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