  * Invoke writeback_inodes_sb if no writeback is currently underway.
1277  * Returns 1 if writeback was started, 0 if not.
1278  */
1279 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1280                                    unsigned long nr,
1281                                    enum wb_reason reason)
1282 {
1283         if (!writeback_in_progress(sb->s_bdi)) {
1284                 down_read(&sb->s_umount);
1285                 writeback_inodes_sb_nr(sb, nr, reason);
1286                 up_read(&sb->s_umount);
1287                 return 1;
1288         } else
1289                 return 0;
1290 }
1291 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1292
1293 /**
1294  * sync_inodes_sb       -       sync sb inode pages
1295  * @sb: the superblock
1296  *
1297  * This function writes and waits on any dirty inode belonging to this
1298  * super_block.
1299  */
1300 void sync_inodes_sb(struct super_block *sb)
1301 {
1302         DECLARE_COMPLETION_ONSTACK(done);
1303         struct wb_writeback_work work = {
1304                 .sb             = sb,
1305                 .sync_mode      = WB_SYNC_ALL,
1306                 .nr_pages       = LONG_MAX,
1307                 .range_cyclic   = 0,
1308                 .done           = &done,
1309                 .reason         = WB_REASON_SYNC,
1310         };
1311
1312         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1313
1314         bdi_queue_work(sb->s_bdi, &work);
1315         wait_for_completion(&done);
1316
1317         wait_sb_inodes(sb);
1318 }
1319 EXPORT_SYMBOL(sync_inodes_sb);
1320
1321 /**
1322  * write_inode_now      -       write an inode to disk
1323  * @inode: inode to write to disk
1324  * @sync: whether the write should be synchronous or not
1325  *
1326  * This function commits an inode to disk immediately if it is dirty. This is
1327  * primarily needed by knfsd.
1328  *
1329  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1330  */
1331 int write_inode_now(struct inode *inode, int sync)
1332 {
1333         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1334         int ret;
1335         struct writeback_control wbc = {
1336                 .nr_to_write = LONG_MAX,
1337                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1338                 .range_start = 0,
1339                 .range_end = LLONG_MAX,
1340         };
1341
1342         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1343                 wbc.nr_to_write = 0;
1344
1345         might_sleep();
1346         spin_lock(&wb->list_lock);
1347         spin_lock(&inode->i_lock);
1348         ret = writeback_single_inode(inode, wb, &wbc);
1349         spin_unlock(&inode->i_lock);
1350         spin_unlock(&wb->list_lock);
1351         return ret;
1352 }
1353 EXPORT_SYMBOL(write_inode_now);
1354
1355 /**
1356  * sync_inode - write an inode and its pages to disk.
1357  * @inode: the inode to sync
1358  * @wbc: controls the writeback mode
1359  *
1360  * sync_inode() will write an inode and its pages to disk.  It will also
1361  * correctly update the inode on its superblock's dirty inode lists and will
1362  * update inode->i_state.
1363  *
1364  * The caller must have a ref on the inode.
1365  */
1366 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1367 {
1368         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1369         int ret;
1370
1371         spin_lock(&wb->list_lock);
1372         spin_lock(&inode->i_lock);
1373         ret = writeback_single_inode(inode, wb, wbc);
1374         spin_unlock(&inode->i_lock);
1375         spin_unlock(&wb->list_lock);
1376         return ret;
1377 }
1378 EXPORT_SYMBOL(sync_inode);
1379
1380 /**
1381  * sync_inode_metadata - write an inode to disk
1382  * @inode: the inode to sync
1383  * @wait: wait for I/O to complete.
1384  *
1385  * Write an inode to disk and adjust its dirty state after completion.
1386  *
1387  * Note: only writes the actual inode, no associated data or other metadata.
1388  */
1389 int sync_inode_metadata(struct inode *inode, int wait)
1390 {
1391         struct writeback_control wbc = {
1392                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1393                 .nr_to_write = 0, /* metadata-only */
1394         };
1395
1396         return sync_inode(inode, &wbc);
1397 }
1398 EXPORT_SYMBOL(sync_inode_metadata);