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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/fs-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
11 *
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
15 */
16
17#include <linux/kernel.h>
18#include <linux/export.h>
19#include <linux/spinlock.h>
20#include <linux/slab.h>
21#include <linux/sched.h>
22#include <linux/fs.h>
23#include <linux/mm.h>
24#include <linux/pagemap.h>
25#include <linux/kthread.h>
26#include <linux/writeback.h>
27#include <linux/blkdev.h>
28#include <linux/backing-dev.h>
29#include <linux/tracepoint.h>
30#include <linux/device.h>
31#include <linux/memcontrol.h>
32#include "internal.h"
33
34/*
35 * 4MB minimal write chunk size
36 */
37#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38
39/*
40 * Passed into wb_writeback(), essentially a subset of writeback_control
41 */
42struct wb_writeback_work {
43 long nr_pages;
44 struct super_block *sb;
45 enum writeback_sync_modes sync_mode;
46 unsigned int tagged_writepages:1;
47 unsigned int for_kupdate:1;
48 unsigned int range_cyclic:1;
49 unsigned int for_background:1;
50 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
51 unsigned int auto_free:1; /* free on completion */
52 enum wb_reason reason; /* why was writeback initiated? */
53
54 struct list_head list; /* pending work list */
55 struct wb_completion *done; /* set if the caller waits */
56};
57
58/*
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
67 */
68unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
70static inline struct inode *wb_inode(struct list_head *head)
71{
72 return list_entry(head, struct inode, i_io_list);
73}
74
75/*
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
79 */
80#define CREATE_TRACE_POINTS
81#include <trace/events/writeback.h>
82
83EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
85static bool wb_io_lists_populated(struct bdi_writeback *wb)
86{
87 if (wb_has_dirty_io(wb)) {
88 return false;
89 } else {
90 set_bit(WB_has_dirty_io, &wb->state);
91 WARN_ON_ONCE(!wb->avg_write_bandwidth);
92 atomic_long_add(wb->avg_write_bandwidth,
93 &wb->bdi->tot_write_bandwidth);
94 return true;
95 }
96}
97
98static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99{
100 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102 clear_bit(WB_has_dirty_io, &wb->state);
103 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 &wb->bdi->tot_write_bandwidth) < 0);
105 }
106}
107
108/**
109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
113 *
114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
117 */
118static bool inode_io_list_move_locked(struct inode *inode,
119 struct bdi_writeback *wb,
120 struct list_head *head)
121{
122 assert_spin_locked(&wb->list_lock);
123
124 list_move(&inode->i_io_list, head);
125
126 /* dirty_time doesn't count as dirty_io until expiration */
127 if (head != &wb->b_dirty_time)
128 return wb_io_lists_populated(wb);
129
130 wb_io_lists_depopulated(wb);
131 return false;
132}
133
134static void wb_wakeup(struct bdi_writeback *wb)
135{
136 spin_lock_bh(&wb->work_lock);
137 if (test_bit(WB_registered, &wb->state))
138 mod_delayed_work(bdi_wq, &wb->dwork, 0);
139 spin_unlock_bh(&wb->work_lock);
140}
141
142static void finish_writeback_work(struct bdi_writeback *wb,
143 struct wb_writeback_work *work)
144{
145 struct wb_completion *done = work->done;
146
147 if (work->auto_free)
148 kfree(work);
149 if (done) {
150 wait_queue_head_t *waitq = done->waitq;
151
152 /* @done can't be accessed after the following dec */
153 if (atomic_dec_and_test(&done->cnt))
154 wake_up_all(waitq);
155 }
156}
157
158static void wb_queue_work(struct bdi_writeback *wb,
159 struct wb_writeback_work *work)
160{
161 trace_writeback_queue(wb, work);
162
163 if (work->done)
164 atomic_inc(&work->done->cnt);
165
166 spin_lock_bh(&wb->work_lock);
167
168 if (test_bit(WB_registered, &wb->state)) {
169 list_add_tail(&work->list, &wb->work_list);
170 mod_delayed_work(bdi_wq, &wb->dwork, 0);
171 } else
172 finish_writeback_work(wb, work);
173
174 spin_unlock_bh(&wb->work_lock);
175}
176
177/**
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
179 * @done: target wb_completion
180 *
181 * Wait for one or more work items issued to @bdi with their ->done field
182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION(). This function returns after all such work items
184 * are completed. Work items which are waited upon aren't freed
185 * automatically on completion.
186 */
187void wb_wait_for_completion(struct wb_completion *done)
188{
189 atomic_dec(&done->cnt); /* put down the initial count */
190 wait_event(*done->waitq, !atomic_read(&done->cnt));
191}
192
193#ifdef CONFIG_CGROUP_WRITEBACK
194
195/*
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
197 * how they're used.
198 *
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy. All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
202 *
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported. As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
208 *
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch. The recording is quantized
211 * to 16 slots. To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
213 */
214#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
215#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
216#define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
217#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
218
219#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
220#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221 /* each slot's duration is 2s / 16 */
222#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
223 /* if foreign slots >= 8, switch */
224#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
225 /* one round can affect upto 5 slots */
226#define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
227
228/*
229 * Maximum inodes per isw. A specific value has been chosen to make
230 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
231 */
232#define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
233 / sizeof(struct inode *))
234
235static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
236static struct workqueue_struct *isw_wq;
237
238void __inode_attach_wb(struct inode *inode, struct page *page)
239{
240 struct backing_dev_info *bdi = inode_to_bdi(inode);
241 struct bdi_writeback *wb = NULL;
242
243 if (inode_cgwb_enabled(inode)) {
244 struct cgroup_subsys_state *memcg_css;
245
246 if (page) {
247 memcg_css = mem_cgroup_css_from_page(page);
248 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
249 } else {
250 /* must pin memcg_css, see wb_get_create() */
251 memcg_css = task_get_css(current, memory_cgrp_id);
252 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
253 css_put(memcg_css);
254 }
255 }
256
257 if (!wb)
258 wb = &bdi->wb;
259
260 /*
261 * There may be multiple instances of this function racing to
262 * update the same inode. Use cmpxchg() to tell the winner.
263 */
264 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
265 wb_put(wb);
266}
267EXPORT_SYMBOL_GPL(__inode_attach_wb);
268
269/**
270 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
271 * @inode: inode of interest with i_lock held
272 * @wb: target bdi_writeback
273 *
274 * Remove the inode from wb's io lists and if necessarily put onto b_attached
275 * list. Only inodes attached to cgwb's are kept on this list.
276 */
277static void inode_cgwb_move_to_attached(struct inode *inode,
278 struct bdi_writeback *wb)
279{
280 assert_spin_locked(&wb->list_lock);
281 assert_spin_locked(&inode->i_lock);
282
283 inode->i_state &= ~I_SYNC_QUEUED;
284 if (wb != &wb->bdi->wb)
285 list_move(&inode->i_io_list, &wb->b_attached);
286 else
287 list_del_init(&inode->i_io_list);
288 wb_io_lists_depopulated(wb);
289}
290
291/**
292 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
293 * @inode: inode of interest with i_lock held
294 *
295 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
296 * held on entry and is released on return. The returned wb is guaranteed
297 * to stay @inode's associated wb until its list_lock is released.
298 */
299static struct bdi_writeback *
300locked_inode_to_wb_and_lock_list(struct inode *inode)
301 __releases(&inode->i_lock)
302 __acquires(&wb->list_lock)
303{
304 while (true) {
305 struct bdi_writeback *wb = inode_to_wb(inode);
306
307 /*
308 * inode_to_wb() association is protected by both
309 * @inode->i_lock and @wb->list_lock but list_lock nests
310 * outside i_lock. Drop i_lock and verify that the
311 * association hasn't changed after acquiring list_lock.
312 */
313 wb_get(wb);
314 spin_unlock(&inode->i_lock);
315 spin_lock(&wb->list_lock);
316
317 /* i_wb may have changed inbetween, can't use inode_to_wb() */
318 if (likely(wb == inode->i_wb)) {
319 wb_put(wb); /* @inode already has ref */
320 return wb;
321 }
322
323 spin_unlock(&wb->list_lock);
324 wb_put(wb);
325 cpu_relax();
326 spin_lock(&inode->i_lock);
327 }
328}
329
330/**
331 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
332 * @inode: inode of interest
333 *
334 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
335 * on entry.
336 */
337static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
338 __acquires(&wb->list_lock)
339{
340 spin_lock(&inode->i_lock);
341 return locked_inode_to_wb_and_lock_list(inode);
342}
343
344struct inode_switch_wbs_context {
345 struct rcu_work work;
346
347 /*
348 * Multiple inodes can be switched at once. The switching procedure
349 * consists of two parts, separated by a RCU grace period. To make
350 * sure that the second part is executed for each inode gone through
351 * the first part, all inode pointers are placed into a NULL-terminated
352 * array embedded into struct inode_switch_wbs_context. Otherwise
353 * an inode could be left in a non-consistent state.
354 */
355 struct bdi_writeback *new_wb;
356 struct inode *inodes[];
357};
358
359static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
360{
361 down_write(&bdi->wb_switch_rwsem);
362}
363
364static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
365{
366 up_write(&bdi->wb_switch_rwsem);
367}
368
369static bool inode_do_switch_wbs(struct inode *inode,
370 struct bdi_writeback *old_wb,
371 struct bdi_writeback *new_wb)
372{
373 struct address_space *mapping = inode->i_mapping;
374 XA_STATE(xas, &mapping->i_pages, 0);
375 struct page *page;
376 bool switched = false;
377
378 spin_lock(&inode->i_lock);
379 xa_lock_irq(&mapping->i_pages);
380
381 /*
382 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
383 * path owns the inode and we shouldn't modify ->i_io_list.
384 */
385 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
386 goto skip_switch;
387
388 trace_inode_switch_wbs(inode, old_wb, new_wb);
389
390 /*
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393 * pages actually under writeback.
394 */
395 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
396 if (PageDirty(page)) {
397 dec_wb_stat(old_wb, WB_RECLAIMABLE);
398 inc_wb_stat(new_wb, WB_RECLAIMABLE);
399 }
400 }
401
402 xas_set(&xas, 0);
403 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
404 WARN_ON_ONCE(!PageWriteback(page));
405 dec_wb_stat(old_wb, WB_WRITEBACK);
406 inc_wb_stat(new_wb, WB_WRITEBACK);
407 }
408
409 wb_get(new_wb);
410
411 /*
412 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
413 * the specific list @inode was on is ignored and the @inode is put on
414 * ->b_dirty which is always correct including from ->b_dirty_time.
415 * The transfer preserves @inode->dirtied_when ordering. If the @inode
416 * was clean, it means it was on the b_attached list, so move it onto
417 * the b_attached list of @new_wb.
418 */
419 if (!list_empty(&inode->i_io_list)) {
420 inode->i_wb = new_wb;
421
422 if (inode->i_state & I_DIRTY_ALL) {
423 struct inode *pos;
424
425 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
426 if (time_after_eq(inode->dirtied_when,
427 pos->dirtied_when))
428 break;
429 inode_io_list_move_locked(inode, new_wb,
430 pos->i_io_list.prev);
431 } else {
432 inode_cgwb_move_to_attached(inode, new_wb);
433 }
434 } else {
435 inode->i_wb = new_wb;
436 }
437
438 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
439 inode->i_wb_frn_winner = 0;
440 inode->i_wb_frn_avg_time = 0;
441 inode->i_wb_frn_history = 0;
442 switched = true;
443skip_switch:
444 /*
445 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
446 * ensures that the new wb is visible if they see !I_WB_SWITCH.
447 */
448 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
449
450 xa_unlock_irq(&mapping->i_pages);
451 spin_unlock(&inode->i_lock);
452
453 return switched;
454}
455
456static void inode_switch_wbs_work_fn(struct work_struct *work)
457{
458 struct inode_switch_wbs_context *isw =
459 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
460 struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
461 struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
462 struct bdi_writeback *new_wb = isw->new_wb;
463 unsigned long nr_switched = 0;
464 struct inode **inodep;
465
466 /*
467 * If @inode switches cgwb membership while sync_inodes_sb() is
468 * being issued, sync_inodes_sb() might miss it. Synchronize.
469 */
470 down_read(&bdi->wb_switch_rwsem);
471
472 /*
473 * By the time control reaches here, RCU grace period has passed
474 * since I_WB_SWITCH assertion and all wb stat update transactions
475 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
476 * synchronizing against the i_pages lock.
477 *
478 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
479 * gives us exclusion against all wb related operations on @inode
480 * including IO list manipulations and stat updates.
481 */
482 if (old_wb < new_wb) {
483 spin_lock(&old_wb->list_lock);
484 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
485 } else {
486 spin_lock(&new_wb->list_lock);
487 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
488 }
489
490 for (inodep = isw->inodes; *inodep; inodep++) {
491 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
492 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
493 nr_switched++;
494 }
495
496 spin_unlock(&new_wb->list_lock);
497 spin_unlock(&old_wb->list_lock);
498
499 up_read(&bdi->wb_switch_rwsem);
500
501 if (nr_switched) {
502 wb_wakeup(new_wb);
503 wb_put_many(old_wb, nr_switched);
504 }
505
506 for (inodep = isw->inodes; *inodep; inodep++)
507 iput(*inodep);
508 wb_put(new_wb);
509 kfree(isw);
510 atomic_dec(&isw_nr_in_flight);
511}
512
513static bool inode_prepare_wbs_switch(struct inode *inode,
514 struct bdi_writeback *new_wb)
515{
516 /*
517 * Paired with smp_mb() in cgroup_writeback_umount().
518 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
519 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
520 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
521 */
522 smp_mb();
523
524 if (IS_DAX(inode))
525 return false;
526
527 /* while holding I_WB_SWITCH, no one else can update the association */
528 spin_lock(&inode->i_lock);
529 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
530 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
531 inode_to_wb(inode) == new_wb) {
532 spin_unlock(&inode->i_lock);
533 return false;
534 }
535 inode->i_state |= I_WB_SWITCH;
536 __iget(inode);
537 spin_unlock(&inode->i_lock);
538
539 return true;
540}
541
542/**
543 * inode_switch_wbs - change the wb association of an inode
544 * @inode: target inode
545 * @new_wb_id: ID of the new wb
546 *
547 * Switch @inode's wb association to the wb identified by @new_wb_id. The
548 * switching is performed asynchronously and may fail silently.
549 */
550static void inode_switch_wbs(struct inode *inode, int new_wb_id)
551{
552 struct backing_dev_info *bdi = inode_to_bdi(inode);
553 struct cgroup_subsys_state *memcg_css;
554 struct inode_switch_wbs_context *isw;
555
556 /* noop if seems to be already in progress */
557 if (inode->i_state & I_WB_SWITCH)
558 return;
559
560 /* avoid queueing a new switch if too many are already in flight */
561 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
562 return;
563
564 isw = kzalloc(sizeof(*isw) + 2 * sizeof(struct inode *), GFP_ATOMIC);
565 if (!isw)
566 return;
567
568 atomic_inc(&isw_nr_in_flight);
569
570 /* find and pin the new wb */
571 rcu_read_lock();
572 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
573 if (memcg_css && !css_tryget(memcg_css))
574 memcg_css = NULL;
575 rcu_read_unlock();
576 if (!memcg_css)
577 goto out_free;
578
579 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
580 css_put(memcg_css);
581 if (!isw->new_wb)
582 goto out_free;
583
584 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
585 goto out_free;
586
587 isw->inodes[0] = inode;
588
589 /*
590 * In addition to synchronizing among switchers, I_WB_SWITCH tells
591 * the RCU protected stat update paths to grab the i_page
592 * lock so that stat transfer can synchronize against them.
593 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
594 */
595 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
596 queue_rcu_work(isw_wq, &isw->work);
597 return;
598
599out_free:
600 atomic_dec(&isw_nr_in_flight);
601 if (isw->new_wb)
602 wb_put(isw->new_wb);
603 kfree(isw);
604}
605
606/**
607 * cleanup_offline_cgwb - detach associated inodes
608 * @wb: target wb
609 *
610 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
611 * to eventually release the dying @wb. Returns %true if not all inodes were
612 * switched and the function has to be restarted.
613 */
614bool cleanup_offline_cgwb(struct bdi_writeback *wb)
615{
616 struct cgroup_subsys_state *memcg_css;
617 struct inode_switch_wbs_context *isw;
618 struct inode *inode;
619 int nr;
620 bool restart = false;
621
622 isw = kzalloc(sizeof(*isw) + WB_MAX_INODES_PER_ISW *
623 sizeof(struct inode *), GFP_KERNEL);
624 if (!isw)
625 return restart;
626
627 atomic_inc(&isw_nr_in_flight);
628
629 for (memcg_css = wb->memcg_css->parent; memcg_css;
630 memcg_css = memcg_css->parent) {
631 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
632 if (isw->new_wb)
633 break;
634 }
635 if (unlikely(!isw->new_wb))
636 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
637
638 nr = 0;
639 spin_lock(&wb->list_lock);
640 list_for_each_entry(inode, &wb->b_attached, i_io_list) {
641 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
642 continue;
643
644 isw->inodes[nr++] = inode;
645
646 if (nr >= WB_MAX_INODES_PER_ISW - 1) {
647 restart = true;
648 break;
649 }
650 }
651 spin_unlock(&wb->list_lock);
652
653 /* no attached inodes? bail out */
654 if (nr == 0) {
655 atomic_dec(&isw_nr_in_flight);
656 wb_put(isw->new_wb);
657 kfree(isw);
658 return restart;
659 }
660
661 /*
662 * In addition to synchronizing among switchers, I_WB_SWITCH tells
663 * the RCU protected stat update paths to grab the i_page
664 * lock so that stat transfer can synchronize against them.
665 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
666 */
667 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
668 queue_rcu_work(isw_wq, &isw->work);
669
670 return restart;
671}
672
673/**
674 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
675 * @wbc: writeback_control of interest
676 * @inode: target inode
677 *
678 * @inode is locked and about to be written back under the control of @wbc.
679 * Record @inode's writeback context into @wbc and unlock the i_lock. On
680 * writeback completion, wbc_detach_inode() should be called. This is used
681 * to track the cgroup writeback context.
682 */
683void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
684 struct inode *inode)
685{
686 if (!inode_cgwb_enabled(inode)) {
687 spin_unlock(&inode->i_lock);
688 return;
689 }
690
691 wbc->wb = inode_to_wb(inode);
692 wbc->inode = inode;
693
694 wbc->wb_id = wbc->wb->memcg_css->id;
695 wbc->wb_lcand_id = inode->i_wb_frn_winner;
696 wbc->wb_tcand_id = 0;
697 wbc->wb_bytes = 0;
698 wbc->wb_lcand_bytes = 0;
699 wbc->wb_tcand_bytes = 0;
700
701 wb_get(wbc->wb);
702 spin_unlock(&inode->i_lock);
703
704 /*
705 * A dying wb indicates that either the blkcg associated with the
706 * memcg changed or the associated memcg is dying. In the first
707 * case, a replacement wb should already be available and we should
708 * refresh the wb immediately. In the second case, trying to
709 * refresh will keep failing.
710 */
711 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
712 inode_switch_wbs(inode, wbc->wb_id);
713}
714EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
715
716/**
717 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
718 * @wbc: writeback_control of the just finished writeback
719 *
720 * To be called after a writeback attempt of an inode finishes and undoes
721 * wbc_attach_and_unlock_inode(). Can be called under any context.
722 *
723 * As concurrent write sharing of an inode is expected to be very rare and
724 * memcg only tracks page ownership on first-use basis severely confining
725 * the usefulness of such sharing, cgroup writeback tracks ownership
726 * per-inode. While the support for concurrent write sharing of an inode
727 * is deemed unnecessary, an inode being written to by different cgroups at
728 * different points in time is a lot more common, and, more importantly,
729 * charging only by first-use can too readily lead to grossly incorrect
730 * behaviors (single foreign page can lead to gigabytes of writeback to be
731 * incorrectly attributed).
732 *
733 * To resolve this issue, cgroup writeback detects the majority dirtier of
734 * an inode and transfers the ownership to it. To avoid unnnecessary
735 * oscillation, the detection mechanism keeps track of history and gives
736 * out the switch verdict only if the foreign usage pattern is stable over
737 * a certain amount of time and/or writeback attempts.
738 *
739 * On each writeback attempt, @wbc tries to detect the majority writer
740 * using Boyer-Moore majority vote algorithm. In addition to the byte
741 * count from the majority voting, it also counts the bytes written for the
742 * current wb and the last round's winner wb (max of last round's current
743 * wb, the winner from two rounds ago, and the last round's majority
744 * candidate). Keeping track of the historical winner helps the algorithm
745 * to semi-reliably detect the most active writer even when it's not the
746 * absolute majority.
747 *
748 * Once the winner of the round is determined, whether the winner is
749 * foreign or not and how much IO time the round consumed is recorded in
750 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
751 * over a certain threshold, the switch verdict is given.
752 */
753void wbc_detach_inode(struct writeback_control *wbc)
754{
755 struct bdi_writeback *wb = wbc->wb;
756 struct inode *inode = wbc->inode;
757 unsigned long avg_time, max_bytes, max_time;
758 u16 history;
759 int max_id;
760
761 if (!wb)
762 return;
763
764 history = inode->i_wb_frn_history;
765 avg_time = inode->i_wb_frn_avg_time;
766
767 /* pick the winner of this round */
768 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
769 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
770 max_id = wbc->wb_id;
771 max_bytes = wbc->wb_bytes;
772 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
773 max_id = wbc->wb_lcand_id;
774 max_bytes = wbc->wb_lcand_bytes;
775 } else {
776 max_id = wbc->wb_tcand_id;
777 max_bytes = wbc->wb_tcand_bytes;
778 }
779
780 /*
781 * Calculate the amount of IO time the winner consumed and fold it
782 * into the running average kept per inode. If the consumed IO
783 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
784 * deciding whether to switch or not. This is to prevent one-off
785 * small dirtiers from skewing the verdict.
786 */
787 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
788 wb->avg_write_bandwidth);
789 if (avg_time)
790 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
791 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
792 else
793 avg_time = max_time; /* immediate catch up on first run */
794
795 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
796 int slots;
797
798 /*
799 * The switch verdict is reached if foreign wb's consume
800 * more than a certain proportion of IO time in a
801 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
802 * history mask where each bit represents one sixteenth of
803 * the period. Determine the number of slots to shift into
804 * history from @max_time.
805 */
806 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
807 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
808 history <<= slots;
809 if (wbc->wb_id != max_id)
810 history |= (1U << slots) - 1;
811
812 if (history)
813 trace_inode_foreign_history(inode, wbc, history);
814
815 /*
816 * Switch if the current wb isn't the consistent winner.
817 * If there are multiple closely competing dirtiers, the
818 * inode may switch across them repeatedly over time, which
819 * is okay. The main goal is avoiding keeping an inode on
820 * the wrong wb for an extended period of time.
821 */
822 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
823 inode_switch_wbs(inode, max_id);
824 }
825
826 /*
827 * Multiple instances of this function may race to update the
828 * following fields but we don't mind occassional inaccuracies.
829 */
830 inode->i_wb_frn_winner = max_id;
831 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
832 inode->i_wb_frn_history = history;
833
834 wb_put(wbc->wb);
835 wbc->wb = NULL;
836}
837EXPORT_SYMBOL_GPL(wbc_detach_inode);
838
839/**
840 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
841 * @wbc: writeback_control of the writeback in progress
842 * @page: page being written out
843 * @bytes: number of bytes being written out
844 *
845 * @bytes from @page are about to written out during the writeback
846 * controlled by @wbc. Keep the book for foreign inode detection. See
847 * wbc_detach_inode().
848 */
849void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
850 size_t bytes)
851{
852 struct cgroup_subsys_state *css;
853 int id;
854
855 /*
856 * pageout() path doesn't attach @wbc to the inode being written
857 * out. This is intentional as we don't want the function to block
858 * behind a slow cgroup. Ultimately, we want pageout() to kick off
859 * regular writeback instead of writing things out itself.
860 */
861 if (!wbc->wb || wbc->no_cgroup_owner)
862 return;
863
864 css = mem_cgroup_css_from_page(page);
865 /* dead cgroups shouldn't contribute to inode ownership arbitration */
866 if (!(css->flags & CSS_ONLINE))
867 return;
868
869 id = css->id;
870
871 if (id == wbc->wb_id) {
872 wbc->wb_bytes += bytes;
873 return;
874 }
875
876 if (id == wbc->wb_lcand_id)
877 wbc->wb_lcand_bytes += bytes;
878
879 /* Boyer-Moore majority vote algorithm */
880 if (!wbc->wb_tcand_bytes)
881 wbc->wb_tcand_id = id;
882 if (id == wbc->wb_tcand_id)
883 wbc->wb_tcand_bytes += bytes;
884 else
885 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
886}
887EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
888
889/**
890 * inode_congested - test whether an inode is congested
891 * @inode: inode to test for congestion (may be NULL)
892 * @cong_bits: mask of WB_[a]sync_congested bits to test
893 *
894 * Tests whether @inode is congested. @cong_bits is the mask of congestion
895 * bits to test and the return value is the mask of set bits.
896 *
897 * If cgroup writeback is enabled for @inode, the congestion state is
898 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
899 * associated with @inode is congested; otherwise, the root wb's congestion
900 * state is used.
901 *
902 * @inode is allowed to be NULL as this function is often called on
903 * mapping->host which is NULL for the swapper space.
904 */
905int inode_congested(struct inode *inode, int cong_bits)
906{
907 /*
908 * Once set, ->i_wb never becomes NULL while the inode is alive.
909 * Start transaction iff ->i_wb is visible.
910 */
911 if (inode && inode_to_wb_is_valid(inode)) {
912 struct bdi_writeback *wb;
913 struct wb_lock_cookie lock_cookie = {};
914 bool congested;
915
916 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
917 congested = wb_congested(wb, cong_bits);
918 unlocked_inode_to_wb_end(inode, &lock_cookie);
919 return congested;
920 }
921
922 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
923}
924EXPORT_SYMBOL_GPL(inode_congested);
925
926/**
927 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
928 * @wb: target bdi_writeback to split @nr_pages to
929 * @nr_pages: number of pages to write for the whole bdi
930 *
931 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
932 * relation to the total write bandwidth of all wb's w/ dirty inodes on
933 * @wb->bdi.
934 */
935static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
936{
937 unsigned long this_bw = wb->avg_write_bandwidth;
938 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
939
940 if (nr_pages == LONG_MAX)
941 return LONG_MAX;
942
943 /*
944 * This may be called on clean wb's and proportional distribution
945 * may not make sense, just use the original @nr_pages in those
946 * cases. In general, we wanna err on the side of writing more.
947 */
948 if (!tot_bw || this_bw >= tot_bw)
949 return nr_pages;
950 else
951 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
952}
953
954/**
955 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
956 * @bdi: target backing_dev_info
957 * @base_work: wb_writeback_work to issue
958 * @skip_if_busy: skip wb's which already have writeback in progress
959 *
960 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
961 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
962 * distributed to the busy wbs according to each wb's proportion in the
963 * total active write bandwidth of @bdi.
964 */
965static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
966 struct wb_writeback_work *base_work,
967 bool skip_if_busy)
968{
969 struct bdi_writeback *last_wb = NULL;
970 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
971 struct bdi_writeback, bdi_node);
972
973 might_sleep();
974restart:
975 rcu_read_lock();
976 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
977 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
978 struct wb_writeback_work fallback_work;
979 struct wb_writeback_work *work;
980 long nr_pages;
981
982 if (last_wb) {
983 wb_put(last_wb);
984 last_wb = NULL;
985 }
986
987 /* SYNC_ALL writes out I_DIRTY_TIME too */
988 if (!wb_has_dirty_io(wb) &&
989 (base_work->sync_mode == WB_SYNC_NONE ||
990 list_empty(&wb->b_dirty_time)))
991 continue;
992 if (skip_if_busy && writeback_in_progress(wb))
993 continue;
994
995 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
996
997 work = kmalloc(sizeof(*work), GFP_ATOMIC);
998 if (work) {
999 *work = *base_work;
1000 work->nr_pages = nr_pages;
1001 work->auto_free = 1;
1002 wb_queue_work(wb, work);
1003 continue;
1004 }
1005
1006 /* alloc failed, execute synchronously using on-stack fallback */
1007 work = &fallback_work;
1008 *work = *base_work;
1009 work->nr_pages = nr_pages;
1010 work->auto_free = 0;
1011 work->done = &fallback_work_done;
1012
1013 wb_queue_work(wb, work);
1014
1015 /*
1016 * Pin @wb so that it stays on @bdi->wb_list. This allows
1017 * continuing iteration from @wb after dropping and
1018 * regrabbing rcu read lock.
1019 */
1020 wb_get(wb);
1021 last_wb = wb;
1022
1023 rcu_read_unlock();
1024 wb_wait_for_completion(&fallback_work_done);
1025 goto restart;
1026 }
1027 rcu_read_unlock();
1028
1029 if (last_wb)
1030 wb_put(last_wb);
1031}
1032
1033/**
1034 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1035 * @bdi_id: target bdi id
1036 * @memcg_id: target memcg css id
1037 * @nr: number of pages to write, 0 for best-effort dirty flushing
1038 * @reason: reason why some writeback work initiated
1039 * @done: target wb_completion
1040 *
1041 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1042 * with the specified parameters.
1043 */
1044int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
1045 enum wb_reason reason, struct wb_completion *done)
1046{
1047 struct backing_dev_info *bdi;
1048 struct cgroup_subsys_state *memcg_css;
1049 struct bdi_writeback *wb;
1050 struct wb_writeback_work *work;
1051 int ret;
1052
1053 /* lookup bdi and memcg */
1054 bdi = bdi_get_by_id(bdi_id);
1055 if (!bdi)
1056 return -ENOENT;
1057
1058 rcu_read_lock();
1059 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1060 if (memcg_css && !css_tryget(memcg_css))
1061 memcg_css = NULL;
1062 rcu_read_unlock();
1063 if (!memcg_css) {
1064 ret = -ENOENT;
1065 goto out_bdi_put;
1066 }
1067
1068 /*
1069 * And find the associated wb. If the wb isn't there already
1070 * there's nothing to flush, don't create one.
1071 */
1072 wb = wb_get_lookup(bdi, memcg_css);
1073 if (!wb) {
1074 ret = -ENOENT;
1075 goto out_css_put;
1076 }
1077
1078 /*
1079 * If @nr is zero, the caller is attempting to write out most of
1080 * the currently dirty pages. Let's take the current dirty page
1081 * count and inflate it by 25% which should be large enough to
1082 * flush out most dirty pages while avoiding getting livelocked by
1083 * concurrent dirtiers.
1084 */
1085 if (!nr) {
1086 unsigned long filepages, headroom, dirty, writeback;
1087
1088 mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
1089 &writeback);
1090 nr = dirty * 10 / 8;
1091 }
1092
1093 /* issue the writeback work */
1094 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1095 if (work) {
1096 work->nr_pages = nr;
1097 work->sync_mode = WB_SYNC_NONE;
1098 work->range_cyclic = 1;
1099 work->reason = reason;
1100 work->done = done;
1101 work->auto_free = 1;
1102 wb_queue_work(wb, work);
1103 ret = 0;
1104 } else {
1105 ret = -ENOMEM;
1106 }
1107
1108 wb_put(wb);
1109out_css_put:
1110 css_put(memcg_css);
1111out_bdi_put:
1112 bdi_put(bdi);
1113 return ret;
1114}
1115
1116/**
1117 * cgroup_writeback_umount - flush inode wb switches for umount
1118 *
1119 * This function is called when a super_block is about to be destroyed and
1120 * flushes in-flight inode wb switches. An inode wb switch goes through
1121 * RCU and then workqueue, so the two need to be flushed in order to ensure
1122 * that all previously scheduled switches are finished. As wb switches are
1123 * rare occurrences and synchronize_rcu() can take a while, perform
1124 * flushing iff wb switches are in flight.
1125 */
1126void cgroup_writeback_umount(void)
1127{
1128 /*
1129 * SB_ACTIVE should be reliably cleared before checking
1130 * isw_nr_in_flight, see generic_shutdown_super().
1131 */
1132 smp_mb();
1133
1134 if (atomic_read(&isw_nr_in_flight)) {
1135 /*
1136 * Use rcu_barrier() to wait for all pending callbacks to
1137 * ensure that all in-flight wb switches are in the workqueue.
1138 */
1139 rcu_barrier();
1140 flush_workqueue(isw_wq);
1141 }
1142}
1143
1144static int __init cgroup_writeback_init(void)
1145{
1146 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1147 if (!isw_wq)
1148 return -ENOMEM;
1149 return 0;
1150}
1151fs_initcall(cgroup_writeback_init);
1152
1153#else /* CONFIG_CGROUP_WRITEBACK */
1154
1155static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1156static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1157
1158static void inode_cgwb_move_to_attached(struct inode *inode,
1159 struct bdi_writeback *wb)
1160{
1161 assert_spin_locked(&wb->list_lock);
1162 assert_spin_locked(&inode->i_lock);
1163
1164 inode->i_state &= ~I_SYNC_QUEUED;
1165 list_del_init(&inode->i_io_list);
1166 wb_io_lists_depopulated(wb);
1167}
1168
1169static struct bdi_writeback *
1170locked_inode_to_wb_and_lock_list(struct inode *inode)
1171 __releases(&inode->i_lock)
1172 __acquires(&wb->list_lock)
1173{
1174 struct bdi_writeback *wb = inode_to_wb(inode);
1175
1176 spin_unlock(&inode->i_lock);
1177 spin_lock(&wb->list_lock);
1178 return wb;
1179}
1180
1181static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1182 __acquires(&wb->list_lock)
1183{
1184 struct bdi_writeback *wb = inode_to_wb(inode);
1185
1186 spin_lock(&wb->list_lock);
1187 return wb;
1188}
1189
1190static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1191{
1192 return nr_pages;
1193}
1194
1195static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1196 struct wb_writeback_work *base_work,
1197 bool skip_if_busy)
1198{
1199 might_sleep();
1200
1201 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1202 base_work->auto_free = 0;
1203 wb_queue_work(&bdi->wb, base_work);
1204 }
1205}
1206
1207#endif /* CONFIG_CGROUP_WRITEBACK */
1208
1209/*
1210 * Add in the number of potentially dirty inodes, because each inode
1211 * write can dirty pagecache in the underlying blockdev.
1212 */
1213static unsigned long get_nr_dirty_pages(void)
1214{
1215 return global_node_page_state(NR_FILE_DIRTY) +
1216 get_nr_dirty_inodes();
1217}
1218
1219static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1220{
1221 if (!wb_has_dirty_io(wb))
1222 return;
1223
1224 /*
1225 * All callers of this function want to start writeback of all
1226 * dirty pages. Places like vmscan can call this at a very
1227 * high frequency, causing pointless allocations of tons of
1228 * work items and keeping the flusher threads busy retrieving
1229 * that work. Ensure that we only allow one of them pending and
1230 * inflight at the time.
1231 */
1232 if (test_bit(WB_start_all, &wb->state) ||
1233 test_and_set_bit(WB_start_all, &wb->state))
1234 return;
1235
1236 wb->start_all_reason = reason;
1237 wb_wakeup(wb);
1238}
1239
1240/**
1241 * wb_start_background_writeback - start background writeback
1242 * @wb: bdi_writback to write from
1243 *
1244 * Description:
1245 * This makes sure WB_SYNC_NONE background writeback happens. When
1246 * this function returns, it is only guaranteed that for given wb
1247 * some IO is happening if we are over background dirty threshold.
1248 * Caller need not hold sb s_umount semaphore.
1249 */
1250void wb_start_background_writeback(struct bdi_writeback *wb)
1251{
1252 /*
1253 * We just wake up the flusher thread. It will perform background
1254 * writeback as soon as there is no other work to do.
1255 */
1256 trace_writeback_wake_background(wb);
1257 wb_wakeup(wb);
1258}
1259
1260/*
1261 * Remove the inode from the writeback list it is on.
1262 */
1263void inode_io_list_del(struct inode *inode)
1264{
1265 struct bdi_writeback *wb;
1266
1267 wb = inode_to_wb_and_lock_list(inode);
1268 spin_lock(&inode->i_lock);
1269
1270 inode->i_state &= ~I_SYNC_QUEUED;
1271 list_del_init(&inode->i_io_list);
1272 wb_io_lists_depopulated(wb);
1273
1274 spin_unlock(&inode->i_lock);
1275 spin_unlock(&wb->list_lock);
1276}
1277EXPORT_SYMBOL(inode_io_list_del);
1278
1279/*
1280 * mark an inode as under writeback on the sb
1281 */
1282void sb_mark_inode_writeback(struct inode *inode)
1283{
1284 struct super_block *sb = inode->i_sb;
1285 unsigned long flags;
1286
1287 if (list_empty(&inode->i_wb_list)) {
1288 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1289 if (list_empty(&inode->i_wb_list)) {
1290 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1291 trace_sb_mark_inode_writeback(inode);
1292 }
1293 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1294 }
1295}
1296
1297/*
1298 * clear an inode as under writeback on the sb
1299 */
1300void sb_clear_inode_writeback(struct inode *inode)
1301{
1302 struct super_block *sb = inode->i_sb;
1303 unsigned long flags;
1304
1305 if (!list_empty(&inode->i_wb_list)) {
1306 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1307 if (!list_empty(&inode->i_wb_list)) {
1308 list_del_init(&inode->i_wb_list);
1309 trace_sb_clear_inode_writeback(inode);
1310 }
1311 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1312 }
1313}
1314
1315/*
1316 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1317 * furthest end of its superblock's dirty-inode list.
1318 *
1319 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1320 * already the most-recently-dirtied inode on the b_dirty list. If that is
1321 * the case then the inode must have been redirtied while it was being written
1322 * out and we don't reset its dirtied_when.
1323 */
1324static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1325{
1326 assert_spin_locked(&inode->i_lock);
1327
1328 if (!list_empty(&wb->b_dirty)) {
1329 struct inode *tail;
1330
1331 tail = wb_inode(wb->b_dirty.next);
1332 if (time_before(inode->dirtied_when, tail->dirtied_when))
1333 inode->dirtied_when = jiffies;
1334 }
1335 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1336 inode->i_state &= ~I_SYNC_QUEUED;
1337}
1338
1339static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1340{
1341 spin_lock(&inode->i_lock);
1342 redirty_tail_locked(inode, wb);
1343 spin_unlock(&inode->i_lock);
1344}
1345
1346/*
1347 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1348 */
1349static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1350{
1351 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1352}
1353
1354static void inode_sync_complete(struct inode *inode)
1355{
1356 inode->i_state &= ~I_SYNC;
1357 /* If inode is clean an unused, put it into LRU now... */
1358 inode_add_lru(inode);
1359 /* Waiters must see I_SYNC cleared before being woken up */
1360 smp_mb();
1361 wake_up_bit(&inode->i_state, __I_SYNC);
1362}
1363
1364static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1365{
1366 bool ret = time_after(inode->dirtied_when, t);
1367#ifndef CONFIG_64BIT
1368 /*
1369 * For inodes being constantly redirtied, dirtied_when can get stuck.
1370 * It _appears_ to be in the future, but is actually in distant past.
1371 * This test is necessary to prevent such wrapped-around relative times
1372 * from permanently stopping the whole bdi writeback.
1373 */
1374 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1375#endif
1376 return ret;
1377}
1378
1379#define EXPIRE_DIRTY_ATIME 0x0001
1380
1381/*
1382 * Move expired (dirtied before dirtied_before) dirty inodes from
1383 * @delaying_queue to @dispatch_queue.
1384 */
1385static int move_expired_inodes(struct list_head *delaying_queue,
1386 struct list_head *dispatch_queue,
1387 unsigned long dirtied_before)
1388{
1389 LIST_HEAD(tmp);
1390 struct list_head *pos, *node;
1391 struct super_block *sb = NULL;
1392 struct inode *inode;
1393 int do_sb_sort = 0;
1394 int moved = 0;
1395
1396 while (!list_empty(delaying_queue)) {
1397 inode = wb_inode(delaying_queue->prev);
1398 if (inode_dirtied_after(inode, dirtied_before))
1399 break;
1400 list_move(&inode->i_io_list, &tmp);
1401 moved++;
1402 spin_lock(&inode->i_lock);
1403 inode->i_state |= I_SYNC_QUEUED;
1404 spin_unlock(&inode->i_lock);
1405 if (sb_is_blkdev_sb(inode->i_sb))
1406 continue;
1407 if (sb && sb != inode->i_sb)
1408 do_sb_sort = 1;
1409 sb = inode->i_sb;
1410 }
1411
1412 /* just one sb in list, splice to dispatch_queue and we're done */
1413 if (!do_sb_sort) {
1414 list_splice(&tmp, dispatch_queue);
1415 goto out;
1416 }
1417
1418 /* Move inodes from one superblock together */
1419 while (!list_empty(&tmp)) {
1420 sb = wb_inode(tmp.prev)->i_sb;
1421 list_for_each_prev_safe(pos, node, &tmp) {
1422 inode = wb_inode(pos);
1423 if (inode->i_sb == sb)
1424 list_move(&inode->i_io_list, dispatch_queue);
1425 }
1426 }
1427out:
1428 return moved;
1429}
1430
1431/*
1432 * Queue all expired dirty inodes for io, eldest first.
1433 * Before
1434 * newly dirtied b_dirty b_io b_more_io
1435 * =============> gf edc BA
1436 * After
1437 * newly dirtied b_dirty b_io b_more_io
1438 * =============> g fBAedc
1439 * |
1440 * +--> dequeue for IO
1441 */
1442static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1443 unsigned long dirtied_before)
1444{
1445 int moved;
1446 unsigned long time_expire_jif = dirtied_before;
1447
1448 assert_spin_locked(&wb->list_lock);
1449 list_splice_init(&wb->b_more_io, &wb->b_io);
1450 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1451 if (!work->for_sync)
1452 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1453 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1454 time_expire_jif);
1455 if (moved)
1456 wb_io_lists_populated(wb);
1457 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1458}
1459
1460static int write_inode(struct inode *inode, struct writeback_control *wbc)
1461{
1462 int ret;
1463
1464 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1465 trace_writeback_write_inode_start(inode, wbc);
1466 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1467 trace_writeback_write_inode(inode, wbc);
1468 return ret;
1469 }
1470 return 0;
1471}
1472
1473/*
1474 * Wait for writeback on an inode to complete. Called with i_lock held.
1475 * Caller must make sure inode cannot go away when we drop i_lock.
1476 */
1477static void __inode_wait_for_writeback(struct inode *inode)
1478 __releases(inode->i_lock)
1479 __acquires(inode->i_lock)
1480{
1481 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1482 wait_queue_head_t *wqh;
1483
1484 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1485 while (inode->i_state & I_SYNC) {
1486 spin_unlock(&inode->i_lock);
1487 __wait_on_bit(wqh, &wq, bit_wait,
1488 TASK_UNINTERRUPTIBLE);
1489 spin_lock(&inode->i_lock);
1490 }
1491}
1492
1493/*
1494 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1495 */
1496void inode_wait_for_writeback(struct inode *inode)
1497{
1498 spin_lock(&inode->i_lock);
1499 __inode_wait_for_writeback(inode);
1500 spin_unlock(&inode->i_lock);
1501}
1502
1503/*
1504 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1505 * held and drops it. It is aimed for callers not holding any inode reference
1506 * so once i_lock is dropped, inode can go away.
1507 */
1508static void inode_sleep_on_writeback(struct inode *inode)
1509 __releases(inode->i_lock)
1510{
1511 DEFINE_WAIT(wait);
1512 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1513 int sleep;
1514
1515 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1516 sleep = inode->i_state & I_SYNC;
1517 spin_unlock(&inode->i_lock);
1518 if (sleep)
1519 schedule();
1520 finish_wait(wqh, &wait);
1521}
1522
1523/*
1524 * Find proper writeback list for the inode depending on its current state and
1525 * possibly also change of its state while we were doing writeback. Here we
1526 * handle things such as livelock prevention or fairness of writeback among
1527 * inodes. This function can be called only by flusher thread - noone else
1528 * processes all inodes in writeback lists and requeueing inodes behind flusher
1529 * thread's back can have unexpected consequences.
1530 */
1531static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1532 struct writeback_control *wbc)
1533{
1534 if (inode->i_state & I_FREEING)
1535 return;
1536
1537 /*
1538 * Sync livelock prevention. Each inode is tagged and synced in one
1539 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1540 * the dirty time to prevent enqueue and sync it again.
1541 */
1542 if ((inode->i_state & I_DIRTY) &&
1543 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1544 inode->dirtied_when = jiffies;
1545
1546 if (wbc->pages_skipped) {
1547 /*
1548 * writeback is not making progress due to locked
1549 * buffers. Skip this inode for now.
1550 */
1551 redirty_tail_locked(inode, wb);
1552 return;
1553 }
1554
1555 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1556 /*
1557 * We didn't write back all the pages. nfs_writepages()
1558 * sometimes bales out without doing anything.
1559 */
1560 if (wbc->nr_to_write <= 0) {
1561 /* Slice used up. Queue for next turn. */
1562 requeue_io(inode, wb);
1563 } else {
1564 /*
1565 * Writeback blocked by something other than
1566 * congestion. Delay the inode for some time to
1567 * avoid spinning on the CPU (100% iowait)
1568 * retrying writeback of the dirty page/inode
1569 * that cannot be performed immediately.
1570 */
1571 redirty_tail_locked(inode, wb);
1572 }
1573 } else if (inode->i_state & I_DIRTY) {
1574 /*
1575 * Filesystems can dirty the inode during writeback operations,
1576 * such as delayed allocation during submission or metadata
1577 * updates after data IO completion.
1578 */
1579 redirty_tail_locked(inode, wb);
1580 } else if (inode->i_state & I_DIRTY_TIME) {
1581 inode->dirtied_when = jiffies;
1582 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1583 inode->i_state &= ~I_SYNC_QUEUED;
1584 } else {
1585 /* The inode is clean. Remove from writeback lists. */
1586 inode_cgwb_move_to_attached(inode, wb);
1587 }
1588}
1589
1590/*
1591 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1592 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1593 *
1594 * This doesn't remove the inode from the writeback list it is on, except
1595 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1596 * expiration. The caller is otherwise responsible for writeback list handling.
1597 *
1598 * The caller is also responsible for setting the I_SYNC flag beforehand and
1599 * calling inode_sync_complete() to clear it afterwards.
1600 */
1601static int
1602__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1603{
1604 struct address_space *mapping = inode->i_mapping;
1605 long nr_to_write = wbc->nr_to_write;
1606 unsigned dirty;
1607 int ret;
1608
1609 WARN_ON(!(inode->i_state & I_SYNC));
1610
1611 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1612
1613 ret = do_writepages(mapping, wbc);
1614
1615 /*
1616 * Make sure to wait on the data before writing out the metadata.
1617 * This is important for filesystems that modify metadata on data
1618 * I/O completion. We don't do it for sync(2) writeback because it has a
1619 * separate, external IO completion path and ->sync_fs for guaranteeing
1620 * inode metadata is written back correctly.
1621 */
1622 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1623 int err = filemap_fdatawait(mapping);
1624 if (ret == 0)
1625 ret = err;
1626 }
1627
1628 /*
1629 * If the inode has dirty timestamps and we need to write them, call
1630 * mark_inode_dirty_sync() to notify the filesystem about it and to
1631 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1632 */
1633 if ((inode->i_state & I_DIRTY_TIME) &&
1634 (wbc->sync_mode == WB_SYNC_ALL ||
1635 time_after(jiffies, inode->dirtied_time_when +
1636 dirtytime_expire_interval * HZ))) {
1637 trace_writeback_lazytime(inode);
1638 mark_inode_dirty_sync(inode);
1639 }
1640
1641 /*
1642 * Get and clear the dirty flags from i_state. This needs to be done
1643 * after calling writepages because some filesystems may redirty the
1644 * inode during writepages due to delalloc. It also needs to be done
1645 * after handling timestamp expiration, as that may dirty the inode too.
1646 */
1647 spin_lock(&inode->i_lock);
1648 dirty = inode->i_state & I_DIRTY;
1649 inode->i_state &= ~dirty;
1650
1651 /*
1652 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1653 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1654 * either they see the I_DIRTY bits cleared or we see the dirtied
1655 * inode.
1656 *
1657 * I_DIRTY_PAGES is always cleared together above even if @mapping
1658 * still has dirty pages. The flag is reinstated after smp_mb() if
1659 * necessary. This guarantees that either __mark_inode_dirty()
1660 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1661 */
1662 smp_mb();
1663
1664 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1665 inode->i_state |= I_DIRTY_PAGES;
1666
1667 spin_unlock(&inode->i_lock);
1668
1669 /* Don't write the inode if only I_DIRTY_PAGES was set */
1670 if (dirty & ~I_DIRTY_PAGES) {
1671 int err = write_inode(inode, wbc);
1672 if (ret == 0)
1673 ret = err;
1674 }
1675 trace_writeback_single_inode(inode, wbc, nr_to_write);
1676 return ret;
1677}
1678
1679/*
1680 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1681 * the regular batched writeback done by the flusher threads in
1682 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1683 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1684 *
1685 * To prevent the inode from going away, either the caller must have a reference
1686 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1687 */
1688static int writeback_single_inode(struct inode *inode,
1689 struct writeback_control *wbc)
1690{
1691 struct bdi_writeback *wb;
1692 int ret = 0;
1693
1694 spin_lock(&inode->i_lock);
1695 if (!atomic_read(&inode->i_count))
1696 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1697 else
1698 WARN_ON(inode->i_state & I_WILL_FREE);
1699
1700 if (inode->i_state & I_SYNC) {
1701 /*
1702 * Writeback is already running on the inode. For WB_SYNC_NONE,
1703 * that's enough and we can just return. For WB_SYNC_ALL, we
1704 * must wait for the existing writeback to complete, then do
1705 * writeback again if there's anything left.
1706 */
1707 if (wbc->sync_mode != WB_SYNC_ALL)
1708 goto out;
1709 __inode_wait_for_writeback(inode);
1710 }
1711 WARN_ON(inode->i_state & I_SYNC);
1712 /*
1713 * If the inode is already fully clean, then there's nothing to do.
1714 *
1715 * For data-integrity syncs we also need to check whether any pages are
1716 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1717 * there are any such pages, we'll need to wait for them.
1718 */
1719 if (!(inode->i_state & I_DIRTY_ALL) &&
1720 (wbc->sync_mode != WB_SYNC_ALL ||
1721 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1722 goto out;
1723 inode->i_state |= I_SYNC;
1724 wbc_attach_and_unlock_inode(wbc, inode);
1725
1726 ret = __writeback_single_inode(inode, wbc);
1727
1728 wbc_detach_inode(wbc);
1729
1730 wb = inode_to_wb_and_lock_list(inode);
1731 spin_lock(&inode->i_lock);
1732 /*
1733 * If the inode is now fully clean, then it can be safely removed from
1734 * its writeback list (if any). Otherwise the flusher threads are
1735 * responsible for the writeback lists.
1736 */
1737 if (!(inode->i_state & I_DIRTY_ALL))
1738 inode_cgwb_move_to_attached(inode, wb);
1739 spin_unlock(&wb->list_lock);
1740 inode_sync_complete(inode);
1741out:
1742 spin_unlock(&inode->i_lock);
1743 return ret;
1744}
1745
1746static long writeback_chunk_size(struct bdi_writeback *wb,
1747 struct wb_writeback_work *work)
1748{
1749 long pages;
1750
1751 /*
1752 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1753 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1754 * here avoids calling into writeback_inodes_wb() more than once.
1755 *
1756 * The intended call sequence for WB_SYNC_ALL writeback is:
1757 *
1758 * wb_writeback()
1759 * writeback_sb_inodes() <== called only once
1760 * write_cache_pages() <== called once for each inode
1761 * (quickly) tag currently dirty pages
1762 * (maybe slowly) sync all tagged pages
1763 */
1764 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1765 pages = LONG_MAX;
1766 else {
1767 pages = min(wb->avg_write_bandwidth / 2,
1768 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1769 pages = min(pages, work->nr_pages);
1770 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1771 MIN_WRITEBACK_PAGES);
1772 }
1773
1774 return pages;
1775}
1776
1777/*
1778 * Write a portion of b_io inodes which belong to @sb.
1779 *
1780 * Return the number of pages and/or inodes written.
1781 *
1782 * NOTE! This is called with wb->list_lock held, and will
1783 * unlock and relock that for each inode it ends up doing
1784 * IO for.
1785 */
1786static long writeback_sb_inodes(struct super_block *sb,
1787 struct bdi_writeback *wb,
1788 struct wb_writeback_work *work)
1789{
1790 struct writeback_control wbc = {
1791 .sync_mode = work->sync_mode,
1792 .tagged_writepages = work->tagged_writepages,
1793 .for_kupdate = work->for_kupdate,
1794 .for_background = work->for_background,
1795 .for_sync = work->for_sync,
1796 .range_cyclic = work->range_cyclic,
1797 .range_start = 0,
1798 .range_end = LLONG_MAX,
1799 };
1800 unsigned long start_time = jiffies;
1801 long write_chunk;
1802 long wrote = 0; /* count both pages and inodes */
1803
1804 while (!list_empty(&wb->b_io)) {
1805 struct inode *inode = wb_inode(wb->b_io.prev);
1806 struct bdi_writeback *tmp_wb;
1807
1808 if (inode->i_sb != sb) {
1809 if (work->sb) {
1810 /*
1811 * We only want to write back data for this
1812 * superblock, move all inodes not belonging
1813 * to it back onto the dirty list.
1814 */
1815 redirty_tail(inode, wb);
1816 continue;
1817 }
1818
1819 /*
1820 * The inode belongs to a different superblock.
1821 * Bounce back to the caller to unpin this and
1822 * pin the next superblock.
1823 */
1824 break;
1825 }
1826
1827 /*
1828 * Don't bother with new inodes or inodes being freed, first
1829 * kind does not need periodic writeout yet, and for the latter
1830 * kind writeout is handled by the freer.
1831 */
1832 spin_lock(&inode->i_lock);
1833 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1834 redirty_tail_locked(inode, wb);
1835 spin_unlock(&inode->i_lock);
1836 continue;
1837 }
1838 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1839 /*
1840 * If this inode is locked for writeback and we are not
1841 * doing writeback-for-data-integrity, move it to
1842 * b_more_io so that writeback can proceed with the
1843 * other inodes on s_io.
1844 *
1845 * We'll have another go at writing back this inode
1846 * when we completed a full scan of b_io.
1847 */
1848 spin_unlock(&inode->i_lock);
1849 requeue_io(inode, wb);
1850 trace_writeback_sb_inodes_requeue(inode);
1851 continue;
1852 }
1853 spin_unlock(&wb->list_lock);
1854
1855 /*
1856 * We already requeued the inode if it had I_SYNC set and we
1857 * are doing WB_SYNC_NONE writeback. So this catches only the
1858 * WB_SYNC_ALL case.
1859 */
1860 if (inode->i_state & I_SYNC) {
1861 /* Wait for I_SYNC. This function drops i_lock... */
1862 inode_sleep_on_writeback(inode);
1863 /* Inode may be gone, start again */
1864 spin_lock(&wb->list_lock);
1865 continue;
1866 }
1867 inode->i_state |= I_SYNC;
1868 wbc_attach_and_unlock_inode(&wbc, inode);
1869
1870 write_chunk = writeback_chunk_size(wb, work);
1871 wbc.nr_to_write = write_chunk;
1872 wbc.pages_skipped = 0;
1873
1874 /*
1875 * We use I_SYNC to pin the inode in memory. While it is set
1876 * evict_inode() will wait so the inode cannot be freed.
1877 */
1878 __writeback_single_inode(inode, &wbc);
1879
1880 wbc_detach_inode(&wbc);
1881 work->nr_pages -= write_chunk - wbc.nr_to_write;
1882 wrote += write_chunk - wbc.nr_to_write;
1883
1884 if (need_resched()) {
1885 /*
1886 * We're trying to balance between building up a nice
1887 * long list of IOs to improve our merge rate, and
1888 * getting those IOs out quickly for anyone throttling
1889 * in balance_dirty_pages(). cond_resched() doesn't
1890 * unplug, so get our IOs out the door before we
1891 * give up the CPU.
1892 */
1893 blk_flush_plug(current);
1894 cond_resched();
1895 }
1896
1897 /*
1898 * Requeue @inode if still dirty. Be careful as @inode may
1899 * have been switched to another wb in the meantime.
1900 */
1901 tmp_wb = inode_to_wb_and_lock_list(inode);
1902 spin_lock(&inode->i_lock);
1903 if (!(inode->i_state & I_DIRTY_ALL))
1904 wrote++;
1905 requeue_inode(inode, tmp_wb, &wbc);
1906 inode_sync_complete(inode);
1907 spin_unlock(&inode->i_lock);
1908
1909 if (unlikely(tmp_wb != wb)) {
1910 spin_unlock(&tmp_wb->list_lock);
1911 spin_lock(&wb->list_lock);
1912 }
1913
1914 /*
1915 * bail out to wb_writeback() often enough to check
1916 * background threshold and other termination conditions.
1917 */
1918 if (wrote) {
1919 if (time_is_before_jiffies(start_time + HZ / 10UL))
1920 break;
1921 if (work->nr_pages <= 0)
1922 break;
1923 }
1924 }
1925 return wrote;
1926}
1927
1928static long __writeback_inodes_wb(struct bdi_writeback *wb,
1929 struct wb_writeback_work *work)
1930{
1931 unsigned long start_time = jiffies;
1932 long wrote = 0;
1933
1934 while (!list_empty(&wb->b_io)) {
1935 struct inode *inode = wb_inode(wb->b_io.prev);
1936 struct super_block *sb = inode->i_sb;
1937
1938 if (!trylock_super(sb)) {
1939 /*
1940 * trylock_super() may fail consistently due to
1941 * s_umount being grabbed by someone else. Don't use
1942 * requeue_io() to avoid busy retrying the inode/sb.
1943 */
1944 redirty_tail(inode, wb);
1945 continue;
1946 }
1947 wrote += writeback_sb_inodes(sb, wb, work);
1948 up_read(&sb->s_umount);
1949
1950 /* refer to the same tests at the end of writeback_sb_inodes */
1951 if (wrote) {
1952 if (time_is_before_jiffies(start_time + HZ / 10UL))
1953 break;
1954 if (work->nr_pages <= 0)
1955 break;
1956 }
1957 }
1958 /* Leave any unwritten inodes on b_io */
1959 return wrote;
1960}
1961
1962static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1963 enum wb_reason reason)
1964{
1965 struct wb_writeback_work work = {
1966 .nr_pages = nr_pages,
1967 .sync_mode = WB_SYNC_NONE,
1968 .range_cyclic = 1,
1969 .reason = reason,
1970 };
1971 struct blk_plug plug;
1972
1973 blk_start_plug(&plug);
1974 spin_lock(&wb->list_lock);
1975 if (list_empty(&wb->b_io))
1976 queue_io(wb, &work, jiffies);
1977 __writeback_inodes_wb(wb, &work);
1978 spin_unlock(&wb->list_lock);
1979 blk_finish_plug(&plug);
1980
1981 return nr_pages - work.nr_pages;
1982}
1983
1984/*
1985 * Explicit flushing or periodic writeback of "old" data.
1986 *
1987 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1988 * dirtying-time in the inode's address_space. So this periodic writeback code
1989 * just walks the superblock inode list, writing back any inodes which are
1990 * older than a specific point in time.
1991 *
1992 * Try to run once per dirty_writeback_interval. But if a writeback event
1993 * takes longer than a dirty_writeback_interval interval, then leave a
1994 * one-second gap.
1995 *
1996 * dirtied_before takes precedence over nr_to_write. So we'll only write back
1997 * all dirty pages if they are all attached to "old" mappings.
1998 */
1999static long wb_writeback(struct bdi_writeback *wb,
2000 struct wb_writeback_work *work)
2001{
2002 unsigned long wb_start = jiffies;
2003 long nr_pages = work->nr_pages;
2004 unsigned long dirtied_before = jiffies;
2005 struct inode *inode;
2006 long progress;
2007 struct blk_plug plug;
2008
2009 blk_start_plug(&plug);
2010 spin_lock(&wb->list_lock);
2011 for (;;) {
2012 /*
2013 * Stop writeback when nr_pages has been consumed
2014 */
2015 if (work->nr_pages <= 0)
2016 break;
2017
2018 /*
2019 * Background writeout and kupdate-style writeback may
2020 * run forever. Stop them if there is other work to do
2021 * so that e.g. sync can proceed. They'll be restarted
2022 * after the other works are all done.
2023 */
2024 if ((work->for_background || work->for_kupdate) &&
2025 !list_empty(&wb->work_list))
2026 break;
2027
2028 /*
2029 * For background writeout, stop when we are below the
2030 * background dirty threshold
2031 */
2032 if (work->for_background && !wb_over_bg_thresh(wb))
2033 break;
2034
2035 /*
2036 * Kupdate and background works are special and we want to
2037 * include all inodes that need writing. Livelock avoidance is
2038 * handled by these works yielding to any other work so we are
2039 * safe.
2040 */
2041 if (work->for_kupdate) {
2042 dirtied_before = jiffies -
2043 msecs_to_jiffies(dirty_expire_interval * 10);
2044 } else if (work->for_background)
2045 dirtied_before = jiffies;
2046
2047 trace_writeback_start(wb, work);
2048 if (list_empty(&wb->b_io))
2049 queue_io(wb, work, dirtied_before);
2050 if (work->sb)
2051 progress = writeback_sb_inodes(work->sb, wb, work);
2052 else
2053 progress = __writeback_inodes_wb(wb, work);
2054 trace_writeback_written(wb, work);
2055
2056 wb_update_bandwidth(wb, wb_start);
2057
2058 /*
2059 * Did we write something? Try for more
2060 *
2061 * Dirty inodes are moved to b_io for writeback in batches.
2062 * The completion of the current batch does not necessarily
2063 * mean the overall work is done. So we keep looping as long
2064 * as made some progress on cleaning pages or inodes.
2065 */
2066 if (progress)
2067 continue;
2068 /*
2069 * No more inodes for IO, bail
2070 */
2071 if (list_empty(&wb->b_more_io))
2072 break;
2073 /*
2074 * Nothing written. Wait for some inode to
2075 * become available for writeback. Otherwise
2076 * we'll just busyloop.
2077 */
2078 trace_writeback_wait(wb, work);
2079 inode = wb_inode(wb->b_more_io.prev);
2080 spin_lock(&inode->i_lock);
2081 spin_unlock(&wb->list_lock);
2082 /* This function drops i_lock... */
2083 inode_sleep_on_writeback(inode);
2084 spin_lock(&wb->list_lock);
2085 }
2086 spin_unlock(&wb->list_lock);
2087 blk_finish_plug(&plug);
2088
2089 return nr_pages - work->nr_pages;
2090}
2091
2092/*
2093 * Return the next wb_writeback_work struct that hasn't been processed yet.
2094 */
2095static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2096{
2097 struct wb_writeback_work *work = NULL;
2098
2099 spin_lock_bh(&wb->work_lock);
2100 if (!list_empty(&wb->work_list)) {
2101 work = list_entry(wb->work_list.next,
2102 struct wb_writeback_work, list);
2103 list_del_init(&work->list);
2104 }
2105 spin_unlock_bh(&wb->work_lock);
2106 return work;
2107}
2108
2109static long wb_check_background_flush(struct bdi_writeback *wb)
2110{
2111 if (wb_over_bg_thresh(wb)) {
2112
2113 struct wb_writeback_work work = {
2114 .nr_pages = LONG_MAX,
2115 .sync_mode = WB_SYNC_NONE,
2116 .for_background = 1,
2117 .range_cyclic = 1,
2118 .reason = WB_REASON_BACKGROUND,
2119 };
2120
2121 return wb_writeback(wb, &work);
2122 }
2123
2124 return 0;
2125}
2126
2127static long wb_check_old_data_flush(struct bdi_writeback *wb)
2128{
2129 unsigned long expired;
2130 long nr_pages;
2131
2132 /*
2133 * When set to zero, disable periodic writeback
2134 */
2135 if (!dirty_writeback_interval)
2136 return 0;
2137
2138 expired = wb->last_old_flush +
2139 msecs_to_jiffies(dirty_writeback_interval * 10);
2140 if (time_before(jiffies, expired))
2141 return 0;
2142
2143 wb->last_old_flush = jiffies;
2144 nr_pages = get_nr_dirty_pages();
2145
2146 if (nr_pages) {
2147 struct wb_writeback_work work = {
2148 .nr_pages = nr_pages,
2149 .sync_mode = WB_SYNC_NONE,
2150 .for_kupdate = 1,
2151 .range_cyclic = 1,
2152 .reason = WB_REASON_PERIODIC,
2153 };
2154
2155 return wb_writeback(wb, &work);
2156 }
2157
2158 return 0;
2159}
2160
2161static long wb_check_start_all(struct bdi_writeback *wb)
2162{
2163 long nr_pages;
2164
2165 if (!test_bit(WB_start_all, &wb->state))
2166 return 0;
2167
2168 nr_pages = get_nr_dirty_pages();
2169 if (nr_pages) {
2170 struct wb_writeback_work work = {
2171 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2172 .sync_mode = WB_SYNC_NONE,
2173 .range_cyclic = 1,
2174 .reason = wb->start_all_reason,
2175 };
2176
2177 nr_pages = wb_writeback(wb, &work);
2178 }
2179
2180 clear_bit(WB_start_all, &wb->state);
2181 return nr_pages;
2182}
2183
2184
2185/*
2186 * Retrieve work items and do the writeback they describe
2187 */
2188static long wb_do_writeback(struct bdi_writeback *wb)
2189{
2190 struct wb_writeback_work *work;
2191 long wrote = 0;
2192
2193 set_bit(WB_writeback_running, &wb->state);
2194 while ((work = get_next_work_item(wb)) != NULL) {
2195 trace_writeback_exec(wb, work);
2196 wrote += wb_writeback(wb, work);
2197 finish_writeback_work(wb, work);
2198 }
2199
2200 /*
2201 * Check for a flush-everything request
2202 */
2203 wrote += wb_check_start_all(wb);
2204
2205 /*
2206 * Check for periodic writeback, kupdated() style
2207 */
2208 wrote += wb_check_old_data_flush(wb);
2209 wrote += wb_check_background_flush(wb);
2210 clear_bit(WB_writeback_running, &wb->state);
2211
2212 return wrote;
2213}
2214
2215/*
2216 * Handle writeback of dirty data for the device backed by this bdi. Also
2217 * reschedules periodically and does kupdated style flushing.
2218 */
2219void wb_workfn(struct work_struct *work)
2220{
2221 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2222 struct bdi_writeback, dwork);
2223 long pages_written;
2224
2225 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2226 current->flags |= PF_SWAPWRITE;
2227
2228 if (likely(!current_is_workqueue_rescuer() ||
2229 !test_bit(WB_registered, &wb->state))) {
2230 /*
2231 * The normal path. Keep writing back @wb until its
2232 * work_list is empty. Note that this path is also taken
2233 * if @wb is shutting down even when we're running off the
2234 * rescuer as work_list needs to be drained.
2235 */
2236 do {
2237 pages_written = wb_do_writeback(wb);
2238 trace_writeback_pages_written(pages_written);
2239 } while (!list_empty(&wb->work_list));
2240 } else {
2241 /*
2242 * bdi_wq can't get enough workers and we're running off
2243 * the emergency worker. Don't hog it. Hopefully, 1024 is
2244 * enough for efficient IO.
2245 */
2246 pages_written = writeback_inodes_wb(wb, 1024,
2247 WB_REASON_FORKER_THREAD);
2248 trace_writeback_pages_written(pages_written);
2249 }
2250
2251 if (!list_empty(&wb->work_list))
2252 wb_wakeup(wb);
2253 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2254 wb_wakeup_delayed(wb);
2255
2256 current->flags &= ~PF_SWAPWRITE;
2257}
2258
2259/*
2260 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2261 * write back the whole world.
2262 */
2263static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2264 enum wb_reason reason)
2265{
2266 struct bdi_writeback *wb;
2267
2268 if (!bdi_has_dirty_io(bdi))
2269 return;
2270
2271 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2272 wb_start_writeback(wb, reason);
2273}
2274
2275void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2276 enum wb_reason reason)
2277{
2278 rcu_read_lock();
2279 __wakeup_flusher_threads_bdi(bdi, reason);
2280 rcu_read_unlock();
2281}
2282
2283/*
2284 * Wakeup the flusher threads to start writeback of all currently dirty pages
2285 */
2286void wakeup_flusher_threads(enum wb_reason reason)
2287{
2288 struct backing_dev_info *bdi;
2289
2290 /*
2291 * If we are expecting writeback progress we must submit plugged IO.
2292 */
2293 if (blk_needs_flush_plug(current))
2294 blk_schedule_flush_plug(current);
2295
2296 rcu_read_lock();
2297 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2298 __wakeup_flusher_threads_bdi(bdi, reason);
2299 rcu_read_unlock();
2300}
2301
2302/*
2303 * Wake up bdi's periodically to make sure dirtytime inodes gets
2304 * written back periodically. We deliberately do *not* check the
2305 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2306 * kernel to be constantly waking up once there are any dirtytime
2307 * inodes on the system. So instead we define a separate delayed work
2308 * function which gets called much more rarely. (By default, only
2309 * once every 12 hours.)
2310 *
2311 * If there is any other write activity going on in the file system,
2312 * this function won't be necessary. But if the only thing that has
2313 * happened on the file system is a dirtytime inode caused by an atime
2314 * update, we need this infrastructure below to make sure that inode
2315 * eventually gets pushed out to disk.
2316 */
2317static void wakeup_dirtytime_writeback(struct work_struct *w);
2318static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2319
2320static void wakeup_dirtytime_writeback(struct work_struct *w)
2321{
2322 struct backing_dev_info *bdi;
2323
2324 rcu_read_lock();
2325 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2326 struct bdi_writeback *wb;
2327
2328 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2329 if (!list_empty(&wb->b_dirty_time))
2330 wb_wakeup(wb);
2331 }
2332 rcu_read_unlock();
2333 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2334}
2335
2336static int __init start_dirtytime_writeback(void)
2337{
2338 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2339 return 0;
2340}
2341__initcall(start_dirtytime_writeback);
2342
2343int dirtytime_interval_handler(struct ctl_table *table, int write,
2344 void *buffer, size_t *lenp, loff_t *ppos)
2345{
2346 int ret;
2347
2348 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2349 if (ret == 0 && write)
2350 mod_delayed_work(system_wq, &dirtytime_work, 0);
2351 return ret;
2352}
2353
2354/**
2355 * __mark_inode_dirty - internal function to mark an inode dirty
2356 *
2357 * @inode: inode to mark
2358 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2359 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2360 * with I_DIRTY_PAGES.
2361 *
2362 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2363 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2364 *
2365 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2366 * instead of calling this directly.
2367 *
2368 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2369 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2370 * even if they are later hashed, as they will have been marked dirty already.
2371 *
2372 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2373 *
2374 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2375 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2376 * the kernel-internal blockdev inode represents the dirtying time of the
2377 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2378 * page->mapping->host, so the page-dirtying time is recorded in the internal
2379 * blockdev inode.
2380 */
2381void __mark_inode_dirty(struct inode *inode, int flags)
2382{
2383 struct super_block *sb = inode->i_sb;
2384 int dirtytime = 0;
2385
2386 trace_writeback_mark_inode_dirty(inode, flags);
2387
2388 if (flags & I_DIRTY_INODE) {
2389 /*
2390 * Notify the filesystem about the inode being dirtied, so that
2391 * (if needed) it can update on-disk fields and journal the
2392 * inode. This is only needed when the inode itself is being
2393 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2394 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2395 */
2396 trace_writeback_dirty_inode_start(inode, flags);
2397 if (sb->s_op->dirty_inode)
2398 sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
2399 trace_writeback_dirty_inode(inode, flags);
2400
2401 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2402 flags &= ~I_DIRTY_TIME;
2403 } else {
2404 /*
2405 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2406 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2407 * in one call to __mark_inode_dirty().)
2408 */
2409 dirtytime = flags & I_DIRTY_TIME;
2410 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2411 }
2412
2413 /*
2414 * Paired with smp_mb() in __writeback_single_inode() for the
2415 * following lockless i_state test. See there for details.
2416 */
2417 smp_mb();
2418
2419 if (((inode->i_state & flags) == flags) ||
2420 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2421 return;
2422
2423 spin_lock(&inode->i_lock);
2424 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2425 goto out_unlock_inode;
2426 if ((inode->i_state & flags) != flags) {
2427 const int was_dirty = inode->i_state & I_DIRTY;
2428
2429 inode_attach_wb(inode, NULL);
2430
2431 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2432 if (flags & I_DIRTY_INODE)
2433 inode->i_state &= ~I_DIRTY_TIME;
2434 inode->i_state |= flags;
2435
2436 /*
2437 * If the inode is queued for writeback by flush worker, just
2438 * update its dirty state. Once the flush worker is done with
2439 * the inode it will place it on the appropriate superblock
2440 * list, based upon its state.
2441 */
2442 if (inode->i_state & I_SYNC_QUEUED)
2443 goto out_unlock_inode;
2444
2445 /*
2446 * Only add valid (hashed) inodes to the superblock's
2447 * dirty list. Add blockdev inodes as well.
2448 */
2449 if (!S_ISBLK(inode->i_mode)) {
2450 if (inode_unhashed(inode))
2451 goto out_unlock_inode;
2452 }
2453 if (inode->i_state & I_FREEING)
2454 goto out_unlock_inode;
2455
2456 /*
2457 * If the inode was already on b_dirty/b_io/b_more_io, don't
2458 * reposition it (that would break b_dirty time-ordering).
2459 */
2460 if (!was_dirty) {
2461 struct bdi_writeback *wb;
2462 struct list_head *dirty_list;
2463 bool wakeup_bdi = false;
2464
2465 wb = locked_inode_to_wb_and_lock_list(inode);
2466
2467 inode->dirtied_when = jiffies;
2468 if (dirtytime)
2469 inode->dirtied_time_when = jiffies;
2470
2471 if (inode->i_state & I_DIRTY)
2472 dirty_list = &wb->b_dirty;
2473 else
2474 dirty_list = &wb->b_dirty_time;
2475
2476 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2477 dirty_list);
2478
2479 spin_unlock(&wb->list_lock);
2480 trace_writeback_dirty_inode_enqueue(inode);
2481
2482 /*
2483 * If this is the first dirty inode for this bdi,
2484 * we have to wake-up the corresponding bdi thread
2485 * to make sure background write-back happens
2486 * later.
2487 */
2488 if (wakeup_bdi &&
2489 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2490 wb_wakeup_delayed(wb);
2491 return;
2492 }
2493 }
2494out_unlock_inode:
2495 spin_unlock(&inode->i_lock);
2496}
2497EXPORT_SYMBOL(__mark_inode_dirty);
2498
2499/*
2500 * The @s_sync_lock is used to serialise concurrent sync operations
2501 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2502 * Concurrent callers will block on the s_sync_lock rather than doing contending
2503 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2504 * has been issued up to the time this function is enter is guaranteed to be
2505 * completed by the time we have gained the lock and waited for all IO that is
2506 * in progress regardless of the order callers are granted the lock.
2507 */
2508static void wait_sb_inodes(struct super_block *sb)
2509{
2510 LIST_HEAD(sync_list);
2511
2512 /*
2513 * We need to be protected against the filesystem going from
2514 * r/o to r/w or vice versa.
2515 */
2516 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2517
2518 mutex_lock(&sb->s_sync_lock);
2519
2520 /*
2521 * Splice the writeback list onto a temporary list to avoid waiting on
2522 * inodes that have started writeback after this point.
2523 *
2524 * Use rcu_read_lock() to keep the inodes around until we have a
2525 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2526 * the local list because inodes can be dropped from either by writeback
2527 * completion.
2528 */
2529 rcu_read_lock();
2530 spin_lock_irq(&sb->s_inode_wblist_lock);
2531 list_splice_init(&sb->s_inodes_wb, &sync_list);
2532
2533 /*
2534 * Data integrity sync. Must wait for all pages under writeback, because
2535 * there may have been pages dirtied before our sync call, but which had
2536 * writeout started before we write it out. In which case, the inode
2537 * may not be on the dirty list, but we still have to wait for that
2538 * writeout.
2539 */
2540 while (!list_empty(&sync_list)) {
2541 struct inode *inode = list_first_entry(&sync_list, struct inode,
2542 i_wb_list);
2543 struct address_space *mapping = inode->i_mapping;
2544
2545 /*
2546 * Move each inode back to the wb list before we drop the lock
2547 * to preserve consistency between i_wb_list and the mapping
2548 * writeback tag. Writeback completion is responsible to remove
2549 * the inode from either list once the writeback tag is cleared.
2550 */
2551 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2552
2553 /*
2554 * The mapping can appear untagged while still on-list since we
2555 * do not have the mapping lock. Skip it here, wb completion
2556 * will remove it.
2557 */
2558 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2559 continue;
2560
2561 spin_unlock_irq(&sb->s_inode_wblist_lock);
2562
2563 spin_lock(&inode->i_lock);
2564 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2565 spin_unlock(&inode->i_lock);
2566
2567 spin_lock_irq(&sb->s_inode_wblist_lock);
2568 continue;
2569 }
2570 __iget(inode);
2571 spin_unlock(&inode->i_lock);
2572 rcu_read_unlock();
2573
2574 /*
2575 * We keep the error status of individual mapping so that
2576 * applications can catch the writeback error using fsync(2).
2577 * See filemap_fdatawait_keep_errors() for details.
2578 */
2579 filemap_fdatawait_keep_errors(mapping);
2580
2581 cond_resched();
2582
2583 iput(inode);
2584
2585 rcu_read_lock();
2586 spin_lock_irq(&sb->s_inode_wblist_lock);
2587 }
2588 spin_unlock_irq(&sb->s_inode_wblist_lock);
2589 rcu_read_unlock();
2590 mutex_unlock(&sb->s_sync_lock);
2591}
2592
2593static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2594 enum wb_reason reason, bool skip_if_busy)
2595{
2596 struct backing_dev_info *bdi = sb->s_bdi;
2597 DEFINE_WB_COMPLETION(done, bdi);
2598 struct wb_writeback_work work = {
2599 .sb = sb,
2600 .sync_mode = WB_SYNC_NONE,
2601 .tagged_writepages = 1,
2602 .done = &done,
2603 .nr_pages = nr,
2604 .reason = reason,
2605 };
2606
2607 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2608 return;
2609 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2610
2611 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2612 wb_wait_for_completion(&done);
2613}
2614
2615/**
2616 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2617 * @sb: the superblock
2618 * @nr: the number of pages to write
2619 * @reason: reason why some writeback work initiated
2620 *
2621 * Start writeback on some inodes on this super_block. No guarantees are made
2622 * on how many (if any) will be written, and this function does not wait
2623 * for IO completion of submitted IO.
2624 */
2625void writeback_inodes_sb_nr(struct super_block *sb,
2626 unsigned long nr,
2627 enum wb_reason reason)
2628{
2629 __writeback_inodes_sb_nr(sb, nr, reason, false);
2630}
2631EXPORT_SYMBOL(writeback_inodes_sb_nr);
2632
2633/**
2634 * writeback_inodes_sb - writeback dirty inodes from given super_block
2635 * @sb: the superblock
2636 * @reason: reason why some writeback work was initiated
2637 *
2638 * Start writeback on some inodes on this super_block. No guarantees are made
2639 * on how many (if any) will be written, and this function does not wait
2640 * for IO completion of submitted IO.
2641 */
2642void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2643{
2644 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2645}
2646EXPORT_SYMBOL(writeback_inodes_sb);
2647
2648/**
2649 * try_to_writeback_inodes_sb - try to start writeback if none underway
2650 * @sb: the superblock
2651 * @reason: reason why some writeback work was initiated
2652 *
2653 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2654 */
2655void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2656{
2657 if (!down_read_trylock(&sb->s_umount))
2658 return;
2659
2660 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2661 up_read(&sb->s_umount);
2662}
2663EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2664
2665/**
2666 * sync_inodes_sb - sync sb inode pages
2667 * @sb: the superblock
2668 *
2669 * This function writes and waits on any dirty inode belonging to this
2670 * super_block.
2671 */
2672void sync_inodes_sb(struct super_block *sb)
2673{
2674 struct backing_dev_info *bdi = sb->s_bdi;
2675 DEFINE_WB_COMPLETION(done, bdi);
2676 struct wb_writeback_work work = {
2677 .sb = sb,
2678 .sync_mode = WB_SYNC_ALL,
2679 .nr_pages = LONG_MAX,
2680 .range_cyclic = 0,
2681 .done = &done,
2682 .reason = WB_REASON_SYNC,
2683 .for_sync = 1,
2684 };
2685
2686 /*
2687 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2688 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2689 * bdi_has_dirty() need to be written out too.
2690 */
2691 if (bdi == &noop_backing_dev_info)
2692 return;
2693 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2694
2695 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2696 bdi_down_write_wb_switch_rwsem(bdi);
2697 bdi_split_work_to_wbs(bdi, &work, false);
2698 wb_wait_for_completion(&done);
2699 bdi_up_write_wb_switch_rwsem(bdi);
2700
2701 wait_sb_inodes(sb);
2702}
2703EXPORT_SYMBOL(sync_inodes_sb);
2704
2705/**
2706 * write_inode_now - write an inode to disk
2707 * @inode: inode to write to disk
2708 * @sync: whether the write should be synchronous or not
2709 *
2710 * This function commits an inode to disk immediately if it is dirty. This is
2711 * primarily needed by knfsd.
2712 *
2713 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2714 */
2715int write_inode_now(struct inode *inode, int sync)
2716{
2717 struct writeback_control wbc = {
2718 .nr_to_write = LONG_MAX,
2719 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2720 .range_start = 0,
2721 .range_end = LLONG_MAX,
2722 };
2723
2724 if (!mapping_can_writeback(inode->i_mapping))
2725 wbc.nr_to_write = 0;
2726
2727 might_sleep();
2728 return writeback_single_inode(inode, &wbc);
2729}
2730EXPORT_SYMBOL(write_inode_now);
2731
2732/**
2733 * sync_inode - write an inode and its pages to disk.
2734 * @inode: the inode to sync
2735 * @wbc: controls the writeback mode
2736 *
2737 * sync_inode() will write an inode and its pages to disk. It will also
2738 * correctly update the inode on its superblock's dirty inode lists and will
2739 * update inode->i_state.
2740 *
2741 * The caller must have a ref on the inode.
2742 */
2743int sync_inode(struct inode *inode, struct writeback_control *wbc)
2744{
2745 return writeback_single_inode(inode, wbc);
2746}
2747EXPORT_SYMBOL(sync_inode);
2748
2749/**
2750 * sync_inode_metadata - write an inode to disk
2751 * @inode: the inode to sync
2752 * @wait: wait for I/O to complete.
2753 *
2754 * Write an inode to disk and adjust its dirty state after completion.
2755 *
2756 * Note: only writes the actual inode, no associated data or other metadata.
2757 */
2758int sync_inode_metadata(struct inode *inode, int wait)
2759{
2760 struct writeback_control wbc = {
2761 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2762 .nr_to_write = 0, /* metadata-only */
2763 };
2764
2765 return sync_inode(inode, &wbc);
2766}
2767EXPORT_SYMBOL(sync_inode_metadata);