blob: 4f680f848c8b5cab0d6b8f2bfcab88c23af89ddf [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * fs/fs-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
11 *
Francois Camie1f8e872008-10-15 22:01:59 -070012 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070013 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
15 */
16
17#include <linux/kernel.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050018#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/sched.h>
22#include <linux/fs.h>
23#include <linux/mm.h>
Wu Fengguangbc31b862012-01-07 20:41:55 -060024#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020025#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/writeback.h>
27#include <linux/blkdev.h>
28#include <linux/backing-dev.h>
Dave Chinner455b2862010-07-07 13:24:06 +100029#include <linux/tracepoint.h>
Al Viro719ea2f2013-09-29 11:24:49 -040030#include <linux/device.h>
Tejun Heo21c6321f2015-05-28 14:50:49 -040031#include <linux/memcontrol.h>
David Howells07f3f052006-09-30 20:52:18 +020032#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
Jens Axboed0bceac2009-05-18 08:20:32 +020034/*
Wu Fengguangbc31b862012-01-07 20:41:55 -060035 * 4MB minimal write chunk size
36 */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +030037#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
Wu Fengguangbc31b862012-01-07 20:41:55 -060038
39/*
Jens Axboec4a77a62009-09-16 15:18:25 +020040 * Passed into wb_writeback(), essentially a subset of writeback_control
41 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020042struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020043 long nr_pages;
44 struct super_block *sb;
45 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060046 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050047 unsigned int for_kupdate:1;
48 unsigned int range_cyclic:1;
49 unsigned int for_background:1;
Dave Chinner7747bd42013-07-02 22:38:35 +100050 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
Tejun Heoac7b19a2015-05-22 17:13:57 -040051 unsigned int auto_free:1; /* free on completion */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060052 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020053
Jens Axboe8010c3b2009-09-15 20:04:57 +020054 struct list_head list; /* pending work list */
Tejun Heocc395d72015-05-22 17:13:58 -040055 struct wb_completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020056};
57
Theodore Ts'oa2f48702015-03-17 12:23:19 -040058/*
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
67 */
68unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
Nick Piggin7ccf19a2010-10-21 11:49:30 +110070static inline struct inode *wb_inode(struct list_head *head)
71{
Dave Chinnerc7f54082015-03-04 14:07:22 -050072 return list_entry(head, struct inode, i_io_list);
Nick Piggin7ccf19a2010-10-21 11:49:30 +110073}
74
Wu Fengguang15eb77a2012-01-17 11:18:56 -060075/*
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
79 */
80#define CREATE_TRACE_POINTS
81#include <trace/events/writeback.h>
82
Steven Whitehouse774016b2014-02-06 15:47:47 +000083EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
Tejun Heod6c10f12015-05-22 17:13:45 -040085static bool wb_io_lists_populated(struct bdi_writeback *wb)
86{
87 if (wb_has_dirty_io(wb)) {
88 return false;
89 } else {
90 set_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -040091 WARN_ON_ONCE(!wb->avg_write_bandwidth);
Tejun Heo766a9d62015-05-22 17:13:46 -040092 atomic_long_add(wb->avg_write_bandwidth,
93 &wb->bdi->tot_write_bandwidth);
Tejun Heod6c10f12015-05-22 17:13:45 -040094 return true;
95 }
96}
97
98static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99{
100 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
Tejun Heo766a9d62015-05-22 17:13:46 -0400101 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
Tejun Heod6c10f12015-05-22 17:13:45 -0400102 clear_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400103 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 &wb->bdi->tot_write_bandwidth) < 0);
Tejun Heo766a9d62015-05-22 17:13:46 -0400105 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400106}
107
108/**
Dave Chinnerc7f54082015-03-04 14:07:22 -0500109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
Tejun Heod6c10f12015-05-22 17:13:45 -0400110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
Wang Longbbbc3c12017-12-05 07:23:19 -0500112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
Tejun Heod6c10f12015-05-22 17:13:45 -0400113 *
Dave Chinnerc7f54082015-03-04 14:07:22 -0500114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
Tejun Heod6c10f12015-05-22 17:13:45 -0400115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
117 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500118static bool inode_io_list_move_locked(struct inode *inode,
Tejun Heod6c10f12015-05-22 17:13:45 -0400119 struct bdi_writeback *wb,
120 struct list_head *head)
121{
122 assert_spin_locked(&wb->list_lock);
123
Dave Chinnerc7f54082015-03-04 14:07:22 -0500124 list_move(&inode->i_io_list, head);
Tejun Heod6c10f12015-05-22 17:13:45 -0400125
126 /* dirty_time doesn't count as dirty_io until expiration */
127 if (head != &wb->b_dirty_time)
128 return wb_io_lists_populated(wb);
129
130 wb_io_lists_depopulated(wb);
131 return false;
132}
133
Tejun Heof0054bb2015-05-22 17:13:30 -0400134static void wb_wakeup(struct bdi_writeback *wb)
Jan Kara5acda9d2014-04-03 14:46:23 -0700135{
Tejun Heof0054bb2015-05-22 17:13:30 -0400136 spin_lock_bh(&wb->work_lock);
137 if (test_bit(WB_registered, &wb->state))
138 mod_delayed_work(bdi_wq, &wb->dwork, 0);
139 spin_unlock_bh(&wb->work_lock);
Jan Kara5acda9d2014-04-03 14:46:23 -0700140}
141
Tahsin Erdogan4a3a4852017-03-10 12:09:49 -0800142static void finish_writeback_work(struct bdi_writeback *wb,
143 struct wb_writeback_work *work)
144{
145 struct wb_completion *done = work->done;
146
147 if (work->auto_free)
148 kfree(work);
Tejun Heo8e00c4e2019-10-06 17:58:09 -0700149 if (done) {
150 wait_queue_head_t *waitq = done->waitq;
151
152 /* @done can't be accessed after the following dec */
153 if (atomic_dec_and_test(&done->cnt))
154 wake_up_all(waitq);
155 }
Tahsin Erdogan4a3a4852017-03-10 12:09:49 -0800156}
157
Tejun Heof0054bb2015-05-22 17:13:30 -0400158static void wb_queue_work(struct bdi_writeback *wb,
159 struct wb_writeback_work *work)
Jan Kara65850272011-01-13 15:45:44 -0800160{
Tejun Heo5634cc22015-08-18 14:54:56 -0700161 trace_writeback_queue(wb, work);
Jan Kara65850272011-01-13 15:45:44 -0800162
Tejun Heocc395d72015-05-22 17:13:58 -0400163 if (work->done)
164 atomic_inc(&work->done->cnt);
Tahsin Erdogan4a3a4852017-03-10 12:09:49 -0800165
166 spin_lock_bh(&wb->work_lock);
167
168 if (test_bit(WB_registered, &wb->state)) {
169 list_add_tail(&work->list, &wb->work_list);
170 mod_delayed_work(bdi_wq, &wb->dwork, 0);
171 } else
172 finish_writeback_work(wb, work);
173
Tejun Heof0054bb2015-05-22 17:13:30 -0400174 spin_unlock_bh(&wb->work_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175}
176
Tejun Heocc395d72015-05-22 17:13:58 -0400177/**
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
Tejun Heocc395d72015-05-22 17:13:58 -0400179 * @done: target wb_completion
180 *
181 * Wait for one or more work items issued to @bdi with their ->done field
Tejun Heo5b9cce42019-08-26 09:06:52 -0700182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION(). This function returns after all such work items
184 * are completed. Work items which are waited upon aren't freed
Tejun Heocc395d72015-05-22 17:13:58 -0400185 * automatically on completion.
186 */
Tejun Heo5b9cce42019-08-26 09:06:52 -0700187void wb_wait_for_completion(struct wb_completion *done)
Tejun Heocc395d72015-05-22 17:13:58 -0400188{
189 atomic_dec(&done->cnt); /* put down the initial count */
Tejun Heo5b9cce42019-08-26 09:06:52 -0700190 wait_event(*done->waitq, !atomic_read(&done->cnt));
Tejun Heocc395d72015-05-22 17:13:58 -0400191}
192
Tejun Heo703c2702015-05-22 17:13:44 -0400193#ifdef CONFIG_CGROUP_WRITEBACK
194
Tejun Heo55a694d2019-08-15 12:25:28 -0700195/*
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
197 * how they're used.
198 *
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy. All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
202 *
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported. As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
208 *
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch. The recording is quantized
211 * to 16 slots. To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
213 */
Tejun Heo2a814902015-05-28 14:50:51 -0400214#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
215#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
Tejun Heo55a694d2019-08-15 12:25:28 -0700216#define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
Tejun Heo2a814902015-05-28 14:50:51 -0400217#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
218
219#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
220#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221 /* each slot's duration is 2s / 16 */
222#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
223 /* if foreign slots >= 8, switch */
224#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
225 /* one round can affect upto 5 slots */
Tejun Heo6444f472019-08-02 12:08:13 -0700226#define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
Tejun Heo2a814902015-05-28 14:50:51 -0400227
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700228/*
229 * Maximum inodes per isw. A specific value has been chosen to make
230 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
231 */
232#define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
233 / sizeof(struct inode *))
234
Tejun Heoa1a0e232016-02-29 18:28:53 -0500235static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
236static struct workqueue_struct *isw_wq;
237
Tejun Heo21c6321f2015-05-28 14:50:49 -0400238void __inode_attach_wb(struct inode *inode, struct page *page)
239{
240 struct backing_dev_info *bdi = inode_to_bdi(inode);
241 struct bdi_writeback *wb = NULL;
242
243 if (inode_cgwb_enabled(inode)) {
244 struct cgroup_subsys_state *memcg_css;
245
246 if (page) {
247 memcg_css = mem_cgroup_css_from_page(page);
248 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
249 } else {
250 /* must pin memcg_css, see wb_get_create() */
251 memcg_css = task_get_css(current, memory_cgrp_id);
252 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
253 css_put(memcg_css);
254 }
255 }
256
257 if (!wb)
258 wb = &bdi->wb;
259
260 /*
261 * There may be multiple instances of this function racing to
262 * update the same inode. Use cmpxchg() to tell the winner.
263 */
264 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
265 wb_put(wb);
266}
Tejun Heo9b0eb692019-06-27 13:39:48 -0700267EXPORT_SYMBOL_GPL(__inode_attach_wb);
Tejun Heo21c6321f2015-05-28 14:50:49 -0400268
Tejun Heo703c2702015-05-22 17:13:44 -0400269/**
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -0700270 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
271 * @inode: inode of interest with i_lock held
272 * @wb: target bdi_writeback
273 *
274 * Remove the inode from wb's io lists and if necessarily put onto b_attached
275 * list. Only inodes attached to cgwb's are kept on this list.
276 */
277static void inode_cgwb_move_to_attached(struct inode *inode,
278 struct bdi_writeback *wb)
279{
280 assert_spin_locked(&wb->list_lock);
281 assert_spin_locked(&inode->i_lock);
282
283 inode->i_state &= ~I_SYNC_QUEUED;
284 if (wb != &wb->bdi->wb)
285 list_move(&inode->i_io_list, &wb->b_attached);
286 else
287 list_del_init(&inode->i_io_list);
288 wb_io_lists_depopulated(wb);
289}
290
291/**
Tejun Heo87e1d782015-05-28 14:50:52 -0400292 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
293 * @inode: inode of interest with i_lock held
294 *
295 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
296 * held on entry and is released on return. The returned wb is guaranteed
297 * to stay @inode's associated wb until its list_lock is released.
298 */
299static struct bdi_writeback *
300locked_inode_to_wb_and_lock_list(struct inode *inode)
301 __releases(&inode->i_lock)
302 __acquires(&wb->list_lock)
303{
304 while (true) {
305 struct bdi_writeback *wb = inode_to_wb(inode);
306
307 /*
308 * inode_to_wb() association is protected by both
309 * @inode->i_lock and @wb->list_lock but list_lock nests
310 * outside i_lock. Drop i_lock and verify that the
311 * association hasn't changed after acquiring list_lock.
312 */
313 wb_get(wb);
314 spin_unlock(&inode->i_lock);
315 spin_lock(&wb->list_lock);
Tejun Heo87e1d782015-05-28 14:50:52 -0400316
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400317 /* i_wb may have changed inbetween, can't use inode_to_wb() */
Tejun Heo614a4e32016-03-18 13:50:03 -0400318 if (likely(wb == inode->i_wb)) {
319 wb_put(wb); /* @inode already has ref */
320 return wb;
321 }
Tejun Heo87e1d782015-05-28 14:50:52 -0400322
323 spin_unlock(&wb->list_lock);
Tejun Heo614a4e32016-03-18 13:50:03 -0400324 wb_put(wb);
Tejun Heo87e1d782015-05-28 14:50:52 -0400325 cpu_relax();
326 spin_lock(&inode->i_lock);
327 }
328}
329
330/**
331 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
332 * @inode: inode of interest
333 *
334 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
335 * on entry.
336 */
337static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
338 __acquires(&wb->list_lock)
339{
340 spin_lock(&inode->i_lock);
341 return locked_inode_to_wb_and_lock_list(inode);
342}
343
Tejun Heo682aa8e2015-05-28 14:50:53 -0400344struct inode_switch_wbs_context {
Roman Gushchin29264d92021-06-28 19:35:50 -0700345 struct rcu_work work;
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700346
347 /*
348 * Multiple inodes can be switched at once. The switching procedure
349 * consists of two parts, separated by a RCU grace period. To make
350 * sure that the second part is executed for each inode gone through
351 * the first part, all inode pointers are placed into a NULL-terminated
352 * array embedded into struct inode_switch_wbs_context. Otherwise
353 * an inode could be left in a non-consistent state.
354 */
355 struct bdi_writeback *new_wb;
356 struct inode *inodes[];
Tejun Heo682aa8e2015-05-28 14:50:53 -0400357};
358
Tejun Heo7fc58542017-12-12 08:38:30 -0800359static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
360{
361 down_write(&bdi->wb_switch_rwsem);
362}
363
364static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
365{
366 up_write(&bdi->wb_switch_rwsem);
367}
368
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700369static bool inode_do_switch_wbs(struct inode *inode,
370 struct bdi_writeback *old_wb,
Roman Gushchin72d45122021-06-28 19:35:56 -0700371 struct bdi_writeback *new_wb)
Tejun Heo682aa8e2015-05-28 14:50:53 -0400372{
Tejun Heod10c8092015-05-28 14:50:56 -0400373 struct address_space *mapping = inode->i_mapping;
Matthew Wilcox04edf022017-12-04 10:46:23 -0500374 XA_STATE(xas, &mapping->i_pages, 0);
Matthew Wilcox (Oracle)22b3c8d662021-03-19 08:58:36 -0400375 struct folio *folio;
Tejun Heod10c8092015-05-28 14:50:56 -0400376 bool switched = false;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400377
Tejun Heo682aa8e2015-05-28 14:50:53 -0400378 spin_lock(&inode->i_lock);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700379 xa_lock_irq(&mapping->i_pages);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400380
Tejun Heod10c8092015-05-28 14:50:56 -0400381 /*
Roman Gushchin4ade5862021-06-28 19:35:41 -0700382 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
383 * path owns the inode and we shouldn't modify ->i_io_list.
Tejun Heod10c8092015-05-28 14:50:56 -0400384 */
Roman Gushchin4ade5862021-06-28 19:35:41 -0700385 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
Tejun Heod10c8092015-05-28 14:50:56 -0400386 goto skip_switch;
387
Tejun Heo3a8e9ac2019-08-29 15:47:19 -0700388 trace_inode_switch_wbs(inode, old_wb, new_wb);
389
Tejun Heod10c8092015-05-28 14:50:56 -0400390 /*
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
Matthew Wilcox (Oracle)22b3c8d662021-03-19 08:58:36 -0400392 * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
393 * folios actually under writeback.
Tejun Heod10c8092015-05-28 14:50:56 -0400394 */
Matthew Wilcox (Oracle)22b3c8d662021-03-19 08:58:36 -0400395 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
396 if (folio_test_dirty(folio)) {
397 long nr = folio_nr_pages(folio);
398 wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
399 wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
Tejun Heod10c8092015-05-28 14:50:56 -0400400 }
401 }
402
Matthew Wilcox04edf022017-12-04 10:46:23 -0500403 xas_set(&xas, 0);
Matthew Wilcox (Oracle)22b3c8d662021-03-19 08:58:36 -0400404 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
405 long nr = folio_nr_pages(folio);
406 WARN_ON_ONCE(!folio_test_writeback(folio));
407 wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
408 wb_stat_mod(new_wb, WB_WRITEBACK, nr);
Tejun Heod10c8092015-05-28 14:50:56 -0400409 }
410
Jan Kara633a2ab2021-09-02 14:53:04 -0700411 if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
412 atomic_dec(&old_wb->writeback_inodes);
413 atomic_inc(&new_wb->writeback_inodes);
414 }
415
Tejun Heod10c8092015-05-28 14:50:56 -0400416 wb_get(new_wb);
417
418 /*
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -0700419 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
420 * the specific list @inode was on is ignored and the @inode is put on
421 * ->b_dirty which is always correct including from ->b_dirty_time.
422 * The transfer preserves @inode->dirtied_when ordering. If the @inode
423 * was clean, it means it was on the b_attached list, so move it onto
424 * the b_attached list of @new_wb.
Tejun Heod10c8092015-05-28 14:50:56 -0400425 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500426 if (!list_empty(&inode->i_io_list)) {
Tejun Heod10c8092015-05-28 14:50:56 -0400427 inode->i_wb = new_wb;
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -0700428
429 if (inode->i_state & I_DIRTY_ALL) {
430 struct inode *pos;
431
432 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
433 if (time_after_eq(inode->dirtied_when,
434 pos->dirtied_when))
435 break;
436 inode_io_list_move_locked(inode, new_wb,
437 pos->i_io_list.prev);
438 } else {
439 inode_cgwb_move_to_attached(inode, new_wb);
440 }
Tejun Heod10c8092015-05-28 14:50:56 -0400441 } else {
442 inode->i_wb = new_wb;
443 }
444
445 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400446 inode->i_wb_frn_winner = 0;
447 inode->i_wb_frn_avg_time = 0;
448 inode->i_wb_frn_history = 0;
Tejun Heod10c8092015-05-28 14:50:56 -0400449 switched = true;
450skip_switch:
Tejun Heo682aa8e2015-05-28 14:50:53 -0400451 /*
452 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
453 * ensures that the new wb is visible if they see !I_WB_SWITCH.
454 */
455 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
456
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700457 xa_unlock_irq(&mapping->i_pages);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400458 spin_unlock(&inode->i_lock);
Tejun Heod10c8092015-05-28 14:50:56 -0400459
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700460 return switched;
Roman Gushchin72d45122021-06-28 19:35:56 -0700461}
Tejun Heo682aa8e2015-05-28 14:50:53 -0400462
Roman Gushchin72d45122021-06-28 19:35:56 -0700463static void inode_switch_wbs_work_fn(struct work_struct *work)
464{
465 struct inode_switch_wbs_context *isw =
466 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700467 struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
468 struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
469 struct bdi_writeback *new_wb = isw->new_wb;
470 unsigned long nr_switched = 0;
471 struct inode **inodep;
Roman Gushchin72d45122021-06-28 19:35:56 -0700472
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700473 /*
474 * If @inode switches cgwb membership while sync_inodes_sb() is
475 * being issued, sync_inodes_sb() might miss it. Synchronize.
476 */
477 down_read(&bdi->wb_switch_rwsem);
478
479 /*
480 * By the time control reaches here, RCU grace period has passed
481 * since I_WB_SWITCH assertion and all wb stat update transactions
482 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
483 * synchronizing against the i_pages lock.
484 *
485 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
486 * gives us exclusion against all wb related operations on @inode
487 * including IO list manipulations and stat updates.
488 */
489 if (old_wb < new_wb) {
490 spin_lock(&old_wb->list_lock);
491 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
492 } else {
493 spin_lock(&new_wb->list_lock);
494 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
495 }
496
497 for (inodep = isw->inodes; *inodep; inodep++) {
498 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
499 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
500 nr_switched++;
501 }
502
503 spin_unlock(&new_wb->list_lock);
504 spin_unlock(&old_wb->list_lock);
505
506 up_read(&bdi->wb_switch_rwsem);
507
508 if (nr_switched) {
509 wb_wakeup(new_wb);
510 wb_put_many(old_wb, nr_switched);
511 }
512
513 for (inodep = isw->inodes; *inodep; inodep++)
514 iput(*inodep);
515 wb_put(new_wb);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400516 kfree(isw);
Tejun Heoa1a0e232016-02-29 18:28:53 -0500517 atomic_dec(&isw_nr_in_flight);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400518}
519
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700520static bool inode_prepare_wbs_switch(struct inode *inode,
521 struct bdi_writeback *new_wb)
522{
523 /*
524 * Paired with smp_mb() in cgroup_writeback_umount().
525 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
526 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
527 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
528 */
529 smp_mb();
530
Roman Gushchin593311e2021-07-23 15:50:32 -0700531 if (IS_DAX(inode))
532 return false;
533
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700534 /* while holding I_WB_SWITCH, no one else can update the association */
535 spin_lock(&inode->i_lock);
536 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
537 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
538 inode_to_wb(inode) == new_wb) {
539 spin_unlock(&inode->i_lock);
540 return false;
541 }
542 inode->i_state |= I_WB_SWITCH;
543 __iget(inode);
544 spin_unlock(&inode->i_lock);
545
546 return true;
547}
548
Tejun Heo682aa8e2015-05-28 14:50:53 -0400549/**
550 * inode_switch_wbs - change the wb association of an inode
551 * @inode: target inode
552 * @new_wb_id: ID of the new wb
553 *
554 * Switch @inode's wb association to the wb identified by @new_wb_id. The
555 * switching is performed asynchronously and may fail silently.
556 */
557static void inode_switch_wbs(struct inode *inode, int new_wb_id)
558{
559 struct backing_dev_info *bdi = inode_to_bdi(inode);
560 struct cgroup_subsys_state *memcg_css;
561 struct inode_switch_wbs_context *isw;
562
563 /* noop if seems to be already in progress */
564 if (inode->i_state & I_WB_SWITCH)
565 return;
566
Tejun Heo6444f472019-08-02 12:08:13 -0700567 /* avoid queueing a new switch if too many are already in flight */
568 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
Tejun Heo7fc58542017-12-12 08:38:30 -0800569 return;
570
Len Baker98b160c2021-09-25 13:43:08 +0200571 isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400572 if (!isw)
Tejun Heo6444f472019-08-02 12:08:13 -0700573 return;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400574
Roman Gushchin8826ee42021-06-28 19:35:47 -0700575 atomic_inc(&isw_nr_in_flight);
576
Tejun Heo682aa8e2015-05-28 14:50:53 -0400577 /* find and pin the new wb */
578 rcu_read_lock();
579 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
Muchun Song8b0ed842021-04-02 17:11:45 +0800580 if (memcg_css && !css_tryget(memcg_css))
581 memcg_css = NULL;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400582 rcu_read_unlock();
Muchun Song8b0ed842021-04-02 17:11:45 +0800583 if (!memcg_css)
584 goto out_free;
585
586 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
587 css_put(memcg_css);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400588 if (!isw->new_wb)
589 goto out_free;
590
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700591 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
Tejun Heoa1a0e232016-02-29 18:28:53 -0500592 goto out_free;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400593
Roman Gushchinf5fbe6b2021-06-28 19:35:59 -0700594 isw->inodes[0] = inode;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400595
596 /*
597 * In addition to synchronizing among switchers, I_WB_SWITCH tells
Matthew Wilcoxb93b0162018-04-10 16:36:56 -0700598 * the RCU protected stat update paths to grab the i_page
599 * lock so that stat transfer can synchronize against them.
Tejun Heo682aa8e2015-05-28 14:50:53 -0400600 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
601 */
Roman Gushchin29264d92021-06-28 19:35:50 -0700602 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
603 queue_rcu_work(isw_wq, &isw->work);
Tejun Heo6444f472019-08-02 12:08:13 -0700604 return;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400605
606out_free:
Roman Gushchin8826ee42021-06-28 19:35:47 -0700607 atomic_dec(&isw_nr_in_flight);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400608 if (isw->new_wb)
609 wb_put(isw->new_wb);
610 kfree(isw);
611}
612
Tejun Heo87e1d782015-05-28 14:50:52 -0400613/**
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700614 * cleanup_offline_cgwb - detach associated inodes
615 * @wb: target wb
616 *
617 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
618 * to eventually release the dying @wb. Returns %true if not all inodes were
619 * switched and the function has to be restarted.
620 */
621bool cleanup_offline_cgwb(struct bdi_writeback *wb)
622{
623 struct cgroup_subsys_state *memcg_css;
624 struct inode_switch_wbs_context *isw;
625 struct inode *inode;
626 int nr;
627 bool restart = false;
628
Len Baker98b160c2021-09-25 13:43:08 +0200629 isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
630 GFP_KERNEL);
Roman Gushchinc22d70a2021-06-28 19:36:03 -0700631 if (!isw)
632 return restart;
633
634 atomic_inc(&isw_nr_in_flight);
635
636 for (memcg_css = wb->memcg_css->parent; memcg_css;
637 memcg_css = memcg_css->parent) {
638 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
639 if (isw->new_wb)
640 break;
641 }
642 if (unlikely(!isw->new_wb))
643 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
644
645 nr = 0;
646 spin_lock(&wb->list_lock);
647 list_for_each_entry(inode, &wb->b_attached, i_io_list) {
648 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
649 continue;
650
651 isw->inodes[nr++] = inode;
652
653 if (nr >= WB_MAX_INODES_PER_ISW - 1) {
654 restart = true;
655 break;
656 }
657 }
658 spin_unlock(&wb->list_lock);
659
660 /* no attached inodes? bail out */
661 if (nr == 0) {
662 atomic_dec(&isw_nr_in_flight);
663 wb_put(isw->new_wb);
664 kfree(isw);
665 return restart;
666 }
667
668 /*
669 * In addition to synchronizing among switchers, I_WB_SWITCH tells
670 * the RCU protected stat update paths to grab the i_page
671 * lock so that stat transfer can synchronize against them.
672 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
673 */
674 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
675 queue_rcu_work(isw_wq, &isw->work);
676
677 return restart;
678}
679
680/**
Tejun Heob16b1de2015-06-02 08:39:48 -0600681 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
682 * @wbc: writeback_control of interest
683 * @inode: target inode
684 *
685 * @inode is locked and about to be written back under the control of @wbc.
686 * Record @inode's writeback context into @wbc and unlock the i_lock. On
687 * writeback completion, wbc_detach_inode() should be called. This is used
688 * to track the cgroup writeback context.
689 */
690void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
691 struct inode *inode)
692{
Tejun Heodd73e4b2015-06-16 18:48:30 -0400693 if (!inode_cgwb_enabled(inode)) {
694 spin_unlock(&inode->i_lock);
695 return;
696 }
697
Tejun Heob16b1de2015-06-02 08:39:48 -0600698 wbc->wb = inode_to_wb(inode);
Tejun Heo2a814902015-05-28 14:50:51 -0400699 wbc->inode = inode;
700
701 wbc->wb_id = wbc->wb->memcg_css->id;
702 wbc->wb_lcand_id = inode->i_wb_frn_winner;
703 wbc->wb_tcand_id = 0;
704 wbc->wb_bytes = 0;
705 wbc->wb_lcand_bytes = 0;
706 wbc->wb_tcand_bytes = 0;
707
Tejun Heob16b1de2015-06-02 08:39:48 -0600708 wb_get(wbc->wb);
709 spin_unlock(&inode->i_lock);
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400710
711 /*
Tejun Heo65de03e2019-11-08 12:18:29 -0800712 * A dying wb indicates that either the blkcg associated with the
713 * memcg changed or the associated memcg is dying. In the first
714 * case, a replacement wb should already be available and we should
715 * refresh the wb immediately. In the second case, trying to
716 * refresh will keep failing.
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400717 */
Tejun Heo65de03e2019-11-08 12:18:29 -0800718 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400719 inode_switch_wbs(inode, wbc->wb_id);
Tejun Heob16b1de2015-06-02 08:39:48 -0600720}
Tejun Heo9b0eb692019-06-27 13:39:48 -0700721EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
Tejun Heob16b1de2015-06-02 08:39:48 -0600722
723/**
Tejun Heo2a814902015-05-28 14:50:51 -0400724 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
725 * @wbc: writeback_control of the just finished writeback
Tejun Heob16b1de2015-06-02 08:39:48 -0600726 *
727 * To be called after a writeback attempt of an inode finishes and undoes
728 * wbc_attach_and_unlock_inode(). Can be called under any context.
Tejun Heo2a814902015-05-28 14:50:51 -0400729 *
730 * As concurrent write sharing of an inode is expected to be very rare and
731 * memcg only tracks page ownership on first-use basis severely confining
732 * the usefulness of such sharing, cgroup writeback tracks ownership
733 * per-inode. While the support for concurrent write sharing of an inode
734 * is deemed unnecessary, an inode being written to by different cgroups at
735 * different points in time is a lot more common, and, more importantly,
736 * charging only by first-use can too readily lead to grossly incorrect
737 * behaviors (single foreign page can lead to gigabytes of writeback to be
738 * incorrectly attributed).
739 *
740 * To resolve this issue, cgroup writeback detects the majority dirtier of
741 * an inode and transfers the ownership to it. To avoid unnnecessary
742 * oscillation, the detection mechanism keeps track of history and gives
743 * out the switch verdict only if the foreign usage pattern is stable over
744 * a certain amount of time and/or writeback attempts.
745 *
746 * On each writeback attempt, @wbc tries to detect the majority writer
747 * using Boyer-Moore majority vote algorithm. In addition to the byte
748 * count from the majority voting, it also counts the bytes written for the
749 * current wb and the last round's winner wb (max of last round's current
750 * wb, the winner from two rounds ago, and the last round's majority
751 * candidate). Keeping track of the historical winner helps the algorithm
752 * to semi-reliably detect the most active writer even when it's not the
753 * absolute majority.
754 *
755 * Once the winner of the round is determined, whether the winner is
756 * foreign or not and how much IO time the round consumed is recorded in
757 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
758 * over a certain threshold, the switch verdict is given.
Tejun Heob16b1de2015-06-02 08:39:48 -0600759 */
760void wbc_detach_inode(struct writeback_control *wbc)
761{
Tejun Heo2a814902015-05-28 14:50:51 -0400762 struct bdi_writeback *wb = wbc->wb;
763 struct inode *inode = wbc->inode;
Tejun Heodd73e4b2015-06-16 18:48:30 -0400764 unsigned long avg_time, max_bytes, max_time;
765 u16 history;
Tejun Heo2a814902015-05-28 14:50:51 -0400766 int max_id;
767
Tejun Heodd73e4b2015-06-16 18:48:30 -0400768 if (!wb)
769 return;
770
771 history = inode->i_wb_frn_history;
772 avg_time = inode->i_wb_frn_avg_time;
773
Tejun Heo2a814902015-05-28 14:50:51 -0400774 /* pick the winner of this round */
775 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
776 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
777 max_id = wbc->wb_id;
778 max_bytes = wbc->wb_bytes;
779 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
780 max_id = wbc->wb_lcand_id;
781 max_bytes = wbc->wb_lcand_bytes;
782 } else {
783 max_id = wbc->wb_tcand_id;
784 max_bytes = wbc->wb_tcand_bytes;
785 }
786
787 /*
788 * Calculate the amount of IO time the winner consumed and fold it
789 * into the running average kept per inode. If the consumed IO
790 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
791 * deciding whether to switch or not. This is to prevent one-off
792 * small dirtiers from skewing the verdict.
793 */
794 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
795 wb->avg_write_bandwidth);
796 if (avg_time)
797 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
798 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
799 else
800 avg_time = max_time; /* immediate catch up on first run */
801
802 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
803 int slots;
804
805 /*
806 * The switch verdict is reached if foreign wb's consume
807 * more than a certain proportion of IO time in a
808 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
809 * history mask where each bit represents one sixteenth of
810 * the period. Determine the number of slots to shift into
811 * history from @max_time.
812 */
813 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
814 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
815 history <<= slots;
816 if (wbc->wb_id != max_id)
817 history |= (1U << slots) - 1;
818
Tejun Heo3a8e9ac2019-08-29 15:47:19 -0700819 if (history)
820 trace_inode_foreign_history(inode, wbc, history);
821
Tejun Heo2a814902015-05-28 14:50:51 -0400822 /*
823 * Switch if the current wb isn't the consistent winner.
824 * If there are multiple closely competing dirtiers, the
825 * inode may switch across them repeatedly over time, which
826 * is okay. The main goal is avoiding keeping an inode on
827 * the wrong wb for an extended period of time.
828 */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400829 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
830 inode_switch_wbs(inode, max_id);
Tejun Heo2a814902015-05-28 14:50:51 -0400831 }
832
833 /*
834 * Multiple instances of this function may race to update the
835 * following fields but we don't mind occassional inaccuracies.
836 */
837 inode->i_wb_frn_winner = max_id;
838 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
839 inode->i_wb_frn_history = history;
840
Tejun Heob16b1de2015-06-02 08:39:48 -0600841 wb_put(wbc->wb);
842 wbc->wb = NULL;
843}
Tejun Heo9b0eb692019-06-27 13:39:48 -0700844EXPORT_SYMBOL_GPL(wbc_detach_inode);
Tejun Heob16b1de2015-06-02 08:39:48 -0600845
846/**
Tejun Heo34e51a52019-06-27 13:39:49 -0700847 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
Tejun Heo2a814902015-05-28 14:50:51 -0400848 * @wbc: writeback_control of the writeback in progress
849 * @page: page being written out
850 * @bytes: number of bytes being written out
851 *
852 * @bytes from @page are about to written out during the writeback
853 * controlled by @wbc. Keep the book for foreign inode detection. See
854 * wbc_detach_inode().
855 */
Tejun Heo34e51a52019-06-27 13:39:49 -0700856void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
857 size_t bytes)
Tejun Heo2a814902015-05-28 14:50:51 -0400858{
Tejun Heo66311422019-06-13 15:30:41 -0700859 struct cgroup_subsys_state *css;
Tejun Heo2a814902015-05-28 14:50:51 -0400860 int id;
861
862 /*
863 * pageout() path doesn't attach @wbc to the inode being written
864 * out. This is intentional as we don't want the function to block
865 * behind a slow cgroup. Ultimately, we want pageout() to kick off
866 * regular writeback instead of writing things out itself.
867 */
Tejun Heo27b36d82019-06-27 13:39:50 -0700868 if (!wbc->wb || wbc->no_cgroup_owner)
Tejun Heo2a814902015-05-28 14:50:51 -0400869 return;
870
Tejun Heo66311422019-06-13 15:30:41 -0700871 css = mem_cgroup_css_from_page(page);
872 /* dead cgroups shouldn't contribute to inode ownership arbitration */
873 if (!(css->flags & CSS_ONLINE))
874 return;
875
876 id = css->id;
Tejun Heo2a814902015-05-28 14:50:51 -0400877
878 if (id == wbc->wb_id) {
879 wbc->wb_bytes += bytes;
880 return;
881 }
882
883 if (id == wbc->wb_lcand_id)
884 wbc->wb_lcand_bytes += bytes;
885
886 /* Boyer-Moore majority vote algorithm */
887 if (!wbc->wb_tcand_bytes)
888 wbc->wb_tcand_id = id;
889 if (id == wbc->wb_tcand_id)
890 wbc->wb_tcand_bytes += bytes;
891 else
892 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
893}
Tejun Heo34e51a52019-06-27 13:39:49 -0700894EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
Tejun Heo2a814902015-05-28 14:50:51 -0400895
896/**
Tejun Heo703c2702015-05-22 17:13:44 -0400897 * inode_congested - test whether an inode is congested
Tejun Heo60292bc2015-08-18 14:54:54 -0700898 * @inode: inode to test for congestion (may be NULL)
Tejun Heo703c2702015-05-22 17:13:44 -0400899 * @cong_bits: mask of WB_[a]sync_congested bits to test
900 *
901 * Tests whether @inode is congested. @cong_bits is the mask of congestion
902 * bits to test and the return value is the mask of set bits.
903 *
904 * If cgroup writeback is enabled for @inode, the congestion state is
905 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
906 * associated with @inode is congested; otherwise, the root wb's congestion
907 * state is used.
Tejun Heo60292bc2015-08-18 14:54:54 -0700908 *
909 * @inode is allowed to be NULL as this function is often called on
910 * mapping->host which is NULL for the swapper space.
Tejun Heo703c2702015-05-22 17:13:44 -0400911 */
912int inode_congested(struct inode *inode, int cong_bits)
913{
Tejun Heo5cb8b822015-05-28 14:50:54 -0400914 /*
915 * Once set, ->i_wb never becomes NULL while the inode is alive.
916 * Start transaction iff ->i_wb is visible.
917 */
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400918 if (inode && inode_to_wb_is_valid(inode)) {
Tejun Heo5cb8b822015-05-28 14:50:54 -0400919 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -0700920 struct wb_lock_cookie lock_cookie = {};
921 bool congested;
Tejun Heo5cb8b822015-05-28 14:50:54 -0400922
Greg Thelen2e898e42018-04-20 14:55:42 -0700923 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
Tejun Heo5cb8b822015-05-28 14:50:54 -0400924 congested = wb_congested(wb, cong_bits);
Greg Thelen2e898e42018-04-20 14:55:42 -0700925 unlocked_inode_to_wb_end(inode, &lock_cookie);
Tejun Heo5cb8b822015-05-28 14:50:54 -0400926 return congested;
Tejun Heo703c2702015-05-22 17:13:44 -0400927 }
928
929 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
930}
931EXPORT_SYMBOL_GPL(inode_congested);
932
Tejun Heof2b65122015-05-22 17:13:55 -0400933/**
934 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
935 * @wb: target bdi_writeback to split @nr_pages to
936 * @nr_pages: number of pages to write for the whole bdi
937 *
938 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
939 * relation to the total write bandwidth of all wb's w/ dirty inodes on
940 * @wb->bdi.
941 */
942static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
943{
944 unsigned long this_bw = wb->avg_write_bandwidth;
945 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
946
947 if (nr_pages == LONG_MAX)
948 return LONG_MAX;
949
950 /*
951 * This may be called on clean wb's and proportional distribution
952 * may not make sense, just use the original @nr_pages in those
953 * cases. In general, we wanna err on the side of writing more.
954 */
955 if (!tot_bw || this_bw >= tot_bw)
956 return nr_pages;
957 else
958 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
959}
960
Tejun Heodb125362015-05-22 17:14:01 -0400961/**
Tejun Heodb125362015-05-22 17:14:01 -0400962 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
963 * @bdi: target backing_dev_info
964 * @base_work: wb_writeback_work to issue
965 * @skip_if_busy: skip wb's which already have writeback in progress
966 *
967 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
968 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
969 * distributed to the busy wbs according to each wb's proportion in the
970 * total active write bandwidth of @bdi.
971 */
972static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
973 struct wb_writeback_work *base_work,
974 bool skip_if_busy)
975{
Tejun Heob8175252015-10-02 14:47:05 -0400976 struct bdi_writeback *last_wb = NULL;
Tejun Heob33e18f2015-10-27 14:19:39 +0900977 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
978 struct bdi_writeback, bdi_node);
Tejun Heodb125362015-05-22 17:14:01 -0400979
980 might_sleep();
Tejun Heodb125362015-05-22 17:14:01 -0400981restart:
982 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -0400983 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
Tejun Heo5b9cce42019-08-26 09:06:52 -0700984 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
Tejun Heo8a1270c2015-08-18 14:54:53 -0700985 struct wb_writeback_work fallback_work;
986 struct wb_writeback_work *work;
987 long nr_pages;
988
Tejun Heob8175252015-10-02 14:47:05 -0400989 if (last_wb) {
990 wb_put(last_wb);
991 last_wb = NULL;
992 }
993
Tejun Heo006a0972015-08-25 14:11:52 -0400994 /* SYNC_ALL writes out I_DIRTY_TIME too */
995 if (!wb_has_dirty_io(wb) &&
996 (base_work->sync_mode == WB_SYNC_NONE ||
997 list_empty(&wb->b_dirty_time)))
998 continue;
999 if (skip_if_busy && writeback_in_progress(wb))
Tejun Heodb125362015-05-22 17:14:01 -04001000 continue;
1001
Tejun Heo8a1270c2015-08-18 14:54:53 -07001002 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1003
1004 work = kmalloc(sizeof(*work), GFP_ATOMIC);
1005 if (work) {
1006 *work = *base_work;
1007 work->nr_pages = nr_pages;
1008 work->auto_free = 1;
1009 wb_queue_work(wb, work);
1010 continue;
Tejun Heodb125362015-05-22 17:14:01 -04001011 }
Tejun Heo8a1270c2015-08-18 14:54:53 -07001012
1013 /* alloc failed, execute synchronously using on-stack fallback */
1014 work = &fallback_work;
1015 *work = *base_work;
1016 work->nr_pages = nr_pages;
1017 work->auto_free = 0;
1018 work->done = &fallback_work_done;
1019
1020 wb_queue_work(wb, work);
1021
Tejun Heob8175252015-10-02 14:47:05 -04001022 /*
1023 * Pin @wb so that it stays on @bdi->wb_list. This allows
1024 * continuing iteration from @wb after dropping and
1025 * regrabbing rcu read lock.
1026 */
1027 wb_get(wb);
1028 last_wb = wb;
1029
Tejun Heo8a1270c2015-08-18 14:54:53 -07001030 rcu_read_unlock();
Tejun Heo5b9cce42019-08-26 09:06:52 -07001031 wb_wait_for_completion(&fallback_work_done);
Tejun Heo8a1270c2015-08-18 14:54:53 -07001032 goto restart;
Tejun Heodb125362015-05-22 17:14:01 -04001033 }
1034 rcu_read_unlock();
Tejun Heob8175252015-10-02 14:47:05 -04001035
1036 if (last_wb)
1037 wb_put(last_wb);
Tejun Heodb125362015-05-22 17:14:01 -04001038}
1039
Tejun Heoa1a0e232016-02-29 18:28:53 -05001040/**
Tejun Heod62241c72019-08-26 09:06:55 -07001041 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1042 * @bdi_id: target bdi id
1043 * @memcg_id: target memcg css id
Tejun Heod62241c72019-08-26 09:06:55 -07001044 * @reason: reason why some writeback work initiated
1045 * @done: target wb_completion
1046 *
1047 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1048 * with the specified parameters.
1049 */
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001050int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
Tejun Heod62241c72019-08-26 09:06:55 -07001051 enum wb_reason reason, struct wb_completion *done)
1052{
1053 struct backing_dev_info *bdi;
1054 struct cgroup_subsys_state *memcg_css;
1055 struct bdi_writeback *wb;
1056 struct wb_writeback_work *work;
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001057 unsigned long dirty;
Tejun Heod62241c72019-08-26 09:06:55 -07001058 int ret;
1059
1060 /* lookup bdi and memcg */
1061 bdi = bdi_get_by_id(bdi_id);
1062 if (!bdi)
1063 return -ENOENT;
1064
1065 rcu_read_lock();
1066 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1067 if (memcg_css && !css_tryget(memcg_css))
1068 memcg_css = NULL;
1069 rcu_read_unlock();
1070 if (!memcg_css) {
1071 ret = -ENOENT;
1072 goto out_bdi_put;
1073 }
1074
1075 /*
1076 * And find the associated wb. If the wb isn't there already
1077 * there's nothing to flush, don't create one.
1078 */
1079 wb = wb_get_lookup(bdi, memcg_css);
1080 if (!wb) {
1081 ret = -ENOENT;
1082 goto out_css_put;
1083 }
1084
1085 /*
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001086 * The caller is attempting to write out most of
Tejun Heod62241c72019-08-26 09:06:55 -07001087 * the currently dirty pages. Let's take the current dirty page
1088 * count and inflate it by 25% which should be large enough to
1089 * flush out most dirty pages while avoiding getting livelocked by
1090 * concurrent dirtiers.
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001091 *
1092 * BTW the memcg stats are flushed periodically and this is best-effort
1093 * estimation, so some potential error is ok.
Tejun Heod62241c72019-08-26 09:06:55 -07001094 */
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001095 dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1096 dirty = dirty * 10 / 8;
Tejun Heod62241c72019-08-26 09:06:55 -07001097
1098 /* issue the writeback work */
1099 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1100 if (work) {
Shakeel Butt7490a2d2021-09-02 14:53:27 -07001101 work->nr_pages = dirty;
Tejun Heod62241c72019-08-26 09:06:55 -07001102 work->sync_mode = WB_SYNC_NONE;
1103 work->range_cyclic = 1;
1104 work->reason = reason;
1105 work->done = done;
1106 work->auto_free = 1;
1107 wb_queue_work(wb, work);
1108 ret = 0;
1109 } else {
1110 ret = -ENOMEM;
1111 }
1112
1113 wb_put(wb);
1114out_css_put:
1115 css_put(memcg_css);
1116out_bdi_put:
1117 bdi_put(bdi);
1118 return ret;
1119}
1120
1121/**
Tejun Heoa1a0e232016-02-29 18:28:53 -05001122 * cgroup_writeback_umount - flush inode wb switches for umount
1123 *
1124 * This function is called when a super_block is about to be destroyed and
1125 * flushes in-flight inode wb switches. An inode wb switch goes through
1126 * RCU and then workqueue, so the two need to be flushed in order to ensure
1127 * that all previously scheduled switches are finished. As wb switches are
1128 * rare occurrences and synchronize_rcu() can take a while, perform
1129 * flushing iff wb switches are in flight.
1130 */
1131void cgroup_writeback_umount(void)
1132{
Roman Gushchin592fa002021-06-28 19:35:44 -07001133 /*
1134 * SB_ACTIVE should be reliably cleared before checking
1135 * isw_nr_in_flight, see generic_shutdown_super().
1136 */
1137 smp_mb();
1138
Tejun Heoa1a0e232016-02-29 18:28:53 -05001139 if (atomic_read(&isw_nr_in_flight)) {
Jiufei Xueec084de2019-05-17 14:31:44 -07001140 /*
1141 * Use rcu_barrier() to wait for all pending callbacks to
1142 * ensure that all in-flight wb switches are in the workqueue.
1143 */
1144 rcu_barrier();
Tejun Heoa1a0e232016-02-29 18:28:53 -05001145 flush_workqueue(isw_wq);
1146 }
1147}
1148
1149static int __init cgroup_writeback_init(void)
1150{
1151 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1152 if (!isw_wq)
1153 return -ENOMEM;
1154 return 0;
1155}
1156fs_initcall(cgroup_writeback_init);
1157
Tejun Heof2b65122015-05-22 17:13:55 -04001158#else /* CONFIG_CGROUP_WRITEBACK */
1159
Tejun Heo7fc58542017-12-12 08:38:30 -08001160static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1161static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1162
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -07001163static void inode_cgwb_move_to_attached(struct inode *inode,
1164 struct bdi_writeback *wb)
1165{
1166 assert_spin_locked(&wb->list_lock);
1167 assert_spin_locked(&inode->i_lock);
1168
1169 inode->i_state &= ~I_SYNC_QUEUED;
1170 list_del_init(&inode->i_io_list);
1171 wb_io_lists_depopulated(wb);
1172}
1173
Tejun Heo87e1d782015-05-28 14:50:52 -04001174static struct bdi_writeback *
1175locked_inode_to_wb_and_lock_list(struct inode *inode)
1176 __releases(&inode->i_lock)
1177 __acquires(&wb->list_lock)
1178{
1179 struct bdi_writeback *wb = inode_to_wb(inode);
1180
1181 spin_unlock(&inode->i_lock);
1182 spin_lock(&wb->list_lock);
1183 return wb;
1184}
1185
1186static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1187 __acquires(&wb->list_lock)
1188{
1189 struct bdi_writeback *wb = inode_to_wb(inode);
1190
1191 spin_lock(&wb->list_lock);
1192 return wb;
1193}
1194
Tejun Heof2b65122015-05-22 17:13:55 -04001195static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1196{
1197 return nr_pages;
1198}
1199
Tejun Heodb125362015-05-22 17:14:01 -04001200static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1201 struct wb_writeback_work *base_work,
1202 bool skip_if_busy)
1203{
1204 might_sleep();
1205
Tejun Heo006a0972015-08-25 14:11:52 -04001206 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
Tejun Heodb125362015-05-22 17:14:01 -04001207 base_work->auto_free = 0;
Tejun Heodb125362015-05-22 17:14:01 -04001208 wb_queue_work(&bdi->wb, base_work);
1209 }
1210}
1211
Tejun Heo703c2702015-05-22 17:13:44 -04001212#endif /* CONFIG_CGROUP_WRITEBACK */
1213
Jens Axboee8e8a0c2017-09-28 11:31:22 -06001214/*
1215 * Add in the number of potentially dirty inodes, because each inode
1216 * write can dirty pagecache in the underlying blockdev.
1217 */
1218static unsigned long get_nr_dirty_pages(void)
1219{
1220 return global_node_page_state(NR_FILE_DIRTY) +
Jens Axboee8e8a0c2017-09-28 11:31:22 -06001221 get_nr_dirty_inodes();
1222}
1223
1224static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +02001225{
Tejun Heoc00ddad2015-05-22 17:13:51 -04001226 if (!wb_has_dirty_io(wb))
1227 return;
1228
1229 /*
Jens Axboeaac8d412017-09-28 11:31:55 -06001230 * All callers of this function want to start writeback of all
1231 * dirty pages. Places like vmscan can call this at a very
1232 * high frequency, causing pointless allocations of tons of
1233 * work items and keeping the flusher threads busy retrieving
1234 * that work. Ensure that we only allow one of them pending and
Jens Axboe85009b42017-09-30 02:09:06 -06001235 * inflight at the time.
Jens Axboeaac8d412017-09-28 11:31:55 -06001236 */
Jens Axboe85009b42017-09-30 02:09:06 -06001237 if (test_bit(WB_start_all, &wb->state) ||
1238 test_and_set_bit(WB_start_all, &wb->state))
Jens Axboeaac8d412017-09-28 11:31:55 -06001239 return;
1240
Jens Axboe85009b42017-09-30 02:09:06 -06001241 wb->start_all_reason = reason;
1242 wb_wakeup(wb);
Christoph Hellwigc5444192010-06-08 18:15:15 +02001243}
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001244
Christoph Hellwigc5444192010-06-08 18:15:15 +02001245/**
Tejun Heo9ecf48662015-05-22 17:13:54 -04001246 * wb_start_background_writeback - start background writeback
1247 * @wb: bdi_writback to write from
Christoph Hellwigc5444192010-06-08 18:15:15 +02001248 *
1249 * Description:
Jan Kara65850272011-01-13 15:45:44 -08001250 * This makes sure WB_SYNC_NONE background writeback happens. When
Tejun Heo9ecf48662015-05-22 17:13:54 -04001251 * this function returns, it is only guaranteed that for given wb
Jan Kara65850272011-01-13 15:45:44 -08001252 * some IO is happening if we are over background dirty threshold.
1253 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +02001254 */
Tejun Heo9ecf48662015-05-22 17:13:54 -04001255void wb_start_background_writeback(struct bdi_writeback *wb)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001256{
Jan Kara65850272011-01-13 15:45:44 -08001257 /*
1258 * We just wake up the flusher thread. It will perform background
1259 * writeback as soon as there is no other work to do.
1260 */
Tejun Heo5634cc22015-08-18 14:54:56 -07001261 trace_writeback_wake_background(wb);
Tejun Heo9ecf48662015-05-22 17:13:54 -04001262 wb_wakeup(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263}
1264
1265/*
Dave Chinnera66979a2011-03-22 22:23:41 +11001266 * Remove the inode from the writeback list it is on.
1267 */
Dave Chinnerc7f54082015-03-04 14:07:22 -05001268void inode_io_list_del(struct inode *inode)
Dave Chinnera66979a2011-03-22 22:23:41 +11001269{
Tejun Heo87e1d782015-05-28 14:50:52 -04001270 struct bdi_writeback *wb;
Dave Chinnera66979a2011-03-22 22:23:41 +11001271
Tejun Heo87e1d782015-05-28 14:50:52 -04001272 wb = inode_to_wb_and_lock_list(inode);
Jan Karab35250c2020-06-10 17:36:03 +02001273 spin_lock(&inode->i_lock);
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -07001274
1275 inode->i_state &= ~I_SYNC_QUEUED;
1276 list_del_init(&inode->i_io_list);
1277 wb_io_lists_depopulated(wb);
1278
Jan Karab35250c2020-06-10 17:36:03 +02001279 spin_unlock(&inode->i_lock);
Tejun Heo52ebea72015-05-22 17:13:37 -04001280 spin_unlock(&wb->list_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001281}
Jan Kara4301efa2020-04-21 10:54:44 +02001282EXPORT_SYMBOL(inode_io_list_del);
Dave Chinnera66979a2011-03-22 22:23:41 +11001283
1284/*
Dave Chinner6c60d2b2016-07-26 15:21:50 -07001285 * mark an inode as under writeback on the sb
1286 */
1287void sb_mark_inode_writeback(struct inode *inode)
1288{
1289 struct super_block *sb = inode->i_sb;
1290 unsigned long flags;
1291
1292 if (list_empty(&inode->i_wb_list)) {
1293 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
Brian Foster9a46b04f2016-07-26 15:21:53 -07001294 if (list_empty(&inode->i_wb_list)) {
Dave Chinner6c60d2b2016-07-26 15:21:50 -07001295 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
Brian Foster9a46b04f2016-07-26 15:21:53 -07001296 trace_sb_mark_inode_writeback(inode);
1297 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07001298 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1299 }
1300}
1301
1302/*
1303 * clear an inode as under writeback on the sb
1304 */
1305void sb_clear_inode_writeback(struct inode *inode)
1306{
1307 struct super_block *sb = inode->i_sb;
1308 unsigned long flags;
1309
1310 if (!list_empty(&inode->i_wb_list)) {
1311 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
Brian Foster9a46b04f2016-07-26 15:21:53 -07001312 if (!list_empty(&inode->i_wb_list)) {
1313 list_del_init(&inode->i_wb_list);
1314 trace_sb_clear_inode_writeback(inode);
1315 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07001316 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1317 }
1318}
1319
1320/*
Andrew Morton6610a0b2007-10-16 23:30:32 -07001321 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1322 * furthest end of its superblock's dirty-inode list.
1323 *
1324 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001325 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -07001326 * the case then the inode must have been redirtied while it was being written
1327 * out and we don't reset its dirtied_when.
1328 */
Jan Karab35250c2020-06-10 17:36:03 +02001329static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -07001330{
Jan Karab35250c2020-06-10 17:36:03 +02001331 assert_spin_locked(&inode->i_lock);
1332
Jens Axboe03ba3782009-09-09 09:08:54 +02001333 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001334 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -07001335
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001336 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001337 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -07001338 inode->dirtied_when = jiffies;
1339 }
Dave Chinnerc7f54082015-03-04 14:07:22 -05001340 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
Jan Kara5afced32020-05-29 15:05:22 +02001341 inode->i_state &= ~I_SYNC_QUEUED;
Andrew Morton6610a0b2007-10-16 23:30:32 -07001342}
1343
Jan Karab35250c2020-06-10 17:36:03 +02001344static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1345{
1346 spin_lock(&inode->i_lock);
1347 redirty_tail_locked(inode, wb);
1348 spin_unlock(&inode->i_lock);
1349}
1350
Andrew Morton6610a0b2007-10-16 23:30:32 -07001351/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001352 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001353 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001354static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001355{
Dave Chinnerc7f54082015-03-04 14:07:22 -05001356 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001357}
1358
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001359static void inode_sync_complete(struct inode *inode)
1360{
Jan Kara365b94ae2012-05-03 14:47:55 +02001361 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -08001362 /* If inode is clean an unused, put it into LRU now... */
1363 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +02001364 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001365 smp_mb();
1366 wake_up_bit(&inode->i_state, __I_SYNC);
1367}
1368
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001369static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1370{
1371 bool ret = time_after(inode->dirtied_when, t);
1372#ifndef CONFIG_64BIT
1373 /*
1374 * For inodes being constantly redirtied, dirtied_when can get stuck.
1375 * It _appears_ to be in the future, but is actually in distant past.
1376 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +02001377 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001378 */
1379 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1380#endif
1381 return ret;
1382}
1383
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001384#define EXPIRE_DIRTY_ATIME 0x0001
1385
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001386/*
Jan Karaf9cae922020-05-29 16:08:58 +02001387 * Move expired (dirtied before dirtied_before) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -08001388 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -07001389 */
Wu Fengguange84d0a42011-04-23 12:27:27 -06001390static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -07001391 struct list_head *dispatch_queue,
Jan Kara5fcd5752020-05-29 16:24:43 +02001392 unsigned long dirtied_before)
Fengguang Wu2c136572007-10-16 23:30:39 -07001393{
Shaohua Li5c034492009-09-24 14:42:33 +02001394 LIST_HEAD(tmp);
1395 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +02001396 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +02001397 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +02001398 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -06001399 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +02001400
Fengguang Wu2c136572007-10-16 23:30:39 -07001401 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001402 inode = wb_inode(delaying_queue->prev);
Jan Karaf9cae922020-05-29 16:08:58 +02001403 if (inode_dirtied_after(inode, dirtied_before))
Fengguang Wu2c136572007-10-16 23:30:39 -07001404 break;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001405 list_move(&inode->i_io_list, &tmp);
Jan Karaa8855992013-07-09 22:36:45 +08001406 moved++;
Jan Kara5afced32020-05-29 15:05:22 +02001407 spin_lock(&inode->i_lock);
Jan Kara5afced32020-05-29 15:05:22 +02001408 inode->i_state |= I_SYNC_QUEUED;
1409 spin_unlock(&inode->i_lock);
Jan Karaa8855992013-07-09 22:36:45 +08001410 if (sb_is_blkdev_sb(inode->i_sb))
1411 continue;
Jens Axboecf137302009-09-24 15:12:57 +02001412 if (sb && sb != inode->i_sb)
1413 do_sb_sort = 1;
1414 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001415 }
1416
Jens Axboecf137302009-09-24 15:12:57 +02001417 /* just one sb in list, splice to dispatch_queue and we're done */
1418 if (!do_sb_sort) {
1419 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -06001420 goto out;
Jens Axboecf137302009-09-24 15:12:57 +02001421 }
1422
Shaohua Li5c034492009-09-24 14:42:33 +02001423 /* Move inodes from one superblock together */
1424 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001425 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001426 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001427 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +02001428 if (inode->i_sb == sb)
Dave Chinnerc7f54082015-03-04 14:07:22 -05001429 list_move(&inode->i_io_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +02001430 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001431 }
Wu Fengguange84d0a42011-04-23 12:27:27 -06001432out:
1433 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -07001434}
1435
1436/*
1437 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001438 * Before
1439 * newly dirtied b_dirty b_io b_more_io
1440 * =============> gf edc BA
1441 * After
1442 * newly dirtied b_dirty b_io b_more_io
1443 * =============> g fBAedc
1444 * |
1445 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -07001446 */
Jan Karaf9cae922020-05-29 16:08:58 +02001447static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1448 unsigned long dirtied_before)
Fengguang Wu2c136572007-10-16 23:30:39 -07001449{
Wu Fengguange84d0a42011-04-23 12:27:27 -06001450 int moved;
Jan Karaf9cae922020-05-29 16:08:58 +02001451 unsigned long time_expire_jif = dirtied_before;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001452
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001453 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001454 list_splice_init(&wb->b_more_io, &wb->b_io);
Jan Kara5fcd5752020-05-29 16:24:43 +02001455 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
Jan Karaf9cae922020-05-29 16:08:58 +02001456 if (!work->for_sync)
1457 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001458 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
Jan Kara5fcd5752020-05-29 16:24:43 +02001459 time_expire_jif);
Tejun Heod6c10f12015-05-22 17:13:45 -04001460 if (moved)
1461 wb_io_lists_populated(wb);
Jan Karaf9cae922020-05-29 16:08:58 +02001462 trace_writeback_queue_io(wb, work, dirtied_before, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001463}
1464
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001465static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001466{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001467 int ret;
1468
1469 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1470 trace_writeback_write_inode_start(inode, wbc);
1471 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1472 trace_writeback_write_inode(inode, wbc);
1473 return ret;
1474 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001475 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -07001476}
1477
1478/*
Jan Kara169ebd92012-05-03 14:48:03 +02001479 * Wait for writeback on an inode to complete. Called with i_lock held.
1480 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +02001481 */
Jan Kara169ebd92012-05-03 14:48:03 +02001482static void __inode_wait_for_writeback(struct inode *inode)
1483 __releases(inode->i_lock)
1484 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +02001485{
1486 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1487 wait_queue_head_t *wqh;
1488
1489 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +11001490 while (inode->i_state & I_SYNC) {
1491 spin_unlock(&inode->i_lock);
NeilBrown74316202014-07-07 15:16:04 +10001492 __wait_on_bit(wqh, &wq, bit_wait,
1493 TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +11001494 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -07001495 }
Christoph Hellwig01c03192009-06-08 13:35:40 +02001496}
1497
1498/*
Jan Kara169ebd92012-05-03 14:48:03 +02001499 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1500 */
1501void inode_wait_for_writeback(struct inode *inode)
1502{
1503 spin_lock(&inode->i_lock);
1504 __inode_wait_for_writeback(inode);
1505 spin_unlock(&inode->i_lock);
1506}
1507
1508/*
1509 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1510 * held and drops it. It is aimed for callers not holding any inode reference
1511 * so once i_lock is dropped, inode can go away.
1512 */
1513static void inode_sleep_on_writeback(struct inode *inode)
1514 __releases(inode->i_lock)
1515{
1516 DEFINE_WAIT(wait);
1517 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1518 int sleep;
1519
1520 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1521 sleep = inode->i_state & I_SYNC;
1522 spin_unlock(&inode->i_lock);
1523 if (sleep)
1524 schedule();
1525 finish_wait(wqh, &wait);
1526}
1527
1528/*
Jan Karaccb26b52012-05-03 14:47:58 +02001529 * Find proper writeback list for the inode depending on its current state and
1530 * possibly also change of its state while we were doing writeback. Here we
1531 * handle things such as livelock prevention or fairness of writeback among
1532 * inodes. This function can be called only by flusher thread - noone else
1533 * processes all inodes in writeback lists and requeueing inodes behind flusher
1534 * thread's back can have unexpected consequences.
1535 */
1536static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1537 struct writeback_control *wbc)
1538{
1539 if (inode->i_state & I_FREEING)
1540 return;
1541
1542 /*
1543 * Sync livelock prevention. Each inode is tagged and synced in one
1544 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1545 * the dirty time to prevent enqueue and sync it again.
1546 */
1547 if ((inode->i_state & I_DIRTY) &&
1548 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1549 inode->dirtied_when = jiffies;
1550
Jan Kara4f8ad652012-05-03 14:48:00 +02001551 if (wbc->pages_skipped) {
1552 /*
1553 * writeback is not making progress due to locked
1554 * buffers. Skip this inode for now.
1555 */
Jan Karab35250c2020-06-10 17:36:03 +02001556 redirty_tail_locked(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001557 return;
1558 }
1559
Jan Karaccb26b52012-05-03 14:47:58 +02001560 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1561 /*
1562 * We didn't write back all the pages. nfs_writepages()
1563 * sometimes bales out without doing anything.
1564 */
1565 if (wbc->nr_to_write <= 0) {
1566 /* Slice used up. Queue for next turn. */
1567 requeue_io(inode, wb);
1568 } else {
1569 /*
1570 * Writeback blocked by something other than
1571 * congestion. Delay the inode for some time to
1572 * avoid spinning on the CPU (100% iowait)
1573 * retrying writeback of the dirty page/inode
1574 * that cannot be performed immediately.
1575 */
Jan Karab35250c2020-06-10 17:36:03 +02001576 redirty_tail_locked(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001577 }
1578 } else if (inode->i_state & I_DIRTY) {
1579 /*
1580 * Filesystems can dirty the inode during writeback operations,
1581 * such as delayed allocation during submission or metadata
1582 * updates after data IO completion.
1583 */
Jan Karab35250c2020-06-10 17:36:03 +02001584 redirty_tail_locked(inode, wb);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001585 } else if (inode->i_state & I_DIRTY_TIME) {
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001586 inode->dirtied_when = jiffies;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001587 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
Jan Kara5afced32020-05-29 15:05:22 +02001588 inode->i_state &= ~I_SYNC_QUEUED;
Jan Karaccb26b52012-05-03 14:47:58 +02001589 } else {
1590 /* The inode is clean. Remove from writeback lists. */
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -07001591 inode_cgwb_move_to_attached(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001592 }
1593}
1594
1595/*
Eric Biggersda0c4c62021-01-12 11:02:51 -08001596 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1597 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1598 *
1599 * This doesn't remove the inode from the writeback list it is on, except
1600 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1601 * expiration. The caller is otherwise responsible for writeback list handling.
1602 *
1603 * The caller is also responsible for setting the I_SYNC flag beforehand and
1604 * calling inode_sync_complete() to clear it afterwards.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 */
1606static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001607__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -06001610 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +02001611 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 int ret;
1613
Jan Kara4f8ad652012-05-03 14:48:00 +02001614 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001616 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1617
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 ret = do_writepages(mapping, wbc);
1619
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001620 /*
1621 * Make sure to wait on the data before writing out the metadata.
1622 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +10001623 * I/O completion. We don't do it for sync(2) writeback because it has a
1624 * separate, external IO completion path and ->sync_fs for guaranteeing
1625 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001626 */
Dave Chinner7747bd42013-07-02 22:38:35 +10001627 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001628 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 if (ret == 0)
1630 ret = err;
1631 }
1632
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001633 /*
Eric Biggers1e249cb2021-01-12 11:02:43 -08001634 * If the inode has dirty timestamps and we need to write them, call
1635 * mark_inode_dirty_sync() to notify the filesystem about it and to
1636 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1637 */
1638 if ((inode->i_state & I_DIRTY_TIME) &&
Eric Biggers83dc8812021-01-12 11:02:50 -08001639 (wbc->sync_mode == WB_SYNC_ALL ||
Eric Biggers1e249cb2021-01-12 11:02:43 -08001640 time_after(jiffies, inode->dirtied_time_when +
1641 dirtytime_expire_interval * HZ))) {
1642 trace_writeback_lazytime(inode);
1643 mark_inode_dirty_sync(inode);
1644 }
1645
1646 /*
Eric Biggersda0c4c62021-01-12 11:02:51 -08001647 * Get and clear the dirty flags from i_state. This needs to be done
1648 * after calling writepages because some filesystems may redirty the
1649 * inode during writepages due to delalloc. It also needs to be done
1650 * after handling timestamp expiration, as that may dirty the inode too.
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001651 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001652 spin_lock(&inode->i_lock);
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001653 dirty = inode->i_state & I_DIRTY;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001654 inode->i_state &= ~dirty;
Tejun Heo9c6ac782014-10-24 15:38:21 -04001655
1656 /*
1657 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1658 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1659 * either they see the I_DIRTY bits cleared or we see the dirtied
1660 * inode.
1661 *
1662 * I_DIRTY_PAGES is always cleared together above even if @mapping
1663 * still has dirty pages. The flag is reinstated after smp_mb() if
1664 * necessary. This guarantees that either __mark_inode_dirty()
1665 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1666 */
1667 smp_mb();
1668
1669 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1670 inode->i_state |= I_DIRTY_PAGES;
1671
Dave Chinner250df6e2011-03-22 22:23:36 +11001672 spin_unlock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001673
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001674 /* Don't write the inode if only I_DIRTY_PAGES was set */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001675 if (dirty & ~I_DIRTY_PAGES) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001676 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 if (ret == 0)
1678 ret = err;
1679 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001680 trace_writeback_single_inode(inode, wbc, nr_to_write);
1681 return ret;
1682}
1683
1684/*
Eric Biggersda0c4c62021-01-12 11:02:51 -08001685 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1686 * the regular batched writeback done by the flusher threads in
1687 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1688 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
Jan Kara4f8ad652012-05-03 14:48:00 +02001689 *
Eric Biggersda0c4c62021-01-12 11:02:51 -08001690 * To prevent the inode from going away, either the caller must have a reference
1691 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
Jan Kara4f8ad652012-05-03 14:48:00 +02001692 */
Tejun Heoaaf25592016-03-18 13:52:04 -04001693static int writeback_single_inode(struct inode *inode,
1694 struct writeback_control *wbc)
Jan Kara4f8ad652012-05-03 14:48:00 +02001695{
Tejun Heoaaf25592016-03-18 13:52:04 -04001696 struct bdi_writeback *wb;
Jan Kara4f8ad652012-05-03 14:48:00 +02001697 int ret = 0;
1698
1699 spin_lock(&inode->i_lock);
1700 if (!atomic_read(&inode->i_count))
1701 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1702 else
1703 WARN_ON(inode->i_state & I_WILL_FREE);
1704
1705 if (inode->i_state & I_SYNC) {
Eric Biggersda0c4c62021-01-12 11:02:51 -08001706 /*
1707 * Writeback is already running on the inode. For WB_SYNC_NONE,
1708 * that's enough and we can just return. For WB_SYNC_ALL, we
1709 * must wait for the existing writeback to complete, then do
1710 * writeback again if there's anything left.
1711 */
Jan Kara4f8ad652012-05-03 14:48:00 +02001712 if (wbc->sync_mode != WB_SYNC_ALL)
1713 goto out;
Jan Kara169ebd92012-05-03 14:48:03 +02001714 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001715 }
1716 WARN_ON(inode->i_state & I_SYNC);
1717 /*
Eric Biggersda0c4c62021-01-12 11:02:51 -08001718 * If the inode is already fully clean, then there's nothing to do.
1719 *
1720 * For data-integrity syncs we also need to check whether any pages are
1721 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1722 * there are any such pages, we'll need to wait for them.
Jan Kara4f8ad652012-05-03 14:48:00 +02001723 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001724 if (!(inode->i_state & I_DIRTY_ALL) &&
Jan Karaf9b0e052013-12-14 04:21:26 +08001725 (wbc->sync_mode != WB_SYNC_ALL ||
1726 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +02001727 goto out;
1728 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001729 wbc_attach_and_unlock_inode(wbc, inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001730
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001731 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
Tejun Heob16b1de2015-06-02 08:39:48 -06001733 wbc_detach_inode(wbc);
Tejun Heoaaf25592016-03-18 13:52:04 -04001734
1735 wb = inode_to_wb_and_lock_list(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11001736 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +02001737 /*
Eric Biggersda0c4c62021-01-12 11:02:51 -08001738 * If the inode is now fully clean, then it can be safely removed from
1739 * its writeback list (if any). Otherwise the flusher threads are
1740 * responsible for the writeback lists.
Jan Kara4f8ad652012-05-03 14:48:00 +02001741 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001742 if (!(inode->i_state & I_DIRTY_ALL))
Roman Gushchinf3b6a6d2021-06-28 19:35:53 -07001743 inode_cgwb_move_to_attached(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001744 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001745 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001746out:
1747 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 return ret;
1749}
1750
Tejun Heoa88a3412015-05-22 17:13:28 -04001751static long writeback_chunk_size(struct bdi_writeback *wb,
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001752 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001753{
1754 long pages;
1755
1756 /*
1757 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1758 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1759 * here avoids calling into writeback_inodes_wb() more than once.
1760 *
1761 * The intended call sequence for WB_SYNC_ALL writeback is:
1762 *
1763 * wb_writeback()
1764 * writeback_sb_inodes() <== called only once
1765 * write_cache_pages() <== called once for each inode
1766 * (quickly) tag currently dirty pages
1767 * (maybe slowly) sync all tagged pages
1768 */
1769 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1770 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001771 else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001772 pages = min(wb->avg_write_bandwidth / 2,
Tejun Heodcc25ae2015-05-22 18:23:22 -04001773 global_wb_domain.dirty_limit / DIRTY_SCOPE);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001774 pages = min(pages, work->nr_pages);
1775 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1776 MIN_WRITEBACK_PAGES);
1777 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001778
1779 return pages;
1780}
1781
Jens Axboe03ba3782009-09-09 09:08:54 +02001782/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001783 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001784 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001785 * Return the number of pages and/or inodes written.
Linus Torvalds0ba13fd2015-09-11 13:26:39 -07001786 *
1787 * NOTE! This is called with wb->list_lock held, and will
1788 * unlock and relock that for each inode it ends up doing
1789 * IO for.
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001790 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001791static long writeback_sb_inodes(struct super_block *sb,
1792 struct bdi_writeback *wb,
1793 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001794{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001795 struct writeback_control wbc = {
1796 .sync_mode = work->sync_mode,
1797 .tagged_writepages = work->tagged_writepages,
1798 .for_kupdate = work->for_kupdate,
1799 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +10001800 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001801 .range_cyclic = work->range_cyclic,
1802 .range_start = 0,
1803 .range_end = LLONG_MAX,
1804 };
1805 unsigned long start_time = jiffies;
1806 long write_chunk;
1807 long wrote = 0; /* count both pages and inodes */
1808
Jens Axboe03ba3782009-09-09 09:08:54 +02001809 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001810 struct inode *inode = wb_inode(wb->b_io.prev);
Tejun Heoaaf25592016-03-18 13:52:04 -04001811 struct bdi_writeback *tmp_wb;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001812
1813 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001814 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001815 /*
1816 * We only want to write back data for this
1817 * superblock, move all inodes not belonging
1818 * to it back onto the dirty list.
1819 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001820 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001821 continue;
1822 }
1823
1824 /*
1825 * The inode belongs to a different superblock.
1826 * Bounce back to the caller to unpin this and
1827 * pin the next superblock.
1828 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001829 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001830 }
1831
Christoph Hellwig9843b76a2010-10-24 19:40:46 +02001832 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +08001833 * Don't bother with new inodes or inodes being freed, first
1834 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b76a2010-10-24 19:40:46 +02001835 * kind writeout is handled by the freer.
1836 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001837 spin_lock(&inode->i_lock);
Christoph Hellwig9843b76a2010-10-24 19:40:46 +02001838 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Jan Karab35250c2020-06-10 17:36:03 +02001839 redirty_tail_locked(inode, wb);
Dave Chinner250df6e2011-03-22 22:23:36 +11001840 spin_unlock(&inode->i_lock);
Nick Piggin7ef0d732009-03-12 14:31:38 -07001841 continue;
1842 }
Jan Karacc1676d2012-05-03 14:47:56 +02001843 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1844 /*
1845 * If this inode is locked for writeback and we are not
1846 * doing writeback-for-data-integrity, move it to
1847 * b_more_io so that writeback can proceed with the
1848 * other inodes on s_io.
1849 *
1850 * We'll have another go at writing back this inode
1851 * when we completed a full scan of b_io.
1852 */
1853 spin_unlock(&inode->i_lock);
1854 requeue_io(inode, wb);
1855 trace_writeback_sb_inodes_requeue(inode);
1856 continue;
1857 }
Jan Karaf0d07b72012-05-03 14:47:59 +02001858 spin_unlock(&wb->list_lock);
1859
Jan Kara4f8ad652012-05-03 14:48:00 +02001860 /*
1861 * We already requeued the inode if it had I_SYNC set and we
1862 * are doing WB_SYNC_NONE writeback. So this catches only the
1863 * WB_SYNC_ALL case.
1864 */
Jan Kara169ebd92012-05-03 14:48:03 +02001865 if (inode->i_state & I_SYNC) {
1866 /* Wait for I_SYNC. This function drops i_lock... */
1867 inode_sleep_on_writeback(inode);
1868 /* Inode may be gone, start again */
Jan Karaead188f92012-06-08 17:07:36 +02001869 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001870 continue;
1871 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001872 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001873 wbc_attach_and_unlock_inode(&wbc, inode);
Jan Kara169ebd92012-05-03 14:48:03 +02001874
Tejun Heoa88a3412015-05-22 17:13:28 -04001875 write_chunk = writeback_chunk_size(wb, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001876 wbc.nr_to_write = write_chunk;
1877 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +11001878
Jan Kara169ebd92012-05-03 14:48:03 +02001879 /*
1880 * We use I_SYNC to pin the inode in memory. While it is set
1881 * evict_inode() will wait so the inode cannot be freed.
1882 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001883 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001884
Tejun Heob16b1de2015-06-02 08:39:48 -06001885 wbc_detach_inode(&wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001886 work->nr_pages -= write_chunk - wbc.nr_to_write;
1887 wrote += write_chunk - wbc.nr_to_write;
Chris Mason590dca32015-09-18 13:35:08 -04001888
1889 if (need_resched()) {
1890 /*
1891 * We're trying to balance between building up a nice
1892 * long list of IOs to improve our merge rate, and
1893 * getting those IOs out quickly for anyone throttling
1894 * in balance_dirty_pages(). cond_resched() doesn't
1895 * unplug, so get our IOs out the door before we
1896 * give up the CPU.
1897 */
Christoph Hellwig008f75a2021-10-20 16:41:19 +02001898 if (current->plug)
1899 blk_flush_plug(current->plug, false);
Chris Mason590dca32015-09-18 13:35:08 -04001900 cond_resched();
1901 }
1902
Tejun Heoaaf25592016-03-18 13:52:04 -04001903 /*
1904 * Requeue @inode if still dirty. Be careful as @inode may
1905 * have been switched to another wb in the meantime.
1906 */
1907 tmp_wb = inode_to_wb_and_lock_list(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001908 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001909 if (!(inode->i_state & I_DIRTY_ALL))
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001910 wrote++;
Tejun Heoaaf25592016-03-18 13:52:04 -04001911 requeue_inode(inode, tmp_wb, &wbc);
Jan Kara4f8ad652012-05-03 14:48:00 +02001912 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001913 spin_unlock(&inode->i_lock);
Chris Mason590dca32015-09-18 13:35:08 -04001914
Tejun Heoaaf25592016-03-18 13:52:04 -04001915 if (unlikely(tmp_wb != wb)) {
1916 spin_unlock(&tmp_wb->list_lock);
1917 spin_lock(&wb->list_lock);
1918 }
1919
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001920 /*
1921 * bail out to wb_writeback() often enough to check
1922 * background threshold and other termination conditions.
1923 */
1924 if (wrote) {
1925 if (time_is_before_jiffies(start_time + HZ / 10UL))
1926 break;
1927 if (work->nr_pages <= 0)
1928 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -08001929 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001931 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001932}
Nick Piggin38f21972009-01-06 14:40:25 -08001933
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001934static long __writeback_inodes_wb(struct bdi_writeback *wb,
1935 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001936{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001937 unsigned long start_time = jiffies;
1938 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001939
1940 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001941 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001942 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001943
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001944 if (!trylock_super(sb)) {
Wu Fengguang0e995812011-07-29 22:14:35 -06001945 /*
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001946 * trylock_super() may fail consistently due to
Wu Fengguang0e995812011-07-29 22:14:35 -06001947 * s_umount being grabbed by someone else. Don't use
1948 * requeue_io() to avoid busy retrying the inode/sb.
1949 */
1950 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001951 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001952 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001953 wrote += writeback_sb_inodes(sb, wb, work);
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001954 up_read(&sb->s_umount);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001955
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001956 /* refer to the same tests at the end of writeback_sb_inodes */
1957 if (wrote) {
1958 if (time_is_before_jiffies(start_time + HZ / 10UL))
1959 break;
1960 if (work->nr_pages <= 0)
1961 break;
1962 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001963 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001964 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001965 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001966}
1967
Wanpeng Li7d9f0732013-09-11 14:22:40 -07001968static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001969 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001970{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001971 struct wb_writeback_work work = {
1972 .nr_pages = nr_pages,
1973 .sync_mode = WB_SYNC_NONE,
1974 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001975 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001976 };
Linus Torvalds505a6662015-09-11 13:37:19 -07001977 struct blk_plug plug;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001978
Linus Torvalds505a6662015-09-11 13:37:19 -07001979 blk_start_plug(&plug);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001980 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -06001981 if (list_empty(&wb->b_io))
Jan Karaf9cae922020-05-29 16:08:58 +02001982 queue_io(wb, &work, jiffies);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001983 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001984 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07001985 blk_finish_plug(&plug);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001986
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001987 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001988}
1989
Jens Axboe03ba3782009-09-09 09:08:54 +02001990/*
1991 * Explicit flushing or periodic writeback of "old" data.
1992 *
1993 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1994 * dirtying-time in the inode's address_space. So this periodic writeback code
1995 * just walks the superblock inode list, writing back any inodes which are
1996 * older than a specific point in time.
1997 *
1998 * Try to run once per dirty_writeback_interval. But if a writeback event
1999 * takes longer than a dirty_writeback_interval interval, then leave a
2000 * one-second gap.
2001 *
Jan Karaf9cae922020-05-29 16:08:58 +02002002 * dirtied_before takes precedence over nr_to_write. So we'll only write back
Jens Axboe03ba3782009-09-09 09:08:54 +02002003 * all dirty pages if they are all attached to "old" mappings.
2004 */
Jens Axboec4a77a62009-09-16 15:18:25 +02002005static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002006 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02002007{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002008 long nr_pages = work->nr_pages;
Jan Karaf9cae922020-05-29 16:08:58 +02002009 unsigned long dirtied_before = jiffies;
Jan Karaa5989bd2009-09-16 19:22:48 +02002010 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002011 long progress;
Linus Torvalds505a6662015-09-11 13:37:19 -07002012 struct blk_plug plug;
Jens Axboe03ba3782009-09-09 09:08:54 +02002013
Linus Torvalds505a6662015-09-11 13:37:19 -07002014 blk_start_plug(&plug);
Wu Fengguange8dfc302011-04-21 12:06:32 -06002015 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002016 for (;;) {
2017 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08002018 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +02002019 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002020 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +02002021 break;
2022
2023 /*
Jan Karaaa373cf2011-01-13 15:45:47 -08002024 * Background writeout and kupdate-style writeback may
2025 * run forever. Stop them if there is other work to do
2026 * so that e.g. sync can proceed. They'll be restarted
2027 * after the other works are all done.
2028 */
2029 if ((work->for_background || work->for_kupdate) &&
Tejun Heof0054bb2015-05-22 17:13:30 -04002030 !list_empty(&wb->work_list))
Jan Karaaa373cf2011-01-13 15:45:47 -08002031 break;
2032
2033 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08002034 * For background writeout, stop when we are below the
2035 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +02002036 */
Tejun Heoaa661bb2015-05-22 18:23:31 -04002037 if (work->for_background && !wb_over_bg_thresh(wb))
Jens Axboe03ba3782009-09-09 09:08:54 +02002038 break;
2039
Jan Kara1bc36b62011-10-19 11:44:41 +02002040 /*
2041 * Kupdate and background works are special and we want to
2042 * include all inodes that need writing. Livelock avoidance is
2043 * handled by these works yielding to any other work so we are
2044 * safe.
2045 */
Wu Fengguangba9aa8392010-07-21 20:32:30 -06002046 if (work->for_kupdate) {
Jan Karaf9cae922020-05-29 16:08:58 +02002047 dirtied_before = jiffies -
Wu Fengguangba9aa8392010-07-21 20:32:30 -06002048 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +02002049 } else if (work->for_background)
Jan Karaf9cae922020-05-29 16:08:58 +02002050 dirtied_before = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +10002051
Tejun Heo5634cc22015-08-18 14:54:56 -07002052 trace_writeback_start(wb, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -06002053 if (list_empty(&wb->b_io))
Jan Karaf9cae922020-05-29 16:08:58 +02002054 queue_io(wb, work, dirtied_before);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002055 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002056 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02002057 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002058 progress = __writeback_inodes_wb(wb, work);
Tejun Heo5634cc22015-08-18 14:54:56 -07002059 trace_writeback_written(wb, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +10002060
Jens Axboe03ba3782009-09-09 09:08:54 +02002061 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +02002062 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -06002063 *
2064 * Dirty inodes are moved to b_io for writeback in batches.
2065 * The completion of the current batch does not necessarily
2066 * mean the overall work is done. So we keep looping as long
2067 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +02002068 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002069 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +02002070 continue;
2071 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -06002072 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +02002073 */
Wu Fengguangb7a24412010-07-21 22:19:51 -06002074 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +02002075 break;
2076 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +02002077 * Nothing written. Wait for some inode to
2078 * become available for writeback. Otherwise
2079 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +02002080 */
Tahsin Erdoganbace9242016-12-12 16:43:20 -08002081 trace_writeback_wait(wb, work);
2082 inode = wb_inode(wb->b_more_io.prev);
2083 spin_lock(&inode->i_lock);
2084 spin_unlock(&wb->list_lock);
2085 /* This function drops i_lock... */
2086 inode_sleep_on_writeback(inode);
2087 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002088 }
Wu Fengguange8dfc302011-04-21 12:06:32 -06002089 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07002090 blk_finish_plug(&plug);
Jens Axboe03ba3782009-09-09 09:08:54 +02002091
Wu Fengguangd46db3d2011-05-04 19:54:37 -06002092 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +02002093}
2094
2095/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002096 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +02002097 */
Tejun Heof0054bb2015-05-22 17:13:30 -04002098static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02002099{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002100 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02002101
Tejun Heof0054bb2015-05-22 17:13:30 -04002102 spin_lock_bh(&wb->work_lock);
2103 if (!list_empty(&wb->work_list)) {
2104 work = list_entry(wb->work_list.next,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002105 struct wb_writeback_work, list);
2106 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +02002107 }
Tejun Heof0054bb2015-05-22 17:13:30 -04002108 spin_unlock_bh(&wb->work_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002109 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +02002110}
2111
Jan Kara65850272011-01-13 15:45:44 -08002112static long wb_check_background_flush(struct bdi_writeback *wb)
2113{
Tejun Heoaa661bb2015-05-22 18:23:31 -04002114 if (wb_over_bg_thresh(wb)) {
Jan Kara65850272011-01-13 15:45:44 -08002115
2116 struct wb_writeback_work work = {
2117 .nr_pages = LONG_MAX,
2118 .sync_mode = WB_SYNC_NONE,
2119 .for_background = 1,
2120 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002121 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -08002122 };
2123
2124 return wb_writeback(wb, &work);
2125 }
2126
2127 return 0;
2128}
2129
Jens Axboe03ba3782009-09-09 09:08:54 +02002130static long wb_check_old_data_flush(struct bdi_writeback *wb)
2131{
2132 unsigned long expired;
2133 long nr_pages;
2134
Jens Axboe69b62d02010-05-17 12:51:03 +02002135 /*
2136 * When set to zero, disable periodic writeback
2137 */
2138 if (!dirty_writeback_interval)
2139 return 0;
2140
Jens Axboe03ba3782009-09-09 09:08:54 +02002141 expired = wb->last_old_flush +
2142 msecs_to_jiffies(dirty_writeback_interval * 10);
2143 if (time_before(jiffies, expired))
2144 return 0;
2145
2146 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07002147 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +02002148
Jens Axboec4a77a62009-09-16 15:18:25 +02002149 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002150 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +02002151 .nr_pages = nr_pages,
2152 .sync_mode = WB_SYNC_NONE,
2153 .for_kupdate = 1,
2154 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002155 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +02002156 };
2157
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002158 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +02002159 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002160
2161 return 0;
2162}
2163
Jens Axboe85009b42017-09-30 02:09:06 -06002164static long wb_check_start_all(struct bdi_writeback *wb)
2165{
2166 long nr_pages;
2167
2168 if (!test_bit(WB_start_all, &wb->state))
2169 return 0;
2170
2171 nr_pages = get_nr_dirty_pages();
2172 if (nr_pages) {
2173 struct wb_writeback_work work = {
2174 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2175 .sync_mode = WB_SYNC_NONE,
2176 .range_cyclic = 1,
2177 .reason = wb->start_all_reason,
2178 };
2179
2180 nr_pages = wb_writeback(wb, &work);
2181 }
2182
2183 clear_bit(WB_start_all, &wb->state);
2184 return nr_pages;
2185}
2186
2187
Jens Axboe03ba3782009-09-09 09:08:54 +02002188/*
2189 * Retrieve work items and do the writeback they describe
2190 */
Wanpeng Li25d130b2013-07-08 16:00:14 -07002191static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02002192{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002193 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +02002194 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +02002195
Tejun Heo44522262015-05-22 17:13:26 -04002196 set_bit(WB_writeback_running, &wb->state);
Tejun Heof0054bb2015-05-22 17:13:30 -04002197 while ((work = get_next_work_item(wb)) != NULL) {
Tejun Heo5634cc22015-08-18 14:54:56 -07002198 trace_writeback_exec(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002199 wrote += wb_writeback(wb, work);
Tahsin Erdogan4a3a4852017-03-10 12:09:49 -08002200 finish_writeback_work(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02002201 }
2202
2203 /*
Jens Axboe85009b42017-09-30 02:09:06 -06002204 * Check for a flush-everything request
2205 */
2206 wrote += wb_check_start_all(wb);
2207
2208 /*
Jens Axboe03ba3782009-09-09 09:08:54 +02002209 * Check for periodic writeback, kupdated() style
2210 */
2211 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08002212 wrote += wb_check_background_flush(wb);
Tejun Heo44522262015-05-22 17:13:26 -04002213 clear_bit(WB_writeback_running, &wb->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02002214
2215 return wrote;
2216}
2217
2218/*
2219 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07002220 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02002221 */
Tejun Heof0054bb2015-05-22 17:13:30 -04002222void wb_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02002223{
Tejun Heo839a8e82013-04-01 19:08:06 -07002224 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2225 struct bdi_writeback, dwork);
Jens Axboe03ba3782009-09-09 09:08:54 +02002226 long pages_written;
2227
Theodore Ts'o68f23b892020-01-30 22:11:04 -08002228 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
Peter Zijlstra766f9162010-10-26 14:22:45 -07002229 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02002230
Tejun Heo839a8e82013-04-01 19:08:06 -07002231 if (likely(!current_is_workqueue_rescuer() ||
Tejun Heo44522262015-05-22 17:13:26 -04002232 !test_bit(WB_registered, &wb->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03002233 /*
Tejun Heof0054bb2015-05-22 17:13:30 -04002234 * The normal path. Keep writing back @wb until its
Tejun Heo839a8e82013-04-01 19:08:06 -07002235 * work_list is empty. Note that this path is also taken
Tejun Heof0054bb2015-05-22 17:13:30 -04002236 * if @wb is shutting down even when we're running off the
Tejun Heo839a8e82013-04-01 19:08:06 -07002237 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03002238 */
Tejun Heo839a8e82013-04-01 19:08:06 -07002239 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07002240 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07002241 trace_writeback_pages_written(pages_written);
Tejun Heof0054bb2015-05-22 17:13:30 -04002242 } while (!list_empty(&wb->work_list));
Tejun Heo839a8e82013-04-01 19:08:06 -07002243 } else {
2244 /*
2245 * bdi_wq can't get enough workers and we're running off
2246 * the emergency worker. Don't hog it. Hopefully, 1024 is
2247 * enough for efficient IO.
2248 */
Tejun Heof0054bb2015-05-22 17:13:30 -04002249 pages_written = writeback_inodes_wb(wb, 1024,
Tejun Heo839a8e82013-04-01 19:08:06 -07002250 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10002251 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02002252 }
2253
Tejun Heof0054bb2015-05-22 17:13:30 -04002254 if (!list_empty(&wb->work_list))
Jan Karab8b78492018-05-03 18:26:26 +02002255 wb_wakeup(wb);
Derek Basehore6ca738d2014-04-03 14:46:22 -07002256 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Tejun Heof0054bb2015-05-22 17:13:30 -04002257 wb_wakeup_delayed(wb);
Dave Chinner455b2862010-07-07 13:24:06 +10002258
Tejun Heo839a8e82013-04-01 19:08:06 -07002259 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02002260}
2261
2262/*
Jens Axboe595043e2017-09-28 11:26:59 -06002263 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2264 * write back the whole world.
2265 */
2266static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
Jens Axboee8e8a0c2017-09-28 11:31:22 -06002267 enum wb_reason reason)
Jens Axboe595043e2017-09-28 11:26:59 -06002268{
2269 struct bdi_writeback *wb;
2270
2271 if (!bdi_has_dirty_io(bdi))
2272 return;
2273
2274 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Jens Axboee8e8a0c2017-09-28 11:31:22 -06002275 wb_start_writeback(wb, reason);
Jens Axboe595043e2017-09-28 11:26:59 -06002276}
2277
2278void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2279 enum wb_reason reason)
2280{
Jens Axboe595043e2017-09-28 11:26:59 -06002281 rcu_read_lock();
Jens Axboee8e8a0c2017-09-28 11:31:22 -06002282 __wakeup_flusher_threads_bdi(bdi, reason);
Jens Axboe595043e2017-09-28 11:26:59 -06002283 rcu_read_unlock();
2284}
2285
2286/*
Jens Axboe9ba4b2d2017-09-20 08:58:25 -06002287 * Wakeup the flusher threads to start writeback of all currently dirty pages
Jens Axboe03ba3782009-09-09 09:08:54 +02002288 */
Jens Axboe9ba4b2d2017-09-20 08:58:25 -06002289void wakeup_flusher_threads(enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02002290{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02002291 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02002292
Konstantin Khlebnikov51350ea2016-08-04 21:36:05 +03002293 /*
2294 * If we are expecting writeback progress we must submit plugged IO.
2295 */
2296 if (blk_needs_flush_plug(current))
Christoph Hellwig008f75a2021-10-20 16:41:19 +02002297 blk_flush_plug(current->plug, true);
Konstantin Khlebnikov51350ea2016-08-04 21:36:05 +03002298
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02002299 rcu_read_lock();
Jens Axboe595043e2017-09-28 11:26:59 -06002300 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
Jens Axboee8e8a0c2017-09-28 11:31:22 -06002301 __wakeup_flusher_threads_bdi(bdi, reason);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02002302 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02002303}
2304
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002305/*
2306 * Wake up bdi's periodically to make sure dirtytime inodes gets
2307 * written back periodically. We deliberately do *not* check the
2308 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2309 * kernel to be constantly waking up once there are any dirtytime
2310 * inodes on the system. So instead we define a separate delayed work
2311 * function which gets called much more rarely. (By default, only
2312 * once every 12 hours.)
2313 *
2314 * If there is any other write activity going on in the file system,
2315 * this function won't be necessary. But if the only thing that has
2316 * happened on the file system is a dirtytime inode caused by an atime
2317 * update, we need this infrastructure below to make sure that inode
2318 * eventually gets pushed out to disk.
2319 */
2320static void wakeup_dirtytime_writeback(struct work_struct *w);
2321static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2322
2323static void wakeup_dirtytime_writeback(struct work_struct *w)
2324{
2325 struct backing_dev_info *bdi;
2326
2327 rcu_read_lock();
2328 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Tejun Heo001fe6f2015-05-22 17:13:56 -04002329 struct bdi_writeback *wb;
Tejun Heo001fe6f2015-05-22 17:13:56 -04002330
Tejun Heob8175252015-10-02 14:47:05 -04002331 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Tejun Heo6fdf8602015-09-29 12:47:51 -04002332 if (!list_empty(&wb->b_dirty_time))
2333 wb_wakeup(wb);
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002334 }
2335 rcu_read_unlock();
2336 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2337}
2338
2339static int __init start_dirtytime_writeback(void)
2340{
2341 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2342 return 0;
2343}
2344__initcall(start_dirtytime_writeback);
2345
Theodore Ts'o1efff912015-03-17 12:23:32 -04002346int dirtytime_interval_handler(struct ctl_table *table, int write,
Tobias Klauser9ca48e22020-09-18 21:20:39 -07002347 void *buffer, size_t *lenp, loff_t *ppos)
Theodore Ts'o1efff912015-03-17 12:23:32 -04002348{
2349 int ret;
2350
2351 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2352 if (ret == 0 && write)
2353 mod_delayed_work(system_wq, &dirtytime_work, 0);
2354 return ret;
2355}
2356
Jens Axboe03ba3782009-09-09 09:08:54 +02002357/**
Eric Biggers35d14f22021-01-12 11:02:49 -08002358 * __mark_inode_dirty - internal function to mark an inode dirty
Mauro Carvalho Chehab0117d422017-05-12 07:45:42 -03002359 *
2360 * @inode: inode to mark
Eric Biggers35d14f22021-01-12 11:02:49 -08002361 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2362 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2363 * with I_DIRTY_PAGES.
Mauro Carvalho Chehab0117d422017-05-12 07:45:42 -03002364 *
Eric Biggers35d14f22021-01-12 11:02:49 -08002365 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2366 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
Jens Axboe03ba3782009-09-09 09:08:54 +02002367 *
Eric Biggers35d14f22021-01-12 11:02:49 -08002368 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2369 * instead of calling this directly.
Jens Axboe03ba3782009-09-09 09:08:54 +02002370 *
Eric Biggers35d14f22021-01-12 11:02:49 -08002371 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2372 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2373 * even if they are later hashed, as they will have been marked dirty already.
Jens Axboe03ba3782009-09-09 09:08:54 +02002374 *
Eric Biggers35d14f22021-01-12 11:02:49 -08002375 * In short, ensure you hash any inodes _before_ you start marking them dirty.
Jens Axboe03ba3782009-09-09 09:08:54 +02002376 *
Jens Axboe03ba3782009-09-09 09:08:54 +02002377 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2378 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2379 * the kernel-internal blockdev inode represents the dirtying time of the
2380 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2381 * page->mapping->host, so the page-dirtying time is recorded in the internal
2382 * blockdev inode.
2383 */
2384void __mark_inode_dirty(struct inode *inode, int flags)
2385{
2386 struct super_block *sb = inode->i_sb;
Eric Biggers35d14f22021-01-12 11:02:49 -08002387 int dirtytime = 0;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002388
2389 trace_writeback_mark_inode_dirty(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002390
Eric Biggerse2728c52021-01-12 11:02:47 -08002391 if (flags & I_DIRTY_INODE) {
Eric Biggers35d14f22021-01-12 11:02:49 -08002392 /*
2393 * Notify the filesystem about the inode being dirtied, so that
2394 * (if needed) it can update on-disk fields and journal the
2395 * inode. This is only needed when the inode itself is being
2396 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2397 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2398 */
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002399 trace_writeback_dirty_inode_start(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002400 if (sb->s_op->dirty_inode)
Eric Biggersa38ed482021-01-12 11:02:48 -08002401 sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002402 trace_writeback_dirty_inode(inode, flags);
Eric Biggerse2728c52021-01-12 11:02:47 -08002403
Eric Biggers35d14f22021-01-12 11:02:49 -08002404 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002405 flags &= ~I_DIRTY_TIME;
Eric Biggers35d14f22021-01-12 11:02:49 -08002406 } else {
2407 /*
2408 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2409 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2410 * in one call to __mark_inode_dirty().)
2411 */
2412 dirtytime = flags & I_DIRTY_TIME;
2413 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
Eric Biggerse2728c52021-01-12 11:02:47 -08002414 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002415
2416 /*
Tejun Heo9c6ac782014-10-24 15:38:21 -04002417 * Paired with smp_mb() in __writeback_single_inode() for the
2418 * following lockless i_state test. See there for details.
Jens Axboe03ba3782009-09-09 09:08:54 +02002419 */
2420 smp_mb();
2421
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002422 if (((inode->i_state & flags) == flags) ||
2423 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
Jens Axboe03ba3782009-09-09 09:08:54 +02002424 return;
2425
Dave Chinner250df6e2011-03-22 22:23:36 +11002426 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002427 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2428 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002429 if ((inode->i_state & flags) != flags) {
2430 const int was_dirty = inode->i_state & I_DIRTY;
2431
Tejun Heo52ebea72015-05-22 17:13:37 -04002432 inode_attach_wb(inode, NULL);
2433
Eric Biggers35d14f22021-01-12 11:02:49 -08002434 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002435 if (flags & I_DIRTY_INODE)
2436 inode->i_state &= ~I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002437 inode->i_state |= flags;
2438
2439 /*
Jan Kara5afced32020-05-29 15:05:22 +02002440 * If the inode is queued for writeback by flush worker, just
2441 * update its dirty state. Once the flush worker is done with
2442 * the inode it will place it on the appropriate superblock
2443 * list, based upon its state.
Jens Axboe03ba3782009-09-09 09:08:54 +02002444 */
Jan Kara5afced32020-05-29 15:05:22 +02002445 if (inode->i_state & I_SYNC_QUEUED)
Dave Chinner250df6e2011-03-22 22:23:36 +11002446 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002447
2448 /*
2449 * Only add valid (hashed) inodes to the superblock's
2450 * dirty list. Add blockdev inodes as well.
2451 */
2452 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -04002453 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11002454 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002455 }
Al Viroa4ffdde2010-06-02 17:38:30 -04002456 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11002457 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002458
2459 /*
2460 * If the inode was already on b_dirty/b_io/b_more_io, don't
2461 * reposition it (that would break b_dirty time-ordering).
2462 */
2463 if (!was_dirty) {
Tejun Heo87e1d782015-05-28 14:50:52 -04002464 struct bdi_writeback *wb;
Tejun Heod6c10f12015-05-22 17:13:45 -04002465 struct list_head *dirty_list;
Dave Chinnera66979a2011-03-22 22:23:41 +11002466 bool wakeup_bdi = false;
Jens Axboe500b0672009-09-09 09:10:25 +02002467
Tejun Heo87e1d782015-05-28 14:50:52 -04002468 wb = locked_inode_to_wb_and_lock_list(inode);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002469
Jens Axboe03ba3782009-09-09 09:08:54 +02002470 inode->dirtied_when = jiffies;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002471 if (dirtytime)
2472 inode->dirtied_time_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04002473
Christoph Hellwig0e11f642018-02-21 07:54:49 -08002474 if (inode->i_state & I_DIRTY)
Tejun Heo07472592015-05-22 17:14:02 -04002475 dirty_list = &wb->b_dirty;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002476 else
Tejun Heo07472592015-05-22 17:14:02 -04002477 dirty_list = &wb->b_dirty_time;
Tejun Heod6c10f12015-05-22 17:13:45 -04002478
Dave Chinnerc7f54082015-03-04 14:07:22 -05002479 wakeup_bdi = inode_io_list_move_locked(inode, wb,
Tejun Heod6c10f12015-05-22 17:13:45 -04002480 dirty_list);
2481
Tejun Heo07472592015-05-22 17:14:02 -04002482 spin_unlock(&wb->list_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002483 trace_writeback_dirty_inode_enqueue(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +11002484
Tejun Heod6c10f12015-05-22 17:13:45 -04002485 /*
2486 * If this is the first dirty inode for this bdi,
2487 * we have to wake-up the corresponding bdi thread
2488 * to make sure background write-back happens
2489 * later.
2490 */
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002491 if (wakeup_bdi &&
2492 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
Tejun Heo07472592015-05-22 17:14:02 -04002493 wb_wakeup_delayed(wb);
Dave Chinnera66979a2011-03-22 22:23:41 +11002494 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02002495 }
2496 }
Dave Chinner250df6e2011-03-22 22:23:36 +11002497out_unlock_inode:
2498 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002499}
2500EXPORT_SYMBOL(__mark_inode_dirty);
2501
Dave Chinnere97fedb2015-03-04 13:40:00 -05002502/*
2503 * The @s_sync_lock is used to serialise concurrent sync operations
2504 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2505 * Concurrent callers will block on the s_sync_lock rather than doing contending
2506 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2507 * has been issued up to the time this function is enter is guaranteed to be
2508 * completed by the time we have gained the lock and waited for all IO that is
2509 * in progress regardless of the order callers are granted the lock.
2510 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002511static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002512{
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002513 LIST_HEAD(sync_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002514
Jens Axboe03ba3782009-09-09 09:08:54 +02002515 /*
2516 * We need to be protected against the filesystem going from
2517 * r/o to r/w or vice versa.
2518 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002519 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002520
Dave Chinnere97fedb2015-03-04 13:40:00 -05002521 mutex_lock(&sb->s_sync_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002522
2523 /*
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002524 * Splice the writeback list onto a temporary list to avoid waiting on
2525 * inodes that have started writeback after this point.
2526 *
2527 * Use rcu_read_lock() to keep the inodes around until we have a
2528 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2529 * the local list because inodes can be dropped from either by writeback
2530 * completion.
Jens Axboe03ba3782009-09-09 09:08:54 +02002531 */
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002532 rcu_read_lock();
2533 spin_lock_irq(&sb->s_inode_wblist_lock);
2534 list_splice_init(&sb->s_inodes_wb, &sync_list);
2535
2536 /*
2537 * Data integrity sync. Must wait for all pages under writeback, because
2538 * there may have been pages dirtied before our sync call, but which had
2539 * writeout started before we write it out. In which case, the inode
2540 * may not be on the dirty list, but we still have to wait for that
2541 * writeout.
2542 */
2543 while (!list_empty(&sync_list)) {
2544 struct inode *inode = list_first_entry(&sync_list, struct inode,
2545 i_wb_list);
Dave Chinner250df6e2011-03-22 22:23:36 +11002546 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02002547
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002548 /*
2549 * Move each inode back to the wb list before we drop the lock
2550 * to preserve consistency between i_wb_list and the mapping
2551 * writeback tag. Writeback completion is responsible to remove
2552 * the inode from either list once the writeback tag is cleared.
2553 */
2554 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2555
2556 /*
2557 * The mapping can appear untagged while still on-list since we
2558 * do not have the mapping lock. Skip it here, wb completion
2559 * will remove it.
2560 */
2561 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2562 continue;
2563
2564 spin_unlock_irq(&sb->s_inode_wblist_lock);
2565
Dave Chinner250df6e2011-03-22 22:23:36 +11002566 spin_lock(&inode->i_lock);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002567 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
Dave Chinner250df6e2011-03-22 22:23:36 +11002568 spin_unlock(&inode->i_lock);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002569
2570 spin_lock_irq(&sb->s_inode_wblist_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002571 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11002572 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002573 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11002574 spin_unlock(&inode->i_lock);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002575 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02002576
Junichi Nomuraaa750fd2015-11-05 18:47:23 -08002577 /*
2578 * We keep the error status of individual mapping so that
2579 * applications can catch the writeback error using fsync(2).
2580 * See filemap_fdatawait_keep_errors() for details.
2581 */
2582 filemap_fdatawait_keep_errors(mapping);
Jens Axboe03ba3782009-09-09 09:08:54 +02002583
2584 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08002585
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002586 iput(inode);
2587
2588 rcu_read_lock();
2589 spin_lock_irq(&sb->s_inode_wblist_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002590 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002591 spin_unlock_irq(&sb->s_inode_wblist_lock);
2592 rcu_read_unlock();
Dave Chinnere97fedb2015-03-04 13:40:00 -05002593 mutex_unlock(&sb->s_sync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594}
2595
Tejun Heof30a7d02015-05-22 17:14:00 -04002596static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2597 enum wb_reason reason, bool skip_if_busy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598{
Tejun Heo5b9cce42019-08-26 09:06:52 -07002599 struct backing_dev_info *bdi = sb->s_bdi;
2600 DEFINE_WB_COMPLETION(done, bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002601 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002602 .sb = sb,
2603 .sync_mode = WB_SYNC_NONE,
2604 .tagged_writepages = 1,
2605 .done = &done,
2606 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002607 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002608 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02002609
Tejun Heoe7972912015-05-22 17:13:48 -04002610 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002611 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002612 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Tejun Heof30a7d02015-05-22 17:14:00 -04002613
Tejun Heodb125362015-05-22 17:14:01 -04002614 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
Tejun Heo5b9cce42019-08-26 09:06:52 -07002615 wb_wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616}
Tejun Heof30a7d02015-05-22 17:14:00 -04002617
2618/**
2619 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2620 * @sb: the superblock
2621 * @nr: the number of pages to write
2622 * @reason: reason why some writeback work initiated
2623 *
2624 * Start writeback on some inodes on this super_block. No guarantees are made
2625 * on how many (if any) will be written, and this function does not wait
2626 * for IO completion of submitted IO.
2627 */
2628void writeback_inodes_sb_nr(struct super_block *sb,
2629 unsigned long nr,
2630 enum wb_reason reason)
2631{
2632 __writeback_inodes_sb_nr(sb, nr, reason, false);
2633}
Chris Mason3259f8b2010-10-29 11:16:17 -04002634EXPORT_SYMBOL(writeback_inodes_sb_nr);
2635
2636/**
2637 * writeback_inodes_sb - writeback dirty inodes from given super_block
2638 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08002639 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04002640 *
2641 * Start writeback on some inodes on this super_block. No guarantees are made
2642 * on how many (if any) will be written, and this function does not wait
2643 * for IO completion of submitted IO.
2644 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002645void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002646{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002647 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04002648}
Jens Axboed8a85592009-09-02 12:34:32 +02002649EXPORT_SYMBOL(writeback_inodes_sb);
2650
2651/**
Miao Xie10ee27a2013-01-10 13:47:57 +08002652 * try_to_writeback_inodes_sb - try to start writeback if none underway
2653 * @sb: the superblock
2654 * @reason: reason why some writeback work was initiated
2655 *
Rakesh Pandit8264c322017-10-09 13:34:41 +03002656 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
Miao Xie10ee27a2013-01-10 13:47:57 +08002657 */
Rakesh Pandit8264c322017-10-09 13:34:41 +03002658void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Miao Xie10ee27a2013-01-10 13:47:57 +08002659{
Rakesh Pandit8264c322017-10-09 13:34:41 +03002660 if (!down_read_trylock(&sb->s_umount))
2661 return;
2662
2663 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2664 up_read(&sb->s_umount);
Miao Xie10ee27a2013-01-10 13:47:57 +08002665}
2666EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04002667
2668/**
Jens Axboed8a85592009-09-02 12:34:32 +02002669 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01002670 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02002671 *
2672 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01002673 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02002674 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01002675void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02002676{
Tejun Heo5b9cce42019-08-26 09:06:52 -07002677 struct backing_dev_info *bdi = sb->s_bdi;
2678 DEFINE_WB_COMPLETION(done, bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002679 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002680 .sb = sb,
2681 .sync_mode = WB_SYNC_ALL,
2682 .nr_pages = LONG_MAX,
2683 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002684 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002685 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10002686 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002687 };
2688
Tejun Heo006a0972015-08-25 14:11:52 -04002689 /*
2690 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2691 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2692 * bdi_has_dirty() need to be written out too.
2693 */
2694 if (bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002695 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002696 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2697
Tejun Heo7fc58542017-12-12 08:38:30 -08002698 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2699 bdi_down_write_wb_switch_rwsem(bdi);
Tejun Heodb125362015-05-22 17:14:01 -04002700 bdi_split_work_to_wbs(bdi, &work, false);
Tejun Heo5b9cce42019-08-26 09:06:52 -07002701 wb_wait_for_completion(&done);
Tejun Heo7fc58542017-12-12 08:38:30 -08002702 bdi_up_write_wb_switch_rwsem(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002703
Jens Axboeb6e51312009-09-16 15:13:54 +02002704 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02002705}
2706EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002709 * write_inode_now - write an inode to disk
2710 * @inode: inode to write to disk
2711 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002713 * This function commits an inode to disk immediately if it is dirty. This is
2714 * primarily needed by knfsd.
2715 *
2716 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718int write_inode_now(struct inode *inode, int sync)
2719{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 struct writeback_control wbc = {
2721 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08002722 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07002723 .range_start = 0,
2724 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 };
2726
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002727 if (!mapping_can_writeback(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08002728 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729
2730 might_sleep();
Tejun Heoaaf25592016-03-18 13:52:04 -04002731 return writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733EXPORT_SYMBOL(write_inode_now);
2734
2735/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002736 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02002737 * @inode: the inode to sync
2738 * @wait: wait for I/O to complete.
2739 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002740 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02002741 *
2742 * Note: only writes the actual inode, no associated data or other metadata.
2743 */
2744int sync_inode_metadata(struct inode *inode, int wait)
2745{
2746 struct writeback_control wbc = {
2747 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2748 .nr_to_write = 0, /* metadata-only */
2749 };
2750
Josef Bacik5662c962021-07-14 14:47:25 -04002751 return writeback_single_inode(inode, &wbc);
Christoph Hellwigc37650162010-10-06 10:48:20 +02002752}
2753EXPORT_SYMBOL(sync_inode_metadata);