Merge branch 'for-linus' of git://git.kernel.dk/linux-2.6-block
[sfrench/cifs-2.6.git] / fs / fs-writeback.c
1 /*
2  * fs/fs-writeback.c
3  *
4  * Copyright (C) 2002, Linus Torvalds.
5  *
6  * Contains all the functions related to writing back and waiting
7  * upon dirty inodes against superblocks, and writing back dirty
8  * pages against inodes.  ie: data writeback.  Writeout of the
9  * inode itself is not handled here.
10  *
11  * 10Apr2002    Andrew Morton
12  *              Split out of fs/inode.c
13  *              Additions for address_space-based writeback
14  */
15
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include "internal.h"
30
31 #define inode_to_bdi(inode)     ((inode)->i_mapping->backing_dev_info)
32
33 /*
34  * We don't actually have pdflush, but this one is exported though /proc...
35  */
36 int nr_pdflush_threads;
37
38 /*
39  * Passed into wb_writeback(), essentially a subset of writeback_control
40  */
41 struct wb_writeback_args {
42         long nr_pages;
43         struct super_block *sb;
44         enum writeback_sync_modes sync_mode;
45         unsigned int for_kupdate:1;
46         unsigned int range_cyclic:1;
47         unsigned int for_background:1;
48 };
49
50 /*
51  * Work items for the bdi_writeback threads
52  */
53 struct bdi_work {
54         struct list_head list;          /* pending work list */
55         struct rcu_head rcu_head;       /* for RCU free/clear of work */
56
57         unsigned long seen;             /* threads that have seen this work */
58         atomic_t pending;               /* number of threads still to do work */
59
60         struct wb_writeback_args args;  /* writeback arguments */
61
62         unsigned long state;            /* flag bits, see WS_* */
63 };
64
65 enum {
66         WS_INPROGRESS = 0,
67         WS_ONSTACK,
68 };
69
70 static inline void bdi_work_init(struct bdi_work *work,
71                                  struct wb_writeback_args *args)
72 {
73         INIT_RCU_HEAD(&work->rcu_head);
74         work->args = *args;
75         __set_bit(WS_INPROGRESS, &work->state);
76 }
77
78 /**
79  * writeback_in_progress - determine whether there is writeback in progress
80  * @bdi: the device's backing_dev_info structure.
81  *
82  * Determine whether there is writeback waiting to be handled against a
83  * backing device.
84  */
85 int writeback_in_progress(struct backing_dev_info *bdi)
86 {
87         return !list_empty(&bdi->work_list);
88 }
89
90 static void bdi_work_free(struct rcu_head *head)
91 {
92         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
93
94         clear_bit(WS_INPROGRESS, &work->state);
95         smp_mb__after_clear_bit();
96         wake_up_bit(&work->state, WS_INPROGRESS);
97
98         if (!test_bit(WS_ONSTACK, &work->state))
99                 kfree(work);
100 }
101
102 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
103 {
104         /*
105          * The caller has retrieved the work arguments from this work,
106          * drop our reference. If this is the last ref, delete and free it
107          */
108         if (atomic_dec_and_test(&work->pending)) {
109                 struct backing_dev_info *bdi = wb->bdi;
110
111                 spin_lock(&bdi->wb_lock);
112                 list_del_rcu(&work->list);
113                 spin_unlock(&bdi->wb_lock);
114
115                 call_rcu(&work->rcu_head, bdi_work_free);
116         }
117 }
118
119 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
120 {
121         work->seen = bdi->wb_mask;
122         BUG_ON(!work->seen);
123         atomic_set(&work->pending, bdi->wb_cnt);
124         BUG_ON(!bdi->wb_cnt);
125
126         /*
127          * list_add_tail_rcu() contains the necessary barriers to
128          * make sure the above stores are seen before the item is
129          * noticed on the list
130          */
131         spin_lock(&bdi->wb_lock);
132         list_add_tail_rcu(&work->list, &bdi->work_list);
133         spin_unlock(&bdi->wb_lock);
134
135         /*
136          * If the default thread isn't there, make sure we add it. When
137          * it gets created and wakes up, we'll run this work.
138          */
139         if (unlikely(list_empty_careful(&bdi->wb_list)))
140                 wake_up_process(default_backing_dev_info.wb.task);
141         else {
142                 struct bdi_writeback *wb = &bdi->wb;
143
144                 if (wb->task)
145                         wake_up_process(wb->task);
146         }
147 }
148
149 /*
150  * Used for on-stack allocated work items. The caller needs to wait until
151  * the wb threads have acked the work before it's safe to continue.
152  */
153 static void bdi_wait_on_work_done(struct bdi_work *work)
154 {
155         wait_on_bit(&work->state, WS_INPROGRESS, bdi_sched_wait,
156                     TASK_UNINTERRUPTIBLE);
157 }
158
159 static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
160                                  struct wb_writeback_args *args)
161 {
162         struct bdi_work *work;
163
164         /*
165          * This is WB_SYNC_NONE writeback, so if allocation fails just
166          * wakeup the thread for old dirty data writeback
167          */
168         work = kmalloc(sizeof(*work), GFP_ATOMIC);
169         if (work) {
170                 bdi_work_init(work, args);
171                 bdi_queue_work(bdi, work);
172         } else {
173                 struct bdi_writeback *wb = &bdi->wb;
174
175                 if (wb->task)
176                         wake_up_process(wb->task);
177         }
178 }
179
180 /**
181  * bdi_queue_work_onstack - start and wait for writeback
182  * @sb: write inodes from this super_block
183  *
184  * Description:
185  *   This function initiates writeback and waits for the operation to
186  *   complete. Callers must hold the sb s_umount semaphore for
187  *   reading, to avoid having the super disappear before we are done.
188  */
189 static void bdi_queue_work_onstack(struct wb_writeback_args *args)
190 {
191         struct bdi_work work;
192
193         bdi_work_init(&work, args);
194         __set_bit(WS_ONSTACK, &work.state);
195
196         bdi_queue_work(args->sb->s_bdi, &work);
197         bdi_wait_on_work_done(&work);
198 }
199
200 /**
201  * bdi_start_writeback - start writeback
202  * @bdi: the backing device to write from
203  * @nr_pages: the number of pages to write
204  *
205  * Description:
206  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
207  *   started when this function returns, we make no guarentees on
208  *   completion. Caller need not hold sb s_umount semaphore.
209  *
210  */
211 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
212 {
213         struct wb_writeback_args args = {
214                 .sync_mode      = WB_SYNC_NONE,
215                 .nr_pages       = nr_pages,
216                 .range_cyclic   = 1,
217         };
218
219         bdi_alloc_queue_work(bdi, &args);
220 }
221
222 /**
223  * bdi_start_background_writeback - start background writeback
224  * @bdi: the backing device to write from
225  *
226  * Description:
227  *   This does WB_SYNC_NONE background writeback. The IO is only
228  *   started when this function returns, we make no guarentees on
229  *   completion. Caller need not hold sb s_umount semaphore.
230  */
231 void bdi_start_background_writeback(struct backing_dev_info *bdi)
232 {
233         struct wb_writeback_args args = {
234                 .sync_mode      = WB_SYNC_NONE,
235                 .nr_pages       = LONG_MAX,
236                 .for_background = 1,
237                 .range_cyclic   = 1,
238         };
239         bdi_alloc_queue_work(bdi, &args);
240 }
241
242 /*
243  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
244  * furthest end of its superblock's dirty-inode list.
245  *
246  * Before stamping the inode's ->dirtied_when, we check to see whether it is
247  * already the most-recently-dirtied inode on the b_dirty list.  If that is
248  * the case then the inode must have been redirtied while it was being written
249  * out and we don't reset its dirtied_when.
250  */
251 static void redirty_tail(struct inode *inode)
252 {
253         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
254
255         if (!list_empty(&wb->b_dirty)) {
256                 struct inode *tail;
257
258                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
259                 if (time_before(inode->dirtied_when, tail->dirtied_when))
260                         inode->dirtied_when = jiffies;
261         }
262         list_move(&inode->i_list, &wb->b_dirty);
263 }
264
265 /*
266  * requeue inode for re-scanning after bdi->b_io list is exhausted.
267  */
268 static void requeue_io(struct inode *inode)
269 {
270         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
271
272         list_move(&inode->i_list, &wb->b_more_io);
273 }
274
275 static void inode_sync_complete(struct inode *inode)
276 {
277         /*
278          * Prevent speculative execution through spin_unlock(&inode_lock);
279          */
280         smp_mb();
281         wake_up_bit(&inode->i_state, __I_SYNC);
282 }
283
284 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
285 {
286         bool ret = time_after(inode->dirtied_when, t);
287 #ifndef CONFIG_64BIT
288         /*
289          * For inodes being constantly redirtied, dirtied_when can get stuck.
290          * It _appears_ to be in the future, but is actually in distant past.
291          * This test is necessary to prevent such wrapped-around relative times
292          * from permanently stopping the whole bdi writeback.
293          */
294         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
295 #endif
296         return ret;
297 }
298
299 /*
300  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
301  */
302 static void move_expired_inodes(struct list_head *delaying_queue,
303                                struct list_head *dispatch_queue,
304                                 unsigned long *older_than_this)
305 {
306         LIST_HEAD(tmp);
307         struct list_head *pos, *node;
308         struct super_block *sb = NULL;
309         struct inode *inode;
310         int do_sb_sort = 0;
311
312         while (!list_empty(delaying_queue)) {
313                 inode = list_entry(delaying_queue->prev, struct inode, i_list);
314                 if (older_than_this &&
315                     inode_dirtied_after(inode, *older_than_this))
316                         break;
317                 if (sb && sb != inode->i_sb)
318                         do_sb_sort = 1;
319                 sb = inode->i_sb;
320                 list_move(&inode->i_list, &tmp);
321         }
322
323         /* just one sb in list, splice to dispatch_queue and we're done */
324         if (!do_sb_sort) {
325                 list_splice(&tmp, dispatch_queue);
326                 return;
327         }
328
329         /* Move inodes from one superblock together */
330         while (!list_empty(&tmp)) {
331                 inode = list_entry(tmp.prev, struct inode, i_list);
332                 sb = inode->i_sb;
333                 list_for_each_prev_safe(pos, node, &tmp) {
334                         inode = list_entry(pos, struct inode, i_list);
335                         if (inode->i_sb == sb)
336                                 list_move(&inode->i_list, dispatch_queue);
337                 }
338         }
339 }
340
341 /*
342  * Queue all expired dirty inodes for io, eldest first.
343  */
344 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
345 {
346         list_splice_init(&wb->b_more_io, wb->b_io.prev);
347         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
348 }
349
350 static int write_inode(struct inode *inode, struct writeback_control *wbc)
351 {
352         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
353                 return inode->i_sb->s_op->write_inode(inode, wbc);
354         return 0;
355 }
356
357 /*
358  * Wait for writeback on an inode to complete.
359  */
360 static void inode_wait_for_writeback(struct inode *inode)
361 {
362         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
363         wait_queue_head_t *wqh;
364
365         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
366          while (inode->i_state & I_SYNC) {
367                 spin_unlock(&inode_lock);
368                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
369                 spin_lock(&inode_lock);
370         }
371 }
372
373 /*
374  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
375  * caller has ref on the inode (either via __iget or via syscall against an fd)
376  * or the inode has I_WILL_FREE set (via generic_forget_inode)
377  *
378  * If `wait' is set, wait on the writeout.
379  *
380  * The whole writeout design is quite complex and fragile.  We want to avoid
381  * starvation of particular inodes when others are being redirtied, prevent
382  * livelocks, etc.
383  *
384  * Called under inode_lock.
385  */
386 static int
387 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
388 {
389         struct address_space *mapping = inode->i_mapping;
390         unsigned dirty;
391         int ret;
392
393         if (!atomic_read(&inode->i_count))
394                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
395         else
396                 WARN_ON(inode->i_state & I_WILL_FREE);
397
398         if (inode->i_state & I_SYNC) {
399                 /*
400                  * If this inode is locked for writeback and we are not doing
401                  * writeback-for-data-integrity, move it to b_more_io so that
402                  * writeback can proceed with the other inodes on s_io.
403                  *
404                  * We'll have another go at writing back this inode when we
405                  * completed a full scan of b_io.
406                  */
407                 if (wbc->sync_mode != WB_SYNC_ALL) {
408                         requeue_io(inode);
409                         return 0;
410                 }
411
412                 /*
413                  * It's a data-integrity sync.  We must wait.
414                  */
415                 inode_wait_for_writeback(inode);
416         }
417
418         BUG_ON(inode->i_state & I_SYNC);
419
420         /* Set I_SYNC, reset I_DIRTY_PAGES */
421         inode->i_state |= I_SYNC;
422         inode->i_state &= ~I_DIRTY_PAGES;
423         spin_unlock(&inode_lock);
424
425         ret = do_writepages(mapping, wbc);
426
427         /*
428          * Make sure to wait on the data before writing out the metadata.
429          * This is important for filesystems that modify metadata on data
430          * I/O completion.
431          */
432         if (wbc->sync_mode == WB_SYNC_ALL) {
433                 int err = filemap_fdatawait(mapping);
434                 if (ret == 0)
435                         ret = err;
436         }
437
438         /*
439          * Some filesystems may redirty the inode during the writeback
440          * due to delalloc, clear dirty metadata flags right before
441          * write_inode()
442          */
443         spin_lock(&inode_lock);
444         dirty = inode->i_state & I_DIRTY;
445         inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
446         spin_unlock(&inode_lock);
447         /* Don't write the inode if only I_DIRTY_PAGES was set */
448         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
449                 int err = write_inode(inode, wbc);
450                 if (ret == 0)
451                         ret = err;
452         }
453
454         spin_lock(&inode_lock);
455         inode->i_state &= ~I_SYNC;
456         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
457                 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
458                         /*
459                          * More pages get dirtied by a fast dirtier.
460                          */
461                         goto select_queue;
462                 } else if (inode->i_state & I_DIRTY) {
463                         /*
464                          * At least XFS will redirty the inode during the
465                          * writeback (delalloc) and on io completion (isize).
466                          */
467                         redirty_tail(inode);
468                 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
469                         /*
470                          * We didn't write back all the pages.  nfs_writepages()
471                          * sometimes bales out without doing anything. Redirty
472                          * the inode; Move it from b_io onto b_more_io/b_dirty.
473                          */
474                         /*
475                          * akpm: if the caller was the kupdate function we put
476                          * this inode at the head of b_dirty so it gets first
477                          * consideration.  Otherwise, move it to the tail, for
478                          * the reasons described there.  I'm not really sure
479                          * how much sense this makes.  Presumably I had a good
480                          * reasons for doing it this way, and I'd rather not
481                          * muck with it at present.
482                          */
483                         if (wbc->for_kupdate) {
484                                 /*
485                                  * For the kupdate function we move the inode
486                                  * to b_more_io so it will get more writeout as
487                                  * soon as the queue becomes uncongested.
488                                  */
489                                 inode->i_state |= I_DIRTY_PAGES;
490 select_queue:
491                                 if (wbc->nr_to_write <= 0) {
492                                         /*
493                                          * slice used up: queue for next turn
494                                          */
495                                         requeue_io(inode);
496                                 } else {
497                                         /*
498                                          * somehow blocked: retry later
499                                          */
500                                         redirty_tail(inode);
501                                 }
502                         } else {
503                                 /*
504                                  * Otherwise fully redirty the inode so that
505                                  * other inodes on this superblock will get some
506                                  * writeout.  Otherwise heavy writing to one
507                                  * file would indefinitely suspend writeout of
508                                  * all the other files.
509                                  */
510                                 inode->i_state |= I_DIRTY_PAGES;
511                                 redirty_tail(inode);
512                         }
513                 } else if (atomic_read(&inode->i_count)) {
514                         /*
515                          * The inode is clean, inuse
516                          */
517                         list_move(&inode->i_list, &inode_in_use);
518                 } else {
519                         /*
520                          * The inode is clean, unused
521                          */
522                         list_move(&inode->i_list, &inode_unused);
523                 }
524         }
525         inode_sync_complete(inode);
526         return ret;
527 }
528
529 /*
530  * For background writeback the caller does not have the sb pinned
531  * before calling writeback. So make sure that we do pin it, so it doesn't
532  * go away while we are writing inodes from it.
533  */
534 static bool pin_sb_for_writeback(struct super_block *sb)
535 {
536         spin_lock(&sb_lock);
537         if (list_empty(&sb->s_instances)) {
538                 spin_unlock(&sb_lock);
539                 return false;
540         }
541
542         sb->s_count++;
543         spin_unlock(&sb_lock);
544
545         if (down_read_trylock(&sb->s_umount)) {
546                 if (sb->s_root)
547                         return true;
548                 up_read(&sb->s_umount);
549         }
550
551         put_super(sb);
552         return false;
553 }
554
555 /*
556  * Write a portion of b_io inodes which belong to @sb.
557  * If @wbc->sb != NULL, then find and write all such
558  * inodes. Otherwise write only ones which go sequentially
559  * in reverse order.
560  * Return 1, if the caller writeback routine should be
561  * interrupted. Otherwise return 0.
562  */
563 static int writeback_sb_inodes(struct super_block *sb,
564                                struct bdi_writeback *wb,
565                                struct writeback_control *wbc)
566 {
567         while (!list_empty(&wb->b_io)) {
568                 long pages_skipped;
569                 struct inode *inode = list_entry(wb->b_io.prev,
570                                                  struct inode, i_list);
571                 if (wbc->sb && sb != inode->i_sb) {
572                         /* super block given and doesn't
573                            match, skip this inode */
574                         redirty_tail(inode);
575                         continue;
576                 }
577                 if (sb != inode->i_sb)
578                         /* finish with this superblock */
579                         return 0;
580                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
581                         requeue_io(inode);
582                         continue;
583                 }
584                 /*
585                  * Was this inode dirtied after sync_sb_inodes was called?
586                  * This keeps sync from extra jobs and livelock.
587                  */
588                 if (inode_dirtied_after(inode, wbc->wb_start))
589                         return 1;
590
591                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
592                 __iget(inode);
593                 pages_skipped = wbc->pages_skipped;
594                 writeback_single_inode(inode, wbc);
595                 if (wbc->pages_skipped != pages_skipped) {
596                         /*
597                          * writeback is not making progress due to locked
598                          * buffers.  Skip this inode for now.
599                          */
600                         redirty_tail(inode);
601                 }
602                 spin_unlock(&inode_lock);
603                 iput(inode);
604                 cond_resched();
605                 spin_lock(&inode_lock);
606                 if (wbc->nr_to_write <= 0) {
607                         wbc->more_io = 1;
608                         return 1;
609                 }
610                 if (!list_empty(&wb->b_more_io))
611                         wbc->more_io = 1;
612         }
613         /* b_io is empty */
614         return 1;
615 }
616
617 static void writeback_inodes_wb(struct bdi_writeback *wb,
618                                 struct writeback_control *wbc)
619 {
620         int ret = 0;
621
622         wbc->wb_start = jiffies; /* livelock avoidance */
623         spin_lock(&inode_lock);
624         if (!wbc->for_kupdate || list_empty(&wb->b_io))
625                 queue_io(wb, wbc->older_than_this);
626
627         while (!list_empty(&wb->b_io)) {
628                 struct inode *inode = list_entry(wb->b_io.prev,
629                                                  struct inode, i_list);
630                 struct super_block *sb = inode->i_sb;
631
632                 if (wbc->sb) {
633                         /*
634                          * We are requested to write out inodes for a specific
635                          * superblock.  This means we already have s_umount
636                          * taken by the caller which also waits for us to
637                          * complete the writeout.
638                          */
639                         if (sb != wbc->sb) {
640                                 redirty_tail(inode);
641                                 continue;
642                         }
643
644                         WARN_ON(!rwsem_is_locked(&sb->s_umount));
645
646                         ret = writeback_sb_inodes(sb, wb, wbc);
647                 } else {
648                         if (!pin_sb_for_writeback(sb)) {
649                                 requeue_io(inode);
650                                 continue;
651                         }
652                         ret = writeback_sb_inodes(sb, wb, wbc);
653                         drop_super(sb);
654                 }
655
656                 if (ret)
657                         break;
658         }
659         spin_unlock(&inode_lock);
660         /* Leave any unwritten inodes on b_io */
661 }
662
663 void writeback_inodes_wbc(struct writeback_control *wbc)
664 {
665         struct backing_dev_info *bdi = wbc->bdi;
666
667         writeback_inodes_wb(&bdi->wb, wbc);
668 }
669
670 /*
671  * The maximum number of pages to writeout in a single bdi flush/kupdate
672  * operation.  We do this so we don't hold I_SYNC against an inode for
673  * enormous amounts of time, which would block a userspace task which has
674  * been forced to throttle against that inode.  Also, the code reevaluates
675  * the dirty each time it has written this many pages.
676  */
677 #define MAX_WRITEBACK_PAGES     1024
678
679 static inline bool over_bground_thresh(void)
680 {
681         unsigned long background_thresh, dirty_thresh;
682
683         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
684
685         return (global_page_state(NR_FILE_DIRTY) +
686                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
687 }
688
689 /*
690  * Explicit flushing or periodic writeback of "old" data.
691  *
692  * Define "old": the first time one of an inode's pages is dirtied, we mark the
693  * dirtying-time in the inode's address_space.  So this periodic writeback code
694  * just walks the superblock inode list, writing back any inodes which are
695  * older than a specific point in time.
696  *
697  * Try to run once per dirty_writeback_interval.  But if a writeback event
698  * takes longer than a dirty_writeback_interval interval, then leave a
699  * one-second gap.
700  *
701  * older_than_this takes precedence over nr_to_write.  So we'll only write back
702  * all dirty pages if they are all attached to "old" mappings.
703  */
704 static long wb_writeback(struct bdi_writeback *wb,
705                          struct wb_writeback_args *args)
706 {
707         struct writeback_control wbc = {
708                 .bdi                    = wb->bdi,
709                 .sb                     = args->sb,
710                 .sync_mode              = args->sync_mode,
711                 .older_than_this        = NULL,
712                 .for_kupdate            = args->for_kupdate,
713                 .for_background         = args->for_background,
714                 .range_cyclic           = args->range_cyclic,
715         };
716         unsigned long oldest_jif;
717         long wrote = 0;
718         struct inode *inode;
719
720         if (wbc.for_kupdate) {
721                 wbc.older_than_this = &oldest_jif;
722                 oldest_jif = jiffies -
723                                 msecs_to_jiffies(dirty_expire_interval * 10);
724         }
725         if (!wbc.range_cyclic) {
726                 wbc.range_start = 0;
727                 wbc.range_end = LLONG_MAX;
728         }
729
730         for (;;) {
731                 /*
732                  * Stop writeback when nr_pages has been consumed
733                  */
734                 if (args->nr_pages <= 0)
735                         break;
736
737                 /*
738                  * For background writeout, stop when we are below the
739                  * background dirty threshold
740                  */
741                 if (args->for_background && !over_bground_thresh())
742                         break;
743
744                 wbc.more_io = 0;
745                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
746                 wbc.pages_skipped = 0;
747                 writeback_inodes_wb(wb, &wbc);
748                 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
749                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
750
751                 /*
752                  * If we consumed everything, see if we have more
753                  */
754                 if (wbc.nr_to_write <= 0)
755                         continue;
756                 /*
757                  * Didn't write everything and we don't have more IO, bail
758                  */
759                 if (!wbc.more_io)
760                         break;
761                 /*
762                  * Did we write something? Try for more
763                  */
764                 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
765                         continue;
766                 /*
767                  * Nothing written. Wait for some inode to
768                  * become available for writeback. Otherwise
769                  * we'll just busyloop.
770                  */
771                 spin_lock(&inode_lock);
772                 if (!list_empty(&wb->b_more_io))  {
773                         inode = list_entry(wb->b_more_io.prev,
774                                                 struct inode, i_list);
775                         inode_wait_for_writeback(inode);
776                 }
777                 spin_unlock(&inode_lock);
778         }
779
780         return wrote;
781 }
782
783 /*
784  * Return the next bdi_work struct that hasn't been processed by this
785  * wb thread yet. ->seen is initially set for each thread that exists
786  * for this device, when a thread first notices a piece of work it
787  * clears its bit. Depending on writeback type, the thread will notify
788  * completion on either receiving the work (WB_SYNC_NONE) or after
789  * it is done (WB_SYNC_ALL).
790  */
791 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
792                                            struct bdi_writeback *wb)
793 {
794         struct bdi_work *work, *ret = NULL;
795
796         rcu_read_lock();
797
798         list_for_each_entry_rcu(work, &bdi->work_list, list) {
799                 if (!test_bit(wb->nr, &work->seen))
800                         continue;
801                 clear_bit(wb->nr, &work->seen);
802
803                 ret = work;
804                 break;
805         }
806
807         rcu_read_unlock();
808         return ret;
809 }
810
811 static long wb_check_old_data_flush(struct bdi_writeback *wb)
812 {
813         unsigned long expired;
814         long nr_pages;
815
816         /*
817          * When set to zero, disable periodic writeback
818          */
819         if (!dirty_writeback_interval)
820                 return 0;
821
822         expired = wb->last_old_flush +
823                         msecs_to_jiffies(dirty_writeback_interval * 10);
824         if (time_before(jiffies, expired))
825                 return 0;
826
827         wb->last_old_flush = jiffies;
828         nr_pages = global_page_state(NR_FILE_DIRTY) +
829                         global_page_state(NR_UNSTABLE_NFS) +
830                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
831
832         if (nr_pages) {
833                 struct wb_writeback_args args = {
834                         .nr_pages       = nr_pages,
835                         .sync_mode      = WB_SYNC_NONE,
836                         .for_kupdate    = 1,
837                         .range_cyclic   = 1,
838                 };
839
840                 return wb_writeback(wb, &args);
841         }
842
843         return 0;
844 }
845
846 /*
847  * Retrieve work items and do the writeback they describe
848  */
849 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
850 {
851         struct backing_dev_info *bdi = wb->bdi;
852         struct bdi_work *work;
853         long wrote = 0;
854
855         while ((work = get_next_work_item(bdi, wb)) != NULL) {
856                 struct wb_writeback_args args = work->args;
857
858                 /*
859                  * Override sync mode, in case we must wait for completion
860                  */
861                 if (force_wait)
862                         work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
863
864                 /*
865                  * If this isn't a data integrity operation, just notify
866                  * that we have seen this work and we are now starting it.
867                  */
868                 if (!test_bit(WS_ONSTACK, &work->state))
869                         wb_clear_pending(wb, work);
870
871                 wrote += wb_writeback(wb, &args);
872
873                 /*
874                  * This is a data integrity writeback, so only do the
875                  * notification when we have completed the work.
876                  */
877                 if (test_bit(WS_ONSTACK, &work->state))
878                         wb_clear_pending(wb, work);
879         }
880
881         /*
882          * Check for periodic writeback, kupdated() style
883          */
884         wrote += wb_check_old_data_flush(wb);
885
886         return wrote;
887 }
888
889 /*
890  * Handle writeback of dirty data for the device backed by this bdi. Also
891  * wakes up periodically and does kupdated style flushing.
892  */
893 int bdi_writeback_task(struct bdi_writeback *wb)
894 {
895         unsigned long last_active = jiffies;
896         unsigned long wait_jiffies = -1UL;
897         long pages_written;
898
899         while (!kthread_should_stop()) {
900                 pages_written = wb_do_writeback(wb, 0);
901
902                 if (pages_written)
903                         last_active = jiffies;
904                 else if (wait_jiffies != -1UL) {
905                         unsigned long max_idle;
906
907                         /*
908                          * Longest period of inactivity that we tolerate. If we
909                          * see dirty data again later, the task will get
910                          * recreated automatically.
911                          */
912                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
913                         if (time_after(jiffies, max_idle + last_active))
914                                 break;
915                 }
916
917                 if (dirty_writeback_interval) {
918                         wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
919                         schedule_timeout_interruptible(wait_jiffies);
920                 } else {
921                         set_current_state(TASK_INTERRUPTIBLE);
922                         if (list_empty_careful(&wb->bdi->work_list) &&
923                             !kthread_should_stop())
924                                 schedule();
925                         __set_current_state(TASK_RUNNING);
926                 }
927
928                 try_to_freeze();
929         }
930
931         return 0;
932 }
933
934 /*
935  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
936  * the whole world.
937  */
938 void wakeup_flusher_threads(long nr_pages)
939 {
940         struct backing_dev_info *bdi;
941         struct wb_writeback_args args = {
942                 .sync_mode      = WB_SYNC_NONE,
943         };
944
945         if (nr_pages) {
946                 args.nr_pages = nr_pages;
947         } else {
948                 args.nr_pages = global_page_state(NR_FILE_DIRTY) +
949                                 global_page_state(NR_UNSTABLE_NFS);
950         }
951
952         rcu_read_lock();
953         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
954                 if (!bdi_has_dirty_io(bdi))
955                         continue;
956                 bdi_alloc_queue_work(bdi, &args);
957         }
958         rcu_read_unlock();
959 }
960
961 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
962 {
963         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
964                 struct dentry *dentry;
965                 const char *name = "?";
966
967                 dentry = d_find_alias(inode);
968                 if (dentry) {
969                         spin_lock(&dentry->d_lock);
970                         name = (const char *) dentry->d_name.name;
971                 }
972                 printk(KERN_DEBUG
973                        "%s(%d): dirtied inode %lu (%s) on %s\n",
974                        current->comm, task_pid_nr(current), inode->i_ino,
975                        name, inode->i_sb->s_id);
976                 if (dentry) {
977                         spin_unlock(&dentry->d_lock);
978                         dput(dentry);
979                 }
980         }
981 }
982
983 /**
984  *      __mark_inode_dirty -    internal function
985  *      @inode: inode to mark
986  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
987  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
988  *      mark_inode_dirty_sync.
989  *
990  * Put the inode on the super block's dirty list.
991  *
992  * CAREFUL! We mark it dirty unconditionally, but move it onto the
993  * dirty list only if it is hashed or if it refers to a blockdev.
994  * If it was not hashed, it will never be added to the dirty list
995  * even if it is later hashed, as it will have been marked dirty already.
996  *
997  * In short, make sure you hash any inodes _before_ you start marking
998  * them dirty.
999  *
1000  * This function *must* be atomic for the I_DIRTY_PAGES case -
1001  * set_page_dirty() is called under spinlock in several places.
1002  *
1003  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1004  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1005  * the kernel-internal blockdev inode represents the dirtying time of the
1006  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1007  * page->mapping->host, so the page-dirtying time is recorded in the internal
1008  * blockdev inode.
1009  */
1010 void __mark_inode_dirty(struct inode *inode, int flags)
1011 {
1012         struct super_block *sb = inode->i_sb;
1013
1014         /*
1015          * Don't do this for I_DIRTY_PAGES - that doesn't actually
1016          * dirty the inode itself
1017          */
1018         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1019                 if (sb->s_op->dirty_inode)
1020                         sb->s_op->dirty_inode(inode);
1021         }
1022
1023         /*
1024          * make sure that changes are seen by all cpus before we test i_state
1025          * -- mikulas
1026          */
1027         smp_mb();
1028
1029         /* avoid the locking if we can */
1030         if ((inode->i_state & flags) == flags)
1031                 return;
1032
1033         if (unlikely(block_dump))
1034                 block_dump___mark_inode_dirty(inode);
1035
1036         spin_lock(&inode_lock);
1037         if ((inode->i_state & flags) != flags) {
1038                 const int was_dirty = inode->i_state & I_DIRTY;
1039
1040                 inode->i_state |= flags;
1041
1042                 /*
1043                  * If the inode is being synced, just update its dirty state.
1044                  * The unlocker will place the inode on the appropriate
1045                  * superblock list, based upon its state.
1046                  */
1047                 if (inode->i_state & I_SYNC)
1048                         goto out;
1049
1050                 /*
1051                  * Only add valid (hashed) inodes to the superblock's
1052                  * dirty list.  Add blockdev inodes as well.
1053                  */
1054                 if (!S_ISBLK(inode->i_mode)) {
1055                         if (hlist_unhashed(&inode->i_hash))
1056                                 goto out;
1057                 }
1058                 if (inode->i_state & (I_FREEING|I_CLEAR))
1059                         goto out;
1060
1061                 /*
1062                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1063                  * reposition it (that would break b_dirty time-ordering).
1064                  */
1065                 if (!was_dirty) {
1066                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1067                         struct backing_dev_info *bdi = wb->bdi;
1068
1069                         if (bdi_cap_writeback_dirty(bdi) &&
1070                             !test_bit(BDI_registered, &bdi->state)) {
1071                                 WARN_ON(1);
1072                                 printk(KERN_ERR "bdi-%s not registered\n",
1073                                                                 bdi->name);
1074                         }
1075
1076                         inode->dirtied_when = jiffies;
1077                         list_move(&inode->i_list, &wb->b_dirty);
1078                 }
1079         }
1080 out:
1081         spin_unlock(&inode_lock);
1082 }
1083 EXPORT_SYMBOL(__mark_inode_dirty);
1084
1085 /*
1086  * Write out a superblock's list of dirty inodes.  A wait will be performed
1087  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1088  *
1089  * If older_than_this is non-NULL, then only write out inodes which
1090  * had their first dirtying at a time earlier than *older_than_this.
1091  *
1092  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1093  * This function assumes that the blockdev superblock's inodes are backed by
1094  * a variety of queues, so all inodes are searched.  For other superblocks,
1095  * assume that all inodes are backed by the same queue.
1096  *
1097  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1098  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1099  * on the writer throttling path, and we get decent balancing between many
1100  * throttled threads: we don't want them all piling up on inode_sync_wait.
1101  */
1102 static void wait_sb_inodes(struct super_block *sb)
1103 {
1104         struct inode *inode, *old_inode = NULL;
1105
1106         /*
1107          * We need to be protected against the filesystem going from
1108          * r/o to r/w or vice versa.
1109          */
1110         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1111
1112         spin_lock(&inode_lock);
1113
1114         /*
1115          * Data integrity sync. Must wait for all pages under writeback,
1116          * because there may have been pages dirtied before our sync
1117          * call, but which had writeout started before we write it out.
1118          * In which case, the inode may not be on the dirty list, but
1119          * we still have to wait for that writeout.
1120          */
1121         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1122                 struct address_space *mapping;
1123
1124                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1125                         continue;
1126                 mapping = inode->i_mapping;
1127                 if (mapping->nrpages == 0)
1128                         continue;
1129                 __iget(inode);
1130                 spin_unlock(&inode_lock);
1131                 /*
1132                  * We hold a reference to 'inode' so it couldn't have
1133                  * been removed from s_inodes list while we dropped the
1134                  * inode_lock.  We cannot iput the inode now as we can
1135                  * be holding the last reference and we cannot iput it
1136                  * under inode_lock. So we keep the reference and iput
1137                  * it later.
1138                  */
1139                 iput(old_inode);
1140                 old_inode = inode;
1141
1142                 filemap_fdatawait(mapping);
1143
1144                 cond_resched();
1145
1146                 spin_lock(&inode_lock);
1147         }
1148         spin_unlock(&inode_lock);
1149         iput(old_inode);
1150 }
1151
1152 /**
1153  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1154  * @sb: the superblock
1155  *
1156  * Start writeback on some inodes on this super_block. No guarantees are made
1157  * on how many (if any) will be written, and this function does not wait
1158  * for IO completion of submitted IO. The number of pages submitted is
1159  * returned.
1160  */
1161 void writeback_inodes_sb(struct super_block *sb)
1162 {
1163         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1164         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1165         struct wb_writeback_args args = {
1166                 .sb             = sb,
1167                 .sync_mode      = WB_SYNC_NONE,
1168         };
1169
1170         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1171
1172         args.nr_pages = nr_dirty + nr_unstable +
1173                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1174
1175         bdi_queue_work_onstack(&args);
1176 }
1177 EXPORT_SYMBOL(writeback_inodes_sb);
1178
1179 /**
1180  * writeback_inodes_sb_if_idle  -       start writeback if none underway
1181  * @sb: the superblock
1182  *
1183  * Invoke writeback_inodes_sb if no writeback is currently underway.
1184  * Returns 1 if writeback was started, 0 if not.
1185  */
1186 int writeback_inodes_sb_if_idle(struct super_block *sb)
1187 {
1188         if (!writeback_in_progress(sb->s_bdi)) {
1189                 down_read(&sb->s_umount);
1190                 writeback_inodes_sb(sb);
1191                 up_read(&sb->s_umount);
1192                 return 1;
1193         } else
1194                 return 0;
1195 }
1196 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1197
1198 /**
1199  * sync_inodes_sb       -       sync sb inode pages
1200  * @sb: the superblock
1201  *
1202  * This function writes and waits on any dirty inode belonging to this
1203  * super_block. The number of pages synced is returned.
1204  */
1205 void sync_inodes_sb(struct super_block *sb)
1206 {
1207         struct wb_writeback_args args = {
1208                 .sb             = sb,
1209                 .sync_mode      = WB_SYNC_ALL,
1210                 .nr_pages       = LONG_MAX,
1211                 .range_cyclic   = 0,
1212         };
1213
1214         WARN_ON(!rwsem_is_locked(&sb->s_umount));
1215
1216         bdi_queue_work_onstack(&args);
1217         wait_sb_inodes(sb);
1218 }
1219 EXPORT_SYMBOL(sync_inodes_sb);
1220
1221 /**
1222  * write_inode_now      -       write an inode to disk
1223  * @inode: inode to write to disk
1224  * @sync: whether the write should be synchronous or not
1225  *
1226  * This function commits an inode to disk immediately if it is dirty. This is
1227  * primarily needed by knfsd.
1228  *
1229  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1230  */
1231 int write_inode_now(struct inode *inode, int sync)
1232 {
1233         int ret;
1234         struct writeback_control wbc = {
1235                 .nr_to_write = LONG_MAX,
1236                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1237                 .range_start = 0,
1238                 .range_end = LLONG_MAX,
1239         };
1240
1241         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1242                 wbc.nr_to_write = 0;
1243
1244         might_sleep();
1245         spin_lock(&inode_lock);
1246         ret = writeback_single_inode(inode, &wbc);
1247         spin_unlock(&inode_lock);
1248         if (sync)
1249                 inode_sync_wait(inode);
1250         return ret;
1251 }
1252 EXPORT_SYMBOL(write_inode_now);
1253
1254 /**
1255  * sync_inode - write an inode and its pages to disk.
1256  * @inode: the inode to sync
1257  * @wbc: controls the writeback mode
1258  *
1259  * sync_inode() will write an inode and its pages to disk.  It will also
1260  * correctly update the inode on its superblock's dirty inode lists and will
1261  * update inode->i_state.
1262  *
1263  * The caller must have a ref on the inode.
1264  */
1265 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1266 {
1267         int ret;
1268
1269         spin_lock(&inode_lock);
1270         ret = writeback_single_inode(inode, wbc);
1271         spin_unlock(&inode_lock);
1272         return ret;
1273 }
1274 EXPORT_SYMBOL(sync_inode);