xref: /openbmc/linux/fs/fs-writeback.c (revision 9d56dd3b083a3bec56e9da35ce07baca81030b03)
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  * @sb: write inodes from this super_block
246  * @nr_pages: the number of pages to write
247  *
248  * Description:
249  *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
250  *   started when this function returns, we make no guarentees on
251  *   completion. Caller need not hold sb s_umount semaphore.
252  *
253  */
254 void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
255 			 long nr_pages)
256 {
257 	struct wb_writeback_args args = {
258 		.sb		= sb,
259 		.sync_mode	= WB_SYNC_NONE,
260 		.nr_pages	= nr_pages,
261 		.range_cyclic	= 1,
262 	};
263 
264 	/*
265 	 * We treat @nr_pages=0 as the special case to do background writeback,
266 	 * ie. to sync pages until the background dirty threshold is reached.
267 	 */
268 	if (!nr_pages) {
269 		args.nr_pages = LONG_MAX;
270 		args.for_background = 1;
271 	}
272 
273 	bdi_alloc_queue_work(bdi, &args);
274 }
275 
276 /*
277  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
278  * furthest end of its superblock's dirty-inode list.
279  *
280  * Before stamping the inode's ->dirtied_when, we check to see whether it is
281  * already the most-recently-dirtied inode on the b_dirty list.  If that is
282  * the case then the inode must have been redirtied while it was being written
283  * out and we don't reset its dirtied_when.
284  */
285 static void redirty_tail(struct inode *inode)
286 {
287 	struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
288 
289 	if (!list_empty(&wb->b_dirty)) {
290 		struct inode *tail;
291 
292 		tail = list_entry(wb->b_dirty.next, struct inode, i_list);
293 		if (time_before(inode->dirtied_when, tail->dirtied_when))
294 			inode->dirtied_when = jiffies;
295 	}
296 	list_move(&inode->i_list, &wb->b_dirty);
297 }
298 
299 /*
300  * requeue inode for re-scanning after bdi->b_io list is exhausted.
301  */
302 static void requeue_io(struct inode *inode)
303 {
304 	struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
305 
306 	list_move(&inode->i_list, &wb->b_more_io);
307 }
308 
309 static void inode_sync_complete(struct inode *inode)
310 {
311 	/*
312 	 * Prevent speculative execution through spin_unlock(&inode_lock);
313 	 */
314 	smp_mb();
315 	wake_up_bit(&inode->i_state, __I_SYNC);
316 }
317 
318 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
319 {
320 	bool ret = time_after(inode->dirtied_when, t);
321 #ifndef CONFIG_64BIT
322 	/*
323 	 * For inodes being constantly redirtied, dirtied_when can get stuck.
324 	 * It _appears_ to be in the future, but is actually in distant past.
325 	 * This test is necessary to prevent such wrapped-around relative times
326 	 * from permanently stopping the whole bdi writeback.
327 	 */
328 	ret = ret && time_before_eq(inode->dirtied_when, jiffies);
329 #endif
330 	return ret;
331 }
332 
333 /*
334  * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
335  */
336 static void move_expired_inodes(struct list_head *delaying_queue,
337 			       struct list_head *dispatch_queue,
338 				unsigned long *older_than_this)
339 {
340 	LIST_HEAD(tmp);
341 	struct list_head *pos, *node;
342 	struct super_block *sb = NULL;
343 	struct inode *inode;
344 	int do_sb_sort = 0;
345 
346 	while (!list_empty(delaying_queue)) {
347 		inode = list_entry(delaying_queue->prev, struct inode, i_list);
348 		if (older_than_this &&
349 		    inode_dirtied_after(inode, *older_than_this))
350 			break;
351 		if (sb && sb != inode->i_sb)
352 			do_sb_sort = 1;
353 		sb = inode->i_sb;
354 		list_move(&inode->i_list, &tmp);
355 	}
356 
357 	/* just one sb in list, splice to dispatch_queue and we're done */
358 	if (!do_sb_sort) {
359 		list_splice(&tmp, dispatch_queue);
360 		return;
361 	}
362 
363 	/* Move inodes from one superblock together */
364 	while (!list_empty(&tmp)) {
365 		inode = list_entry(tmp.prev, struct inode, i_list);
366 		sb = inode->i_sb;
367 		list_for_each_prev_safe(pos, node, &tmp) {
368 			inode = list_entry(pos, struct inode, i_list);
369 			if (inode->i_sb == sb)
370 				list_move(&inode->i_list, dispatch_queue);
371 		}
372 	}
373 }
374 
375 /*
376  * Queue all expired dirty inodes for io, eldest first.
377  */
378 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
379 {
380 	list_splice_init(&wb->b_more_io, wb->b_io.prev);
381 	move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
382 }
383 
384 static int write_inode(struct inode *inode, int sync)
385 {
386 	if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
387 		return inode->i_sb->s_op->write_inode(inode, sync);
388 	return 0;
389 }
390 
391 /*
392  * Wait for writeback on an inode to complete.
393  */
394 static void inode_wait_for_writeback(struct inode *inode)
395 {
396 	DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
397 	wait_queue_head_t *wqh;
398 
399 	wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
400 	do {
401 		spin_unlock(&inode_lock);
402 		__wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
403 		spin_lock(&inode_lock);
404 	} while (inode->i_state & I_SYNC);
405 }
406 
407 /*
408  * Write out an inode's dirty pages.  Called under inode_lock.  Either the
409  * caller has ref on the inode (either via __iget or via syscall against an fd)
410  * or the inode has I_WILL_FREE set (via generic_forget_inode)
411  *
412  * If `wait' is set, wait on the writeout.
413  *
414  * The whole writeout design is quite complex and fragile.  We want to avoid
415  * starvation of particular inodes when others are being redirtied, prevent
416  * livelocks, etc.
417  *
418  * Called under inode_lock.
419  */
420 static int
421 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
422 {
423 	struct address_space *mapping = inode->i_mapping;
424 	int wait = wbc->sync_mode == WB_SYNC_ALL;
425 	unsigned dirty;
426 	int ret;
427 
428 	if (!atomic_read(&inode->i_count))
429 		WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
430 	else
431 		WARN_ON(inode->i_state & I_WILL_FREE);
432 
433 	if (inode->i_state & I_SYNC) {
434 		/*
435 		 * If this inode is locked for writeback and we are not doing
436 		 * writeback-for-data-integrity, move it to b_more_io so that
437 		 * writeback can proceed with the other inodes on s_io.
438 		 *
439 		 * We'll have another go at writing back this inode when we
440 		 * completed a full scan of b_io.
441 		 */
442 		if (!wait) {
443 			requeue_io(inode);
444 			return 0;
445 		}
446 
447 		/*
448 		 * It's a data-integrity sync.  We must wait.
449 		 */
450 		inode_wait_for_writeback(inode);
451 	}
452 
453 	BUG_ON(inode->i_state & I_SYNC);
454 
455 	/* Set I_SYNC, reset I_DIRTY */
456 	dirty = inode->i_state & I_DIRTY;
457 	inode->i_state |= I_SYNC;
458 	inode->i_state &= ~I_DIRTY;
459 
460 	spin_unlock(&inode_lock);
461 
462 	ret = do_writepages(mapping, wbc);
463 
464 	/* Don't write the inode if only I_DIRTY_PAGES was set */
465 	if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
466 		int err = write_inode(inode, wait);
467 		if (ret == 0)
468 			ret = err;
469 	}
470 
471 	if (wait) {
472 		int err = filemap_fdatawait(mapping);
473 		if (ret == 0)
474 			ret = err;
475 	}
476 
477 	spin_lock(&inode_lock);
478 	inode->i_state &= ~I_SYNC;
479 	if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
480 		if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
481 			/*
482 			 * More pages get dirtied by a fast dirtier.
483 			 */
484 			goto select_queue;
485 		} else if (inode->i_state & I_DIRTY) {
486 			/*
487 			 * At least XFS will redirty the inode during the
488 			 * writeback (delalloc) and on io completion (isize).
489 			 */
490 			redirty_tail(inode);
491 		} else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
492 			/*
493 			 * We didn't write back all the pages.  nfs_writepages()
494 			 * sometimes bales out without doing anything. Redirty
495 			 * the inode; Move it from b_io onto b_more_io/b_dirty.
496 			 */
497 			/*
498 			 * akpm: if the caller was the kupdate function we put
499 			 * this inode at the head of b_dirty so it gets first
500 			 * consideration.  Otherwise, move it to the tail, for
501 			 * the reasons described there.  I'm not really sure
502 			 * how much sense this makes.  Presumably I had a good
503 			 * reasons for doing it this way, and I'd rather not
504 			 * muck with it at present.
505 			 */
506 			if (wbc->for_kupdate) {
507 				/*
508 				 * For the kupdate function we move the inode
509 				 * to b_more_io so it will get more writeout as
510 				 * soon as the queue becomes uncongested.
511 				 */
512 				inode->i_state |= I_DIRTY_PAGES;
513 select_queue:
514 				if (wbc->nr_to_write <= 0) {
515 					/*
516 					 * slice used up: queue for next turn
517 					 */
518 					requeue_io(inode);
519 				} else {
520 					/*
521 					 * somehow blocked: retry later
522 					 */
523 					redirty_tail(inode);
524 				}
525 			} else {
526 				/*
527 				 * Otherwise fully redirty the inode so that
528 				 * other inodes on this superblock will get some
529 				 * writeout.  Otherwise heavy writing to one
530 				 * file would indefinitely suspend writeout of
531 				 * all the other files.
532 				 */
533 				inode->i_state |= I_DIRTY_PAGES;
534 				redirty_tail(inode);
535 			}
536 		} else if (atomic_read(&inode->i_count)) {
537 			/*
538 			 * The inode is clean, inuse
539 			 */
540 			list_move(&inode->i_list, &inode_in_use);
541 		} else {
542 			/*
543 			 * The inode is clean, unused
544 			 */
545 			list_move(&inode->i_list, &inode_unused);
546 		}
547 	}
548 	inode_sync_complete(inode);
549 	return ret;
550 }
551 
552 static void unpin_sb_for_writeback(struct super_block **psb)
553 {
554 	struct super_block *sb = *psb;
555 
556 	if (sb) {
557 		up_read(&sb->s_umount);
558 		put_super(sb);
559 		*psb = NULL;
560 	}
561 }
562 
563 /*
564  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
565  * before calling writeback. So make sure that we do pin it, so it doesn't
566  * go away while we are writing inodes from it.
567  *
568  * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
569  * 1 if we failed.
570  */
571 static int pin_sb_for_writeback(struct writeback_control *wbc,
572 				struct inode *inode, struct super_block **psb)
573 {
574 	struct super_block *sb = inode->i_sb;
575 
576 	/*
577 	 * If this sb is already pinned, nothing more to do. If not and
578 	 * *psb is non-NULL, unpin the old one first
579 	 */
580 	if (sb == *psb)
581 		return 0;
582 	else if (*psb)
583 		unpin_sb_for_writeback(psb);
584 
585 	/*
586 	 * Caller must already hold the ref for this
587 	 */
588 	if (wbc->sync_mode == WB_SYNC_ALL) {
589 		WARN_ON(!rwsem_is_locked(&sb->s_umount));
590 		return 0;
591 	}
592 
593 	spin_lock(&sb_lock);
594 	sb->s_count++;
595 	if (down_read_trylock(&sb->s_umount)) {
596 		if (sb->s_root) {
597 			spin_unlock(&sb_lock);
598 			goto pinned;
599 		}
600 		/*
601 		 * umounted, drop rwsem again and fall through to failure
602 		 */
603 		up_read(&sb->s_umount);
604 	}
605 
606 	sb->s_count--;
607 	spin_unlock(&sb_lock);
608 	return 1;
609 pinned:
610 	*psb = sb;
611 	return 0;
612 }
613 
614 static void writeback_inodes_wb(struct bdi_writeback *wb,
615 				struct writeback_control *wbc)
616 {
617 	struct super_block *sb = wbc->sb, *pin_sb = NULL;
618 	const unsigned long start = jiffies;	/* livelock avoidance */
619 
620 	spin_lock(&inode_lock);
621 
622 	if (!wbc->for_kupdate || list_empty(&wb->b_io))
623 		queue_io(wb, wbc->older_than_this);
624 
625 	while (!list_empty(&wb->b_io)) {
626 		struct inode *inode = list_entry(wb->b_io.prev,
627 						struct inode, i_list);
628 		long pages_skipped;
629 
630 		/*
631 		 * super block given and doesn't match, skip this inode
632 		 */
633 		if (sb && sb != inode->i_sb) {
634 			redirty_tail(inode);
635 			continue;
636 		}
637 
638 		if (inode->i_state & (I_NEW | I_WILL_FREE)) {
639 			requeue_io(inode);
640 			continue;
641 		}
642 
643 		/*
644 		 * Was this inode dirtied after sync_sb_inodes was called?
645 		 * This keeps sync from extra jobs and livelock.
646 		 */
647 		if (inode_dirtied_after(inode, start))
648 			break;
649 
650 		if (pin_sb_for_writeback(wbc, inode, &pin_sb)) {
651 			requeue_io(inode);
652 			continue;
653 		}
654 
655 		BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
656 		__iget(inode);
657 		pages_skipped = wbc->pages_skipped;
658 		writeback_single_inode(inode, wbc);
659 		if (wbc->pages_skipped != pages_skipped) {
660 			/*
661 			 * writeback is not making progress due to locked
662 			 * buffers.  Skip this inode for now.
663 			 */
664 			redirty_tail(inode);
665 		}
666 		spin_unlock(&inode_lock);
667 		iput(inode);
668 		cond_resched();
669 		spin_lock(&inode_lock);
670 		if (wbc->nr_to_write <= 0) {
671 			wbc->more_io = 1;
672 			break;
673 		}
674 		if (!list_empty(&wb->b_more_io))
675 			wbc->more_io = 1;
676 	}
677 
678 	unpin_sb_for_writeback(&pin_sb);
679 
680 	spin_unlock(&inode_lock);
681 	/* Leave any unwritten inodes on b_io */
682 }
683 
684 void writeback_inodes_wbc(struct writeback_control *wbc)
685 {
686 	struct backing_dev_info *bdi = wbc->bdi;
687 
688 	writeback_inodes_wb(&bdi->wb, wbc);
689 }
690 
691 /*
692  * The maximum number of pages to writeout in a single bdi flush/kupdate
693  * operation.  We do this so we don't hold I_SYNC against an inode for
694  * enormous amounts of time, which would block a userspace task which has
695  * been forced to throttle against that inode.  Also, the code reevaluates
696  * the dirty each time it has written this many pages.
697  */
698 #define MAX_WRITEBACK_PAGES     1024
699 
700 static inline bool over_bground_thresh(void)
701 {
702 	unsigned long background_thresh, dirty_thresh;
703 
704 	get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
705 
706 	return (global_page_state(NR_FILE_DIRTY) +
707 		global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
708 }
709 
710 /*
711  * Explicit flushing or periodic writeback of "old" data.
712  *
713  * Define "old": the first time one of an inode's pages is dirtied, we mark the
714  * dirtying-time in the inode's address_space.  So this periodic writeback code
715  * just walks the superblock inode list, writing back any inodes which are
716  * older than a specific point in time.
717  *
718  * Try to run once per dirty_writeback_interval.  But if a writeback event
719  * takes longer than a dirty_writeback_interval interval, then leave a
720  * one-second gap.
721  *
722  * older_than_this takes precedence over nr_to_write.  So we'll only write back
723  * all dirty pages if they are all attached to "old" mappings.
724  */
725 static long wb_writeback(struct bdi_writeback *wb,
726 			 struct wb_writeback_args *args)
727 {
728 	struct writeback_control wbc = {
729 		.bdi			= wb->bdi,
730 		.sb			= args->sb,
731 		.sync_mode		= args->sync_mode,
732 		.older_than_this	= NULL,
733 		.for_kupdate		= args->for_kupdate,
734 		.for_background		= args->for_background,
735 		.range_cyclic		= args->range_cyclic,
736 	};
737 	unsigned long oldest_jif;
738 	long wrote = 0;
739 	struct inode *inode;
740 
741 	if (wbc.for_kupdate) {
742 		wbc.older_than_this = &oldest_jif;
743 		oldest_jif = jiffies -
744 				msecs_to_jiffies(dirty_expire_interval * 10);
745 	}
746 	if (!wbc.range_cyclic) {
747 		wbc.range_start = 0;
748 		wbc.range_end = LLONG_MAX;
749 	}
750 
751 	for (;;) {
752 		/*
753 		 * Stop writeback when nr_pages has been consumed
754 		 */
755 		if (args->nr_pages <= 0)
756 			break;
757 
758 		/*
759 		 * For background writeout, stop when we are below the
760 		 * background dirty threshold
761 		 */
762 		if (args->for_background && !over_bground_thresh())
763 			break;
764 
765 		wbc.more_io = 0;
766 		wbc.nr_to_write = MAX_WRITEBACK_PAGES;
767 		wbc.pages_skipped = 0;
768 		writeback_inodes_wb(wb, &wbc);
769 		args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
770 		wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
771 
772 		/*
773 		 * If we consumed everything, see if we have more
774 		 */
775 		if (wbc.nr_to_write <= 0)
776 			continue;
777 		/*
778 		 * Didn't write everything and we don't have more IO, bail
779 		 */
780 		if (!wbc.more_io)
781 			break;
782 		/*
783 		 * Did we write something? Try for more
784 		 */
785 		if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
786 			continue;
787 		/*
788 		 * Nothing written. Wait for some inode to
789 		 * become available for writeback. Otherwise
790 		 * we'll just busyloop.
791 		 */
792 		spin_lock(&inode_lock);
793 		if (!list_empty(&wb->b_more_io))  {
794 			inode = list_entry(wb->b_more_io.prev,
795 						struct inode, i_list);
796 			inode_wait_for_writeback(inode);
797 		}
798 		spin_unlock(&inode_lock);
799 	}
800 
801 	return wrote;
802 }
803 
804 /*
805  * Return the next bdi_work struct that hasn't been processed by this
806  * wb thread yet. ->seen is initially set for each thread that exists
807  * for this device, when a thread first notices a piece of work it
808  * clears its bit. Depending on writeback type, the thread will notify
809  * completion on either receiving the work (WB_SYNC_NONE) or after
810  * it is done (WB_SYNC_ALL).
811  */
812 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
813 					   struct bdi_writeback *wb)
814 {
815 	struct bdi_work *work, *ret = NULL;
816 
817 	rcu_read_lock();
818 
819 	list_for_each_entry_rcu(work, &bdi->work_list, list) {
820 		if (!test_bit(wb->nr, &work->seen))
821 			continue;
822 		clear_bit(wb->nr, &work->seen);
823 
824 		ret = work;
825 		break;
826 	}
827 
828 	rcu_read_unlock();
829 	return ret;
830 }
831 
832 static long wb_check_old_data_flush(struct bdi_writeback *wb)
833 {
834 	unsigned long expired;
835 	long nr_pages;
836 
837 	expired = wb->last_old_flush +
838 			msecs_to_jiffies(dirty_writeback_interval * 10);
839 	if (time_before(jiffies, expired))
840 		return 0;
841 
842 	wb->last_old_flush = jiffies;
843 	nr_pages = global_page_state(NR_FILE_DIRTY) +
844 			global_page_state(NR_UNSTABLE_NFS) +
845 			(inodes_stat.nr_inodes - inodes_stat.nr_unused);
846 
847 	if (nr_pages) {
848 		struct wb_writeback_args args = {
849 			.nr_pages	= nr_pages,
850 			.sync_mode	= WB_SYNC_NONE,
851 			.for_kupdate	= 1,
852 			.range_cyclic	= 1,
853 		};
854 
855 		return wb_writeback(wb, &args);
856 	}
857 
858 	return 0;
859 }
860 
861 /*
862  * Retrieve work items and do the writeback they describe
863  */
864 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
865 {
866 	struct backing_dev_info *bdi = wb->bdi;
867 	struct bdi_work *work;
868 	long wrote = 0;
869 
870 	while ((work = get_next_work_item(bdi, wb)) != NULL) {
871 		struct wb_writeback_args args = work->args;
872 
873 		/*
874 		 * Override sync mode, in case we must wait for completion
875 		 */
876 		if (force_wait)
877 			work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
878 
879 		/*
880 		 * If this isn't a data integrity operation, just notify
881 		 * that we have seen this work and we are now starting it.
882 		 */
883 		if (args.sync_mode == WB_SYNC_NONE)
884 			wb_clear_pending(wb, work);
885 
886 		wrote += wb_writeback(wb, &args);
887 
888 		/*
889 		 * This is a data integrity writeback, so only do the
890 		 * notification when we have completed the work.
891 		 */
892 		if (args.sync_mode == WB_SYNC_ALL)
893 			wb_clear_pending(wb, work);
894 	}
895 
896 	/*
897 	 * Check for periodic writeback, kupdated() style
898 	 */
899 	wrote += wb_check_old_data_flush(wb);
900 
901 	return wrote;
902 }
903 
904 /*
905  * Handle writeback of dirty data for the device backed by this bdi. Also
906  * wakes up periodically and does kupdated style flushing.
907  */
908 int bdi_writeback_task(struct bdi_writeback *wb)
909 {
910 	unsigned long last_active = jiffies;
911 	unsigned long wait_jiffies = -1UL;
912 	long pages_written;
913 
914 	while (!kthread_should_stop()) {
915 		pages_written = wb_do_writeback(wb, 0);
916 
917 		if (pages_written)
918 			last_active = jiffies;
919 		else if (wait_jiffies != -1UL) {
920 			unsigned long max_idle;
921 
922 			/*
923 			 * Longest period of inactivity that we tolerate. If we
924 			 * see dirty data again later, the task will get
925 			 * recreated automatically.
926 			 */
927 			max_idle = max(5UL * 60 * HZ, wait_jiffies);
928 			if (time_after(jiffies, max_idle + last_active))
929 				break;
930 		}
931 
932 		wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
933 		schedule_timeout_interruptible(wait_jiffies);
934 		try_to_freeze();
935 	}
936 
937 	return 0;
938 }
939 
940 /*
941  * Schedule writeback for all backing devices. This does WB_SYNC_NONE
942  * writeback, for integrity writeback see bdi_sync_writeback().
943  */
944 static void bdi_writeback_all(struct super_block *sb, long nr_pages)
945 {
946 	struct wb_writeback_args args = {
947 		.sb		= sb,
948 		.nr_pages	= nr_pages,
949 		.sync_mode	= WB_SYNC_NONE,
950 	};
951 	struct backing_dev_info *bdi;
952 
953 	rcu_read_lock();
954 
955 	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
956 		if (!bdi_has_dirty_io(bdi))
957 			continue;
958 
959 		bdi_alloc_queue_work(bdi, &args);
960 	}
961 
962 	rcu_read_unlock();
963 }
964 
965 /*
966  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
967  * the whole world.
968  */
969 void wakeup_flusher_threads(long nr_pages)
970 {
971 	if (nr_pages == 0)
972 		nr_pages = global_page_state(NR_FILE_DIRTY) +
973 				global_page_state(NR_UNSTABLE_NFS);
974 	bdi_writeback_all(NULL, nr_pages);
975 }
976 
977 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
978 {
979 	if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
980 		struct dentry *dentry;
981 		const char *name = "?";
982 
983 		dentry = d_find_alias(inode);
984 		if (dentry) {
985 			spin_lock(&dentry->d_lock);
986 			name = (const char *) dentry->d_name.name;
987 		}
988 		printk(KERN_DEBUG
989 		       "%s(%d): dirtied inode %lu (%s) on %s\n",
990 		       current->comm, task_pid_nr(current), inode->i_ino,
991 		       name, inode->i_sb->s_id);
992 		if (dentry) {
993 			spin_unlock(&dentry->d_lock);
994 			dput(dentry);
995 		}
996 	}
997 }
998 
999 /**
1000  *	__mark_inode_dirty -	internal function
1001  *	@inode: inode to mark
1002  *	@flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1003  *	Mark an inode as dirty. Callers should use mark_inode_dirty or
1004  *  	mark_inode_dirty_sync.
1005  *
1006  * Put the inode on the super block's dirty list.
1007  *
1008  * CAREFUL! We mark it dirty unconditionally, but move it onto the
1009  * dirty list only if it is hashed or if it refers to a blockdev.
1010  * If it was not hashed, it will never be added to the dirty list
1011  * even if it is later hashed, as it will have been marked dirty already.
1012  *
1013  * In short, make sure you hash any inodes _before_ you start marking
1014  * them dirty.
1015  *
1016  * This function *must* be atomic for the I_DIRTY_PAGES case -
1017  * set_page_dirty() is called under spinlock in several places.
1018  *
1019  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1020  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1021  * the kernel-internal blockdev inode represents the dirtying time of the
1022  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1023  * page->mapping->host, so the page-dirtying time is recorded in the internal
1024  * blockdev inode.
1025  */
1026 void __mark_inode_dirty(struct inode *inode, int flags)
1027 {
1028 	struct super_block *sb = inode->i_sb;
1029 
1030 	/*
1031 	 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1032 	 * dirty the inode itself
1033 	 */
1034 	if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1035 		if (sb->s_op->dirty_inode)
1036 			sb->s_op->dirty_inode(inode);
1037 	}
1038 
1039 	/*
1040 	 * make sure that changes are seen by all cpus before we test i_state
1041 	 * -- mikulas
1042 	 */
1043 	smp_mb();
1044 
1045 	/* avoid the locking if we can */
1046 	if ((inode->i_state & flags) == flags)
1047 		return;
1048 
1049 	if (unlikely(block_dump))
1050 		block_dump___mark_inode_dirty(inode);
1051 
1052 	spin_lock(&inode_lock);
1053 	if ((inode->i_state & flags) != flags) {
1054 		const int was_dirty = inode->i_state & I_DIRTY;
1055 
1056 		inode->i_state |= flags;
1057 
1058 		/*
1059 		 * If the inode is being synced, just update its dirty state.
1060 		 * The unlocker will place the inode on the appropriate
1061 		 * superblock list, based upon its state.
1062 		 */
1063 		if (inode->i_state & I_SYNC)
1064 			goto out;
1065 
1066 		/*
1067 		 * Only add valid (hashed) inodes to the superblock's
1068 		 * dirty list.  Add blockdev inodes as well.
1069 		 */
1070 		if (!S_ISBLK(inode->i_mode)) {
1071 			if (hlist_unhashed(&inode->i_hash))
1072 				goto out;
1073 		}
1074 		if (inode->i_state & (I_FREEING|I_CLEAR))
1075 			goto out;
1076 
1077 		/*
1078 		 * If the inode was already on b_dirty/b_io/b_more_io, don't
1079 		 * reposition it (that would break b_dirty time-ordering).
1080 		 */
1081 		if (!was_dirty) {
1082 			struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1083 			struct backing_dev_info *bdi = wb->bdi;
1084 
1085 			if (bdi_cap_writeback_dirty(bdi) &&
1086 			    !test_bit(BDI_registered, &bdi->state)) {
1087 				WARN_ON(1);
1088 				printk(KERN_ERR "bdi-%s not registered\n",
1089 								bdi->name);
1090 			}
1091 
1092 			inode->dirtied_when = jiffies;
1093 			list_move(&inode->i_list, &wb->b_dirty);
1094 		}
1095 	}
1096 out:
1097 	spin_unlock(&inode_lock);
1098 }
1099 EXPORT_SYMBOL(__mark_inode_dirty);
1100 
1101 /*
1102  * Write out a superblock's list of dirty inodes.  A wait will be performed
1103  * upon no inodes, all inodes or the final one, depending upon sync_mode.
1104  *
1105  * If older_than_this is non-NULL, then only write out inodes which
1106  * had their first dirtying at a time earlier than *older_than_this.
1107  *
1108  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1109  * This function assumes that the blockdev superblock's inodes are backed by
1110  * a variety of queues, so all inodes are searched.  For other superblocks,
1111  * assume that all inodes are backed by the same queue.
1112  *
1113  * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1114  * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1115  * on the writer throttling path, and we get decent balancing between many
1116  * throttled threads: we don't want them all piling up on inode_sync_wait.
1117  */
1118 static void wait_sb_inodes(struct super_block *sb)
1119 {
1120 	struct inode *inode, *old_inode = NULL;
1121 
1122 	/*
1123 	 * We need to be protected against the filesystem going from
1124 	 * r/o to r/w or vice versa.
1125 	 */
1126 	WARN_ON(!rwsem_is_locked(&sb->s_umount));
1127 
1128 	spin_lock(&inode_lock);
1129 
1130 	/*
1131 	 * Data integrity sync. Must wait for all pages under writeback,
1132 	 * because there may have been pages dirtied before our sync
1133 	 * call, but which had writeout started before we write it out.
1134 	 * In which case, the inode may not be on the dirty list, but
1135 	 * we still have to wait for that writeout.
1136 	 */
1137 	list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1138 		struct address_space *mapping;
1139 
1140 		if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1141 			continue;
1142 		mapping = inode->i_mapping;
1143 		if (mapping->nrpages == 0)
1144 			continue;
1145 		__iget(inode);
1146 		spin_unlock(&inode_lock);
1147 		/*
1148 		 * We hold a reference to 'inode' so it couldn't have
1149 		 * been removed from s_inodes list while we dropped the
1150 		 * inode_lock.  We cannot iput the inode now as we can
1151 		 * be holding the last reference and we cannot iput it
1152 		 * under inode_lock. So we keep the reference and iput
1153 		 * it later.
1154 		 */
1155 		iput(old_inode);
1156 		old_inode = inode;
1157 
1158 		filemap_fdatawait(mapping);
1159 
1160 		cond_resched();
1161 
1162 		spin_lock(&inode_lock);
1163 	}
1164 	spin_unlock(&inode_lock);
1165 	iput(old_inode);
1166 }
1167 
1168 /**
1169  * writeback_inodes_sb	-	writeback dirty inodes from given super_block
1170  * @sb: the superblock
1171  *
1172  * Start writeback on some inodes on this super_block. No guarantees are made
1173  * on how many (if any) will be written, and this function does not wait
1174  * for IO completion of submitted IO. The number of pages submitted is
1175  * returned.
1176  */
1177 void writeback_inodes_sb(struct super_block *sb)
1178 {
1179 	unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1180 	unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1181 	long nr_to_write;
1182 
1183 	nr_to_write = nr_dirty + nr_unstable +
1184 			(inodes_stat.nr_inodes - inodes_stat.nr_unused);
1185 
1186 	bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
1187 }
1188 EXPORT_SYMBOL(writeback_inodes_sb);
1189 
1190 /**
1191  * writeback_inodes_sb_if_idle	-	start writeback if none underway
1192  * @sb: the superblock
1193  *
1194  * Invoke writeback_inodes_sb if no writeback is currently underway.
1195  * Returns 1 if writeback was started, 0 if not.
1196  */
1197 int writeback_inodes_sb_if_idle(struct super_block *sb)
1198 {
1199 	if (!writeback_in_progress(sb->s_bdi)) {
1200 		writeback_inodes_sb(sb);
1201 		return 1;
1202 	} else
1203 		return 0;
1204 }
1205 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1206 
1207 /**
1208  * sync_inodes_sb	-	sync sb inode pages
1209  * @sb: the superblock
1210  *
1211  * This function writes and waits on any dirty inode belonging to this
1212  * super_block. The number of pages synced is returned.
1213  */
1214 void sync_inodes_sb(struct super_block *sb)
1215 {
1216 	bdi_sync_writeback(sb->s_bdi, sb);
1217 	wait_sb_inodes(sb);
1218 }
1219 EXPORT_SYMBOL(sync_inodes_sb);
1220 
1221 /**
1222  * write_inode_now	-	write an inode to disk
1223  * @inode: inode to write to disk
1224  * @sync: whether the write should be synchronous or not
1225  *
1226  * This function commits an inode to disk immediately if it is dirty. This is
1227  * primarily needed by knfsd.
1228  *
1229  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1230  */
1231 int write_inode_now(struct inode *inode, int sync)
1232 {
1233 	int ret;
1234 	struct writeback_control wbc = {
1235 		.nr_to_write = LONG_MAX,
1236 		.sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1237 		.range_start = 0,
1238 		.range_end = LLONG_MAX,
1239 	};
1240 
1241 	if (!mapping_cap_writeback_dirty(inode->i_mapping))
1242 		wbc.nr_to_write = 0;
1243 
1244 	might_sleep();
1245 	spin_lock(&inode_lock);
1246 	ret = writeback_single_inode(inode, &wbc);
1247 	spin_unlock(&inode_lock);
1248 	if (sync)
1249 		inode_sync_wait(inode);
1250 	return ret;
1251 }
1252 EXPORT_SYMBOL(write_inode_now);
1253 
1254 /**
1255  * sync_inode - write an inode and its pages to disk.
1256  * @inode: the inode to sync
1257  * @wbc: controls the writeback mode
1258  *
1259  * sync_inode() will write an inode and its pages to disk.  It will also
1260  * correctly update the inode on its superblock's dirty inode lists and will
1261  * update inode->i_state.
1262  *
1263  * The caller must have a ref on the inode.
1264  */
1265 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1266 {
1267 	int ret;
1268 
1269 	spin_lock(&inode_lock);
1270 	ret = writeback_single_inode(inode, wbc);
1271 	spin_unlock(&inode_lock);
1272 	return ret;
1273 }
1274 EXPORT_SYMBOL(sync_inode);
1275