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