writeback: avoid livelocking WB_SYNC_ALL writeback
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / fs / fs-writeback.c
1 /*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
31
32 /*
33 * Passed into wb_writeback(), essentially a subset of writeback_control
34 */
35 struct wb_writeback_work {
36 long nr_pages;
37 struct super_block *sb;
38 enum writeback_sync_modes sync_mode;
39 unsigned int for_kupdate:1;
40 unsigned int range_cyclic:1;
41 unsigned int for_background:1;
42
43 struct list_head list; /* pending work list */
44 struct completion *done; /* set if the caller waits */
45 };
46
47 /*
48 * Include the creation of the trace points after defining the
49 * wb_writeback_work structure so that the definition remains local to this
50 * file.
51 */
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/writeback.h>
54
55 /*
56 * We don't actually have pdflush, but this one is exported though /proc...
57 */
58 int nr_pdflush_threads;
59
60 /**
61 * writeback_in_progress - determine whether there is writeback in progress
62 * @bdi: the device's backing_dev_info structure.
63 *
64 * Determine whether there is writeback waiting to be handled against a
65 * backing device.
66 */
67 int writeback_in_progress(struct backing_dev_info *bdi)
68 {
69 return test_bit(BDI_writeback_running, &bdi->state);
70 }
71
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
73 {
74 struct super_block *sb = inode->i_sb;
75
76 if (strcmp(sb->s_type->name, "bdev") == 0)
77 return inode->i_mapping->backing_dev_info;
78
79 return sb->s_bdi;
80 }
81
82 static inline struct inode *wb_inode(struct list_head *head)
83 {
84 return list_entry(head, struct inode, i_wb_list);
85 }
86
87 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
88 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
89 {
90 if (bdi->wb.task) {
91 wake_up_process(bdi->wb.task);
92 } else {
93 /*
94 * The bdi thread isn't there, wake up the forker thread which
95 * will create and run it.
96 */
97 wake_up_process(default_backing_dev_info.wb.task);
98 }
99 }
100
101 static void bdi_queue_work(struct backing_dev_info *bdi,
102 struct wb_writeback_work *work)
103 {
104 trace_writeback_queue(bdi, work);
105
106 spin_lock_bh(&bdi->wb_lock);
107 list_add_tail(&work->list, &bdi->work_list);
108 if (!bdi->wb.task)
109 trace_writeback_nothread(bdi, work);
110 bdi_wakeup_flusher(bdi);
111 spin_unlock_bh(&bdi->wb_lock);
112 }
113
114 static void
115 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
116 bool range_cyclic)
117 {
118 struct wb_writeback_work *work;
119
120 /*
121 * This is WB_SYNC_NONE writeback, so if allocation fails just
122 * wakeup the thread for old dirty data writeback
123 */
124 work = kzalloc(sizeof(*work), GFP_ATOMIC);
125 if (!work) {
126 if (bdi->wb.task) {
127 trace_writeback_nowork(bdi);
128 wake_up_process(bdi->wb.task);
129 }
130 return;
131 }
132
133 work->sync_mode = WB_SYNC_NONE;
134 work->nr_pages = nr_pages;
135 work->range_cyclic = range_cyclic;
136
137 bdi_queue_work(bdi, work);
138 }
139
140 /**
141 * bdi_start_writeback - start writeback
142 * @bdi: the backing device to write from
143 * @nr_pages: the number of pages to write
144 *
145 * Description:
146 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
147 * started when this function returns, we make no guarentees on
148 * completion. Caller need not hold sb s_umount semaphore.
149 *
150 */
151 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
152 {
153 __bdi_start_writeback(bdi, nr_pages, true);
154 }
155
156 /**
157 * bdi_start_background_writeback - start background writeback
158 * @bdi: the backing device to write from
159 *
160 * Description:
161 * This makes sure WB_SYNC_NONE background writeback happens. When
162 * this function returns, it is only guaranteed that for given BDI
163 * some IO is happening if we are over background dirty threshold.
164 * Caller need not hold sb s_umount semaphore.
165 */
166 void bdi_start_background_writeback(struct backing_dev_info *bdi)
167 {
168 /*
169 * We just wake up the flusher thread. It will perform background
170 * writeback as soon as there is no other work to do.
171 */
172 trace_writeback_wake_background(bdi);
173 spin_lock_bh(&bdi->wb_lock);
174 bdi_wakeup_flusher(bdi);
175 spin_unlock_bh(&bdi->wb_lock);
176 }
177
178 /*
179 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
180 * furthest end of its superblock's dirty-inode list.
181 *
182 * Before stamping the inode's ->dirtied_when, we check to see whether it is
183 * already the most-recently-dirtied inode on the b_dirty list. If that is
184 * the case then the inode must have been redirtied while it was being written
185 * out and we don't reset its dirtied_when.
186 */
187 static void redirty_tail(struct inode *inode)
188 {
189 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
190
191 if (!list_empty(&wb->b_dirty)) {
192 struct inode *tail;
193
194 tail = wb_inode(wb->b_dirty.next);
195 if (time_before(inode->dirtied_when, tail->dirtied_when))
196 inode->dirtied_when = jiffies;
197 }
198 list_move(&inode->i_wb_list, &wb->b_dirty);
199 }
200
201 /*
202 * requeue inode for re-scanning after bdi->b_io list is exhausted.
203 */
204 static void requeue_io(struct inode *inode)
205 {
206 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
207
208 list_move(&inode->i_wb_list, &wb->b_more_io);
209 }
210
211 static void inode_sync_complete(struct inode *inode)
212 {
213 /*
214 * Prevent speculative execution through spin_unlock(&inode_lock);
215 */
216 smp_mb();
217 wake_up_bit(&inode->i_state, __I_SYNC);
218 }
219
220 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
221 {
222 bool ret = time_after(inode->dirtied_when, t);
223 #ifndef CONFIG_64BIT
224 /*
225 * For inodes being constantly redirtied, dirtied_when can get stuck.
226 * It _appears_ to be in the future, but is actually in distant past.
227 * This test is necessary to prevent such wrapped-around relative times
228 * from permanently stopping the whole bdi writeback.
229 */
230 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
231 #endif
232 return ret;
233 }
234
235 /*
236 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
237 */
238 static void move_expired_inodes(struct list_head *delaying_queue,
239 struct list_head *dispatch_queue,
240 unsigned long *older_than_this)
241 {
242 LIST_HEAD(tmp);
243 struct list_head *pos, *node;
244 struct super_block *sb = NULL;
245 struct inode *inode;
246 int do_sb_sort = 0;
247
248 while (!list_empty(delaying_queue)) {
249 inode = wb_inode(delaying_queue->prev);
250 if (older_than_this &&
251 inode_dirtied_after(inode, *older_than_this))
252 break;
253 if (sb && sb != inode->i_sb)
254 do_sb_sort = 1;
255 sb = inode->i_sb;
256 list_move(&inode->i_wb_list, &tmp);
257 }
258
259 /* just one sb in list, splice to dispatch_queue and we're done */
260 if (!do_sb_sort) {
261 list_splice(&tmp, dispatch_queue);
262 return;
263 }
264
265 /* Move inodes from one superblock together */
266 while (!list_empty(&tmp)) {
267 sb = wb_inode(tmp.prev)->i_sb;
268 list_for_each_prev_safe(pos, node, &tmp) {
269 inode = wb_inode(pos);
270 if (inode->i_sb == sb)
271 list_move(&inode->i_wb_list, dispatch_queue);
272 }
273 }
274 }
275
276 /*
277 * Queue all expired dirty inodes for io, eldest first.
278 * Before
279 * newly dirtied b_dirty b_io b_more_io
280 * =============> gf edc BA
281 * After
282 * newly dirtied b_dirty b_io b_more_io
283 * =============> g fBAedc
284 * |
285 * +--> dequeue for IO
286 */
287 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
288 {
289 list_splice_init(&wb->b_more_io, &wb->b_io);
290 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
291 }
292
293 static int write_inode(struct inode *inode, struct writeback_control *wbc)
294 {
295 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
296 return inode->i_sb->s_op->write_inode(inode, wbc);
297 return 0;
298 }
299
300 /*
301 * Wait for writeback on an inode to complete.
302 */
303 static void inode_wait_for_writeback(struct inode *inode)
304 {
305 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
306 wait_queue_head_t *wqh;
307
308 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
309 while (inode->i_state & I_SYNC) {
310 spin_unlock(&inode_lock);
311 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
312 spin_lock(&inode_lock);
313 }
314 }
315
316 /*
317 * Write out an inode's dirty pages. Called under inode_lock. Either the
318 * caller has ref on the inode (either via __iget or via syscall against an fd)
319 * or the inode has I_WILL_FREE set (via generic_forget_inode)
320 *
321 * If `wait' is set, wait on the writeout.
322 *
323 * The whole writeout design is quite complex and fragile. We want to avoid
324 * starvation of particular inodes when others are being redirtied, prevent
325 * livelocks, etc.
326 *
327 * Called under inode_lock.
328 */
329 static int
330 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
331 {
332 struct address_space *mapping = inode->i_mapping;
333 unsigned dirty;
334 int ret;
335
336 if (!atomic_read(&inode->i_count))
337 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
338 else
339 WARN_ON(inode->i_state & I_WILL_FREE);
340
341 if (inode->i_state & I_SYNC) {
342 /*
343 * If this inode is locked for writeback and we are not doing
344 * writeback-for-data-integrity, move it to b_more_io so that
345 * writeback can proceed with the other inodes on s_io.
346 *
347 * We'll have another go at writing back this inode when we
348 * completed a full scan of b_io.
349 */
350 if (wbc->sync_mode != WB_SYNC_ALL) {
351 requeue_io(inode);
352 return 0;
353 }
354
355 /*
356 * It's a data-integrity sync. We must wait.
357 */
358 inode_wait_for_writeback(inode);
359 }
360
361 BUG_ON(inode->i_state & I_SYNC);
362
363 /* Set I_SYNC, reset I_DIRTY_PAGES */
364 inode->i_state |= I_SYNC;
365 inode->i_state &= ~I_DIRTY_PAGES;
366 spin_unlock(&inode_lock);
367
368 ret = do_writepages(mapping, wbc);
369
370 /*
371 * Make sure to wait on the data before writing out the metadata.
372 * This is important for filesystems that modify metadata on data
373 * I/O completion.
374 */
375 if (wbc->sync_mode == WB_SYNC_ALL) {
376 int err = filemap_fdatawait(mapping);
377 if (ret == 0)
378 ret = err;
379 }
380
381 /*
382 * Some filesystems may redirty the inode during the writeback
383 * due to delalloc, clear dirty metadata flags right before
384 * write_inode()
385 */
386 spin_lock(&inode_lock);
387 dirty = inode->i_state & I_DIRTY;
388 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
389 spin_unlock(&inode_lock);
390 /* Don't write the inode if only I_DIRTY_PAGES was set */
391 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
392 int err = write_inode(inode, wbc);
393 if (ret == 0)
394 ret = err;
395 }
396
397 spin_lock(&inode_lock);
398 inode->i_state &= ~I_SYNC;
399 if (!(inode->i_state & I_FREEING)) {
400 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
401 /*
402 * We didn't write back all the pages. nfs_writepages()
403 * sometimes bales out without doing anything.
404 */
405 inode->i_state |= I_DIRTY_PAGES;
406 if (wbc->nr_to_write <= 0) {
407 /*
408 * slice used up: queue for next turn
409 */
410 requeue_io(inode);
411 } else {
412 /*
413 * Writeback blocked by something other than
414 * congestion. Delay the inode for some time to
415 * avoid spinning on the CPU (100% iowait)
416 * retrying writeback of the dirty page/inode
417 * that cannot be performed immediately.
418 */
419 redirty_tail(inode);
420 }
421 } else if (inode->i_state & I_DIRTY) {
422 /*
423 * Filesystems can dirty the inode during writeback
424 * operations, such as delayed allocation during
425 * submission or metadata updates after data IO
426 * completion.
427 */
428 redirty_tail(inode);
429 } else {
430 /*
431 * The inode is clean. At this point we either have
432 * a reference to the inode or it's on it's way out.
433 * No need to add it back to the LRU.
434 */
435 list_del_init(&inode->i_wb_list);
436 }
437 }
438 inode_sync_complete(inode);
439 return ret;
440 }
441
442 /*
443 * For background writeback the caller does not have the sb pinned
444 * before calling writeback. So make sure that we do pin it, so it doesn't
445 * go away while we are writing inodes from it.
446 */
447 static bool pin_sb_for_writeback(struct super_block *sb)
448 {
449 spin_lock(&sb_lock);
450 if (list_empty(&sb->s_instances)) {
451 spin_unlock(&sb_lock);
452 return false;
453 }
454
455 sb->s_count++;
456 spin_unlock(&sb_lock);
457
458 if (down_read_trylock(&sb->s_umount)) {
459 if (sb->s_root)
460 return true;
461 up_read(&sb->s_umount);
462 }
463
464 put_super(sb);
465 return false;
466 }
467
468 /*
469 * Write a portion of b_io inodes which belong to @sb.
470 *
471 * If @only_this_sb is true, then find and write all such
472 * inodes. Otherwise write only ones which go sequentially
473 * in reverse order.
474 *
475 * Return 1, if the caller writeback routine should be
476 * interrupted. Otherwise return 0.
477 */
478 static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
479 struct writeback_control *wbc, bool only_this_sb)
480 {
481 while (!list_empty(&wb->b_io)) {
482 long pages_skipped;
483 struct inode *inode = wb_inode(wb->b_io.prev);
484
485 if (inode->i_sb != sb) {
486 if (only_this_sb) {
487 /*
488 * We only want to write back data for this
489 * superblock, move all inodes not belonging
490 * to it back onto the dirty list.
491 */
492 redirty_tail(inode);
493 continue;
494 }
495
496 /*
497 * The inode belongs to a different superblock.
498 * Bounce back to the caller to unpin this and
499 * pin the next superblock.
500 */
501 return 0;
502 }
503
504 /*
505 * Don't bother with new inodes or inodes beeing freed, first
506 * kind does not need peridic writeout yet, and for the latter
507 * kind writeout is handled by the freer.
508 */
509 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
510 requeue_io(inode);
511 continue;
512 }
513
514 /*
515 * Was this inode dirtied after sync_sb_inodes was called?
516 * This keeps sync from extra jobs and livelock.
517 */
518 if (inode_dirtied_after(inode, wbc->wb_start))
519 return 1;
520
521 __iget(inode);
522 pages_skipped = wbc->pages_skipped;
523 writeback_single_inode(inode, wbc);
524 if (wbc->pages_skipped != pages_skipped) {
525 /*
526 * writeback is not making progress due to locked
527 * buffers. Skip this inode for now.
528 */
529 redirty_tail(inode);
530 }
531 spin_unlock(&inode_lock);
532 iput(inode);
533 cond_resched();
534 spin_lock(&inode_lock);
535 if (wbc->nr_to_write <= 0) {
536 wbc->more_io = 1;
537 return 1;
538 }
539 if (!list_empty(&wb->b_more_io))
540 wbc->more_io = 1;
541 }
542 /* b_io is empty */
543 return 1;
544 }
545
546 void writeback_inodes_wb(struct bdi_writeback *wb,
547 struct writeback_control *wbc)
548 {
549 int ret = 0;
550
551 if (!wbc->wb_start)
552 wbc->wb_start = jiffies; /* livelock avoidance */
553 spin_lock(&inode_lock);
554 if (!wbc->for_kupdate || list_empty(&wb->b_io))
555 queue_io(wb, wbc->older_than_this);
556
557 while (!list_empty(&wb->b_io)) {
558 struct inode *inode = wb_inode(wb->b_io.prev);
559 struct super_block *sb = inode->i_sb;
560
561 if (!pin_sb_for_writeback(sb)) {
562 requeue_io(inode);
563 continue;
564 }
565 ret = writeback_sb_inodes(sb, wb, wbc, false);
566 drop_super(sb);
567
568 if (ret)
569 break;
570 }
571 spin_unlock(&inode_lock);
572 /* Leave any unwritten inodes on b_io */
573 }
574
575 static void __writeback_inodes_sb(struct super_block *sb,
576 struct bdi_writeback *wb, struct writeback_control *wbc)
577 {
578 WARN_ON(!rwsem_is_locked(&sb->s_umount));
579
580 spin_lock(&inode_lock);
581 if (!wbc->for_kupdate || list_empty(&wb->b_io))
582 queue_io(wb, wbc->older_than_this);
583 writeback_sb_inodes(sb, wb, wbc, true);
584 spin_unlock(&inode_lock);
585 }
586
587 /*
588 * The maximum number of pages to writeout in a single bdi flush/kupdate
589 * operation. We do this so we don't hold I_SYNC against an inode for
590 * enormous amounts of time, which would block a userspace task which has
591 * been forced to throttle against that inode. Also, the code reevaluates
592 * the dirty each time it has written this many pages.
593 */
594 #define MAX_WRITEBACK_PAGES 1024
595
596 static inline bool over_bground_thresh(void)
597 {
598 unsigned long background_thresh, dirty_thresh;
599
600 global_dirty_limits(&background_thresh, &dirty_thresh);
601
602 return (global_page_state(NR_FILE_DIRTY) +
603 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
604 }
605
606 /*
607 * Explicit flushing or periodic writeback of "old" data.
608 *
609 * Define "old": the first time one of an inode's pages is dirtied, we mark the
610 * dirtying-time in the inode's address_space. So this periodic writeback code
611 * just walks the superblock inode list, writing back any inodes which are
612 * older than a specific point in time.
613 *
614 * Try to run once per dirty_writeback_interval. But if a writeback event
615 * takes longer than a dirty_writeback_interval interval, then leave a
616 * one-second gap.
617 *
618 * older_than_this takes precedence over nr_to_write. So we'll only write back
619 * all dirty pages if they are all attached to "old" mappings.
620 */
621 static long wb_writeback(struct bdi_writeback *wb,
622 struct wb_writeback_work *work)
623 {
624 struct writeback_control wbc = {
625 .sync_mode = work->sync_mode,
626 .older_than_this = NULL,
627 .for_kupdate = work->for_kupdate,
628 .for_background = work->for_background,
629 .range_cyclic = work->range_cyclic,
630 };
631 unsigned long oldest_jif;
632 long wrote = 0;
633 long write_chunk;
634 struct inode *inode;
635
636 if (wbc.for_kupdate) {
637 wbc.older_than_this = &oldest_jif;
638 oldest_jif = jiffies -
639 msecs_to_jiffies(dirty_expire_interval * 10);
640 }
641 if (!wbc.range_cyclic) {
642 wbc.range_start = 0;
643 wbc.range_end = LLONG_MAX;
644 }
645
646 /*
647 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
648 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
649 * here avoids calling into writeback_inodes_wb() more than once.
650 *
651 * The intended call sequence for WB_SYNC_ALL writeback is:
652 *
653 * wb_writeback()
654 * __writeback_inodes_sb() <== called only once
655 * write_cache_pages() <== called once for each inode
656 * (quickly) tag currently dirty pages
657 * (maybe slowly) sync all tagged pages
658 */
659 if (wbc.sync_mode == WB_SYNC_NONE)
660 write_chunk = MAX_WRITEBACK_PAGES;
661 else
662 write_chunk = LONG_MAX;
663
664 wbc.wb_start = jiffies; /* livelock avoidance */
665 for (;;) {
666 /*
667 * Stop writeback when nr_pages has been consumed
668 */
669 if (work->nr_pages <= 0)
670 break;
671
672 /*
673 * Background writeout and kupdate-style writeback may
674 * run forever. Stop them if there is other work to do
675 * so that e.g. sync can proceed. They'll be restarted
676 * after the other works are all done.
677 */
678 if ((work->for_background || work->for_kupdate) &&
679 !list_empty(&wb->bdi->work_list))
680 break;
681
682 /*
683 * For background writeout, stop when we are below the
684 * background dirty threshold
685 */
686 if (work->for_background && !over_bground_thresh())
687 break;
688
689 wbc.more_io = 0;
690 wbc.nr_to_write = write_chunk;
691 wbc.pages_skipped = 0;
692
693 trace_wbc_writeback_start(&wbc, wb->bdi);
694 if (work->sb)
695 __writeback_inodes_sb(work->sb, wb, &wbc);
696 else
697 writeback_inodes_wb(wb, &wbc);
698 trace_wbc_writeback_written(&wbc, wb->bdi);
699
700 work->nr_pages -= write_chunk - wbc.nr_to_write;
701 wrote += write_chunk - wbc.nr_to_write;
702
703 /*
704 * If we consumed everything, see if we have more
705 */
706 if (wbc.nr_to_write <= 0)
707 continue;
708 /*
709 * Didn't write everything and we don't have more IO, bail
710 */
711 if (!wbc.more_io)
712 break;
713 /*
714 * Did we write something? Try for more
715 */
716 if (wbc.nr_to_write < write_chunk)
717 continue;
718 /*
719 * Nothing written. Wait for some inode to
720 * become available for writeback. Otherwise
721 * we'll just busyloop.
722 */
723 spin_lock(&inode_lock);
724 if (!list_empty(&wb->b_more_io)) {
725 inode = wb_inode(wb->b_more_io.prev);
726 trace_wbc_writeback_wait(&wbc, wb->bdi);
727 inode_wait_for_writeback(inode);
728 }
729 spin_unlock(&inode_lock);
730 }
731
732 return wrote;
733 }
734
735 /*
736 * Return the next wb_writeback_work struct that hasn't been processed yet.
737 */
738 static struct wb_writeback_work *
739 get_next_work_item(struct backing_dev_info *bdi)
740 {
741 struct wb_writeback_work *work = NULL;
742
743 spin_lock_bh(&bdi->wb_lock);
744 if (!list_empty(&bdi->work_list)) {
745 work = list_entry(bdi->work_list.next,
746 struct wb_writeback_work, list);
747 list_del_init(&work->list);
748 }
749 spin_unlock_bh(&bdi->wb_lock);
750 return work;
751 }
752
753 /*
754 * Add in the number of potentially dirty inodes, because each inode
755 * write can dirty pagecache in the underlying blockdev.
756 */
757 static unsigned long get_nr_dirty_pages(void)
758 {
759 return global_page_state(NR_FILE_DIRTY) +
760 global_page_state(NR_UNSTABLE_NFS) +
761 get_nr_dirty_inodes();
762 }
763
764 static long wb_check_background_flush(struct bdi_writeback *wb)
765 {
766 if (over_bground_thresh()) {
767
768 struct wb_writeback_work work = {
769 .nr_pages = LONG_MAX,
770 .sync_mode = WB_SYNC_NONE,
771 .for_background = 1,
772 .range_cyclic = 1,
773 };
774
775 return wb_writeback(wb, &work);
776 }
777
778 return 0;
779 }
780
781 static long wb_check_old_data_flush(struct bdi_writeback *wb)
782 {
783 unsigned long expired;
784 long nr_pages;
785
786 /*
787 * When set to zero, disable periodic writeback
788 */
789 if (!dirty_writeback_interval)
790 return 0;
791
792 expired = wb->last_old_flush +
793 msecs_to_jiffies(dirty_writeback_interval * 10);
794 if (time_before(jiffies, expired))
795 return 0;
796
797 wb->last_old_flush = jiffies;
798 nr_pages = get_nr_dirty_pages();
799
800 if (nr_pages) {
801 struct wb_writeback_work work = {
802 .nr_pages = nr_pages,
803 .sync_mode = WB_SYNC_NONE,
804 .for_kupdate = 1,
805 .range_cyclic = 1,
806 };
807
808 return wb_writeback(wb, &work);
809 }
810
811 return 0;
812 }
813
814 /*
815 * Retrieve work items and do the writeback they describe
816 */
817 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
818 {
819 struct backing_dev_info *bdi = wb->bdi;
820 struct wb_writeback_work *work;
821 long wrote = 0;
822
823 set_bit(BDI_writeback_running, &wb->bdi->state);
824 while ((work = get_next_work_item(bdi)) != NULL) {
825 /*
826 * Override sync mode, in case we must wait for completion
827 * because this thread is exiting now.
828 */
829 if (force_wait)
830 work->sync_mode = WB_SYNC_ALL;
831
832 trace_writeback_exec(bdi, work);
833
834 wrote += wb_writeback(wb, work);
835
836 /*
837 * Notify the caller of completion if this is a synchronous
838 * work item, otherwise just free it.
839 */
840 if (work->done)
841 complete(work->done);
842 else
843 kfree(work);
844 }
845
846 /*
847 * Check for periodic writeback, kupdated() style
848 */
849 wrote += wb_check_old_data_flush(wb);
850 wrote += wb_check_background_flush(wb);
851 clear_bit(BDI_writeback_running, &wb->bdi->state);
852
853 return wrote;
854 }
855
856 /*
857 * Handle writeback of dirty data for the device backed by this bdi. Also
858 * wakes up periodically and does kupdated style flushing.
859 */
860 int bdi_writeback_thread(void *data)
861 {
862 struct bdi_writeback *wb = data;
863 struct backing_dev_info *bdi = wb->bdi;
864 long pages_written;
865
866 current->flags |= PF_SWAPWRITE;
867 set_freezable();
868 wb->last_active = jiffies;
869
870 /*
871 * Our parent may run at a different priority, just set us to normal
872 */
873 set_user_nice(current, 0);
874
875 trace_writeback_thread_start(bdi);
876
877 while (!kthread_should_stop()) {
878 /*
879 * Remove own delayed wake-up timer, since we are already awake
880 * and we'll take care of the preriodic write-back.
881 */
882 del_timer(&wb->wakeup_timer);
883
884 pages_written = wb_do_writeback(wb, 0);
885
886 trace_writeback_pages_written(pages_written);
887
888 if (pages_written)
889 wb->last_active = jiffies;
890
891 set_current_state(TASK_INTERRUPTIBLE);
892 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
893 __set_current_state(TASK_RUNNING);
894 continue;
895 }
896
897 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
898 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
899 else {
900 /*
901 * We have nothing to do, so can go sleep without any
902 * timeout and save power. When a work is queued or
903 * something is made dirty - we will be woken up.
904 */
905 schedule();
906 }
907
908 try_to_freeze();
909 }
910
911 /* Flush any work that raced with us exiting */
912 if (!list_empty(&bdi->work_list))
913 wb_do_writeback(wb, 1);
914
915 trace_writeback_thread_stop(bdi);
916 return 0;
917 }
918
919
920 /*
921 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
922 * the whole world.
923 */
924 void wakeup_flusher_threads(long nr_pages)
925 {
926 struct backing_dev_info *bdi;
927
928 if (!nr_pages) {
929 nr_pages = global_page_state(NR_FILE_DIRTY) +
930 global_page_state(NR_UNSTABLE_NFS);
931 }
932
933 rcu_read_lock();
934 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
935 if (!bdi_has_dirty_io(bdi))
936 continue;
937 __bdi_start_writeback(bdi, nr_pages, false);
938 }
939 rcu_read_unlock();
940 }
941
942 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
943 {
944 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
945 struct dentry *dentry;
946 const char *name = "?";
947
948 dentry = d_find_alias(inode);
949 if (dentry) {
950 spin_lock(&dentry->d_lock);
951 name = (const char *) dentry->d_name.name;
952 }
953 printk(KERN_DEBUG
954 "%s(%d): dirtied inode %lu (%s) on %s\n",
955 current->comm, task_pid_nr(current), inode->i_ino,
956 name, inode->i_sb->s_id);
957 if (dentry) {
958 spin_unlock(&dentry->d_lock);
959 dput(dentry);
960 }
961 }
962 }
963
964 /**
965 * __mark_inode_dirty - internal function
966 * @inode: inode to mark
967 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
968 * Mark an inode as dirty. Callers should use mark_inode_dirty or
969 * mark_inode_dirty_sync.
970 *
971 * Put the inode on the super block's dirty list.
972 *
973 * CAREFUL! We mark it dirty unconditionally, but move it onto the
974 * dirty list only if it is hashed or if it refers to a blockdev.
975 * If it was not hashed, it will never be added to the dirty list
976 * even if it is later hashed, as it will have been marked dirty already.
977 *
978 * In short, make sure you hash any inodes _before_ you start marking
979 * them dirty.
980 *
981 * This function *must* be atomic for the I_DIRTY_PAGES case -
982 * set_page_dirty() is called under spinlock in several places.
983 *
984 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
985 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
986 * the kernel-internal blockdev inode represents the dirtying time of the
987 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
988 * page->mapping->host, so the page-dirtying time is recorded in the internal
989 * blockdev inode.
990 */
991 void __mark_inode_dirty(struct inode *inode, int flags)
992 {
993 struct super_block *sb = inode->i_sb;
994 struct backing_dev_info *bdi = NULL;
995 bool wakeup_bdi = false;
996
997 /*
998 * Don't do this for I_DIRTY_PAGES - that doesn't actually
999 * dirty the inode itself
1000 */
1001 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1002 if (sb->s_op->dirty_inode)
1003 sb->s_op->dirty_inode(inode);
1004 }
1005
1006 /*
1007 * make sure that changes are seen by all cpus before we test i_state
1008 * -- mikulas
1009 */
1010 smp_mb();
1011
1012 /* avoid the locking if we can */
1013 if ((inode->i_state & flags) == flags)
1014 return;
1015
1016 if (unlikely(block_dump))
1017 block_dump___mark_inode_dirty(inode);
1018
1019 spin_lock(&inode_lock);
1020 if ((inode->i_state & flags) != flags) {
1021 const int was_dirty = inode->i_state & I_DIRTY;
1022
1023 inode->i_state |= flags;
1024
1025 /*
1026 * If the inode is being synced, just update its dirty state.
1027 * The unlocker will place the inode on the appropriate
1028 * superblock list, based upon its state.
1029 */
1030 if (inode->i_state & I_SYNC)
1031 goto out;
1032
1033 /*
1034 * Only add valid (hashed) inodes to the superblock's
1035 * dirty list. Add blockdev inodes as well.
1036 */
1037 if (!S_ISBLK(inode->i_mode)) {
1038 if (inode_unhashed(inode))
1039 goto out;
1040 }
1041 if (inode->i_state & I_FREEING)
1042 goto out;
1043
1044 /*
1045 * If the inode was already on b_dirty/b_io/b_more_io, don't
1046 * reposition it (that would break b_dirty time-ordering).
1047 */
1048 if (!was_dirty) {
1049 bdi = inode_to_bdi(inode);
1050
1051 if (bdi_cap_writeback_dirty(bdi)) {
1052 WARN(!test_bit(BDI_registered, &bdi->state),
1053 "bdi-%s not registered\n", bdi->name);
1054
1055 /*
1056 * If this is the first dirty inode for this
1057 * bdi, we have to wake-up the corresponding
1058 * bdi thread to make sure background
1059 * write-back happens later.
1060 */
1061 if (!wb_has_dirty_io(&bdi->wb))
1062 wakeup_bdi = true;
1063 }
1064
1065 inode->dirtied_when = jiffies;
1066 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1067 }
1068 }
1069 out:
1070 spin_unlock(&inode_lock);
1071
1072 if (wakeup_bdi)
1073 bdi_wakeup_thread_delayed(bdi);
1074 }
1075 EXPORT_SYMBOL(__mark_inode_dirty);
1076
1077 /*
1078 * Write out a superblock's list of dirty inodes. A wait will be performed
1079 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1080 *
1081 * If older_than_this is non-NULL, then only write out inodes which
1082 * had their first dirtying at a time earlier than *older_than_this.
1083 *
1084 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1085 * This function assumes that the blockdev superblock's inodes are backed by
1086 * a variety of queues, so all inodes are searched. For other superblocks,
1087 * assume that all inodes are backed by the same queue.
1088 *
1089 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1090 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1091 * on the writer throttling path, and we get decent balancing between many
1092 * throttled threads: we don't want them all piling up on inode_sync_wait.
1093 */
1094 static void wait_sb_inodes(struct super_block *sb)
1095 {
1096 struct inode *inode, *old_inode = NULL;
1097
1098 /*
1099 * We need to be protected against the filesystem going from
1100 * r/o to r/w or vice versa.
1101 */
1102 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1103
1104 spin_lock(&inode_lock);
1105
1106 /*
1107 * Data integrity sync. Must wait for all pages under writeback,
1108 * because there may have been pages dirtied before our sync
1109 * call, but which had writeout started before we write it out.
1110 * In which case, the inode may not be on the dirty list, but
1111 * we still have to wait for that writeout.
1112 */
1113 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1114 struct address_space *mapping;
1115
1116 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW))
1117 continue;
1118 mapping = inode->i_mapping;
1119 if (mapping->nrpages == 0)
1120 continue;
1121 __iget(inode);
1122 spin_unlock(&inode_lock);
1123 /*
1124 * We hold a reference to 'inode' so it couldn't have
1125 * been removed from s_inodes list while we dropped the
1126 * inode_lock. We cannot iput the inode now as we can
1127 * be holding the last reference and we cannot iput it
1128 * under inode_lock. So we keep the reference and iput
1129 * it later.
1130 */
1131 iput(old_inode);
1132 old_inode = inode;
1133
1134 filemap_fdatawait(mapping);
1135
1136 cond_resched();
1137
1138 spin_lock(&inode_lock);
1139 }
1140 spin_unlock(&inode_lock);
1141 iput(old_inode);
1142 }
1143
1144 /**
1145 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1146 * @sb: the superblock
1147 * @nr: the number of pages to write
1148 *
1149 * Start writeback on some inodes on this super_block. No guarantees are made
1150 * on how many (if any) will be written, and this function does not wait
1151 * for IO completion of submitted IO.
1152 */
1153 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1154 {
1155 DECLARE_COMPLETION_ONSTACK(done);
1156 struct wb_writeback_work work = {
1157 .sb = sb,
1158 .sync_mode = WB_SYNC_NONE,
1159 .done = &done,
1160 .nr_pages = nr,
1161 };
1162
1163 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1164 bdi_queue_work(sb->s_bdi, &work);
1165 wait_for_completion(&done);
1166 }
1167 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1168
1169 /**
1170 * writeback_inodes_sb - writeback dirty inodes from given super_block
1171 * @sb: the superblock
1172 *
1173 * Start writeback on some inodes on this super_block. No guarantees are made
1174 * on how many (if any) will be written, and this function does not wait
1175 * for IO completion of submitted IO.
1176 */
1177 void writeback_inodes_sb(struct super_block *sb)
1178 {
1179 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1180 }
1181 EXPORT_SYMBOL(writeback_inodes_sb);
1182
1183 /**
1184 * writeback_inodes_sb_if_idle - start writeback if none underway
1185 * @sb: the superblock
1186 *
1187 * Invoke writeback_inodes_sb if no writeback is currently underway.
1188 * Returns 1 if writeback was started, 0 if not.
1189 */
1190 int writeback_inodes_sb_if_idle(struct super_block *sb)
1191 {
1192 if (!writeback_in_progress(sb->s_bdi)) {
1193 down_read(&sb->s_umount);
1194 writeback_inodes_sb(sb);
1195 up_read(&sb->s_umount);
1196 return 1;
1197 } else
1198 return 0;
1199 }
1200 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1201
1202 /**
1203 * writeback_inodes_sb_if_idle - start writeback if none underway
1204 * @sb: the superblock
1205 * @nr: the number of pages to write
1206 *
1207 * Invoke writeback_inodes_sb if no writeback is currently underway.
1208 * Returns 1 if writeback was started, 0 if not.
1209 */
1210 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1211 unsigned long nr)
1212 {
1213 if (!writeback_in_progress(sb->s_bdi)) {
1214 down_read(&sb->s_umount);
1215 writeback_inodes_sb_nr(sb, nr);
1216 up_read(&sb->s_umount);
1217 return 1;
1218 } else
1219 return 0;
1220 }
1221 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1222
1223 /**
1224 * sync_inodes_sb - sync sb inode pages
1225 * @sb: the superblock
1226 *
1227 * This function writes and waits on any dirty inode belonging to this
1228 * super_block. The number of pages synced is returned.
1229 */
1230 void sync_inodes_sb(struct super_block *sb)
1231 {
1232 DECLARE_COMPLETION_ONSTACK(done);
1233 struct wb_writeback_work work = {
1234 .sb = sb,
1235 .sync_mode = WB_SYNC_ALL,
1236 .nr_pages = LONG_MAX,
1237 .range_cyclic = 0,
1238 .done = &done,
1239 };
1240
1241 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1242
1243 bdi_queue_work(sb->s_bdi, &work);
1244 wait_for_completion(&done);
1245
1246 wait_sb_inodes(sb);
1247 }
1248 EXPORT_SYMBOL(sync_inodes_sb);
1249
1250 /**
1251 * write_inode_now - write an inode to disk
1252 * @inode: inode to write to disk
1253 * @sync: whether the write should be synchronous or not
1254 *
1255 * This function commits an inode to disk immediately if it is dirty. This is
1256 * primarily needed by knfsd.
1257 *
1258 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1259 */
1260 int write_inode_now(struct inode *inode, int sync)
1261 {
1262 int ret;
1263 struct writeback_control wbc = {
1264 .nr_to_write = LONG_MAX,
1265 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1266 .range_start = 0,
1267 .range_end = LLONG_MAX,
1268 };
1269
1270 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1271 wbc.nr_to_write = 0;
1272
1273 might_sleep();
1274 spin_lock(&inode_lock);
1275 ret = writeback_single_inode(inode, &wbc);
1276 spin_unlock(&inode_lock);
1277 if (sync)
1278 inode_sync_wait(inode);
1279 return ret;
1280 }
1281 EXPORT_SYMBOL(write_inode_now);
1282
1283 /**
1284 * sync_inode - write an inode and its pages to disk.
1285 * @inode: the inode to sync
1286 * @wbc: controls the writeback mode
1287 *
1288 * sync_inode() will write an inode and its pages to disk. It will also
1289 * correctly update the inode on its superblock's dirty inode lists and will
1290 * update inode->i_state.
1291 *
1292 * The caller must have a ref on the inode.
1293 */
1294 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1295 {
1296 int ret;
1297
1298 spin_lock(&inode_lock);
1299 ret = writeback_single_inode(inode, wbc);
1300 spin_unlock(&inode_lock);
1301 return ret;
1302 }
1303 EXPORT_SYMBOL(sync_inode);
1304
1305 /**
1306 * sync_inode - write an inode to disk
1307 * @inode: the inode to sync
1308 * @wait: wait for I/O to complete.
1309 *
1310 * Write an inode to disk and adjust it's dirty state after completion.
1311 *
1312 * Note: only writes the actual inode, no associated data or other metadata.
1313 */
1314 int sync_inode_metadata(struct inode *inode, int wait)
1315 {
1316 struct writeback_control wbc = {
1317 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1318 .nr_to_write = 0, /* metadata-only */
1319 };
1320
1321 return sync_inode(inode, &wbc);
1322 }
1323 EXPORT_SYMBOL(sync_inode_metadata);