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