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/export.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/pagemap.h>
24#include <linux/kthread.h>
25#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
28#include <linux/tracepoint.h>
29#include <linux/device.h>
30#include <linux/memcontrol.h>
31#include "internal.h"
32
33/*
34 * 4MB minimal write chunk size
35 */
36#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
37
38struct wb_completion {
39 atomic_t cnt;
40};
41
42/*
43 * Passed into wb_writeback(), essentially a subset of writeback_control
44 */
45struct wb_writeback_work {
46 long nr_pages;
47 struct super_block *sb;
48 unsigned long *older_than_this;
49 enum writeback_sync_modes sync_mode;
50 unsigned int tagged_writepages:1;
51 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
54 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
55 unsigned int auto_free:1; /* free on completion */
56 enum wb_reason reason; /* why was writeback initiated? */
57
58 struct list_head list; /* pending work list */
59 struct wb_completion *done; /* set if the caller waits */
60};
61
62/*
63 * If one wants to wait for one or more wb_writeback_works, each work's
64 * ->done should be set to a wb_completion defined using the following
65 * macro. Once all work items are issued with wb_queue_work(), the caller
66 * can wait for the completion of all using wb_wait_for_completion(). Work
67 * items which are waited upon aren't freed automatically on completion.
68 */
69#define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
70 struct wb_completion cmpl = { \
71 .cnt = ATOMIC_INIT(1), \
72 }
73
74
75/*
76 * If an inode is constantly having its pages dirtied, but then the
77 * updates stop dirtytime_expire_interval seconds in the past, it's
78 * possible for the worst case time between when an inode has its
79 * timestamps updated and when they finally get written out to be two
80 * dirtytime_expire_intervals. We set the default to 12 hours (in
81 * seconds), which means most of the time inodes will have their
82 * timestamps written to disk after 12 hours, but in the worst case a
83 * few inodes might not their timestamps updated for 24 hours.
84 */
85unsigned int dirtytime_expire_interval = 12 * 60 * 60;
86
87static inline struct inode *wb_inode(struct list_head *head)
88{
89 return list_entry(head, struct inode, i_io_list);
90}
91
92/*
93 * Include the creation of the trace points after defining the
94 * wb_writeback_work structure and inline functions so that the definition
95 * remains local to this file.
96 */
97#define CREATE_TRACE_POINTS
98#include <trace/events/writeback.h>
99
100EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
101
102static bool wb_io_lists_populated(struct bdi_writeback *wb)
103{
104 if (wb_has_dirty_io(wb)) {
105 return false;
106 } else {
107 set_bit(WB_has_dirty_io, &wb->state);
108 WARN_ON_ONCE(!wb->avg_write_bandwidth);
109 atomic_long_add(wb->avg_write_bandwidth,
110 &wb->bdi->tot_write_bandwidth);
111 return true;
112 }
113}
114
115static void wb_io_lists_depopulated(struct bdi_writeback *wb)
116{
117 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
118 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
119 clear_bit(WB_has_dirty_io, &wb->state);
120 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
121 &wb->bdi->tot_write_bandwidth) < 0);
122 }
123}
124
125/**
126 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
127 * @inode: inode to be moved
128 * @wb: target bdi_writeback
129 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
130 *
131 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
132 * Returns %true if @inode is the first occupant of the !dirty_time IO
133 * lists; otherwise, %false.
134 */
135static bool inode_io_list_move_locked(struct inode *inode,
136 struct bdi_writeback *wb,
137 struct list_head *head)
138{
139 assert_spin_locked(&wb->list_lock);
140
141 list_move(&inode->i_io_list, head);
142
143 /* dirty_time doesn't count as dirty_io until expiration */
144 if (head != &wb->b_dirty_time)
145 return wb_io_lists_populated(wb);
146
147 wb_io_lists_depopulated(wb);
148 return false;
149}
150
151/**
152 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
153 * @inode: inode to be removed
154 * @wb: bdi_writeback @inode is being removed from
155 *
156 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157 * clear %WB_has_dirty_io if all are empty afterwards.
158 */
159static void inode_io_list_del_locked(struct inode *inode,
160 struct bdi_writeback *wb)
161{
162 assert_spin_locked(&wb->list_lock);
163
164 list_del_init(&inode->i_io_list);
165 wb_io_lists_depopulated(wb);
166}
167
168static void wb_wakeup(struct bdi_writeback *wb)
169{
170 spin_lock_bh(&wb->work_lock);
171 if (test_bit(WB_registered, &wb->state))
172 mod_delayed_work(bdi_wq, &wb->dwork, 0);
173 spin_unlock_bh(&wb->work_lock);
174}
175
176static void finish_writeback_work(struct bdi_writeback *wb,
177 struct wb_writeback_work *work)
178{
179 struct wb_completion *done = work->done;
180
181 if (work->auto_free)
182 kfree(work);
183 if (done && atomic_dec_and_test(&done->cnt))
184 wake_up_all(&wb->bdi->wb_waitq);
185}
186
187static void wb_queue_work(struct bdi_writeback *wb,
188 struct wb_writeback_work *work)
189{
190 trace_writeback_queue(wb, work);
191
192 if (work->done)
193 atomic_inc(&work->done->cnt);
194
195 spin_lock_bh(&wb->work_lock);
196
197 if (test_bit(WB_registered, &wb->state)) {
198 list_add_tail(&work->list, &wb->work_list);
199 mod_delayed_work(bdi_wq, &wb->dwork, 0);
200 } else
201 finish_writeback_work(wb, work);
202
203 spin_unlock_bh(&wb->work_lock);
204}
205
206/**
207 * wb_wait_for_completion - wait for completion of bdi_writeback_works
208 * @bdi: bdi work items were issued to
209 * @done: target wb_completion
210 *
211 * Wait for one or more work items issued to @bdi with their ->done field
212 * set to @done, which should have been defined with
213 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
214 * work items are completed. Work items which are waited upon aren't freed
215 * automatically on completion.
216 */
217static void wb_wait_for_completion(struct backing_dev_info *bdi,
218 struct wb_completion *done)
219{
220 atomic_dec(&done->cnt); /* put down the initial count */
221 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
222}
223
224#ifdef CONFIG_CGROUP_WRITEBACK
225
226/* parameters for foreign inode detection, see wb_detach_inode() */
227#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
228#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
229#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
230#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
231
232#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
233#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
234 /* each slot's duration is 2s / 16 */
235#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
236 /* if foreign slots >= 8, switch */
237#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
238 /* one round can affect upto 5 slots */
239
240static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
241static struct workqueue_struct *isw_wq;
242
243void __inode_attach_wb(struct inode *inode, struct page *page)
244{
245 struct backing_dev_info *bdi = inode_to_bdi(inode);
246 struct bdi_writeback *wb = NULL;
247
248 if (inode_cgwb_enabled(inode)) {
249 struct cgroup_subsys_state *memcg_css;
250
251 if (page) {
252 memcg_css = mem_cgroup_css_from_page(page);
253 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254 } else {
255 /* must pin memcg_css, see wb_get_create() */
256 memcg_css = task_get_css(current, memory_cgrp_id);
257 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
258 css_put(memcg_css);
259 }
260 }
261
262 if (!wb)
263 wb = &bdi->wb;
264
265 /*
266 * There may be multiple instances of this function racing to
267 * update the same inode. Use cmpxchg() to tell the winner.
268 */
269 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
270 wb_put(wb);
271}
272
273/**
274 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
275 * @inode: inode of interest with i_lock held
276 *
277 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
278 * held on entry and is released on return. The returned wb is guaranteed
279 * to stay @inode's associated wb until its list_lock is released.
280 */
281static struct bdi_writeback *
282locked_inode_to_wb_and_lock_list(struct inode *inode)
283 __releases(&inode->i_lock)
284 __acquires(&wb->list_lock)
285{
286 while (true) {
287 struct bdi_writeback *wb = inode_to_wb(inode);
288
289 /*
290 * inode_to_wb() association is protected by both
291 * @inode->i_lock and @wb->list_lock but list_lock nests
292 * outside i_lock. Drop i_lock and verify that the
293 * association hasn't changed after acquiring list_lock.
294 */
295 wb_get(wb);
296 spin_unlock(&inode->i_lock);
297 spin_lock(&wb->list_lock);
298
299 /* i_wb may have changed inbetween, can't use inode_to_wb() */
300 if (likely(wb == inode->i_wb)) {
301 wb_put(wb); /* @inode already has ref */
302 return wb;
303 }
304
305 spin_unlock(&wb->list_lock);
306 wb_put(wb);
307 cpu_relax();
308 spin_lock(&inode->i_lock);
309 }
310}
311
312/**
313 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
314 * @inode: inode of interest
315 *
316 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
317 * on entry.
318 */
319static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
320 __acquires(&wb->list_lock)
321{
322 spin_lock(&inode->i_lock);
323 return locked_inode_to_wb_and_lock_list(inode);
324}
325
326struct inode_switch_wbs_context {
327 struct inode *inode;
328 struct bdi_writeback *new_wb;
329
330 struct rcu_head rcu_head;
331 struct work_struct work;
332};
333
334static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
335{
336 down_write(&bdi->wb_switch_rwsem);
337}
338
339static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
340{
341 up_write(&bdi->wb_switch_rwsem);
342}
343
344static void inode_switch_wbs_work_fn(struct work_struct *work)
345{
346 struct inode_switch_wbs_context *isw =
347 container_of(work, struct inode_switch_wbs_context, work);
348 struct inode *inode = isw->inode;
349 struct backing_dev_info *bdi = inode_to_bdi(inode);
350 struct address_space *mapping = inode->i_mapping;
351 struct bdi_writeback *old_wb = inode->i_wb;
352 struct bdi_writeback *new_wb = isw->new_wb;
353 XA_STATE(xas, &mapping->i_pages, 0);
354 struct page *page;
355 bool switched = false;
356
357 /*
358 * If @inode switches cgwb membership while sync_inodes_sb() is
359 * being issued, sync_inodes_sb() might miss it. Synchronize.
360 */
361 down_read(&bdi->wb_switch_rwsem);
362
363 /*
364 * By the time control reaches here, RCU grace period has passed
365 * since I_WB_SWITCH assertion and all wb stat update transactions
366 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
367 * synchronizing against the i_pages lock.
368 *
369 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
370 * gives us exclusion against all wb related operations on @inode
371 * including IO list manipulations and stat updates.
372 */
373 if (old_wb < new_wb) {
374 spin_lock(&old_wb->list_lock);
375 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
376 } else {
377 spin_lock(&new_wb->list_lock);
378 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
379 }
380 spin_lock(&inode->i_lock);
381 xa_lock_irq(&mapping->i_pages);
382
383 /*
384 * Once I_FREEING is visible under i_lock, the eviction path owns
385 * the inode and we shouldn't modify ->i_io_list.
386 */
387 if (unlikely(inode->i_state & I_FREEING))
388 goto skip_switch;
389
390 /*
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393 * pages actually under writeback.
394 */
395 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
396 if (PageDirty(page)) {
397 dec_wb_stat(old_wb, WB_RECLAIMABLE);
398 inc_wb_stat(new_wb, WB_RECLAIMABLE);
399 }
400 }
401
402 xas_set(&xas, 0);
403 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
404 WARN_ON_ONCE(!PageWriteback(page));
405 dec_wb_stat(old_wb, WB_WRITEBACK);
406 inc_wb_stat(new_wb, WB_WRITEBACK);
407 }
408
409 wb_get(new_wb);
410
411 /*
412 * Transfer to @new_wb's IO list if necessary. The specific list
413 * @inode was on is ignored and the inode is put on ->b_dirty which
414 * is always correct including from ->b_dirty_time. The transfer
415 * preserves @inode->dirtied_when ordering.
416 */
417 if (!list_empty(&inode->i_io_list)) {
418 struct inode *pos;
419
420 inode_io_list_del_locked(inode, old_wb);
421 inode->i_wb = new_wb;
422 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
423 if (time_after_eq(inode->dirtied_when,
424 pos->dirtied_when))
425 break;
426 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
427 } else {
428 inode->i_wb = new_wb;
429 }
430
431 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
432 inode->i_wb_frn_winner = 0;
433 inode->i_wb_frn_avg_time = 0;
434 inode->i_wb_frn_history = 0;
435 switched = true;
436skip_switch:
437 /*
438 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
439 * ensures that the new wb is visible if they see !I_WB_SWITCH.
440 */
441 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
442
443 xa_unlock_irq(&mapping->i_pages);
444 spin_unlock(&inode->i_lock);
445 spin_unlock(&new_wb->list_lock);
446 spin_unlock(&old_wb->list_lock);
447
448 up_read(&bdi->wb_switch_rwsem);
449
450 if (switched) {
451 wb_wakeup(new_wb);
452 wb_put(old_wb);
453 }
454 wb_put(new_wb);
455
456 iput(inode);
457 kfree(isw);
458
459 atomic_dec(&isw_nr_in_flight);
460}
461
462static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
463{
464 struct inode_switch_wbs_context *isw = container_of(rcu_head,
465 struct inode_switch_wbs_context, rcu_head);
466
467 /* needs to grab bh-unsafe locks, bounce to work item */
468 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
469 queue_work(isw_wq, &isw->work);
470}
471
472/**
473 * inode_switch_wbs - change the wb association of an inode
474 * @inode: target inode
475 * @new_wb_id: ID of the new wb
476 *
477 * Switch @inode's wb association to the wb identified by @new_wb_id. The
478 * switching is performed asynchronously and may fail silently.
479 */
480static void inode_switch_wbs(struct inode *inode, int new_wb_id)
481{
482 struct backing_dev_info *bdi = inode_to_bdi(inode);
483 struct cgroup_subsys_state *memcg_css;
484 struct inode_switch_wbs_context *isw;
485
486 /* noop if seems to be already in progress */
487 if (inode->i_state & I_WB_SWITCH)
488 return;
489
490 /*
491 * Avoid starting new switches while sync_inodes_sb() is in
492 * progress. Otherwise, if the down_write protected issue path
493 * blocks heavily, we might end up starting a large number of
494 * switches which will block on the rwsem.
495 */
496 if (!down_read_trylock(&bdi->wb_switch_rwsem))
497 return;
498
499 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
500 if (!isw)
501 goto out_unlock;
502
503 /* find and pin the new wb */
504 rcu_read_lock();
505 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
506 if (memcg_css)
507 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
508 rcu_read_unlock();
509 if (!isw->new_wb)
510 goto out_free;
511
512 /* while holding I_WB_SWITCH, no one else can update the association */
513 spin_lock(&inode->i_lock);
514 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
515 inode->i_state & (I_WB_SWITCH | I_FREEING) ||
516 inode_to_wb(inode) == isw->new_wb) {
517 spin_unlock(&inode->i_lock);
518 goto out_free;
519 }
520 inode->i_state |= I_WB_SWITCH;
521 __iget(inode);
522 spin_unlock(&inode->i_lock);
523
524 isw->inode = inode;
525
526 atomic_inc(&isw_nr_in_flight);
527
528 /*
529 * In addition to synchronizing among switchers, I_WB_SWITCH tells
530 * the RCU protected stat update paths to grab the i_page
531 * lock so that stat transfer can synchronize against them.
532 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
533 */
534 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
535 goto out_unlock;
536
537out_free:
538 if (isw->new_wb)
539 wb_put(isw->new_wb);
540 kfree(isw);
541out_unlock:
542 up_read(&bdi->wb_switch_rwsem);
543}
544
545/**
546 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
547 * @wbc: writeback_control of interest
548 * @inode: target inode
549 *
550 * @inode is locked and about to be written back under the control of @wbc.
551 * Record @inode's writeback context into @wbc and unlock the i_lock. On
552 * writeback completion, wbc_detach_inode() should be called. This is used
553 * to track the cgroup writeback context.
554 */
555void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
556 struct inode *inode)
557{
558 if (!inode_cgwb_enabled(inode)) {
559 spin_unlock(&inode->i_lock);
560 return;
561 }
562
563 wbc->wb = inode_to_wb(inode);
564 wbc->inode = inode;
565
566 wbc->wb_id = wbc->wb->memcg_css->id;
567 wbc->wb_lcand_id = inode->i_wb_frn_winner;
568 wbc->wb_tcand_id = 0;
569 wbc->wb_bytes = 0;
570 wbc->wb_lcand_bytes = 0;
571 wbc->wb_tcand_bytes = 0;
572
573 wb_get(wbc->wb);
574 spin_unlock(&inode->i_lock);
575
576 /*
577 * A dying wb indicates that the memcg-blkcg mapping has changed
578 * and a new wb is already serving the memcg. Switch immediately.
579 */
580 if (unlikely(wb_dying(wbc->wb)))
581 inode_switch_wbs(inode, wbc->wb_id);
582}
583
584/**
585 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
586 * @wbc: writeback_control of the just finished writeback
587 *
588 * To be called after a writeback attempt of an inode finishes and undoes
589 * wbc_attach_and_unlock_inode(). Can be called under any context.
590 *
591 * As concurrent write sharing of an inode is expected to be very rare and
592 * memcg only tracks page ownership on first-use basis severely confining
593 * the usefulness of such sharing, cgroup writeback tracks ownership
594 * per-inode. While the support for concurrent write sharing of an inode
595 * is deemed unnecessary, an inode being written to by different cgroups at
596 * different points in time is a lot more common, and, more importantly,
597 * charging only by first-use can too readily lead to grossly incorrect
598 * behaviors (single foreign page can lead to gigabytes of writeback to be
599 * incorrectly attributed).
600 *
601 * To resolve this issue, cgroup writeback detects the majority dirtier of
602 * an inode and transfers the ownership to it. To avoid unnnecessary
603 * oscillation, the detection mechanism keeps track of history and gives
604 * out the switch verdict only if the foreign usage pattern is stable over
605 * a certain amount of time and/or writeback attempts.
606 *
607 * On each writeback attempt, @wbc tries to detect the majority writer
608 * using Boyer-Moore majority vote algorithm. In addition to the byte
609 * count from the majority voting, it also counts the bytes written for the
610 * current wb and the last round's winner wb (max of last round's current
611 * wb, the winner from two rounds ago, and the last round's majority
612 * candidate). Keeping track of the historical winner helps the algorithm
613 * to semi-reliably detect the most active writer even when it's not the
614 * absolute majority.
615 *
616 * Once the winner of the round is determined, whether the winner is
617 * foreign or not and how much IO time the round consumed is recorded in
618 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
619 * over a certain threshold, the switch verdict is given.
620 */
621void wbc_detach_inode(struct writeback_control *wbc)
622{
623 struct bdi_writeback *wb = wbc->wb;
624 struct inode *inode = wbc->inode;
625 unsigned long avg_time, max_bytes, max_time;
626 u16 history;
627 int max_id;
628
629 if (!wb)
630 return;
631
632 history = inode->i_wb_frn_history;
633 avg_time = inode->i_wb_frn_avg_time;
634
635 /* pick the winner of this round */
636 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
637 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
638 max_id = wbc->wb_id;
639 max_bytes = wbc->wb_bytes;
640 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
641 max_id = wbc->wb_lcand_id;
642 max_bytes = wbc->wb_lcand_bytes;
643 } else {
644 max_id = wbc->wb_tcand_id;
645 max_bytes = wbc->wb_tcand_bytes;
646 }
647
648 /*
649 * Calculate the amount of IO time the winner consumed and fold it
650 * into the running average kept per inode. If the consumed IO
651 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
652 * deciding whether to switch or not. This is to prevent one-off
653 * small dirtiers from skewing the verdict.
654 */
655 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
656 wb->avg_write_bandwidth);
657 if (avg_time)
658 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
659 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
660 else
661 avg_time = max_time; /* immediate catch up on first run */
662
663 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
664 int slots;
665
666 /*
667 * The switch verdict is reached if foreign wb's consume
668 * more than a certain proportion of IO time in a
669 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
670 * history mask where each bit represents one sixteenth of
671 * the period. Determine the number of slots to shift into
672 * history from @max_time.
673 */
674 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
675 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
676 history <<= slots;
677 if (wbc->wb_id != max_id)
678 history |= (1U << slots) - 1;
679
680 /*
681 * Switch if the current wb isn't the consistent winner.
682 * If there are multiple closely competing dirtiers, the
683 * inode may switch across them repeatedly over time, which
684 * is okay. The main goal is avoiding keeping an inode on
685 * the wrong wb for an extended period of time.
686 */
687 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
688 inode_switch_wbs(inode, max_id);
689 }
690
691 /*
692 * Multiple instances of this function may race to update the
693 * following fields but we don't mind occassional inaccuracies.
694 */
695 inode->i_wb_frn_winner = max_id;
696 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
697 inode->i_wb_frn_history = history;
698
699 wb_put(wbc->wb);
700 wbc->wb = NULL;
701}
702
703/**
704 * wbc_account_io - account IO issued during writeback
705 * @wbc: writeback_control of the writeback in progress
706 * @page: page being written out
707 * @bytes: number of bytes being written out
708 *
709 * @bytes from @page are about to written out during the writeback
710 * controlled by @wbc. Keep the book for foreign inode detection. See
711 * wbc_detach_inode().
712 */
713void wbc_account_io(struct writeback_control *wbc, struct page *page,
714 size_t bytes)
715{
716 int id;
717
718 /*
719 * pageout() path doesn't attach @wbc to the inode being written
720 * out. This is intentional as we don't want the function to block
721 * behind a slow cgroup. Ultimately, we want pageout() to kick off
722 * regular writeback instead of writing things out itself.
723 */
724 if (!wbc->wb)
725 return;
726
727 id = mem_cgroup_css_from_page(page)->id;
728
729 if (id == wbc->wb_id) {
730 wbc->wb_bytes += bytes;
731 return;
732 }
733
734 if (id == wbc->wb_lcand_id)
735 wbc->wb_lcand_bytes += bytes;
736
737 /* Boyer-Moore majority vote algorithm */
738 if (!wbc->wb_tcand_bytes)
739 wbc->wb_tcand_id = id;
740 if (id == wbc->wb_tcand_id)
741 wbc->wb_tcand_bytes += bytes;
742 else
743 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
744}
745EXPORT_SYMBOL_GPL(wbc_account_io);
746
747/**
748 * inode_congested - test whether an inode is congested
749 * @inode: inode to test for congestion (may be NULL)
750 * @cong_bits: mask of WB_[a]sync_congested bits to test
751 *
752 * Tests whether @inode is congested. @cong_bits is the mask of congestion
753 * bits to test and the return value is the mask of set bits.
754 *
755 * If cgroup writeback is enabled for @inode, the congestion state is
756 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
757 * associated with @inode is congested; otherwise, the root wb's congestion
758 * state is used.
759 *
760 * @inode is allowed to be NULL as this function is often called on
761 * mapping->host which is NULL for the swapper space.
762 */
763int inode_congested(struct inode *inode, int cong_bits)
764{
765 /*
766 * Once set, ->i_wb never becomes NULL while the inode is alive.
767 * Start transaction iff ->i_wb is visible.
768 */
769 if (inode && inode_to_wb_is_valid(inode)) {
770 struct bdi_writeback *wb;
771 struct wb_lock_cookie lock_cookie = {};
772 bool congested;
773
774 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
775 congested = wb_congested(wb, cong_bits);
776 unlocked_inode_to_wb_end(inode, &lock_cookie);
777 return congested;
778 }
779
780 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
781}
782EXPORT_SYMBOL_GPL(inode_congested);
783
784/**
785 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
786 * @wb: target bdi_writeback to split @nr_pages to
787 * @nr_pages: number of pages to write for the whole bdi
788 *
789 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
790 * relation to the total write bandwidth of all wb's w/ dirty inodes on
791 * @wb->bdi.
792 */
793static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
794{
795 unsigned long this_bw = wb->avg_write_bandwidth;
796 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
797
798 if (nr_pages == LONG_MAX)
799 return LONG_MAX;
800
801 /*
802 * This may be called on clean wb's and proportional distribution
803 * may not make sense, just use the original @nr_pages in those
804 * cases. In general, we wanna err on the side of writing more.
805 */
806 if (!tot_bw || this_bw >= tot_bw)
807 return nr_pages;
808 else
809 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
810}
811
812/**
813 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
814 * @bdi: target backing_dev_info
815 * @base_work: wb_writeback_work to issue
816 * @skip_if_busy: skip wb's which already have writeback in progress
817 *
818 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
819 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
820 * distributed to the busy wbs according to each wb's proportion in the
821 * total active write bandwidth of @bdi.
822 */
823static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
824 struct wb_writeback_work *base_work,
825 bool skip_if_busy)
826{
827 struct bdi_writeback *last_wb = NULL;
828 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
829 struct bdi_writeback, bdi_node);
830
831 might_sleep();
832restart:
833 rcu_read_lock();
834 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
835 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
836 struct wb_writeback_work fallback_work;
837 struct wb_writeback_work *work;
838 long nr_pages;
839
840 if (last_wb) {
841 wb_put(last_wb);
842 last_wb = NULL;
843 }
844
845 /* SYNC_ALL writes out I_DIRTY_TIME too */
846 if (!wb_has_dirty_io(wb) &&
847 (base_work->sync_mode == WB_SYNC_NONE ||
848 list_empty(&wb->b_dirty_time)))
849 continue;
850 if (skip_if_busy && writeback_in_progress(wb))
851 continue;
852
853 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
854
855 work = kmalloc(sizeof(*work), GFP_ATOMIC);
856 if (work) {
857 *work = *base_work;
858 work->nr_pages = nr_pages;
859 work->auto_free = 1;
860 wb_queue_work(wb, work);
861 continue;
862 }
863
864 /* alloc failed, execute synchronously using on-stack fallback */
865 work = &fallback_work;
866 *work = *base_work;
867 work->nr_pages = nr_pages;
868 work->auto_free = 0;
869 work->done = &fallback_work_done;
870
871 wb_queue_work(wb, work);
872
873 /*
874 * Pin @wb so that it stays on @bdi->wb_list. This allows
875 * continuing iteration from @wb after dropping and
876 * regrabbing rcu read lock.
877 */
878 wb_get(wb);
879 last_wb = wb;
880
881 rcu_read_unlock();
882 wb_wait_for_completion(bdi, &fallback_work_done);
883 goto restart;
884 }
885 rcu_read_unlock();
886
887 if (last_wb)
888 wb_put(last_wb);
889}
890
891/**
892 * cgroup_writeback_umount - flush inode wb switches for umount
893 *
894 * This function is called when a super_block is about to be destroyed and
895 * flushes in-flight inode wb switches. An inode wb switch goes through
896 * RCU and then workqueue, so the two need to be flushed in order to ensure
897 * that all previously scheduled switches are finished. As wb switches are
898 * rare occurrences and synchronize_rcu() can take a while, perform
899 * flushing iff wb switches are in flight.
900 */
901void cgroup_writeback_umount(void)
902{
903 if (atomic_read(&isw_nr_in_flight)) {
904 synchronize_rcu();
905 flush_workqueue(isw_wq);
906 }
907}
908
909static int __init cgroup_writeback_init(void)
910{
911 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
912 if (!isw_wq)
913 return -ENOMEM;
914 return 0;
915}
916fs_initcall(cgroup_writeback_init);
917
918#else /* CONFIG_CGROUP_WRITEBACK */
919
920static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
921static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
922
923static struct bdi_writeback *
924locked_inode_to_wb_and_lock_list(struct inode *inode)
925 __releases(&inode->i_lock)
926 __acquires(&wb->list_lock)
927{
928 struct bdi_writeback *wb = inode_to_wb(inode);
929
930 spin_unlock(&inode->i_lock);
931 spin_lock(&wb->list_lock);
932 return wb;
933}
934
935static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
936 __acquires(&wb->list_lock)
937{
938 struct bdi_writeback *wb = inode_to_wb(inode);
939
940 spin_lock(&wb->list_lock);
941 return wb;
942}
943
944static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
945{
946 return nr_pages;
947}
948
949static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
950 struct wb_writeback_work *base_work,
951 bool skip_if_busy)
952{
953 might_sleep();
954
955 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
956 base_work->auto_free = 0;
957 wb_queue_work(&bdi->wb, base_work);
958 }
959}
960
961#endif /* CONFIG_CGROUP_WRITEBACK */
962
963/*
964 * Add in the number of potentially dirty inodes, because each inode
965 * write can dirty pagecache in the underlying blockdev.
966 */
967static unsigned long get_nr_dirty_pages(void)
968{
969 return global_node_page_state(NR_FILE_DIRTY) +
970 global_node_page_state(NR_UNSTABLE_NFS) +
971 get_nr_dirty_inodes();
972}
973
974static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
975{
976 if (!wb_has_dirty_io(wb))
977 return;
978
979 /*
980 * All callers of this function want to start writeback of all
981 * dirty pages. Places like vmscan can call this at a very
982 * high frequency, causing pointless allocations of tons of
983 * work items and keeping the flusher threads busy retrieving
984 * that work. Ensure that we only allow one of them pending and
985 * inflight at the time.
986 */
987 if (test_bit(WB_start_all, &wb->state) ||
988 test_and_set_bit(WB_start_all, &wb->state))
989 return;
990
991 wb->start_all_reason = reason;
992 wb_wakeup(wb);
993}
994
995/**
996 * wb_start_background_writeback - start background writeback
997 * @wb: bdi_writback to write from
998 *
999 * Description:
1000 * This makes sure WB_SYNC_NONE background writeback happens. When
1001 * this function returns, it is only guaranteed that for given wb
1002 * some IO is happening if we are over background dirty threshold.
1003 * Caller need not hold sb s_umount semaphore.
1004 */
1005void wb_start_background_writeback(struct bdi_writeback *wb)
1006{
1007 /*
1008 * We just wake up the flusher thread. It will perform background
1009 * writeback as soon as there is no other work to do.
1010 */
1011 trace_writeback_wake_background(wb);
1012 wb_wakeup(wb);
1013}
1014
1015/*
1016 * Remove the inode from the writeback list it is on.
1017 */
1018void inode_io_list_del(struct inode *inode)
1019{
1020 struct bdi_writeback *wb;
1021
1022 wb = inode_to_wb_and_lock_list(inode);
1023 inode_io_list_del_locked(inode, wb);
1024 spin_unlock(&wb->list_lock);
1025}
1026
1027/*
1028 * mark an inode as under writeback on the sb
1029 */
1030void sb_mark_inode_writeback(struct inode *inode)
1031{
1032 struct super_block *sb = inode->i_sb;
1033 unsigned long flags;
1034
1035 if (list_empty(&inode->i_wb_list)) {
1036 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1037 if (list_empty(&inode->i_wb_list)) {
1038 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1039 trace_sb_mark_inode_writeback(inode);
1040 }
1041 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1042 }
1043}
1044
1045/*
1046 * clear an inode as under writeback on the sb
1047 */
1048void sb_clear_inode_writeback(struct inode *inode)
1049{
1050 struct super_block *sb = inode->i_sb;
1051 unsigned long flags;
1052
1053 if (!list_empty(&inode->i_wb_list)) {
1054 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1055 if (!list_empty(&inode->i_wb_list)) {
1056 list_del_init(&inode->i_wb_list);
1057 trace_sb_clear_inode_writeback(inode);
1058 }
1059 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1060 }
1061}
1062
1063/*
1064 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1065 * furthest end of its superblock's dirty-inode list.
1066 *
1067 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1068 * already the most-recently-dirtied inode on the b_dirty list. If that is
1069 * the case then the inode must have been redirtied while it was being written
1070 * out and we don't reset its dirtied_when.
1071 */
1072static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1073{
1074 if (!list_empty(&wb->b_dirty)) {
1075 struct inode *tail;
1076
1077 tail = wb_inode(wb->b_dirty.next);
1078 if (time_before(inode->dirtied_when, tail->dirtied_when))
1079 inode->dirtied_when = jiffies;
1080 }
1081 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1082}
1083
1084/*
1085 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1086 */
1087static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1088{
1089 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1090}
1091
1092static void inode_sync_complete(struct inode *inode)
1093{
1094 inode->i_state &= ~I_SYNC;
1095 /* If inode is clean an unused, put it into LRU now... */
1096 inode_add_lru(inode);
1097 /* Waiters must see I_SYNC cleared before being woken up */
1098 smp_mb();
1099 wake_up_bit(&inode->i_state, __I_SYNC);
1100}
1101
1102static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1103{
1104 bool ret = time_after(inode->dirtied_when, t);
1105#ifndef CONFIG_64BIT
1106 /*
1107 * For inodes being constantly redirtied, dirtied_when can get stuck.
1108 * It _appears_ to be in the future, but is actually in distant past.
1109 * This test is necessary to prevent such wrapped-around relative times
1110 * from permanently stopping the whole bdi writeback.
1111 */
1112 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1113#endif
1114 return ret;
1115}
1116
1117#define EXPIRE_DIRTY_ATIME 0x0001
1118
1119/*
1120 * Move expired (dirtied before work->older_than_this) dirty inodes from
1121 * @delaying_queue to @dispatch_queue.
1122 */
1123static int move_expired_inodes(struct list_head *delaying_queue,
1124 struct list_head *dispatch_queue,
1125 int flags,
1126 struct wb_writeback_work *work)
1127{
1128 unsigned long *older_than_this = NULL;
1129 unsigned long expire_time;
1130 LIST_HEAD(tmp);
1131 struct list_head *pos, *node;
1132 struct super_block *sb = NULL;
1133 struct inode *inode;
1134 int do_sb_sort = 0;
1135 int moved = 0;
1136
1137 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1138 older_than_this = work->older_than_this;
1139 else if (!work->for_sync) {
1140 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1141 older_than_this = &expire_time;
1142 }
1143 while (!list_empty(delaying_queue)) {
1144 inode = wb_inode(delaying_queue->prev);
1145 if (older_than_this &&
1146 inode_dirtied_after(inode, *older_than_this))
1147 break;
1148 list_move(&inode->i_io_list, &tmp);
1149 moved++;
1150 if (flags & EXPIRE_DIRTY_ATIME)
1151 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1152 if (sb_is_blkdev_sb(inode->i_sb))
1153 continue;
1154 if (sb && sb != inode->i_sb)
1155 do_sb_sort = 1;
1156 sb = inode->i_sb;
1157 }
1158
1159 /* just one sb in list, splice to dispatch_queue and we're done */
1160 if (!do_sb_sort) {
1161 list_splice(&tmp, dispatch_queue);
1162 goto out;
1163 }
1164
1165 /* Move inodes from one superblock together */
1166 while (!list_empty(&tmp)) {
1167 sb = wb_inode(tmp.prev)->i_sb;
1168 list_for_each_prev_safe(pos, node, &tmp) {
1169 inode = wb_inode(pos);
1170 if (inode->i_sb == sb)
1171 list_move(&inode->i_io_list, dispatch_queue);
1172 }
1173 }
1174out:
1175 return moved;
1176}
1177
1178/*
1179 * Queue all expired dirty inodes for io, eldest first.
1180 * Before
1181 * newly dirtied b_dirty b_io b_more_io
1182 * =============> gf edc BA
1183 * After
1184 * newly dirtied b_dirty b_io b_more_io
1185 * =============> g fBAedc
1186 * |
1187 * +--> dequeue for IO
1188 */
1189static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1190{
1191 int moved;
1192
1193 assert_spin_locked(&wb->list_lock);
1194 list_splice_init(&wb->b_more_io, &wb->b_io);
1195 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1196 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1197 EXPIRE_DIRTY_ATIME, work);
1198 if (moved)
1199 wb_io_lists_populated(wb);
1200 trace_writeback_queue_io(wb, work, moved);
1201}
1202
1203static int write_inode(struct inode *inode, struct writeback_control *wbc)
1204{
1205 int ret;
1206
1207 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1208 trace_writeback_write_inode_start(inode, wbc);
1209 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1210 trace_writeback_write_inode(inode, wbc);
1211 return ret;
1212 }
1213 return 0;
1214}
1215
1216/*
1217 * Wait for writeback on an inode to complete. Called with i_lock held.
1218 * Caller must make sure inode cannot go away when we drop i_lock.
1219 */
1220static void __inode_wait_for_writeback(struct inode *inode)
1221 __releases(inode->i_lock)
1222 __acquires(inode->i_lock)
1223{
1224 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1225 wait_queue_head_t *wqh;
1226
1227 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1228 while (inode->i_state & I_SYNC) {
1229 spin_unlock(&inode->i_lock);
1230 __wait_on_bit(wqh, &wq, bit_wait,
1231 TASK_UNINTERRUPTIBLE);
1232 spin_lock(&inode->i_lock);
1233 }
1234}
1235
1236/*
1237 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1238 */
1239void inode_wait_for_writeback(struct inode *inode)
1240{
1241 spin_lock(&inode->i_lock);
1242 __inode_wait_for_writeback(inode);
1243 spin_unlock(&inode->i_lock);
1244}
1245
1246/*
1247 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1248 * held and drops it. It is aimed for callers not holding any inode reference
1249 * so once i_lock is dropped, inode can go away.
1250 */
1251static void inode_sleep_on_writeback(struct inode *inode)
1252 __releases(inode->i_lock)
1253{
1254 DEFINE_WAIT(wait);
1255 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1256 int sleep;
1257
1258 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1259 sleep = inode->i_state & I_SYNC;
1260 spin_unlock(&inode->i_lock);
1261 if (sleep)
1262 schedule();
1263 finish_wait(wqh, &wait);
1264}
1265
1266/*
1267 * Find proper writeback list for the inode depending on its current state and
1268 * possibly also change of its state while we were doing writeback. Here we
1269 * handle things such as livelock prevention or fairness of writeback among
1270 * inodes. This function can be called only by flusher thread - noone else
1271 * processes all inodes in writeback lists and requeueing inodes behind flusher
1272 * thread's back can have unexpected consequences.
1273 */
1274static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1275 struct writeback_control *wbc)
1276{
1277 if (inode->i_state & I_FREEING)
1278 return;
1279
1280 /*
1281 * Sync livelock prevention. Each inode is tagged and synced in one
1282 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1283 * the dirty time to prevent enqueue and sync it again.
1284 */
1285 if ((inode->i_state & I_DIRTY) &&
1286 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1287 inode->dirtied_when = jiffies;
1288
1289 if (wbc->pages_skipped) {
1290 /*
1291 * writeback is not making progress due to locked
1292 * buffers. Skip this inode for now.
1293 */
1294 redirty_tail(inode, wb);
1295 return;
1296 }
1297
1298 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1299 /*
1300 * We didn't write back all the pages. nfs_writepages()
1301 * sometimes bales out without doing anything.
1302 */
1303 if (wbc->nr_to_write <= 0) {
1304 /* Slice used up. Queue for next turn. */
1305 requeue_io(inode, wb);
1306 } else {
1307 /*
1308 * Writeback blocked by something other than
1309 * congestion. Delay the inode for some time to
1310 * avoid spinning on the CPU (100% iowait)
1311 * retrying writeback of the dirty page/inode
1312 * that cannot be performed immediately.
1313 */
1314 redirty_tail(inode, wb);
1315 }
1316 } else if (inode->i_state & I_DIRTY) {
1317 /*
1318 * Filesystems can dirty the inode during writeback operations,
1319 * such as delayed allocation during submission or metadata
1320 * updates after data IO completion.
1321 */
1322 redirty_tail(inode, wb);
1323 } else if (inode->i_state & I_DIRTY_TIME) {
1324 inode->dirtied_when = jiffies;
1325 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1326 } else {
1327 /* The inode is clean. Remove from writeback lists. */
1328 inode_io_list_del_locked(inode, wb);
1329 }
1330}
1331
1332/*
1333 * Write out an inode and its dirty pages. Do not update the writeback list
1334 * linkage. That is left to the caller. The caller is also responsible for
1335 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1336 */
1337static int
1338__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1339{
1340 struct address_space *mapping = inode->i_mapping;
1341 long nr_to_write = wbc->nr_to_write;
1342 unsigned dirty;
1343 int ret;
1344
1345 WARN_ON(!(inode->i_state & I_SYNC));
1346
1347 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1348
1349 ret = do_writepages(mapping, wbc);
1350
1351 /*
1352 * Make sure to wait on the data before writing out the metadata.
1353 * This is important for filesystems that modify metadata on data
1354 * I/O completion. We don't do it for sync(2) writeback because it has a
1355 * separate, external IO completion path and ->sync_fs for guaranteeing
1356 * inode metadata is written back correctly.
1357 */
1358 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1359 int err = filemap_fdatawait(mapping);
1360 if (ret == 0)
1361 ret = err;
1362 }
1363
1364 /*
1365 * Some filesystems may redirty the inode during the writeback
1366 * due to delalloc, clear dirty metadata flags right before
1367 * write_inode()
1368 */
1369 spin_lock(&inode->i_lock);
1370
1371 dirty = inode->i_state & I_DIRTY;
1372 if (inode->i_state & I_DIRTY_TIME) {
1373 if ((dirty & I_DIRTY_INODE) ||
1374 wbc->sync_mode == WB_SYNC_ALL ||
1375 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1376 unlikely(time_after(jiffies,
1377 (inode->dirtied_time_when +
1378 dirtytime_expire_interval * HZ)))) {
1379 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1380 trace_writeback_lazytime(inode);
1381 }
1382 } else
1383 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1384 inode->i_state &= ~dirty;
1385
1386 /*
1387 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1388 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1389 * either they see the I_DIRTY bits cleared or we see the dirtied
1390 * inode.
1391 *
1392 * I_DIRTY_PAGES is always cleared together above even if @mapping
1393 * still has dirty pages. The flag is reinstated after smp_mb() if
1394 * necessary. This guarantees that either __mark_inode_dirty()
1395 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1396 */
1397 smp_mb();
1398
1399 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1400 inode->i_state |= I_DIRTY_PAGES;
1401
1402 spin_unlock(&inode->i_lock);
1403
1404 if (dirty & I_DIRTY_TIME)
1405 mark_inode_dirty_sync(inode);
1406 /* Don't write the inode if only I_DIRTY_PAGES was set */
1407 if (dirty & ~I_DIRTY_PAGES) {
1408 int err = write_inode(inode, wbc);
1409 if (ret == 0)
1410 ret = err;
1411 }
1412 trace_writeback_single_inode(inode, wbc, nr_to_write);
1413 return ret;
1414}
1415
1416/*
1417 * Write out an inode's dirty pages. Either the caller has an active reference
1418 * on the inode or the inode has I_WILL_FREE set.
1419 *
1420 * This function is designed to be called for writing back one inode which
1421 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1422 * and does more profound writeback list handling in writeback_sb_inodes().
1423 */
1424static int writeback_single_inode(struct inode *inode,
1425 struct writeback_control *wbc)
1426{
1427 struct bdi_writeback *wb;
1428 int ret = 0;
1429
1430 spin_lock(&inode->i_lock);
1431 if (!atomic_read(&inode->i_count))
1432 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1433 else
1434 WARN_ON(inode->i_state & I_WILL_FREE);
1435
1436 if (inode->i_state & I_SYNC) {
1437 if (wbc->sync_mode != WB_SYNC_ALL)
1438 goto out;
1439 /*
1440 * It's a data-integrity sync. We must wait. Since callers hold
1441 * inode reference or inode has I_WILL_FREE set, it cannot go
1442 * away under us.
1443 */
1444 __inode_wait_for_writeback(inode);
1445 }
1446 WARN_ON(inode->i_state & I_SYNC);
1447 /*
1448 * Skip inode if it is clean and we have no outstanding writeback in
1449 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1450 * function since flusher thread may be doing for example sync in
1451 * parallel and if we move the inode, it could get skipped. So here we
1452 * make sure inode is on some writeback list and leave it there unless
1453 * we have completely cleaned the inode.
1454 */
1455 if (!(inode->i_state & I_DIRTY_ALL) &&
1456 (wbc->sync_mode != WB_SYNC_ALL ||
1457 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1458 goto out;
1459 inode->i_state |= I_SYNC;
1460 wbc_attach_and_unlock_inode(wbc, inode);
1461
1462 ret = __writeback_single_inode(inode, wbc);
1463
1464 wbc_detach_inode(wbc);
1465
1466 wb = inode_to_wb_and_lock_list(inode);
1467 spin_lock(&inode->i_lock);
1468 /*
1469 * If inode is clean, remove it from writeback lists. Otherwise don't
1470 * touch it. See comment above for explanation.
1471 */
1472 if (!(inode->i_state & I_DIRTY_ALL))
1473 inode_io_list_del_locked(inode, wb);
1474 spin_unlock(&wb->list_lock);
1475 inode_sync_complete(inode);
1476out:
1477 spin_unlock(&inode->i_lock);
1478 return ret;
1479}
1480
1481static long writeback_chunk_size(struct bdi_writeback *wb,
1482 struct wb_writeback_work *work)
1483{
1484 long pages;
1485
1486 /*
1487 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1488 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1489 * here avoids calling into writeback_inodes_wb() more than once.
1490 *
1491 * The intended call sequence for WB_SYNC_ALL writeback is:
1492 *
1493 * wb_writeback()
1494 * writeback_sb_inodes() <== called only once
1495 * write_cache_pages() <== called once for each inode
1496 * (quickly) tag currently dirty pages
1497 * (maybe slowly) sync all tagged pages
1498 */
1499 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1500 pages = LONG_MAX;
1501 else {
1502 pages = min(wb->avg_write_bandwidth / 2,
1503 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1504 pages = min(pages, work->nr_pages);
1505 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1506 MIN_WRITEBACK_PAGES);
1507 }
1508
1509 return pages;
1510}
1511
1512/*
1513 * Write a portion of b_io inodes which belong to @sb.
1514 *
1515 * Return the number of pages and/or inodes written.
1516 *
1517 * NOTE! This is called with wb->list_lock held, and will
1518 * unlock and relock that for each inode it ends up doing
1519 * IO for.
1520 */
1521static long writeback_sb_inodes(struct super_block *sb,
1522 struct bdi_writeback *wb,
1523 struct wb_writeback_work *work)
1524{
1525 struct writeback_control wbc = {
1526 .sync_mode = work->sync_mode,
1527 .tagged_writepages = work->tagged_writepages,
1528 .for_kupdate = work->for_kupdate,
1529 .for_background = work->for_background,
1530 .for_sync = work->for_sync,
1531 .range_cyclic = work->range_cyclic,
1532 .range_start = 0,
1533 .range_end = LLONG_MAX,
1534 };
1535 unsigned long start_time = jiffies;
1536 long write_chunk;
1537 long wrote = 0; /* count both pages and inodes */
1538
1539 while (!list_empty(&wb->b_io)) {
1540 struct inode *inode = wb_inode(wb->b_io.prev);
1541 struct bdi_writeback *tmp_wb;
1542
1543 if (inode->i_sb != sb) {
1544 if (work->sb) {
1545 /*
1546 * We only want to write back data for this
1547 * superblock, move all inodes not belonging
1548 * to it back onto the dirty list.
1549 */
1550 redirty_tail(inode, wb);
1551 continue;
1552 }
1553
1554 /*
1555 * The inode belongs to a different superblock.
1556 * Bounce back to the caller to unpin this and
1557 * pin the next superblock.
1558 */
1559 break;
1560 }
1561
1562 /*
1563 * Don't bother with new inodes or inodes being freed, first
1564 * kind does not need periodic writeout yet, and for the latter
1565 * kind writeout is handled by the freer.
1566 */
1567 spin_lock(&inode->i_lock);
1568 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1569 spin_unlock(&inode->i_lock);
1570 redirty_tail(inode, wb);
1571 continue;
1572 }
1573 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1574 /*
1575 * If this inode is locked for writeback and we are not
1576 * doing writeback-for-data-integrity, move it to
1577 * b_more_io so that writeback can proceed with the
1578 * other inodes on s_io.
1579 *
1580 * We'll have another go at writing back this inode
1581 * when we completed a full scan of b_io.
1582 */
1583 spin_unlock(&inode->i_lock);
1584 requeue_io(inode, wb);
1585 trace_writeback_sb_inodes_requeue(inode);
1586 continue;
1587 }
1588 spin_unlock(&wb->list_lock);
1589
1590 /*
1591 * We already requeued the inode if it had I_SYNC set and we
1592 * are doing WB_SYNC_NONE writeback. So this catches only the
1593 * WB_SYNC_ALL case.
1594 */
1595 if (inode->i_state & I_SYNC) {
1596 /* Wait for I_SYNC. This function drops i_lock... */
1597 inode_sleep_on_writeback(inode);
1598 /* Inode may be gone, start again */
1599 spin_lock(&wb->list_lock);
1600 continue;
1601 }
1602 inode->i_state |= I_SYNC;
1603 wbc_attach_and_unlock_inode(&wbc, inode);
1604
1605 write_chunk = writeback_chunk_size(wb, work);
1606 wbc.nr_to_write = write_chunk;
1607 wbc.pages_skipped = 0;
1608
1609 /*
1610 * We use I_SYNC to pin the inode in memory. While it is set
1611 * evict_inode() will wait so the inode cannot be freed.
1612 */
1613 __writeback_single_inode(inode, &wbc);
1614
1615 wbc_detach_inode(&wbc);
1616 work->nr_pages -= write_chunk - wbc.nr_to_write;
1617 wrote += write_chunk - wbc.nr_to_write;
1618
1619 if (need_resched()) {
1620 /*
1621 * We're trying to balance between building up a nice
1622 * long list of IOs to improve our merge rate, and
1623 * getting those IOs out quickly for anyone throttling
1624 * in balance_dirty_pages(). cond_resched() doesn't
1625 * unplug, so get our IOs out the door before we
1626 * give up the CPU.
1627 */
1628 blk_flush_plug(current);
1629 cond_resched();
1630 }
1631
1632 /*
1633 * Requeue @inode if still dirty. Be careful as @inode may
1634 * have been switched to another wb in the meantime.
1635 */
1636 tmp_wb = inode_to_wb_and_lock_list(inode);
1637 spin_lock(&inode->i_lock);
1638 if (!(inode->i_state & I_DIRTY_ALL))
1639 wrote++;
1640 requeue_inode(inode, tmp_wb, &wbc);
1641 inode_sync_complete(inode);
1642 spin_unlock(&inode->i_lock);
1643
1644 if (unlikely(tmp_wb != wb)) {
1645 spin_unlock(&tmp_wb->list_lock);
1646 spin_lock(&wb->list_lock);
1647 }
1648
1649 /*
1650 * bail out to wb_writeback() often enough to check
1651 * background threshold and other termination conditions.
1652 */
1653 if (wrote) {
1654 if (time_is_before_jiffies(start_time + HZ / 10UL))
1655 break;
1656 if (work->nr_pages <= 0)
1657 break;
1658 }
1659 }
1660 return wrote;
1661}
1662
1663static long __writeback_inodes_wb(struct bdi_writeback *wb,
1664 struct wb_writeback_work *work)
1665{
1666 unsigned long start_time = jiffies;
1667 long wrote = 0;
1668
1669 while (!list_empty(&wb->b_io)) {
1670 struct inode *inode = wb_inode(wb->b_io.prev);
1671 struct super_block *sb = inode->i_sb;
1672
1673 if (!trylock_super(sb)) {
1674 /*
1675 * trylock_super() may fail consistently due to
1676 * s_umount being grabbed by someone else. Don't use
1677 * requeue_io() to avoid busy retrying the inode/sb.
1678 */
1679 redirty_tail(inode, wb);
1680 continue;
1681 }
1682 wrote += writeback_sb_inodes(sb, wb, work);
1683 up_read(&sb->s_umount);
1684
1685 /* refer to the same tests at the end of writeback_sb_inodes */
1686 if (wrote) {
1687 if (time_is_before_jiffies(start_time + HZ / 10UL))
1688 break;
1689 if (work->nr_pages <= 0)
1690 break;
1691 }
1692 }
1693 /* Leave any unwritten inodes on b_io */
1694 return wrote;
1695}
1696
1697static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1698 enum wb_reason reason)
1699{
1700 struct wb_writeback_work work = {
1701 .nr_pages = nr_pages,
1702 .sync_mode = WB_SYNC_NONE,
1703 .range_cyclic = 1,
1704 .reason = reason,
1705 };
1706 struct blk_plug plug;
1707
1708 blk_start_plug(&plug);
1709 spin_lock(&wb->list_lock);
1710 if (list_empty(&wb->b_io))
1711 queue_io(wb, &work);
1712 __writeback_inodes_wb(wb, &work);
1713 spin_unlock(&wb->list_lock);
1714 blk_finish_plug(&plug);
1715
1716 return nr_pages - work.nr_pages;
1717}
1718
1719/*
1720 * Explicit flushing or periodic writeback of "old" data.
1721 *
1722 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1723 * dirtying-time in the inode's address_space. So this periodic writeback code
1724 * just walks the superblock inode list, writing back any inodes which are
1725 * older than a specific point in time.
1726 *
1727 * Try to run once per dirty_writeback_interval. But if a writeback event
1728 * takes longer than a dirty_writeback_interval interval, then leave a
1729 * one-second gap.
1730 *
1731 * older_than_this takes precedence over nr_to_write. So we'll only write back
1732 * all dirty pages if they are all attached to "old" mappings.
1733 */
1734static long wb_writeback(struct bdi_writeback *wb,
1735 struct wb_writeback_work *work)
1736{
1737 unsigned long wb_start = jiffies;
1738 long nr_pages = work->nr_pages;
1739 unsigned long oldest_jif;
1740 struct inode *inode;
1741 long progress;
1742 struct blk_plug plug;
1743
1744 oldest_jif = jiffies;
1745 work->older_than_this = &oldest_jif;
1746
1747 blk_start_plug(&plug);
1748 spin_lock(&wb->list_lock);
1749 for (;;) {
1750 /*
1751 * Stop writeback when nr_pages has been consumed
1752 */
1753 if (work->nr_pages <= 0)
1754 break;
1755
1756 /*
1757 * Background writeout and kupdate-style writeback may
1758 * run forever. Stop them if there is other work to do
1759 * so that e.g. sync can proceed. They'll be restarted
1760 * after the other works are all done.
1761 */
1762 if ((work->for_background || work->for_kupdate) &&
1763 !list_empty(&wb->work_list))
1764 break;
1765
1766 /*
1767 * For background writeout, stop when we are below the
1768 * background dirty threshold
1769 */
1770 if (work->for_background && !wb_over_bg_thresh(wb))
1771 break;
1772
1773 /*
1774 * Kupdate and background works are special and we want to
1775 * include all inodes that need writing. Livelock avoidance is
1776 * handled by these works yielding to any other work so we are
1777 * safe.
1778 */
1779 if (work->for_kupdate) {
1780 oldest_jif = jiffies -
1781 msecs_to_jiffies(dirty_expire_interval * 10);
1782 } else if (work->for_background)
1783 oldest_jif = jiffies;
1784
1785 trace_writeback_start(wb, work);
1786 if (list_empty(&wb->b_io))
1787 queue_io(wb, work);
1788 if (work->sb)
1789 progress = writeback_sb_inodes(work->sb, wb, work);
1790 else
1791 progress = __writeback_inodes_wb(wb, work);
1792 trace_writeback_written(wb, work);
1793
1794 wb_update_bandwidth(wb, wb_start);
1795
1796 /*
1797 * Did we write something? Try for more
1798 *
1799 * Dirty inodes are moved to b_io for writeback in batches.
1800 * The completion of the current batch does not necessarily
1801 * mean the overall work is done. So we keep looping as long
1802 * as made some progress on cleaning pages or inodes.
1803 */
1804 if (progress)
1805 continue;
1806 /*
1807 * No more inodes for IO, bail
1808 */
1809 if (list_empty(&wb->b_more_io))
1810 break;
1811 /*
1812 * Nothing written. Wait for some inode to
1813 * become available for writeback. Otherwise
1814 * we'll just busyloop.
1815 */
1816 trace_writeback_wait(wb, work);
1817 inode = wb_inode(wb->b_more_io.prev);
1818 spin_lock(&inode->i_lock);
1819 spin_unlock(&wb->list_lock);
1820 /* This function drops i_lock... */
1821 inode_sleep_on_writeback(inode);
1822 spin_lock(&wb->list_lock);
1823 }
1824 spin_unlock(&wb->list_lock);
1825 blk_finish_plug(&plug);
1826
1827 return nr_pages - work->nr_pages;
1828}
1829
1830/*
1831 * Return the next wb_writeback_work struct that hasn't been processed yet.
1832 */
1833static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1834{
1835 struct wb_writeback_work *work = NULL;
1836
1837 spin_lock_bh(&wb->work_lock);
1838 if (!list_empty(&wb->work_list)) {
1839 work = list_entry(wb->work_list.next,
1840 struct wb_writeback_work, list);
1841 list_del_init(&work->list);
1842 }
1843 spin_unlock_bh(&wb->work_lock);
1844 return work;
1845}
1846
1847static long wb_check_background_flush(struct bdi_writeback *wb)
1848{
1849 if (wb_over_bg_thresh(wb)) {
1850
1851 struct wb_writeback_work work = {
1852 .nr_pages = LONG_MAX,
1853 .sync_mode = WB_SYNC_NONE,
1854 .for_background = 1,
1855 .range_cyclic = 1,
1856 .reason = WB_REASON_BACKGROUND,
1857 };
1858
1859 return wb_writeback(wb, &work);
1860 }
1861
1862 return 0;
1863}
1864
1865static long wb_check_old_data_flush(struct bdi_writeback *wb)
1866{
1867 unsigned long expired;
1868 long nr_pages;
1869
1870 /*
1871 * When set to zero, disable periodic writeback
1872 */
1873 if (!dirty_writeback_interval)
1874 return 0;
1875
1876 expired = wb->last_old_flush +
1877 msecs_to_jiffies(dirty_writeback_interval * 10);
1878 if (time_before(jiffies, expired))
1879 return 0;
1880
1881 wb->last_old_flush = jiffies;
1882 nr_pages = get_nr_dirty_pages();
1883
1884 if (nr_pages) {
1885 struct wb_writeback_work work = {
1886 .nr_pages = nr_pages,
1887 .sync_mode = WB_SYNC_NONE,
1888 .for_kupdate = 1,
1889 .range_cyclic = 1,
1890 .reason = WB_REASON_PERIODIC,
1891 };
1892
1893 return wb_writeback(wb, &work);
1894 }
1895
1896 return 0;
1897}
1898
1899static long wb_check_start_all(struct bdi_writeback *wb)
1900{
1901 long nr_pages;
1902
1903 if (!test_bit(WB_start_all, &wb->state))
1904 return 0;
1905
1906 nr_pages = get_nr_dirty_pages();
1907 if (nr_pages) {
1908 struct wb_writeback_work work = {
1909 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
1910 .sync_mode = WB_SYNC_NONE,
1911 .range_cyclic = 1,
1912 .reason = wb->start_all_reason,
1913 };
1914
1915 nr_pages = wb_writeback(wb, &work);
1916 }
1917
1918 clear_bit(WB_start_all, &wb->state);
1919 return nr_pages;
1920}
1921
1922
1923/*
1924 * Retrieve work items and do the writeback they describe
1925 */
1926static long wb_do_writeback(struct bdi_writeback *wb)
1927{
1928 struct wb_writeback_work *work;
1929 long wrote = 0;
1930
1931 set_bit(WB_writeback_running, &wb->state);
1932 while ((work = get_next_work_item(wb)) != NULL) {
1933 trace_writeback_exec(wb, work);
1934 wrote += wb_writeback(wb, work);
1935 finish_writeback_work(wb, work);
1936 }
1937
1938 /*
1939 * Check for a flush-everything request
1940 */
1941 wrote += wb_check_start_all(wb);
1942
1943 /*
1944 * Check for periodic writeback, kupdated() style
1945 */
1946 wrote += wb_check_old_data_flush(wb);
1947 wrote += wb_check_background_flush(wb);
1948 clear_bit(WB_writeback_running, &wb->state);
1949
1950 return wrote;
1951}
1952
1953/*
1954 * Handle writeback of dirty data for the device backed by this bdi. Also
1955 * reschedules periodically and does kupdated style flushing.
1956 */
1957void wb_workfn(struct work_struct *work)
1958{
1959 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1960 struct bdi_writeback, dwork);
1961 long pages_written;
1962
1963 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1964 current->flags |= PF_SWAPWRITE;
1965
1966 if (likely(!current_is_workqueue_rescuer() ||
1967 !test_bit(WB_registered, &wb->state))) {
1968 /*
1969 * The normal path. Keep writing back @wb until its
1970 * work_list is empty. Note that this path is also taken
1971 * if @wb is shutting down even when we're running off the
1972 * rescuer as work_list needs to be drained.
1973 */
1974 do {
1975 pages_written = wb_do_writeback(wb);
1976 trace_writeback_pages_written(pages_written);
1977 } while (!list_empty(&wb->work_list));
1978 } else {
1979 /*
1980 * bdi_wq can't get enough workers and we're running off
1981 * the emergency worker. Don't hog it. Hopefully, 1024 is
1982 * enough for efficient IO.
1983 */
1984 pages_written = writeback_inodes_wb(wb, 1024,
1985 WB_REASON_FORKER_THREAD);
1986 trace_writeback_pages_written(pages_written);
1987 }
1988
1989 if (!list_empty(&wb->work_list))
1990 wb_wakeup(wb);
1991 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1992 wb_wakeup_delayed(wb);
1993
1994 current->flags &= ~PF_SWAPWRITE;
1995}
1996
1997/*
1998 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
1999 * write back the whole world.
2000 */
2001static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2002 enum wb_reason reason)
2003{
2004 struct bdi_writeback *wb;
2005
2006 if (!bdi_has_dirty_io(bdi))
2007 return;
2008
2009 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2010 wb_start_writeback(wb, reason);
2011}
2012
2013void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2014 enum wb_reason reason)
2015{
2016 rcu_read_lock();
2017 __wakeup_flusher_threads_bdi(bdi, reason);
2018 rcu_read_unlock();
2019}
2020
2021/*
2022 * Wakeup the flusher threads to start writeback of all currently dirty pages
2023 */
2024void wakeup_flusher_threads(enum wb_reason reason)
2025{
2026 struct backing_dev_info *bdi;
2027
2028 /*
2029 * If we are expecting writeback progress we must submit plugged IO.
2030 */
2031 if (blk_needs_flush_plug(current))
2032 blk_schedule_flush_plug(current);
2033
2034 rcu_read_lock();
2035 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2036 __wakeup_flusher_threads_bdi(bdi, reason);
2037 rcu_read_unlock();
2038}
2039
2040/*
2041 * Wake up bdi's periodically to make sure dirtytime inodes gets
2042 * written back periodically. We deliberately do *not* check the
2043 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2044 * kernel to be constantly waking up once there are any dirtytime
2045 * inodes on the system. So instead we define a separate delayed work
2046 * function which gets called much more rarely. (By default, only
2047 * once every 12 hours.)
2048 *
2049 * If there is any other write activity going on in the file system,
2050 * this function won't be necessary. But if the only thing that has
2051 * happened on the file system is a dirtytime inode caused by an atime
2052 * update, we need this infrastructure below to make sure that inode
2053 * eventually gets pushed out to disk.
2054 */
2055static void wakeup_dirtytime_writeback(struct work_struct *w);
2056static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2057
2058static void wakeup_dirtytime_writeback(struct work_struct *w)
2059{
2060 struct backing_dev_info *bdi;
2061
2062 rcu_read_lock();
2063 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2064 struct bdi_writeback *wb;
2065
2066 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2067 if (!list_empty(&wb->b_dirty_time))
2068 wb_wakeup(wb);
2069 }
2070 rcu_read_unlock();
2071 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2072}
2073
2074static int __init start_dirtytime_writeback(void)
2075{
2076 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2077 return 0;
2078}
2079__initcall(start_dirtytime_writeback);
2080
2081int dirtytime_interval_handler(struct ctl_table *table, int write,
2082 void __user *buffer, size_t *lenp, loff_t *ppos)
2083{
2084 int ret;
2085
2086 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2087 if (ret == 0 && write)
2088 mod_delayed_work(system_wq, &dirtytime_work, 0);
2089 return ret;
2090}
2091
2092static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2093{
2094 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2095 struct dentry *dentry;
2096 const char *name = "?";
2097
2098 dentry = d_find_alias(inode);
2099 if (dentry) {
2100 spin_lock(&dentry->d_lock);
2101 name = (const char *) dentry->d_name.name;
2102 }
2103 printk(KERN_DEBUG
2104 "%s(%d): dirtied inode %lu (%s) on %s\n",
2105 current->comm, task_pid_nr(current), inode->i_ino,
2106 name, inode->i_sb->s_id);
2107 if (dentry) {
2108 spin_unlock(&dentry->d_lock);
2109 dput(dentry);
2110 }
2111 }
2112}
2113
2114/**
2115 * __mark_inode_dirty - internal function
2116 *
2117 * @inode: inode to mark
2118 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2119 *
2120 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2121 * mark_inode_dirty_sync.
2122 *
2123 * Put the inode on the super block's dirty list.
2124 *
2125 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2126 * dirty list only if it is hashed or if it refers to a blockdev.
2127 * If it was not hashed, it will never be added to the dirty list
2128 * even if it is later hashed, as it will have been marked dirty already.
2129 *
2130 * In short, make sure you hash any inodes _before_ you start marking
2131 * them dirty.
2132 *
2133 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2134 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2135 * the kernel-internal blockdev inode represents the dirtying time of the
2136 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2137 * page->mapping->host, so the page-dirtying time is recorded in the internal
2138 * blockdev inode.
2139 */
2140void __mark_inode_dirty(struct inode *inode, int flags)
2141{
2142 struct super_block *sb = inode->i_sb;
2143 int dirtytime;
2144
2145 trace_writeback_mark_inode_dirty(inode, flags);
2146
2147 /*
2148 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2149 * dirty the inode itself
2150 */
2151 if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) {
2152 trace_writeback_dirty_inode_start(inode, flags);
2153
2154 if (sb->s_op->dirty_inode)
2155 sb->s_op->dirty_inode(inode, flags);
2156
2157 trace_writeback_dirty_inode(inode, flags);
2158 }
2159 if (flags & I_DIRTY_INODE)
2160 flags &= ~I_DIRTY_TIME;
2161 dirtytime = flags & I_DIRTY_TIME;
2162
2163 /*
2164 * Paired with smp_mb() in __writeback_single_inode() for the
2165 * following lockless i_state test. See there for details.
2166 */
2167 smp_mb();
2168
2169 if (((inode->i_state & flags) == flags) ||
2170 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2171 return;
2172
2173 if (unlikely(block_dump))
2174 block_dump___mark_inode_dirty(inode);
2175
2176 spin_lock(&inode->i_lock);
2177 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2178 goto out_unlock_inode;
2179 if ((inode->i_state & flags) != flags) {
2180 const int was_dirty = inode->i_state & I_DIRTY;
2181
2182 inode_attach_wb(inode, NULL);
2183
2184 if (flags & I_DIRTY_INODE)
2185 inode->i_state &= ~I_DIRTY_TIME;
2186 inode->i_state |= flags;
2187
2188 /*
2189 * If the inode is being synced, just update its dirty state.
2190 * The unlocker will place the inode on the appropriate
2191 * superblock list, based upon its state.
2192 */
2193 if (inode->i_state & I_SYNC)
2194 goto out_unlock_inode;
2195
2196 /*
2197 * Only add valid (hashed) inodes to the superblock's
2198 * dirty list. Add blockdev inodes as well.
2199 */
2200 if (!S_ISBLK(inode->i_mode)) {
2201 if (inode_unhashed(inode))
2202 goto out_unlock_inode;
2203 }
2204 if (inode->i_state & I_FREEING)
2205 goto out_unlock_inode;
2206
2207 /*
2208 * If the inode was already on b_dirty/b_io/b_more_io, don't
2209 * reposition it (that would break b_dirty time-ordering).
2210 */
2211 if (!was_dirty) {
2212 struct bdi_writeback *wb;
2213 struct list_head *dirty_list;
2214 bool wakeup_bdi = false;
2215
2216 wb = locked_inode_to_wb_and_lock_list(inode);
2217
2218 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2219 !test_bit(WB_registered, &wb->state),
2220 "bdi-%s not registered\n", wb->bdi->name);
2221
2222 inode->dirtied_when = jiffies;
2223 if (dirtytime)
2224 inode->dirtied_time_when = jiffies;
2225
2226 if (inode->i_state & I_DIRTY)
2227 dirty_list = &wb->b_dirty;
2228 else
2229 dirty_list = &wb->b_dirty_time;
2230
2231 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2232 dirty_list);
2233
2234 spin_unlock(&wb->list_lock);
2235 trace_writeback_dirty_inode_enqueue(inode);
2236
2237 /*
2238 * If this is the first dirty inode for this bdi,
2239 * we have to wake-up the corresponding bdi thread
2240 * to make sure background write-back happens
2241 * later.
2242 */
2243 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2244 wb_wakeup_delayed(wb);
2245 return;
2246 }
2247 }
2248out_unlock_inode:
2249 spin_unlock(&inode->i_lock);
2250}
2251EXPORT_SYMBOL(__mark_inode_dirty);
2252
2253/*
2254 * The @s_sync_lock is used to serialise concurrent sync operations
2255 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2256 * Concurrent callers will block on the s_sync_lock rather than doing contending
2257 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2258 * has been issued up to the time this function is enter is guaranteed to be
2259 * completed by the time we have gained the lock and waited for all IO that is
2260 * in progress regardless of the order callers are granted the lock.
2261 */
2262static void wait_sb_inodes(struct super_block *sb)
2263{
2264 LIST_HEAD(sync_list);
2265
2266 /*
2267 * We need to be protected against the filesystem going from
2268 * r/o to r/w or vice versa.
2269 */
2270 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2271
2272 mutex_lock(&sb->s_sync_lock);
2273
2274 /*
2275 * Splice the writeback list onto a temporary list to avoid waiting on
2276 * inodes that have started writeback after this point.
2277 *
2278 * Use rcu_read_lock() to keep the inodes around until we have a
2279 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2280 * the local list because inodes can be dropped from either by writeback
2281 * completion.
2282 */
2283 rcu_read_lock();
2284 spin_lock_irq(&sb->s_inode_wblist_lock);
2285 list_splice_init(&sb->s_inodes_wb, &sync_list);
2286
2287 /*
2288 * Data integrity sync. Must wait for all pages under writeback, because
2289 * there may have been pages dirtied before our sync call, but which had
2290 * writeout started before we write it out. In which case, the inode
2291 * may not be on the dirty list, but we still have to wait for that
2292 * writeout.
2293 */
2294 while (!list_empty(&sync_list)) {
2295 struct inode *inode = list_first_entry(&sync_list, struct inode,
2296 i_wb_list);
2297 struct address_space *mapping = inode->i_mapping;
2298
2299 /*
2300 * Move each inode back to the wb list before we drop the lock
2301 * to preserve consistency between i_wb_list and the mapping
2302 * writeback tag. Writeback completion is responsible to remove
2303 * the inode from either list once the writeback tag is cleared.
2304 */
2305 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2306
2307 /*
2308 * The mapping can appear untagged while still on-list since we
2309 * do not have the mapping lock. Skip it here, wb completion
2310 * will remove it.
2311 */
2312 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2313 continue;
2314
2315 spin_unlock_irq(&sb->s_inode_wblist_lock);
2316
2317 spin_lock(&inode->i_lock);
2318 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2319 spin_unlock(&inode->i_lock);
2320
2321 spin_lock_irq(&sb->s_inode_wblist_lock);
2322 continue;
2323 }
2324 __iget(inode);
2325 spin_unlock(&inode->i_lock);
2326 rcu_read_unlock();
2327
2328 /*
2329 * We keep the error status of individual mapping so that
2330 * applications can catch the writeback error using fsync(2).
2331 * See filemap_fdatawait_keep_errors() for details.
2332 */
2333 filemap_fdatawait_keep_errors(mapping);
2334
2335 cond_resched();
2336
2337 iput(inode);
2338
2339 rcu_read_lock();
2340 spin_lock_irq(&sb->s_inode_wblist_lock);
2341 }
2342 spin_unlock_irq(&sb->s_inode_wblist_lock);
2343 rcu_read_unlock();
2344 mutex_unlock(&sb->s_sync_lock);
2345}
2346
2347static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2348 enum wb_reason reason, bool skip_if_busy)
2349{
2350 DEFINE_WB_COMPLETION_ONSTACK(done);
2351 struct wb_writeback_work work = {
2352 .sb = sb,
2353 .sync_mode = WB_SYNC_NONE,
2354 .tagged_writepages = 1,
2355 .done = &done,
2356 .nr_pages = nr,
2357 .reason = reason,
2358 };
2359 struct backing_dev_info *bdi = sb->s_bdi;
2360
2361 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2362 return;
2363 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2364
2365 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2366 wb_wait_for_completion(bdi, &done);
2367}
2368
2369/**
2370 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2371 * @sb: the superblock
2372 * @nr: the number of pages to write
2373 * @reason: reason why some writeback work initiated
2374 *
2375 * Start writeback on some inodes on this super_block. No guarantees are made
2376 * on how many (if any) will be written, and this function does not wait
2377 * for IO completion of submitted IO.
2378 */
2379void writeback_inodes_sb_nr(struct super_block *sb,
2380 unsigned long nr,
2381 enum wb_reason reason)
2382{
2383 __writeback_inodes_sb_nr(sb, nr, reason, false);
2384}
2385EXPORT_SYMBOL(writeback_inodes_sb_nr);
2386
2387/**
2388 * writeback_inodes_sb - writeback dirty inodes from given super_block
2389 * @sb: the superblock
2390 * @reason: reason why some writeback work was initiated
2391 *
2392 * Start writeback on some inodes on this super_block. No guarantees are made
2393 * on how many (if any) will be written, and this function does not wait
2394 * for IO completion of submitted IO.
2395 */
2396void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2397{
2398 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2399}
2400EXPORT_SYMBOL(writeback_inodes_sb);
2401
2402/**
2403 * try_to_writeback_inodes_sb - try to start writeback if none underway
2404 * @sb: the superblock
2405 * @reason: reason why some writeback work was initiated
2406 *
2407 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2408 */
2409void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2410{
2411 if (!down_read_trylock(&sb->s_umount))
2412 return;
2413
2414 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2415 up_read(&sb->s_umount);
2416}
2417EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2418
2419/**
2420 * sync_inodes_sb - sync sb inode pages
2421 * @sb: the superblock
2422 *
2423 * This function writes and waits on any dirty inode belonging to this
2424 * super_block.
2425 */
2426void sync_inodes_sb(struct super_block *sb)
2427{
2428 DEFINE_WB_COMPLETION_ONSTACK(done);
2429 struct wb_writeback_work work = {
2430 .sb = sb,
2431 .sync_mode = WB_SYNC_ALL,
2432 .nr_pages = LONG_MAX,
2433 .range_cyclic = 0,
2434 .done = &done,
2435 .reason = WB_REASON_SYNC,
2436 .for_sync = 1,
2437 };
2438 struct backing_dev_info *bdi = sb->s_bdi;
2439
2440 /*
2441 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2442 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2443 * bdi_has_dirty() need to be written out too.
2444 */
2445 if (bdi == &noop_backing_dev_info)
2446 return;
2447 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2448
2449 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2450 bdi_down_write_wb_switch_rwsem(bdi);
2451 bdi_split_work_to_wbs(bdi, &work, false);
2452 wb_wait_for_completion(bdi, &done);
2453 bdi_up_write_wb_switch_rwsem(bdi);
2454
2455 wait_sb_inodes(sb);
2456}
2457EXPORT_SYMBOL(sync_inodes_sb);
2458
2459/**
2460 * write_inode_now - write an inode to disk
2461 * @inode: inode to write to disk
2462 * @sync: whether the write should be synchronous or not
2463 *
2464 * This function commits an inode to disk immediately if it is dirty. This is
2465 * primarily needed by knfsd.
2466 *
2467 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2468 */
2469int write_inode_now(struct inode *inode, int sync)
2470{
2471 struct writeback_control wbc = {
2472 .nr_to_write = LONG_MAX,
2473 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2474 .range_start = 0,
2475 .range_end = LLONG_MAX,
2476 };
2477
2478 if (!mapping_cap_writeback_dirty(inode->i_mapping))
2479 wbc.nr_to_write = 0;
2480
2481 might_sleep();
2482 return writeback_single_inode(inode, &wbc);
2483}
2484EXPORT_SYMBOL(write_inode_now);
2485
2486/**
2487 * sync_inode - write an inode and its pages to disk.
2488 * @inode: the inode to sync
2489 * @wbc: controls the writeback mode
2490 *
2491 * sync_inode() will write an inode and its pages to disk. It will also
2492 * correctly update the inode on its superblock's dirty inode lists and will
2493 * update inode->i_state.
2494 *
2495 * The caller must have a ref on the inode.
2496 */
2497int sync_inode(struct inode *inode, struct writeback_control *wbc)
2498{
2499 return writeback_single_inode(inode, wbc);
2500}
2501EXPORT_SYMBOL(sync_inode);
2502
2503/**
2504 * sync_inode_metadata - write an inode to disk
2505 * @inode: the inode to sync
2506 * @wait: wait for I/O to complete.
2507 *
2508 * Write an inode to disk and adjust its dirty state after completion.
2509 *
2510 * Note: only writes the actual inode, no associated data or other metadata.
2511 */
2512int sync_inode_metadata(struct inode *inode, int wait)
2513{
2514 struct writeback_control wbc = {
2515 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2516 .nr_to_write = 0, /* metadata-only */
2517 };
2518
2519 return sync_inode(inode, &wbc);
2520}
2521EXPORT_SYMBOL(sync_inode_metadata);
2522