760dc8d0b4ff0863326b5afcb0ba10dc19da13b9
[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 };
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 {
81         INIT_RCU_HEAD(&work->rcu_head);
82         work->args = *args;
83         work->state = WS_USED;
84 }
85
86 /**
87  * writeback_in_progress - determine whether there is writeback in progress
88  * @bdi: the device's backing_dev_info structure.
89  *
90  * Determine whether there is writeback waiting to be handled against a
91  * backing device.
92  */
93 int writeback_in_progress(struct backing_dev_info *bdi)
94 {
95         return !list_empty(&bdi->work_list);
96 }
97
98 static void bdi_work_clear(struct bdi_work *work)
99 {
100         clear_bit(WS_USED_B, &work->state);
101         smp_mb__after_clear_bit();
102         /*
103          * work can have disappeared at this point. bit waitq functions
104          * should be able to tolerate this, provided bdi_sched_wait does
105          * not dereference it's pointer argument.
106         */
107         wake_up_bit(&work->state, WS_USED_B);
108 }
109
110 static void bdi_work_free(struct rcu_head *head)
111 {
112         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
113
114         if (!bdi_work_on_stack(work))
115                 kfree(work);
116         else
117                 bdi_work_clear(work);
118 }
119
120 static void wb_work_complete(struct bdi_work *work)
121 {
122         const enum writeback_sync_modes sync_mode = work->args.sync_mode;
123         int onstack = bdi_work_on_stack(work);
124
125         /*
126          * For allocated work, we can clear the done/seen bit right here.
127          * For on-stack work, we need to postpone both the clear and free
128          * to after the RCU grace period, since the stack could be invalidated
129          * as soon as bdi_work_clear() has done the wakeup.
130          */
131         if (!onstack)
132                 bdi_work_clear(work);
133         if (sync_mode == WB_SYNC_NONE || onstack)
134                 call_rcu(&work->rcu_head, bdi_work_free);
135 }
136
137 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
138 {
139         /*
140          * The caller has retrieved the work arguments from this work,
141          * drop our reference. If this is the last ref, delete and free it
142          */
143         if (atomic_dec_and_test(&work->pending)) {
144                 struct backing_dev_info *bdi = wb->bdi;
145
146                 spin_lock(&bdi->wb_lock);
147                 list_del_rcu(&work->list);
148                 spin_unlock(&bdi->wb_lock);
149
150                 wb_work_complete(work);
151         }
152 }
153
154 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
155 {
156         work->seen = bdi->wb_mask;
157         BUG_ON(!work->seen);
158         atomic_set(&work->pending, bdi->wb_cnt);
159         BUG_ON(!bdi->wb_cnt);
160
161         /*
162          * list_add_tail_rcu() contains the necessary barriers to
163          * make sure the above stores are seen before the item is
164          * noticed on the list
165          */
166         spin_lock(&bdi->wb_lock);
167         list_add_tail_rcu(&work->list, &bdi->work_list);
168         spin_unlock(&bdi->wb_lock);
169
170         /*
171          * If the default thread isn't there, make sure we add it. When
172          * it gets created and wakes up, we'll run this work.
173          */
174         if (unlikely(list_empty_careful(&bdi->wb_list)))
175                 wake_up_process(default_backing_dev_info.wb.task);
176         else {
177                 struct bdi_writeback *wb = &bdi->wb;
178
179                 if (wb->task)
180                         wake_up_process(wb->task);
181         }
182 }
183
184 /*
185  * Used for on-stack allocated work items. The caller needs to wait until
186  * the wb threads have acked the work before it's safe to continue.
187  */
188 static void bdi_wait_on_work_clear(struct bdi_work *work)
189 {
190         wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
191                     TASK_UNINTERRUPTIBLE);
192 }
193
194 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
195                                  struct wb_writeback_args *args)
196 {
197         struct bdi_work *work;
198
199         /*
200          * This is WB_SYNC_NONE writeback, so if allocation fails just
201          * wakeup the thread for old dirty data writeback
202          */
203         work = kmalloc(sizeof(*work), GFP_ATOMIC);
204         if (work) {
205                 bdi_work_init(work, args);
206                 bdi_queue_work(bdi, work);
207         } else {
208                 struct bdi_writeback *wb = &bdi->wb;
209
210                 if (wb->task)
211                         wake_up_process(wb->task);
212         }
213 }
214
215 /**
216  * bdi_sync_writeback - start and wait for writeback
217  * @bdi: the backing device to write from
218  * @sb: write inodes from this super_block
219  *
220  * Description:
221  *   This does WB_SYNC_ALL data integrity writeback and waits for the
222  *   IO to complete. Callers must hold the sb s_umount semaphore for
223  *   reading, to avoid having the super disappear before we are done.
224  */
225 static void bdi_sync_writeback(struct backing_dev_info *bdi,
226                                struct super_block *sb)
227 {
228         struct wb_writeback_args args = {
229                 .sb             = sb,
230                 .sync_mode      = WB_SYNC_ALL,
231                 .nr_pages       = LONG_MAX,
232                 .range_cyclic   = 0,
233         };
234         struct bdi_work work;
235
236         bdi_work_init(&work, &args);
237         work.state |= WS_ONSTACK;
238
239         bdi_queue_work(bdi, &work);
240         bdi_wait_on_work_clear(&work);
241 }
242
243 /**
244  * bdi_start_writeback - start writeback
245  * @bdi: the backing device to write from
246  * @sb: write inodes from this super_block
247  * @nr_pages: the number of pages to write
248  *
249  * Description:
250  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
251  *   started when this function returns, we make no guarentees on
252  *   completion. Caller need not hold sb s_umount semaphore.
253  *
254  */
255 void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
256                          long nr_pages)
257 {
258         struct wb_writeback_args args = {
259                 .sb             = sb,
260                 .sync_mode      = WB_SYNC_NONE,
261                 .nr_pages       = nr_pages,
262                 .range_cyclic   = 1,
263         };
264
265         /*
266          * We treat @nr_pages=0 as the special case to do background writeback,
267          * ie. to sync pages until the background dirty threshold is reached.
268          */
269         if (!nr_pages) {
270                 args.nr_pages = LONG_MAX;
271                 args.for_background = 1;
272         }
273
274         bdi_alloc_queue_work(bdi, &args);
275 }
276
277 /*
278  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
279  * furthest end of its superblock's dirty-inode list.
280  *
281  * Before stamping the inode's ->dirtied_when, we check to see whether it is
282  * already the most-recently-dirtied inode on the b_dirty list.  If that is
283  * the case then the inode must have been redirtied while it was being written
284  * out and we don't reset its dirtied_when.
285  */
286 static void redirty_tail(struct inode *inode)
287 {
288         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
289
290         if (!list_empty(&wb->b_dirty)) {
291                 struct inode *tail;
292
293                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
294                 if (time_before(inode->dirtied_when, tail->dirtied_when))
295                         inode->dirtied_when = jiffies;
296         }
297         list_move(&inode->i_list, &wb->b_dirty);
298 }
299
300 /*
301  * requeue inode for re-scanning after bdi->b_io list is exhausted.
302  */
303 static void requeue_io(struct inode *inode)
304 {
305         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
306
307         list_move(&inode->i_list, &wb->b_more_io);
308 }
309
310 static void inode_sync_complete(struct inode *inode)
311 {
312         /*
313          * Prevent speculative execution through spin_unlock(&inode_lock);
314          */
315         smp_mb();
316         wake_up_bit(&inode->i_state, __I_SYNC);
317 }
318
319 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
320 {
321         bool ret = time_after(inode->dirtied_when, t);
322 #ifndef CONFIG_64BIT
323         /*
324          * For inodes being constantly redirtied, dirtied_when can get stuck.
325          * It _appears_ to be in the future, but is actually in distant past.
326          * This test is necessary to prevent such wrapped-around relative times
327          * from permanently stopping the whole bdi writeback.
328          */
329         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
330 #endif
331         return ret;
332 }
333
334 /*
335  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
336  */
337 static void move_expired_inodes(struct list_head *delaying_queue,
338                                struct list_head *dispatch_queue,
339                                 unsigned long *older_than_this)
340 {
341         LIST_HEAD(tmp);
342         struct list_head *pos, *node;
343         struct super_block *sb = NULL;
344         struct inode *inode;
345         int do_sb_sort = 0;
346
347         while (!list_empty(delaying_queue)) {
348                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
349                 if (older_than_this &&
350                     inode_dirtied_after(inode, *older_than_this))
351                         break;
352                 if (sb && sb != inode->i_sb)
353                         do_sb_sort = 1;
354                 sb = inode->i_sb;
355                 list_move(&inode->i_list, &tmp);
356         }
357
358         /* just one sb in list, splice to dispatch_queue and we're done */
359         if (!do_sb_sort) {
360                 list_splice(&tmp, dispatch_queue);
361                 return;
362         }
363
364         /* Move inodes from one superblock together */
365         while (!list_empty(&tmp)) {
366                 inode = list_entry(tmp.prev, struct inode, i_list);
367                 sb = inode->i_sb;
368                 list_for_each_prev_safe(pos, node, &tmp) {
369                         inode = list_entry(pos, struct inode, i_list);
370                         if (inode->i_sb == sb)
371                                 list_move(&inode->i_list, dispatch_queue);
372                 }
373         }
374 }
375
376 /*
377  * Queue all expired dirty inodes for io, eldest first.
378  */
379 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
380 {
381         list_splice_init(&wb->b_more_io, wb->b_io.prev);
382         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
383 }
384
385 static int write_inode(struct inode *inode, struct writeback_control *wbc)
386 {
387         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
388                 return inode->i_sb->s_op->write_inode(inode, wbc);
389         return 0;
390 }
391
392 /*
393  * Wait for writeback on an inode to complete.
394  */
395 static void inode_wait_for_writeback(struct inode *inode)
396 {
397         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
398         wait_queue_head_t *wqh;
399
400         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
401         do {
402                 spin_unlock(&inode_lock);
403                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
404                 spin_lock(&inode_lock);
405         } while (inode->i_state & I_SYNC);
406 }
407
408 /*
409  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
410  * caller has ref on the inode (either via __iget or via syscall against an fd)
411  * or the inode has I_WILL_FREE set (via generic_forget_inode)
412  *
413  * If `wait' is set, wait on the writeout.
414  *
415  * The whole writeout design is quite complex and fragile.  We want to avoid
416  * starvation of particular inodes when others are being redirtied, prevent
417  * livelocks, etc.
418  *
419  * Called under inode_lock.
420  */
421 static int
422 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
423 {
424         struct address_space *mapping = inode->i_mapping;
425         unsigned dirty;
426         int ret;
427
428         if (!atomic_read(&inode->i_count))
429                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
430         else
431                 WARN_ON(inode->i_state & I_WILL_FREE);
432
433         if (inode->i_state & I_SYNC) {
434                 /*
435                  * If this inode is locked for writeback and we are not doing
436                  * writeback-for-data-integrity, move it to b_more_io so that
437                  * writeback can proceed with the other inodes on s_io.
438                  *
439                  * We'll have another go at writing back this inode when we
440                  * completed a full scan of b_io.
441                  */
442                 if (wbc->sync_mode != WB_SYNC_ALL) {
443                         requeue_io(inode);
444                         return 0;
445                 }
446
447                 /*
448                  * It's a data-integrity sync.  We must wait.
449                  */
450                 inode_wait_for_writeback(inode);
451         }
452
453         BUG_ON(inode->i_state & I_SYNC);
454
455         /* Set I_SYNC, reset I_DIRTY */
456         dirty = inode->i_state & I_DIRTY;
457         inode->i_state |= I_SYNC;
458         inode->i_state &= ~I_DIRTY;
459
460         spin_unlock(&inode_lock);
461
462         ret = do_writepages(mapping, wbc);
463
464         /*
465          * Make sure to wait on the data before writing out the metadata.
466          * This is important for filesystems that modify metadata on data
467          * I/O completion.
468          */
469         if (wbc->sync_mode == WB_SYNC_ALL) {
470                 int err = filemap_fdatawait(mapping);
471                 if (ret == 0)
472                         ret = err;
473         }
474
475         /* Don't write the inode if only I_DIRTY_PAGES was set */
476         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
477                 int err = write_inode(inode, wbc);
478                 if (ret == 0)
479                         ret = err;
480         }
481
482         spin_lock(&inode_lock);
483         inode->i_state &= ~I_SYNC;
484         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
485                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
486                         /*
487                          * More pages get dirtied by a fast dirtier.
488                          */
489                         goto select_queue;
490                 } else if (inode->i_state & I_DIRTY) {
491                         /*
492                          * At least XFS will redirty the inode during the
493                          * writeback (delalloc) and on io completion (isize).
494                          */
495                         redirty_tail(inode);
496                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
497                         /*
498                          * We didn't write back all the pages.  nfs_writepages()
499                          * sometimes bales out without doing anything. Redirty
500                          * the inode; Move it from b_io onto b_more_io/b_dirty.
501                          */
502                         /*
503                          * akpm: if the caller was the kupdate function we put
504                          * this inode at the head of b_dirty so it gets first
505                          * consideration.  Otherwise, move it to the tail, for
506                          * the reasons described there.  I'm not really sure
507                          * how much sense this makes.  Presumably I had a good
508                          * reasons for doing it this way, and I'd rather not
509                          * muck with it at present.
510                          */
511                         if (wbc->for_kupdate) {
512                                 /*
513                                  * For the kupdate function we move the inode
514                                  * to b_more_io so it will get more writeout as
515                                  * soon as the queue becomes uncongested.
516                                  */
517                                 inode->i_state |= I_DIRTY_PAGES;
518 select_queue:
519                                 if (wbc->nr_to_write <= 0) {
520                                         /*
521                                          * slice used up: queue for next turn
522                                          */
523                                         requeue_io(inode);
524                                 } else {
525                                         /*
526                                          * somehow blocked: retry later
527                                          */
528                                         redirty_tail(inode);
529                                 }
530                         } else {
531                                 /*
532                                  * Otherwise fully redirty the inode so that
533                                  * other inodes on this superblock will get some
534                                  * writeout.  Otherwise heavy writing to one
535                                  * file would indefinitely suspend writeout of
536                                  * all the other files.
537                                  */
538                                 inode->i_state |= I_DIRTY_PAGES;
539                                 redirty_tail(inode);
540                         }
541                 } else if (atomic_read(&inode->i_count)) {
542                         /*
543                          * The inode is clean, inuse
544                          */
545                         list_move(&inode->i_list, &inode_in_use);
546                 } else {
547                         /*
548                          * The inode is clean, unused
549                          */
550                         list_move(&inode->i_list, &inode_unused);
551                 }
552         }
553         inode_sync_complete(inode);
554         return ret;
555 }
556
557 static void unpin_sb_for_writeback(struct super_block *sb)
558 {
559         up_read(&sb->s_umount);
560         put_super(sb);
561 }
562
563 enum sb_pin_state {
564         SB_PINNED,
565         SB_NOT_PINNED,
566         SB_PIN_FAILED
567 };
568
569 /*
570  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
571  * before calling writeback. So make sure that we do pin it, so it doesn't
572  * go away while we are writing inodes from it.
573  */
574 static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
575                                               struct super_block *sb)
576 {
577         /*
578          * Caller must already hold the ref for this
579          */
580         if (wbc->sync_mode == WB_SYNC_ALL) {
581                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
582                 return SB_NOT_PINNED;
583         }
584         spin_lock(&sb_lock);
585         sb->s_count++;
586         if (down_read_trylock(&sb->s_umount)) {
587                 if (sb->s_root) {
588                         spin_unlock(&sb_lock);
589                         return SB_PINNED;
590                 }
591                 /*
592                  * umounted, drop rwsem again and fall through to failure
593                  */
594                 up_read(&sb->s_umount);
595         }
596         sb->s_count--;
597         spin_unlock(&sb_lock);
598         return SB_PIN_FAILED;
599 }
600
601 /*
602  * Write a portion of b_io inodes which belong to @sb.
603  * If @wbc->sb != NULL, then find and write all such
604  * inodes. Otherwise write only ones which go sequentially
605  * in reverse order.
606  * Return 1, if the caller writeback routine should be
607  * interrupted. Otherwise return 0.
608  */
609 static int writeback_sb_inodes(struct super_block *sb,
610                                struct bdi_writeback *wb,
611                                struct writeback_control *wbc)
612 {
613         while (!list_empty(&wb->b_io)) {
614                 long pages_skipped;
615                 struct inode *inode = list_entry(wb->b_io.prev,
616                                                  struct inode, i_list);
617                 if (wbc->sb && sb != inode->i_sb) {
618                         /* super block given and doesn't
619                            match, skip this inode */
620                         redirty_tail(inode);
621                         continue;
622                 }
623                 if (sb != inode->i_sb)
624                         /* finish with this superblock */
625                         return 0;
626                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
627                         requeue_io(inode);
628                         continue;
629                 }
630                 /*
631                  * Was this inode dirtied after sync_sb_inodes was called?
632                  * This keeps sync from extra jobs and livelock.
633                  */
634                 if (inode_dirtied_after(inode, wbc->wb_start))
635                         return 1;
636
637                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
638                 __iget(inode);
639                 pages_skipped = wbc->pages_skipped;
640                 writeback_single_inode(inode, wbc);
641                 if (wbc->pages_skipped != pages_skipped) {
642                         /*
643                          * writeback is not making progress due to locked
644                          * buffers.  Skip this inode for now.
645                          */
646                         redirty_tail(inode);
647                 }
648                 spin_unlock(&inode_lock);
649                 iput(inode);
650                 cond_resched();
651                 spin_lock(&inode_lock);
652                 if (wbc->nr_to_write <= 0) {
653                         wbc->more_io = 1;
654                         return 1;
655                 }
656                 if (!list_empty(&wb->b_more_io))
657                         wbc->more_io = 1;
658         }
659         /* b_io is empty */
660         return 1;
661 }
662
663 static void writeback_inodes_wb(struct bdi_writeback *wb,
664                                 struct writeback_control *wbc)
665 {
666         int ret = 0;
667
668         wbc->wb_start = jiffies; /* livelock avoidance */
669         spin_lock(&inode_lock);
670         if (!wbc->for_kupdate || list_empty(&wb->b_io))
671                 queue_io(wb, wbc->older_than_this);
672
673         while (!list_empty(&wb->b_io)) {
674                 struct inode *inode = list_entry(wb->b_io.prev,
675                                                  struct inode, i_list);
676                 struct super_block *sb = inode->i_sb;
677                 enum sb_pin_state state;
678
679                 if (wbc->sb && sb != wbc->sb) {
680                         /* super block given and doesn't
681                            match, skip this inode */
682                         redirty_tail(inode);
683                         continue;
684                 }
685                 state = pin_sb_for_writeback(wbc, sb);
686
687                 if (state == SB_PIN_FAILED) {
688                         requeue_io(inode);
689                         continue;
690                 }
691                 ret = writeback_sb_inodes(sb, wb, wbc);
692
693                 if (state == SB_PINNED)
694                         unpin_sb_for_writeback(sb);
695                 if (ret)
696                         break;
697         }
698         spin_unlock(&inode_lock);
699         /* Leave any unwritten inodes on b_io */
700 }
701
702 void writeback_inodes_wbc(struct writeback_control *wbc)
703 {
704         struct backing_dev_info *bdi = wbc->bdi;
705
706         writeback_inodes_wb(&bdi->wb, wbc);
707 }
708
709 /*
710  * The maximum number of pages to writeout in a single bdi flush/kupdate
711  * operation.  We do this so we don't hold I_SYNC against an inode for
712  * enormous amounts of time, which would block a userspace task which has
713  * been forced to throttle against that inode.  Also, the code reevaluates
714  * the dirty each time it has written this many pages.
715  */
716 #define MAX_WRITEBACK_PAGES     1024
717
718 static inline bool over_bground_thresh(void)
719 {
720         unsigned long background_thresh, dirty_thresh;
721
722         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
723
724         return (global_page_state(NR_FILE_DIRTY) +
725                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
726 }
727
728 /*
729  * Explicit flushing or periodic writeback of "old" data.
730  *
731  * Define "old": the first time one of an inode's pages is dirtied, we mark the
732  * dirtying-time in the inode's address_space.  So this periodic writeback code
733  * just walks the superblock inode list, writing back any inodes which are
734  * older than a specific point in time.
735  *
736  * Try to run once per dirty_writeback_interval.  But if a writeback event
737  * takes longer than a dirty_writeback_interval interval, then leave a
738  * one-second gap.
739  *
740  * older_than_this takes precedence over nr_to_write.  So we'll only write back
741  * all dirty pages if they are all attached to "old" mappings.
742  */
743 static long wb_writeback(struct bdi_writeback *wb,
744                          struct wb_writeback_args *args)
745 {
746         struct writeback_control wbc = {
747                 .bdi                    = wb->bdi,
748                 .sb                     = args->sb,
749                 .sync_mode              = args->sync_mode,
750                 .older_than_this        = NULL,
751                 .for_kupdate            = args->for_kupdate,
752                 .for_background         = args->for_background,
753                 .range_cyclic           = args->range_cyclic,
754         };
755         unsigned long oldest_jif;
756         long wrote = 0;
757         struct inode *inode;
758
759         if (wbc.for_kupdate) {
760                 wbc.older_than_this = &oldest_jif;
761                 oldest_jif = jiffies -
762                                 msecs_to_jiffies(dirty_expire_interval * 10);
763         }
764         if (!wbc.range_cyclic) {
765                 wbc.range_start = 0;
766                 wbc.range_end = LLONG_MAX;
767         }
768
769         for (;;) {
770                 /*
771                  * Stop writeback when nr_pages has been consumed
772                  */
773                 if (args->nr_pages <= 0)
774                         break;
775
776                 /*
777                  * For background writeout, stop when we are below the
778                  * background dirty threshold
779                  */
780                 if (args->for_background && !over_bground_thresh())
781                         break;
782
783                 wbc.more_io = 0;
784                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
785                 wbc.pages_skipped = 0;
786                 writeback_inodes_wb(wb, &wbc);
787                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
788                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
789
790                 /*
791                  * If we consumed everything, see if we have more
792                  */
793                 if (wbc.nr_to_write <= 0)
794                         continue;
795                 /*
796                  * Didn't write everything and we don't have more IO, bail
797                  */
798                 if (!wbc.more_io)
799                         break;
800                 /*
801                  * Did we write something? Try for more
802                  */
803                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
804                         continue;
805                 /*
806                  * Nothing written. Wait for some inode to
807                  * become available for writeback. Otherwise
808                  * we'll just busyloop.
809                  */
810                 spin_lock(&inode_lock);
811                 if (!list_empty(&wb->b_more_io))  {
812                         inode = list_entry(wb->b_more_io.prev,
813                                                 struct inode, i_list);
814                         inode_wait_for_writeback(inode);
815                 }
816                 spin_unlock(&inode_lock);
817         }
818
819         return wrote;
820 }
821
822 /*
823  * Return the next bdi_work struct that hasn't been processed by this
824  * wb thread yet. ->seen is initially set for each thread that exists
825  * for this device, when a thread first notices a piece of work it
826  * clears its bit. Depending on writeback type, the thread will notify
827  * completion on either receiving the work (WB_SYNC_NONE) or after
828  * it is done (WB_SYNC_ALL).
829  */
830 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
831                                            struct bdi_writeback *wb)
832 {
833         struct bdi_work *work, *ret = NULL;
834
835         rcu_read_lock();
836
837         list_for_each_entry_rcu(work, &bdi->work_list, list) {
838                 if (!test_bit(wb->nr, &work->seen))
839                         continue;
840                 clear_bit(wb->nr, &work->seen);
841
842                 ret = work;
843                 break;
844         }
845
846         rcu_read_unlock();
847         return ret;
848 }
849
850 static long wb_check_old_data_flush(struct bdi_writeback *wb)
851 {
852         unsigned long expired;
853         long nr_pages;
854
855         /*
856          * When set to zero, disable periodic writeback
857          */
858         if (!dirty_writeback_interval)
859                 return 0;
860
861         expired = wb->last_old_flush +
862                         msecs_to_jiffies(dirty_writeback_interval * 10);
863         if (time_before(jiffies, expired))
864                 return 0;
865
866         wb->last_old_flush = jiffies;
867         nr_pages = global_page_state(NR_FILE_DIRTY) +
868                         global_page_state(NR_UNSTABLE_NFS) +
869                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
870
871         if (nr_pages) {
872                 struct wb_writeback_args args = {
873                         .nr_pages       = nr_pages,
874                         .sync_mode      = WB_SYNC_NONE,
875                         .for_kupdate    = 1,
876                         .range_cyclic   = 1,
877                 };
878
879                 return wb_writeback(wb, &args);
880         }
881
882         return 0;
883 }
884
885 /*
886  * Retrieve work items and do the writeback they describe
887  */
888 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
889 {
890         struct backing_dev_info *bdi = wb->bdi;
891         struct bdi_work *work;
892         long wrote = 0;
893
894         while ((work = get_next_work_item(bdi, wb)) != NULL) {
895                 struct wb_writeback_args args = work->args;
896
897                 /*
898                  * Override sync mode, in case we must wait for completion
899                  */
900                 if (force_wait)
901                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
902
903                 /*
904                  * If this isn't a data integrity operation, just notify
905                  * that we have seen this work and we are now starting it.
906                  */
907                 if (args.sync_mode == WB_SYNC_NONE)
908                         wb_clear_pending(wb, work);
909
910                 wrote += wb_writeback(wb, &args);
911
912                 /*
913                  * This is a data integrity writeback, so only do the
914                  * notification when we have completed the work.
915                  */
916                 if (args.sync_mode == WB_SYNC_ALL)
917                         wb_clear_pending(wb, work);
918         }
919
920         /*
921          * Check for periodic writeback, kupdated() style
922          */
923         wrote += wb_check_old_data_flush(wb);
924
925         return wrote;
926 }
927
928 /*
929  * Handle writeback of dirty data for the device backed by this bdi. Also
930  * wakes up periodically and does kupdated style flushing.
931  */
932 int bdi_writeback_task(struct bdi_writeback *wb)
933 {
934         unsigned long last_active = jiffies;
935         unsigned long wait_jiffies = -1UL;
936         long pages_written;
937
938         while (!kthread_should_stop()) {
939                 pages_written = wb_do_writeback(wb, 0);
940
941                 if (pages_written)
942                         last_active = jiffies;
943                 else if (wait_jiffies != -1UL) {
944                         unsigned long max_idle;
945
946                         /*
947                          * Longest period of inactivity that we tolerate. If we
948                          * see dirty data again later, the task will get
949                          * recreated automatically.
950                          */
951                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
952                         if (time_after(jiffies, max_idle + last_active))
953                                 break;
954                 }
955
956                 if (dirty_writeback_interval) {
957                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
958                         schedule_timeout_interruptible(wait_jiffies);
959                 } else
960                         schedule();
961
962                 try_to_freeze();
963         }
964
965         return 0;
966 }
967
968 /*
969  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
970  * writeback, for integrity writeback see bdi_sync_writeback().
971  */
972 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
973 {
974         struct wb_writeback_args args = {
975                 .sb             = sb,
976                 .nr_pages       = nr_pages,
977                 .sync_mode      = WB_SYNC_NONE,
978         };
979         struct backing_dev_info *bdi;
980
981         rcu_read_lock();
982
983         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
984                 if (!bdi_has_dirty_io(bdi))
985                         continue;
986
987                 bdi_alloc_queue_work(bdi, &args);
988         }
989
990         rcu_read_unlock();
991 }
992
993 /*
994  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
995  * the whole world.
996  */
997 void wakeup_flusher_threads(long nr_pages)
998 {
999         if (nr_pages == 0)
1000                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1001                                 global_page_state(NR_UNSTABLE_NFS);
1002         bdi_writeback_all(NULL, nr_pages);
1003 }
1004
1005 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1006 {
1007         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1008                 struct dentry *dentry;
1009                 const char *name = "?";
1010
1011                 dentry = d_find_alias(inode);
1012                 if (dentry) {
1013                         spin_lock(&dentry->d_lock);
1014                         name = (const char *) dentry->d_name.name;
1015                 }
1016                 printk(KERN_DEBUG
1017                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1018                        current->comm, task_pid_nr(current), inode->i_ino,
1019                        name, inode->i_sb->s_id);
1020                 if (dentry) {
1021                         spin_unlock(&dentry->d_lock);
1022                         dput(dentry);
1023                 }
1024         }
1025 }
1026
1027 /**
1028  *      __mark_inode_dirty -    internal function
1029  *      @inode: inode to mark
1030  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1031  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1032  *      mark_inode_dirty_sync.
1033  *
1034  * Put the inode on the super block's dirty list.
1035  *
1036  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1037  * dirty list only if it is hashed or if it refers to a blockdev.
1038  * If it was not hashed, it will never be added to the dirty list
1039  * even if it is later hashed, as it will have been marked dirty already.
1040  *
1041  * In short, make sure you hash any inodes _before_ you start marking
1042  * them dirty.
1043  *
1044  * This function *must* be atomic for the I_DIRTY_PAGES case -
1045  * set_page_dirty() is called under spinlock in several places.
1046  *
1047  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1048  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1049  * the kernel-internal blockdev inode represents the dirtying time of the
1050  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1051  * page->mapping->host, so the page-dirtying time is recorded in the internal
1052  * blockdev inode.
1053  */
1054 void __mark_inode_dirty(struct inode *inode, int flags)
1055 {
1056         struct super_block *sb = inode->i_sb;
1057
1058         /*
1059          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1060          * dirty the inode itself
1061          */
1062         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1063                 if (sb->s_op->dirty_inode)
1064                         sb->s_op->dirty_inode(inode);
1065         }
1066
1067         /*
1068          * make sure that changes are seen by all cpus before we test i_state
1069          * -- mikulas
1070          */
1071         smp_mb();
1072
1073         /* avoid the locking if we can */
1074         if ((inode->i_state & flags) == flags)
1075                 return;
1076
1077         if (unlikely(block_dump))
1078                 block_dump___mark_inode_dirty(inode);
1079
1080         spin_lock(&inode_lock);
1081         if ((inode->i_state & flags) != flags) {
1082                 const int was_dirty = inode->i_state & I_DIRTY;
1083
1084                 inode->i_state |= flags;
1085
1086                 /*
1087                  * If the inode is being synced, just update its dirty state.
1088                  * The unlocker will place the inode on the appropriate
1089                  * superblock list, based upon its state.
1090                  */
1091                 if (inode->i_state & I_SYNC)
1092                         goto out;
1093
1094                 /*
1095                  * Only add valid (hashed) inodes to the superblock's
1096                  * dirty list.  Add blockdev inodes as well.
1097                  */
1098                 if (!S_ISBLK(inode->i_mode)) {
1099                         if (hlist_unhashed(&inode->i_hash))
1100                                 goto out;
1101                 }
1102                 if (inode->i_state & (I_FREEING|I_CLEAR))
1103                         goto out;
1104
1105                 /*
1106                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1107                  * reposition it (that would break b_dirty time-ordering).
1108                  */
1109                 if (!was_dirty) {
1110                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1111                         struct backing_dev_info *bdi = wb->bdi;
1112
1113                         if (bdi_cap_writeback_dirty(bdi) &&
1114                             !test_bit(BDI_registered, &bdi->state)) {
1115                                 WARN_ON(1);
1116                                 printk(KERN_ERR "bdi-%s not registered\n",
1117                                                                 bdi->name);
1118                         }
1119
1120                         inode->dirtied_when = jiffies;
1121                         list_move(&inode->i_list, &wb->b_dirty);
1122                 }
1123         }
1124 out:
1125         spin_unlock(&inode_lock);
1126 }
1127 EXPORT_SYMBOL(__mark_inode_dirty);
1128
1129 /*
1130  * Write out a superblock's list of dirty inodes.  A wait will be performed
1131  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1132  *
1133  * If older_than_this is non-NULL, then only write out inodes which
1134  * had their first dirtying at a time earlier than *older_than_this.
1135  *
1136  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1137  * This function assumes that the blockdev superblock's inodes are backed by
1138  * a variety of queues, so all inodes are searched.  For other superblocks,
1139  * assume that all inodes are backed by the same queue.
1140  *
1141  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1142  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1143  * on the writer throttling path, and we get decent balancing between many
1144  * throttled threads: we don't want them all piling up on inode_sync_wait.
1145  */
1146 static void wait_sb_inodes(struct super_block *sb)
1147 {
1148         struct inode *inode, *old_inode = NULL;
1149
1150         /*
1151          * We need to be protected against the filesystem going from
1152          * r/o to r/w or vice versa.
1153          */
1154         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1155
1156         spin_lock(&inode_lock);
1157
1158         /*
1159          * Data integrity sync. Must wait for all pages under writeback,
1160          * because there may have been pages dirtied before our sync
1161          * call, but which had writeout started before we write it out.
1162          * In which case, the inode may not be on the dirty list, but
1163          * we still have to wait for that writeout.
1164          */
1165         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1166                 struct address_space *mapping;
1167
1168                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1169                         continue;
1170                 mapping = inode->i_mapping;
1171                 if (mapping->nrpages == 0)
1172                         continue;
1173                 __iget(inode);
1174                 spin_unlock(&inode_lock);
1175                 /*
1176                  * We hold a reference to 'inode' so it couldn't have
1177                  * been removed from s_inodes list while we dropped the
1178                  * inode_lock.  We cannot iput the inode now as we can
1179                  * be holding the last reference and we cannot iput it
1180                  * under inode_lock. So we keep the reference and iput
1181                  * it later.
1182                  */
1183                 iput(old_inode);
1184                 old_inode = inode;
1185
1186                 filemap_fdatawait(mapping);
1187
1188                 cond_resched();
1189
1190                 spin_lock(&inode_lock);
1191         }
1192         spin_unlock(&inode_lock);
1193         iput(old_inode);
1194 }
1195
1196 /**
1197  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1198  * @sb: the superblock
1199  *
1200  * Start writeback on some inodes on this super_block. No guarantees are made
1201  * on how many (if any) will be written, and this function does not wait
1202  * for IO completion of submitted IO. The number of pages submitted is
1203  * returned.
1204  */
1205 void writeback_inodes_sb(struct super_block *sb)
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);
1215 }
1216 EXPORT_SYMBOL(writeback_inodes_sb);
1217
1218 /**
1219  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1220  * @sb: the superblock
1221  *
1222  * Invoke writeback_inodes_sb if no writeback is currently underway.
1223  * Returns 1 if writeback was started, 0 if not.
1224  */
1225 int writeback_inodes_sb_if_idle(struct super_block *sb)
1226 {
1227         if (!writeback_in_progress(sb->s_bdi)) {
1228                 writeback_inodes_sb(sb);
1229                 return 1;
1230         } else
1231                 return 0;
1232 }
1233 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1234
1235 /**
1236  * sync_inodes_sb       -       sync sb inode pages
1237  * @sb: the superblock
1238  *
1239  * This function writes and waits on any dirty inode belonging to this
1240  * super_block. The number of pages synced is returned.
1241  */
1242 void sync_inodes_sb(struct super_block *sb)
1243 {
1244         bdi_sync_writeback(sb->s_bdi, sb);
1245         wait_sb_inodes(sb);
1246 }
1247 EXPORT_SYMBOL(sync_inodes_sb);
1248
1249 /**
1250  * write_inode_now      -       write an inode to disk
1251  * @inode: inode to write to disk
1252  * @sync: whether the write should be synchronous or not
1253  *
1254  * This function commits an inode to disk immediately if it is dirty. This is
1255  * primarily needed by knfsd.
1256  *
1257  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1258  */
1259 int write_inode_now(struct inode *inode, int sync)
1260 {
1261         int ret;
1262         struct writeback_control wbc = {
1263                 .nr_to_write = LONG_MAX,
1264                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1265                 .range_start = 0,
1266                 .range_end = LLONG_MAX,
1267         };
1268
1269         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1270                 wbc.nr_to_write = 0;
1271
1272         might_sleep();
1273         spin_lock(&inode_lock);
1274         ret = writeback_single_inode(inode, &wbc);
1275         spin_unlock(&inode_lock);
1276         if (sync)
1277                 inode_sync_wait(inode);
1278         return ret;
1279 }
1280 EXPORT_SYMBOL(write_inode_now);
1281
1282 /**
1283  * sync_inode - write an inode and its pages to disk.
1284  * @inode: the inode to sync
1285  * @wbc: controls the writeback mode
1286  *
1287  * sync_inode() will write an inode and its pages to disk.  It will also
1288  * correctly update the inode on its superblock's dirty inode lists and will
1289  * update inode->i_state.
1290  *
1291  * The caller must have a ref on the inode.
1292  */
1293 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1294 {
1295         int ret;
1296
1297         spin_lock(&inode_lock);
1298         ret = writeback_single_inode(inode, wbc);
1299         spin_unlock(&inode_lock);
1300         return ret;
1301 }
1302 EXPORT_SYMBOL(sync_inode);