blob: 25458fa4753dc10ddb56c5a9be61c2aa224c0d5f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050017#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Wu Fengguangbc31b862012-01-07 20:41:55 -060023#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020024#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
Dave Chinner455b2862010-07-07 13:24:06 +100028#include <linux/tracepoint.h>
Al Viro719ea2f2013-09-29 11:24:49 -040029#include <linux/device.h>
Tejun Heo21c63212015-05-28 14:50:49 -040030#include <linux/memcontrol.h>
David Howells07f3f052006-09-30 20:52:18 +020031#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
Jens Axboed0bceac2009-05-18 08:20:32 +020033/*
Wu Fengguangbc31b862012-01-07 20:41:55 -060034 * 4MB minimal write chunk size
35 */
36#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
37
Tejun Heocc395d72015-05-22 17:13:58 -040038struct wb_completion {
39 atomic_t cnt;
40};
41
Wu Fengguangbc31b862012-01-07 20:41:55 -060042/*
Jens Axboec4a77a62009-09-16 15:18:25 +020043 * Passed into wb_writeback(), essentially a subset of writeback_control
44 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020045struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020046 long nr_pages;
47 struct super_block *sb;
Jan Kara0dc83bd2014-02-21 11:19:04 +010048 unsigned long *older_than_this;
Jens Axboec4a77a62009-09-16 15:18:25 +020049 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060050 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050051 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
Dave Chinner7747bd42013-07-02 22:38:35 +100054 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
Tejun Heoac7b19a2015-05-22 17:13:57 -040055 unsigned int auto_free:1; /* free on completion */
Tejun Heo98754bf2015-05-22 17:13:59 -040056 unsigned int single_wait:1;
57 unsigned int single_done:1;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060058 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020059
Jens Axboe8010c3b2009-09-15 20:04:57 +020060 struct list_head list; /* pending work list */
Tejun Heocc395d72015-05-22 17:13:58 -040061 struct wb_completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020062};
63
Theodore Ts'oa2f48702015-03-17 12:23:19 -040064/*
Tejun Heocc395d72015-05-22 17:13:58 -040065 * If one wants to wait for one or more wb_writeback_works, each work's
66 * ->done should be set to a wb_completion defined using the following
67 * macro. Once all work items are issued with wb_queue_work(), the caller
68 * can wait for the completion of all using wb_wait_for_completion(). Work
69 * items which are waited upon aren't freed automatically on completion.
70 */
71#define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
72 struct wb_completion cmpl = { \
73 .cnt = ATOMIC_INIT(1), \
74 }
75
76
77/*
Theodore Ts'oa2f48702015-03-17 12:23:19 -040078 * If an inode is constantly having its pages dirtied, but then the
79 * updates stop dirtytime_expire_interval seconds in the past, it's
80 * possible for the worst case time between when an inode has its
81 * timestamps updated and when they finally get written out to be two
82 * dirtytime_expire_intervals. We set the default to 12 hours (in
83 * seconds), which means most of the time inodes will have their
84 * timestamps written to disk after 12 hours, but in the worst case a
85 * few inodes might not their timestamps updated for 24 hours.
86 */
87unsigned int dirtytime_expire_interval = 12 * 60 * 60;
88
Nick Piggin7ccf19a2010-10-21 11:49:30 +110089static inline struct inode *wb_inode(struct list_head *head)
90{
91 return list_entry(head, struct inode, i_wb_list);
92}
93
Wu Fengguang15eb77a2012-01-17 11:18:56 -060094/*
95 * Include the creation of the trace points after defining the
96 * wb_writeback_work structure and inline functions so that the definition
97 * remains local to this file.
98 */
99#define CREATE_TRACE_POINTS
100#include <trace/events/writeback.h>
101
Steven Whitehouse774016b2014-02-06 15:47:47 +0000102EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
103
Tejun Heod6c10f12015-05-22 17:13:45 -0400104static bool wb_io_lists_populated(struct bdi_writeback *wb)
105{
106 if (wb_has_dirty_io(wb)) {
107 return false;
108 } else {
109 set_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400110 WARN_ON_ONCE(!wb->avg_write_bandwidth);
Tejun Heo766a9d62015-05-22 17:13:46 -0400111 atomic_long_add(wb->avg_write_bandwidth,
112 &wb->bdi->tot_write_bandwidth);
Tejun Heod6c10f12015-05-22 17:13:45 -0400113 return true;
114 }
115}
116
117static void wb_io_lists_depopulated(struct bdi_writeback *wb)
118{
119 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
Tejun Heo766a9d62015-05-22 17:13:46 -0400120 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
Tejun Heod6c10f12015-05-22 17:13:45 -0400121 clear_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400122 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
123 &wb->bdi->tot_write_bandwidth) < 0);
Tejun Heo766a9d62015-05-22 17:13:46 -0400124 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400125}
126
127/**
128 * inode_wb_list_move_locked - move an inode onto a bdi_writeback IO list
129 * @inode: inode to be moved
130 * @wb: target bdi_writeback
131 * @head: one of @wb->b_{dirty|io|more_io}
132 *
133 * Move @inode->i_wb_list to @list of @wb and set %WB_has_dirty_io.
134 * Returns %true if @inode is the first occupant of the !dirty_time IO
135 * lists; otherwise, %false.
136 */
137static bool inode_wb_list_move_locked(struct inode *inode,
138 struct bdi_writeback *wb,
139 struct list_head *head)
140{
141 assert_spin_locked(&wb->list_lock);
142
143 list_move(&inode->i_wb_list, head);
144
145 /* dirty_time doesn't count as dirty_io until expiration */
146 if (head != &wb->b_dirty_time)
147 return wb_io_lists_populated(wb);
148
149 wb_io_lists_depopulated(wb);
150 return false;
151}
152
153/**
154 * inode_wb_list_del_locked - remove an inode from its bdi_writeback IO list
155 * @inode: inode to be removed
156 * @wb: bdi_writeback @inode is being removed from
157 *
158 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
159 * clear %WB_has_dirty_io if all are empty afterwards.
160 */
161static void inode_wb_list_del_locked(struct inode *inode,
162 struct bdi_writeback *wb)
163{
164 assert_spin_locked(&wb->list_lock);
165
166 list_del_init(&inode->i_wb_list);
167 wb_io_lists_depopulated(wb);
168}
169
Tejun Heof0054bb2015-05-22 17:13:30 -0400170static void wb_wakeup(struct bdi_writeback *wb)
Jan Kara5acda9d2014-04-03 14:46:23 -0700171{
Tejun Heof0054bb2015-05-22 17:13:30 -0400172 spin_lock_bh(&wb->work_lock);
173 if (test_bit(WB_registered, &wb->state))
174 mod_delayed_work(bdi_wq, &wb->dwork, 0);
175 spin_unlock_bh(&wb->work_lock);
Jan Kara5acda9d2014-04-03 14:46:23 -0700176}
177
Tejun Heof0054bb2015-05-22 17:13:30 -0400178static void wb_queue_work(struct bdi_writeback *wb,
179 struct wb_writeback_work *work)
Jan Kara65850272011-01-13 15:45:44 -0800180{
Tejun Heof0054bb2015-05-22 17:13:30 -0400181 trace_writeback_queue(wb->bdi, work);
Jan Kara65850272011-01-13 15:45:44 -0800182
Tejun Heof0054bb2015-05-22 17:13:30 -0400183 spin_lock_bh(&wb->work_lock);
Tejun Heo98754bf2015-05-22 17:13:59 -0400184 if (!test_bit(WB_registered, &wb->state)) {
185 if (work->single_wait)
186 work->single_done = 1;
Jan Kara5acda9d2014-04-03 14:46:23 -0700187 goto out_unlock;
Tejun Heo98754bf2015-05-22 17:13:59 -0400188 }
Tejun Heocc395d72015-05-22 17:13:58 -0400189 if (work->done)
190 atomic_inc(&work->done->cnt);
Tejun Heof0054bb2015-05-22 17:13:30 -0400191 list_add_tail(&work->list, &wb->work_list);
192 mod_delayed_work(bdi_wq, &wb->dwork, 0);
Jan Kara5acda9d2014-04-03 14:46:23 -0700193out_unlock:
Tejun Heof0054bb2015-05-22 17:13:30 -0400194 spin_unlock_bh(&wb->work_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
Tejun Heocc395d72015-05-22 17:13:58 -0400197/**
198 * wb_wait_for_completion - wait for completion of bdi_writeback_works
199 * @bdi: bdi work items were issued to
200 * @done: target wb_completion
201 *
202 * Wait for one or more work items issued to @bdi with their ->done field
203 * set to @done, which should have been defined with
204 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
205 * work items are completed. Work items which are waited upon aren't freed
206 * automatically on completion.
207 */
208static void wb_wait_for_completion(struct backing_dev_info *bdi,
209 struct wb_completion *done)
210{
211 atomic_dec(&done->cnt); /* put down the initial count */
212 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
213}
214
Tejun Heo703c2702015-05-22 17:13:44 -0400215#ifdef CONFIG_CGROUP_WRITEBACK
216
Tejun Heo2a814902015-05-28 14:50:51 -0400217/* parameters for foreign inode detection, see wb_detach_inode() */
218#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
219#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
220#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
221#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
222
223#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
224#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
225 /* each slot's duration is 2s / 16 */
226#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
227 /* if foreign slots >= 8, switch */
228#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
229 /* one round can affect upto 5 slots */
230
Tejun Heo21c63212015-05-28 14:50:49 -0400231void __inode_attach_wb(struct inode *inode, struct page *page)
232{
233 struct backing_dev_info *bdi = inode_to_bdi(inode);
234 struct bdi_writeback *wb = NULL;
235
236 if (inode_cgwb_enabled(inode)) {
237 struct cgroup_subsys_state *memcg_css;
238
239 if (page) {
240 memcg_css = mem_cgroup_css_from_page(page);
241 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
242 } else {
243 /* must pin memcg_css, see wb_get_create() */
244 memcg_css = task_get_css(current, memory_cgrp_id);
245 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
246 css_put(memcg_css);
247 }
248 }
249
250 if (!wb)
251 wb = &bdi->wb;
252
253 /*
254 * There may be multiple instances of this function racing to
255 * update the same inode. Use cmpxchg() to tell the winner.
256 */
257 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
258 wb_put(wb);
259}
260
Tejun Heo703c2702015-05-22 17:13:44 -0400261/**
Tejun Heo87e1d782015-05-28 14:50:52 -0400262 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
263 * @inode: inode of interest with i_lock held
264 *
265 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
266 * held on entry and is released on return. The returned wb is guaranteed
267 * to stay @inode's associated wb until its list_lock is released.
268 */
269static struct bdi_writeback *
270locked_inode_to_wb_and_lock_list(struct inode *inode)
271 __releases(&inode->i_lock)
272 __acquires(&wb->list_lock)
273{
274 while (true) {
275 struct bdi_writeback *wb = inode_to_wb(inode);
276
277 /*
278 * inode_to_wb() association is protected by both
279 * @inode->i_lock and @wb->list_lock but list_lock nests
280 * outside i_lock. Drop i_lock and verify that the
281 * association hasn't changed after acquiring list_lock.
282 */
283 wb_get(wb);
284 spin_unlock(&inode->i_lock);
285 spin_lock(&wb->list_lock);
286 wb_put(wb); /* not gonna deref it anymore */
287
288 if (likely(wb == inode_to_wb(inode)))
289 return wb; /* @inode already has ref */
290
291 spin_unlock(&wb->list_lock);
292 cpu_relax();
293 spin_lock(&inode->i_lock);
294 }
295}
296
297/**
298 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
299 * @inode: inode of interest
300 *
301 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
302 * on entry.
303 */
304static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
305 __acquires(&wb->list_lock)
306{
307 spin_lock(&inode->i_lock);
308 return locked_inode_to_wb_and_lock_list(inode);
309}
310
Tejun Heo682aa8e2015-05-28 14:50:53 -0400311struct inode_switch_wbs_context {
312 struct inode *inode;
313 struct bdi_writeback *new_wb;
314
315 struct rcu_head rcu_head;
316 struct work_struct work;
317};
318
319static void inode_switch_wbs_work_fn(struct work_struct *work)
320{
321 struct inode_switch_wbs_context *isw =
322 container_of(work, struct inode_switch_wbs_context, work);
323 struct inode *inode = isw->inode;
324 struct bdi_writeback *new_wb = isw->new_wb;
325
326 /*
327 * By the time control reaches here, RCU grace period has passed
328 * since I_WB_SWITCH assertion and all wb stat update transactions
329 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
330 * synchronizing against mapping->tree_lock.
331 */
332 spin_lock(&inode->i_lock);
333
334 inode->i_wb_frn_winner = 0;
335 inode->i_wb_frn_avg_time = 0;
336 inode->i_wb_frn_history = 0;
337
338 /*
339 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
340 * ensures that the new wb is visible if they see !I_WB_SWITCH.
341 */
342 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
343
344 spin_unlock(&inode->i_lock);
345
346 iput(inode);
347 wb_put(new_wb);
348 kfree(isw);
349}
350
351static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
352{
353 struct inode_switch_wbs_context *isw = container_of(rcu_head,
354 struct inode_switch_wbs_context, rcu_head);
355
356 /* needs to grab bh-unsafe locks, bounce to work item */
357 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
358 schedule_work(&isw->work);
359}
360
361/**
362 * inode_switch_wbs - change the wb association of an inode
363 * @inode: target inode
364 * @new_wb_id: ID of the new wb
365 *
366 * Switch @inode's wb association to the wb identified by @new_wb_id. The
367 * switching is performed asynchronously and may fail silently.
368 */
369static void inode_switch_wbs(struct inode *inode, int new_wb_id)
370{
371 struct backing_dev_info *bdi = inode_to_bdi(inode);
372 struct cgroup_subsys_state *memcg_css;
373 struct inode_switch_wbs_context *isw;
374
375 /* noop if seems to be already in progress */
376 if (inode->i_state & I_WB_SWITCH)
377 return;
378
379 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
380 if (!isw)
381 return;
382
383 /* find and pin the new wb */
384 rcu_read_lock();
385 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
386 if (memcg_css)
387 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
388 rcu_read_unlock();
389 if (!isw->new_wb)
390 goto out_free;
391
392 /* while holding I_WB_SWITCH, no one else can update the association */
393 spin_lock(&inode->i_lock);
394 if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
395 inode_to_wb(inode) == isw->new_wb) {
396 spin_unlock(&inode->i_lock);
397 goto out_free;
398 }
399 inode->i_state |= I_WB_SWITCH;
400 spin_unlock(&inode->i_lock);
401
402 ihold(inode);
403 isw->inode = inode;
404
405 /*
406 * In addition to synchronizing among switchers, I_WB_SWITCH tells
407 * the RCU protected stat update paths to grab the mapping's
408 * tree_lock so that stat transfer can synchronize against them.
409 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
410 */
411 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
412 return;
413
414out_free:
415 if (isw->new_wb)
416 wb_put(isw->new_wb);
417 kfree(isw);
418}
419
Tejun Heo87e1d782015-05-28 14:50:52 -0400420/**
Tejun Heob16b1de2015-06-02 08:39:48 -0600421 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
422 * @wbc: writeback_control of interest
423 * @inode: target inode
424 *
425 * @inode is locked and about to be written back under the control of @wbc.
426 * Record @inode's writeback context into @wbc and unlock the i_lock. On
427 * writeback completion, wbc_detach_inode() should be called. This is used
428 * to track the cgroup writeback context.
429 */
430void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
431 struct inode *inode)
432{
433 wbc->wb = inode_to_wb(inode);
Tejun Heo2a814902015-05-28 14:50:51 -0400434 wbc->inode = inode;
435
436 wbc->wb_id = wbc->wb->memcg_css->id;
437 wbc->wb_lcand_id = inode->i_wb_frn_winner;
438 wbc->wb_tcand_id = 0;
439 wbc->wb_bytes = 0;
440 wbc->wb_lcand_bytes = 0;
441 wbc->wb_tcand_bytes = 0;
442
Tejun Heob16b1de2015-06-02 08:39:48 -0600443 wb_get(wbc->wb);
444 spin_unlock(&inode->i_lock);
445}
446
447/**
Tejun Heo2a814902015-05-28 14:50:51 -0400448 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
449 * @wbc: writeback_control of the just finished writeback
Tejun Heob16b1de2015-06-02 08:39:48 -0600450 *
451 * To be called after a writeback attempt of an inode finishes and undoes
452 * wbc_attach_and_unlock_inode(). Can be called under any context.
Tejun Heo2a814902015-05-28 14:50:51 -0400453 *
454 * As concurrent write sharing of an inode is expected to be very rare and
455 * memcg only tracks page ownership on first-use basis severely confining
456 * the usefulness of such sharing, cgroup writeback tracks ownership
457 * per-inode. While the support for concurrent write sharing of an inode
458 * is deemed unnecessary, an inode being written to by different cgroups at
459 * different points in time is a lot more common, and, more importantly,
460 * charging only by first-use can too readily lead to grossly incorrect
461 * behaviors (single foreign page can lead to gigabytes of writeback to be
462 * incorrectly attributed).
463 *
464 * To resolve this issue, cgroup writeback detects the majority dirtier of
465 * an inode and transfers the ownership to it. To avoid unnnecessary
466 * oscillation, the detection mechanism keeps track of history and gives
467 * out the switch verdict only if the foreign usage pattern is stable over
468 * a certain amount of time and/or writeback attempts.
469 *
470 * On each writeback attempt, @wbc tries to detect the majority writer
471 * using Boyer-Moore majority vote algorithm. In addition to the byte
472 * count from the majority voting, it also counts the bytes written for the
473 * current wb and the last round's winner wb (max of last round's current
474 * wb, the winner from two rounds ago, and the last round's majority
475 * candidate). Keeping track of the historical winner helps the algorithm
476 * to semi-reliably detect the most active writer even when it's not the
477 * absolute majority.
478 *
479 * Once the winner of the round is determined, whether the winner is
480 * foreign or not and how much IO time the round consumed is recorded in
481 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
482 * over a certain threshold, the switch verdict is given.
Tejun Heob16b1de2015-06-02 08:39:48 -0600483 */
484void wbc_detach_inode(struct writeback_control *wbc)
485{
Tejun Heo2a814902015-05-28 14:50:51 -0400486 struct bdi_writeback *wb = wbc->wb;
487 struct inode *inode = wbc->inode;
488 u16 history = inode->i_wb_frn_history;
489 unsigned long avg_time = inode->i_wb_frn_avg_time;
490 unsigned long max_bytes, max_time;
491 int max_id;
492
493 /* pick the winner of this round */
494 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
495 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
496 max_id = wbc->wb_id;
497 max_bytes = wbc->wb_bytes;
498 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
499 max_id = wbc->wb_lcand_id;
500 max_bytes = wbc->wb_lcand_bytes;
501 } else {
502 max_id = wbc->wb_tcand_id;
503 max_bytes = wbc->wb_tcand_bytes;
504 }
505
506 /*
507 * Calculate the amount of IO time the winner consumed and fold it
508 * into the running average kept per inode. If the consumed IO
509 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
510 * deciding whether to switch or not. This is to prevent one-off
511 * small dirtiers from skewing the verdict.
512 */
513 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
514 wb->avg_write_bandwidth);
515 if (avg_time)
516 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
517 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
518 else
519 avg_time = max_time; /* immediate catch up on first run */
520
521 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
522 int slots;
523
524 /*
525 * The switch verdict is reached if foreign wb's consume
526 * more than a certain proportion of IO time in a
527 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
528 * history mask where each bit represents one sixteenth of
529 * the period. Determine the number of slots to shift into
530 * history from @max_time.
531 */
532 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
533 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
534 history <<= slots;
535 if (wbc->wb_id != max_id)
536 history |= (1U << slots) - 1;
537
538 /*
539 * Switch if the current wb isn't the consistent winner.
540 * If there are multiple closely competing dirtiers, the
541 * inode may switch across them repeatedly over time, which
542 * is okay. The main goal is avoiding keeping an inode on
543 * the wrong wb for an extended period of time.
544 */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400545 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
546 inode_switch_wbs(inode, max_id);
Tejun Heo2a814902015-05-28 14:50:51 -0400547 }
548
549 /*
550 * Multiple instances of this function may race to update the
551 * following fields but we don't mind occassional inaccuracies.
552 */
553 inode->i_wb_frn_winner = max_id;
554 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
555 inode->i_wb_frn_history = history;
556
Tejun Heob16b1de2015-06-02 08:39:48 -0600557 wb_put(wbc->wb);
558 wbc->wb = NULL;
559}
560
561/**
Tejun Heo2a814902015-05-28 14:50:51 -0400562 * wbc_account_io - account IO issued during writeback
563 * @wbc: writeback_control of the writeback in progress
564 * @page: page being written out
565 * @bytes: number of bytes being written out
566 *
567 * @bytes from @page are about to written out during the writeback
568 * controlled by @wbc. Keep the book for foreign inode detection. See
569 * wbc_detach_inode().
570 */
571void wbc_account_io(struct writeback_control *wbc, struct page *page,
572 size_t bytes)
573{
574 int id;
575
576 /*
577 * pageout() path doesn't attach @wbc to the inode being written
578 * out. This is intentional as we don't want the function to block
579 * behind a slow cgroup. Ultimately, we want pageout() to kick off
580 * regular writeback instead of writing things out itself.
581 */
582 if (!wbc->wb)
583 return;
584
585 rcu_read_lock();
586 id = mem_cgroup_css_from_page(page)->id;
587 rcu_read_unlock();
588
589 if (id == wbc->wb_id) {
590 wbc->wb_bytes += bytes;
591 return;
592 }
593
594 if (id == wbc->wb_lcand_id)
595 wbc->wb_lcand_bytes += bytes;
596
597 /* Boyer-Moore majority vote algorithm */
598 if (!wbc->wb_tcand_bytes)
599 wbc->wb_tcand_id = id;
600 if (id == wbc->wb_tcand_id)
601 wbc->wb_tcand_bytes += bytes;
602 else
603 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
604}
605
606/**
Tejun Heo703c2702015-05-22 17:13:44 -0400607 * inode_congested - test whether an inode is congested
608 * @inode: inode to test for congestion
609 * @cong_bits: mask of WB_[a]sync_congested bits to test
610 *
611 * Tests whether @inode is congested. @cong_bits is the mask of congestion
612 * bits to test and the return value is the mask of set bits.
613 *
614 * If cgroup writeback is enabled for @inode, the congestion state is
615 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
616 * associated with @inode is congested; otherwise, the root wb's congestion
617 * state is used.
618 */
619int inode_congested(struct inode *inode, int cong_bits)
620{
Tejun Heo5cb8b822015-05-28 14:50:54 -0400621 /*
622 * Once set, ->i_wb never becomes NULL while the inode is alive.
623 * Start transaction iff ->i_wb is visible.
624 */
625 if (inode && inode_to_wb(inode)) {
626 struct bdi_writeback *wb;
627 bool locked, congested;
628
629 wb = unlocked_inode_to_wb_begin(inode, &locked);
630 congested = wb_congested(wb, cong_bits);
631 unlocked_inode_to_wb_end(inode, locked);
632 return congested;
Tejun Heo703c2702015-05-22 17:13:44 -0400633 }
634
635 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
636}
637EXPORT_SYMBOL_GPL(inode_congested);
638
Tejun Heof2b65122015-05-22 17:13:55 -0400639/**
Tejun Heo98754bf2015-05-22 17:13:59 -0400640 * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work
641 * @bdi: bdi the work item was issued to
642 * @work: work item to wait for
643 *
644 * Wait for the completion of @work which was issued to one of @bdi's
645 * bdi_writeback's. The caller must have set @work->single_wait before
646 * issuing it. This wait operates independently fo
647 * wb_wait_for_completion() and also disables automatic freeing of @work.
648 */
649static void wb_wait_for_single_work(struct backing_dev_info *bdi,
650 struct wb_writeback_work *work)
651{
652 if (WARN_ON_ONCE(!work->single_wait))
653 return;
654
655 wait_event(bdi->wb_waitq, work->single_done);
656
657 /*
658 * Paired with smp_wmb() in wb_do_writeback() and ensures that all
659 * modifications to @work prior to assertion of ->single_done is
660 * visible to the caller once this function returns.
661 */
662 smp_rmb();
663}
664
665/**
Tejun Heof2b65122015-05-22 17:13:55 -0400666 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
667 * @wb: target bdi_writeback to split @nr_pages to
668 * @nr_pages: number of pages to write for the whole bdi
669 *
670 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
671 * relation to the total write bandwidth of all wb's w/ dirty inodes on
672 * @wb->bdi.
673 */
674static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
675{
676 unsigned long this_bw = wb->avg_write_bandwidth;
677 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
678
679 if (nr_pages == LONG_MAX)
680 return LONG_MAX;
681
682 /*
683 * This may be called on clean wb's and proportional distribution
684 * may not make sense, just use the original @nr_pages in those
685 * cases. In general, we wanna err on the side of writing more.
686 */
687 if (!tot_bw || this_bw >= tot_bw)
688 return nr_pages;
689 else
690 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
691}
692
Tejun Heodb125362015-05-22 17:14:01 -0400693/**
694 * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb
695 * @wb: target bdi_writeback
696 * @base_work: source wb_writeback_work
697 *
698 * Try to make a clone of @base_work and issue it to @wb. If cloning
699 * succeeds, %true is returned; otherwise, @base_work is issued directly
700 * and %false is returned. In the latter case, the caller is required to
701 * wait for @base_work's completion using wb_wait_for_single_work().
702 *
703 * A clone is auto-freed on completion. @base_work never is.
704 */
705static bool wb_clone_and_queue_work(struct bdi_writeback *wb,
706 struct wb_writeback_work *base_work)
707{
708 struct wb_writeback_work *work;
709
710 work = kmalloc(sizeof(*work), GFP_ATOMIC);
711 if (work) {
712 *work = *base_work;
713 work->auto_free = 1;
714 work->single_wait = 0;
715 } else {
716 work = base_work;
717 work->auto_free = 0;
718 work->single_wait = 1;
719 }
720 work->single_done = 0;
721 wb_queue_work(wb, work);
722 return work != base_work;
723}
724
725/**
726 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
727 * @bdi: target backing_dev_info
728 * @base_work: wb_writeback_work to issue
729 * @skip_if_busy: skip wb's which already have writeback in progress
730 *
731 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
732 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
733 * distributed to the busy wbs according to each wb's proportion in the
734 * total active write bandwidth of @bdi.
735 */
736static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
737 struct wb_writeback_work *base_work,
738 bool skip_if_busy)
739{
740 long nr_pages = base_work->nr_pages;
741 int next_blkcg_id = 0;
742 struct bdi_writeback *wb;
743 struct wb_iter iter;
744
745 might_sleep();
746
747 if (!bdi_has_dirty_io(bdi))
748 return;
749restart:
750 rcu_read_lock();
751 bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) {
752 if (!wb_has_dirty_io(wb) ||
753 (skip_if_busy && writeback_in_progress(wb)))
754 continue;
755
756 base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages);
757 if (!wb_clone_and_queue_work(wb, base_work)) {
758 next_blkcg_id = wb->blkcg_css->id + 1;
759 rcu_read_unlock();
760 wb_wait_for_single_work(bdi, base_work);
761 goto restart;
762 }
763 }
764 rcu_read_unlock();
765}
766
Tejun Heof2b65122015-05-22 17:13:55 -0400767#else /* CONFIG_CGROUP_WRITEBACK */
768
Tejun Heo87e1d782015-05-28 14:50:52 -0400769static struct bdi_writeback *
770locked_inode_to_wb_and_lock_list(struct inode *inode)
771 __releases(&inode->i_lock)
772 __acquires(&wb->list_lock)
773{
774 struct bdi_writeback *wb = inode_to_wb(inode);
775
776 spin_unlock(&inode->i_lock);
777 spin_lock(&wb->list_lock);
778 return wb;
779}
780
781static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
782 __acquires(&wb->list_lock)
783{
784 struct bdi_writeback *wb = inode_to_wb(inode);
785
786 spin_lock(&wb->list_lock);
787 return wb;
788}
789
Tejun Heof2b65122015-05-22 17:13:55 -0400790static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
791{
792 return nr_pages;
793}
794
Tejun Heodb125362015-05-22 17:14:01 -0400795static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
796 struct wb_writeback_work *base_work,
797 bool skip_if_busy)
798{
799 might_sleep();
800
801 if (bdi_has_dirty_io(bdi) &&
802 (!skip_if_busy || !writeback_in_progress(&bdi->wb))) {
803 base_work->auto_free = 0;
804 base_work->single_wait = 0;
805 base_work->single_done = 0;
806 wb_queue_work(&bdi->wb, base_work);
807 }
808}
809
Tejun Heo703c2702015-05-22 17:13:44 -0400810#endif /* CONFIG_CGROUP_WRITEBACK */
811
Tejun Heoc00ddad2015-05-22 17:13:51 -0400812void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
813 bool range_cyclic, enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200814{
Tejun Heoc00ddad2015-05-22 17:13:51 -0400815 struct wb_writeback_work *work;
816
817 if (!wb_has_dirty_io(wb))
818 return;
819
820 /*
821 * This is WB_SYNC_NONE writeback, so if allocation fails just
822 * wakeup the thread for old dirty data writeback
823 */
824 work = kzalloc(sizeof(*work), GFP_ATOMIC);
825 if (!work) {
826 trace_writeback_nowork(wb->bdi);
827 wb_wakeup(wb);
828 return;
829 }
830
831 work->sync_mode = WB_SYNC_NONE;
832 work->nr_pages = nr_pages;
833 work->range_cyclic = range_cyclic;
834 work->reason = reason;
Tejun Heoac7b19a2015-05-22 17:13:57 -0400835 work->auto_free = 1;
Tejun Heoc00ddad2015-05-22 17:13:51 -0400836
837 wb_queue_work(wb, work);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200838}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800839
Christoph Hellwigc5444192010-06-08 18:15:15 +0200840/**
Tejun Heo9ecf48662015-05-22 17:13:54 -0400841 * wb_start_background_writeback - start background writeback
842 * @wb: bdi_writback to write from
Christoph Hellwigc5444192010-06-08 18:15:15 +0200843 *
844 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800845 * This makes sure WB_SYNC_NONE background writeback happens. When
Tejun Heo9ecf48662015-05-22 17:13:54 -0400846 * this function returns, it is only guaranteed that for given wb
Jan Kara65850272011-01-13 15:45:44 -0800847 * some IO is happening if we are over background dirty threshold.
848 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200849 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400850void wb_start_background_writeback(struct bdi_writeback *wb)
Christoph Hellwigc5444192010-06-08 18:15:15 +0200851{
Jan Kara65850272011-01-13 15:45:44 -0800852 /*
853 * We just wake up the flusher thread. It will perform background
854 * writeback as soon as there is no other work to do.
855 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400856 trace_writeback_wake_background(wb->bdi);
857 wb_wakeup(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858}
859
860/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100861 * Remove the inode from the writeback list it is on.
862 */
863void inode_wb_list_del(struct inode *inode)
864{
Tejun Heo87e1d782015-05-28 14:50:52 -0400865 struct bdi_writeback *wb;
Dave Chinnera66979a2011-03-22 22:23:41 +1100866
Tejun Heo87e1d782015-05-28 14:50:52 -0400867 wb = inode_to_wb_and_lock_list(inode);
Tejun Heod6c10f12015-05-22 17:13:45 -0400868 inode_wb_list_del_locked(inode, wb);
Tejun Heo52ebea72015-05-22 17:13:37 -0400869 spin_unlock(&wb->list_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600870}
Dave Chinnera66979a2011-03-22 22:23:41 +1100871
872/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700873 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
874 * furthest end of its superblock's dirty-inode list.
875 *
876 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200877 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700878 * the case then the inode must have been redirtied while it was being written
879 * out and we don't reset its dirtied_when.
880 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600881static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700882{
Jens Axboe03ba3782009-09-09 09:08:54 +0200883 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200884 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700885
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100886 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200887 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700888 inode->dirtied_when = jiffies;
889 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400890 inode_wb_list_move_locked(inode, wb, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700891}
892
893/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200894 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700895 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600896static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700897{
Tejun Heod6c10f12015-05-22 17:13:45 -0400898 inode_wb_list_move_locked(inode, wb, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700899}
900
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700901static void inode_sync_complete(struct inode *inode)
902{
Jan Kara365b94ae2012-05-03 14:47:55 +0200903 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -0800904 /* If inode is clean an unused, put it into LRU now... */
905 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +0200906 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700907 smp_mb();
908 wake_up_bit(&inode->i_state, __I_SYNC);
909}
910
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700911static bool inode_dirtied_after(struct inode *inode, unsigned long t)
912{
913 bool ret = time_after(inode->dirtied_when, t);
914#ifndef CONFIG_64BIT
915 /*
916 * For inodes being constantly redirtied, dirtied_when can get stuck.
917 * It _appears_ to be in the future, but is actually in distant past.
918 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200919 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700920 */
921 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
922#endif
923 return ret;
924}
925
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500926#define EXPIRE_DIRTY_ATIME 0x0001
927
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700928/*
Wang Sheng-Hui0e2f2b22012-09-11 08:28:18 +0800929 * Move expired (dirtied before work->older_than_this) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -0800930 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -0700931 */
Wu Fengguange84d0a42011-04-23 12:27:27 -0600932static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -0700933 struct list_head *dispatch_queue,
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500934 int flags,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600935 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700936{
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500937 unsigned long *older_than_this = NULL;
938 unsigned long expire_time;
Shaohua Li5c034492009-09-24 14:42:33 +0200939 LIST_HEAD(tmp);
940 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200941 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200942 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200943 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -0600944 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200945
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500946 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
947 older_than_this = work->older_than_this;
Theodore Ts'oa2f48702015-03-17 12:23:19 -0400948 else if (!work->for_sync) {
949 expire_time = jiffies - (dirtytime_expire_interval * HZ);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500950 older_than_this = &expire_time;
951 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700952 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100953 inode = wb_inode(delaying_queue->prev);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500954 if (older_than_this &&
955 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700956 break;
Jan Karaa8855992013-07-09 22:36:45 +0800957 list_move(&inode->i_wb_list, &tmp);
958 moved++;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -0500959 if (flags & EXPIRE_DIRTY_ATIME)
960 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
Jan Karaa8855992013-07-09 22:36:45 +0800961 if (sb_is_blkdev_sb(inode->i_sb))
962 continue;
Jens Axboecf137302009-09-24 15:12:57 +0200963 if (sb && sb != inode->i_sb)
964 do_sb_sort = 1;
965 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200966 }
967
Jens Axboecf137302009-09-24 15:12:57 +0200968 /* just one sb in list, splice to dispatch_queue and we're done */
969 if (!do_sb_sort) {
970 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -0600971 goto out;
Jens Axboecf137302009-09-24 15:12:57 +0200972 }
973
Shaohua Li5c034492009-09-24 14:42:33 +0200974 /* Move inodes from one superblock together */
975 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100976 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200977 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100978 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +0200979 if (inode->i_sb == sb)
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100980 list_move(&inode->i_wb_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +0200981 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700982 }
Wu Fengguange84d0a42011-04-23 12:27:27 -0600983out:
984 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -0700985}
986
987/*
988 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700989 * Before
990 * newly dirtied b_dirty b_io b_more_io
991 * =============> gf edc BA
992 * After
993 * newly dirtied b_dirty b_io b_more_io
994 * =============> g fBAedc
995 * |
996 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -0700997 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600998static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700999{
Wu Fengguange84d0a42011-04-23 12:27:27 -06001000 int moved;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001001
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001002 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001003 list_splice_init(&wb->b_more_io, &wb->b_io);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001004 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1005 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1006 EXPIRE_DIRTY_ATIME, work);
Tejun Heod6c10f12015-05-22 17:13:45 -04001007 if (moved)
1008 wb_io_lists_populated(wb);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001009 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001010}
1011
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001012static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001013{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001014 int ret;
1015
1016 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1017 trace_writeback_write_inode_start(inode, wbc);
1018 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1019 trace_writeback_write_inode(inode, wbc);
1020 return ret;
1021 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001022 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -07001023}
1024
1025/*
Jan Kara169ebd92012-05-03 14:48:03 +02001026 * Wait for writeback on an inode to complete. Called with i_lock held.
1027 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +02001028 */
Jan Kara169ebd92012-05-03 14:48:03 +02001029static void __inode_wait_for_writeback(struct inode *inode)
1030 __releases(inode->i_lock)
1031 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +02001032{
1033 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1034 wait_queue_head_t *wqh;
1035
1036 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +11001037 while (inode->i_state & I_SYNC) {
1038 spin_unlock(&inode->i_lock);
NeilBrown74316202014-07-07 15:16:04 +10001039 __wait_on_bit(wqh, &wq, bit_wait,
1040 TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +11001041 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -07001042 }
Christoph Hellwig01c03192009-06-08 13:35:40 +02001043}
1044
1045/*
Jan Kara169ebd92012-05-03 14:48:03 +02001046 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1047 */
1048void inode_wait_for_writeback(struct inode *inode)
1049{
1050 spin_lock(&inode->i_lock);
1051 __inode_wait_for_writeback(inode);
1052 spin_unlock(&inode->i_lock);
1053}
1054
1055/*
1056 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1057 * held and drops it. It is aimed for callers not holding any inode reference
1058 * so once i_lock is dropped, inode can go away.
1059 */
1060static void inode_sleep_on_writeback(struct inode *inode)
1061 __releases(inode->i_lock)
1062{
1063 DEFINE_WAIT(wait);
1064 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1065 int sleep;
1066
1067 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1068 sleep = inode->i_state & I_SYNC;
1069 spin_unlock(&inode->i_lock);
1070 if (sleep)
1071 schedule();
1072 finish_wait(wqh, &wait);
1073}
1074
1075/*
Jan Karaccb26b52012-05-03 14:47:58 +02001076 * Find proper writeback list for the inode depending on its current state and
1077 * possibly also change of its state while we were doing writeback. Here we
1078 * handle things such as livelock prevention or fairness of writeback among
1079 * inodes. This function can be called only by flusher thread - noone else
1080 * processes all inodes in writeback lists and requeueing inodes behind flusher
1081 * thread's back can have unexpected consequences.
1082 */
1083static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1084 struct writeback_control *wbc)
1085{
1086 if (inode->i_state & I_FREEING)
1087 return;
1088
1089 /*
1090 * Sync livelock prevention. Each inode is tagged and synced in one
1091 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1092 * the dirty time to prevent enqueue and sync it again.
1093 */
1094 if ((inode->i_state & I_DIRTY) &&
1095 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1096 inode->dirtied_when = jiffies;
1097
Jan Kara4f8ad652012-05-03 14:48:00 +02001098 if (wbc->pages_skipped) {
1099 /*
1100 * writeback is not making progress due to locked
1101 * buffers. Skip this inode for now.
1102 */
1103 redirty_tail(inode, wb);
1104 return;
1105 }
1106
Jan Karaccb26b52012-05-03 14:47:58 +02001107 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1108 /*
1109 * We didn't write back all the pages. nfs_writepages()
1110 * sometimes bales out without doing anything.
1111 */
1112 if (wbc->nr_to_write <= 0) {
1113 /* Slice used up. Queue for next turn. */
1114 requeue_io(inode, wb);
1115 } else {
1116 /*
1117 * Writeback blocked by something other than
1118 * congestion. Delay the inode for some time to
1119 * avoid spinning on the CPU (100% iowait)
1120 * retrying writeback of the dirty page/inode
1121 * that cannot be performed immediately.
1122 */
1123 redirty_tail(inode, wb);
1124 }
1125 } else if (inode->i_state & I_DIRTY) {
1126 /*
1127 * Filesystems can dirty the inode during writeback operations,
1128 * such as delayed allocation during submission or metadata
1129 * updates after data IO completion.
1130 */
1131 redirty_tail(inode, wb);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001132 } else if (inode->i_state & I_DIRTY_TIME) {
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001133 inode->dirtied_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04001134 inode_wb_list_move_locked(inode, wb, &wb->b_dirty_time);
Jan Karaccb26b52012-05-03 14:47:58 +02001135 } else {
1136 /* The inode is clean. Remove from writeback lists. */
Tejun Heod6c10f12015-05-22 17:13:45 -04001137 inode_wb_list_del_locked(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001138 }
1139}
1140
1141/*
Jan Kara4f8ad652012-05-03 14:48:00 +02001142 * Write out an inode and its dirty pages. Do not update the writeback list
1143 * linkage. That is left to the caller. The caller is also responsible for
1144 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 */
1146static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001147__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -06001150 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +02001151 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 int ret;
1153
Jan Kara4f8ad652012-05-03 14:48:00 +02001154 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001156 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1157
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 ret = do_writepages(mapping, wbc);
1159
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001160 /*
1161 * Make sure to wait on the data before writing out the metadata.
1162 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +10001163 * I/O completion. We don't do it for sync(2) writeback because it has a
1164 * separate, external IO completion path and ->sync_fs for guaranteeing
1165 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001166 */
Dave Chinner7747bd42013-07-02 22:38:35 +10001167 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001168 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 if (ret == 0)
1170 ret = err;
1171 }
1172
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001173 /*
1174 * Some filesystems may redirty the inode during the writeback
1175 * due to delalloc, clear dirty metadata flags right before
1176 * write_inode()
1177 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001178 spin_lock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001179
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001180 dirty = inode->i_state & I_DIRTY;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001181 if (inode->i_state & I_DIRTY_TIME) {
1182 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1183 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1184 unlikely(time_after(jiffies,
1185 (inode->dirtied_time_when +
1186 dirtytime_expire_interval * HZ)))) {
1187 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1188 trace_writeback_lazytime(inode);
1189 }
1190 } else
1191 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001192 inode->i_state &= ~dirty;
Tejun Heo9c6ac782014-10-24 15:38:21 -04001193
1194 /*
1195 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1196 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1197 * either they see the I_DIRTY bits cleared or we see the dirtied
1198 * inode.
1199 *
1200 * I_DIRTY_PAGES is always cleared together above even if @mapping
1201 * still has dirty pages. The flag is reinstated after smp_mb() if
1202 * necessary. This guarantees that either __mark_inode_dirty()
1203 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1204 */
1205 smp_mb();
1206
1207 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1208 inode->i_state |= I_DIRTY_PAGES;
1209
Dave Chinner250df6e2011-03-22 22:23:36 +11001210 spin_unlock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001211
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001212 if (dirty & I_DIRTY_TIME)
1213 mark_inode_dirty_sync(inode);
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001214 /* Don't write the inode if only I_DIRTY_PAGES was set */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001215 if (dirty & ~I_DIRTY_PAGES) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001216 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 if (ret == 0)
1218 ret = err;
1219 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001220 trace_writeback_single_inode(inode, wbc, nr_to_write);
1221 return ret;
1222}
1223
1224/*
1225 * Write out an inode's dirty pages. Either the caller has an active reference
1226 * on the inode or the inode has I_WILL_FREE set.
1227 *
1228 * This function is designed to be called for writing back one inode which
1229 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1230 * and does more profound writeback list handling in writeback_sb_inodes().
1231 */
1232static int
1233writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1234 struct writeback_control *wbc)
1235{
1236 int ret = 0;
1237
1238 spin_lock(&inode->i_lock);
1239 if (!atomic_read(&inode->i_count))
1240 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1241 else
1242 WARN_ON(inode->i_state & I_WILL_FREE);
1243
1244 if (inode->i_state & I_SYNC) {
1245 if (wbc->sync_mode != WB_SYNC_ALL)
1246 goto out;
1247 /*
Jan Kara169ebd92012-05-03 14:48:03 +02001248 * It's a data-integrity sync. We must wait. Since callers hold
1249 * inode reference or inode has I_WILL_FREE set, it cannot go
1250 * away under us.
Jan Kara4f8ad652012-05-03 14:48:00 +02001251 */
Jan Kara169ebd92012-05-03 14:48:03 +02001252 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001253 }
1254 WARN_ON(inode->i_state & I_SYNC);
1255 /*
Jan Karaf9b0e052013-12-14 04:21:26 +08001256 * Skip inode if it is clean and we have no outstanding writeback in
1257 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1258 * function since flusher thread may be doing for example sync in
1259 * parallel and if we move the inode, it could get skipped. So here we
1260 * make sure inode is on some writeback list and leave it there unless
1261 * we have completely cleaned the inode.
Jan Kara4f8ad652012-05-03 14:48:00 +02001262 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001263 if (!(inode->i_state & I_DIRTY_ALL) &&
Jan Karaf9b0e052013-12-14 04:21:26 +08001264 (wbc->sync_mode != WB_SYNC_ALL ||
1265 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +02001266 goto out;
1267 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001268 wbc_attach_and_unlock_inode(wbc, inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001269
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001270 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
Tejun Heob16b1de2015-06-02 08:39:48 -06001272 wbc_detach_inode(wbc);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001273 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +11001274 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +02001275 /*
1276 * If inode is clean, remove it from writeback lists. Otherwise don't
1277 * touch it. See comment above for explanation.
1278 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001279 if (!(inode->i_state & I_DIRTY_ALL))
Tejun Heod6c10f12015-05-22 17:13:45 -04001280 inode_wb_list_del_locked(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001281 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001282 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001283out:
1284 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 return ret;
1286}
1287
Tejun Heoa88a3412015-05-22 17:13:28 -04001288static long writeback_chunk_size(struct bdi_writeback *wb,
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001289 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001290{
1291 long pages;
1292
1293 /*
1294 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1295 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1296 * here avoids calling into writeback_inodes_wb() more than once.
1297 *
1298 * The intended call sequence for WB_SYNC_ALL writeback is:
1299 *
1300 * wb_writeback()
1301 * writeback_sb_inodes() <== called only once
1302 * write_cache_pages() <== called once for each inode
1303 * (quickly) tag currently dirty pages
1304 * (maybe slowly) sync all tagged pages
1305 */
1306 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1307 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001308 else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001309 pages = min(wb->avg_write_bandwidth / 2,
Tejun Heodcc25ae2015-05-22 18:23:22 -04001310 global_wb_domain.dirty_limit / DIRTY_SCOPE);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001311 pages = min(pages, work->nr_pages);
1312 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1313 MIN_WRITEBACK_PAGES);
1314 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001315
1316 return pages;
1317}
1318
Jens Axboe03ba3782009-09-09 09:08:54 +02001319/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001320 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001321 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001322 * Return the number of pages and/or inodes written.
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001323 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001324static long writeback_sb_inodes(struct super_block *sb,
1325 struct bdi_writeback *wb,
1326 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001327{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001328 struct writeback_control wbc = {
1329 .sync_mode = work->sync_mode,
1330 .tagged_writepages = work->tagged_writepages,
1331 .for_kupdate = work->for_kupdate,
1332 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +10001333 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001334 .range_cyclic = work->range_cyclic,
1335 .range_start = 0,
1336 .range_end = LLONG_MAX,
1337 };
1338 unsigned long start_time = jiffies;
1339 long write_chunk;
1340 long wrote = 0; /* count both pages and inodes */
1341
Jens Axboe03ba3782009-09-09 09:08:54 +02001342 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001343 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001344
1345 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001346 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001347 /*
1348 * We only want to write back data for this
1349 * superblock, move all inodes not belonging
1350 * to it back onto the dirty list.
1351 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001352 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001353 continue;
1354 }
1355
1356 /*
1357 * The inode belongs to a different superblock.
1358 * Bounce back to the caller to unpin this and
1359 * pin the next superblock.
1360 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001361 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001362 }
1363
Christoph Hellwig9843b762010-10-24 19:40:46 +02001364 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +08001365 * Don't bother with new inodes or inodes being freed, first
1366 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b762010-10-24 19:40:46 +02001367 * kind writeout is handled by the freer.
1368 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001369 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +02001370 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001371 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -07001372 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -07001373 continue;
1374 }
Jan Karacc1676d2012-05-03 14:47:56 +02001375 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1376 /*
1377 * If this inode is locked for writeback and we are not
1378 * doing writeback-for-data-integrity, move it to
1379 * b_more_io so that writeback can proceed with the
1380 * other inodes on s_io.
1381 *
1382 * We'll have another go at writing back this inode
1383 * when we completed a full scan of b_io.
1384 */
1385 spin_unlock(&inode->i_lock);
1386 requeue_io(inode, wb);
1387 trace_writeback_sb_inodes_requeue(inode);
1388 continue;
1389 }
Jan Karaf0d07b72012-05-03 14:47:59 +02001390 spin_unlock(&wb->list_lock);
1391
Jan Kara4f8ad652012-05-03 14:48:00 +02001392 /*
1393 * We already requeued the inode if it had I_SYNC set and we
1394 * are doing WB_SYNC_NONE writeback. So this catches only the
1395 * WB_SYNC_ALL case.
1396 */
Jan Kara169ebd92012-05-03 14:48:03 +02001397 if (inode->i_state & I_SYNC) {
1398 /* Wait for I_SYNC. This function drops i_lock... */
1399 inode_sleep_on_writeback(inode);
1400 /* Inode may be gone, start again */
Jan Karaead188f2012-06-08 17:07:36 +02001401 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001402 continue;
1403 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001404 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001405 wbc_attach_and_unlock_inode(&wbc, inode);
Jan Kara169ebd92012-05-03 14:48:03 +02001406
Tejun Heoa88a3412015-05-22 17:13:28 -04001407 write_chunk = writeback_chunk_size(wb, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001408 wbc.nr_to_write = write_chunk;
1409 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +11001410
Jan Kara169ebd92012-05-03 14:48:03 +02001411 /*
1412 * We use I_SYNC to pin the inode in memory. While it is set
1413 * evict_inode() will wait so the inode cannot be freed.
1414 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001415 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001416
Tejun Heob16b1de2015-06-02 08:39:48 -06001417 wbc_detach_inode(&wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001418 work->nr_pages -= write_chunk - wbc.nr_to_write;
1419 wrote += write_chunk - wbc.nr_to_write;
Jan Kara4f8ad652012-05-03 14:48:00 +02001420 spin_lock(&wb->list_lock);
1421 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001422 if (!(inode->i_state & I_DIRTY_ALL))
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001423 wrote++;
Jan Kara4f8ad652012-05-03 14:48:00 +02001424 requeue_inode(inode, wb, &wbc);
1425 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001426 spin_unlock(&inode->i_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001427 cond_resched_lock(&wb->list_lock);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001428 /*
1429 * bail out to wb_writeback() often enough to check
1430 * background threshold and other termination conditions.
1431 */
1432 if (wrote) {
1433 if (time_is_before_jiffies(start_time + HZ / 10UL))
1434 break;
1435 if (work->nr_pages <= 0)
1436 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -08001437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001439 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001440}
Nick Piggin38f21972009-01-06 14:40:25 -08001441
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001442static long __writeback_inodes_wb(struct bdi_writeback *wb,
1443 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001444{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001445 unsigned long start_time = jiffies;
1446 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001447
1448 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001449 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001450 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001451
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001452 if (!trylock_super(sb)) {
Wu Fengguang0e995812011-07-29 22:14:35 -06001453 /*
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001454 * trylock_super() may fail consistently due to
Wu Fengguang0e995812011-07-29 22:14:35 -06001455 * s_umount being grabbed by someone else. Don't use
1456 * requeue_io() to avoid busy retrying the inode/sb.
1457 */
1458 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001459 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001460 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001461 wrote += writeback_sb_inodes(sb, wb, work);
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001462 up_read(&sb->s_umount);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001463
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001464 /* refer to the same tests at the end of writeback_sb_inodes */
1465 if (wrote) {
1466 if (time_is_before_jiffies(start_time + HZ / 10UL))
1467 break;
1468 if (work->nr_pages <= 0)
1469 break;
1470 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001471 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001472 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001473 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001474}
1475
Wanpeng Li7d9f0732013-09-11 14:22:40 -07001476static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001477 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001478{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001479 struct wb_writeback_work work = {
1480 .nr_pages = nr_pages,
1481 .sync_mode = WB_SYNC_NONE,
1482 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001483 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001484 };
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001485
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001486 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -06001487 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001488 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001489 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001490 spin_unlock(&wb->list_lock);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001491
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001492 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001493}
1494
Jens Axboe03ba3782009-09-09 09:08:54 +02001495/*
1496 * Explicit flushing or periodic writeback of "old" data.
1497 *
1498 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1499 * dirtying-time in the inode's address_space. So this periodic writeback code
1500 * just walks the superblock inode list, writing back any inodes which are
1501 * older than a specific point in time.
1502 *
1503 * Try to run once per dirty_writeback_interval. But if a writeback event
1504 * takes longer than a dirty_writeback_interval interval, then leave a
1505 * one-second gap.
1506 *
1507 * older_than_this takes precedence over nr_to_write. So we'll only write back
1508 * all dirty pages if they are all attached to "old" mappings.
1509 */
Jens Axboec4a77a62009-09-16 15:18:25 +02001510static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001511 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001512{
Wu Fengguange98be2d2010-08-29 11:22:30 -06001513 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001514 long nr_pages = work->nr_pages;
Jan Kara0dc83bd2014-02-21 11:19:04 +01001515 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +02001516 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001517 long progress;
Jens Axboe03ba3782009-09-09 09:08:54 +02001518
Jan Kara0dc83bd2014-02-21 11:19:04 +01001519 oldest_jif = jiffies;
1520 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +02001521
Wu Fengguange8dfc302011-04-21 12:06:32 -06001522 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001523 for (;;) {
1524 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001525 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +02001526 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001527 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +02001528 break;
1529
1530 /*
Jan Karaaa373cf2011-01-13 15:45:47 -08001531 * Background writeout and kupdate-style writeback may
1532 * run forever. Stop them if there is other work to do
1533 * so that e.g. sync can proceed. They'll be restarted
1534 * after the other works are all done.
1535 */
1536 if ((work->for_background || work->for_kupdate) &&
Tejun Heof0054bb2015-05-22 17:13:30 -04001537 !list_empty(&wb->work_list))
Jan Karaaa373cf2011-01-13 15:45:47 -08001538 break;
1539
1540 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001541 * For background writeout, stop when we are below the
1542 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +02001543 */
Tejun Heoaa661bb2015-05-22 18:23:31 -04001544 if (work->for_background && !wb_over_bg_thresh(wb))
Jens Axboe03ba3782009-09-09 09:08:54 +02001545 break;
1546
Jan Kara1bc36b62011-10-19 11:44:41 +02001547 /*
1548 * Kupdate and background works are special and we want to
1549 * include all inodes that need writing. Livelock avoidance is
1550 * handled by these works yielding to any other work so we are
1551 * safe.
1552 */
Wu Fengguangba9aa832010-07-21 20:32:30 -06001553 if (work->for_kupdate) {
Jan Kara0dc83bd2014-02-21 11:19:04 +01001554 oldest_jif = jiffies -
Wu Fengguangba9aa832010-07-21 20:32:30 -06001555 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +02001556 } else if (work->for_background)
Jan Kara0dc83bd2014-02-21 11:19:04 +01001557 oldest_jif = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +10001558
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001559 trace_writeback_start(wb->bdi, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001560 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001561 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001562 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001563 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001564 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001565 progress = __writeback_inodes_wb(wb, work);
1566 trace_writeback_written(wb->bdi, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +10001567
Wu Fengguange98be2d2010-08-29 11:22:30 -06001568 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +02001569
1570 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +02001571 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001572 *
1573 * Dirty inodes are moved to b_io for writeback in batches.
1574 * The completion of the current batch does not necessarily
1575 * mean the overall work is done. So we keep looping as long
1576 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +02001577 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001578 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +02001579 continue;
1580 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001581 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +02001582 */
Wu Fengguangb7a24412010-07-21 22:19:51 -06001583 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +02001584 break;
1585 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +02001586 * Nothing written. Wait for some inode to
1587 * become available for writeback. Otherwise
1588 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +02001589 */
Jens Axboe03ba3782009-09-09 09:08:54 +02001590 if (!list_empty(&wb->b_more_io)) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001591 trace_writeback_wait(wb->bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001592 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +11001593 spin_lock(&inode->i_lock);
Jan Karaf0d07b72012-05-03 14:47:59 +02001594 spin_unlock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001595 /* This function drops i_lock... */
1596 inode_sleep_on_writeback(inode);
Jan Karaf0d07b72012-05-03 14:47:59 +02001597 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001598 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001599 }
Wu Fengguange8dfc302011-04-21 12:06:32 -06001600 spin_unlock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001601
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001602 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +02001603}
1604
1605/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001606 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +02001607 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001608static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001609{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001610 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001611
Tejun Heof0054bb2015-05-22 17:13:30 -04001612 spin_lock_bh(&wb->work_lock);
1613 if (!list_empty(&wb->work_list)) {
1614 work = list_entry(wb->work_list.next,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001615 struct wb_writeback_work, list);
1616 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +02001617 }
Tejun Heof0054bb2015-05-22 17:13:30 -04001618 spin_unlock_bh(&wb->work_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001619 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +02001620}
1621
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001622/*
1623 * Add in the number of potentially dirty inodes, because each inode
1624 * write can dirty pagecache in the underlying blockdev.
1625 */
1626static unsigned long get_nr_dirty_pages(void)
1627{
1628 return global_page_state(NR_FILE_DIRTY) +
1629 global_page_state(NR_UNSTABLE_NFS) +
1630 get_nr_dirty_inodes();
1631}
1632
Jan Kara65850272011-01-13 15:45:44 -08001633static long wb_check_background_flush(struct bdi_writeback *wb)
1634{
Tejun Heoaa661bb2015-05-22 18:23:31 -04001635 if (wb_over_bg_thresh(wb)) {
Jan Kara65850272011-01-13 15:45:44 -08001636
1637 struct wb_writeback_work work = {
1638 .nr_pages = LONG_MAX,
1639 .sync_mode = WB_SYNC_NONE,
1640 .for_background = 1,
1641 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001642 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -08001643 };
1644
1645 return wb_writeback(wb, &work);
1646 }
1647
1648 return 0;
1649}
1650
Jens Axboe03ba3782009-09-09 09:08:54 +02001651static long wb_check_old_data_flush(struct bdi_writeback *wb)
1652{
1653 unsigned long expired;
1654 long nr_pages;
1655
Jens Axboe69b62d02010-05-17 12:51:03 +02001656 /*
1657 * When set to zero, disable periodic writeback
1658 */
1659 if (!dirty_writeback_interval)
1660 return 0;
1661
Jens Axboe03ba3782009-09-09 09:08:54 +02001662 expired = wb->last_old_flush +
1663 msecs_to_jiffies(dirty_writeback_interval * 10);
1664 if (time_before(jiffies, expired))
1665 return 0;
1666
1667 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001668 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +02001669
Jens Axboec4a77a62009-09-16 15:18:25 +02001670 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001671 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +02001672 .nr_pages = nr_pages,
1673 .sync_mode = WB_SYNC_NONE,
1674 .for_kupdate = 1,
1675 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001676 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +02001677 };
1678
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001679 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +02001680 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001681
1682 return 0;
1683}
1684
1685/*
1686 * Retrieve work items and do the writeback they describe
1687 */
Wanpeng Li25d130b2013-07-08 16:00:14 -07001688static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001689{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001690 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +02001691 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +02001692
Tejun Heo44522262015-05-22 17:13:26 -04001693 set_bit(WB_writeback_running, &wb->state);
Tejun Heof0054bb2015-05-22 17:13:30 -04001694 while ((work = get_next_work_item(wb)) != NULL) {
Tejun Heocc395d72015-05-22 17:13:58 -04001695 struct wb_completion *done = work->done;
Tejun Heo98754bf2015-05-22 17:13:59 -04001696 bool need_wake_up = false;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001697
Tejun Heof0054bb2015-05-22 17:13:30 -04001698 trace_writeback_exec(wb->bdi, work);
Dave Chinner455b2862010-07-07 13:24:06 +10001699
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001700 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001701
Tejun Heo98754bf2015-05-22 17:13:59 -04001702 if (work->single_wait) {
1703 WARN_ON_ONCE(work->auto_free);
1704 /* paired w/ rmb in wb_wait_for_single_work() */
1705 smp_wmb();
1706 work->single_done = 1;
1707 need_wake_up = true;
1708 } else if (work->auto_free) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001709 kfree(work);
Tejun Heo98754bf2015-05-22 17:13:59 -04001710 }
1711
Tejun Heocc395d72015-05-22 17:13:58 -04001712 if (done && atomic_dec_and_test(&done->cnt))
Tejun Heo98754bf2015-05-22 17:13:59 -04001713 need_wake_up = true;
1714
1715 if (need_wake_up)
Tejun Heocc395d72015-05-22 17:13:58 -04001716 wake_up_all(&wb->bdi->wb_waitq);
Jens Axboe03ba3782009-09-09 09:08:54 +02001717 }
1718
1719 /*
1720 * Check for periodic writeback, kupdated() style
1721 */
1722 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08001723 wrote += wb_check_background_flush(wb);
Tejun Heo44522262015-05-22 17:13:26 -04001724 clear_bit(WB_writeback_running, &wb->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02001725
1726 return wrote;
1727}
1728
1729/*
1730 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07001731 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02001732 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001733void wb_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001734{
Tejun Heo839a8e82013-04-01 19:08:06 -07001735 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1736 struct bdi_writeback, dwork);
Jens Axboe03ba3782009-09-09 09:08:54 +02001737 long pages_written;
1738
Tejun Heof0054bb2015-05-22 17:13:30 -04001739 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
Peter Zijlstra766f9162010-10-26 14:22:45 -07001740 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02001741
Tejun Heo839a8e82013-04-01 19:08:06 -07001742 if (likely(!current_is_workqueue_rescuer() ||
Tejun Heo44522262015-05-22 17:13:26 -04001743 !test_bit(WB_registered, &wb->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03001744 /*
Tejun Heof0054bb2015-05-22 17:13:30 -04001745 * The normal path. Keep writing back @wb until its
Tejun Heo839a8e82013-04-01 19:08:06 -07001746 * work_list is empty. Note that this path is also taken
Tejun Heof0054bb2015-05-22 17:13:30 -04001747 * if @wb is shutting down even when we're running off the
Tejun Heo839a8e82013-04-01 19:08:06 -07001748 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03001749 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001750 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07001751 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07001752 trace_writeback_pages_written(pages_written);
Tejun Heof0054bb2015-05-22 17:13:30 -04001753 } while (!list_empty(&wb->work_list));
Tejun Heo839a8e82013-04-01 19:08:06 -07001754 } else {
1755 /*
1756 * bdi_wq can't get enough workers and we're running off
1757 * the emergency worker. Don't hog it. Hopefully, 1024 is
1758 * enough for efficient IO.
1759 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001760 pages_written = writeback_inodes_wb(wb, 1024,
Tejun Heo839a8e82013-04-01 19:08:06 -07001761 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10001762 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02001763 }
1764
Tejun Heof0054bb2015-05-22 17:13:30 -04001765 if (!list_empty(&wb->work_list))
Derek Basehore6ca738d2014-04-03 14:46:22 -07001766 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1767 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Tejun Heof0054bb2015-05-22 17:13:30 -04001768 wb_wakeup_delayed(wb);
Dave Chinner455b2862010-07-07 13:24:06 +10001769
Tejun Heo839a8e82013-04-01 19:08:06 -07001770 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02001771}
1772
1773/*
Jens Axboe03ba3782009-09-09 09:08:54 +02001774 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1775 * the whole world.
1776 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001777void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02001778{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001779 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001780
Jan Kara47df3dd2013-09-11 14:22:22 -07001781 if (!nr_pages)
1782 nr_pages = get_nr_dirty_pages();
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001783
1784 rcu_read_lock();
Tejun Heof2b65122015-05-22 17:13:55 -04001785 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1786 struct bdi_writeback *wb;
1787 struct wb_iter iter;
1788
1789 if (!bdi_has_dirty_io(bdi))
1790 continue;
1791
1792 bdi_for_each_wb(wb, bdi, &iter, 0)
1793 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1794 false, reason);
1795 }
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001796 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001797}
1798
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001799/*
1800 * Wake up bdi's periodically to make sure dirtytime inodes gets
1801 * written back periodically. We deliberately do *not* check the
1802 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1803 * kernel to be constantly waking up once there are any dirtytime
1804 * inodes on the system. So instead we define a separate delayed work
1805 * function which gets called much more rarely. (By default, only
1806 * once every 12 hours.)
1807 *
1808 * If there is any other write activity going on in the file system,
1809 * this function won't be necessary. But if the only thing that has
1810 * happened on the file system is a dirtytime inode caused by an atime
1811 * update, we need this infrastructure below to make sure that inode
1812 * eventually gets pushed out to disk.
1813 */
1814static void wakeup_dirtytime_writeback(struct work_struct *w);
1815static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1816
1817static void wakeup_dirtytime_writeback(struct work_struct *w)
1818{
1819 struct backing_dev_info *bdi;
1820
1821 rcu_read_lock();
1822 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Tejun Heo001fe6f2015-05-22 17:13:56 -04001823 struct bdi_writeback *wb;
1824 struct wb_iter iter;
1825
1826 bdi_for_each_wb(wb, bdi, &iter, 0)
1827 if (!list_empty(&bdi->wb.b_dirty_time))
1828 wb_wakeup(&bdi->wb);
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001829 }
1830 rcu_read_unlock();
1831 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1832}
1833
1834static int __init start_dirtytime_writeback(void)
1835{
1836 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1837 return 0;
1838}
1839__initcall(start_dirtytime_writeback);
1840
Theodore Ts'o1efff912015-03-17 12:23:32 -04001841int dirtytime_interval_handler(struct ctl_table *table, int write,
1842 void __user *buffer, size_t *lenp, loff_t *ppos)
1843{
1844 int ret;
1845
1846 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1847 if (ret == 0 && write)
1848 mod_delayed_work(system_wq, &dirtytime_work, 0);
1849 return ret;
1850}
1851
Jens Axboe03ba3782009-09-09 09:08:54 +02001852static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1853{
1854 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1855 struct dentry *dentry;
1856 const char *name = "?";
1857
1858 dentry = d_find_alias(inode);
1859 if (dentry) {
1860 spin_lock(&dentry->d_lock);
1861 name = (const char *) dentry->d_name.name;
1862 }
1863 printk(KERN_DEBUG
1864 "%s(%d): dirtied inode %lu (%s) on %s\n",
1865 current->comm, task_pid_nr(current), inode->i_ino,
1866 name, inode->i_sb->s_id);
1867 if (dentry) {
1868 spin_unlock(&dentry->d_lock);
1869 dput(dentry);
1870 }
1871 }
1872}
1873
1874/**
1875 * __mark_inode_dirty - internal function
1876 * @inode: inode to mark
1877 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1878 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1879 * mark_inode_dirty_sync.
1880 *
1881 * Put the inode on the super block's dirty list.
1882 *
1883 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1884 * dirty list only if it is hashed or if it refers to a blockdev.
1885 * If it was not hashed, it will never be added to the dirty list
1886 * even if it is later hashed, as it will have been marked dirty already.
1887 *
1888 * In short, make sure you hash any inodes _before_ you start marking
1889 * them dirty.
1890 *
Jens Axboe03ba3782009-09-09 09:08:54 +02001891 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1892 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1893 * the kernel-internal blockdev inode represents the dirtying time of the
1894 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1895 * page->mapping->host, so the page-dirtying time is recorded in the internal
1896 * blockdev inode.
1897 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001898#define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
Jens Axboe03ba3782009-09-09 09:08:54 +02001899void __mark_inode_dirty(struct inode *inode, int flags)
1900{
1901 struct super_block *sb = inode->i_sb;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001902 int dirtytime;
1903
1904 trace_writeback_mark_inode_dirty(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02001905
1906 /*
1907 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1908 * dirty the inode itself
1909 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001910 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001911 trace_writeback_dirty_inode_start(inode, flags);
1912
Jens Axboe03ba3782009-09-09 09:08:54 +02001913 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04001914 sb->s_op->dirty_inode(inode, flags);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001915
1916 trace_writeback_dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02001917 }
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001918 if (flags & I_DIRTY_INODE)
1919 flags &= ~I_DIRTY_TIME;
1920 dirtytime = flags & I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02001921
1922 /*
Tejun Heo9c6ac782014-10-24 15:38:21 -04001923 * Paired with smp_mb() in __writeback_single_inode() for the
1924 * following lockless i_state test. See there for details.
Jens Axboe03ba3782009-09-09 09:08:54 +02001925 */
1926 smp_mb();
1927
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001928 if (((inode->i_state & flags) == flags) ||
1929 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
Jens Axboe03ba3782009-09-09 09:08:54 +02001930 return;
1931
1932 if (unlikely(block_dump))
1933 block_dump___mark_inode_dirty(inode);
1934
Dave Chinner250df6e2011-03-22 22:23:36 +11001935 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001936 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
1937 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001938 if ((inode->i_state & flags) != flags) {
1939 const int was_dirty = inode->i_state & I_DIRTY;
1940
Tejun Heo52ebea72015-05-22 17:13:37 -04001941 inode_attach_wb(inode, NULL);
1942
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001943 if (flags & I_DIRTY_INODE)
1944 inode->i_state &= ~I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02001945 inode->i_state |= flags;
1946
1947 /*
1948 * If the inode is being synced, just update its dirty state.
1949 * The unlocker will place the inode on the appropriate
1950 * superblock list, based upon its state.
1951 */
1952 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11001953 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001954
1955 /*
1956 * Only add valid (hashed) inodes to the superblock's
1957 * dirty list. Add blockdev inodes as well.
1958 */
1959 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -04001960 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11001961 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001962 }
Al Viroa4ffdde2010-06-02 17:38:30 -04001963 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11001964 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001965
1966 /*
1967 * If the inode was already on b_dirty/b_io/b_more_io, don't
1968 * reposition it (that would break b_dirty time-ordering).
1969 */
1970 if (!was_dirty) {
Tejun Heo87e1d782015-05-28 14:50:52 -04001971 struct bdi_writeback *wb;
Tejun Heod6c10f12015-05-22 17:13:45 -04001972 struct list_head *dirty_list;
Dave Chinnera66979a2011-03-22 22:23:41 +11001973 bool wakeup_bdi = false;
Jens Axboe500b0672009-09-09 09:10:25 +02001974
Tejun Heo87e1d782015-05-28 14:50:52 -04001975 wb = locked_inode_to_wb_and_lock_list(inode);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001976
Tejun Heo07472592015-05-22 17:14:02 -04001977 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
1978 !test_bit(WB_registered, &wb->state),
1979 "bdi-%s not registered\n", wb->bdi->name);
Jens Axboe03ba3782009-09-09 09:08:54 +02001980
1981 inode->dirtied_when = jiffies;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001982 if (dirtytime)
1983 inode->dirtied_time_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04001984
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001985 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
Tejun Heo07472592015-05-22 17:14:02 -04001986 dirty_list = &wb->b_dirty;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001987 else
Tejun Heo07472592015-05-22 17:14:02 -04001988 dirty_list = &wb->b_dirty_time;
Tejun Heod6c10f12015-05-22 17:13:45 -04001989
Tejun Heo07472592015-05-22 17:14:02 -04001990 wakeup_bdi = inode_wb_list_move_locked(inode, wb,
Tejun Heod6c10f12015-05-22 17:13:45 -04001991 dirty_list);
1992
Tejun Heo07472592015-05-22 17:14:02 -04001993 spin_unlock(&wb->list_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001994 trace_writeback_dirty_inode_enqueue(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +11001995
Tejun Heod6c10f12015-05-22 17:13:45 -04001996 /*
1997 * If this is the first dirty inode for this bdi,
1998 * we have to wake-up the corresponding bdi thread
1999 * to make sure background write-back happens
2000 * later.
2001 */
Tejun Heo07472592015-05-22 17:14:02 -04002002 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2003 wb_wakeup_delayed(wb);
Dave Chinnera66979a2011-03-22 22:23:41 +11002004 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02002005 }
2006 }
Dave Chinner250df6e2011-03-22 22:23:36 +11002007out_unlock_inode:
2008 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002009
Jens Axboe03ba3782009-09-09 09:08:54 +02002010}
2011EXPORT_SYMBOL(__mark_inode_dirty);
2012
Jens Axboeb6e51312009-09-16 15:13:54 +02002013static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002014{
Jens Axboe03ba3782009-09-09 09:08:54 +02002015 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002016
Jens Axboe03ba3782009-09-09 09:08:54 +02002017 /*
2018 * We need to be protected against the filesystem going from
2019 * r/o to r/w or vice versa.
2020 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002021 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002022
Dave Chinner55fa6092011-03-22 22:23:40 +11002023 spin_lock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002024
2025 /*
2026 * Data integrity sync. Must wait for all pages under writeback,
2027 * because there may have been pages dirtied before our sync
2028 * call, but which had writeout started before we write it out.
2029 * In which case, the inode may not be on the dirty list, but
2030 * we still have to wait for that writeout.
2031 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002032 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11002033 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02002034
Dave Chinner250df6e2011-03-22 22:23:36 +11002035 spin_lock(&inode->i_lock);
2036 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2037 (mapping->nrpages == 0)) {
2038 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002039 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11002040 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002041 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11002042 spin_unlock(&inode->i_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11002043 spin_unlock(&inode_sb_list_lock);
2044
Jens Axboe03ba3782009-09-09 09:08:54 +02002045 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11002046 * We hold a reference to 'inode' so it couldn't have been
2047 * removed from s_inodes list while we dropped the
2048 * inode_sb_list_lock. We cannot iput the inode now as we can
2049 * be holding the last reference and we cannot iput it under
2050 * inode_sb_list_lock. So we keep the reference and iput it
2051 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02002052 */
2053 iput(old_inode);
2054 old_inode = inode;
2055
2056 filemap_fdatawait(mapping);
2057
2058 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08002059
Dave Chinner55fa6092011-03-22 22:23:40 +11002060 spin_lock(&inode_sb_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002061 }
Dave Chinner55fa6092011-03-22 22:23:40 +11002062 spin_unlock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002063 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064}
2065
Tejun Heof30a7d02015-05-22 17:14:00 -04002066static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2067 enum wb_reason reason, bool skip_if_busy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068{
Tejun Heocc395d72015-05-22 17:13:58 -04002069 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002070 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002071 .sb = sb,
2072 .sync_mode = WB_SYNC_NONE,
2073 .tagged_writepages = 1,
2074 .done = &done,
2075 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002076 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002077 };
Tejun Heoe7972912015-05-22 17:13:48 -04002078 struct backing_dev_info *bdi = sb->s_bdi;
Jens Axboe0e3c9a22010-06-01 11:08:43 +02002079
Tejun Heoe7972912015-05-22 17:13:48 -04002080 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002081 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002082 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Tejun Heof30a7d02015-05-22 17:14:00 -04002083
Tejun Heodb125362015-05-22 17:14:01 -04002084 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
Tejun Heocc395d72015-05-22 17:13:58 -04002085 wb_wait_for_completion(bdi, &done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086}
Tejun Heof30a7d02015-05-22 17:14:00 -04002087
2088/**
2089 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2090 * @sb: the superblock
2091 * @nr: the number of pages to write
2092 * @reason: reason why some writeback work initiated
2093 *
2094 * Start writeback on some inodes on this super_block. No guarantees are made
2095 * on how many (if any) will be written, and this function does not wait
2096 * for IO completion of submitted IO.
2097 */
2098void writeback_inodes_sb_nr(struct super_block *sb,
2099 unsigned long nr,
2100 enum wb_reason reason)
2101{
2102 __writeback_inodes_sb_nr(sb, nr, reason, false);
2103}
Chris Mason3259f8b2010-10-29 11:16:17 -04002104EXPORT_SYMBOL(writeback_inodes_sb_nr);
2105
2106/**
2107 * writeback_inodes_sb - writeback dirty inodes from given super_block
2108 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08002109 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04002110 *
2111 * Start writeback on some inodes on this super_block. No guarantees are made
2112 * on how many (if any) will be written, and this function does not wait
2113 * for IO completion of submitted IO.
2114 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002115void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002116{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002117 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04002118}
Jens Axboed8a85592009-09-02 12:34:32 +02002119EXPORT_SYMBOL(writeback_inodes_sb);
2120
2121/**
Miao Xie10ee27a2013-01-10 13:47:57 +08002122 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
Chris Mason3259f8b2010-10-29 11:16:17 -04002123 * @sb: the superblock
2124 * @nr: the number of pages to write
Miao Xie10ee27a2013-01-10 13:47:57 +08002125 * @reason: the reason of writeback
Chris Mason3259f8b2010-10-29 11:16:17 -04002126 *
Miao Xie10ee27a2013-01-10 13:47:57 +08002127 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
Chris Mason3259f8b2010-10-29 11:16:17 -04002128 * Returns 1 if writeback was started, 0 if not.
2129 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002130bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2131 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002132{
Miao Xie10ee27a2013-01-10 13:47:57 +08002133 if (!down_read_trylock(&sb->s_umount))
Tejun Heof30a7d02015-05-22 17:14:00 -04002134 return false;
Miao Xie10ee27a2013-01-10 13:47:57 +08002135
Tejun Heof30a7d02015-05-22 17:14:00 -04002136 __writeback_inodes_sb_nr(sb, nr, reason, true);
Miao Xie10ee27a2013-01-10 13:47:57 +08002137 up_read(&sb->s_umount);
Tejun Heof30a7d02015-05-22 17:14:00 -04002138 return true;
Chris Mason3259f8b2010-10-29 11:16:17 -04002139}
Miao Xie10ee27a2013-01-10 13:47:57 +08002140EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2141
2142/**
2143 * try_to_writeback_inodes_sb - try to start writeback if none underway
2144 * @sb: the superblock
2145 * @reason: reason why some writeback work was initiated
2146 *
2147 * Implement by try_to_writeback_inodes_sb_nr()
2148 * Returns 1 if writeback was started, 0 if not.
2149 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002150bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Miao Xie10ee27a2013-01-10 13:47:57 +08002151{
2152 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2153}
2154EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04002155
2156/**
Jens Axboed8a85592009-09-02 12:34:32 +02002157 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01002158 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02002159 *
2160 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01002161 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02002162 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01002163void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02002164{
Tejun Heocc395d72015-05-22 17:13:58 -04002165 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002166 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002167 .sb = sb,
2168 .sync_mode = WB_SYNC_ALL,
2169 .nr_pages = LONG_MAX,
2170 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002171 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002172 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10002173 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002174 };
Tejun Heoe7972912015-05-22 17:13:48 -04002175 struct backing_dev_info *bdi = sb->s_bdi;
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002176
Jan Kara6eedc702012-07-03 16:45:27 +02002177 /* Nothing to do? */
Tejun Heoe7972912015-05-22 17:13:48 -04002178 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002179 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002180 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2181
Tejun Heodb125362015-05-22 17:14:01 -04002182 bdi_split_work_to_wbs(bdi, &work, false);
Tejun Heocc395d72015-05-22 17:13:58 -04002183 wb_wait_for_completion(bdi, &done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002184
Jens Axboeb6e51312009-09-16 15:13:54 +02002185 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02002186}
2187EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002190 * write_inode_now - write an inode to disk
2191 * @inode: inode to write to disk
2192 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002194 * This function commits an inode to disk immediately if it is dirty. This is
2195 * primarily needed by knfsd.
2196 *
2197 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199int write_inode_now(struct inode *inode, int sync)
2200{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06002201 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 struct writeback_control wbc = {
2203 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08002204 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07002205 .range_start = 0,
2206 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 };
2208
2209 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08002210 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211
2212 might_sleep();
Jan Kara4f8ad652012-05-03 14:48:00 +02002213 return writeback_single_inode(inode, wb, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
2215EXPORT_SYMBOL(write_inode_now);
2216
2217/**
2218 * sync_inode - write an inode and its pages to disk.
2219 * @inode: the inode to sync
2220 * @wbc: controls the writeback mode
2221 *
2222 * sync_inode() will write an inode and its pages to disk. It will also
2223 * correctly update the inode on its superblock's dirty inode lists and will
2224 * update inode->i_state.
2225 *
2226 * The caller must have a ref on the inode.
2227 */
2228int sync_inode(struct inode *inode, struct writeback_control *wbc)
2229{
Jan Kara4f8ad652012-05-03 14:48:00 +02002230 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231}
2232EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02002233
2234/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002235 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02002236 * @inode: the inode to sync
2237 * @wait: wait for I/O to complete.
2238 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002239 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02002240 *
2241 * Note: only writes the actual inode, no associated data or other metadata.
2242 */
2243int sync_inode_metadata(struct inode *inode, int wait)
2244{
2245 struct writeback_control wbc = {
2246 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2247 .nr_to_write = 0, /* metadata-only */
2248 };
2249
2250 return sync_inode(inode, &wbc);
2251}
2252EXPORT_SYMBOL(sync_inode_metadata);