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