blob: 73c3992b2bb4aa765d3dc2418e907ca0a7ba78f2 [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>
Jens Axboef5ff8422007-09-21 09:19:54 +020017#include <linux/module.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>
Jens Axboe03ba3782009-09-09 09:08:54 +020023#include <linux/kthread.h>
24#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
28#include <linux/buffer_head.h>
Dave Chinner455b2862010-07-07 13:24:06 +100029#include <linux/tracepoint.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/*
Jens Axboec4a77a62009-09-16 15:18:25 +020033 * Passed into wb_writeback(), essentially a subset of writeback_control
34 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020035struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020036 long nr_pages;
37 struct super_block *sb;
Wu Fengguangd46db3d2011-05-04 19:54:37 -060038 unsigned long *older_than_this;
Jens Axboec4a77a62009-09-16 15:18:25 +020039 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060040 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050041 unsigned int for_kupdate:1;
42 unsigned int range_cyclic:1;
43 unsigned int for_background:1;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060044 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020045
Jens Axboe8010c3b2009-09-15 20:04:57 +020046 struct list_head list; /* pending work list */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020047 struct completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020048};
49
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060050const char *wb_reason_name[] = {
51 [WB_REASON_BACKGROUND] = "background",
52 [WB_REASON_TRY_TO_FREE_PAGES] = "try_to_free_pages",
53 [WB_REASON_SYNC] = "sync",
54 [WB_REASON_PERIODIC] = "periodic",
55 [WB_REASON_LAPTOP_TIMER] = "laptop_timer",
56 [WB_REASON_FREE_MORE_MEM] = "free_more_memory",
57 [WB_REASON_FS_FREE_SPACE] = "fs_free_space",
58 [WB_REASON_FORKER_THREAD] = "forker_thread"
59};
60
Dave Chinner455b2862010-07-07 13:24:06 +100061/*
62 * Include the creation of the trace points after defining the
63 * wb_writeback_work structure so that the definition remains local to this
64 * file.
65 */
66#define CREATE_TRACE_POINTS
67#include <trace/events/writeback.h>
68
Dave Chinner455b2862010-07-07 13:24:06 +100069/*
70 * We don't actually have pdflush, but this one is exported though /proc...
71 */
72int nr_pdflush_threads;
73
Adrian Bunkf11b00f2008-04-29 00:58:56 -070074/**
75 * writeback_in_progress - determine whether there is writeback in progress
76 * @bdi: the device's backing_dev_info structure.
77 *
Jens Axboe03ba3782009-09-09 09:08:54 +020078 * Determine whether there is writeback waiting to be handled against a
79 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070080 */
81int writeback_in_progress(struct backing_dev_info *bdi)
82{
Jan Kara81d73a32010-08-11 14:17:44 -070083 return test_bit(BDI_writeback_running, &bdi->state);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070084}
85
Jan Kara692ebd12010-09-21 11:51:01 +020086static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
87{
88 struct super_block *sb = inode->i_sb;
Jan Kara692ebd12010-09-21 11:51:01 +020089
Christoph Hellwigaaead252010-10-04 14:25:33 +020090 if (strcmp(sb->s_type->name, "bdev") == 0)
91 return inode->i_mapping->backing_dev_info;
92
93 return sb->s_bdi;
Jan Kara692ebd12010-09-21 11:51:01 +020094}
95
Nick Piggin7ccf19a2010-10-21 11:49:30 +110096static inline struct inode *wb_inode(struct list_head *head)
97{
98 return list_entry(head, struct inode, i_wb_list);
99}
100
Jan Kara65850272011-01-13 15:45:44 -0800101/* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
102static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
Nick Piggin4195f732009-05-28 09:01:15 +0200103{
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300104 if (bdi->wb.task) {
105 wake_up_process(bdi->wb.task);
106 } else {
107 /*
108 * The bdi thread isn't there, wake up the forker thread which
109 * will create and run it.
110 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200111 wake_up_process(default_backing_dev_info.wb.task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 }
Jan Kara65850272011-01-13 15:45:44 -0800113}
114
115static void bdi_queue_work(struct backing_dev_info *bdi,
116 struct wb_writeback_work *work)
117{
118 trace_writeback_queue(bdi, work);
119
120 spin_lock_bh(&bdi->wb_lock);
121 list_add_tail(&work->list, &bdi->work_list);
122 if (!bdi->wb.task)
123 trace_writeback_nothread(bdi, work);
124 bdi_wakeup_flusher(bdi);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300125 spin_unlock_bh(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126}
127
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200128static void
129__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600130 bool range_cyclic, enum wb_reason reason)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200132 struct wb_writeback_work *work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200133
Jens Axboebcddc3f2009-09-13 20:07:36 +0200134 /*
135 * This is WB_SYNC_NONE writeback, so if allocation fails just
136 * wakeup the thread for old dirty data writeback
137 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200138 work = kzalloc(sizeof(*work), GFP_ATOMIC);
139 if (!work) {
Dave Chinner455b2862010-07-07 13:24:06 +1000140 if (bdi->wb.task) {
141 trace_writeback_nowork(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200142 wake_up_process(bdi->wb.task);
Dave Chinner455b2862010-07-07 13:24:06 +1000143 }
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200144 return;
Jens Axboebcddc3f2009-09-13 20:07:36 +0200145 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200146
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200147 work->sync_mode = WB_SYNC_NONE;
148 work->nr_pages = nr_pages;
149 work->range_cyclic = range_cyclic;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600150 work->reason = reason;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200151
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200152 bdi_queue_work(bdi, work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200153}
154
155/**
156 * bdi_start_writeback - start writeback
157 * @bdi: the backing device to write from
158 * @nr_pages: the number of pages to write
159 *
160 * Description:
161 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300162 * started when this function returns, we make no guarantees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200163 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200164 *
165 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600166void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
167 enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200168{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600169 __bdi_start_writeback(bdi, nr_pages, true, reason);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200170}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800171
Christoph Hellwigc5444192010-06-08 18:15:15 +0200172/**
173 * bdi_start_background_writeback - start background writeback
174 * @bdi: the backing device to write from
175 *
176 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800177 * This makes sure WB_SYNC_NONE background writeback happens. When
178 * this function returns, it is only guaranteed that for given BDI
179 * some IO is happening if we are over background dirty threshold.
180 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200181 */
182void bdi_start_background_writeback(struct backing_dev_info *bdi)
183{
Jan Kara65850272011-01-13 15:45:44 -0800184 /*
185 * We just wake up the flusher thread. It will perform background
186 * writeback as soon as there is no other work to do.
187 */
Wu Fengguang71927e82011-01-13 15:45:46 -0800188 trace_writeback_wake_background(bdi);
Jan Kara65850272011-01-13 15:45:44 -0800189 spin_lock_bh(&bdi->wb_lock);
190 bdi_wakeup_flusher(bdi);
191 spin_unlock_bh(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192}
193
194/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100195 * Remove the inode from the writeback list it is on.
196 */
197void inode_wb_list_del(struct inode *inode)
198{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600199 struct backing_dev_info *bdi = inode_to_bdi(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +1100200
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600201 spin_lock(&bdi->wb.list_lock);
202 list_del_init(&inode->i_wb_list);
203 spin_unlock(&bdi->wb.list_lock);
204}
Dave Chinnera66979a2011-03-22 22:23:41 +1100205
206/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700207 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
208 * furthest end of its superblock's dirty-inode list.
209 *
210 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200211 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700212 * the case then the inode must have been redirtied while it was being written
213 * out and we don't reset its dirtied_when.
214 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600215static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700216{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600217 assert_spin_locked(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200218 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200219 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700220
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100221 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200222 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700223 inode->dirtied_when = jiffies;
224 }
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100225 list_move(&inode->i_wb_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700226}
227
228/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200229 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700230 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600231static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700232{
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600233 assert_spin_locked(&wb->list_lock);
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100234 list_move(&inode->i_wb_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700235}
236
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700237static void inode_sync_complete(struct inode *inode)
238{
239 /*
Dave Chinnera66979a2011-03-22 22:23:41 +1100240 * Prevent speculative execution through
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600241 * spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700242 */
Dave Chinnera66979a2011-03-22 22:23:41 +1100243
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700244 smp_mb();
245 wake_up_bit(&inode->i_state, __I_SYNC);
246}
247
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700248static bool inode_dirtied_after(struct inode *inode, unsigned long t)
249{
250 bool ret = time_after(inode->dirtied_when, t);
251#ifndef CONFIG_64BIT
252 /*
253 * For inodes being constantly redirtied, dirtied_when can get stuck.
254 * It _appears_ to be in the future, but is actually in distant past.
255 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200256 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700257 */
258 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
259#endif
260 return ret;
261}
262
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700263/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700264 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
265 */
Wu Fengguange84d0a42011-04-23 12:27:27 -0600266static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -0700267 struct list_head *dispatch_queue,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600268 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700269{
Shaohua Li5c034492009-09-24 14:42:33 +0200270 LIST_HEAD(tmp);
271 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200272 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200273 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200274 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -0600275 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200276
Fengguang Wu2c136572007-10-16 23:30:39 -0700277 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100278 inode = wb_inode(delaying_queue->prev);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600279 if (work->older_than_this &&
280 inode_dirtied_after(inode, *work->older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700281 break;
Jens Axboecf137302009-09-24 15:12:57 +0200282 if (sb && sb != inode->i_sb)
283 do_sb_sort = 1;
284 sb = inode->i_sb;
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100285 list_move(&inode->i_wb_list, &tmp);
Wu Fengguange84d0a42011-04-23 12:27:27 -0600286 moved++;
Shaohua Li5c034492009-09-24 14:42:33 +0200287 }
288
Jens Axboecf137302009-09-24 15:12:57 +0200289 /* just one sb in list, splice to dispatch_queue and we're done */
290 if (!do_sb_sort) {
291 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -0600292 goto out;
Jens Axboecf137302009-09-24 15:12:57 +0200293 }
294
Shaohua Li5c034492009-09-24 14:42:33 +0200295 /* Move inodes from one superblock together */
296 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100297 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200298 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100299 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +0200300 if (inode->i_sb == sb)
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100301 list_move(&inode->i_wb_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +0200302 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700303 }
Wu Fengguange84d0a42011-04-23 12:27:27 -0600304out:
305 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -0700306}
307
308/*
309 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700310 * Before
311 * newly dirtied b_dirty b_io b_more_io
312 * =============> gf edc BA
313 * After
314 * newly dirtied b_dirty b_io b_more_io
315 * =============> g fBAedc
316 * |
317 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -0700318 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600319static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -0700320{
Wu Fengguange84d0a42011-04-23 12:27:27 -0600321 int moved;
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600322 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700323 list_splice_init(&wb->b_more_io, &wb->b_io);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600324 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
325 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200326}
327
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100328static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200329{
Jens Axboe03ba3782009-09-09 09:08:54 +0200330 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100331 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200332 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700333}
334
335/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200336 * Wait for writeback on an inode to complete.
337 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600338static void inode_wait_for_writeback(struct inode *inode,
339 struct bdi_writeback *wb)
Christoph Hellwig01c03192009-06-08 13:35:40 +0200340{
341 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
342 wait_queue_head_t *wqh;
343
344 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +1100345 while (inode->i_state & I_SYNC) {
346 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600347 spin_unlock(&wb->list_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +0200348 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600349 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +1100350 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d2010-05-24 14:32:38 -0700351 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200352}
353
354/*
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600355 * Write out an inode's dirty pages. Called under wb->list_lock and
Dave Chinner0f1b1fd2011-03-22 22:23:43 +1100356 * inode->i_lock. Either the caller has an active reference on the inode or
357 * the inode has I_WILL_FREE set.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200358 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 * If `wait' is set, wait on the writeout.
360 *
361 * The whole writeout design is quite complex and fragile. We want to avoid
362 * starvation of particular inodes when others are being redirtied, prevent
363 * livelocks, etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 */
365static int
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600366writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
367 struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -0600370 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200371 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 int ret;
373
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600374 assert_spin_locked(&wb->list_lock);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +1100375 assert_spin_locked(&inode->i_lock);
376
Christoph Hellwig01c03192009-06-08 13:35:40 +0200377 if (!atomic_read(&inode->i_count))
378 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
379 else
380 WARN_ON(inode->i_state & I_WILL_FREE);
381
382 if (inode->i_state & I_SYNC) {
383 /*
384 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200385 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200386 * writeback can proceed with the other inodes on s_io.
387 *
388 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200389 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200390 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100391 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600392 requeue_io(inode, wb);
Wu Fengguang251d6a42010-12-01 17:33:37 -0600393 trace_writeback_single_inode_requeue(inode, wbc,
394 nr_to_write);
Christoph Hellwig01c03192009-06-08 13:35:40 +0200395 return 0;
396 }
397
398 /*
399 * It's a data-integrity sync. We must wait.
400 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600401 inode_wait_for_writeback(inode, wb);
Christoph Hellwig01c03192009-06-08 13:35:40 +0200402 }
403
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700404 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400406 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700407 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400408 inode->i_state &= ~I_DIRTY_PAGES;
Dave Chinner250df6e2011-03-22 22:23:36 +1100409 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600410 spin_unlock(&wb->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
412 ret = do_writepages(mapping, wbc);
413
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100414 /*
415 * Make sure to wait on the data before writing out the metadata.
416 * This is important for filesystems that modify metadata on data
417 * I/O completion.
418 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100419 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100420 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 if (ret == 0)
422 ret = err;
423 }
424
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400425 /*
426 * Some filesystems may redirty the inode during the writeback
427 * due to delalloc, clear dirty metadata flags right before
428 * write_inode()
429 */
Dave Chinner250df6e2011-03-22 22:23:36 +1100430 spin_lock(&inode->i_lock);
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400431 dirty = inode->i_state & I_DIRTY;
432 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +1100433 spin_unlock(&inode->i_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100434 /* Don't write the inode if only I_DIRTY_PAGES was set */
435 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100436 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 if (ret == 0)
438 ret = err;
439 }
440
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600441 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +1100442 spin_lock(&inode->i_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700443 inode->i_state &= ~I_SYNC;
Al Viroa4ffdde2010-06-02 17:38:30 -0400444 if (!(inode->i_state & I_FREEING)) {
Wu Fengguang94c3dcb2011-04-27 19:05:21 -0600445 /*
446 * Sync livelock prevention. Each inode is tagged and synced in
447 * one shot. If still dirty, it will be redirty_tail()'ed below.
448 * Update the dirty time to prevent enqueue and sync it again.
449 */
450 if ((inode->i_state & I_DIRTY) &&
451 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
452 inode->dirtied_when = jiffies;
453
Wu Fengguang23539af2010-08-11 14:17:41 -0700454 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 /*
456 * We didn't write back all the pages. nfs_writepages()
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700457 * sometimes bales out without doing anything.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700458 */
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700459 inode->i_state |= I_DIRTY_PAGES;
460 if (wbc->nr_to_write <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700462 * slice used up: queue for next turn
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600464 requeue_io(inode, wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 } else {
466 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700467 * Writeback blocked by something other than
468 * congestion. Delay the inode for some time to
469 * avoid spinning on the CPU (100% iowait)
470 * retrying writeback of the dirty page/inode
471 * that cannot be performed immediately.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600473 redirty_tail(inode, wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 }
Wu Fengguang23539af2010-08-11 14:17:41 -0700475 } else if (inode->i_state & I_DIRTY) {
476 /*
477 * Filesystems can dirty the inode during writeback
478 * operations, such as delayed allocation during
479 * submission or metadata updates after data IO
480 * completion.
481 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600482 redirty_tail(inode, wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 } else {
484 /*
Nick Piggin9e38d862010-10-23 06:55:17 -0400485 * The inode is clean. At this point we either have
486 * a reference to the inode or it's on it's way out.
487 * No need to add it back to the LRU.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 */
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100489 list_del_init(&inode->i_wb_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 }
491 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700492 inode_sync_complete(inode);
Wu Fengguang251d6a42010-12-01 17:33:37 -0600493 trace_writeback_single_inode(inode, wbc, nr_to_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 return ret;
495}
496
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600497static long writeback_chunk_size(struct backing_dev_info *bdi,
498 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600499{
500 long pages;
501
502 /*
503 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
504 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
505 * here avoids calling into writeback_inodes_wb() more than once.
506 *
507 * The intended call sequence for WB_SYNC_ALL writeback is:
508 *
509 * wb_writeback()
510 * writeback_sb_inodes() <== called only once
511 * write_cache_pages() <== called once for each inode
512 * (quickly) tag currently dirty pages
513 * (maybe slowly) sync all tagged pages
514 */
515 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
516 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600517 else {
518 pages = min(bdi->avg_write_bandwidth / 2,
519 global_dirty_limit / DIRTY_SCOPE);
520 pages = min(pages, work->nr_pages);
521 pages = round_down(pages + MIN_WRITEBACK_PAGES,
522 MIN_WRITEBACK_PAGES);
523 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600524
525 return pages;
526}
527
Jens Axboe03ba3782009-09-09 09:08:54 +0200528/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800529 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200530 *
531 * If @only_this_sb is true, then find and write all such
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800532 * inodes. Otherwise write only ones which go sequentially
533 * in reverse order.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200534 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600535 * Return the number of pages and/or inodes written.
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800536 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600537static long writeback_sb_inodes(struct super_block *sb,
538 struct bdi_writeback *wb,
539 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200540{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600541 struct writeback_control wbc = {
542 .sync_mode = work->sync_mode,
543 .tagged_writepages = work->tagged_writepages,
544 .for_kupdate = work->for_kupdate,
545 .for_background = work->for_background,
546 .range_cyclic = work->range_cyclic,
547 .range_start = 0,
548 .range_end = LLONG_MAX,
549 };
550 unsigned long start_time = jiffies;
551 long write_chunk;
552 long wrote = 0; /* count both pages and inodes */
553
Jens Axboe03ba3782009-09-09 09:08:54 +0200554 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100555 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200556
557 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600558 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200559 /*
560 * We only want to write back data for this
561 * superblock, move all inodes not belonging
562 * to it back onto the dirty list.
563 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600564 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200565 continue;
566 }
567
568 /*
569 * The inode belongs to a different superblock.
570 * Bounce back to the caller to unpin this and
571 * pin the next superblock.
572 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600573 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200574 }
575
Christoph Hellwig9843b762010-10-24 19:40:46 +0200576 /*
577 * Don't bother with new inodes or inodes beeing freed, first
578 * kind does not need peridic writeout yet, and for the latter
579 * kind writeout is handled by the freer.
580 */
Dave Chinner250df6e2011-03-22 22:23:36 +1100581 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +0200582 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +1100583 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -0700584 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -0700585 continue;
586 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 __iget(inode);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -0600588 write_chunk = writeback_chunk_size(wb->bdi, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600589 wbc.nr_to_write = write_chunk;
590 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +1100591
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600592 writeback_single_inode(inode, wb, &wbc);
593
594 work->nr_pages -= write_chunk - wbc.nr_to_write;
595 wrote += write_chunk - wbc.nr_to_write;
596 if (!(inode->i_state & I_DIRTY))
597 wrote++;
598 if (wbc.pages_skipped) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 /*
600 * writeback is not making progress due to locked
601 * buffers. Skip this inode for now.
602 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600603 redirty_tail(inode, wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 }
Dave Chinner0f1b1fd2011-03-22 22:23:43 +1100605 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600606 spin_unlock(&wb->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800608 cond_resched();
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600609 spin_lock(&wb->list_lock);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600610 /*
611 * bail out to wb_writeback() often enough to check
612 * background threshold and other termination conditions.
613 */
614 if (wrote) {
615 if (time_is_before_jiffies(start_time + HZ / 10UL))
616 break;
617 if (work->nr_pages <= 0)
618 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800619 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600621 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800622}
Nick Piggin38f21972009-01-06 14:40:25 -0800623
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600624static long __writeback_inodes_wb(struct bdi_writeback *wb,
625 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800626{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600627 unsigned long start_time = jiffies;
628 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800629
630 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100631 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800632 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800633
Dave Chinner12ad3ab2011-07-08 14:14:41 +1000634 if (!grab_super_passive(sb)) {
Wu Fengguang0e995812011-07-29 22:14:35 -0600635 /*
636 * grab_super_passive() may fail consistently due to
637 * s_umount being grabbed by someone else. Don't use
638 * requeue_io() to avoid busy retrying the inode/sb.
639 */
640 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200641 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800642 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600643 wrote += writeback_sb_inodes(sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200644 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800645
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600646 /* refer to the same tests at the end of writeback_sb_inodes */
647 if (wrote) {
648 if (time_is_before_jiffies(start_time + HZ / 10UL))
649 break;
650 if (work->nr_pages <= 0)
651 break;
652 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800653 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200654 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600655 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200656}
657
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600658long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
659 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200660{
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600661 struct wb_writeback_work work = {
662 .nr_pages = nr_pages,
663 .sync_mode = WB_SYNC_NONE,
664 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600665 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600666 };
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200667
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600668 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -0600669 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600670 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600671 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600672 spin_unlock(&wb->list_lock);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200673
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600674 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200675}
676
Wu Fengguangb00949a2010-11-18 14:38:33 -0600677static bool over_bground_thresh(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200678{
679 unsigned long background_thresh, dirty_thresh;
680
Wu Fengguang16c40422010-08-11 14:17:39 -0700681 global_dirty_limits(&background_thresh, &dirty_thresh);
Jens Axboe03ba3782009-09-09 09:08:54 +0200682
Wu Fengguangb00949a2010-11-18 14:38:33 -0600683 if (global_page_state(NR_FILE_DIRTY) +
684 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
685 return true;
686
687 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
688 bdi_dirty_limit(bdi, background_thresh))
689 return true;
690
691 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200692}
693
694/*
Wu Fengguange98be2d2010-08-29 11:22:30 -0600695 * Called under wb->list_lock. If there are multiple wb per bdi,
696 * only the flusher working on the first wb should do it.
697 */
698static void wb_update_bandwidth(struct bdi_writeback *wb,
699 unsigned long start_time)
700{
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600701 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600702}
703
704/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200705 * Explicit flushing or periodic writeback of "old" data.
706 *
707 * Define "old": the first time one of an inode's pages is dirtied, we mark the
708 * dirtying-time in the inode's address_space. So this periodic writeback code
709 * just walks the superblock inode list, writing back any inodes which are
710 * older than a specific point in time.
711 *
712 * Try to run once per dirty_writeback_interval. But if a writeback event
713 * takes longer than a dirty_writeback_interval interval, then leave a
714 * one-second gap.
715 *
716 * older_than_this takes precedence over nr_to_write. So we'll only write back
717 * all dirty pages if they are all attached to "old" mappings.
718 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200719static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200720 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200721{
Wu Fengguange98be2d2010-08-29 11:22:30 -0600722 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600723 long nr_pages = work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +0200724 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +0200725 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600726 long progress;
Jens Axboe03ba3782009-09-09 09:08:54 +0200727
Wu Fengguange185dda2011-04-23 11:26:07 -0600728 oldest_jif = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600729 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +0200730
Wu Fengguange8dfc302011-04-21 12:06:32 -0600731 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200732 for (;;) {
733 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800734 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200735 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200736 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200737 break;
738
739 /*
Jan Karaaa373cf2011-01-13 15:45:47 -0800740 * Background writeout and kupdate-style writeback may
741 * run forever. Stop them if there is other work to do
742 * so that e.g. sync can proceed. They'll be restarted
743 * after the other works are all done.
744 */
745 if ((work->for_background || work->for_kupdate) &&
746 !list_empty(&wb->bdi->work_list))
747 break;
748
749 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800750 * For background writeout, stop when we are below the
751 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200752 */
Wu Fengguangb00949a2010-11-18 14:38:33 -0600753 if (work->for_background && !over_bground_thresh(wb->bdi))
Jens Axboe03ba3782009-09-09 09:08:54 +0200754 break;
755
Wu Fengguangba9aa832010-07-21 20:32:30 -0600756 if (work->for_kupdate) {
757 oldest_jif = jiffies -
758 msecs_to_jiffies(dirty_expire_interval * 10);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600759 work->older_than_this = &oldest_jif;
Wu Fengguangba9aa832010-07-21 20:32:30 -0600760 }
Dave Chinner028c2dd2010-07-07 13:24:07 +1000761
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600762 trace_writeback_start(wb->bdi, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -0600763 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -0600764 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200765 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600766 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200767 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600768 progress = __writeback_inodes_wb(wb, work);
769 trace_writeback_written(wb->bdi, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000770
Wu Fengguange98be2d2010-08-29 11:22:30 -0600771 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +0200772
773 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200774 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -0600775 *
776 * Dirty inodes are moved to b_io for writeback in batches.
777 * The completion of the current batch does not necessarily
778 * mean the overall work is done. So we keep looping as long
779 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +0200780 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600781 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +0200782 continue;
783 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -0600784 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +0200785 */
Wu Fengguangb7a24412010-07-21 22:19:51 -0600786 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +0200787 break;
788 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +0200789 * Nothing written. Wait for some inode to
790 * become available for writeback. Otherwise
791 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +0200792 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200793 if (!list_empty(&wb->b_more_io)) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600794 trace_writeback_wait(wb->bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200795 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +1100796 spin_lock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600797 inode_wait_for_writeback(inode, wb);
Dave Chinner250df6e2011-03-22 22:23:36 +1100798 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200799 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200800 }
Wu Fengguange8dfc302011-04-21 12:06:32 -0600801 spin_unlock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200802
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600803 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +0200804}
805
806/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200807 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +0200808 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200809static struct wb_writeback_work *
Minchan Kim08852b62010-08-03 12:51:16 +0200810get_next_work_item(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200811{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200812 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200813
Artem Bityutskiy64677162010-07-25 14:29:22 +0300814 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200815 if (!list_empty(&bdi->work_list)) {
816 work = list_entry(bdi->work_list.next,
817 struct wb_writeback_work, list);
818 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200819 }
Artem Bityutskiy64677162010-07-25 14:29:22 +0300820 spin_unlock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200821 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200822}
823
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700824/*
825 * Add in the number of potentially dirty inodes, because each inode
826 * write can dirty pagecache in the underlying blockdev.
827 */
828static unsigned long get_nr_dirty_pages(void)
829{
830 return global_page_state(NR_FILE_DIRTY) +
831 global_page_state(NR_UNSTABLE_NFS) +
832 get_nr_dirty_inodes();
833}
834
Jan Kara65850272011-01-13 15:45:44 -0800835static long wb_check_background_flush(struct bdi_writeback *wb)
836{
Wu Fengguangb00949a2010-11-18 14:38:33 -0600837 if (over_bground_thresh(wb->bdi)) {
Jan Kara65850272011-01-13 15:45:44 -0800838
839 struct wb_writeback_work work = {
840 .nr_pages = LONG_MAX,
841 .sync_mode = WB_SYNC_NONE,
842 .for_background = 1,
843 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600844 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -0800845 };
846
847 return wb_writeback(wb, &work);
848 }
849
850 return 0;
851}
852
Jens Axboe03ba3782009-09-09 09:08:54 +0200853static long wb_check_old_data_flush(struct bdi_writeback *wb)
854{
855 unsigned long expired;
856 long nr_pages;
857
Jens Axboe69b62d02010-05-17 12:51:03 +0200858 /*
859 * When set to zero, disable periodic writeback
860 */
861 if (!dirty_writeback_interval)
862 return 0;
863
Jens Axboe03ba3782009-09-09 09:08:54 +0200864 expired = wb->last_old_flush +
865 msecs_to_jiffies(dirty_writeback_interval * 10);
866 if (time_before(jiffies, expired))
867 return 0;
868
869 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700870 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +0200871
Jens Axboec4a77a62009-09-16 15:18:25 +0200872 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200873 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200874 .nr_pages = nr_pages,
875 .sync_mode = WB_SYNC_NONE,
876 .for_kupdate = 1,
877 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600878 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +0200879 };
880
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200881 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +0200882 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200883
884 return 0;
885}
886
887/*
888 * Retrieve work items and do the writeback they describe
889 */
890long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
891{
892 struct backing_dev_info *bdi = wb->bdi;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200893 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200894 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200895
Jan Kara81d73a32010-08-11 14:17:44 -0700896 set_bit(BDI_writeback_running, &wb->bdi->state);
Minchan Kim08852b62010-08-03 12:51:16 +0200897 while ((work = get_next_work_item(bdi)) != NULL) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200898 /*
899 * Override sync mode, in case we must wait for completion
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200900 * because this thread is exiting now.
Jens Axboe03ba3782009-09-09 09:08:54 +0200901 */
902 if (force_wait)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200903 work->sync_mode = WB_SYNC_ALL;
904
Dave Chinner455b2862010-07-07 13:24:06 +1000905 trace_writeback_exec(bdi, work);
906
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200907 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200908
909 /*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200910 * Notify the caller of completion if this is a synchronous
911 * work item, otherwise just free it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200912 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200913 if (work->done)
914 complete(work->done);
915 else
916 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200917 }
918
919 /*
920 * Check for periodic writeback, kupdated() style
921 */
922 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -0800923 wrote += wb_check_background_flush(wb);
Jan Kara81d73a32010-08-11 14:17:44 -0700924 clear_bit(BDI_writeback_running, &wb->bdi->state);
Jens Axboe03ba3782009-09-09 09:08:54 +0200925
926 return wrote;
927}
928
929/*
930 * Handle writeback of dirty data for the device backed by this bdi. Also
931 * wakes up periodically and does kupdated style flushing.
932 */
Christoph Hellwig08243902010-06-19 23:08:22 +0200933int bdi_writeback_thread(void *data)
Jens Axboe03ba3782009-09-09 09:08:54 +0200934{
Christoph Hellwig08243902010-06-19 23:08:22 +0200935 struct bdi_writeback *wb = data;
936 struct backing_dev_info *bdi = wb->bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200937 long pages_written;
938
Peter Zijlstra766f9162010-10-26 14:22:45 -0700939 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +0200940 set_freezable();
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300941 wb->last_active = jiffies;
Christoph Hellwig08243902010-06-19 23:08:22 +0200942
943 /*
944 * Our parent may run at a different priority, just set us to normal
945 */
946 set_user_nice(current, 0);
947
Dave Chinner455b2862010-07-07 13:24:06 +1000948 trace_writeback_thread_start(bdi);
949
Jens Axboe03ba3782009-09-09 09:08:54 +0200950 while (!kthread_should_stop()) {
Artem Bityutskiy64677162010-07-25 14:29:22 +0300951 /*
952 * Remove own delayed wake-up timer, since we are already awake
953 * and we'll take care of the preriodic write-back.
954 */
955 del_timer(&wb->wakeup_timer);
956
Jens Axboe03ba3782009-09-09 09:08:54 +0200957 pages_written = wb_do_writeback(wb, 0);
958
Dave Chinner455b2862010-07-07 13:24:06 +1000959 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +0200960
Jens Axboe03ba3782009-09-09 09:08:54 +0200961 if (pages_written)
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300962 wb->last_active = jiffies;
Jens Axboe03ba3782009-09-09 09:08:54 +0200963
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300964 set_current_state(TASK_INTERRUPTIBLE);
J. Bruce Fieldsb76b4012010-08-28 08:52:10 +0200965 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300966 __set_current_state(TASK_RUNNING);
967 continue;
Jens Axboe03ba3782009-09-09 09:08:54 +0200968 }
969
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300970 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300971 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300972 else {
973 /*
974 * We have nothing to do, so can go sleep without any
975 * timeout and save power. When a work is queued or
976 * something is made dirty - we will be woken up.
977 */
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300978 schedule();
Jens Axboef9eadbb2010-05-18 14:31:45 +0200979 }
Jens Axboe69b62d02010-05-17 12:51:03 +0200980
Jens Axboe03ba3782009-09-09 09:08:54 +0200981 try_to_freeze();
982 }
983
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300984 /* Flush any work that raced with us exiting */
Christoph Hellwig08243902010-06-19 23:08:22 +0200985 if (!list_empty(&bdi->work_list))
986 wb_do_writeback(wb, 1);
Dave Chinner455b2862010-07-07 13:24:06 +1000987
988 trace_writeback_thread_stop(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +0200989 return 0;
990}
991
Christoph Hellwig08243902010-06-19 23:08:22 +0200992
Jens Axboe03ba3782009-09-09 09:08:54 +0200993/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200994 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
995 * the whole world.
996 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600997void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +0200998{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200999 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001000
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001001 if (!nr_pages) {
1002 nr_pages = global_page_state(NR_FILE_DIRTY) +
Jens Axboe03ba3782009-09-09 09:08:54 +02001003 global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001004 }
1005
1006 rcu_read_lock();
1007 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1008 if (!bdi_has_dirty_io(bdi))
1009 continue;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001010 __bdi_start_writeback(bdi, nr_pages, false, reason);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001011 }
1012 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001013}
1014
1015static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1016{
1017 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1018 struct dentry *dentry;
1019 const char *name = "?";
1020
1021 dentry = d_find_alias(inode);
1022 if (dentry) {
1023 spin_lock(&dentry->d_lock);
1024 name = (const char *) dentry->d_name.name;
1025 }
1026 printk(KERN_DEBUG
1027 "%s(%d): dirtied inode %lu (%s) on %s\n",
1028 current->comm, task_pid_nr(current), inode->i_ino,
1029 name, inode->i_sb->s_id);
1030 if (dentry) {
1031 spin_unlock(&dentry->d_lock);
1032 dput(dentry);
1033 }
1034 }
1035}
1036
1037/**
1038 * __mark_inode_dirty - internal function
1039 * @inode: inode to mark
1040 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1041 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1042 * mark_inode_dirty_sync.
1043 *
1044 * Put the inode on the super block's dirty list.
1045 *
1046 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1047 * dirty list only if it is hashed or if it refers to a blockdev.
1048 * If it was not hashed, it will never be added to the dirty list
1049 * even if it is later hashed, as it will have been marked dirty already.
1050 *
1051 * In short, make sure you hash any inodes _before_ you start marking
1052 * them dirty.
1053 *
Jens Axboe03ba3782009-09-09 09:08:54 +02001054 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1055 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1056 * the kernel-internal blockdev inode represents the dirtying time of the
1057 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1058 * page->mapping->host, so the page-dirtying time is recorded in the internal
1059 * blockdev inode.
1060 */
1061void __mark_inode_dirty(struct inode *inode, int flags)
1062{
1063 struct super_block *sb = inode->i_sb;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001064 struct backing_dev_info *bdi = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001065
1066 /*
1067 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1068 * dirty the inode itself
1069 */
1070 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1071 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04001072 sb->s_op->dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02001073 }
1074
1075 /*
1076 * make sure that changes are seen by all cpus before we test i_state
1077 * -- mikulas
1078 */
1079 smp_mb();
1080
1081 /* avoid the locking if we can */
1082 if ((inode->i_state & flags) == flags)
1083 return;
1084
1085 if (unlikely(block_dump))
1086 block_dump___mark_inode_dirty(inode);
1087
Dave Chinner250df6e2011-03-22 22:23:36 +11001088 spin_lock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001089 if ((inode->i_state & flags) != flags) {
1090 const int was_dirty = inode->i_state & I_DIRTY;
1091
1092 inode->i_state |= flags;
1093
1094 /*
1095 * If the inode is being synced, just update its dirty state.
1096 * The unlocker will place the inode on the appropriate
1097 * superblock list, based upon its state.
1098 */
1099 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11001100 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001101
1102 /*
1103 * Only add valid (hashed) inodes to the superblock's
1104 * dirty list. Add blockdev inodes as well.
1105 */
1106 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382c2010-10-23 15:19:20 -04001107 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11001108 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001109 }
Al Viroa4ffdde2010-06-02 17:38:30 -04001110 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11001111 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02001112
1113 /*
1114 * If the inode was already on b_dirty/b_io/b_more_io, don't
1115 * reposition it (that would break b_dirty time-ordering).
1116 */
1117 if (!was_dirty) {
Dave Chinnera66979a2011-03-22 22:23:41 +11001118 bool wakeup_bdi = false;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001119 bdi = inode_to_bdi(inode);
Jens Axboe500b0672009-09-09 09:10:25 +02001120
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001121 if (bdi_cap_writeback_dirty(bdi)) {
1122 WARN(!test_bit(BDI_registered, &bdi->state),
1123 "bdi-%s not registered\n", bdi->name);
1124
1125 /*
1126 * If this is the first dirty inode for this
1127 * bdi, we have to wake-up the corresponding
1128 * bdi thread to make sure background
1129 * write-back happens later.
1130 */
1131 if (!wb_has_dirty_io(&bdi->wb))
1132 wakeup_bdi = true;
Jens Axboe500b0672009-09-09 09:10:25 +02001133 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001134
Dave Chinnera66979a2011-03-22 22:23:41 +11001135 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001136 spin_lock(&bdi->wb.list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001137 inode->dirtied_when = jiffies;
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001138 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001139 spin_unlock(&bdi->wb.list_lock);
Dave Chinnera66979a2011-03-22 22:23:41 +11001140
1141 if (wakeup_bdi)
1142 bdi_wakeup_thread_delayed(bdi);
1143 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02001144 }
1145 }
Dave Chinner250df6e2011-03-22 22:23:36 +11001146out_unlock_inode:
1147 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001148
Jens Axboe03ba3782009-09-09 09:08:54 +02001149}
1150EXPORT_SYMBOL(__mark_inode_dirty);
1151
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001152/*
1153 * Write out a superblock's list of dirty inodes. A wait will be performed
1154 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1155 *
1156 * If older_than_this is non-NULL, then only write out inodes which
1157 * had their first dirtying at a time earlier than *older_than_this.
1158 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001159 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1160 * This function assumes that the blockdev superblock's inodes are backed by
1161 * a variety of queues, so all inodes are searched. For other superblocks,
1162 * assume that all inodes are backed by the same queue.
1163 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001164 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1165 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1166 * on the writer throttling path, and we get decent balancing between many
1167 * throttled threads: we don't want them all piling up on inode_sync_wait.
1168 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001169static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001170{
Jens Axboe03ba3782009-09-09 09:08:54 +02001171 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001172
Jens Axboe03ba3782009-09-09 09:08:54 +02001173 /*
1174 * We need to be protected against the filesystem going from
1175 * r/o to r/w or vice versa.
1176 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001177 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001178
Dave Chinner55fa6092011-03-22 22:23:40 +11001179 spin_lock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001180
1181 /*
1182 * Data integrity sync. Must wait for all pages under writeback,
1183 * because there may have been pages dirtied before our sync
1184 * call, but which had writeout started before we write it out.
1185 * In which case, the inode may not be on the dirty list, but
1186 * we still have to wait for that writeout.
1187 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001188 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001189 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02001190
Dave Chinner250df6e2011-03-22 22:23:36 +11001191 spin_lock(&inode->i_lock);
1192 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1193 (mapping->nrpages == 0)) {
1194 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001195 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11001196 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001197 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11001198 spin_unlock(&inode->i_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11001199 spin_unlock(&inode_sb_list_lock);
1200
Jens Axboe03ba3782009-09-09 09:08:54 +02001201 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11001202 * We hold a reference to 'inode' so it couldn't have been
1203 * removed from s_inodes list while we dropped the
1204 * inode_sb_list_lock. We cannot iput the inode now as we can
1205 * be holding the last reference and we cannot iput it under
1206 * inode_sb_list_lock. So we keep the reference and iput it
1207 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02001208 */
1209 iput(old_inode);
1210 old_inode = inode;
1211
1212 filemap_fdatawait(mapping);
1213
1214 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001215
Dave Chinner55fa6092011-03-22 22:23:40 +11001216 spin_lock(&inode_sb_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001217 }
Dave Chinner55fa6092011-03-22 22:23:40 +11001218 spin_unlock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001219 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220}
1221
Jens Axboed8a85592009-09-02 12:34:32 +02001222/**
Chris Mason3259f8b2010-10-29 11:16:17 -04001223 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
Jens Axboed8a85592009-09-02 12:34:32 +02001224 * @sb: the superblock
Chris Mason3259f8b2010-10-29 11:16:17 -04001225 * @nr: the number of pages to write
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 *
Jens Axboed8a85592009-09-02 12:34:32 +02001227 * Start writeback on some inodes on this super_block. No guarantees are made
1228 * on how many (if any) will be written, and this function does not wait
Chris Mason3259f8b2010-10-29 11:16:17 -04001229 * for IO completion of submitted IO.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001231void writeback_inodes_sb_nr(struct super_block *sb,
1232 unsigned long nr,
1233 enum wb_reason reason)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001235 DECLARE_COMPLETION_ONSTACK(done);
1236 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001237 .sb = sb,
1238 .sync_mode = WB_SYNC_NONE,
1239 .tagged_writepages = 1,
1240 .done = &done,
1241 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001242 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001243 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001244
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001245 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001246 bdi_queue_work(sb->s_bdi, &work);
1247 wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248}
Chris Mason3259f8b2010-10-29 11:16:17 -04001249EXPORT_SYMBOL(writeback_inodes_sb_nr);
1250
1251/**
1252 * writeback_inodes_sb - writeback dirty inodes from given super_block
1253 * @sb: the superblock
1254 *
1255 * Start writeback on some inodes on this super_block. No guarantees are made
1256 * on how many (if any) will be written, and this function does not wait
1257 * for IO completion of submitted IO.
1258 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001259void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04001260{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001261 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04001262}
Jens Axboed8a85592009-09-02 12:34:32 +02001263EXPORT_SYMBOL(writeback_inodes_sb);
1264
1265/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001266 * writeback_inodes_sb_if_idle - start writeback if none underway
1267 * @sb: the superblock
1268 *
1269 * Invoke writeback_inodes_sb if no writeback is currently underway.
1270 * Returns 1 if writeback was started, 0 if not.
1271 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001272int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001273{
1274 if (!writeback_in_progress(sb->s_bdi)) {
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001275 down_read(&sb->s_umount);
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001276 writeback_inodes_sb(sb, reason);
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001277 up_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001278 return 1;
1279 } else
1280 return 0;
1281}
1282EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1283
1284/**
Chris Mason3259f8b2010-10-29 11:16:17 -04001285 * writeback_inodes_sb_if_idle - start writeback if none underway
1286 * @sb: the superblock
1287 * @nr: the number of pages to write
1288 *
1289 * Invoke writeback_inodes_sb if no writeback is currently underway.
1290 * Returns 1 if writeback was started, 0 if not.
1291 */
1292int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001293 unsigned long nr,
1294 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04001295{
1296 if (!writeback_in_progress(sb->s_bdi)) {
1297 down_read(&sb->s_umount);
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001298 writeback_inodes_sb_nr(sb, nr, reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04001299 up_read(&sb->s_umount);
1300 return 1;
1301 } else
1302 return 0;
1303}
1304EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1305
1306/**
Jens Axboed8a85592009-09-02 12:34:32 +02001307 * sync_inodes_sb - sync sb inode pages
1308 * @sb: the superblock
1309 *
1310 * This function writes and waits on any dirty inode belonging to this
Stefan Hajnoczicb9ef8d2011-01-13 15:47:26 -08001311 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02001312 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001313void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001314{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001315 DECLARE_COMPLETION_ONSTACK(done);
1316 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001317 .sb = sb,
1318 .sync_mode = WB_SYNC_ALL,
1319 .nr_pages = LONG_MAX,
1320 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001321 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001322 .reason = WB_REASON_SYNC,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001323 };
1324
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001325 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1326
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001327 bdi_queue_work(sb->s_bdi, &work);
1328 wait_for_completion(&done);
1329
Jens Axboeb6e51312009-09-16 15:13:54 +02001330 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001331}
1332EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001335 * write_inode_now - write an inode to disk
1336 * @inode: inode to write to disk
1337 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001339 * This function commits an inode to disk immediately if it is dirty. This is
1340 * primarily needed by knfsd.
1341 *
1342 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344int write_inode_now(struct inode *inode, int sync)
1345{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001346 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 int ret;
1348 struct writeback_control wbc = {
1349 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001350 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001351 .range_start = 0,
1352 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 };
1354
1355 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001356 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357
1358 might_sleep();
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001359 spin_lock(&wb->list_lock);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001360 spin_lock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001361 ret = writeback_single_inode(inode, wb, &wbc);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001362 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001363 spin_unlock(&wb->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001365 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 return ret;
1367}
1368EXPORT_SYMBOL(write_inode_now);
1369
1370/**
1371 * sync_inode - write an inode and its pages to disk.
1372 * @inode: the inode to sync
1373 * @wbc: controls the writeback mode
1374 *
1375 * sync_inode() will write an inode and its pages to disk. It will also
1376 * correctly update the inode on its superblock's dirty inode lists and will
1377 * update inode->i_state.
1378 *
1379 * The caller must have a ref on the inode.
1380 */
1381int sync_inode(struct inode *inode, struct writeback_control *wbc)
1382{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001383 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 int ret;
1385
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001386 spin_lock(&wb->list_lock);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001387 spin_lock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001388 ret = writeback_single_inode(inode, wb, wbc);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001389 spin_unlock(&inode->i_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001390 spin_unlock(&wb->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 return ret;
1392}
1393EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc3765012010-10-06 10:48:20 +02001394
1395/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08001396 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc3765012010-10-06 10:48:20 +02001397 * @inode: the inode to sync
1398 * @wait: wait for I/O to complete.
1399 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08001400 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc3765012010-10-06 10:48:20 +02001401 *
1402 * Note: only writes the actual inode, no associated data or other metadata.
1403 */
1404int sync_inode_metadata(struct inode *inode, int wait)
1405{
1406 struct writeback_control wbc = {
1407 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1408 .nr_to_write = 0, /* metadata-only */
1409 };
1410
1411 return sync_inode(inode, &wbc);
1412}
1413EXPORT_SYMBOL(sync_inode_metadata);