476be9b10881ca947cfcd7e79d55c84563dba353
[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/sched.h>
20 #include <linux/fs.h>
21 #include <linux/mm.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/backing-dev.h>
27 #include <linux/buffer_head.h>
28 #include "internal.h"
29
30 #define inode_to_bdi(inode)     ((inode)->i_mapping->backing_dev_info)
31
32 /*
33  * We don't actually have pdflush, but this one is exported though /proc...
34  */
35 int nr_pdflush_threads;
36
37 /*
38  * Passed into wb_writeback(), essentially a subset of writeback_control
39  */
40 struct wb_writeback_args {
41         long nr_pages;
42         struct super_block *sb;
43         enum writeback_sync_modes sync_mode;
44         int for_kupdate:1;
45         int range_cyclic:1;
46         int for_background:1;
47 };
48
49 /*
50  * Work items for the bdi_writeback threads
51  */
52 struct bdi_work {
53         struct list_head list;          /* pending work list */
54         struct rcu_head rcu_head;       /* for RCU free/clear of work */
55
56         unsigned long seen;             /* threads that have seen this work */
57         atomic_t pending;               /* number of threads still to do work */
58
59         struct wb_writeback_args args;  /* writeback arguments */
60
61         unsigned long state;            /* flag bits, see WS_* */
62 };
63
64 enum {
65         WS_USED_B = 0,
66         WS_ONSTACK_B,
67 };
68
69 #define WS_USED (1 << WS_USED_B)
70 #define WS_ONSTACK (1 << WS_ONSTACK_B)
71
72 static inline bool bdi_work_on_stack(struct bdi_work *work)
73 {
74         return test_bit(WS_ONSTACK_B, &work->state);
75 }
76
77 static inline void bdi_work_init(struct bdi_work *work,
78                                  struct wb_writeback_args *args)
79 {
80         INIT_RCU_HEAD(&work->rcu_head);
81         work->args = *args;
82         work->state = WS_USED;
83 }
84
85 /**
86  * writeback_in_progress - determine whether there is writeback in progress
87  * @bdi: the device's backing_dev_info structure.
88  *
89  * Determine whether there is writeback waiting to be handled against a
90  * backing device.
91  */
92 int writeback_in_progress(struct backing_dev_info *bdi)
93 {
94         return !list_empty(&bdi->work_list);
95 }
96
97 static void bdi_work_clear(struct bdi_work *work)
98 {
99         clear_bit(WS_USED_B, &work->state);
100         smp_mb__after_clear_bit();
101         /*
102          * work can have disappeared at this point. bit waitq functions
103          * should be able to tolerate this, provided bdi_sched_wait does
104          * not dereference it's pointer argument.
105         */
106         wake_up_bit(&work->state, WS_USED_B);
107 }
108
109 static void bdi_work_free(struct rcu_head *head)
110 {
111         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
112
113         if (!bdi_work_on_stack(work))
114                 kfree(work);
115         else
116                 bdi_work_clear(work);
117 }
118
119 static void wb_work_complete(struct bdi_work *work)
120 {
121         const enum writeback_sync_modes sync_mode = work->args.sync_mode;
122         int onstack = bdi_work_on_stack(work);
123
124         /*
125          * For allocated work, we can clear the done/seen bit right here.
126          * For on-stack work, we need to postpone both the clear and free
127          * to after the RCU grace period, since the stack could be invalidated
128          * as soon as bdi_work_clear() has done the wakeup.
129          */
130         if (!onstack)
131                 bdi_work_clear(work);
132         if (sync_mode == WB_SYNC_NONE || onstack)
133                 call_rcu(&work->rcu_head, bdi_work_free);
134 }
135
136 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
137 {
138         /*
139          * The caller has retrieved the work arguments from this work,
140          * drop our reference. If this is the last ref, delete and free it
141          */
142         if (atomic_dec_and_test(&work->pending)) {
143                 struct backing_dev_info *bdi = wb->bdi;
144
145                 spin_lock(&bdi->wb_lock);
146                 list_del_rcu(&work->list);
147                 spin_unlock(&bdi->wb_lock);
148
149                 wb_work_complete(work);
150         }
151 }
152
153 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
154 {
155         work->seen = bdi->wb_mask;
156         BUG_ON(!work->seen);
157         atomic_set(&work->pending, bdi->wb_cnt);
158         BUG_ON(!bdi->wb_cnt);
159
160         /*
161          * list_add_tail_rcu() contains the necessary barriers to
162          * make sure the above stores are seen before the item is
163          * noticed on the list
164          */
165         spin_lock(&bdi->wb_lock);
166         list_add_tail_rcu(&work->list, &bdi->work_list);
167         spin_unlock(&bdi->wb_lock);
168
169         /*
170          * If the default thread isn't there, make sure we add it. When
171          * it gets created and wakes up, we'll run this work.
172          */
173         if (unlikely(list_empty_careful(&bdi->wb_list)))
174                 wake_up_process(default_backing_dev_info.wb.task);
175         else {
176                 struct bdi_writeback *wb = &bdi->wb;
177
178                 if (wb->task)
179                         wake_up_process(wb->task);
180         }
181 }
182
183 /*
184  * Used for on-stack allocated work items. The caller needs to wait until
185  * the wb threads have acked the work before it's safe to continue.
186  */
187 static void bdi_wait_on_work_clear(struct bdi_work *work)
188 {
189         wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
190                     TASK_UNINTERRUPTIBLE);
191 }
192
193 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
194                                  struct wb_writeback_args *args)
195 {
196         struct bdi_work *work;
197
198         /*
199          * This is WB_SYNC_NONE writeback, so if allocation fails just
200          * wakeup the thread for old dirty data writeback
201          */
202         work = kmalloc(sizeof(*work), GFP_ATOMIC);
203         if (work) {
204                 bdi_work_init(work, args);
205                 bdi_queue_work(bdi, work);
206         } else {
207                 struct bdi_writeback *wb = &bdi->wb;
208
209                 if (wb->task)
210                         wake_up_process(wb->task);
211         }
212 }
213
214 /**
215  * bdi_sync_writeback - start and wait for writeback
216  * @bdi: the backing device to write from
217  * @sb: write inodes from this super_block
218  *
219  * Description:
220  *   This does WB_SYNC_ALL data integrity writeback and waits for the
221  *   IO to complete. Callers must hold the sb s_umount semaphore for
222  *   reading, to avoid having the super disappear before we are done.
223  */
224 static void bdi_sync_writeback(struct backing_dev_info *bdi,
225                                struct super_block *sb)
226 {
227         struct wb_writeback_args args = {
228                 .sb             = sb,
229                 .sync_mode      = WB_SYNC_ALL,
230                 .nr_pages       = LONG_MAX,
231                 .range_cyclic   = 0,
232         };
233         struct bdi_work work;
234
235         bdi_work_init(&work, &args);
236         work.state |= WS_ONSTACK;
237
238         bdi_queue_work(bdi, &work);
239         bdi_wait_on_work_clear(&work);
240 }
241
242 /**
243  * bdi_start_writeback - start writeback
244  * @bdi: the backing device to write from
245  * @nr_pages: the number of pages to write
246  *
247  * Description:
248  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
249  *   started when this function returns, we make no guarentees on
250  *   completion. Caller need not hold sb s_umount semaphore.
251  *
252  */
253 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
254 {
255         struct wb_writeback_args args = {
256                 .sync_mode      = WB_SYNC_NONE,
257                 .nr_pages       = nr_pages,
258                 .range_cyclic   = 1,
259         };
260
261         /*
262          * We treat @nr_pages=0 as the special case to do background writeback,
263          * ie. to sync pages until the background dirty threshold is reached.
264          */
265         if (!nr_pages) {
266                 args.nr_pages = LONG_MAX;
267                 args.for_background = 1;
268         }
269
270         bdi_alloc_queue_work(bdi, &args);
271 }
272
273 /*
274  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
275  * furthest end of its superblock's dirty-inode list.
276  *
277  * Before stamping the inode's ->dirtied_when, we check to see whether it is
278  * already the most-recently-dirtied inode on the b_dirty list.  If that is
279  * the case then the inode must have been redirtied while it was being written
280  * out and we don't reset its dirtied_when.
281  */
282 static void redirty_tail(struct inode *inode)
283 {
284         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
285
286         if (!list_empty(&wb->b_dirty)) {
287                 struct inode *tail;
288
289                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
290                 if (time_before(inode->dirtied_when, tail->dirtied_when))
291                         inode->dirtied_when = jiffies;
292         }
293         list_move(&inode->i_list, &wb->b_dirty);
294 }
295
296 /*
297  * requeue inode for re-scanning after bdi->b_io list is exhausted.
298  */
299 static void requeue_io(struct inode *inode)
300 {
301         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
302
303         list_move(&inode->i_list, &wb->b_more_io);
304 }
305
306 static void inode_sync_complete(struct inode *inode)
307 {
308         /*
309          * Prevent speculative execution through spin_unlock(&inode_lock);
310          */
311         smp_mb();
312         wake_up_bit(&inode->i_state, __I_SYNC);
313 }
314
315 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
316 {
317         bool ret = time_after(inode->dirtied_when, t);
318 #ifndef CONFIG_64BIT
319         /*
320          * For inodes being constantly redirtied, dirtied_when can get stuck.
321          * It _appears_ to be in the future, but is actually in distant past.
322          * This test is necessary to prevent such wrapped-around relative times
323          * from permanently stopping the whole pdflush writeback.
324          */
325         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
326 #endif
327         return ret;
328 }
329
330 /*
331  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
332  */
333 static void move_expired_inodes(struct list_head *delaying_queue,
334                                struct list_head *dispatch_queue,
335                                 unsigned long *older_than_this)
336 {
337         while (!list_empty(delaying_queue)) {
338                 struct inode *inode = list_entry(delaying_queue->prev,
339                                                 struct inode, i_list);
340                 if (older_than_this &&
341                     inode_dirtied_after(inode, *older_than_this))
342                         break;
343                 list_move(&inode->i_list, dispatch_queue);
344         }
345 }
346
347 /*
348  * Queue all expired dirty inodes for io, eldest first.
349  */
350 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
351 {
352         list_splice_init(&wb->b_more_io, wb->b_io.prev);
353         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
354 }
355
356 static int write_inode(struct inode *inode, int sync)
357 {
358         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
359                 return inode->i_sb->s_op->write_inode(inode, sync);
360         return 0;
361 }
362
363 /*
364  * Wait for writeback on an inode to complete.
365  */
366 static void inode_wait_for_writeback(struct inode *inode)
367 {
368         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
369         wait_queue_head_t *wqh;
370
371         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
372         do {
373                 spin_unlock(&inode_lock);
374                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
375                 spin_lock(&inode_lock);
376         } while (inode->i_state & I_SYNC);
377 }
378
379 /*
380  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
381  * caller has ref on the inode (either via __iget or via syscall against an fd)
382  * or the inode has I_WILL_FREE set (via generic_forget_inode)
383  *
384  * If `wait' is set, wait on the writeout.
385  *
386  * The whole writeout design is quite complex and fragile.  We want to avoid
387  * starvation of particular inodes when others are being redirtied, prevent
388  * livelocks, etc.
389  *
390  * Called under inode_lock.
391  */
392 static int
393 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
394 {
395         struct address_space *mapping = inode->i_mapping;
396         int wait = wbc->sync_mode == WB_SYNC_ALL;
397         unsigned dirty;
398         int ret;
399
400         if (!atomic_read(&inode->i_count))
401                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
402         else
403                 WARN_ON(inode->i_state & I_WILL_FREE);
404
405         if (inode->i_state & I_SYNC) {
406                 /*
407                  * If this inode is locked for writeback and we are not doing
408                  * writeback-for-data-integrity, move it to b_more_io so that
409                  * writeback can proceed with the other inodes on s_io.
410                  *
411                  * We'll have another go at writing back this inode when we
412                  * completed a full scan of b_io.
413                  */
414                 if (!wait) {
415                         requeue_io(inode);
416                         return 0;
417                 }
418
419                 /*
420                  * It's a data-integrity sync.  We must wait.
421                  */
422                 inode_wait_for_writeback(inode);
423         }
424
425         BUG_ON(inode->i_state & I_SYNC);
426
427         /* Set I_SYNC, reset I_DIRTY */
428         dirty = inode->i_state & I_DIRTY;
429         inode->i_state |= I_SYNC;
430         inode->i_state &= ~I_DIRTY;
431
432         spin_unlock(&inode_lock);
433
434         ret = do_writepages(mapping, wbc);
435
436         /* Don't write the inode if only I_DIRTY_PAGES was set */
437         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
438                 int err = write_inode(inode, wait);
439                 if (ret == 0)
440                         ret = err;
441         }
442
443         if (wait) {
444                 int err = filemap_fdatawait(mapping);
445                 if (ret == 0)
446                         ret = err;
447         }
448
449         spin_lock(&inode_lock);
450         inode->i_state &= ~I_SYNC;
451         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
452                 if (!(inode->i_state & I_DIRTY) &&
453                     mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
454                         /*
455                          * We didn't write back all the pages.  nfs_writepages()
456                          * sometimes bales out without doing anything. Redirty
457                          * the inode; Move it from b_io onto b_more_io/b_dirty.
458                          */
459                         /*
460                          * akpm: if the caller was the kupdate function we put
461                          * this inode at the head of b_dirty so it gets first
462                          * consideration.  Otherwise, move it to the tail, for
463                          * the reasons described there.  I'm not really sure
464                          * how much sense this makes.  Presumably I had a good
465                          * reasons for doing it this way, and I'd rather not
466                          * muck with it at present.
467                          */
468                         if (wbc->for_kupdate) {
469                                 /*
470                                  * For the kupdate function we move the inode
471                                  * to b_more_io so it will get more writeout as
472                                  * soon as the queue becomes uncongested.
473                                  */
474                                 inode->i_state |= I_DIRTY_PAGES;
475                                 if (wbc->nr_to_write <= 0) {
476                                         /*
477                                          * slice used up: queue for next turn
478                                          */
479                                         requeue_io(inode);
480                                 } else {
481                                         /*
482                                          * somehow blocked: retry later
483                                          */
484                                         redirty_tail(inode);
485                                 }
486                         } else {
487                                 /*
488                                  * Otherwise fully redirty the inode so that
489                                  * other inodes on this superblock will get some
490                                  * writeout.  Otherwise heavy writing to one
491                                  * file would indefinitely suspend writeout of
492                                  * all the other files.
493                                  */
494                                 inode->i_state |= I_DIRTY_PAGES;
495                                 redirty_tail(inode);
496                         }
497                 } else if (inode->i_state & I_DIRTY) {
498                         /*
499                          * Someone redirtied the inode while were writing back
500                          * the pages.
501                          */
502                         redirty_tail(inode);
503                 } else if (atomic_read(&inode->i_count)) {
504                         /*
505                          * The inode is clean, inuse
506                          */
507                         list_move(&inode->i_list, &inode_in_use);
508                 } else {
509                         /*
510                          * The inode is clean, unused
511                          */
512                         list_move(&inode->i_list, &inode_unused);
513                 }
514         }
515         inode_sync_complete(inode);
516         return ret;
517 }
518
519 /*
520  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
521  * before calling writeback. So make sure that we do pin it, so it doesn't
522  * go away while we are writing inodes from it.
523  *
524  * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
525  * 1 if we failed.
526  */
527 static int pin_sb_for_writeback(struct writeback_control *wbc,
528                                    struct inode *inode)
529 {
530         struct super_block *sb = inode->i_sb;
531
532         /*
533          * Caller must already hold the ref for this
534          */
535         if (wbc->sync_mode == WB_SYNC_ALL) {
536                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
537                 return 0;
538         }
539
540         spin_lock(&sb_lock);
541         sb->s_count++;
542         if (down_read_trylock(&sb->s_umount)) {
543                 if (sb->s_root) {
544                         spin_unlock(&sb_lock);
545                         return 0;
546                 }
547                 /*
548                  * umounted, drop rwsem again and fall through to failure
549                  */
550                 up_read(&sb->s_umount);
551         }
552
553         sb->s_count--;
554         spin_unlock(&sb_lock);
555         return 1;
556 }
557
558 static void unpin_sb_for_writeback(struct writeback_control *wbc,
559                                    struct inode *inode)
560 {
561         struct super_block *sb = inode->i_sb;
562
563         if (wbc->sync_mode == WB_SYNC_ALL)
564                 return;
565
566         up_read(&sb->s_umount);
567         put_super(sb);
568 }
569
570 static void writeback_inodes_wb(struct bdi_writeback *wb,
571                                 struct writeback_control *wbc)
572 {
573         struct super_block *sb = wbc->sb;
574         const int is_blkdev_sb = sb_is_blkdev_sb(sb);
575         const unsigned long start = jiffies;    /* livelock avoidance */
576
577         spin_lock(&inode_lock);
578
579         if (!wbc->for_kupdate || list_empty(&wb->b_io))
580                 queue_io(wb, wbc->older_than_this);
581
582         while (!list_empty(&wb->b_io)) {
583                 struct inode *inode = list_entry(wb->b_io.prev,
584                                                 struct inode, i_list);
585                 long pages_skipped;
586
587                 /*
588                  * super block given and doesn't match, skip this inode
589                  */
590                 if (sb && sb != inode->i_sb) {
591                         redirty_tail(inode);
592                         continue;
593                 }
594
595                 if (!bdi_cap_writeback_dirty(wb->bdi)) {
596                         redirty_tail(inode);
597                         if (is_blkdev_sb) {
598                                 /*
599                                  * Dirty memory-backed blockdev: the ramdisk
600                                  * driver does this.  Skip just this inode
601                                  */
602                                 continue;
603                         }
604                         /*
605                          * Dirty memory-backed inode against a filesystem other
606                          * than the kernel-internal bdev filesystem.  Skip the
607                          * entire superblock.
608                          */
609                         break;
610                 }
611
612                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
613                         requeue_io(inode);
614                         continue;
615                 }
616
617                 if (wbc->nonblocking && bdi_write_congested(wb->bdi)) {
618                         wbc->encountered_congestion = 1;
619                         if (!is_blkdev_sb)
620                                 break;          /* Skip a congested fs */
621                         requeue_io(inode);
622                         continue;               /* Skip a congested blockdev */
623                 }
624
625                 /*
626                  * Was this inode dirtied after sync_sb_inodes was called?
627                  * This keeps sync from extra jobs and livelock.
628                  */
629                 if (inode_dirtied_after(inode, start))
630                         break;
631
632                 if (pin_sb_for_writeback(wbc, inode)) {
633                         requeue_io(inode);
634                         continue;
635                 }
636
637                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
638                 __iget(inode);
639                 pages_skipped = wbc->pages_skipped;
640                 writeback_single_inode(inode, wbc);
641                 unpin_sb_for_writeback(wbc, inode);
642                 if (wbc->pages_skipped != pages_skipped) {
643                         /*
644                          * writeback is not making progress due to locked
645                          * buffers.  Skip this inode for now.
646                          */
647                         redirty_tail(inode);
648                 }
649                 spin_unlock(&inode_lock);
650                 iput(inode);
651                 cond_resched();
652                 spin_lock(&inode_lock);
653                 if (wbc->nr_to_write <= 0) {
654                         wbc->more_io = 1;
655                         break;
656                 }
657                 if (!list_empty(&wb->b_more_io))
658                         wbc->more_io = 1;
659         }
660
661         spin_unlock(&inode_lock);
662         /* Leave any unwritten inodes on b_io */
663 }
664
665 void writeback_inodes_wbc(struct writeback_control *wbc)
666 {
667         struct backing_dev_info *bdi = wbc->bdi;
668
669         writeback_inodes_wb(&bdi->wb, wbc);
670 }
671
672 /*
673  * The maximum number of pages to writeout in a single bdi flush/kupdate
674  * operation.  We do this so we don't hold I_SYNC against an inode for
675  * enormous amounts of time, which would block a userspace task which has
676  * been forced to throttle against that inode.  Also, the code reevaluates
677  * the dirty each time it has written this many pages.
678  */
679 #define MAX_WRITEBACK_PAGES     1024
680
681 static inline bool over_bground_thresh(void)
682 {
683         unsigned long background_thresh, dirty_thresh;
684
685         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
686
687         return (global_page_state(NR_FILE_DIRTY) +
688                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
689 }
690
691 /*
692  * Explicit flushing or periodic writeback of "old" data.
693  *
694  * Define "old": the first time one of an inode's pages is dirtied, we mark the
695  * dirtying-time in the inode's address_space.  So this periodic writeback code
696  * just walks the superblock inode list, writing back any inodes which are
697  * older than a specific point in time.
698  *
699  * Try to run once per dirty_writeback_interval.  But if a writeback event
700  * takes longer than a dirty_writeback_interval interval, then leave a
701  * one-second gap.
702  *
703  * older_than_this takes precedence over nr_to_write.  So we'll only write back
704  * all dirty pages if they are all attached to "old" mappings.
705  */
706 static long wb_writeback(struct bdi_writeback *wb,
707                          struct wb_writeback_args *args)
708 {
709         struct writeback_control wbc = {
710                 .bdi                    = wb->bdi,
711                 .sb                     = args->sb,
712                 .sync_mode              = args->sync_mode,
713                 .older_than_this        = NULL,
714                 .for_kupdate            = args->for_kupdate,
715                 .range_cyclic           = args->range_cyclic,
716         };
717         unsigned long oldest_jif;
718         long wrote = 0;
719         struct inode *inode;
720
721         if (wbc.for_kupdate) {
722                 wbc.older_than_this = &oldest_jif;
723                 oldest_jif = jiffies -
724                                 msecs_to_jiffies(dirty_expire_interval * 10);
725         }
726         if (!wbc.range_cyclic) {
727                 wbc.range_start = 0;
728                 wbc.range_end = LLONG_MAX;
729         }
730
731         for (;;) {
732                 /*
733                  * Stop writeback when nr_pages has been consumed
734                  */
735                 if (args->nr_pages <= 0)
736                         break;
737
738                 /*
739                  * For background writeout, stop when we are below the
740                  * background dirty threshold
741                  */
742                 if (args->for_background && !over_bground_thresh())
743                         break;
744
745                 wbc.more_io = 0;
746                 wbc.encountered_congestion = 0;
747                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
748                 wbc.pages_skipped = 0;
749                 writeback_inodes_wb(wb, &wbc);
750                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
751                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
752
753                 /*
754                  * If we ran out of stuff to write, bail unless more_io got set
755                  */
756                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
757                         if (wbc.more_io && !wbc.for_kupdate) {
758                                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
759                                         continue;
760                                 /*
761                                  * Nothing written. Wait for some inode to
762                                  * become available for writeback. Otherwise
763                                  * we'll just busyloop.
764                                  */
765                                 spin_lock(&inode_lock);
766                                 if (!list_empty(&wb->b_more_io))  {
767                                         inode = list_entry(
768                                                         wb->b_more_io.prev,
769                                                         struct inode, i_list);
770                                         inode_wait_for_writeback(inode);
771                                 }
772                                 spin_unlock(&inode_lock);
773                                 continue;
774                         }
775                         break;
776                 }
777         }
778
779         return wrote;
780 }
781
782 /*
783  * Return the next bdi_work struct that hasn't been processed by this
784  * wb thread yet. ->seen is initially set for each thread that exists
785  * for this device, when a thread first notices a piece of work it
786  * clears its bit. Depending on writeback type, the thread will notify
787  * completion on either receiving the work (WB_SYNC_NONE) or after
788  * it is done (WB_SYNC_ALL).
789  */
790 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
791                                            struct bdi_writeback *wb)
792 {
793         struct bdi_work *work, *ret = NULL;
794
795         rcu_read_lock();
796
797         list_for_each_entry_rcu(work, &bdi->work_list, list) {
798                 if (!test_bit(wb->nr, &work->seen))
799                         continue;
800                 clear_bit(wb->nr, &work->seen);
801
802                 ret = work;
803                 break;
804         }
805
806         rcu_read_unlock();
807         return ret;
808 }
809
810 static long wb_check_old_data_flush(struct bdi_writeback *wb)
811 {
812         unsigned long expired;
813         long nr_pages;
814
815         expired = wb->last_old_flush +
816                         msecs_to_jiffies(dirty_writeback_interval * 10);
817         if (time_before(jiffies, expired))
818                 return 0;
819
820         wb->last_old_flush = jiffies;
821         nr_pages = global_page_state(NR_FILE_DIRTY) +
822                         global_page_state(NR_UNSTABLE_NFS) +
823                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
824
825         if (nr_pages) {
826                 struct wb_writeback_args args = {
827                         .nr_pages       = nr_pages,
828                         .sync_mode      = WB_SYNC_NONE,
829                         .for_kupdate    = 1,
830                         .range_cyclic   = 1,
831                 };
832
833                 return wb_writeback(wb, &args);
834         }
835
836         return 0;
837 }
838
839 /*
840  * Retrieve work items and do the writeback they describe
841  */
842 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
843 {
844         struct backing_dev_info *bdi = wb->bdi;
845         struct bdi_work *work;
846         long wrote = 0;
847
848         while ((work = get_next_work_item(bdi, wb)) != NULL) {
849                 struct wb_writeback_args args = work->args;
850
851                 /*
852                  * Override sync mode, in case we must wait for completion
853                  */
854                 if (force_wait)
855                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
856
857                 /*
858                  * If this isn't a data integrity operation, just notify
859                  * that we have seen this work and we are now starting it.
860                  */
861                 if (args.sync_mode == WB_SYNC_NONE)
862                         wb_clear_pending(wb, work);
863
864                 wrote += wb_writeback(wb, &args);
865
866                 /*
867                  * This is a data integrity writeback, so only do the
868                  * notification when we have completed the work.
869                  */
870                 if (args.sync_mode == WB_SYNC_ALL)
871                         wb_clear_pending(wb, work);
872         }
873
874         /*
875          * Check for periodic writeback, kupdated() style
876          */
877         wrote += wb_check_old_data_flush(wb);
878
879         return wrote;
880 }
881
882 /*
883  * Handle writeback of dirty data for the device backed by this bdi. Also
884  * wakes up periodically and does kupdated style flushing.
885  */
886 int bdi_writeback_task(struct bdi_writeback *wb)
887 {
888         unsigned long last_active = jiffies;
889         unsigned long wait_jiffies = -1UL;
890         long pages_written;
891
892         while (!kthread_should_stop()) {
893                 pages_written = wb_do_writeback(wb, 0);
894
895                 if (pages_written)
896                         last_active = jiffies;
897                 else if (wait_jiffies != -1UL) {
898                         unsigned long max_idle;
899
900                         /*
901                          * Longest period of inactivity that we tolerate. If we
902                          * see dirty data again later, the task will get
903                          * recreated automatically.
904                          */
905                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
906                         if (time_after(jiffies, max_idle + last_active))
907                                 break;
908                 }
909
910                 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
911                 schedule_timeout_interruptible(wait_jiffies);
912                 try_to_freeze();
913         }
914
915         return 0;
916 }
917
918 /*
919  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
920  * writeback, for integrity writeback see bdi_sync_writeback().
921  */
922 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
923 {
924         struct wb_writeback_args args = {
925                 .sb             = sb,
926                 .nr_pages       = nr_pages,
927                 .sync_mode      = WB_SYNC_NONE,
928         };
929         struct backing_dev_info *bdi;
930
931         rcu_read_lock();
932
933         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
934                 if (!bdi_has_dirty_io(bdi))
935                         continue;
936
937                 bdi_alloc_queue_work(bdi, &args);
938         }
939
940         rcu_read_unlock();
941 }
942
943 /*
944  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
945  * the whole world.
946  */
947 void wakeup_flusher_threads(long nr_pages)
948 {
949         if (nr_pages == 0)
950                 nr_pages = global_page_state(NR_FILE_DIRTY) +
951                                 global_page_state(NR_UNSTABLE_NFS);
952         bdi_writeback_all(NULL, nr_pages);
953 }
954
955 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
956 {
957         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
958                 struct dentry *dentry;
959                 const char *name = "?";
960
961                 dentry = d_find_alias(inode);
962                 if (dentry) {
963                         spin_lock(&dentry->d_lock);
964                         name = (const char *) dentry->d_name.name;
965                 }
966                 printk(KERN_DEBUG
967                        "%s(%d): dirtied inode %lu (%s) on %s\n",
968                        current->comm, task_pid_nr(current), inode->i_ino,
969                        name, inode->i_sb->s_id);
970                 if (dentry) {
971                         spin_unlock(&dentry->d_lock);
972                         dput(dentry);
973                 }
974         }
975 }
976
977 /**
978  *      __mark_inode_dirty -    internal function
979  *      @inode: inode to mark
980  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
981  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
982  *      mark_inode_dirty_sync.
983  *
984  * Put the inode on the super block's dirty list.
985  *
986  * CAREFUL! We mark it dirty unconditionally, but move it onto the
987  * dirty list only if it is hashed or if it refers to a blockdev.
988  * If it was not hashed, it will never be added to the dirty list
989  * even if it is later hashed, as it will have been marked dirty already.
990  *
991  * In short, make sure you hash any inodes _before_ you start marking
992  * them dirty.
993  *
994  * This function *must* be atomic for the I_DIRTY_PAGES case -
995  * set_page_dirty() is called under spinlock in several places.
996  *
997  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
998  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
999  * the kernel-internal blockdev inode represents the dirtying time of the
1000  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1001  * page->mapping->host, so the page-dirtying time is recorded in the internal
1002  * blockdev inode.
1003  */
1004 void __mark_inode_dirty(struct inode *inode, int flags)
1005 {
1006         struct super_block *sb = inode->i_sb;
1007
1008         /*
1009          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1010          * dirty the inode itself
1011          */
1012         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1013                 if (sb->s_op->dirty_inode)
1014                         sb->s_op->dirty_inode(inode);
1015         }
1016
1017         /*
1018          * make sure that changes are seen by all cpus before we test i_state
1019          * -- mikulas
1020          */
1021         smp_mb();
1022
1023         /* avoid the locking if we can */
1024         if ((inode->i_state & flags) == flags)
1025                 return;
1026
1027         if (unlikely(block_dump))
1028                 block_dump___mark_inode_dirty(inode);
1029
1030         spin_lock(&inode_lock);
1031         if ((inode->i_state & flags) != flags) {
1032                 const int was_dirty = inode->i_state & I_DIRTY;
1033
1034                 inode->i_state |= flags;
1035
1036                 /*
1037                  * If the inode is being synced, just update its dirty state.
1038                  * The unlocker will place the inode on the appropriate
1039                  * superblock list, based upon its state.
1040                  */
1041                 if (inode->i_state & I_SYNC)
1042                         goto out;
1043
1044                 /*
1045                  * Only add valid (hashed) inodes to the superblock's
1046                  * dirty list.  Add blockdev inodes as well.
1047                  */
1048                 if (!S_ISBLK(inode->i_mode)) {
1049                         if (hlist_unhashed(&inode->i_hash))
1050                                 goto out;
1051                 }
1052                 if (inode->i_state & (I_FREEING|I_CLEAR))
1053                         goto out;
1054
1055                 /*
1056                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1057                  * reposition it (that would break b_dirty time-ordering).
1058                  */
1059                 if (!was_dirty) {
1060                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1061                         struct backing_dev_info *bdi = wb->bdi;
1062
1063                         if (bdi_cap_writeback_dirty(bdi) &&
1064                             !test_bit(BDI_registered, &bdi->state)) {
1065                                 WARN_ON(1);
1066                                 printk(KERN_ERR "bdi-%s not registered\n",
1067                                                                 bdi->name);
1068                         }
1069
1070                         inode->dirtied_when = jiffies;
1071                         list_move(&inode->i_list, &wb->b_dirty);
1072                 }
1073         }
1074 out:
1075         spin_unlock(&inode_lock);
1076 }
1077 EXPORT_SYMBOL(__mark_inode_dirty);
1078
1079 /*
1080  * Write out a superblock's list of dirty inodes.  A wait will be performed
1081  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1082  *
1083  * If older_than_this is non-NULL, then only write out inodes which
1084  * had their first dirtying at a time earlier than *older_than_this.
1085  *
1086  * If we're a pdlfush thread, then implement pdflush collision avoidance
1087  * against the entire list.
1088  *
1089  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1090  * This function assumes that the blockdev superblock's inodes are backed by
1091  * a variety of queues, so all inodes are searched.  For other superblocks,
1092  * assume that all inodes are backed by the same queue.
1093  *
1094  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1095  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1096  * on the writer throttling path, and we get decent balancing between many
1097  * throttled threads: we don't want them all piling up on inode_sync_wait.
1098  */
1099 static void wait_sb_inodes(struct super_block *sb)
1100 {
1101         struct inode *inode, *old_inode = NULL;
1102
1103         /*
1104          * We need to be protected against the filesystem going from
1105          * r/o to r/w or vice versa.
1106          */
1107         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1108
1109         spin_lock(&inode_lock);
1110
1111         /*
1112          * Data integrity sync. Must wait for all pages under writeback,
1113          * because there may have been pages dirtied before our sync
1114          * call, but which had writeout started before we write it out.
1115          * In which case, the inode may not be on the dirty list, but
1116          * we still have to wait for that writeout.
1117          */
1118         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1119                 struct address_space *mapping;
1120
1121                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1122                         continue;
1123                 mapping = inode->i_mapping;
1124                 if (mapping->nrpages == 0)
1125                         continue;
1126                 __iget(inode);
1127                 spin_unlock(&inode_lock);
1128                 /*
1129                  * We hold a reference to 'inode' so it couldn't have
1130                  * been removed from s_inodes list while we dropped the
1131                  * inode_lock.  We cannot iput the inode now as we can
1132                  * be holding the last reference and we cannot iput it
1133                  * under inode_lock. So we keep the reference and iput
1134                  * it later.
1135                  */
1136                 iput(old_inode);
1137                 old_inode = inode;
1138
1139                 filemap_fdatawait(mapping);
1140
1141                 cond_resched();
1142
1143                 spin_lock(&inode_lock);
1144         }
1145         spin_unlock(&inode_lock);
1146         iput(old_inode);
1147 }
1148
1149 /**
1150  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1151  * @sb: the superblock
1152  *
1153  * Start writeback on some inodes on this super_block. No guarantees are made
1154  * on how many (if any) will be written, and this function does not wait
1155  * for IO completion of submitted IO. The number of pages submitted is
1156  * returned.
1157  */
1158 void writeback_inodes_sb(struct super_block *sb)
1159 {
1160         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1161         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1162         long nr_to_write;
1163
1164         nr_to_write = nr_dirty + nr_unstable +
1165                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1166
1167         bdi_writeback_all(sb, nr_to_write);
1168 }
1169 EXPORT_SYMBOL(writeback_inodes_sb);
1170
1171 /**
1172  * sync_inodes_sb       -       sync sb inode pages
1173  * @sb: the superblock
1174  *
1175  * This function writes and waits on any dirty inode belonging to this
1176  * super_block. The number of pages synced is returned.
1177  */
1178 void sync_inodes_sb(struct super_block *sb)
1179 {
1180         bdi_sync_writeback(sb->s_bdi, sb);
1181         wait_sb_inodes(sb);
1182 }
1183 EXPORT_SYMBOL(sync_inodes_sb);
1184
1185 /**
1186  * write_inode_now      -       write an inode to disk
1187  * @inode: inode to write to disk
1188  * @sync: whether the write should be synchronous or not
1189  *
1190  * This function commits an inode to disk immediately if it is dirty. This is
1191  * primarily needed by knfsd.
1192  *
1193  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1194  */
1195 int write_inode_now(struct inode *inode, int sync)
1196 {
1197         int ret;
1198         struct writeback_control wbc = {
1199                 .nr_to_write = LONG_MAX,
1200                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1201                 .range_start = 0,
1202                 .range_end = LLONG_MAX,
1203         };
1204
1205         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1206                 wbc.nr_to_write = 0;
1207
1208         might_sleep();
1209         spin_lock(&inode_lock);
1210         ret = writeback_single_inode(inode, &wbc);
1211         spin_unlock(&inode_lock);
1212         if (sync)
1213                 inode_sync_wait(inode);
1214         return ret;
1215 }
1216 EXPORT_SYMBOL(write_inode_now);
1217
1218 /**
1219  * sync_inode - write an inode and its pages to disk.
1220  * @inode: the inode to sync
1221  * @wbc: controls the writeback mode
1222  *
1223  * sync_inode() will write an inode and its pages to disk.  It will also
1224  * correctly update the inode on its superblock's dirty inode lists and will
1225  * update inode->i_state.
1226  *
1227  * The caller must have a ref on the inode.
1228  */
1229 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1230 {
1231         int ret;
1232
1233         spin_lock(&inode_lock);
1234         ret = writeback_single_inode(inode, wbc);
1235         spin_unlock(&inode_lock);
1236         return ret;
1237 }
1238 EXPORT_SYMBOL(sync_inode);