nfs4 use mandatory attribute file type in nfs4_get_root
[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         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 {
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          while (inode->i_state & I_SYNC) {
402                 spin_unlock(&inode_lock);
403                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
404                 spin_lock(&inode_lock);
405         }
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_PAGES */
456         inode->i_state |= I_SYNC;
457         inode->i_state &= ~I_DIRTY_PAGES;
458         spin_unlock(&inode_lock);
459
460         ret = do_writepages(mapping, wbc);
461
462         /*
463          * Make sure to wait on the data before writing out the metadata.
464          * This is important for filesystems that modify metadata on data
465          * I/O completion.
466          */
467         if (wbc->sync_mode == WB_SYNC_ALL) {
468                 int err = filemap_fdatawait(mapping);
469                 if (ret == 0)
470                         ret = err;
471         }
472
473         /*
474          * Some filesystems may redirty the inode during the writeback
475          * due to delalloc, clear dirty metadata flags right before
476          * write_inode()
477          */
478         spin_lock(&inode_lock);
479         dirty = inode->i_state & I_DIRTY;
480         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
481         spin_unlock(&inode_lock);
482         /* Don't write the inode if only I_DIRTY_PAGES was set */
483         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
484                 int err = write_inode(inode, wbc);
485                 if (ret == 0)
486                         ret = err;
487         }
488
489         spin_lock(&inode_lock);
490         inode->i_state &= ~I_SYNC;
491         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
492                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
493                         /*
494                          * More pages get dirtied by a fast dirtier.
495                          */
496                         goto select_queue;
497                 } else if (inode->i_state & I_DIRTY) {
498                         /*
499                          * At least XFS will redirty the inode during the
500                          * writeback (delalloc) and on io completion (isize).
501                          */
502                         redirty_tail(inode);
503                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
504                         /*
505                          * We didn't write back all the pages.  nfs_writepages()
506                          * sometimes bales out without doing anything. Redirty
507                          * the inode; Move it from b_io onto b_more_io/b_dirty.
508                          */
509                         /*
510                          * akpm: if the caller was the kupdate function we put
511                          * this inode at the head of b_dirty so it gets first
512                          * consideration.  Otherwise, move it to the tail, for
513                          * the reasons described there.  I'm not really sure
514                          * how much sense this makes.  Presumably I had a good
515                          * reasons for doing it this way, and I'd rather not
516                          * muck with it at present.
517                          */
518                         if (wbc->for_kupdate) {
519                                 /*
520                                  * For the kupdate function we move the inode
521                                  * to b_more_io so it will get more writeout as
522                                  * soon as the queue becomes uncongested.
523                                  */
524                                 inode->i_state |= I_DIRTY_PAGES;
525 select_queue:
526                                 if (wbc->nr_to_write <= 0) {
527                                         /*
528                                          * slice used up: queue for next turn
529                                          */
530                                         requeue_io(inode);
531                                 } else {
532                                         /*
533                                          * somehow blocked: retry later
534                                          */
535                                         redirty_tail(inode);
536                                 }
537                         } else {
538                                 /*
539                                  * Otherwise fully redirty the inode so that
540                                  * other inodes on this superblock will get some
541                                  * writeout.  Otherwise heavy writing to one
542                                  * file would indefinitely suspend writeout of
543                                  * all the other files.
544                                  */
545                                 inode->i_state |= I_DIRTY_PAGES;
546                                 redirty_tail(inode);
547                         }
548                 } else if (atomic_read(&inode->i_count)) {
549                         /*
550                          * The inode is clean, inuse
551                          */
552                         list_move(&inode->i_list, &inode_in_use);
553                 } else {
554                         /*
555                          * The inode is clean, unused
556                          */
557                         list_move(&inode->i_list, &inode_unused);
558                 }
559         }
560         inode_sync_complete(inode);
561         return ret;
562 }
563
564 static void unpin_sb_for_writeback(struct super_block *sb)
565 {
566         up_read(&sb->s_umount);
567         put_super(sb);
568 }
569
570 enum sb_pin_state {
571         SB_PINNED,
572         SB_NOT_PINNED,
573         SB_PIN_FAILED
574 };
575
576 /*
577  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
578  * before calling writeback. So make sure that we do pin it, so it doesn't
579  * go away while we are writing inodes from it.
580  */
581 static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
582                                               struct super_block *sb)
583 {
584         /*
585          * Caller must already hold the ref for this
586          */
587         if (wbc->sync_mode == WB_SYNC_ALL) {
588                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
589                 return SB_NOT_PINNED;
590         }
591         spin_lock(&sb_lock);
592         sb->s_count++;
593         if (down_read_trylock(&sb->s_umount)) {
594                 if (sb->s_root) {
595                         spin_unlock(&sb_lock);
596                         return SB_PINNED;
597                 }
598                 /*
599                  * umounted, drop rwsem again and fall through to failure
600                  */
601                 up_read(&sb->s_umount);
602         }
603         sb->s_count--;
604         spin_unlock(&sb_lock);
605         return SB_PIN_FAILED;
606 }
607
608 /*
609  * Write a portion of b_io inodes which belong to @sb.
610  * If @wbc->sb != NULL, then find and write all such
611  * inodes. Otherwise write only ones which go sequentially
612  * in reverse order.
613  * Return 1, if the caller writeback routine should be
614  * interrupted. Otherwise return 0.
615  */
616 static int writeback_sb_inodes(struct super_block *sb,
617                                struct bdi_writeback *wb,
618                                struct writeback_control *wbc)
619 {
620         while (!list_empty(&wb->b_io)) {
621                 long pages_skipped;
622                 struct inode *inode = list_entry(wb->b_io.prev,
623                                                  struct inode, i_list);
624                 if (wbc->sb && sb != inode->i_sb) {
625                         /* super block given and doesn't
626                            match, skip this inode */
627                         redirty_tail(inode);
628                         continue;
629                 }
630                 if (sb != inode->i_sb)
631                         /* finish with this superblock */
632                         return 0;
633                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
634                         requeue_io(inode);
635                         continue;
636                 }
637                 /*
638                  * Was this inode dirtied after sync_sb_inodes was called?
639                  * This keeps sync from extra jobs and livelock.
640                  */
641                 if (inode_dirtied_after(inode, wbc->wb_start))
642                         return 1;
643
644                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
645                 __iget(inode);
646                 pages_skipped = wbc->pages_skipped;
647                 writeback_single_inode(inode, wbc);
648                 if (wbc->pages_skipped != pages_skipped) {
649                         /*
650                          * writeback is not making progress due to locked
651                          * buffers.  Skip this inode for now.
652                          */
653                         redirty_tail(inode);
654                 }
655                 spin_unlock(&inode_lock);
656                 iput(inode);
657                 cond_resched();
658                 spin_lock(&inode_lock);
659                 if (wbc->nr_to_write <= 0) {
660                         wbc->more_io = 1;
661                         return 1;
662                 }
663                 if (!list_empty(&wb->b_more_io))
664                         wbc->more_io = 1;
665         }
666         /* b_io is empty */
667         return 1;
668 }
669
670 static void writeback_inodes_wb(struct bdi_writeback *wb,
671                                 struct writeback_control *wbc)
672 {
673         int ret = 0;
674
675         wbc->wb_start = jiffies; /* livelock avoidance */
676         spin_lock(&inode_lock);
677         if (!wbc->for_kupdate || list_empty(&wb->b_io))
678                 queue_io(wb, wbc->older_than_this);
679
680         while (!list_empty(&wb->b_io)) {
681                 struct inode *inode = list_entry(wb->b_io.prev,
682                                                  struct inode, i_list);
683                 struct super_block *sb = inode->i_sb;
684                 enum sb_pin_state state;
685
686                 if (wbc->sb && sb != wbc->sb) {
687                         /* super block given and doesn't
688                            match, skip this inode */
689                         redirty_tail(inode);
690                         continue;
691                 }
692                 state = pin_sb_for_writeback(wbc, sb);
693
694                 if (state == SB_PIN_FAILED) {
695                         requeue_io(inode);
696                         continue;
697                 }
698                 ret = writeback_sb_inodes(sb, wb, wbc);
699
700                 if (state == SB_PINNED)
701                         unpin_sb_for_writeback(sb);
702                 if (ret)
703                         break;
704         }
705         spin_unlock(&inode_lock);
706         /* Leave any unwritten inodes on b_io */
707 }
708
709 void writeback_inodes_wbc(struct writeback_control *wbc)
710 {
711         struct backing_dev_info *bdi = wbc->bdi;
712
713         writeback_inodes_wb(&bdi->wb, wbc);
714 }
715
716 /*
717  * The maximum number of pages to writeout in a single bdi flush/kupdate
718  * operation.  We do this so we don't hold I_SYNC against an inode for
719  * enormous amounts of time, which would block a userspace task which has
720  * been forced to throttle against that inode.  Also, the code reevaluates
721  * the dirty each time it has written this many pages.
722  */
723 #define MAX_WRITEBACK_PAGES     1024
724
725 static inline bool over_bground_thresh(void)
726 {
727         unsigned long background_thresh, dirty_thresh;
728
729         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
730
731         return (global_page_state(NR_FILE_DIRTY) +
732                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
733 }
734
735 /*
736  * Explicit flushing or periodic writeback of "old" data.
737  *
738  * Define "old": the first time one of an inode's pages is dirtied, we mark the
739  * dirtying-time in the inode's address_space.  So this periodic writeback code
740  * just walks the superblock inode list, writing back any inodes which are
741  * older than a specific point in time.
742  *
743  * Try to run once per dirty_writeback_interval.  But if a writeback event
744  * takes longer than a dirty_writeback_interval interval, then leave a
745  * one-second gap.
746  *
747  * older_than_this takes precedence over nr_to_write.  So we'll only write back
748  * all dirty pages if they are all attached to "old" mappings.
749  */
750 static long wb_writeback(struct bdi_writeback *wb,
751                          struct wb_writeback_args *args)
752 {
753         struct writeback_control wbc = {
754                 .bdi                    = wb->bdi,
755                 .sb                     = args->sb,
756                 .sync_mode              = args->sync_mode,
757                 .older_than_this        = NULL,
758                 .for_kupdate            = args->for_kupdate,
759                 .for_background         = args->for_background,
760                 .range_cyclic           = args->range_cyclic,
761         };
762         unsigned long oldest_jif;
763         long wrote = 0;
764         struct inode *inode;
765
766         if (wbc.for_kupdate) {
767                 wbc.older_than_this = &oldest_jif;
768                 oldest_jif = jiffies -
769                                 msecs_to_jiffies(dirty_expire_interval * 10);
770         }
771         if (!wbc.range_cyclic) {
772                 wbc.range_start = 0;
773                 wbc.range_end = LLONG_MAX;
774         }
775
776         for (;;) {
777                 /*
778                  * Stop writeback when nr_pages has been consumed
779                  */
780                 if (args->nr_pages <= 0)
781                         break;
782
783                 /*
784                  * For background writeout, stop when we are below the
785                  * background dirty threshold
786                  */
787                 if (args->for_background && !over_bground_thresh())
788                         break;
789
790                 wbc.more_io = 0;
791                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
792                 wbc.pages_skipped = 0;
793                 writeback_inodes_wb(wb, &wbc);
794                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
795                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
796
797                 /*
798                  * If we consumed everything, see if we have more
799                  */
800                 if (wbc.nr_to_write <= 0)
801                         continue;
802                 /*
803                  * Didn't write everything and we don't have more IO, bail
804                  */
805                 if (!wbc.more_io)
806                         break;
807                 /*
808                  * Did we write something? Try for more
809                  */
810                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
811                         continue;
812                 /*
813                  * Nothing written. Wait for some inode to
814                  * become available for writeback. Otherwise
815                  * we'll just busyloop.
816                  */
817                 spin_lock(&inode_lock);
818                 if (!list_empty(&wb->b_more_io))  {
819                         inode = list_entry(wb->b_more_io.prev,
820                                                 struct inode, i_list);
821                         inode_wait_for_writeback(inode);
822                 }
823                 spin_unlock(&inode_lock);
824         }
825
826         return wrote;
827 }
828
829 /*
830  * Return the next bdi_work struct that hasn't been processed by this
831  * wb thread yet. ->seen is initially set for each thread that exists
832  * for this device, when a thread first notices a piece of work it
833  * clears its bit. Depending on writeback type, the thread will notify
834  * completion on either receiving the work (WB_SYNC_NONE) or after
835  * it is done (WB_SYNC_ALL).
836  */
837 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
838                                            struct bdi_writeback *wb)
839 {
840         struct bdi_work *work, *ret = NULL;
841
842         rcu_read_lock();
843
844         list_for_each_entry_rcu(work, &bdi->work_list, list) {
845                 if (!test_bit(wb->nr, &work->seen))
846                         continue;
847                 clear_bit(wb->nr, &work->seen);
848
849                 ret = work;
850                 break;
851         }
852
853         rcu_read_unlock();
854         return ret;
855 }
856
857 static long wb_check_old_data_flush(struct bdi_writeback *wb)
858 {
859         unsigned long expired;
860         long nr_pages;
861
862         /*
863          * When set to zero, disable periodic writeback
864          */
865         if (!dirty_writeback_interval)
866                 return 0;
867
868         expired = wb->last_old_flush +
869                         msecs_to_jiffies(dirty_writeback_interval * 10);
870         if (time_before(jiffies, expired))
871                 return 0;
872
873         wb->last_old_flush = jiffies;
874         nr_pages = global_page_state(NR_FILE_DIRTY) +
875                         global_page_state(NR_UNSTABLE_NFS) +
876                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
877
878         if (nr_pages) {
879                 struct wb_writeback_args args = {
880                         .nr_pages       = nr_pages,
881                         .sync_mode      = WB_SYNC_NONE,
882                         .for_kupdate    = 1,
883                         .range_cyclic   = 1,
884                 };
885
886                 return wb_writeback(wb, &args);
887         }
888
889         return 0;
890 }
891
892 /*
893  * Retrieve work items and do the writeback they describe
894  */
895 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
896 {
897         struct backing_dev_info *bdi = wb->bdi;
898         struct bdi_work *work;
899         long wrote = 0;
900
901         while ((work = get_next_work_item(bdi, wb)) != NULL) {
902                 struct wb_writeback_args args = work->args;
903
904                 /*
905                  * Override sync mode, in case we must wait for completion
906                  */
907                 if (force_wait)
908                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
909
910                 /*
911                  * If this isn't a data integrity operation, just notify
912                  * that we have seen this work and we are now starting it.
913                  */
914                 if (args.sync_mode == WB_SYNC_NONE)
915                         wb_clear_pending(wb, work);
916
917                 wrote += wb_writeback(wb, &args);
918
919                 /*
920                  * This is a data integrity writeback, so only do the
921                  * notification when we have completed the work.
922                  */
923                 if (args.sync_mode == WB_SYNC_ALL)
924                         wb_clear_pending(wb, work);
925         }
926
927         /*
928          * Check for periodic writeback, kupdated() style
929          */
930         wrote += wb_check_old_data_flush(wb);
931
932         return wrote;
933 }
934
935 /*
936  * Handle writeback of dirty data for the device backed by this bdi. Also
937  * wakes up periodically and does kupdated style flushing.
938  */
939 int bdi_writeback_task(struct bdi_writeback *wb)
940 {
941         unsigned long last_active = jiffies;
942         unsigned long wait_jiffies = -1UL;
943         long pages_written;
944
945         while (!kthread_should_stop()) {
946                 pages_written = wb_do_writeback(wb, 0);
947
948                 if (pages_written)
949                         last_active = jiffies;
950                 else if (wait_jiffies != -1UL) {
951                         unsigned long max_idle;
952
953                         /*
954                          * Longest period of inactivity that we tolerate. If we
955                          * see dirty data again later, the task will get
956                          * recreated automatically.
957                          */
958                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
959                         if (time_after(jiffies, max_idle + last_active))
960                                 break;
961                 }
962
963                 if (dirty_writeback_interval) {
964                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
965                         schedule_timeout_interruptible(wait_jiffies);
966                 } else {
967                         set_current_state(TASK_INTERRUPTIBLE);
968                         if (list_empty_careful(&wb->bdi->work_list) &&
969                             !kthread_should_stop())
970                                 schedule();
971                         __set_current_state(TASK_RUNNING);
972                 }
973
974                 try_to_freeze();
975         }
976
977         return 0;
978 }
979
980 /*
981  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
982  * writeback, for integrity writeback see bdi_sync_writeback().
983  */
984 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
985 {
986         struct wb_writeback_args args = {
987                 .sb             = sb,
988                 .nr_pages       = nr_pages,
989                 .sync_mode      = WB_SYNC_NONE,
990         };
991         struct backing_dev_info *bdi;
992
993         rcu_read_lock();
994
995         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
996                 if (!bdi_has_dirty_io(bdi))
997                         continue;
998
999                 bdi_alloc_queue_work(bdi, &args);
1000         }
1001
1002         rcu_read_unlock();
1003 }
1004
1005 /*
1006  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1007  * the whole world.
1008  */
1009 void wakeup_flusher_threads(long nr_pages)
1010 {
1011         if (nr_pages == 0)
1012                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1013                                 global_page_state(NR_UNSTABLE_NFS);
1014         bdi_writeback_all(NULL, nr_pages);
1015 }
1016
1017 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1018 {
1019         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1020                 struct dentry *dentry;
1021                 const char *name = "?";
1022
1023                 dentry = d_find_alias(inode);
1024                 if (dentry) {
1025                         spin_lock(&dentry->d_lock);
1026                         name = (const char *) dentry->d_name.name;
1027                 }
1028                 printk(KERN_DEBUG
1029                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1030                        current->comm, task_pid_nr(current), inode->i_ino,
1031                        name, inode->i_sb->s_id);
1032                 if (dentry) {
1033                         spin_unlock(&dentry->d_lock);
1034                         dput(dentry);
1035                 }
1036         }
1037 }
1038
1039 /**
1040  *      __mark_inode_dirty -    internal function
1041  *      @inode: inode to mark
1042  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1043  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1044  *      mark_inode_dirty_sync.
1045  *
1046  * Put the inode on the super block's dirty list.
1047  *
1048  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1049  * dirty list only if it is hashed or if it refers to a blockdev.
1050  * If it was not hashed, it will never be added to the dirty list
1051  * even if it is later hashed, as it will have been marked dirty already.
1052  *
1053  * In short, make sure you hash any inodes _before_ you start marking
1054  * them dirty.
1055  *
1056  * This function *must* be atomic for the I_DIRTY_PAGES case -
1057  * set_page_dirty() is called under spinlock in several places.
1058  *
1059  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1060  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1061  * the kernel-internal blockdev inode represents the dirtying time of the
1062  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1063  * page->mapping->host, so the page-dirtying time is recorded in the internal
1064  * blockdev inode.
1065  */
1066 void __mark_inode_dirty(struct inode *inode, int flags)
1067 {
1068         struct super_block *sb = inode->i_sb;
1069
1070         /*
1071          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1072          * dirty the inode itself
1073          */
1074         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1075                 if (sb->s_op->dirty_inode)
1076                         sb->s_op->dirty_inode(inode);
1077         }
1078
1079         /*
1080          * make sure that changes are seen by all cpus before we test i_state
1081          * -- mikulas
1082          */
1083         smp_mb();
1084
1085         /* avoid the locking if we can */
1086         if ((inode->i_state & flags) == flags)
1087                 return;
1088
1089         if (unlikely(block_dump))
1090                 block_dump___mark_inode_dirty(inode);
1091
1092         spin_lock(&inode_lock);
1093         if ((inode->i_state & flags) != flags) {
1094                 const int was_dirty = inode->i_state & I_DIRTY;
1095
1096                 inode->i_state |= flags;
1097
1098                 /*
1099                  * If the inode is being synced, just update its dirty state.
1100                  * The unlocker will place the inode on the appropriate
1101                  * superblock list, based upon its state.
1102                  */
1103                 if (inode->i_state & I_SYNC)
1104                         goto out;
1105
1106                 /*
1107                  * Only add valid (hashed) inodes to the superblock's
1108                  * dirty list.  Add blockdev inodes as well.
1109                  */
1110                 if (!S_ISBLK(inode->i_mode)) {
1111                         if (hlist_unhashed(&inode->i_hash))
1112                                 goto out;
1113                 }
1114                 if (inode->i_state & (I_FREEING|I_CLEAR))
1115                         goto out;
1116
1117                 /*
1118                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1119                  * reposition it (that would break b_dirty time-ordering).
1120                  */
1121                 if (!was_dirty) {
1122                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1123                         struct backing_dev_info *bdi = wb->bdi;
1124
1125                         if (bdi_cap_writeback_dirty(bdi) &&
1126                             !test_bit(BDI_registered, &bdi->state)) {
1127                                 WARN_ON(1);
1128                                 printk(KERN_ERR "bdi-%s not registered\n",
1129                                                                 bdi->name);
1130                         }
1131
1132                         inode->dirtied_when = jiffies;
1133                         list_move(&inode->i_list, &wb->b_dirty);
1134                 }
1135         }
1136 out:
1137         spin_unlock(&inode_lock);
1138 }
1139 EXPORT_SYMBOL(__mark_inode_dirty);
1140
1141 /*
1142  * Write out a superblock's list of dirty inodes.  A wait will be performed
1143  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1144  *
1145  * If older_than_this is non-NULL, then only write out inodes which
1146  * had their first dirtying at a time earlier than *older_than_this.
1147  *
1148  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1149  * This function assumes that the blockdev superblock's inodes are backed by
1150  * a variety of queues, so all inodes are searched.  For other superblocks,
1151  * assume that all inodes are backed by the same queue.
1152  *
1153  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1154  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1155  * on the writer throttling path, and we get decent balancing between many
1156  * throttled threads: we don't want them all piling up on inode_sync_wait.
1157  */
1158 static void wait_sb_inodes(struct super_block *sb)
1159 {
1160         struct inode *inode, *old_inode = NULL;
1161
1162         /*
1163          * We need to be protected against the filesystem going from
1164          * r/o to r/w or vice versa.
1165          */
1166         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1167
1168         spin_lock(&inode_lock);
1169
1170         /*
1171          * Data integrity sync. Must wait for all pages under writeback,
1172          * because there may have been pages dirtied before our sync
1173          * call, but which had writeout started before we write it out.
1174          * In which case, the inode may not be on the dirty list, but
1175          * we still have to wait for that writeout.
1176          */
1177         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1178                 struct address_space *mapping;
1179
1180                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1181                         continue;
1182                 mapping = inode->i_mapping;
1183                 if (mapping->nrpages == 0)
1184                         continue;
1185                 __iget(inode);
1186                 spin_unlock(&inode_lock);
1187                 /*
1188                  * We hold a reference to 'inode' so it couldn't have
1189                  * been removed from s_inodes list while we dropped the
1190                  * inode_lock.  We cannot iput the inode now as we can
1191                  * be holding the last reference and we cannot iput it
1192                  * under inode_lock. So we keep the reference and iput
1193                  * it later.
1194                  */
1195                 iput(old_inode);
1196                 old_inode = inode;
1197
1198                 filemap_fdatawait(mapping);
1199
1200                 cond_resched();
1201
1202                 spin_lock(&inode_lock);
1203         }
1204         spin_unlock(&inode_lock);
1205         iput(old_inode);
1206 }
1207
1208 /**
1209  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1210  * @sb: the superblock
1211  *
1212  * Start writeback on some inodes on this super_block. No guarantees are made
1213  * on how many (if any) will be written, and this function does not wait
1214  * for IO completion of submitted IO. The number of pages submitted is
1215  * returned.
1216  */
1217 void writeback_inodes_sb(struct super_block *sb)
1218 {
1219         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1220         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1221         long nr_to_write;
1222
1223         nr_to_write = nr_dirty + nr_unstable +
1224                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1225
1226         bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
1227 }
1228 EXPORT_SYMBOL(writeback_inodes_sb);
1229
1230 /**
1231  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1232  * @sb: the superblock
1233  *
1234  * Invoke writeback_inodes_sb if no writeback is currently underway.
1235  * Returns 1 if writeback was started, 0 if not.
1236  */
1237 int writeback_inodes_sb_if_idle(struct super_block *sb)
1238 {
1239         if (!writeback_in_progress(sb->s_bdi)) {
1240                 writeback_inodes_sb(sb);
1241                 return 1;
1242         } else
1243                 return 0;
1244 }
1245 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1246
1247 /**
1248  * sync_inodes_sb       -       sync sb inode pages
1249  * @sb: the superblock
1250  *
1251  * This function writes and waits on any dirty inode belonging to this
1252  * super_block. The number of pages synced is returned.
1253  */
1254 void sync_inodes_sb(struct super_block *sb)
1255 {
1256         bdi_sync_writeback(sb->s_bdi, sb);
1257         wait_sb_inodes(sb);
1258 }
1259 EXPORT_SYMBOL(sync_inodes_sb);
1260
1261 /**
1262  * write_inode_now      -       write an inode to disk
1263  * @inode: inode to write to disk
1264  * @sync: whether the write should be synchronous or not
1265  *
1266  * This function commits an inode to disk immediately if it is dirty. This is
1267  * primarily needed by knfsd.
1268  *
1269  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1270  */
1271 int write_inode_now(struct inode *inode, int sync)
1272 {
1273         int ret;
1274         struct writeback_control wbc = {
1275                 .nr_to_write = LONG_MAX,
1276                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1277                 .range_start = 0,
1278                 .range_end = LLONG_MAX,
1279         };
1280
1281         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1282                 wbc.nr_to_write = 0;
1283
1284         might_sleep();
1285         spin_lock(&inode_lock);
1286         ret = writeback_single_inode(inode, &wbc);
1287         spin_unlock(&inode_lock);
1288         if (sync)
1289                 inode_sync_wait(inode);
1290         return ret;
1291 }
1292 EXPORT_SYMBOL(write_inode_now);
1293
1294 /**
1295  * sync_inode - write an inode and its pages to disk.
1296  * @inode: the inode to sync
1297  * @wbc: controls the writeback mode
1298  *
1299  * sync_inode() will write an inode and its pages to disk.  It will also
1300  * correctly update the inode on its superblock's dirty inode lists and will
1301  * update inode->i_state.
1302  *
1303  * The caller must have a ref on the inode.
1304  */
1305 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1306 {
1307         int ret;
1308
1309         spin_lock(&inode_lock);
1310         ret = writeback_single_inode(inode, wbc);
1311         spin_unlock(&inode_lock);
1312         return ret;
1313 }
1314 EXPORT_SYMBOL(sync_inode);