vfs: add fs.h to define struct file
[pandora-kernel.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         unsigned int for_kupdate:1;
46         unsigned int range_cyclic:1;
47         unsigned int for_background:1;
48 };
49
50 /*
51  * Work items for the bdi_writeback threads
52  */
53 struct bdi_work {
54         struct list_head list;          /* pending work list */
55         struct rcu_head rcu_head;       /* for RCU free/clear of work */
56
57         unsigned long seen;             /* threads that have seen this work */
58         atomic_t pending;               /* number of threads still to do work */
59
60         struct wb_writeback_args args;  /* writeback arguments */
61
62         unsigned long state;            /* flag bits, see WS_* */
63 };
64
65 enum {
66         WS_USED_B = 0,
67         WS_ONSTACK_B,
68 };
69
70 #define WS_USED (1 << WS_USED_B)
71 #define WS_ONSTACK (1 << WS_ONSTACK_B)
72
73 static inline bool bdi_work_on_stack(struct bdi_work *work)
74 {
75         return test_bit(WS_ONSTACK_B, &work->state);
76 }
77
78 static inline void __bdi_work_init(struct bdi_work *work,
79                                    struct wb_writeback_args *args,
80                                    int on_stack)
81 {
82         work->args = *args;
83         work->state = WS_USED;
84         if (on_stack) {
85                 work->state |= WS_ONSTACK;
86                 init_rcu_head_on_stack(&work->rcu_head);
87         }
88 }
89
90 static inline void bdi_work_init(struct bdi_work *work,
91                                  struct wb_writeback_args *args)
92 {
93         __bdi_work_init(work, args, false);
94 }
95
96 static inline void bdi_work_init_on_stack(struct bdi_work *work,
97                                           struct wb_writeback_args *args)
98 {
99         __bdi_work_init(work, args, true);
100 }
101
102 static inline void bdi_destroy_work_on_stack(struct bdi_work *work)
103 {
104         destroy_rcu_head_on_stack(&work->rcu_head);
105 }
106
107 /**
108  * writeback_in_progress - determine whether there is writeback in progress
109  * @bdi: the device's backing_dev_info structure.
110  *
111  * Determine whether there is writeback waiting to be handled against a
112  * backing device.
113  */
114 int writeback_in_progress(struct backing_dev_info *bdi)
115 {
116         return !list_empty(&bdi->work_list);
117 }
118
119 static void bdi_work_clear(struct bdi_work *work)
120 {
121         clear_bit(WS_USED_B, &work->state);
122         smp_mb__after_clear_bit();
123         /*
124          * work can have disappeared at this point. bit waitq functions
125          * should be able to tolerate this, provided bdi_sched_wait does
126          * not dereference it's pointer argument.
127         */
128         wake_up_bit(&work->state, WS_USED_B);
129 }
130
131 static void bdi_work_free(struct rcu_head *head)
132 {
133         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
134
135         if (!bdi_work_on_stack(work))
136                 kfree(work);
137         else
138                 bdi_work_clear(work);
139 }
140
141 static void wb_work_complete(struct bdi_work *work)
142 {
143         const enum writeback_sync_modes sync_mode = work->args.sync_mode;
144         int onstack = bdi_work_on_stack(work);
145
146         /*
147          * For allocated work, we can clear the done/seen bit right here.
148          * For on-stack work, we need to postpone both the clear and free
149          * to after the RCU grace period, since the stack could be invalidated
150          * as soon as bdi_work_clear() has done the wakeup.
151          */
152         if (!onstack)
153                 bdi_work_clear(work);
154         if (sync_mode == WB_SYNC_NONE || onstack)
155                 call_rcu(&work->rcu_head, bdi_work_free);
156 }
157
158 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
159 {
160         /*
161          * The caller has retrieved the work arguments from this work,
162          * drop our reference. If this is the last ref, delete and free it
163          */
164         if (atomic_dec_and_test(&work->pending)) {
165                 struct backing_dev_info *bdi = wb->bdi;
166
167                 spin_lock(&bdi->wb_lock);
168                 list_del_rcu(&work->list);
169                 spin_unlock(&bdi->wb_lock);
170
171                 wb_work_complete(work);
172         }
173 }
174
175 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
176 {
177         work->seen = bdi->wb_mask;
178         BUG_ON(!work->seen);
179         atomic_set(&work->pending, bdi->wb_cnt);
180         BUG_ON(!bdi->wb_cnt);
181
182         /*
183          * list_add_tail_rcu() contains the necessary barriers to
184          * make sure the above stores are seen before the item is
185          * noticed on the list
186          */
187         spin_lock(&bdi->wb_lock);
188         list_add_tail_rcu(&work->list, &bdi->work_list);
189         spin_unlock(&bdi->wb_lock);
190
191         /*
192          * If the default thread isn't there, make sure we add it. When
193          * it gets created and wakes up, we'll run this work.
194          */
195         if (unlikely(list_empty_careful(&bdi->wb_list)))
196                 wake_up_process(default_backing_dev_info.wb.task);
197         else {
198                 struct bdi_writeback *wb = &bdi->wb;
199
200                 if (wb->task)
201                         wake_up_process(wb->task);
202         }
203 }
204
205 /*
206  * Used for on-stack allocated work items. The caller needs to wait until
207  * the wb threads have acked the work before it's safe to continue.
208  */
209 static void bdi_wait_on_work_clear(struct bdi_work *work)
210 {
211         wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
212                     TASK_UNINTERRUPTIBLE);
213 }
214
215 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
216                                  struct wb_writeback_args *args)
217 {
218         struct bdi_work *work;
219
220         /*
221          * This is WB_SYNC_NONE writeback, so if allocation fails just
222          * wakeup the thread for old dirty data writeback
223          */
224         work = kmalloc(sizeof(*work), GFP_ATOMIC);
225         if (work) {
226                 bdi_work_init(work, args);
227                 bdi_queue_work(bdi, work);
228         } else {
229                 struct bdi_writeback *wb = &bdi->wb;
230
231                 if (wb->task)
232                         wake_up_process(wb->task);
233         }
234 }
235
236 /**
237  * bdi_sync_writeback - start and wait for writeback
238  * @bdi: the backing device to write from
239  * @sb: write inodes from this super_block
240  *
241  * Description:
242  *   This does WB_SYNC_ALL data integrity writeback and waits for the
243  *   IO to complete. Callers must hold the sb s_umount semaphore for
244  *   reading, to avoid having the super disappear before we are done.
245  */
246 static void bdi_sync_writeback(struct backing_dev_info *bdi,
247                                struct super_block *sb)
248 {
249         struct wb_writeback_args args = {
250                 .sb             = sb,
251                 .sync_mode      = WB_SYNC_ALL,
252                 .nr_pages       = LONG_MAX,
253                 .range_cyclic   = 0,
254         };
255         struct bdi_work work;
256
257         bdi_work_init_on_stack(&work, &args);
258
259         bdi_queue_work(bdi, &work);
260         bdi_wait_on_work_clear(&work);
261         bdi_destroy_work_on_stack(&work);
262 }
263
264 /**
265  * bdi_start_writeback - start writeback
266  * @bdi: the backing device to write from
267  * @sb: write inodes from this super_block
268  * @nr_pages: the number of pages to write
269  *
270  * Description:
271  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
272  *   started when this function returns, we make no guarentees on
273  *   completion. Caller need not hold sb s_umount semaphore.
274  *
275  */
276 void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
277                          long nr_pages)
278 {
279         struct wb_writeback_args args = {
280                 .sb             = sb,
281                 .sync_mode      = WB_SYNC_NONE,
282                 .nr_pages       = nr_pages,
283                 .range_cyclic   = 1,
284         };
285
286         /*
287          * We treat @nr_pages=0 as the special case to do background writeback,
288          * ie. to sync pages until the background dirty threshold is reached.
289          */
290         if (!nr_pages) {
291                 args.nr_pages = LONG_MAX;
292                 args.for_background = 1;
293         }
294
295         bdi_alloc_queue_work(bdi, &args);
296 }
297
298 /*
299  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
300  * furthest end of its superblock's dirty-inode list.
301  *
302  * Before stamping the inode's ->dirtied_when, we check to see whether it is
303  * already the most-recently-dirtied inode on the b_dirty list.  If that is
304  * the case then the inode must have been redirtied while it was being written
305  * out and we don't reset its dirtied_when.
306  */
307 static void redirty_tail(struct inode *inode)
308 {
309         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
310
311         if (!list_empty(&wb->b_dirty)) {
312                 struct inode *tail;
313
314                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
315                 if (time_before(inode->dirtied_when, tail->dirtied_when))
316                         inode->dirtied_when = jiffies;
317         }
318         list_move(&inode->i_list, &wb->b_dirty);
319 }
320
321 /*
322  * requeue inode for re-scanning after bdi->b_io list is exhausted.
323  */
324 static void requeue_io(struct inode *inode)
325 {
326         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
327
328         list_move(&inode->i_list, &wb->b_more_io);
329 }
330
331 static void inode_sync_complete(struct inode *inode)
332 {
333         /*
334          * Prevent speculative execution through spin_unlock(&inode_lock);
335          */
336         smp_mb();
337         wake_up_bit(&inode->i_state, __I_SYNC);
338 }
339
340 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
341 {
342         bool ret = time_after(inode->dirtied_when, t);
343 #ifndef CONFIG_64BIT
344         /*
345          * For inodes being constantly redirtied, dirtied_when can get stuck.
346          * It _appears_ to be in the future, but is actually in distant past.
347          * This test is necessary to prevent such wrapped-around relative times
348          * from permanently stopping the whole bdi writeback.
349          */
350         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
351 #endif
352         return ret;
353 }
354
355 /*
356  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
357  */
358 static void move_expired_inodes(struct list_head *delaying_queue,
359                                struct list_head *dispatch_queue,
360                                 unsigned long *older_than_this)
361 {
362         LIST_HEAD(tmp);
363         struct list_head *pos, *node;
364         struct super_block *sb = NULL;
365         struct inode *inode;
366         int do_sb_sort = 0;
367
368         while (!list_empty(delaying_queue)) {
369                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
370                 if (older_than_this &&
371                     inode_dirtied_after(inode, *older_than_this))
372                         break;
373                 if (sb && sb != inode->i_sb)
374                         do_sb_sort = 1;
375                 sb = inode->i_sb;
376                 list_move(&inode->i_list, &tmp);
377         }
378
379         /* just one sb in list, splice to dispatch_queue and we're done */
380         if (!do_sb_sort) {
381                 list_splice(&tmp, dispatch_queue);
382                 return;
383         }
384
385         /* Move inodes from one superblock together */
386         while (!list_empty(&tmp)) {
387                 inode = list_entry(tmp.prev, struct inode, i_list);
388                 sb = inode->i_sb;
389                 list_for_each_prev_safe(pos, node, &tmp) {
390                         inode = list_entry(pos, struct inode, i_list);
391                         if (inode->i_sb == sb)
392                                 list_move(&inode->i_list, dispatch_queue);
393                 }
394         }
395 }
396
397 /*
398  * Queue all expired dirty inodes for io, eldest first.
399  */
400 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
401 {
402         list_splice_init(&wb->b_more_io, wb->b_io.prev);
403         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
404 }
405
406 static int write_inode(struct inode *inode, struct writeback_control *wbc)
407 {
408         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
409                 return inode->i_sb->s_op->write_inode(inode, wbc);
410         return 0;
411 }
412
413 /*
414  * Wait for writeback on an inode to complete.
415  */
416 static void inode_wait_for_writeback(struct inode *inode)
417 {
418         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
419         wait_queue_head_t *wqh;
420
421         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
422          while (inode->i_state & I_SYNC) {
423                 spin_unlock(&inode_lock);
424                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
425                 spin_lock(&inode_lock);
426         }
427 }
428
429 /*
430  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
431  * caller has ref on the inode (either via __iget or via syscall against an fd)
432  * or the inode has I_WILL_FREE set (via generic_forget_inode)
433  *
434  * If `wait' is set, wait on the writeout.
435  *
436  * The whole writeout design is quite complex and fragile.  We want to avoid
437  * starvation of particular inodes when others are being redirtied, prevent
438  * livelocks, etc.
439  *
440  * Called under inode_lock.
441  */
442 static int
443 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
444 {
445         struct address_space *mapping = inode->i_mapping;
446         unsigned dirty;
447         int ret;
448
449         if (!atomic_read(&inode->i_count))
450                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
451         else
452                 WARN_ON(inode->i_state & I_WILL_FREE);
453
454         if (inode->i_state & I_SYNC) {
455                 /*
456                  * If this inode is locked for writeback and we are not doing
457                  * writeback-for-data-integrity, move it to b_more_io so that
458                  * writeback can proceed with the other inodes on s_io.
459                  *
460                  * We'll have another go at writing back this inode when we
461                  * completed a full scan of b_io.
462                  */
463                 if (wbc->sync_mode != WB_SYNC_ALL) {
464                         requeue_io(inode);
465                         return 0;
466                 }
467
468                 /*
469                  * It's a data-integrity sync.  We must wait.
470                  */
471                 inode_wait_for_writeback(inode);
472         }
473
474         BUG_ON(inode->i_state & I_SYNC);
475
476         /* Set I_SYNC, reset I_DIRTY_PAGES */
477         inode->i_state |= I_SYNC;
478         inode->i_state &= ~I_DIRTY_PAGES;
479         spin_unlock(&inode_lock);
480
481         ret = do_writepages(mapping, wbc);
482
483         /*
484          * Make sure to wait on the data before writing out the metadata.
485          * This is important for filesystems that modify metadata on data
486          * I/O completion.
487          */
488         if (wbc->sync_mode == WB_SYNC_ALL) {
489                 int err = filemap_fdatawait(mapping);
490                 if (ret == 0)
491                         ret = err;
492         }
493
494         /*
495          * Some filesystems may redirty the inode during the writeback
496          * due to delalloc, clear dirty metadata flags right before
497          * write_inode()
498          */
499         spin_lock(&inode_lock);
500         dirty = inode->i_state & I_DIRTY;
501         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
502         spin_unlock(&inode_lock);
503         /* Don't write the inode if only I_DIRTY_PAGES was set */
504         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
505                 int err = write_inode(inode, wbc);
506                 if (ret == 0)
507                         ret = err;
508         }
509
510         spin_lock(&inode_lock);
511         inode->i_state &= ~I_SYNC;
512         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
513                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
514                         /*
515                          * More pages get dirtied by a fast dirtier.
516                          */
517                         goto select_queue;
518                 } else if (inode->i_state & I_DIRTY) {
519                         /*
520                          * At least XFS will redirty the inode during the
521                          * writeback (delalloc) and on io completion (isize).
522                          */
523                         redirty_tail(inode);
524                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
525                         /*
526                          * We didn't write back all the pages.  nfs_writepages()
527                          * sometimes bales out without doing anything. Redirty
528                          * the inode; Move it from b_io onto b_more_io/b_dirty.
529                          */
530                         /*
531                          * akpm: if the caller was the kupdate function we put
532                          * this inode at the head of b_dirty so it gets first
533                          * consideration.  Otherwise, move it to the tail, for
534                          * the reasons described there.  I'm not really sure
535                          * how much sense this makes.  Presumably I had a good
536                          * reasons for doing it this way, and I'd rather not
537                          * muck with it at present.
538                          */
539                         if (wbc->for_kupdate) {
540                                 /*
541                                  * For the kupdate function we move the inode
542                                  * to b_more_io so it will get more writeout as
543                                  * soon as the queue becomes uncongested.
544                                  */
545                                 inode->i_state |= I_DIRTY_PAGES;
546 select_queue:
547                                 if (wbc->nr_to_write <= 0) {
548                                         /*
549                                          * slice used up: queue for next turn
550                                          */
551                                         requeue_io(inode);
552                                 } else {
553                                         /*
554                                          * somehow blocked: retry later
555                                          */
556                                         redirty_tail(inode);
557                                 }
558                         } else {
559                                 /*
560                                  * Otherwise fully redirty the inode so that
561                                  * other inodes on this superblock will get some
562                                  * writeout.  Otherwise heavy writing to one
563                                  * file would indefinitely suspend writeout of
564                                  * all the other files.
565                                  */
566                                 inode->i_state |= I_DIRTY_PAGES;
567                                 redirty_tail(inode);
568                         }
569                 } else if (atomic_read(&inode->i_count)) {
570                         /*
571                          * The inode is clean, inuse
572                          */
573                         list_move(&inode->i_list, &inode_in_use);
574                 } else {
575                         /*
576                          * The inode is clean, unused
577                          */
578                         list_move(&inode->i_list, &inode_unused);
579                 }
580         }
581         inode_sync_complete(inode);
582         return ret;
583 }
584
585 static void unpin_sb_for_writeback(struct super_block *sb)
586 {
587         up_read(&sb->s_umount);
588         put_super(sb);
589 }
590
591 enum sb_pin_state {
592         SB_PINNED,
593         SB_NOT_PINNED,
594         SB_PIN_FAILED
595 };
596
597 /*
598  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
599  * before calling writeback. So make sure that we do pin it, so it doesn't
600  * go away while we are writing inodes from it.
601  */
602 static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
603                                               struct super_block *sb)
604 {
605         /*
606          * Caller must already hold the ref for this
607          */
608         if (wbc->sync_mode == WB_SYNC_ALL) {
609                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
610                 return SB_NOT_PINNED;
611         }
612         spin_lock(&sb_lock);
613         sb->s_count++;
614         if (down_read_trylock(&sb->s_umount)) {
615                 if (sb->s_root) {
616                         spin_unlock(&sb_lock);
617                         return SB_PINNED;
618                 }
619                 /*
620                  * umounted, drop rwsem again and fall through to failure
621                  */
622                 up_read(&sb->s_umount);
623         }
624         sb->s_count--;
625         spin_unlock(&sb_lock);
626         return SB_PIN_FAILED;
627 }
628
629 /*
630  * Write a portion of b_io inodes which belong to @sb.
631  * If @wbc->sb != NULL, then find and write all such
632  * inodes. Otherwise write only ones which go sequentially
633  * in reverse order.
634  * Return 1, if the caller writeback routine should be
635  * interrupted. Otherwise return 0.
636  */
637 static int writeback_sb_inodes(struct super_block *sb,
638                                struct bdi_writeback *wb,
639                                struct writeback_control *wbc)
640 {
641         while (!list_empty(&wb->b_io)) {
642                 long pages_skipped;
643                 struct inode *inode = list_entry(wb->b_io.prev,
644                                                  struct inode, i_list);
645                 if (wbc->sb && sb != inode->i_sb) {
646                         /* super block given and doesn't
647                            match, skip this inode */
648                         redirty_tail(inode);
649                         continue;
650                 }
651                 if (sb != inode->i_sb)
652                         /* finish with this superblock */
653                         return 0;
654                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
655                         requeue_io(inode);
656                         continue;
657                 }
658                 /*
659                  * Was this inode dirtied after sync_sb_inodes was called?
660                  * This keeps sync from extra jobs and livelock.
661                  */
662                 if (inode_dirtied_after(inode, wbc->wb_start))
663                         return 1;
664
665                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
666                 __iget(inode);
667                 pages_skipped = wbc->pages_skipped;
668                 writeback_single_inode(inode, wbc);
669                 if (wbc->pages_skipped != pages_skipped) {
670                         /*
671                          * writeback is not making progress due to locked
672                          * buffers.  Skip this inode for now.
673                          */
674                         redirty_tail(inode);
675                 }
676                 spin_unlock(&inode_lock);
677                 iput(inode);
678                 cond_resched();
679                 spin_lock(&inode_lock);
680                 if (wbc->nr_to_write <= 0) {
681                         wbc->more_io = 1;
682                         return 1;
683                 }
684                 if (!list_empty(&wb->b_more_io))
685                         wbc->more_io = 1;
686         }
687         /* b_io is empty */
688         return 1;
689 }
690
691 static void writeback_inodes_wb(struct bdi_writeback *wb,
692                                 struct writeback_control *wbc)
693 {
694         int ret = 0;
695
696         wbc->wb_start = jiffies; /* livelock avoidance */
697         spin_lock(&inode_lock);
698         if (!wbc->for_kupdate || list_empty(&wb->b_io))
699                 queue_io(wb, wbc->older_than_this);
700
701         while (!list_empty(&wb->b_io)) {
702                 struct inode *inode = list_entry(wb->b_io.prev,
703                                                  struct inode, i_list);
704                 struct super_block *sb = inode->i_sb;
705                 enum sb_pin_state state;
706
707                 if (wbc->sb && sb != wbc->sb) {
708                         /* super block given and doesn't
709                            match, skip this inode */
710                         redirty_tail(inode);
711                         continue;
712                 }
713                 state = pin_sb_for_writeback(wbc, sb);
714
715                 if (state == SB_PIN_FAILED) {
716                         requeue_io(inode);
717                         continue;
718                 }
719                 ret = writeback_sb_inodes(sb, wb, wbc);
720
721                 if (state == SB_PINNED)
722                         unpin_sb_for_writeback(sb);
723                 if (ret)
724                         break;
725         }
726         spin_unlock(&inode_lock);
727         /* Leave any unwritten inodes on b_io */
728 }
729
730 void writeback_inodes_wbc(struct writeback_control *wbc)
731 {
732         struct backing_dev_info *bdi = wbc->bdi;
733
734         writeback_inodes_wb(&bdi->wb, wbc);
735 }
736
737 /*
738  * The maximum number of pages to writeout in a single bdi flush/kupdate
739  * operation.  We do this so we don't hold I_SYNC against an inode for
740  * enormous amounts of time, which would block a userspace task which has
741  * been forced to throttle against that inode.  Also, the code reevaluates
742  * the dirty each time it has written this many pages.
743  */
744 #define MAX_WRITEBACK_PAGES     1024
745
746 static inline bool over_bground_thresh(void)
747 {
748         unsigned long background_thresh, dirty_thresh;
749
750         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
751
752         return (global_page_state(NR_FILE_DIRTY) +
753                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
754 }
755
756 /*
757  * Explicit flushing or periodic writeback of "old" data.
758  *
759  * Define "old": the first time one of an inode's pages is dirtied, we mark the
760  * dirtying-time in the inode's address_space.  So this periodic writeback code
761  * just walks the superblock inode list, writing back any inodes which are
762  * older than a specific point in time.
763  *
764  * Try to run once per dirty_writeback_interval.  But if a writeback event
765  * takes longer than a dirty_writeback_interval interval, then leave a
766  * one-second gap.
767  *
768  * older_than_this takes precedence over nr_to_write.  So we'll only write back
769  * all dirty pages if they are all attached to "old" mappings.
770  */
771 static long wb_writeback(struct bdi_writeback *wb,
772                          struct wb_writeback_args *args)
773 {
774         struct writeback_control wbc = {
775                 .bdi                    = wb->bdi,
776                 .sb                     = args->sb,
777                 .sync_mode              = args->sync_mode,
778                 .older_than_this        = NULL,
779                 .for_kupdate            = args->for_kupdate,
780                 .for_background         = args->for_background,
781                 .range_cyclic           = args->range_cyclic,
782         };
783         unsigned long oldest_jif;
784         long wrote = 0;
785         struct inode *inode;
786
787         if (wbc.for_kupdate) {
788                 wbc.older_than_this = &oldest_jif;
789                 oldest_jif = jiffies -
790                                 msecs_to_jiffies(dirty_expire_interval * 10);
791         }
792         if (!wbc.range_cyclic) {
793                 wbc.range_start = 0;
794                 wbc.range_end = LLONG_MAX;
795         }
796
797         for (;;) {
798                 /*
799                  * Stop writeback when nr_pages has been consumed
800                  */
801                 if (args->nr_pages <= 0)
802                         break;
803
804                 /*
805                  * For background writeout, stop when we are below the
806                  * background dirty threshold
807                  */
808                 if (args->for_background && !over_bground_thresh())
809                         break;
810
811                 wbc.more_io = 0;
812                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
813                 wbc.pages_skipped = 0;
814                 writeback_inodes_wb(wb, &wbc);
815                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
816                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
817
818                 /*
819                  * If we consumed everything, see if we have more
820                  */
821                 if (wbc.nr_to_write <= 0)
822                         continue;
823                 /*
824                  * Didn't write everything and we don't have more IO, bail
825                  */
826                 if (!wbc.more_io)
827                         break;
828                 /*
829                  * Did we write something? Try for more
830                  */
831                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
832                         continue;
833                 /*
834                  * Nothing written. Wait for some inode to
835                  * become available for writeback. Otherwise
836                  * we'll just busyloop.
837                  */
838                 spin_lock(&inode_lock);
839                 if (!list_empty(&wb->b_more_io))  {
840                         inode = list_entry(wb->b_more_io.prev,
841                                                 struct inode, i_list);
842                         inode_wait_for_writeback(inode);
843                 }
844                 spin_unlock(&inode_lock);
845         }
846
847         return wrote;
848 }
849
850 /*
851  * Return the next bdi_work struct that hasn't been processed by this
852  * wb thread yet. ->seen is initially set for each thread that exists
853  * for this device, when a thread first notices a piece of work it
854  * clears its bit. Depending on writeback type, the thread will notify
855  * completion on either receiving the work (WB_SYNC_NONE) or after
856  * it is done (WB_SYNC_ALL).
857  */
858 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
859                                            struct bdi_writeback *wb)
860 {
861         struct bdi_work *work, *ret = NULL;
862
863         rcu_read_lock();
864
865         list_for_each_entry_rcu(work, &bdi->work_list, list) {
866                 if (!test_bit(wb->nr, &work->seen))
867                         continue;
868                 clear_bit(wb->nr, &work->seen);
869
870                 ret = work;
871                 break;
872         }
873
874         rcu_read_unlock();
875         return ret;
876 }
877
878 static long wb_check_old_data_flush(struct bdi_writeback *wb)
879 {
880         unsigned long expired;
881         long nr_pages;
882
883         /*
884          * When set to zero, disable periodic writeback
885          */
886         if (!dirty_writeback_interval)
887                 return 0;
888
889         expired = wb->last_old_flush +
890                         msecs_to_jiffies(dirty_writeback_interval * 10);
891         if (time_before(jiffies, expired))
892                 return 0;
893
894         wb->last_old_flush = jiffies;
895         nr_pages = global_page_state(NR_FILE_DIRTY) +
896                         global_page_state(NR_UNSTABLE_NFS) +
897                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
898
899         if (nr_pages) {
900                 struct wb_writeback_args args = {
901                         .nr_pages       = nr_pages,
902                         .sync_mode      = WB_SYNC_NONE,
903                         .for_kupdate    = 1,
904                         .range_cyclic   = 1,
905                 };
906
907                 return wb_writeback(wb, &args);
908         }
909
910         return 0;
911 }
912
913 /*
914  * Retrieve work items and do the writeback they describe
915  */
916 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
917 {
918         struct backing_dev_info *bdi = wb->bdi;
919         struct bdi_work *work;
920         long wrote = 0;
921
922         while ((work = get_next_work_item(bdi, wb)) != NULL) {
923                 struct wb_writeback_args args = work->args;
924
925                 /*
926                  * Override sync mode, in case we must wait for completion
927                  */
928                 if (force_wait)
929                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
930
931                 /*
932                  * If this isn't a data integrity operation, just notify
933                  * that we have seen this work and we are now starting it.
934                  */
935                 if (args.sync_mode == WB_SYNC_NONE)
936                         wb_clear_pending(wb, work);
937
938                 wrote += wb_writeback(wb, &args);
939
940                 /*
941                  * This is a data integrity writeback, so only do the
942                  * notification when we have completed the work.
943                  */
944                 if (args.sync_mode == WB_SYNC_ALL)
945                         wb_clear_pending(wb, work);
946         }
947
948         /*
949          * Check for periodic writeback, kupdated() style
950          */
951         wrote += wb_check_old_data_flush(wb);
952
953         return wrote;
954 }
955
956 /*
957  * Handle writeback of dirty data for the device backed by this bdi. Also
958  * wakes up periodically and does kupdated style flushing.
959  */
960 int bdi_writeback_task(struct bdi_writeback *wb)
961 {
962         unsigned long last_active = jiffies;
963         unsigned long wait_jiffies = -1UL;
964         long pages_written;
965
966         while (!kthread_should_stop()) {
967                 pages_written = wb_do_writeback(wb, 0);
968
969                 if (pages_written)
970                         last_active = jiffies;
971                 else if (wait_jiffies != -1UL) {
972                         unsigned long max_idle;
973
974                         /*
975                          * Longest period of inactivity that we tolerate. If we
976                          * see dirty data again later, the task will get
977                          * recreated automatically.
978                          */
979                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
980                         if (time_after(jiffies, max_idle + last_active))
981                                 break;
982                 }
983
984                 if (dirty_writeback_interval) {
985                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
986                         schedule_timeout_interruptible(wait_jiffies);
987                 } else {
988                         set_current_state(TASK_INTERRUPTIBLE);
989                         if (list_empty_careful(&wb->bdi->work_list) &&
990                             !kthread_should_stop())
991                                 schedule();
992                         __set_current_state(TASK_RUNNING);
993                 }
994
995                 try_to_freeze();
996         }
997
998         return 0;
999 }
1000
1001 /*
1002  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
1003  * writeback, for integrity writeback see bdi_sync_writeback().
1004  */
1005 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
1006 {
1007         struct wb_writeback_args args = {
1008                 .sb             = sb,
1009                 .nr_pages       = nr_pages,
1010                 .sync_mode      = WB_SYNC_NONE,
1011         };
1012         struct backing_dev_info *bdi;
1013
1014         rcu_read_lock();
1015
1016         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1017                 if (!bdi_has_dirty_io(bdi))
1018                         continue;
1019
1020                 bdi_alloc_queue_work(bdi, &args);
1021         }
1022
1023         rcu_read_unlock();
1024 }
1025
1026 /*
1027  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1028  * the whole world.
1029  */
1030 void wakeup_flusher_threads(long nr_pages)
1031 {
1032         if (nr_pages == 0)
1033                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1034                                 global_page_state(NR_UNSTABLE_NFS);
1035         bdi_writeback_all(NULL, nr_pages);
1036 }
1037
1038 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1039 {
1040         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1041                 struct dentry *dentry;
1042                 const char *name = "?";
1043
1044                 dentry = d_find_alias(inode);
1045                 if (dentry) {
1046                         spin_lock(&dentry->d_lock);
1047                         name = (const char *) dentry->d_name.name;
1048                 }
1049                 printk(KERN_DEBUG
1050                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1051                        current->comm, task_pid_nr(current), inode->i_ino,
1052                        name, inode->i_sb->s_id);
1053                 if (dentry) {
1054                         spin_unlock(&dentry->d_lock);
1055                         dput(dentry);
1056                 }
1057         }
1058 }
1059
1060 /**
1061  *      __mark_inode_dirty -    internal function
1062  *      @inode: inode to mark
1063  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1064  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1065  *      mark_inode_dirty_sync.
1066  *
1067  * Put the inode on the super block's dirty list.
1068  *
1069  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1070  * dirty list only if it is hashed or if it refers to a blockdev.
1071  * If it was not hashed, it will never be added to the dirty list
1072  * even if it is later hashed, as it will have been marked dirty already.
1073  *
1074  * In short, make sure you hash any inodes _before_ you start marking
1075  * them dirty.
1076  *
1077  * This function *must* be atomic for the I_DIRTY_PAGES case -
1078  * set_page_dirty() is called under spinlock in several places.
1079  *
1080  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1081  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1082  * the kernel-internal blockdev inode represents the dirtying time of the
1083  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1084  * page->mapping->host, so the page-dirtying time is recorded in the internal
1085  * blockdev inode.
1086  */
1087 void __mark_inode_dirty(struct inode *inode, int flags)
1088 {
1089         struct super_block *sb = inode->i_sb;
1090
1091         /*
1092          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1093          * dirty the inode itself
1094          */
1095         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1096                 if (sb->s_op->dirty_inode)
1097                         sb->s_op->dirty_inode(inode);
1098         }
1099
1100         /*
1101          * make sure that changes are seen by all cpus before we test i_state
1102          * -- mikulas
1103          */
1104         smp_mb();
1105
1106         /* avoid the locking if we can */
1107         if ((inode->i_state & flags) == flags)
1108                 return;
1109
1110         if (unlikely(block_dump))
1111                 block_dump___mark_inode_dirty(inode);
1112
1113         spin_lock(&inode_lock);
1114         if ((inode->i_state & flags) != flags) {
1115                 const int was_dirty = inode->i_state & I_DIRTY;
1116
1117                 inode->i_state |= flags;
1118
1119                 /*
1120                  * If the inode is being synced, just update its dirty state.
1121                  * The unlocker will place the inode on the appropriate
1122                  * superblock list, based upon its state.
1123                  */
1124                 if (inode->i_state & I_SYNC)
1125                         goto out;
1126
1127                 /*
1128                  * Only add valid (hashed) inodes to the superblock's
1129                  * dirty list.  Add blockdev inodes as well.
1130                  */
1131                 if (!S_ISBLK(inode->i_mode)) {
1132                         if (hlist_unhashed(&inode->i_hash))
1133                                 goto out;
1134                 }
1135                 if (inode->i_state & (I_FREEING|I_CLEAR))
1136                         goto out;
1137
1138                 /*
1139                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1140                  * reposition it (that would break b_dirty time-ordering).
1141                  */
1142                 if (!was_dirty) {
1143                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1144                         struct backing_dev_info *bdi = wb->bdi;
1145
1146                         if (bdi_cap_writeback_dirty(bdi) &&
1147                             !test_bit(BDI_registered, &bdi->state)) {
1148                                 WARN_ON(1);
1149                                 printk(KERN_ERR "bdi-%s not registered\n",
1150                                                                 bdi->name);
1151                         }
1152
1153                         inode->dirtied_when = jiffies;
1154                         list_move(&inode->i_list, &wb->b_dirty);
1155                 }
1156         }
1157 out:
1158         spin_unlock(&inode_lock);
1159 }
1160 EXPORT_SYMBOL(__mark_inode_dirty);
1161
1162 /*
1163  * Write out a superblock's list of dirty inodes.  A wait will be performed
1164  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1165  *
1166  * If older_than_this is non-NULL, then only write out inodes which
1167  * had their first dirtying at a time earlier than *older_than_this.
1168  *
1169  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1170  * This function assumes that the blockdev superblock's inodes are backed by
1171  * a variety of queues, so all inodes are searched.  For other superblocks,
1172  * assume that all inodes are backed by the same queue.
1173  *
1174  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1175  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1176  * on the writer throttling path, and we get decent balancing between many
1177  * throttled threads: we don't want them all piling up on inode_sync_wait.
1178  */
1179 static void wait_sb_inodes(struct super_block *sb)
1180 {
1181         struct inode *inode, *old_inode = NULL;
1182
1183         /*
1184          * We need to be protected against the filesystem going from
1185          * r/o to r/w or vice versa.
1186          */
1187         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1188
1189         spin_lock(&inode_lock);
1190
1191         /*
1192          * Data integrity sync. Must wait for all pages under writeback,
1193          * because there may have been pages dirtied before our sync
1194          * call, but which had writeout started before we write it out.
1195          * In which case, the inode may not be on the dirty list, but
1196          * we still have to wait for that writeout.
1197          */
1198         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1199                 struct address_space *mapping;
1200
1201                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1202                         continue;
1203                 mapping = inode->i_mapping;
1204                 if (mapping->nrpages == 0)
1205                         continue;
1206                 __iget(inode);
1207                 spin_unlock(&inode_lock);
1208                 /*
1209                  * We hold a reference to 'inode' so it couldn't have
1210                  * been removed from s_inodes list while we dropped the
1211                  * inode_lock.  We cannot iput the inode now as we can
1212                  * be holding the last reference and we cannot iput it
1213                  * under inode_lock. So we keep the reference and iput
1214                  * it later.
1215                  */
1216                 iput(old_inode);
1217                 old_inode = inode;
1218
1219                 filemap_fdatawait(mapping);
1220
1221                 cond_resched();
1222
1223                 spin_lock(&inode_lock);
1224         }
1225         spin_unlock(&inode_lock);
1226         iput(old_inode);
1227 }
1228
1229 /**
1230  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1231  * @sb: the superblock
1232  *
1233  * Start writeback on some inodes on this super_block. No guarantees are made
1234  * on how many (if any) will be written, and this function does not wait
1235  * for IO completion of submitted IO. The number of pages submitted is
1236  * returned.
1237  */
1238 void writeback_inodes_sb(struct super_block *sb)
1239 {
1240         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1241         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1242         long nr_to_write;
1243
1244         nr_to_write = nr_dirty + nr_unstable +
1245                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1246
1247         bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
1248 }
1249 EXPORT_SYMBOL(writeback_inodes_sb);
1250
1251 /**
1252  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1253  * @sb: the superblock
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)
1259 {
1260         if (!writeback_in_progress(sb->s_bdi)) {
1261                 writeback_inodes_sb(sb);
1262                 return 1;
1263         } else
1264                 return 0;
1265 }
1266 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1267
1268 /**
1269  * sync_inodes_sb       -       sync sb inode pages
1270  * @sb: the superblock
1271  *
1272  * This function writes and waits on any dirty inode belonging to this
1273  * super_block. The number of pages synced is returned.
1274  */
1275 void sync_inodes_sb(struct super_block *sb)
1276 {
1277         bdi_sync_writeback(sb->s_bdi, sb);
1278         wait_sb_inodes(sb);
1279 }
1280 EXPORT_SYMBOL(sync_inodes_sb);
1281
1282 /**
1283  * write_inode_now      -       write an inode to disk
1284  * @inode: inode to write to disk
1285  * @sync: whether the write should be synchronous or not
1286  *
1287  * This function commits an inode to disk immediately if it is dirty. This is
1288  * primarily needed by knfsd.
1289  *
1290  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1291  */
1292 int write_inode_now(struct inode *inode, int sync)
1293 {
1294         int ret;
1295         struct writeback_control wbc = {
1296                 .nr_to_write = LONG_MAX,
1297                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1298                 .range_start = 0,
1299                 .range_end = LLONG_MAX,
1300         };
1301
1302         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1303                 wbc.nr_to_write = 0;
1304
1305         might_sleep();
1306         spin_lock(&inode_lock);
1307         ret = writeback_single_inode(inode, &wbc);
1308         spin_unlock(&inode_lock);
1309         if (sync)
1310                 inode_sync_wait(inode);
1311         return ret;
1312 }
1313 EXPORT_SYMBOL(write_inode_now);
1314
1315 /**
1316  * sync_inode - write an inode and its pages to disk.
1317  * @inode: the inode to sync
1318  * @wbc: controls the writeback mode
1319  *
1320  * sync_inode() will write an inode and its pages to disk.  It will also
1321  * correctly update the inode on its superblock's dirty inode lists and will
1322  * update inode->i_state.
1323  *
1324  * The caller must have a ref on the inode.
1325  */
1326 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1327 {
1328         int ret;
1329
1330         spin_lock(&inode_lock);
1331         ret = writeback_single_inode(inode, wbc);
1332         spin_unlock(&inode_lock);
1333         return ret;
1334 }
1335 EXPORT_SYMBOL(sync_inode);