blob: be568b7311d6f5236099dc23bc84d3ecab750801 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050017#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Wu Fengguangbc31b862012-01-07 20:41:55 -060023#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020024#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
Dave Chinner455b2862010-07-07 13:24:06 +100028#include <linux/tracepoint.h>
Al Viro719ea2f2013-09-29 11:24:49 -040029#include <linux/device.h>
David Howells07f3f052006-09-30 20:52:18 +020030#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
Jens Axboed0bceac2009-05-18 08:20:32 +020032/*
Wu Fengguangbc31b862012-01-07 20:41:55 -060033 * 4MB minimal write chunk size
34 */
35#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
36
37/*
Jens Axboec4a77a62009-09-16 15:18:25 +020038 * Passed into wb_writeback(), essentially a subset of writeback_control
39 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020040struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020041 long nr_pages;
42 struct super_block *sb;
Jan Kara0dc83bd2014-02-21 11:19:04 +010043 unsigned long *older_than_this;
Jens Axboec4a77a62009-09-16 15:18:25 +020044 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060045 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050046 unsigned int for_kupdate:1;
47 unsigned int range_cyclic:1;
48 unsigned int for_background:1;
Dave Chinner7747bd42013-07-02 22:38:35 +100049 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060050 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020051
Jens Axboe8010c3b2009-09-15 20:04:57 +020052 struct list_head list; /* pending work list */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020053 struct completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020054};
55
Adrian Bunkf11b00f2008-04-29 00:58:56 -070056/**
57 * writeback_in_progress - determine whether there is writeback in progress
58 * @bdi: the device's backing_dev_info structure.
59 *
Jens Axboe03ba3782009-09-09 09:08:54 +020060 * Determine whether there is writeback waiting to be handled against a
61 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070062 */
63int writeback_in_progress(struct backing_dev_info *bdi)
64{
Jan Kara81d73a32010-08-11 14:17:44 -070065 return test_bit(BDI_writeback_running, &bdi->state);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070066}
Theodore Ts'o00d4e732012-09-19 22:42:36 -040067EXPORT_SYMBOL(writeback_in_progress);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070068
Jan Kara692ebd12010-09-21 11:51:01 +020069static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
70{
71 struct super_block *sb = inode->i_sb;
Jan Kara692ebd12010-09-21 11:51:01 +020072
Jan Karaa8855992013-07-09 22:36:45 +080073 if (sb_is_blkdev_sb(sb))
Christoph Hellwigaaead252010-10-04 14:25:33 +020074 return inode->i_mapping->backing_dev_info;
75
76 return sb->s_bdi;
Jan Kara692ebd12010-09-21 11:51:01 +020077}
78
Nick Piggin7ccf19a2010-10-21 11:49:30 +110079static inline struct inode *wb_inode(struct list_head *head)
80{
81 return list_entry(head, struct inode, i_wb_list);
82}
83
Wu Fengguang15eb77a2012-01-17 11:18:56 -060084/*
85 * Include the creation of the trace points after defining the
86 * wb_writeback_work structure and inline functions so that the definition
87 * remains local to this file.
88 */
89#define CREATE_TRACE_POINTS
90#include <trace/events/writeback.h>
91
Steven Whitehouse774016b2014-02-06 15:47:47 +000092EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
93
Jan Kara5acda9d2014-04-03 14:46:23 -070094static void bdi_wakeup_thread(struct backing_dev_info *bdi)
95{
96 spin_lock_bh(&bdi->wb_lock);
97 if (test_bit(BDI_registered, &bdi->state))
98 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
99 spin_unlock_bh(&bdi->wb_lock);
100}
101
Jan Kara65850272011-01-13 15:45:44 -0800102static void bdi_queue_work(struct backing_dev_info *bdi,
103 struct wb_writeback_work *work)
104{
105 trace_writeback_queue(bdi, work);
106
107 spin_lock_bh(&bdi->wb_lock);
Jan Kara5acda9d2014-04-03 14:46:23 -0700108 if (!test_bit(BDI_registered, &bdi->state)) {
109 if (work->done)
110 complete(work->done);
111 goto out_unlock;
112 }
Jan Kara65850272011-01-13 15:45:44 -0800113 list_add_tail(&work->list, &bdi->work_list);
Tejun Heo839a8e82013-04-01 19:08:06 -0700114 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
Jan Kara5acda9d2014-04-03 14:46:23 -0700115out_unlock:
116 spin_unlock_bh(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117}
118
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200119static void
120__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600121 bool range_cyclic, enum wb_reason reason)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200123 struct wb_writeback_work *work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200124
Jens Axboebcddc3f2009-09-13 20:07:36 +0200125 /*
126 * This is WB_SYNC_NONE writeback, so if allocation fails just
127 * wakeup the thread for old dirty data writeback
128 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200129 work = kzalloc(sizeof(*work), GFP_ATOMIC);
130 if (!work) {
Tejun Heo839a8e82013-04-01 19:08:06 -0700131 trace_writeback_nowork(bdi);
Jan Kara5acda9d2014-04-03 14:46:23 -0700132 bdi_wakeup_thread(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200133 return;
Jens Axboebcddc3f2009-09-13 20:07:36 +0200134 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200135
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200136 work->sync_mode = WB_SYNC_NONE;
137 work->nr_pages = nr_pages;
138 work->range_cyclic = range_cyclic;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600139 work->reason = reason;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200140
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200141 bdi_queue_work(bdi, work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200142}
143
144/**
145 * bdi_start_writeback - start writeback
146 * @bdi: the backing device to write from
147 * @nr_pages: the number of pages to write
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +0800148 * @reason: reason why some writeback work was initiated
Jens Axboeb6e51312009-09-16 15:13:54 +0200149 *
150 * Description:
151 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300152 * started when this function returns, we make no guarantees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200153 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200154 *
155 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600156void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
157 enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200158{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600159 __bdi_start_writeback(bdi, nr_pages, true, reason);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200160}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800161
Christoph Hellwigc5444192010-06-08 18:15:15 +0200162/**
163 * bdi_start_background_writeback - start background writeback
164 * @bdi: the backing device to write from
165 *
166 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800167 * This makes sure WB_SYNC_NONE background writeback happens. When
168 * this function returns, it is only guaranteed that for given BDI
169 * some IO is happening if we are over background dirty threshold.
170 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200171 */
172void bdi_start_background_writeback(struct backing_dev_info *bdi)
173{
Jan Kara65850272011-01-13 15:45:44 -0800174 /*
175 * We just wake up the flusher thread. It will perform background
176 * writeback as soon as there is no other work to do.
177 */
Wu Fengguang71927e82011-01-13 15:45:46 -0800178 trace_writeback_wake_background(bdi);
Jan Kara5acda9d2014-04-03 14:46:23 -0700179 bdi_wakeup_thread(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180}
181
182/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100183 * Remove the inode from the writeback list it is on.
184 */
185void inode_wb_list_del(struct inode *inode)
186{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600187 struct backing_dev_info *bdi = inode_to_bdi(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +1100188
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600189 spin_lock(&bdi->wb.list_lock);
190 list_del_init(&inode->i_wb_list);
191 spin_unlock(&bdi->wb.list_lock);
192}
Dave Chinnera66979a2011-03-22 22:23:41 +1100193
194/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700195 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
196 * furthest end of its superblock's dirty-inode list.
197 *
198 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200199 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700200 * the case then the inode must have been redirtied while it was being written
201 * out and we don't reset its dirtied_when.
202 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600203static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700204{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600205 assert_spin_locked(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200206 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200207 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700208
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100209 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200210 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700211 inode->dirtied_when = jiffies;
212 }
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100213 list_move(&inode->i_wb_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700214}
215
216/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200217 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700218 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600219static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700220{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600221 assert_spin_locked(&wb->list_lock);
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100222 list_move(&inode->i_wb_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700223}
224
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700225static void inode_sync_complete(struct inode *inode)
226{
Jan Kara365b94ae2012-05-03 14:47:55 +0200227 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -0800228 /* If inode is clean an unused, put it into LRU now... */
229 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +0200230 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700231 smp_mb();
232 wake_up_bit(&inode->i_state, __I_SYNC);
233}
234
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700235static bool inode_dirtied_after(struct inode *inode, unsigned long t)
236{
237 bool ret = time_after(inode->dirtied_when, t);
238#ifndef CONFIG_64BIT
239 /*
240 * For inodes being constantly redirtied, dirtied_when can get stuck.
241 * It _appears_ to be in the future, but is actually in distant past.
242 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200243 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700244 */
245 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
246#endif
247 return ret;
248}
249
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700250/*
Wang Sheng-Hui0e2f2b22012-09-11 08:28:18 +0800251 * Move expired (dirtied before work->older_than_this) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -0800252 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -0700253 */
Wu Fengguange84d0a42011-04-23 12:27:27 -0600254static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -0700255 struct list_head *dispatch_queue,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600256 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700257{
Shaohua Li5c034492009-09-24 14:42:33 +0200258 LIST_HEAD(tmp);
259 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200260 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200261 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200262 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -0600263 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200264
Fengguang Wu2c136572007-10-16 23:30:39 -0700265 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100266 inode = wb_inode(delaying_queue->prev);
Jan Kara0dc83bd2014-02-21 11:19:04 +0100267 if (work->older_than_this &&
268 inode_dirtied_after(inode, *work->older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700269 break;
Jan Karaa8855992013-07-09 22:36:45 +0800270 list_move(&inode->i_wb_list, &tmp);
271 moved++;
272 if (sb_is_blkdev_sb(inode->i_sb))
273 continue;
Jens Axboecf137302009-09-24 15:12:57 +0200274 if (sb && sb != inode->i_sb)
275 do_sb_sort = 1;
276 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200277 }
278
Jens Axboecf137302009-09-24 15:12:57 +0200279 /* just one sb in list, splice to dispatch_queue and we're done */
280 if (!do_sb_sort) {
281 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -0600282 goto out;
Jens Axboecf137302009-09-24 15:12:57 +0200283 }
284
Shaohua Li5c034492009-09-24 14:42:33 +0200285 /* Move inodes from one superblock together */
286 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100287 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200288 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100289 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +0200290 if (inode->i_sb == sb)
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100291 list_move(&inode->i_wb_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +0200292 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700293 }
Wu Fengguange84d0a42011-04-23 12:27:27 -0600294out:
295 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -0700296}
297
298/*
299 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700300 * Before
301 * newly dirtied b_dirty b_io b_more_io
302 * =============> gf edc BA
303 * After
304 * newly dirtied b_dirty b_io b_more_io
305 * =============> g fBAedc
306 * |
307 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -0700308 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600309static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700310{
Wu Fengguange84d0a42011-04-23 12:27:27 -0600311 int moved;
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600312 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700313 list_splice_init(&wb->b_more_io, &wb->b_io);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600314 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
315 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200316}
317
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100318static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200319{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -0800320 int ret;
321
322 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
323 trace_writeback_write_inode_start(inode, wbc);
324 ret = inode->i_sb->s_op->write_inode(inode, wbc);
325 trace_writeback_write_inode(inode, wbc);
326 return ret;
327 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200328 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700329}
330
331/*
Jan Kara169ebd92012-05-03 14:48:03 +0200332 * Wait for writeback on an inode to complete. Called with i_lock held.
333 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200334 */
Jan Kara169ebd92012-05-03 14:48:03 +0200335static void __inode_wait_for_writeback(struct inode *inode)
336 __releases(inode->i_lock)
337 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +0200338{
339 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
340 wait_queue_head_t *wqh;
341
342 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +1100343 while (inode->i_state & I_SYNC) {
344 spin_unlock(&inode->i_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +0200345 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +1100346 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700347 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200348}
349
350/*
Jan Kara169ebd92012-05-03 14:48:03 +0200351 * Wait for writeback on an inode to complete. Caller must have inode pinned.
352 */
353void inode_wait_for_writeback(struct inode *inode)
354{
355 spin_lock(&inode->i_lock);
356 __inode_wait_for_writeback(inode);
357 spin_unlock(&inode->i_lock);
358}
359
360/*
361 * Sleep until I_SYNC is cleared. This function must be called with i_lock
362 * held and drops it. It is aimed for callers not holding any inode reference
363 * so once i_lock is dropped, inode can go away.
364 */
365static void inode_sleep_on_writeback(struct inode *inode)
366 __releases(inode->i_lock)
367{
368 DEFINE_WAIT(wait);
369 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
370 int sleep;
371
372 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
373 sleep = inode->i_state & I_SYNC;
374 spin_unlock(&inode->i_lock);
375 if (sleep)
376 schedule();
377 finish_wait(wqh, &wait);
378}
379
380/*
Jan Karaccb26b52012-05-03 14:47:58 +0200381 * Find proper writeback list for the inode depending on its current state and
382 * possibly also change of its state while we were doing writeback. Here we
383 * handle things such as livelock prevention or fairness of writeback among
384 * inodes. This function can be called only by flusher thread - noone else
385 * processes all inodes in writeback lists and requeueing inodes behind flusher
386 * thread's back can have unexpected consequences.
387 */
388static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
389 struct writeback_control *wbc)
390{
391 if (inode->i_state & I_FREEING)
392 return;
393
394 /*
395 * Sync livelock prevention. Each inode is tagged and synced in one
396 * shot. If still dirty, it will be redirty_tail()'ed below. Update
397 * the dirty time to prevent enqueue and sync it again.
398 */
399 if ((inode->i_state & I_DIRTY) &&
400 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
401 inode->dirtied_when = jiffies;
402
Jan Kara4f8ad652012-05-03 14:48:00 +0200403 if (wbc->pages_skipped) {
404 /*
405 * writeback is not making progress due to locked
406 * buffers. Skip this inode for now.
407 */
408 redirty_tail(inode, wb);
409 return;
410 }
411
Jan Karaccb26b52012-05-03 14:47:58 +0200412 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
413 /*
414 * We didn't write back all the pages. nfs_writepages()
415 * sometimes bales out without doing anything.
416 */
417 if (wbc->nr_to_write <= 0) {
418 /* Slice used up. Queue for next turn. */
419 requeue_io(inode, wb);
420 } else {
421 /*
422 * Writeback blocked by something other than
423 * congestion. Delay the inode for some time to
424 * avoid spinning on the CPU (100% iowait)
425 * retrying writeback of the dirty page/inode
426 * that cannot be performed immediately.
427 */
428 redirty_tail(inode, wb);
429 }
430 } else if (inode->i_state & I_DIRTY) {
431 /*
432 * Filesystems can dirty the inode during writeback operations,
433 * such as delayed allocation during submission or metadata
434 * updates after data IO completion.
435 */
436 redirty_tail(inode, wb);
437 } else {
438 /* The inode is clean. Remove from writeback lists. */
439 list_del_init(&inode->i_wb_list);
440 }
441}
442
443/*
Jan Kara4f8ad652012-05-03 14:48:00 +0200444 * Write out an inode and its dirty pages. Do not update the writeback list
445 * linkage. That is left to the caller. The caller is also responsible for
446 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 */
448static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -0700449__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -0600452 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200453 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 int ret;
455
Jan Kara4f8ad652012-05-03 14:48:00 +0200456 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457
Tejun Heo9fb0a7d2013-01-11 13:06:37 -0800458 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 ret = do_writepages(mapping, wbc);
461
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100462 /*
463 * Make sure to wait on the data before writing out the metadata.
464 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +1000465 * I/O completion. We don't do it for sync(2) writeback because it has a
466 * separate, external IO completion path and ->sync_fs for guaranteeing
467 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100468 */
Dave Chinner7747bd42013-07-02 22:38:35 +1000469 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100470 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 if (ret == 0)
472 ret = err;
473 }
474
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400475 /*
476 * Some filesystems may redirty the inode during the writeback
477 * due to delalloc, clear dirty metadata flags right before
478 * write_inode()
479 */
Dave Chinner250df6e2011-03-22 22:23:36 +1100480 spin_lock(&inode->i_lock);
Jan Kara6290be12012-05-03 14:47:57 +0200481 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
482 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
483 inode->i_state &= ~I_DIRTY_PAGES;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400484 dirty = inode->i_state & I_DIRTY;
485 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +1100486 spin_unlock(&inode->i_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100487 /* Don't write the inode if only I_DIRTY_PAGES was set */
488 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100489 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 if (ret == 0)
491 ret = err;
492 }
Jan Kara4f8ad652012-05-03 14:48:00 +0200493 trace_writeback_single_inode(inode, wbc, nr_to_write);
494 return ret;
495}
496
497/*
498 * Write out an inode's dirty pages. Either the caller has an active reference
499 * on the inode or the inode has I_WILL_FREE set.
500 *
501 * This function is designed to be called for writing back one inode which
502 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
503 * and does more profound writeback list handling in writeback_sb_inodes().
504 */
505static int
506writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
507 struct writeback_control *wbc)
508{
509 int ret = 0;
510
511 spin_lock(&inode->i_lock);
512 if (!atomic_read(&inode->i_count))
513 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
514 else
515 WARN_ON(inode->i_state & I_WILL_FREE);
516
517 if (inode->i_state & I_SYNC) {
518 if (wbc->sync_mode != WB_SYNC_ALL)
519 goto out;
520 /*
Jan Kara169ebd92012-05-03 14:48:03 +0200521 * It's a data-integrity sync. We must wait. Since callers hold
522 * inode reference or inode has I_WILL_FREE set, it cannot go
523 * away under us.
Jan Kara4f8ad652012-05-03 14:48:00 +0200524 */
Jan Kara169ebd92012-05-03 14:48:03 +0200525 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +0200526 }
527 WARN_ON(inode->i_state & I_SYNC);
528 /*
Jan Karaf9b0e052013-12-14 04:21:26 +0800529 * Skip inode if it is clean and we have no outstanding writeback in
530 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
531 * function since flusher thread may be doing for example sync in
532 * parallel and if we move the inode, it could get skipped. So here we
533 * make sure inode is on some writeback list and leave it there unless
534 * we have completely cleaned the inode.
Jan Kara4f8ad652012-05-03 14:48:00 +0200535 */
Jan Karaf9b0e052013-12-14 04:21:26 +0800536 if (!(inode->i_state & I_DIRTY) &&
537 (wbc->sync_mode != WB_SYNC_ALL ||
538 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +0200539 goto out;
540 inode->i_state |= I_SYNC;
541 spin_unlock(&inode->i_lock);
542
Yan Hongcd8ed2a2012-10-08 16:33:45 -0700543 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600545 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +1100546 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +0200547 /*
548 * If inode is clean, remove it from writeback lists. Otherwise don't
549 * touch it. See comment above for explanation.
550 */
551 if (!(inode->i_state & I_DIRTY))
552 list_del_init(&inode->i_wb_list);
553 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700554 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +0200555out:
556 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 return ret;
558}
559
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600560static long writeback_chunk_size(struct backing_dev_info *bdi,
561 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600562{
563 long pages;
564
565 /*
566 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
567 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
568 * here avoids calling into writeback_inodes_wb() more than once.
569 *
570 * The intended call sequence for WB_SYNC_ALL writeback is:
571 *
572 * wb_writeback()
573 * writeback_sb_inodes() <== called only once
574 * write_cache_pages() <== called once for each inode
575 * (quickly) tag currently dirty pages
576 * (maybe slowly) sync all tagged pages
577 */
578 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
579 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600580 else {
581 pages = min(bdi->avg_write_bandwidth / 2,
582 global_dirty_limit / DIRTY_SCOPE);
583 pages = min(pages, work->nr_pages);
584 pages = round_down(pages + MIN_WRITEBACK_PAGES,
585 MIN_WRITEBACK_PAGES);
586 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600587
588 return pages;
589}
590
Jens Axboe03ba3782009-09-09 09:08:54 +0200591/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800592 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200593 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600594 * Return the number of pages and/or inodes written.
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800595 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600596static long writeback_sb_inodes(struct super_block *sb,
597 struct bdi_writeback *wb,
598 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200599{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600600 struct writeback_control wbc = {
601 .sync_mode = work->sync_mode,
602 .tagged_writepages = work->tagged_writepages,
603 .for_kupdate = work->for_kupdate,
604 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +1000605 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600606 .range_cyclic = work->range_cyclic,
607 .range_start = 0,
608 .range_end = LLONG_MAX,
609 };
610 unsigned long start_time = jiffies;
611 long write_chunk;
612 long wrote = 0; /* count both pages and inodes */
613
Jens Axboe03ba3782009-09-09 09:08:54 +0200614 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100615 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200616
617 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600618 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200619 /*
620 * We only want to write back data for this
621 * superblock, move all inodes not belonging
622 * to it back onto the dirty list.
623 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600624 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200625 continue;
626 }
627
628 /*
629 * The inode belongs to a different superblock.
630 * Bounce back to the caller to unpin this and
631 * pin the next superblock.
632 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600633 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200634 }
635
Christoph Hellwig9843b762010-10-24 19:40:46 +0200636 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +0800637 * Don't bother with new inodes or inodes being freed, first
638 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b762010-10-24 19:40:46 +0200639 * kind writeout is handled by the freer.
640 */
Dave Chinner250df6e2011-03-22 22:23:36 +1100641 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +0200642 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +1100643 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -0700644 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -0700645 continue;
646 }
Jan Karacc1676d2012-05-03 14:47:56 +0200647 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
648 /*
649 * If this inode is locked for writeback and we are not
650 * doing writeback-for-data-integrity, move it to
651 * b_more_io so that writeback can proceed with the
652 * other inodes on s_io.
653 *
654 * We'll have another go at writing back this inode
655 * when we completed a full scan of b_io.
656 */
657 spin_unlock(&inode->i_lock);
658 requeue_io(inode, wb);
659 trace_writeback_sb_inodes_requeue(inode);
660 continue;
661 }
Jan Karaf0d07b72012-05-03 14:47:59 +0200662 spin_unlock(&wb->list_lock);
663
Jan Kara4f8ad652012-05-03 14:48:00 +0200664 /*
665 * We already requeued the inode if it had I_SYNC set and we
666 * are doing WB_SYNC_NONE writeback. So this catches only the
667 * WB_SYNC_ALL case.
668 */
Jan Kara169ebd92012-05-03 14:48:03 +0200669 if (inode->i_state & I_SYNC) {
670 /* Wait for I_SYNC. This function drops i_lock... */
671 inode_sleep_on_writeback(inode);
672 /* Inode may be gone, start again */
Jan Karaead188f2012-06-08 17:07:36 +0200673 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +0200674 continue;
675 }
Jan Kara4f8ad652012-05-03 14:48:00 +0200676 inode->i_state |= I_SYNC;
677 spin_unlock(&inode->i_lock);
Jan Kara169ebd92012-05-03 14:48:03 +0200678
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600679 write_chunk = writeback_chunk_size(wb->bdi, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600680 wbc.nr_to_write = write_chunk;
681 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +1100682
Jan Kara169ebd92012-05-03 14:48:03 +0200683 /*
684 * We use I_SYNC to pin the inode in memory. While it is set
685 * evict_inode() will wait so the inode cannot be freed.
686 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -0700687 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600688
689 work->nr_pages -= write_chunk - wbc.nr_to_write;
690 wrote += write_chunk - wbc.nr_to_write;
Jan Kara4f8ad652012-05-03 14:48:00 +0200691 spin_lock(&wb->list_lock);
692 spin_lock(&inode->i_lock);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600693 if (!(inode->i_state & I_DIRTY))
694 wrote++;
Jan Kara4f8ad652012-05-03 14:48:00 +0200695 requeue_inode(inode, wb, &wbc);
696 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +1100697 spin_unlock(&inode->i_lock);
Jan Kara169ebd92012-05-03 14:48:03 +0200698 cond_resched_lock(&wb->list_lock);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600699 /*
700 * bail out to wb_writeback() often enough to check
701 * background threshold and other termination conditions.
702 */
703 if (wrote) {
704 if (time_is_before_jiffies(start_time + HZ / 10UL))
705 break;
706 if (work->nr_pages <= 0)
707 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600710 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800711}
Nick Piggin38f21972009-01-06 14:40:25 -0800712
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600713static long __writeback_inodes_wb(struct bdi_writeback *wb,
714 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800715{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600716 unsigned long start_time = jiffies;
717 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800718
719 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100720 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800721 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800722
Dave Chinner12ad3ab2011-07-08 14:14:41 +1000723 if (!grab_super_passive(sb)) {
Wu Fengguang0e995812011-07-29 22:14:35 -0600724 /*
725 * grab_super_passive() may fail consistently due to
726 * s_umount being grabbed by someone else. Don't use
727 * requeue_io() to avoid busy retrying the inode/sb.
728 */
729 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200730 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800731 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600732 wrote += writeback_sb_inodes(sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200733 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800734
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600735 /* refer to the same tests at the end of writeback_sb_inodes */
736 if (wrote) {
737 if (time_is_before_jiffies(start_time + HZ / 10UL))
738 break;
739 if (work->nr_pages <= 0)
740 break;
741 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800742 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200743 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600744 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200745}
746
Wanpeng Li7d9f0732013-09-11 14:22:40 -0700747static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600748 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200749{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600750 struct wb_writeback_work work = {
751 .nr_pages = nr_pages,
752 .sync_mode = WB_SYNC_NONE,
753 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600754 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600755 };
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200756
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600757 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -0600758 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600759 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600760 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600761 spin_unlock(&wb->list_lock);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200762
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600763 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200764}
765
Wu Fengguangb00949a2010-11-18 14:38:33 -0600766static bool over_bground_thresh(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200767{
768 unsigned long background_thresh, dirty_thresh;
769
Wu Fengguang16c40422010-08-11 14:17:39 -0700770 global_dirty_limits(&background_thresh, &dirty_thresh);
Jens Axboe03ba3782009-09-09 09:08:54 +0200771
Wu Fengguangb00949a2010-11-18 14:38:33 -0600772 if (global_page_state(NR_FILE_DIRTY) +
773 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
774 return true;
775
776 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
777 bdi_dirty_limit(bdi, background_thresh))
778 return true;
779
780 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200781}
782
783/*
Wu Fengguange98be2d2010-08-29 11:22:30 -0600784 * Called under wb->list_lock. If there are multiple wb per bdi,
785 * only the flusher working on the first wb should do it.
786 */
787static void wb_update_bandwidth(struct bdi_writeback *wb,
788 unsigned long start_time)
789{
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600790 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600791}
792
793/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200794 * Explicit flushing or periodic writeback of "old" data.
795 *
796 * Define "old": the first time one of an inode's pages is dirtied, we mark the
797 * dirtying-time in the inode's address_space. So this periodic writeback code
798 * just walks the superblock inode list, writing back any inodes which are
799 * older than a specific point in time.
800 *
801 * Try to run once per dirty_writeback_interval. But if a writeback event
802 * takes longer than a dirty_writeback_interval interval, then leave a
803 * one-second gap.
804 *
805 * older_than_this takes precedence over nr_to_write. So we'll only write back
806 * all dirty pages if they are all attached to "old" mappings.
807 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200808static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200809 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200810{
Wu Fengguange98be2d2010-08-29 11:22:30 -0600811 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600812 long nr_pages = work->nr_pages;
Jan Kara0dc83bd2014-02-21 11:19:04 +0100813 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +0200814 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600815 long progress;
Jens Axboe03ba3782009-09-09 09:08:54 +0200816
Jan Kara0dc83bd2014-02-21 11:19:04 +0100817 oldest_jif = jiffies;
818 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +0200819
Wu Fengguange8dfc302011-04-21 12:06:32 -0600820 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200821 for (;;) {
822 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800823 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200824 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200825 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200826 break;
827
828 /*
Jan Karaaa373cf2011-01-13 15:45:47 -0800829 * Background writeout and kupdate-style writeback may
830 * run forever. Stop them if there is other work to do
831 * so that e.g. sync can proceed. They'll be restarted
832 * after the other works are all done.
833 */
834 if ((work->for_background || work->for_kupdate) &&
835 !list_empty(&wb->bdi->work_list))
836 break;
837
838 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800839 * For background writeout, stop when we are below the
840 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200841 */
Wu Fengguangb00949a2010-11-18 14:38:33 -0600842 if (work->for_background && !over_bground_thresh(wb->bdi))
Jens Axboe03ba3782009-09-09 09:08:54 +0200843 break;
844
Jan Kara1bc36b62011-10-19 11:44:41 +0200845 /*
846 * Kupdate and background works are special and we want to
847 * include all inodes that need writing. Livelock avoidance is
848 * handled by these works yielding to any other work so we are
849 * safe.
850 */
Wu Fengguangba9aa832010-07-21 20:32:30 -0600851 if (work->for_kupdate) {
Jan Kara0dc83bd2014-02-21 11:19:04 +0100852 oldest_jif = jiffies -
Wu Fengguangba9aa832010-07-21 20:32:30 -0600853 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +0200854 } else if (work->for_background)
Jan Kara0dc83bd2014-02-21 11:19:04 +0100855 oldest_jif = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +1000856
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600857 trace_writeback_start(wb->bdi, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -0600858 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600859 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200860 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600861 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200862 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600863 progress = __writeback_inodes_wb(wb, work);
864 trace_writeback_written(wb->bdi, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000865
Wu Fengguange98be2d2010-08-29 11:22:30 -0600866 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +0200867
868 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200869 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -0600870 *
871 * Dirty inodes are moved to b_io for writeback in batches.
872 * The completion of the current batch does not necessarily
873 * mean the overall work is done. So we keep looping as long
874 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +0200875 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600876 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +0200877 continue;
878 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -0600879 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +0200880 */
Wu Fengguangb7a24412010-07-21 22:19:51 -0600881 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +0200882 break;
883 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +0200884 * Nothing written. Wait for some inode to
885 * become available for writeback. Otherwise
886 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +0200887 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200888 if (!list_empty(&wb->b_more_io)) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600889 trace_writeback_wait(wb->bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200890 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +1100891 spin_lock(&inode->i_lock);
Jan Karaf0d07b72012-05-03 14:47:59 +0200892 spin_unlock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +0200893 /* This function drops i_lock... */
894 inode_sleep_on_writeback(inode);
Jan Karaf0d07b72012-05-03 14:47:59 +0200895 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200896 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200897 }
Wu Fengguange8dfc302011-04-21 12:06:32 -0600898 spin_unlock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200899
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600900 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +0200901}
902
903/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200904 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +0200905 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200906static struct wb_writeback_work *
Minchan Kim08852b62010-08-03 12:51:16 +0200907get_next_work_item(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200908{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200909 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200910
Artem Bityutskiy64677162010-07-25 14:29:22 +0300911 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200912 if (!list_empty(&bdi->work_list)) {
913 work = list_entry(bdi->work_list.next,
914 struct wb_writeback_work, list);
915 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200916 }
Artem Bityutskiy64677162010-07-25 14:29:22 +0300917 spin_unlock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200918 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200919}
920
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700921/*
922 * Add in the number of potentially dirty inodes, because each inode
923 * write can dirty pagecache in the underlying blockdev.
924 */
925static unsigned long get_nr_dirty_pages(void)
926{
927 return global_page_state(NR_FILE_DIRTY) +
928 global_page_state(NR_UNSTABLE_NFS) +
929 get_nr_dirty_inodes();
930}
931
Jan Kara65850272011-01-13 15:45:44 -0800932static long wb_check_background_flush(struct bdi_writeback *wb)
933{
Wu Fengguangb00949a2010-11-18 14:38:33 -0600934 if (over_bground_thresh(wb->bdi)) {
Jan Kara65850272011-01-13 15:45:44 -0800935
936 struct wb_writeback_work work = {
937 .nr_pages = LONG_MAX,
938 .sync_mode = WB_SYNC_NONE,
939 .for_background = 1,
940 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600941 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -0800942 };
943
944 return wb_writeback(wb, &work);
945 }
946
947 return 0;
948}
949
Jens Axboe03ba3782009-09-09 09:08:54 +0200950static long wb_check_old_data_flush(struct bdi_writeback *wb)
951{
952 unsigned long expired;
953 long nr_pages;
954
Jens Axboe69b62d02010-05-17 12:51:03 +0200955 /*
956 * When set to zero, disable periodic writeback
957 */
958 if (!dirty_writeback_interval)
959 return 0;
960
Jens Axboe03ba3782009-09-09 09:08:54 +0200961 expired = wb->last_old_flush +
962 msecs_to_jiffies(dirty_writeback_interval * 10);
963 if (time_before(jiffies, expired))
964 return 0;
965
966 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700967 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +0200968
Jens Axboec4a77a62009-09-16 15:18:25 +0200969 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200970 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200971 .nr_pages = nr_pages,
972 .sync_mode = WB_SYNC_NONE,
973 .for_kupdate = 1,
974 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600975 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +0200976 };
977
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200978 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +0200979 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200980
981 return 0;
982}
983
984/*
985 * Retrieve work items and do the writeback they describe
986 */
Wanpeng Li25d130b2013-07-08 16:00:14 -0700987static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200988{
989 struct backing_dev_info *bdi = wb->bdi;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200990 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200991 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200992
Jan Kara81d73a32010-08-11 14:17:44 -0700993 set_bit(BDI_writeback_running, &wb->bdi->state);
Minchan Kim08852b62010-08-03 12:51:16 +0200994 while ((work = get_next_work_item(bdi)) != NULL) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200995
Dave Chinner455b2862010-07-07 13:24:06 +1000996 trace_writeback_exec(bdi, work);
997
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200998 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200999
1000 /*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001001 * Notify the caller of completion if this is a synchronous
1002 * work item, otherwise just free it.
Jens Axboe03ba3782009-09-09 09:08:54 +02001003 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001004 if (work->done)
1005 complete(work->done);
1006 else
1007 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001008 }
1009
1010 /*
1011 * Check for periodic writeback, kupdated() style
1012 */
1013 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08001014 wrote += wb_check_background_flush(wb);
Jan Kara81d73a32010-08-11 14:17:44 -07001015 clear_bit(BDI_writeback_running, &wb->bdi->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02001016
1017 return wrote;
1018}
1019
1020/*
1021 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07001022 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02001023 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001024void bdi_writeback_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001025{
Tejun Heo839a8e82013-04-01 19:08:06 -07001026 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1027 struct bdi_writeback, dwork);
Christoph Hellwig08243902010-06-19 23:08:22 +02001028 struct backing_dev_info *bdi = wb->bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +02001029 long pages_written;
1030
Tejun Heoef3b1012013-04-30 15:27:24 -07001031 set_worker_desc("flush-%s", dev_name(bdi->dev));
Peter Zijlstra766f9162010-10-26 14:22:45 -07001032 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02001033
Tejun Heo839a8e82013-04-01 19:08:06 -07001034 if (likely(!current_is_workqueue_rescuer() ||
Jan Kara5acda9d2014-04-03 14:46:23 -07001035 !test_bit(BDI_registered, &bdi->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03001036 /*
Tejun Heo839a8e82013-04-01 19:08:06 -07001037 * The normal path. Keep writing back @bdi until its
1038 * work_list is empty. Note that this path is also taken
1039 * if @bdi is shutting down even when we're running off the
1040 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03001041 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001042 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07001043 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07001044 trace_writeback_pages_written(pages_written);
1045 } while (!list_empty(&bdi->work_list));
1046 } else {
1047 /*
1048 * bdi_wq can't get enough workers and we're running off
1049 * the emergency worker. Don't hog it. Hopefully, 1024 is
1050 * enough for efficient IO.
1051 */
1052 pages_written = writeback_inodes_wb(&bdi->wb, 1024,
1053 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10001054 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02001055 }
1056
Derek Basehore6ca738d2014-04-03 14:46:22 -07001057 if (!list_empty(&bdi->work_list))
1058 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1059 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1060 bdi_wakeup_thread_delayed(bdi);
Dave Chinner455b2862010-07-07 13:24:06 +10001061
Tejun Heo839a8e82013-04-01 19:08:06 -07001062 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02001063}
1064
1065/*
Jens Axboe03ba3782009-09-09 09:08:54 +02001066 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1067 * the whole world.
1068 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001069void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02001070{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001071 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001072
Jan Kara47df3dd2013-09-11 14:22:22 -07001073 if (!nr_pages)
1074 nr_pages = get_nr_dirty_pages();
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001075
1076 rcu_read_lock();
1077 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1078 if (!bdi_has_dirty_io(bdi))
1079 continue;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001080 __bdi_start_writeback(bdi, nr_pages, false, reason);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001081 }
1082 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001083}
1084
1085static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1086{
1087 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1088 struct dentry *dentry;
1089 const char *name = "?";
1090
1091 dentry = d_find_alias(inode);
1092 if (dentry) {
1093 spin_lock(&dentry->d_lock);
1094 name = (const char *) dentry->d_name.name;
1095 }
1096 printk(KERN_DEBUG
1097 "%s(%d): dirtied inode %lu (%s) on %s\n",
1098 current->comm, task_pid_nr(current), inode->i_ino,
1099 name, inode->i_sb->s_id);
1100 if (dentry) {
1101 spin_unlock(&dentry->d_lock);
1102 dput(dentry);
1103 }
1104 }
1105}
1106
1107/**
1108 * __mark_inode_dirty - internal function
1109 * @inode: inode to mark
1110 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1111 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1112 * mark_inode_dirty_sync.
1113 *
1114 * Put the inode on the super block's dirty list.
1115 *
1116 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1117 * dirty list only if it is hashed or if it refers to a blockdev.
1118 * If it was not hashed, it will never be added to the dirty list
1119 * even if it is later hashed, as it will have been marked dirty already.
1120 *
1121 * In short, make sure you hash any inodes _before_ you start marking
1122 * them dirty.
1123 *
Jens Axboe03ba3782009-09-09 09:08:54 +02001124 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1125 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1126 * the kernel-internal blockdev inode represents the dirtying time of the
1127 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1128 * page->mapping->host, so the page-dirtying time is recorded in the internal
1129 * blockdev inode.
1130 */
1131void __mark_inode_dirty(struct inode *inode, int flags)
1132{
1133 struct super_block *sb = inode->i_sb;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001134 struct backing_dev_info *bdi = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001135
1136 /*
1137 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1138 * dirty the inode itself
1139 */
1140 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001141 trace_writeback_dirty_inode_start(inode, flags);
1142
Jens Axboe03ba3782009-09-09 09:08:54 +02001143 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04001144 sb->s_op->dirty_inode(inode, flags);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001145
1146 trace_writeback_dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02001147 }
1148
1149 /*
1150 * make sure that changes are seen by all cpus before we test i_state
1151 * -- mikulas
1152 */
1153 smp_mb();
1154
1155 /* avoid the locking if we can */
1156 if ((inode->i_state & flags) == flags)
1157 return;
1158
1159 if (unlikely(block_dump))
1160 block_dump___mark_inode_dirty(inode);
1161
Dave Chinner250df6e2011-03-22 22:23:36 +11001162 spin_lock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001163 if ((inode->i_state & flags) != flags) {
1164 const int was_dirty = inode->i_state & I_DIRTY;
1165
1166 inode->i_state |= flags;
1167
1168 /*
1169 * If the inode is being synced, just update its dirty state.
1170 * The unlocker will place the inode on the appropriate
1171 * superblock list, based upon its state.
1172 */
1173 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11001174 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001175
1176 /*
1177 * Only add valid (hashed) inodes to the superblock's
1178 * dirty list. Add blockdev inodes as well.
1179 */
1180 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382c2010-10-23 15:19:20 -04001181 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11001182 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001183 }
Al Viroa4ffdde2010-06-02 17:38:30 -04001184 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11001185 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001186
1187 /*
1188 * If the inode was already on b_dirty/b_io/b_more_io, don't
1189 * reposition it (that would break b_dirty time-ordering).
1190 */
1191 if (!was_dirty) {
Dave Chinnera66979a2011-03-22 22:23:41 +11001192 bool wakeup_bdi = false;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001193 bdi = inode_to_bdi(inode);
Jens Axboe500b0672009-09-09 09:10:25 +02001194
Junxiao Bi146d7002013-09-11 14:23:04 -07001195 spin_unlock(&inode->i_lock);
1196 spin_lock(&bdi->wb.list_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001197 if (bdi_cap_writeback_dirty(bdi)) {
1198 WARN(!test_bit(BDI_registered, &bdi->state),
1199 "bdi-%s not registered\n", bdi->name);
1200
1201 /*
1202 * If this is the first dirty inode for this
1203 * bdi, we have to wake-up the corresponding
1204 * bdi thread to make sure background
1205 * write-back happens later.
1206 */
1207 if (!wb_has_dirty_io(&bdi->wb))
1208 wakeup_bdi = true;
Jens Axboe500b0672009-09-09 09:10:25 +02001209 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001210
1211 inode->dirtied_when = jiffies;
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001212 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001213 spin_unlock(&bdi->wb.list_lock);
Dave Chinnera66979a2011-03-22 22:23:41 +11001214
1215 if (wakeup_bdi)
1216 bdi_wakeup_thread_delayed(bdi);
1217 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02001218 }
1219 }
Dave Chinner250df6e2011-03-22 22:23:36 +11001220out_unlock_inode:
1221 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001222
Jens Axboe03ba3782009-09-09 09:08:54 +02001223}
1224EXPORT_SYMBOL(__mark_inode_dirty);
1225
Jens Axboeb6e51312009-09-16 15:13:54 +02001226static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001227{
Jens Axboe03ba3782009-09-09 09:08:54 +02001228 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001229
Jens Axboe03ba3782009-09-09 09:08:54 +02001230 /*
1231 * We need to be protected against the filesystem going from
1232 * r/o to r/w or vice versa.
1233 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001234 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001235
Dave Chinner55fa6092011-03-22 22:23:40 +11001236 spin_lock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001237
1238 /*
1239 * Data integrity sync. Must wait for all pages under writeback,
1240 * because there may have been pages dirtied before our sync
1241 * call, but which had writeout started before we write it out.
1242 * In which case, the inode may not be on the dirty list, but
1243 * we still have to wait for that writeout.
1244 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001245 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001246 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02001247
Dave Chinner250df6e2011-03-22 22:23:36 +11001248 spin_lock(&inode->i_lock);
1249 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1250 (mapping->nrpages == 0)) {
1251 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001252 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11001253 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001254 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11001255 spin_unlock(&inode->i_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11001256 spin_unlock(&inode_sb_list_lock);
1257
Jens Axboe03ba3782009-09-09 09:08:54 +02001258 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11001259 * We hold a reference to 'inode' so it couldn't have been
1260 * removed from s_inodes list while we dropped the
1261 * inode_sb_list_lock. We cannot iput the inode now as we can
1262 * be holding the last reference and we cannot iput it under
1263 * inode_sb_list_lock. So we keep the reference and iput it
1264 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02001265 */
1266 iput(old_inode);
1267 old_inode = inode;
1268
1269 filemap_fdatawait(mapping);
1270
1271 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001272
Dave Chinner55fa6092011-03-22 22:23:40 +11001273 spin_lock(&inode_sb_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001274 }
Dave Chinner55fa6092011-03-22 22:23:40 +11001275 spin_unlock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001276 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277}
1278
Jens Axboed8a85592009-09-02 12:34:32 +02001279/**
Chris Mason3259f8b2010-10-29 11:16:17 -04001280 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
Jens Axboed8a85592009-09-02 12:34:32 +02001281 * @sb: the superblock
Chris Mason3259f8b2010-10-29 11:16:17 -04001282 * @nr: the number of pages to write
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08001283 * @reason: reason why some writeback work initiated
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 *
Jens Axboed8a85592009-09-02 12:34:32 +02001285 * Start writeback on some inodes on this super_block. No guarantees are made
1286 * on how many (if any) will be written, and this function does not wait
Chris Mason3259f8b2010-10-29 11:16:17 -04001287 * for IO completion of submitted IO.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001289void writeback_inodes_sb_nr(struct super_block *sb,
1290 unsigned long nr,
1291 enum wb_reason reason)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001293 DECLARE_COMPLETION_ONSTACK(done);
1294 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001295 .sb = sb,
1296 .sync_mode = WB_SYNC_NONE,
1297 .tagged_writepages = 1,
1298 .done = &done,
1299 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001300 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001301 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001302
Jan Kara6eedc702012-07-03 16:45:27 +02001303 if (sb->s_bdi == &noop_backing_dev_info)
1304 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001305 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001306 bdi_queue_work(sb->s_bdi, &work);
1307 wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308}
Chris Mason3259f8b2010-10-29 11:16:17 -04001309EXPORT_SYMBOL(writeback_inodes_sb_nr);
1310
1311/**
1312 * writeback_inodes_sb - writeback dirty inodes from given super_block
1313 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08001314 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04001315 *
1316 * Start writeback on some inodes on this super_block. No guarantees are made
1317 * on how many (if any) will be written, and this function does not wait
1318 * for IO completion of submitted IO.
1319 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001320void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04001321{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001322 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04001323}
Jens Axboed8a85592009-09-02 12:34:32 +02001324EXPORT_SYMBOL(writeback_inodes_sb);
1325
1326/**
Miao Xie10ee27a2013-01-10 13:47:57 +08001327 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
Chris Mason3259f8b2010-10-29 11:16:17 -04001328 * @sb: the superblock
1329 * @nr: the number of pages to write
Miao Xie10ee27a2013-01-10 13:47:57 +08001330 * @reason: the reason of writeback
Chris Mason3259f8b2010-10-29 11:16:17 -04001331 *
Miao Xie10ee27a2013-01-10 13:47:57 +08001332 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
Chris Mason3259f8b2010-10-29 11:16:17 -04001333 * Returns 1 if writeback was started, 0 if not.
1334 */
Miao Xie10ee27a2013-01-10 13:47:57 +08001335int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1336 unsigned long nr,
1337 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04001338{
Miao Xie10ee27a2013-01-10 13:47:57 +08001339 if (writeback_in_progress(sb->s_bdi))
Chris Mason3259f8b2010-10-29 11:16:17 -04001340 return 1;
Miao Xie10ee27a2013-01-10 13:47:57 +08001341
1342 if (!down_read_trylock(&sb->s_umount))
Chris Mason3259f8b2010-10-29 11:16:17 -04001343 return 0;
Miao Xie10ee27a2013-01-10 13:47:57 +08001344
1345 writeback_inodes_sb_nr(sb, nr, reason);
1346 up_read(&sb->s_umount);
1347 return 1;
Chris Mason3259f8b2010-10-29 11:16:17 -04001348}
Miao Xie10ee27a2013-01-10 13:47:57 +08001349EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1350
1351/**
1352 * try_to_writeback_inodes_sb - try to start writeback if none underway
1353 * @sb: the superblock
1354 * @reason: reason why some writeback work was initiated
1355 *
1356 * Implement by try_to_writeback_inodes_sb_nr()
1357 * Returns 1 if writeback was started, 0 if not.
1358 */
1359int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1360{
1361 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1362}
1363EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04001364
1365/**
Jens Axboed8a85592009-09-02 12:34:32 +02001366 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01001367 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02001368 *
1369 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01001370 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02001371 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01001372void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001373{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001374 DECLARE_COMPLETION_ONSTACK(done);
1375 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001376 .sb = sb,
1377 .sync_mode = WB_SYNC_ALL,
1378 .nr_pages = LONG_MAX,
1379 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001380 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001381 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10001382 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001383 };
1384
Jan Kara6eedc702012-07-03 16:45:27 +02001385 /* Nothing to do? */
1386 if (sb->s_bdi == &noop_backing_dev_info)
1387 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001388 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1389
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001390 bdi_queue_work(sb->s_bdi, &work);
1391 wait_for_completion(&done);
1392
Jens Axboeb6e51312009-09-16 15:13:54 +02001393 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001394}
1395EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001398 * write_inode_now - write an inode to disk
1399 * @inode: inode to write to disk
1400 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001402 * This function commits an inode to disk immediately if it is dirty. This is
1403 * primarily needed by knfsd.
1404 *
1405 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407int write_inode_now(struct inode *inode, int sync)
1408{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001409 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 struct writeback_control wbc = {
1411 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001412 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001413 .range_start = 0,
1414 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 };
1416
1417 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001418 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419
1420 might_sleep();
Jan Kara4f8ad652012-05-03 14:48:00 +02001421 return writeback_single_inode(inode, wb, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422}
1423EXPORT_SYMBOL(write_inode_now);
1424
1425/**
1426 * sync_inode - write an inode and its pages to disk.
1427 * @inode: the inode to sync
1428 * @wbc: controls the writeback mode
1429 *
1430 * sync_inode() will write an inode and its pages to disk. It will also
1431 * correctly update the inode on its superblock's dirty inode lists and will
1432 * update inode->i_state.
1433 *
1434 * The caller must have a ref on the inode.
1435 */
1436int sync_inode(struct inode *inode, struct writeback_control *wbc)
1437{
Jan Kara4f8ad652012-05-03 14:48:00 +02001438 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439}
1440EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02001441
1442/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08001443 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02001444 * @inode: the inode to sync
1445 * @wait: wait for I/O to complete.
1446 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08001447 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02001448 *
1449 * Note: only writes the actual inode, no associated data or other metadata.
1450 */
1451int sync_inode_metadata(struct inode *inode, int wait)
1452{
1453 struct writeback_control wbc = {
1454 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1455 .nr_to_write = 0, /* metadata-only */
1456 };
1457
1458 return sync_inode(inode, &wbc);
1459}
1460EXPORT_SYMBOL(sync_inode_metadata);