writeback: switch to per-bdi threads for flushing data
[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  * Work items for the bdi_writeback threads
34  */
35 struct bdi_work {
36         struct list_head list;
37         struct list_head wait_list;
38         struct rcu_head rcu_head;
39
40         unsigned long seen;
41         atomic_t pending;
42
43         struct super_block *sb;
44         unsigned long nr_pages;
45         enum writeback_sync_modes sync_mode;
46
47         unsigned long state;
48 };
49
50 enum {
51         WS_USED_B = 0,
52         WS_ONSTACK_B,
53 };
54
55 #define WS_USED (1 << WS_USED_B)
56 #define WS_ONSTACK (1 << WS_ONSTACK_B)
57
58 static inline bool bdi_work_on_stack(struct bdi_work *work)
59 {
60         return test_bit(WS_ONSTACK_B, &work->state);
61 }
62
63 static inline void bdi_work_init(struct bdi_work *work,
64                                  struct writeback_control *wbc)
65 {
66         INIT_RCU_HEAD(&work->rcu_head);
67         work->sb = wbc->sb;
68         work->nr_pages = wbc->nr_to_write;
69         work->sync_mode = wbc->sync_mode;
70         work->state = WS_USED;
71 }
72
73 static inline void bdi_work_init_on_stack(struct bdi_work *work,
74                                           struct writeback_control *wbc)
75 {
76         bdi_work_init(work, wbc);
77         work->state |= WS_ONSTACK;
78 }
79
80 /**
81  * writeback_in_progress - determine whether there is writeback in progress
82  * @bdi: the device's backing_dev_info structure.
83  *
84  * Determine whether there is writeback waiting to be handled against a
85  * backing device.
86  */
87 int writeback_in_progress(struct backing_dev_info *bdi)
88 {
89         return !list_empty(&bdi->work_list);
90 }
91
92 static void bdi_work_clear(struct bdi_work *work)
93 {
94         clear_bit(WS_USED_B, &work->state);
95         smp_mb__after_clear_bit();
96         wake_up_bit(&work->state, WS_USED_B);
97 }
98
99 static void bdi_work_free(struct rcu_head *head)
100 {
101         struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
102
103         if (!bdi_work_on_stack(work))
104                 kfree(work);
105         else
106                 bdi_work_clear(work);
107 }
108
109 static void wb_work_complete(struct bdi_work *work)
110 {
111         const enum writeback_sync_modes sync_mode = work->sync_mode;
112
113         /*
114          * For allocated work, we can clear the done/seen bit right here.
115          * For on-stack work, we need to postpone both the clear and free
116          * to after the RCU grace period, since the stack could be invalidated
117          * as soon as bdi_work_clear() has done the wakeup.
118          */
119         if (!bdi_work_on_stack(work))
120                 bdi_work_clear(work);
121         if (sync_mode == WB_SYNC_NONE || bdi_work_on_stack(work))
122                 call_rcu(&work->rcu_head, bdi_work_free);
123 }
124
125 static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
126 {
127         /*
128          * The caller has retrieved the work arguments from this work,
129          * drop our reference. If this is the last ref, delete and free it
130          */
131         if (atomic_dec_and_test(&work->pending)) {
132                 struct backing_dev_info *bdi = wb->bdi;
133
134                 spin_lock(&bdi->wb_lock);
135                 list_del_rcu(&work->list);
136                 spin_unlock(&bdi->wb_lock);
137
138                 wb_work_complete(work);
139         }
140 }
141
142 static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
143 {
144         if (work) {
145                 work->seen = bdi->wb_mask;
146                 BUG_ON(!work->seen);
147                 atomic_set(&work->pending, bdi->wb_cnt);
148                 BUG_ON(!bdi->wb_cnt);
149
150                 /*
151                  * Make sure stores are seen before it appears on the list
152                  */
153                 smp_mb();
154
155                 spin_lock(&bdi->wb_lock);
156                 list_add_tail_rcu(&work->list, &bdi->work_list);
157                 spin_unlock(&bdi->wb_lock);
158         }
159
160         /*
161          * If the default thread isn't there, make sure we add it. When
162          * it gets created and wakes up, we'll run this work.
163          */
164         if (unlikely(list_empty_careful(&bdi->wb_list)))
165                 wake_up_process(default_backing_dev_info.wb.task);
166         else {
167                 struct bdi_writeback *wb = &bdi->wb;
168
169                 /*
170                  * If we failed allocating the bdi work item, wake up the wb
171                  * thread always. As a safety precaution, it'll flush out
172                  * everything
173                  */
174                 if (!wb_has_dirty_io(wb)) {
175                         if (work)
176                                 wb_clear_pending(wb, work);
177                 } else if (wb->task)
178                         wake_up_process(wb->task);
179         }
180 }
181
182 /*
183  * Used for on-stack allocated work items. The caller needs to wait until
184  * the wb threads have acked the work before it's safe to continue.
185  */
186 static void bdi_wait_on_work_clear(struct bdi_work *work)
187 {
188         wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
189                     TASK_UNINTERRUPTIBLE);
190 }
191
192 static struct bdi_work *bdi_alloc_work(struct writeback_control *wbc)
193 {
194         struct bdi_work *work;
195
196         work = kmalloc(sizeof(*work), GFP_ATOMIC);
197         if (work)
198                 bdi_work_init(work, wbc);
199
200         return work;
201 }
202
203 void bdi_start_writeback(struct writeback_control *wbc)
204 {
205         const bool must_wait = wbc->sync_mode == WB_SYNC_ALL;
206         struct bdi_work work_stack, *work = NULL;
207
208         if (!must_wait)
209                 work = bdi_alloc_work(wbc);
210
211         if (!work) {
212                 work = &work_stack;
213                 bdi_work_init_on_stack(work, wbc);
214         }
215
216         bdi_queue_work(wbc->bdi, work);
217
218         /*
219          * If the sync mode is WB_SYNC_ALL, block waiting for the work to
220          * complete. If not, we only need to wait for the work to be started,
221          * if we allocated it on-stack. We use the same mechanism, if the
222          * wait bit is set in the bdi_work struct, then threads will not
223          * clear pending until after they are done.
224          *
225          * Note that work == &work_stack if must_wait is true, so we don't
226          * need to do call_rcu() here ever, since the completion path will
227          * have done that for us.
228          */
229         if (must_wait || work == &work_stack) {
230                 bdi_wait_on_work_clear(work);
231                 if (work != &work_stack)
232                         call_rcu(&work->rcu_head, bdi_work_free);
233         }
234 }
235
236 /*
237  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
238  * furthest end of its superblock's dirty-inode list.
239  *
240  * Before stamping the inode's ->dirtied_when, we check to see whether it is
241  * already the most-recently-dirtied inode on the b_dirty list.  If that is
242  * the case then the inode must have been redirtied while it was being written
243  * out and we don't reset its dirtied_when.
244  */
245 static void redirty_tail(struct inode *inode)
246 {
247         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
248
249         if (!list_empty(&wb->b_dirty)) {
250                 struct inode *tail;
251
252                 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
253                 if (time_before(inode->dirtied_when, tail->dirtied_when))
254                         inode->dirtied_when = jiffies;
255         }
256         list_move(&inode->i_list, &wb->b_dirty);
257 }
258
259 /*
260  * requeue inode for re-scanning after bdi->b_io list is exhausted.
261  */
262 static void requeue_io(struct inode *inode)
263 {
264         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
265
266         list_move(&inode->i_list, &wb->b_more_io);
267 }
268
269 static void inode_sync_complete(struct inode *inode)
270 {
271         /*
272          * Prevent speculative execution through spin_unlock(&inode_lock);
273          */
274         smp_mb();
275         wake_up_bit(&inode->i_state, __I_SYNC);
276 }
277
278 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
279 {
280         bool ret = time_after(inode->dirtied_when, t);
281 #ifndef CONFIG_64BIT
282         /*
283          * For inodes being constantly redirtied, dirtied_when can get stuck.
284          * It _appears_ to be in the future, but is actually in distant past.
285          * This test is necessary to prevent such wrapped-around relative times
286          * from permanently stopping the whole pdflush writeback.
287          */
288         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
289 #endif
290         return ret;
291 }
292
293 /*
294  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
295  */
296 static void move_expired_inodes(struct list_head *delaying_queue,
297                                struct list_head *dispatch_queue,
298                                 unsigned long *older_than_this)
299 {
300         while (!list_empty(delaying_queue)) {
301                 struct inode *inode = list_entry(delaying_queue->prev,
302                                                 struct inode, i_list);
303                 if (older_than_this &&
304                     inode_dirtied_after(inode, *older_than_this))
305                         break;
306                 list_move(&inode->i_list, dispatch_queue);
307         }
308 }
309
310 /*
311  * Queue all expired dirty inodes for io, eldest first.
312  */
313 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
314 {
315         list_splice_init(&wb->b_more_io, wb->b_io.prev);
316         move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
317 }
318
319 static int write_inode(struct inode *inode, int sync)
320 {
321         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
322                 return inode->i_sb->s_op->write_inode(inode, sync);
323         return 0;
324 }
325
326 /*
327  * Wait for writeback on an inode to complete.
328  */
329 static void inode_wait_for_writeback(struct inode *inode)
330 {
331         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
332         wait_queue_head_t *wqh;
333
334         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
335         do {
336                 spin_unlock(&inode_lock);
337                 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
338                 spin_lock(&inode_lock);
339         } while (inode->i_state & I_SYNC);
340 }
341
342 /*
343  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
344  * caller has ref on the inode (either via __iget or via syscall against an fd)
345  * or the inode has I_WILL_FREE set (via generic_forget_inode)
346  *
347  * If `wait' is set, wait on the writeout.
348  *
349  * The whole writeout design is quite complex and fragile.  We want to avoid
350  * starvation of particular inodes when others are being redirtied, prevent
351  * livelocks, etc.
352  *
353  * Called under inode_lock.
354  */
355 static int
356 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
357 {
358         struct address_space *mapping = inode->i_mapping;
359         int wait = wbc->sync_mode == WB_SYNC_ALL;
360         unsigned dirty;
361         int ret;
362
363         if (!atomic_read(&inode->i_count))
364                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
365         else
366                 WARN_ON(inode->i_state & I_WILL_FREE);
367
368         if (inode->i_state & I_SYNC) {
369                 /*
370                  * If this inode is locked for writeback and we are not doing
371                  * writeback-for-data-integrity, move it to b_more_io so that
372                  * writeback can proceed with the other inodes on s_io.
373                  *
374                  * We'll have another go at writing back this inode when we
375                  * completed a full scan of b_io.
376                  */
377                 if (!wait) {
378                         requeue_io(inode);
379                         return 0;
380                 }
381
382                 /*
383                  * It's a data-integrity sync.  We must wait.
384                  */
385                 inode_wait_for_writeback(inode);
386         }
387
388         BUG_ON(inode->i_state & I_SYNC);
389
390         /* Set I_SYNC, reset I_DIRTY */
391         dirty = inode->i_state & I_DIRTY;
392         inode->i_state |= I_SYNC;
393         inode->i_state &= ~I_DIRTY;
394
395         spin_unlock(&inode_lock);
396
397         ret = do_writepages(mapping, wbc);
398
399         /* Don't write the inode if only I_DIRTY_PAGES was set */
400         if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
401                 int err = write_inode(inode, wait);
402                 if (ret == 0)
403                         ret = err;
404         }
405
406         if (wait) {
407                 int err = filemap_fdatawait(mapping);
408                 if (ret == 0)
409                         ret = err;
410         }
411
412         spin_lock(&inode_lock);
413         inode->i_state &= ~I_SYNC;
414         if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
415                 if (!(inode->i_state & I_DIRTY) &&
416                     mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
417                         /*
418                          * We didn't write back all the pages.  nfs_writepages()
419                          * sometimes bales out without doing anything. Redirty
420                          * the inode; Move it from b_io onto b_more_io/b_dirty.
421                          */
422                         /*
423                          * akpm: if the caller was the kupdate function we put
424                          * this inode at the head of b_dirty so it gets first
425                          * consideration.  Otherwise, move it to the tail, for
426                          * the reasons described there.  I'm not really sure
427                          * how much sense this makes.  Presumably I had a good
428                          * reasons for doing it this way, and I'd rather not
429                          * muck with it at present.
430                          */
431                         if (wbc->for_kupdate) {
432                                 /*
433                                  * For the kupdate function we move the inode
434                                  * to b_more_io so it will get more writeout as
435                                  * soon as the queue becomes uncongested.
436                                  */
437                                 inode->i_state |= I_DIRTY_PAGES;
438                                 if (wbc->nr_to_write <= 0) {
439                                         /*
440                                          * slice used up: queue for next turn
441                                          */
442                                         requeue_io(inode);
443                                 } else {
444                                         /*
445                                          * somehow blocked: retry later
446                                          */
447                                         redirty_tail(inode);
448                                 }
449                         } else {
450                                 /*
451                                  * Otherwise fully redirty the inode so that
452                                  * other inodes on this superblock will get some
453                                  * writeout.  Otherwise heavy writing to one
454                                  * file would indefinitely suspend writeout of
455                                  * all the other files.
456                                  */
457                                 inode->i_state |= I_DIRTY_PAGES;
458                                 redirty_tail(inode);
459                         }
460                 } else if (inode->i_state & I_DIRTY) {
461                         /*
462                          * Someone redirtied the inode while were writing back
463                          * the pages.
464                          */
465                         redirty_tail(inode);
466                 } else if (atomic_read(&inode->i_count)) {
467                         /*
468                          * The inode is clean, inuse
469                          */
470                         list_move(&inode->i_list, &inode_in_use);
471                 } else {
472                         /*
473                          * The inode is clean, unused
474                          */
475                         list_move(&inode->i_list, &inode_unused);
476                 }
477         }
478         inode_sync_complete(inode);
479         return ret;
480 }
481
482 /*
483  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
484  * before calling writeback. So make sure that we do pin it, so it doesn't
485  * go away while we are writing inodes from it.
486  *
487  * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
488  * 1 if we failed.
489  */
490 static int pin_sb_for_writeback(struct writeback_control *wbc,
491                                    struct inode *inode)
492 {
493         struct super_block *sb = inode->i_sb;
494
495         /*
496          * Caller must already hold the ref for this
497          */
498         if (wbc->sync_mode == WB_SYNC_ALL) {
499                 WARN_ON(!rwsem_is_locked(&sb->s_umount));
500                 return 0;
501         }
502
503         spin_lock(&sb_lock);
504         sb->s_count++;
505         if (down_read_trylock(&sb->s_umount)) {
506                 if (sb->s_root) {
507                         spin_unlock(&sb_lock);
508                         return 0;
509                 }
510                 /*
511                  * umounted, drop rwsem again and fall through to failure
512                  */
513                 up_read(&sb->s_umount);
514         }
515
516         sb->s_count--;
517         spin_unlock(&sb_lock);
518         return 1;
519 }
520
521 static void unpin_sb_for_writeback(struct writeback_control *wbc,
522                                    struct inode *inode)
523 {
524         struct super_block *sb = inode->i_sb;
525
526         if (wbc->sync_mode == WB_SYNC_ALL)
527                 return;
528
529         up_read(&sb->s_umount);
530         put_super(sb);
531 }
532
533 static void writeback_inodes_wb(struct bdi_writeback *wb,
534                                 struct writeback_control *wbc)
535 {
536         struct super_block *sb = wbc->sb;
537         const int is_blkdev_sb = sb_is_blkdev_sb(sb);
538         const unsigned long start = jiffies;    /* livelock avoidance */
539
540         spin_lock(&inode_lock);
541
542         if (!wbc->for_kupdate || list_empty(&wb->b_io))
543                 queue_io(wb, wbc->older_than_this);
544
545         while (!list_empty(&wb->b_io)) {
546                 struct inode *inode = list_entry(wb->b_io.prev,
547                                                 struct inode, i_list);
548                 long pages_skipped;
549
550                 /*
551                  * super block given and doesn't match, skip this inode
552                  */
553                 if (sb && sb != inode->i_sb) {
554                         redirty_tail(inode);
555                         continue;
556                 }
557
558                 if (!bdi_cap_writeback_dirty(wb->bdi)) {
559                         redirty_tail(inode);
560                         if (is_blkdev_sb) {
561                                 /*
562                                  * Dirty memory-backed blockdev: the ramdisk
563                                  * driver does this.  Skip just this inode
564                                  */
565                                 continue;
566                         }
567                         /*
568                          * Dirty memory-backed inode against a filesystem other
569                          * than the kernel-internal bdev filesystem.  Skip the
570                          * entire superblock.
571                          */
572                         break;
573                 }
574
575                 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
576                         requeue_io(inode);
577                         continue;
578                 }
579
580                 if (wbc->nonblocking && bdi_write_congested(wb->bdi)) {
581                         wbc->encountered_congestion = 1;
582                         if (!is_blkdev_sb)
583                                 break;          /* Skip a congested fs */
584                         requeue_io(inode);
585                         continue;               /* Skip a congested blockdev */
586                 }
587
588                 /*
589                  * Was this inode dirtied after sync_sb_inodes was called?
590                  * This keeps sync from extra jobs and livelock.
591                  */
592                 if (inode_dirtied_after(inode, start))
593                         break;
594
595                 if (pin_sb_for_writeback(wbc, inode)) {
596                         requeue_io(inode);
597                         continue;
598                 }
599
600                 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
601                 __iget(inode);
602                 pages_skipped = wbc->pages_skipped;
603                 writeback_single_inode(inode, wbc);
604                 unpin_sb_for_writeback(wbc, inode);
605                 if (wbc->pages_skipped != pages_skipped) {
606                         /*
607                          * writeback is not making progress due to locked
608                          * buffers.  Skip this inode for now.
609                          */
610                         redirty_tail(inode);
611                 }
612                 spin_unlock(&inode_lock);
613                 iput(inode);
614                 cond_resched();
615                 spin_lock(&inode_lock);
616                 if (wbc->nr_to_write <= 0) {
617                         wbc->more_io = 1;
618                         break;
619                 }
620                 if (!list_empty(&wb->b_more_io))
621                         wbc->more_io = 1;
622         }
623
624         spin_unlock(&inode_lock);
625         /* Leave any unwritten inodes on b_io */
626 }
627
628 void writeback_inodes_wbc(struct writeback_control *wbc)
629 {
630         struct backing_dev_info *bdi = wbc->bdi;
631
632         writeback_inodes_wb(&bdi->wb, wbc);
633 }
634
635 /*
636  * The maximum number of pages to writeout in a single bdi flush/kupdate
637  * operation.  We do this so we don't hold I_SYNC against an inode for
638  * enormous amounts of time, which would block a userspace task which has
639  * been forced to throttle against that inode.  Also, the code reevaluates
640  * the dirty each time it has written this many pages.
641  */
642 #define MAX_WRITEBACK_PAGES     1024
643
644 static inline bool over_bground_thresh(void)
645 {
646         unsigned long background_thresh, dirty_thresh;
647
648         get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
649
650         return (global_page_state(NR_FILE_DIRTY) +
651                 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
652 }
653
654 /*
655  * Explicit flushing or periodic writeback of "old" data.
656  *
657  * Define "old": the first time one of an inode's pages is dirtied, we mark the
658  * dirtying-time in the inode's address_space.  So this periodic writeback code
659  * just walks the superblock inode list, writing back any inodes which are
660  * older than a specific point in time.
661  *
662  * Try to run once per dirty_writeback_interval.  But if a writeback event
663  * takes longer than a dirty_writeback_interval interval, then leave a
664  * one-second gap.
665  *
666  * older_than_this takes precedence over nr_to_write.  So we'll only write back
667  * all dirty pages if they are all attached to "old" mappings.
668  */
669 static long wb_writeback(struct bdi_writeback *wb, long nr_pages,
670                          struct super_block *sb,
671                          enum writeback_sync_modes sync_mode, int for_kupdate)
672 {
673         struct writeback_control wbc = {
674                 .bdi                    = wb->bdi,
675                 .sb                     = sb,
676                 .sync_mode              = sync_mode,
677                 .older_than_this        = NULL,
678                 .for_kupdate            = for_kupdate,
679                 .range_cyclic           = 1,
680         };
681         unsigned long oldest_jif;
682         long wrote = 0;
683
684         if (wbc.for_kupdate) {
685                 wbc.older_than_this = &oldest_jif;
686                 oldest_jif = jiffies -
687                                 msecs_to_jiffies(dirty_expire_interval * 10);
688         }
689
690         for (;;) {
691                 /*
692                  * Don't flush anything for non-integrity writeback where
693                  * no nr_pages was given
694                  */
695                 if (!for_kupdate && nr_pages <= 0 && sync_mode == WB_SYNC_NONE)
696                         break;
697
698                 /*
699                  * If no specific pages were given and this is just a
700                  * periodic background writeout and we are below the
701                  * background dirty threshold, don't do anything
702                  */
703                 if (for_kupdate && nr_pages <= 0 && !over_bground_thresh())
704                         break;
705
706                 wbc.more_io = 0;
707                 wbc.encountered_congestion = 0;
708                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
709                 wbc.pages_skipped = 0;
710                 writeback_inodes_wb(wb, &wbc);
711                 nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
712                 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
713
714                 /*
715                  * If we ran out of stuff to write, bail unless more_io got set
716                  */
717                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
718                         if (wbc.more_io && !wbc.for_kupdate)
719                                 continue;
720                         break;
721                 }
722         }
723
724         return wrote;
725 }
726
727 /*
728  * Return the next bdi_work struct that hasn't been processed by this
729  * wb thread yet
730  */
731 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
732                                            struct bdi_writeback *wb)
733 {
734         struct bdi_work *work, *ret = NULL;
735
736         rcu_read_lock();
737
738         list_for_each_entry_rcu(work, &bdi->work_list, list) {
739                 if (!test_and_clear_bit(wb->nr, &work->seen))
740                         continue;
741
742                 ret = work;
743                 break;
744         }
745
746         rcu_read_unlock();
747         return ret;
748 }
749
750 static long wb_check_old_data_flush(struct bdi_writeback *wb)
751 {
752         unsigned long expired;
753         long nr_pages;
754
755         expired = wb->last_old_flush +
756                         msecs_to_jiffies(dirty_writeback_interval * 10);
757         if (time_before(jiffies, expired))
758                 return 0;
759
760         wb->last_old_flush = jiffies;
761         nr_pages = global_page_state(NR_FILE_DIRTY) +
762                         global_page_state(NR_UNSTABLE_NFS) +
763                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
764
765         if (nr_pages)
766                 return wb_writeback(wb, nr_pages, NULL, WB_SYNC_NONE, 1);
767
768         return 0;
769 }
770
771 /*
772  * Retrieve work items and do the writeback they describe
773  */
774 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
775 {
776         struct backing_dev_info *bdi = wb->bdi;
777         struct bdi_work *work;
778         long nr_pages, wrote = 0;
779
780         while ((work = get_next_work_item(bdi, wb)) != NULL) {
781                 enum writeback_sync_modes sync_mode;
782
783                 nr_pages = work->nr_pages;
784
785                 /*
786                  * Override sync mode, in case we must wait for completion
787                  */
788                 if (force_wait)
789                         work->sync_mode = sync_mode = WB_SYNC_ALL;
790                 else
791                         sync_mode = work->sync_mode;
792
793                 /*
794                  * If this isn't a data integrity operation, just notify
795                  * that we have seen this work and we are now starting it.
796                  */
797                 if (sync_mode == WB_SYNC_NONE)
798                         wb_clear_pending(wb, work);
799
800                 wrote += wb_writeback(wb, nr_pages, work->sb, sync_mode, 0);
801
802                 /*
803                  * This is a data integrity writeback, so only do the
804                  * notification when we have completed the work.
805                  */
806                 if (sync_mode == WB_SYNC_ALL)
807                         wb_clear_pending(wb, work);
808         }
809
810         /*
811          * Check for periodic writeback, kupdated() style
812          */
813         wrote += wb_check_old_data_flush(wb);
814
815         return wrote;
816 }
817
818 /*
819  * Handle writeback of dirty data for the device backed by this bdi. Also
820  * wakes up periodically and does kupdated style flushing.
821  */
822 int bdi_writeback_task(struct bdi_writeback *wb)
823 {
824         unsigned long last_active = jiffies;
825         unsigned long wait_jiffies = -1UL;
826         long pages_written;
827
828         while (!kthread_should_stop()) {
829                 pages_written = wb_do_writeback(wb, 0);
830
831                 if (pages_written)
832                         last_active = jiffies;
833                 else if (wait_jiffies != -1UL) {
834                         unsigned long max_idle;
835
836                         /*
837                          * Longest period of inactivity that we tolerate. If we
838                          * see dirty data again later, the task will get
839                          * recreated automatically.
840                          */
841                         max_idle = max(5UL * 60 * HZ, wait_jiffies);
842                         if (time_after(jiffies, max_idle + last_active))
843                                 break;
844                 }
845
846                 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
847                 set_current_state(TASK_INTERRUPTIBLE);
848                 schedule_timeout(wait_jiffies);
849                 try_to_freeze();
850         }
851
852         return 0;
853 }
854
855 /*
856  * Schedule writeback for all backing devices. Expensive! If this is a data
857  * integrity operation, writeback will be complete when this returns. If
858  * we are simply called for WB_SYNC_NONE, then writeback will merely be
859  * scheduled to run.
860  */
861 static void bdi_writeback_all(struct writeback_control *wbc)
862 {
863         const bool must_wait = wbc->sync_mode == WB_SYNC_ALL;
864         struct backing_dev_info *bdi;
865         struct bdi_work *work;
866         LIST_HEAD(list);
867
868 restart:
869         spin_lock(&bdi_lock);
870
871         list_for_each_entry(bdi, &bdi_list, bdi_list) {
872                 struct bdi_work *work;
873
874                 if (!bdi_has_dirty_io(bdi))
875                         continue;
876
877                 /*
878                  * If work allocation fails, do the writes inline. We drop
879                  * the lock and restart the list writeout. This should be OK,
880                  * since this happens rarely and because the writeout should
881                  * eventually make more free memory available.
882                  */
883                 work = bdi_alloc_work(wbc);
884                 if (!work) {
885                         struct writeback_control __wbc;
886
887                         /*
888                          * Not a data integrity writeout, just continue
889                          */
890                         if (!must_wait)
891                                 continue;
892
893                         spin_unlock(&bdi_lock);
894                         __wbc = *wbc;
895                         __wbc.bdi = bdi;
896                         writeback_inodes_wbc(&__wbc);
897                         goto restart;
898                 }
899                 if (must_wait)
900                         list_add_tail(&work->wait_list, &list);
901
902                 bdi_queue_work(bdi, work);
903         }
904
905         spin_unlock(&bdi_lock);
906
907         /*
908          * If this is for WB_SYNC_ALL, wait for pending work to complete
909          * before returning.
910          */
911         while (!list_empty(&list)) {
912                 work = list_entry(list.next, struct bdi_work, wait_list);
913                 list_del(&work->wait_list);
914                 bdi_wait_on_work_clear(work);
915                 call_rcu(&work->rcu_head, bdi_work_free);
916         }
917 }
918
919 /*
920  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
921  * the whole world.
922  */
923 void wakeup_flusher_threads(long nr_pages)
924 {
925         struct writeback_control wbc = {
926                 .sync_mode      = WB_SYNC_NONE,
927                 .older_than_this = NULL,
928                 .range_cyclic   = 1,
929         };
930
931         if (nr_pages == 0)
932                 nr_pages = global_page_state(NR_FILE_DIRTY) +
933                                 global_page_state(NR_UNSTABLE_NFS);
934         wbc.nr_to_write = nr_pages;
935         bdi_writeback_all(&wbc);
936 }
937
938 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
939 {
940         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
941                 struct dentry *dentry;
942                 const char *name = "?";
943
944                 dentry = d_find_alias(inode);
945                 if (dentry) {
946                         spin_lock(&dentry->d_lock);
947                         name = (const char *) dentry->d_name.name;
948                 }
949                 printk(KERN_DEBUG
950                        "%s(%d): dirtied inode %lu (%s) on %s\n",
951                        current->comm, task_pid_nr(current), inode->i_ino,
952                        name, inode->i_sb->s_id);
953                 if (dentry) {
954                         spin_unlock(&dentry->d_lock);
955                         dput(dentry);
956                 }
957         }
958 }
959
960 /**
961  *      __mark_inode_dirty -    internal function
962  *      @inode: inode to mark
963  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
964  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
965  *      mark_inode_dirty_sync.
966  *
967  * Put the inode on the super block's dirty list.
968  *
969  * CAREFUL! We mark it dirty unconditionally, but move it onto the
970  * dirty list only if it is hashed or if it refers to a blockdev.
971  * If it was not hashed, it will never be added to the dirty list
972  * even if it is later hashed, as it will have been marked dirty already.
973  *
974  * In short, make sure you hash any inodes _before_ you start marking
975  * them dirty.
976  *
977  * This function *must* be atomic for the I_DIRTY_PAGES case -
978  * set_page_dirty() is called under spinlock in several places.
979  *
980  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
981  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
982  * the kernel-internal blockdev inode represents the dirtying time of the
983  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
984  * page->mapping->host, so the page-dirtying time is recorded in the internal
985  * blockdev inode.
986  */
987 void __mark_inode_dirty(struct inode *inode, int flags)
988 {
989         struct super_block *sb = inode->i_sb;
990
991         /*
992          * Don't do this for I_DIRTY_PAGES - that doesn't actually
993          * dirty the inode itself
994          */
995         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
996                 if (sb->s_op->dirty_inode)
997                         sb->s_op->dirty_inode(inode);
998         }
999
1000         /*
1001          * make sure that changes are seen by all cpus before we test i_state
1002          * -- mikulas
1003          */
1004         smp_mb();
1005
1006         /* avoid the locking if we can */
1007         if ((inode->i_state & flags) == flags)
1008                 return;
1009
1010         if (unlikely(block_dump))
1011                 block_dump___mark_inode_dirty(inode);
1012
1013         spin_lock(&inode_lock);
1014         if ((inode->i_state & flags) != flags) {
1015                 const int was_dirty = inode->i_state & I_DIRTY;
1016
1017                 inode->i_state |= flags;
1018
1019                 /*
1020                  * If the inode is being synced, just update its dirty state.
1021                  * The unlocker will place the inode on the appropriate
1022                  * superblock list, based upon its state.
1023                  */
1024                 if (inode->i_state & I_SYNC)
1025                         goto out;
1026
1027                 /*
1028                  * Only add valid (hashed) inodes to the superblock's
1029                  * dirty list.  Add blockdev inodes as well.
1030                  */
1031                 if (!S_ISBLK(inode->i_mode)) {
1032                         if (hlist_unhashed(&inode->i_hash))
1033                                 goto out;
1034                 }
1035                 if (inode->i_state & (I_FREEING|I_CLEAR))
1036                         goto out;
1037
1038                 /*
1039                  * If the inode was already on b_dirty/b_io/b_more_io, don't
1040                  * reposition it (that would break b_dirty time-ordering).
1041                  */
1042                 if (!was_dirty) {
1043                         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1044
1045                         inode->dirtied_when = jiffies;
1046                         list_move(&inode->i_list, &wb->b_dirty);
1047                 }
1048         }
1049 out:
1050         spin_unlock(&inode_lock);
1051 }
1052 EXPORT_SYMBOL(__mark_inode_dirty);
1053
1054 /*
1055  * Write out a superblock's list of dirty inodes.  A wait will be performed
1056  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1057  *
1058  * If older_than_this is non-NULL, then only write out inodes which
1059  * had their first dirtying at a time earlier than *older_than_this.
1060  *
1061  * If we're a pdlfush thread, then implement pdflush collision avoidance
1062  * against the entire list.
1063  *
1064  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1065  * This function assumes that the blockdev superblock's inodes are backed by
1066  * a variety of queues, so all inodes are searched.  For other superblocks,
1067  * assume that all inodes are backed by the same queue.
1068  *
1069  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1070  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1071  * on the writer throttling path, and we get decent balancing between many
1072  * throttled threads: we don't want them all piling up on inode_sync_wait.
1073  */
1074 static void wait_sb_inodes(struct writeback_control *wbc)
1075 {
1076         struct inode *inode, *old_inode = NULL;
1077
1078         /*
1079          * We need to be protected against the filesystem going from
1080          * r/o to r/w or vice versa.
1081          */
1082         WARN_ON(!rwsem_is_locked(&wbc->sb->s_umount));
1083
1084         spin_lock(&inode_lock);
1085
1086         /*
1087          * Data integrity sync. Must wait for all pages under writeback,
1088          * because there may have been pages dirtied before our sync
1089          * call, but which had writeout started before we write it out.
1090          * In which case, the inode may not be on the dirty list, but
1091          * we still have to wait for that writeout.
1092          */
1093         list_for_each_entry(inode, &wbc->sb->s_inodes, i_sb_list) {
1094                 struct address_space *mapping;
1095
1096                 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1097                         continue;
1098                 mapping = inode->i_mapping;
1099                 if (mapping->nrpages == 0)
1100                         continue;
1101                 __iget(inode);
1102                 spin_unlock(&inode_lock);
1103                 /*
1104                  * We hold a reference to 'inode' so it couldn't have
1105                  * been removed from s_inodes list while we dropped the
1106                  * inode_lock.  We cannot iput the inode now as we can
1107                  * be holding the last reference and we cannot iput it
1108                  * under inode_lock. So we keep the reference and iput
1109                  * it later.
1110                  */
1111                 iput(old_inode);
1112                 old_inode = inode;
1113
1114                 filemap_fdatawait(mapping);
1115
1116                 cond_resched();
1117
1118                 spin_lock(&inode_lock);
1119         }
1120         spin_unlock(&inode_lock);
1121         iput(old_inode);
1122 }
1123
1124 /**
1125  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1126  * @sb: the superblock
1127  *
1128  * Start writeback on some inodes on this super_block. No guarantees are made
1129  * on how many (if any) will be written, and this function does not wait
1130  * for IO completion of submitted IO. The number of pages submitted is
1131  * returned.
1132  */
1133 long writeback_inodes_sb(struct super_block *sb)
1134 {
1135         struct writeback_control wbc = {
1136                 .sb             = sb,
1137                 .sync_mode      = WB_SYNC_NONE,
1138                 .range_start    = 0,
1139                 .range_end      = LLONG_MAX,
1140         };
1141         unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1142         unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1143         long nr_to_write;
1144
1145         nr_to_write = nr_dirty + nr_unstable +
1146                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1147
1148         wbc.nr_to_write = nr_to_write;
1149         bdi_writeback_all(&wbc);
1150         return nr_to_write - wbc.nr_to_write;
1151 }
1152 EXPORT_SYMBOL(writeback_inodes_sb);
1153
1154 /**
1155  * sync_inodes_sb       -       sync sb inode pages
1156  * @sb: the superblock
1157  *
1158  * This function writes and waits on any dirty inode belonging to this
1159  * super_block. The number of pages synced is returned.
1160  */
1161 long sync_inodes_sb(struct super_block *sb)
1162 {
1163         struct writeback_control wbc = {
1164                 .sb             = sb,
1165                 .sync_mode      = WB_SYNC_ALL,
1166                 .range_start    = 0,
1167                 .range_end      = LLONG_MAX,
1168         };
1169         long nr_to_write = LONG_MAX; /* doesn't actually matter */
1170
1171         wbc.nr_to_write = nr_to_write;
1172         bdi_writeback_all(&wbc);
1173         wait_sb_inodes(&wbc);
1174         return nr_to_write - wbc.nr_to_write;
1175 }
1176 EXPORT_SYMBOL(sync_inodes_sb);
1177
1178 /**
1179  * write_inode_now      -       write an inode to disk
1180  * @inode: inode to write to disk
1181  * @sync: whether the write should be synchronous or not
1182  *
1183  * This function commits an inode to disk immediately if it is dirty. This is
1184  * primarily needed by knfsd.
1185  *
1186  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1187  */
1188 int write_inode_now(struct inode *inode, int sync)
1189 {
1190         int ret;
1191         struct writeback_control wbc = {
1192                 .nr_to_write = LONG_MAX,
1193                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1194                 .range_start = 0,
1195                 .range_end = LLONG_MAX,
1196         };
1197
1198         if (!mapping_cap_writeback_dirty(inode->i_mapping))
1199                 wbc.nr_to_write = 0;
1200
1201         might_sleep();
1202         spin_lock(&inode_lock);
1203         ret = writeback_single_inode(inode, &wbc);
1204         spin_unlock(&inode_lock);
1205         if (sync)
1206                 inode_sync_wait(inode);
1207         return ret;
1208 }
1209 EXPORT_SYMBOL(write_inode_now);
1210
1211 /**
1212  * sync_inode - write an inode and its pages to disk.
1213  * @inode: the inode to sync
1214  * @wbc: controls the writeback mode
1215  *
1216  * sync_inode() will write an inode and its pages to disk.  It will also
1217  * correctly update the inode on its superblock's dirty inode lists and will
1218  * update inode->i_state.
1219  *
1220  * The caller must have a ref on the inode.
1221  */
1222 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1223 {
1224         int ret;
1225
1226         spin_lock(&inode_lock);
1227         ret = writeback_single_inode(inode, wbc);
1228         spin_unlock(&inode_lock);
1229         return ret;
1230 }
1231 EXPORT_SYMBOL(sync_inode);
1232
1233 /**
1234  * generic_osync_inode - flush all dirty data for a given inode to disk
1235  * @inode: inode to write
1236  * @mapping: the address_space that should be flushed
1237  * @what:  what to write and wait upon
1238  *
1239  * This can be called by file_write functions for files which have the
1240  * O_SYNC flag set, to flush dirty writes to disk.
1241  *
1242  * @what is a bitmask, specifying which part of the inode's data should be
1243  * written and waited upon.
1244  *
1245  *    OSYNC_DATA:     i_mapping's dirty data
1246  *    OSYNC_METADATA: the buffers at i_mapping->private_list
1247  *    OSYNC_INODE:    the inode itself
1248  */
1249
1250 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
1251 {
1252         int err = 0;
1253         int need_write_inode_now = 0;
1254         int err2;
1255
1256         if (what & OSYNC_DATA)
1257                 err = filemap_fdatawrite(mapping);
1258         if (what & (OSYNC_METADATA|OSYNC_DATA)) {
1259                 err2 = sync_mapping_buffers(mapping);
1260                 if (!err)
1261                         err = err2;
1262         }
1263         if (what & OSYNC_DATA) {
1264                 err2 = filemap_fdatawait(mapping);
1265                 if (!err)
1266                         err = err2;
1267         }
1268
1269         spin_lock(&inode_lock);
1270         if ((inode->i_state & I_DIRTY) &&
1271             ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
1272                 need_write_inode_now = 1;
1273         spin_unlock(&inode_lock);
1274
1275         if (need_write_inode_now) {
1276                 err2 = write_inode_now(inode, 1);
1277                 if (!err)
1278                         err = err2;
1279         }
1280         else
1281                 inode_sync_wait(inode);
1282
1283         return err;
1284 }
1285 EXPORT_SYMBOL(generic_osync_inode);