6753912641b4b1846776d4b72eaca555a516f9fb
[pandora-kernel.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include "internal.h"
30
31 #define inode_to_bdi(inode)     ((inode)->i_mapping->backing_dev_info)
32
33 /*
34  * We don't actually have pdflush, but this one is exported though /proc...
35  */
36 int nr_pdflush_threads;
37
38 /*
39  * Passed into wb_writeback(), essentially a subset of writeback_control
40  */
41 struct wb_writeback_args {
42         long nr_pages;
43         struct super_block *sb;
44         enum writeback_sync_modes sync_mode;
45         unsigned int for_kupdate:1;
46         unsigned int range_cyclic:1;
47         unsigned int for_background:1;
48         unsigned int sb_pinned:1;
49 };
50
51 /*
52  * Work items for the bdi_writeback threads
53  */
54 struct bdi_work {
55         struct list_head list;          /* pending work list */
56         struct rcu_head rcu_head;       /* for RCU free/clear of work */
57
58         unsigned long seen;             /* threads that have seen this work */
59         atomic_t pending;               /* number of threads still to do work */
60
61         struct wb_writeback_args args;  /* writeback arguments */
62
63         unsigned long state;            /* flag bits, see WS_* */
64 };
65
66 enum {
67         WS_USED_B = 0,
68         WS_ONSTACK_B,
69 };
70
71 #define WS_USED (1 << WS_USED_B)
72 #define WS_ONSTACK (1 << WS_ONSTACK_B)
73
74 static inline bool bdi_work_on_stack(struct bdi_work *work)
75 {
76         return test_bit(WS_ONSTACK_B, &work->state);
77 }
78
79 static inline void bdi_work_init(struct bdi_work *work,
80                                  struct wb_writeback_args *args)
81 {
82         INIT_RCU_HEAD(&work->rcu_head);
83         work->args = *args;
84         work->state = WS_USED;
85 }
86
87 /**
88  * writeback_in_progress - determine whether there is writeback in progress
89  * @bdi: the device's backing_dev_info structure.
90  *
91  * Determine whether there is writeback waiting to be handled against a
92  * backing device.
93  */
94 int writeback_in_progress(struct backing_dev_info *bdi)
95 {
96         return !list_empty(&bdi->work_list);
97 }
98
99 static void bdi_work_clear(struct bdi_work *work)
100 {
101         clear_bit(WS_USED_B, &work->state);
102         smp_mb__after_clear_bit();
103         /*
104          * work can have disappeared at this point. bit waitq functions
105          * should be able to tolerate this, provided bdi_sched_wait does
106          * not dereference it's pointer argument.
107         */
108         wake_up_bit(&work->state, WS_USED_B);
109 }
110
111 static void bdi_work_free(struct rcu_head *head)
112 {
113         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
114
115         if (!bdi_work_on_stack(work))
116                 kfree(work);
117         else
118                 bdi_work_clear(work);
119 }
120
121 static void wb_work_complete(struct bdi_work *work)
122 {
123         const enum writeback_sync_modes sync_mode = work->args.sync_mode;
124         int onstack = bdi_work_on_stack(work);
125
126         /*
127          * For allocated work, we can clear the done/seen bit right here.
128          * For on-stack work, we need to postpone both the clear and free
129          * to after the RCU grace period, since the stack could be invalidated
130          * as soon as bdi_work_clear() has done the wakeup.
131          */
132         if (!onstack)
133                 bdi_work_clear(work);
134         if (sync_mode == WB_SYNC_NONE || onstack)
135                 call_rcu(&work->rcu_head, bdi_work_free);
136 }
137
138 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
139 {
140         /*
141          * The caller has retrieved the work arguments from this work,
142          * drop our reference. If this is the last ref, delete and free it
143          */
144         if (atomic_dec_and_test(&work->pending)) {
145                 struct backing_dev_info *bdi = wb->bdi;
146
147                 spin_lock(&bdi->wb_lock);
148                 list_del_rcu(&work->list);
149                 spin_unlock(&bdi->wb_lock);
150
151                 wb_work_complete(work);
152         }
153 }
154
155 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
156 {
157         work->seen = bdi->wb_mask;
158         BUG_ON(!work->seen);
159         atomic_set(&work->pending, bdi->wb_cnt);
160         BUG_ON(!bdi->wb_cnt);
161
162         /*
163          * list_add_tail_rcu() contains the necessary barriers to
164          * make sure the above stores are seen before the item is
165          * noticed on the list
166          */
167         spin_lock(&bdi->wb_lock);
168         list_add_tail_rcu(&work->list, &bdi->work_list);
169         spin_unlock(&bdi->wb_lock);
170
171         /*
172          * If the default thread isn't there, make sure we add it. When
173          * it gets created and wakes up, we'll run this work.
174          */
175         if (unlikely(list_empty_careful(&bdi->wb_list)))
176                 wake_up_process(default_backing_dev_info.wb.task);
177         else {
178                 struct bdi_writeback *wb = &bdi->wb;
179
180                 if (wb->task)
181                         wake_up_process(wb->task);
182         }
183 }
184
185 /*
186  * Used for on-stack allocated work items. The caller needs to wait until
187  * the wb threads have acked the work before it's safe to continue.
188  */
189 static void bdi_wait_on_work_clear(struct bdi_work *work)
190 {
191         wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
192                     TASK_UNINTERRUPTIBLE);
193 }
194
195 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
196                                  struct wb_writeback_args *args)
197 {
198         struct bdi_work *work;
199
200         /*
201          * This is WB_SYNC_NONE writeback, so if allocation fails just
202          * wakeup the thread for old dirty data writeback
203          */
204         work = kmalloc(sizeof(*work), GFP_ATOMIC);
205         if (work) {
206                 bdi_work_init(work, args);
207                 bdi_queue_work(bdi, work);
208         } else {
209                 struct bdi_writeback *wb = &bdi->wb;
210
211                 if (wb->task)
212                         wake_up_process(wb->task);
213         }
214 }
215
216 /**
217  * bdi_sync_writeback - start and wait for writeback
218  * @bdi: the backing device to write from
219  * @sb: write inodes from this super_block
220  *
221  * Description:
222  *   This does WB_SYNC_ALL data integrity writeback and waits for the
223  *   IO to complete. Callers must hold the sb s_umount semaphore for
224  *   reading, to avoid having the super disappear before we are done.
225  */
226 static void bdi_sync_writeback(struct backing_dev_info *bdi,
227                                struct super_block *sb)
228 {
229         struct wb_writeback_args args = {
230                 .sb             = sb,
231                 .sync_mode      = WB_SYNC_ALL,
232                 .nr_pages       = LONG_MAX,
233                 .range_cyclic   = 0,
234                 /*
235                  * Setting sb_pinned is not necessary for WB_SYNC_ALL, but
236                  * lets make it explicitly clear.
237                  */
238                 .sb_pinned      = 1,
239         };
240         struct bdi_work work;
241
242         bdi_work_init(&work, &args);
243         work.state |= WS_ONSTACK;
244
245         bdi_queue_work(bdi, &work);
246         bdi_wait_on_work_clear(&work);
247 }
248
249 /**
250  * bdi_start_writeback - start writeback
251  * @bdi: the backing device to write from
252  * @sb: write inodes from this super_block
253  * @nr_pages: the number of pages to write
254  * @sb_locked: caller already holds sb umount sem.
255  *
256  * Description:
257  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
258  *   started when this function returns, we make no guarentees on
259  *   completion. Caller specifies whether sb umount sem is held already or not.
260  *
261  */
262 void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
263                          long nr_pages, int sb_locked)
264 {
265         struct wb_writeback_args args = {
266                 .sb             = sb,
267                 .sync_mode      = WB_SYNC_NONE,
268                 .nr_pages       = nr_pages,
269                 .range_cyclic   = 1,
270                 .sb_pinned      = sb_locked,
271         };
272
273         /*
274          * We treat @nr_pages=0 as the special case to do background writeback,
275          * ie. to sync pages until the background dirty threshold is reached.
276          */
277         if (!nr_pages) {
278                 args.nr_pages = LONG_MAX;
279                 args.for_background = 1;
280         }
281
282         bdi_alloc_queue_work(bdi, &args);
283 }
284
285 /*
286  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
287  * furthest end of its superblock's dirty-inode list.
288  *
289  * Before stamping the inode's ->dirtied_when, we check to see whether it is
290  * already the most-recently-dirtied inode on the b_dirty list.  If that is
291  * the case then the inode must have been redirtied while it was being written
292  * out and we don't reset its dirtied_when.
293  */
294 static void redirty_tail(struct inode *inode)
295 {
296         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
297
298         if (!list_empty(&wb->b_dirty)) {
299                 struct inode *tail;
300
301                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
302                 if (time_before(inode->dirtied_when, tail->dirtied_when))
303                         inode->dirtied_when = jiffies;
304         }
305         list_move(&inode->i_list, &wb->b_dirty);
306 }
307
308 /*
309  * requeue inode for re-scanning after bdi->b_io list is exhausted.
310  */
311 static void requeue_io(struct inode *inode)
312 {
313         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
314
315         list_move(&inode->i_list, &wb->b_more_io);
316 }
317
318 static void inode_sync_complete(struct inode *inode)
319 {
320         /*
321          * Prevent speculative execution through spin_unlock(&inode_lock);
322          */
323         smp_mb();
324         wake_up_bit(&inode->i_state, __I_SYNC);
325 }
326
327 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
328 {
329         bool ret = time_after(inode->dirtied_when, t);
330 #ifndef CONFIG_64BIT
331         /*
332          * For inodes being constantly redirtied, dirtied_when can get stuck.
333          * It _appears_ to be in the future, but is actually in distant past.
334          * This test is necessary to prevent such wrapped-around relative times
335          * from permanently stopping the whole bdi writeback.
336          */
337         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
338 #endif
339         return ret;
340 }
341
342 /*
343  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
344  */
345 static void move_expired_inodes(struct list_head *delaying_queue,
346                                struct list_head *dispatch_queue,
347                                 unsigned long *older_than_this)
348 {
349         LIST_HEAD(tmp);
350         struct list_head *pos, *node;
351         struct super_block *sb = NULL;
352         struct inode *inode;
353         int do_sb_sort = 0;
354
355         while (!list_empty(delaying_queue)) {
356                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
357                 if (older_than_this &&
358                     inode_dirtied_after(inode, *older_than_this))
359                         break;
360                 if (sb && sb != inode->i_sb)
361                         do_sb_sort = 1;
362                 sb = inode->i_sb;
363                 list_move(&inode->i_list, &tmp);
364         }
365
366         /* just one sb in list, splice to dispatch_queue and we're done */
367         if (!do_sb_sort) {
368                 list_splice(&tmp, dispatch_queue);
369                 return;
370         }
371
372         /* Move inodes from one superblock together */
373         while (!list_empty(&tmp)) {
374                 inode = list_entry(tmp.prev, struct inode, i_list);
375                 sb = inode->i_sb;
376                 list_for_each_prev_safe(pos, node, &tmp) {
377                         inode = list_entry(pos, struct inode, i_list);
378                         if (inode->i_sb == sb)
379                                 list_move(&inode->i_list, dispatch_queue);
380                 }
381         }
382 }
383
384 /*
385  * Queue all expired dirty inodes for io, eldest first.
386  */
387 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
388 {
389         list_splice_init(&wb->b_more_io, wb->b_io.prev);
390         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
391 }
392
393 static int write_inode(struct inode *inode, struct writeback_control *wbc)
394 {
395         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
396                 return inode->i_sb->s_op->write_inode(inode, wbc);
397         return 0;
398 }
399
400 /*
401  * Wait for writeback on an inode to complete.
402  */
403 static void inode_wait_for_writeback(struct inode *inode)
404 {
405         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
406         wait_queue_head_t *wqh;
407
408         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
409         do {
410                 spin_unlock(&inode_lock);
411                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
412                 spin_lock(&inode_lock);
413         } while (inode->i_state & I_SYNC);
414 }
415
416 /*
417  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
418  * caller has ref on the inode (either via __iget or via syscall against an fd)
419  * or the inode has I_WILL_FREE set (via generic_forget_inode)
420  *
421  * If `wait' is set, wait on the writeout.
422  *
423  * The whole writeout design is quite complex and fragile.  We want to avoid
424  * starvation of particular inodes when others are being redirtied, prevent
425  * livelocks, etc.
426  *
427  * Called under inode_lock.
428  */
429 static int
430 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
431 {
432         struct address_space *mapping = inode->i_mapping;
433         unsigned dirty;
434         int ret;
435
436         if (!atomic_read(&inode->i_count))
437                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
438         else
439                 WARN_ON(inode->i_state & I_WILL_FREE);
440
441         if (inode->i_state & I_SYNC) {
442                 /*
443                  * If this inode is locked for writeback and we are not doing
444                  * writeback-for-data-integrity, move it to b_more_io so that
445                  * writeback can proceed with the other inodes on s_io.
446                  *
447                  * We'll have another go at writing back this inode when we
448                  * completed a full scan of b_io.
449                  */
450                 if (wbc->sync_mode != WB_SYNC_ALL) {
451                         requeue_io(inode);
452                         return 0;
453                 }
454
455                 /*
456                  * It's a data-integrity sync.  We must wait.
457                  */
458                 inode_wait_for_writeback(inode);
459         }
460
461         BUG_ON(inode->i_state & I_SYNC);
462
463         /* Set I_SYNC, reset I_DIRTY_PAGES */
464         inode->i_state |= I_SYNC;
465         inode->i_state &= ~I_DIRTY_PAGES;
466         spin_unlock(&inode_lock);
467
468         ret = do_writepages(mapping, wbc);
469
470         /*
471          * Make sure to wait on the data before writing out the metadata.
472          * This is important for filesystems that modify metadata on data
473          * I/O completion.
474          */
475         if (wbc->sync_mode == WB_SYNC_ALL) {
476                 int err = filemap_fdatawait(mapping);
477                 if (ret == 0)
478                         ret = err;
479         }
480
481         /*
482          * Some filesystems may redirty the inode during the writeback
483          * due to delalloc, clear dirty metadata flags right before
484          * write_inode()
485          */
486         spin_lock(&inode_lock);
487         dirty = inode->i_state & I_DIRTY;
488         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
489         spin_unlock(&inode_lock);
490         /* Don't write the inode if only I_DIRTY_PAGES was set */
491         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
492                 int err = write_inode(inode, wbc);
493                 if (ret == 0)
494                         ret = err;
495         }
496
497         spin_lock(&inode_lock);
498         inode->i_state &= ~I_SYNC;
499         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
500                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
501                         /*
502                          * More pages get dirtied by a fast dirtier.
503                          */
504                         goto select_queue;
505                 } else if (inode->i_state & I_DIRTY) {
506                         /*
507                          * At least XFS will redirty the inode during the
508                          * writeback (delalloc) and on io completion (isize).
509                          */
510                         redirty_tail(inode);
511                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
512                         /*
513                          * We didn't write back all the pages.  nfs_writepages()
514                          * sometimes bales out without doing anything. Redirty
515                          * the inode; Move it from b_io onto b_more_io/b_dirty.
516                          */
517                         /*
518                          * akpm: if the caller was the kupdate function we put
519                          * this inode at the head of b_dirty so it gets first
520                          * consideration.  Otherwise, move it to the tail, for
521                          * the reasons described there.  I'm not really sure
522                          * how much sense this makes.  Presumably I had a good
523                          * reasons for doing it this way, and I'd rather not
524                          * muck with it at present.
525                          */
526                         if (wbc->for_kupdate) {
527                                 /*
528                                  * For the kupdate function we move the inode
529                                  * to b_more_io so it will get more writeout as
530                                  * soon as the queue becomes uncongested.
531                                  */
532                                 inode->i_state |= I_DIRTY_PAGES;
533 select_queue:
534                                 if (wbc->nr_to_write <= 0) {
535                                         /*
536                                          * slice used up: queue for next turn
537                                          */
538                                         requeue_io(inode);
539                                 } else {
540                                         /*
541                                          * somehow blocked: retry later
542                                          */
543                                         redirty_tail(inode);
544                                 }
545                         } else {
546                                 /*
547                                  * Otherwise fully redirty the inode so that
548                                  * other inodes on this superblock will get some
549                                  * writeout.  Otherwise heavy writing to one
550                                  * file would indefinitely suspend writeout of
551                                  * all the other files.
552                                  */
553                                 inode->i_state |= I_DIRTY_PAGES;
554                                 redirty_tail(inode);
555                         }
556                 } else if (atomic_read(&inode->i_count)) {
557                         /*
558                          * The inode is clean, inuse
559                          */
560                         list_move(&inode->i_list, &inode_in_use);
561                 } else {
562                         /*
563                          * The inode is clean, unused
564                          */
565                         list_move(&inode->i_list, &inode_unused);
566                 }
567         }
568         inode_sync_complete(inode);
569         return ret;
570 }
571
572 static void unpin_sb_for_writeback(struct super_block *sb)
573 {
574         up_read(&sb->s_umount);
575         put_super(sb);
576 }
577
578 enum sb_pin_state {
579         SB_PINNED,
580         SB_NOT_PINNED,
581         SB_PIN_FAILED
582 };
583
584 /*
585  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
586  * before calling writeback. So make sure that we do pin it, so it doesn't
587  * go away while we are writing inodes from it.
588  */
589 static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
590                                               struct super_block *sb)
591 {
592         /*
593          * Caller must already hold the ref for this
594          */
595         if (wbc->sync_mode == WB_SYNC_ALL || wbc->sb_pinned) {
596                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
597                 return SB_NOT_PINNED;
598         }
599         spin_lock(&sb_lock);
600         sb->s_count++;
601         if (down_read_trylock(&sb->s_umount)) {
602                 if (sb->s_root) {
603                         spin_unlock(&sb_lock);
604                         return SB_PINNED;
605                 }
606                 /*
607                  * umounted, drop rwsem again and fall through to failure
608                  */
609                 up_read(&sb->s_umount);
610         }
611         sb->s_count--;
612         spin_unlock(&sb_lock);
613         return SB_PIN_FAILED;
614 }
615
616 /*
617  * Write a portion of b_io inodes which belong to @sb.
618  * If @wbc->sb != NULL, then find and write all such
619  * inodes. Otherwise write only ones which go sequentially
620  * in reverse order.
621  * Return 1, if the caller writeback routine should be
622  * interrupted. Otherwise return 0.
623  */
624 static int writeback_sb_inodes(struct super_block *sb,
625                                struct bdi_writeback *wb,
626                                struct writeback_control *wbc)
627 {
628         while (!list_empty(&wb->b_io)) {
629                 long pages_skipped;
630                 struct inode *inode = list_entry(wb->b_io.prev,
631                                                  struct inode, i_list);
632                 if (wbc->sb && sb != inode->i_sb) {
633                         /* super block given and doesn't
634                            match, skip this inode */
635                         redirty_tail(inode);
636                         continue;
637                 }
638                 if (sb != inode->i_sb)
639                         /* finish with this superblock */
640                         return 0;
641                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
642                         requeue_io(inode);
643                         continue;
644                 }
645                 /*
646                  * Was this inode dirtied after sync_sb_inodes was called?
647                  * This keeps sync from extra jobs and livelock.
648                  */
649                 if (inode_dirtied_after(inode, wbc->wb_start))
650                         return 1;
651
652                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
653                 __iget(inode);
654                 pages_skipped = wbc->pages_skipped;
655                 writeback_single_inode(inode, wbc);
656                 if (wbc->pages_skipped != pages_skipped) {
657                         /*
658                          * writeback is not making progress due to locked
659                          * buffers.  Skip this inode for now.
660                          */
661                         redirty_tail(inode);
662                 }
663                 spin_unlock(&inode_lock);
664                 iput(inode);
665                 cond_resched();
666                 spin_lock(&inode_lock);
667                 if (wbc->nr_to_write <= 0) {
668                         wbc->more_io = 1;
669                         return 1;
670                 }
671                 if (!list_empty(&wb->b_more_io))
672                         wbc->more_io = 1;
673         }
674         /* b_io is empty */
675         return 1;
676 }
677
678 static void writeback_inodes_wb(struct bdi_writeback *wb,
679                                 struct writeback_control *wbc)
680 {
681         int ret = 0;
682
683         wbc->wb_start = jiffies; /* livelock avoidance */
684         spin_lock(&inode_lock);
685         if (!wbc->for_kupdate || list_empty(&wb->b_io))
686                 queue_io(wb, wbc->older_than_this);
687
688         while (!list_empty(&wb->b_io)) {
689                 struct inode *inode = list_entry(wb->b_io.prev,
690                                                  struct inode, i_list);
691                 struct super_block *sb = inode->i_sb;
692                 enum sb_pin_state state;
693
694                 if (wbc->sb && sb != wbc->sb) {
695                         /* super block given and doesn't
696                            match, skip this inode */
697                         redirty_tail(inode);
698                         continue;
699                 }
700                 state = pin_sb_for_writeback(wbc, sb);
701
702                 if (state == SB_PIN_FAILED) {
703                         requeue_io(inode);
704                         continue;
705                 }
706                 ret = writeback_sb_inodes(sb, wb, wbc);
707
708                 if (state == SB_PINNED)
709                         unpin_sb_for_writeback(sb);
710                 if (ret)
711                         break;
712         }
713         spin_unlock(&inode_lock);
714         /* Leave any unwritten inodes on b_io */
715 }
716
717 void writeback_inodes_wbc(struct writeback_control *wbc)
718 {
719         struct backing_dev_info *bdi = wbc->bdi;
720
721         writeback_inodes_wb(&bdi->wb, wbc);
722 }
723
724 /*
725  * The maximum number of pages to writeout in a single bdi flush/kupdate
726  * operation.  We do this so we don't hold I_SYNC against an inode for
727  * enormous amounts of time, which would block a userspace task which has
728  * been forced to throttle against that inode.  Also, the code reevaluates
729  * the dirty each time it has written this many pages.
730  */
731 #define MAX_WRITEBACK_PAGES     1024
732
733 static inline bool over_bground_thresh(void)
734 {
735         unsigned long background_thresh, dirty_thresh;
736
737         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
738
739         return (global_page_state(NR_FILE_DIRTY) +
740                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
741 }
742
743 /*
744  * Explicit flushing or periodic writeback of "old" data.
745  *
746  * Define "old": the first time one of an inode's pages is dirtied, we mark the
747  * dirtying-time in the inode's address_space.  So this periodic writeback code
748  * just walks the superblock inode list, writing back any inodes which are
749  * older than a specific point in time.
750  *
751  * Try to run once per dirty_writeback_interval.  But if a writeback event
752  * takes longer than a dirty_writeback_interval interval, then leave a
753  * one-second gap.
754  *
755  * older_than_this takes precedence over nr_to_write.  So we'll only write back
756  * all dirty pages if they are all attached to "old" mappings.
757  */
758 static long wb_writeback(struct bdi_writeback *wb,
759                          struct wb_writeback_args *args)
760 {
761         struct writeback_control wbc = {
762                 .bdi                    = wb->bdi,
763                 .sb                     = args->sb,
764                 .sync_mode              = args->sync_mode,
765                 .older_than_this        = NULL,
766                 .for_kupdate            = args->for_kupdate,
767                 .for_background         = args->for_background,
768                 .range_cyclic           = args->range_cyclic,
769                 .sb_pinned              = args->sb_pinned,
770         };
771         unsigned long oldest_jif;
772         long wrote = 0;
773         struct inode *inode;
774
775         if (wbc.for_kupdate) {
776                 wbc.older_than_this = &oldest_jif;
777                 oldest_jif = jiffies -
778                                 msecs_to_jiffies(dirty_expire_interval * 10);
779         }
780         if (!wbc.range_cyclic) {
781                 wbc.range_start = 0;
782                 wbc.range_end = LLONG_MAX;
783         }
784
785         for (;;) {
786                 /*
787                  * Stop writeback when nr_pages has been consumed
788                  */
789                 if (args->nr_pages <= 0)
790                         break;
791
792                 /*
793                  * For background writeout, stop when we are below the
794                  * background dirty threshold
795                  */
796                 if (args->for_background && !over_bground_thresh())
797                         break;
798
799                 wbc.more_io = 0;
800                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
801                 wbc.pages_skipped = 0;
802                 writeback_inodes_wb(wb, &wbc);
803                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
804                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
805
806                 /*
807                  * If we consumed everything, see if we have more
808                  */
809                 if (wbc.nr_to_write <= 0)
810                         continue;
811                 /*
812                  * Didn't write everything and we don't have more IO, bail
813                  */
814                 if (!wbc.more_io)
815                         break;
816                 /*
817                  * Did we write something? Try for more
818                  */
819                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
820                         continue;
821                 /*
822                  * Nothing written. Wait for some inode to
823                  * become available for writeback. Otherwise
824                  * we'll just busyloop.
825                  */
826                 spin_lock(&inode_lock);
827                 if (!list_empty(&wb->b_more_io))  {
828                         inode = list_entry(wb->b_more_io.prev,
829                                                 struct inode, i_list);
830                         inode_wait_for_writeback(inode);
831                 }
832                 spin_unlock(&inode_lock);
833         }
834
835         return wrote;
836 }
837
838 /*
839  * Return the next bdi_work struct that hasn't been processed by this
840  * wb thread yet. ->seen is initially set for each thread that exists
841  * for this device, when a thread first notices a piece of work it
842  * clears its bit. Depending on writeback type, the thread will notify
843  * completion on either receiving the work (WB_SYNC_NONE) or after
844  * it is done (WB_SYNC_ALL).
845  */
846 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
847                                            struct bdi_writeback *wb)
848 {
849         struct bdi_work *work, *ret = NULL;
850
851         rcu_read_lock();
852
853         list_for_each_entry_rcu(work, &bdi->work_list, list) {
854                 if (!test_bit(wb->nr, &work->seen))
855                         continue;
856                 clear_bit(wb->nr, &work->seen);
857
858                 ret = work;
859                 break;
860         }
861
862         rcu_read_unlock();
863         return ret;
864 }
865
866 static long wb_check_old_data_flush(struct bdi_writeback *wb)
867 {
868         unsigned long expired;
869         long nr_pages;
870
871         /*
872          * When set to zero, disable periodic writeback
873          */
874         if (!dirty_writeback_interval)
875                 return 0;
876
877         expired = wb->last_old_flush +
878                         msecs_to_jiffies(dirty_writeback_interval * 10);
879         if (time_before(jiffies, expired))
880                 return 0;
881
882         wb->last_old_flush = jiffies;
883         nr_pages = global_page_state(NR_FILE_DIRTY) +
884                         global_page_state(NR_UNSTABLE_NFS) +
885                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
886
887         if (nr_pages) {
888                 struct wb_writeback_args args = {
889                         .nr_pages       = nr_pages,
890                         .sync_mode      = WB_SYNC_NONE,
891                         .for_kupdate    = 1,
892                         .range_cyclic   = 1,
893                 };
894
895                 return wb_writeback(wb, &args);
896         }
897
898         return 0;
899 }
900
901 /*
902  * Retrieve work items and do the writeback they describe
903  */
904 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
905 {
906         struct backing_dev_info *bdi = wb->bdi;
907         struct bdi_work *work;
908         long wrote = 0;
909
910         while ((work = get_next_work_item(bdi, wb)) != NULL) {
911                 struct wb_writeback_args args = work->args;
912
913                 /*
914                  * Override sync mode, in case we must wait for completion
915                  */
916                 if (force_wait)
917                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
918
919                 /*
920                  * If this isn't a data integrity operation, just notify
921                  * that we have seen this work and we are now starting it.
922                  */
923                 if (args.sync_mode == WB_SYNC_NONE)
924                         wb_clear_pending(wb, work);
925
926                 wrote += wb_writeback(wb, &args);
927
928                 /*
929                  * This is a data integrity writeback, so only do the
930                  * notification when we have completed the work.
931                  */
932                 if (args.sync_mode == WB_SYNC_ALL)
933                         wb_clear_pending(wb, work);
934         }
935
936         /*
937          * Check for periodic writeback, kupdated() style
938          */
939         wrote += wb_check_old_data_flush(wb);
940
941         return wrote;
942 }
943
944 /*
945  * Handle writeback of dirty data for the device backed by this bdi. Also
946  * wakes up periodically and does kupdated style flushing.
947  */
948 int bdi_writeback_task(struct bdi_writeback *wb)
949 {
950         unsigned long last_active = jiffies;
951         unsigned long wait_jiffies = -1UL;
952         long pages_written;
953
954         while (!kthread_should_stop()) {
955                 pages_written = wb_do_writeback(wb, 0);
956
957                 if (pages_written)
958                         last_active = jiffies;
959                 else if (wait_jiffies != -1UL) {
960                         unsigned long max_idle;
961
962                         /*
963                          * Longest period of inactivity that we tolerate. If we
964                          * see dirty data again later, the task will get
965                          * recreated automatically.
966                          */
967                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
968                         if (time_after(jiffies, max_idle + last_active))
969                                 break;
970                 }
971
972                 if (dirty_writeback_interval) {
973                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
974                         schedule_timeout_interruptible(wait_jiffies);
975                 } else {
976                         set_current_state(TASK_INTERRUPTIBLE);
977                         if (list_empty_careful(&wb->bdi->work_list) &&
978                             !kthread_should_stop())
979                                 schedule();
980                         __set_current_state(TASK_RUNNING);
981                 }
982
983                 try_to_freeze();
984         }
985
986         return 0;
987 }
988
989 /*
990  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
991  * writeback, for integrity writeback see bdi_sync_writeback().
992  */
993 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
994 {
995         struct wb_writeback_args args = {
996                 .sb             = sb,
997                 .nr_pages       = nr_pages,
998                 .sync_mode      = WB_SYNC_NONE,
999         };
1000         struct backing_dev_info *bdi;
1001
1002         rcu_read_lock();
1003
1004         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1005                 if (!bdi_has_dirty_io(bdi))
1006                         continue;
1007
1008                 bdi_alloc_queue_work(bdi, &args);
1009         }
1010
1011         rcu_read_unlock();
1012 }
1013
1014 /*
1015  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1016  * the whole world.
1017  */
1018 void wakeup_flusher_threads(long nr_pages)
1019 {
1020         if (nr_pages == 0)
1021                 nr_pages = global_page_state(NR_FILE_DIRTY) +
1022                                 global_page_state(NR_UNSTABLE_NFS);
1023         bdi_writeback_all(NULL, nr_pages);
1024 }
1025
1026 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1027 {
1028         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1029                 struct dentry *dentry;
1030                 const char *name = "?";
1031
1032                 dentry = d_find_alias(inode);
1033                 if (dentry) {
1034                         spin_lock(&dentry->d_lock);
1035                         name = (const char *) dentry->d_name.name;
1036                 }
1037                 printk(KERN_DEBUG
1038                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1039                        current->comm, task_pid_nr(current), inode->i_ino,
1040                        name, inode->i_sb->s_id);
1041                 if (dentry) {
1042                         spin_unlock(&dentry->d_lock);
1043                         dput(dentry);
1044                 }
1045         }
1046 }
1047
1048 /**
1049  *      __mark_inode_dirty -    internal function
1050  *      @inode: inode to mark
1051  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1052  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1053  *      mark_inode_dirty_sync.
1054  *
1055  * Put the inode on the super block's dirty list.
1056  *
1057  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1058  * dirty list only if it is hashed or if it refers to a blockdev.
1059  * If it was not hashed, it will never be added to the dirty list
1060  * even if it is later hashed, as it will have been marked dirty already.
1061  *
1062  * In short, make sure you hash any inodes _before_ you start marking
1063  * them dirty.
1064  *
1065  * This function *must* be atomic for the I_DIRTY_PAGES case -
1066  * set_page_dirty() is called under spinlock in several places.
1067  *
1068  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1069  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1070  * the kernel-internal blockdev inode represents the dirtying time of the
1071  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1072  * page->mapping->host, so the page-dirtying time is recorded in the internal
1073  * blockdev inode.
1074  */
1075 void __mark_inode_dirty(struct inode *inode, int flags)
1076 {
1077         struct super_block *sb = inode->i_sb;
1078
1079         /*
1080          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1081          * dirty the inode itself
1082          */
1083         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1084                 if (sb->s_op->dirty_inode)
1085                         sb->s_op->dirty_inode(inode);
1086         }
1087
1088         /*
1089          * make sure that changes are seen by all cpus before we test i_state
1090          * -- mikulas
1091          */
1092         smp_mb();
1093
1094         /* avoid the locking if we can */
1095         if ((inode->i_state & flags) == flags)
1096                 return;
1097
1098         if (unlikely(block_dump))
1099                 block_dump___mark_inode_dirty(inode);
1100
1101         spin_lock(&inode_lock);
1102         if ((inode->i_state & flags) != flags) {
1103                 const int was_dirty = inode->i_state & I_DIRTY;
1104
1105                 inode->i_state |= flags;
1106
1107                 /*
1108                  * If the inode is being synced, just update its dirty state.
1109                  * The unlocker will place the inode on the appropriate
1110                  * superblock list, based upon its state.
1111                  */
1112                 if (inode->i_state & I_SYNC)
1113                         goto out;
1114
1115                 /*
1116                  * Only add valid (hashed) inodes to the superblock's
1117                  * dirty list.  Add blockdev inodes as well.
1118                  */
1119                 if (!S_ISBLK(inode->i_mode)) {
1120                         if (hlist_unhashed(&inode->i_hash))
1121                                 goto out;
1122                 }
1123                 if (inode->i_state & (I_FREEING|I_CLEAR))
1124                         goto out;
1125
1126                 /*
1127                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1128                  * reposition it (that would break b_dirty time-ordering).
1129                  */
1130                 if (!was_dirty) {
1131                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1132                         struct backing_dev_info *bdi = wb->bdi;
1133
1134                         if (bdi_cap_writeback_dirty(bdi) &&
1135                             !test_bit(BDI_registered, &bdi->state)) {
1136                                 WARN_ON(1);
1137                                 printk(KERN_ERR "bdi-%s not registered\n",
1138                                                                 bdi->name);
1139                         }
1140
1141                         inode->dirtied_when = jiffies;
1142                         list_move(&inode->i_list, &wb->b_dirty);
1143                 }
1144         }
1145 out:
1146         spin_unlock(&inode_lock);
1147 }
1148 EXPORT_SYMBOL(__mark_inode_dirty);
1149
1150 /*
1151  * Write out a superblock's list of dirty inodes.  A wait will be performed
1152  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1153  *
1154  * If older_than_this is non-NULL, then only write out inodes which
1155  * had their first dirtying at a time earlier than *older_than_this.
1156  *
1157  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1158  * This function assumes that the blockdev superblock's inodes are backed by
1159  * a variety of queues, so all inodes are searched.  For other superblocks,
1160  * assume that all inodes are backed by the same queue.
1161  *
1162  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1163  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1164  * on the writer throttling path, and we get decent balancing between many
1165  * throttled threads: we don't want them all piling up on inode_sync_wait.
1166  */
1167 static void wait_sb_inodes(struct super_block *sb)
1168 {
1169         struct inode *inode, *old_inode = NULL;
1170
1171         /*
1172          * We need to be protected against the filesystem going from
1173          * r/o to r/w or vice versa.
1174          */
1175         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1176
1177         spin_lock(&inode_lock);
1178
1179         /*
1180          * Data integrity sync. Must wait for all pages under writeback,
1181          * because there may have been pages dirtied before our sync
1182          * call, but which had writeout started before we write it out.
1183          * In which case, the inode may not be on the dirty list, but
1184          * we still have to wait for that writeout.
1185          */
1186         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1187                 struct address_space *mapping;
1188
1189                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1190                         continue;
1191                 mapping = inode->i_mapping;
1192                 if (mapping->nrpages == 0)
1193                         continue;
1194                 __iget(inode);
1195                 spin_unlock(&inode_lock);
1196                 /*
1197                  * We hold a reference to 'inode' so it couldn't have
1198                  * been removed from s_inodes list while we dropped the
1199                  * inode_lock.  We cannot iput the inode now as we can
1200                  * be holding the last reference and we cannot iput it
1201                  * under inode_lock. So we keep the reference and iput
1202                  * it later.
1203                  */
1204                 iput(old_inode);
1205                 old_inode = inode;
1206
1207                 filemap_fdatawait(mapping);
1208
1209                 cond_resched();
1210
1211                 spin_lock(&inode_lock);
1212         }
1213         spin_unlock(&inode_lock);
1214         iput(old_inode);
1215 }
1216
1217 static void __writeback_inodes_sb(struct super_block *sb, int sb_locked)
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, sb_locked);
1227 }
1228
1229 /**
1230  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1231  * @sb: the superblock
1232  *
1233  * Start writeback on some inodes on this super_block. No guarantees are made
1234  * on how many (if any) will be written, and this function does not wait
1235  * for IO completion of submitted IO. The number of pages submitted is
1236  * returned.
1237  */
1238 void writeback_inodes_sb(struct super_block *sb)
1239 {
1240         __writeback_inodes_sb(sb, 0);
1241 }
1242 EXPORT_SYMBOL(writeback_inodes_sb);
1243
1244 /**
1245  * writeback_inodes_sb_locked   - writeback dirty inodes from given super_block
1246  * @sb: the superblock
1247  *
1248  * Like writeback_inodes_sb(), except the caller already holds the
1249  * sb umount sem.
1250  */
1251 void writeback_inodes_sb_locked(struct super_block *sb)
1252 {
1253         __writeback_inodes_sb(sb, 1);
1254 }
1255
1256 /**
1257  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1258  * @sb: the superblock
1259  *
1260  * Invoke writeback_inodes_sb if no writeback is currently underway.
1261  * Returns 1 if writeback was started, 0 if not.
1262  */
1263 int writeback_inodes_sb_if_idle(struct super_block *sb)
1264 {
1265         if (!writeback_in_progress(sb->s_bdi)) {
1266                 writeback_inodes_sb(sb);
1267                 return 1;
1268         } else
1269                 return 0;
1270 }
1271 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1272
1273 /**
1274  * sync_inodes_sb       -       sync sb inode pages
1275  * @sb: the superblock
1276  *
1277  * This function writes and waits on any dirty inode belonging to this
1278  * super_block. The number of pages synced is returned.
1279  */
1280 void sync_inodes_sb(struct super_block *sb)
1281 {
1282         bdi_sync_writeback(sb->s_bdi, sb);
1283         wait_sb_inodes(sb);
1284 }
1285 EXPORT_SYMBOL(sync_inodes_sb);
1286
1287 /**
1288  * write_inode_now      -       write an inode to disk
1289  * @inode: inode to write to disk
1290  * @sync: whether the write should be synchronous or not
1291  *
1292  * This function commits an inode to disk immediately if it is dirty. This is
1293  * primarily needed by knfsd.
1294  *
1295  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1296  */
1297 int write_inode_now(struct inode *inode, int sync)
1298 {
1299         int ret;
1300         struct writeback_control wbc = {
1301                 .nr_to_write = LONG_MAX,
1302                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1303                 .range_start = 0,
1304                 .range_end = LLONG_MAX,
1305         };
1306
1307         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1308                 wbc.nr_to_write = 0;
1309
1310         might_sleep();
1311         spin_lock(&inode_lock);
1312         ret = writeback_single_inode(inode, &wbc);
1313         spin_unlock(&inode_lock);
1314         if (sync)
1315                 inode_sync_wait(inode);
1316         return ret;
1317 }
1318 EXPORT_SYMBOL(write_inode_now);
1319
1320 /**
1321  * sync_inode - write an inode and its pages to disk.
1322  * @inode: the inode to sync
1323  * @wbc: controls the writeback mode
1324  *
1325  * sync_inode() will write an inode and its pages to disk.  It will also
1326  * correctly update the inode on its superblock's dirty inode lists and will
1327  * update inode->i_state.
1328  *
1329  * The caller must have a ref on the inode.
1330  */
1331 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1332 {
1333         int ret;
1334
1335         spin_lock(&inode_lock);
1336         ret = writeback_single_inode(inode, wbc);
1337         spin_unlock(&inode_lock);
1338         return ret;
1339 }
1340 EXPORT_SYMBOL(sync_inode);