blob: 8cc06d5432b53d5eadec4a998197a575bbacfcec [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>
David Howells07f3f052006-09-30 20:52:18 +020029#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070030
Jens Axboe66f3b8e2009-09-02 09:19:46 +020031#define inode_to_bdi(inode) ((inode)->i_mapping->backing_dev_info)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070032
Jens Axboe03ba3782009-09-09 09:08:54 +020033/*
Jens Axboed0bceac2009-05-18 08:20:32 +020034 * We don't actually have pdflush, but this one is exported though /proc...
35 */
36int nr_pdflush_threads;
37
38/*
Jens Axboec4a77a62009-09-16 15:18:25 +020039 * Passed into wb_writeback(), essentially a subset of writeback_control
40 */
41struct wb_writeback_args {
42 long nr_pages;
43 struct super_block *sb;
44 enum writeback_sync_modes sync_mode;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050045 unsigned int for_kupdate:1;
46 unsigned int range_cyclic:1;
47 unsigned int for_background:1;
Jens Axboec4a77a62009-09-16 15:18:25 +020048};
49
50/*
Jens Axboe03ba3782009-09-09 09:08:54 +020051 * Work items for the bdi_writeback threads
Adrian Bunkf11b00f2008-04-29 00:58:56 -070052 */
Jens Axboe03ba3782009-09-09 09:08:54 +020053struct bdi_work {
Jens Axboe8010c3b2009-09-15 20:04:57 +020054 struct list_head list; /* pending work list */
55 struct rcu_head rcu_head; /* for RCU free/clear of work */
Jens Axboe03ba3782009-09-09 09:08:54 +020056
Jens Axboe8010c3b2009-09-15 20:04:57 +020057 unsigned long seen; /* threads that have seen this work */
58 atomic_t pending; /* number of threads still to do work */
Jens Axboe03ba3782009-09-09 09:08:54 +020059
Jens Axboe8010c3b2009-09-15 20:04:57 +020060 struct wb_writeback_args args; /* writeback arguments */
Jens Axboe03ba3782009-09-09 09:08:54 +020061
Jens Axboe8010c3b2009-09-15 20:04:57 +020062 unsigned long state; /* flag bits, see WS_* */
Jens Axboe03ba3782009-09-09 09:08:54 +020063};
64
65enum {
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +020066 WS_INPROGRESS = 0,
67 WS_ONSTACK,
Jens Axboe03ba3782009-09-09 09:08:54 +020068};
69
Jens Axboe03ba3782009-09-09 09:08:54 +020070static inline void bdi_work_init(struct bdi_work *work,
Jens Axboeb6e51312009-09-16 15:13:54 +020071 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +020072{
73 INIT_RCU_HEAD(&work->rcu_head);
Jens Axboeb6e51312009-09-16 15:13:54 +020074 work->args = *args;
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +020075 __set_bit(WS_INPROGRESS, &work->state);
Jens Axboe03ba3782009-09-09 09:08:54 +020076}
77
Adrian Bunkf11b00f2008-04-29 00:58:56 -070078/**
79 * writeback_in_progress - determine whether there is writeback in progress
80 * @bdi: the device's backing_dev_info structure.
81 *
Jens Axboe03ba3782009-09-09 09:08:54 +020082 * Determine whether there is writeback waiting to be handled against a
83 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070084 */
85int writeback_in_progress(struct backing_dev_info *bdi)
86{
Jens Axboe03ba3782009-09-09 09:08:54 +020087 return !list_empty(&bdi->work_list);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070088}
89
Jens Axboe03ba3782009-09-09 09:08:54 +020090static void bdi_work_free(struct rcu_head *head)
Nick Piggin4195f732009-05-28 09:01:15 +020091{
Jens Axboe03ba3782009-09-09 09:08:54 +020092 struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
Nick Piggin4195f732009-05-28 09:01:15 +020093
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +020094 clear_bit(WS_INPROGRESS, &work->state);
95 smp_mb__after_clear_bit();
96 wake_up_bit(&work->state, WS_INPROGRESS);
97
98 if (!test_bit(WS_ONSTACK, &work->state))
Jens Axboe03ba3782009-09-09 09:08:54 +020099 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200100}
101
102static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
103{
104 /*
105 * The caller has retrieved the work arguments from this work,
106 * drop our reference. If this is the last ref, delete and free it
107 */
108 if (atomic_dec_and_test(&work->pending)) {
109 struct backing_dev_info *bdi = wb->bdi;
110
111 spin_lock(&bdi->wb_lock);
112 list_del_rcu(&work->list);
113 spin_unlock(&bdi->wb_lock);
114
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200115 call_rcu(&work->rcu_head, bdi_work_free);
Nick Piggin4195f732009-05-28 09:01:15 +0200116 }
117}
118
Jens Axboe03ba3782009-09-09 09:08:54 +0200119static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120{
Jens Axboebcddc3f2009-09-13 20:07:36 +0200121 work->seen = bdi->wb_mask;
122 BUG_ON(!work->seen);
123 atomic_set(&work->pending, bdi->wb_cnt);
124 BUG_ON(!bdi->wb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
Jens Axboebcddc3f2009-09-13 20:07:36 +0200126 /*
Nick Piggindeed62e2009-09-15 21:32:58 +0200127 * list_add_tail_rcu() contains the necessary barriers to
128 * make sure the above stores are seen before the item is
129 * noticed on the list
Jens Axboebcddc3f2009-09-13 20:07:36 +0200130 */
Jens Axboebcddc3f2009-09-13 20:07:36 +0200131 spin_lock(&bdi->wb_lock);
132 list_add_tail_rcu(&work->list, &bdi->work_list);
133 spin_unlock(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200136 * If the default thread isn't there, make sure we add it. When
137 * it gets created and wakes up, we'll run this work.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200139 if (unlikely(list_empty_careful(&bdi->wb_list)))
140 wake_up_process(default_backing_dev_info.wb.task);
141 else {
142 struct bdi_writeback *wb = &bdi->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200144 if (wb->task)
Jens Axboe03ba3782009-09-09 09:08:54 +0200145 wake_up_process(wb->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147}
148
Jens Axboe03ba3782009-09-09 09:08:54 +0200149/*
150 * Used for on-stack allocated work items. The caller needs to wait until
151 * the wb threads have acked the work before it's safe to continue.
152 */
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200153static void bdi_wait_on_work_done(struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154{
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200155 wait_on_bit(&work->state, WS_INPROGRESS, bdi_sched_wait,
Jens Axboe03ba3782009-09-09 09:08:54 +0200156 TASK_UNINTERRUPTIBLE);
157}
158
Jens Axboef11fcae2009-09-15 09:53:35 +0200159static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
Jens Axboef17625b2010-06-01 11:05:22 +0200160 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200161{
162 struct bdi_work *work;
163
Jens Axboebcddc3f2009-09-13 20:07:36 +0200164 /*
165 * This is WB_SYNC_NONE writeback, so if allocation fails just
166 * wakeup the thread for old dirty data writeback
167 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200168 work = kmalloc(sizeof(*work), GFP_ATOMIC);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200169 if (work) {
Jens Axboeb6e51312009-09-16 15:13:54 +0200170 bdi_work_init(work, args);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200171 bdi_queue_work(bdi, work);
172 } else {
173 struct bdi_writeback *wb = &bdi->wb;
Jens Axboe03ba3782009-09-09 09:08:54 +0200174
Jens Axboebcddc3f2009-09-13 20:07:36 +0200175 if (wb->task)
176 wake_up_process(wb->task);
177 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200178}
179
Jens Axboeb6e51312009-09-16 15:13:54 +0200180/**
Christoph Hellwig3c4d7162010-06-08 18:14:43 +0200181 * bdi_queue_work_onstack - start and wait for writeback
Randy Dunlap06d738f2010-07-01 08:26:34 +0200182 * @args: parameters to control the work queue writeback
Jens Axboeb6e51312009-09-16 15:13:54 +0200183 *
184 * Description:
Christoph Hellwig3c4d7162010-06-08 18:14:43 +0200185 * This function initiates writeback and waits for the operation to
186 * complete. Callers must hold the sb s_umount semaphore for
Jens Axboeb6e51312009-09-16 15:13:54 +0200187 * reading, to avoid having the super disappear before we are done.
188 */
Christoph Hellwig3c4d7162010-06-08 18:14:43 +0200189static void bdi_queue_work_onstack(struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200190{
Jens Axboeb6e51312009-09-16 15:13:54 +0200191 struct bdi_work work;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200192
Christoph Hellwig3c4d7162010-06-08 18:14:43 +0200193 bdi_work_init(&work, args);
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200194 __set_bit(WS_ONSTACK, &work.state);
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200195
Christoph Hellwig3c4d7162010-06-08 18:14:43 +0200196 bdi_queue_work(args->sb->s_bdi, &work);
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200197 bdi_wait_on_work_done(&work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200198}
199
200/**
201 * bdi_start_writeback - start writeback
202 * @bdi: the backing device to write from
203 * @nr_pages: the number of pages to write
204 *
205 * Description:
206 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
207 * started when this function returns, we make no guarentees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200208 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200209 *
210 */
Christoph Hellwigc5444192010-06-08 18:15:15 +0200211void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200212{
213 struct wb_writeback_args args = {
214 .sync_mode = WB_SYNC_NONE,
215 .nr_pages = nr_pages,
216 .range_cyclic = 1,
217 };
218
Christoph Hellwigc5444192010-06-08 18:15:15 +0200219 bdi_alloc_queue_work(bdi, &args);
220}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800221
Christoph Hellwigc5444192010-06-08 18:15:15 +0200222/**
223 * bdi_start_background_writeback - start background writeback
224 * @bdi: the backing device to write from
225 *
226 * Description:
227 * This does WB_SYNC_NONE background writeback. The IO is only
228 * started when this function returns, we make no guarentees on
229 * completion. Caller need not hold sb s_umount semaphore.
230 */
231void bdi_start_background_writeback(struct backing_dev_info *bdi)
232{
233 struct wb_writeback_args args = {
234 .sync_mode = WB_SYNC_NONE,
235 .nr_pages = LONG_MAX,
236 .for_background = 1,
237 .range_cyclic = 1,
238 };
Jens Axboef17625b2010-06-01 11:05:22 +0200239 bdi_alloc_queue_work(bdi, &args);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240}
241
242/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700243 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
244 * furthest end of its superblock's dirty-inode list.
245 *
246 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200247 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700248 * the case then the inode must have been redirtied while it was being written
249 * out and we don't reset its dirtied_when.
250 */
251static void redirty_tail(struct inode *inode)
252{
Jens Axboe03ba3782009-09-09 09:08:54 +0200253 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700254
Jens Axboe03ba3782009-09-09 09:08:54 +0200255 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200256 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700257
Jens Axboe03ba3782009-09-09 09:08:54 +0200258 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200259 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700260 inode->dirtied_when = jiffies;
261 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200262 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700263}
264
265/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200266 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700267 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700268static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700269{
Jens Axboe03ba3782009-09-09 09:08:54 +0200270 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
271
272 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700273}
274
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700275static void inode_sync_complete(struct inode *inode)
276{
277 /*
278 * Prevent speculative execution through spin_unlock(&inode_lock);
279 */
280 smp_mb();
281 wake_up_bit(&inode->i_state, __I_SYNC);
282}
283
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700284static bool inode_dirtied_after(struct inode *inode, unsigned long t)
285{
286 bool ret = time_after(inode->dirtied_when, t);
287#ifndef CONFIG_64BIT
288 /*
289 * For inodes being constantly redirtied, dirtied_when can get stuck.
290 * It _appears_ to be in the future, but is actually in distant past.
291 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200292 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700293 */
294 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
295#endif
296 return ret;
297}
298
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700299/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700300 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
301 */
302static void move_expired_inodes(struct list_head *delaying_queue,
303 struct list_head *dispatch_queue,
304 unsigned long *older_than_this)
305{
Shaohua Li5c034492009-09-24 14:42:33 +0200306 LIST_HEAD(tmp);
307 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200308 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200309 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200310 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200311
Fengguang Wu2c136572007-10-16 23:30:39 -0700312 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200313 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700314 if (older_than_this &&
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700315 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700316 break;
Jens Axboecf137302009-09-24 15:12:57 +0200317 if (sb && sb != inode->i_sb)
318 do_sb_sort = 1;
319 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200320 list_move(&inode->i_list, &tmp);
321 }
322
Jens Axboecf137302009-09-24 15:12:57 +0200323 /* just one sb in list, splice to dispatch_queue and we're done */
324 if (!do_sb_sort) {
325 list_splice(&tmp, dispatch_queue);
326 return;
327 }
328
Shaohua Li5c034492009-09-24 14:42:33 +0200329 /* Move inodes from one superblock together */
330 while (!list_empty(&tmp)) {
331 inode = list_entry(tmp.prev, struct inode, i_list);
332 sb = inode->i_sb;
333 list_for_each_prev_safe(pos, node, &tmp) {
334 inode = list_entry(pos, struct inode, i_list);
335 if (inode->i_sb == sb)
336 list_move(&inode->i_list, dispatch_queue);
337 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700338 }
339}
340
341/*
342 * Queue all expired dirty inodes for io, eldest first.
343 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200344static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700345{
Jens Axboe03ba3782009-09-09 09:08:54 +0200346 list_splice_init(&wb->b_more_io, wb->b_io.prev);
347 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200348}
349
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100350static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200351{
Jens Axboe03ba3782009-09-09 09:08:54 +0200352 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100353 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200354 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700355}
356
357/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200358 * Wait for writeback on an inode to complete.
359 */
360static void inode_wait_for_writeback(struct inode *inode)
361{
362 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
363 wait_queue_head_t *wqh;
364
365 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700366 while (inode->i_state & I_SYNC) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200367 spin_unlock(&inode_lock);
368 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
369 spin_lock(&inode_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700370 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200371}
372
373/*
374 * Write out an inode's dirty pages. Called under inode_lock. Either the
375 * caller has ref on the inode (either via __iget or via syscall against an fd)
376 * or the inode has I_WILL_FREE set (via generic_forget_inode)
377 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 * If `wait' is set, wait on the writeout.
379 *
380 * The whole writeout design is quite complex and fragile. We want to avoid
381 * starvation of particular inodes when others are being redirtied, prevent
382 * livelocks, etc.
383 *
384 * Called under inode_lock.
385 */
386static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200387writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200390 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 int ret;
392
Christoph Hellwig01c03192009-06-08 13:35:40 +0200393 if (!atomic_read(&inode->i_count))
394 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
395 else
396 WARN_ON(inode->i_state & I_WILL_FREE);
397
398 if (inode->i_state & I_SYNC) {
399 /*
400 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200401 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200402 * writeback can proceed with the other inodes on s_io.
403 *
404 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200405 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200406 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100407 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200408 requeue_io(inode);
409 return 0;
410 }
411
412 /*
413 * It's a data-integrity sync. We must wait.
414 */
415 inode_wait_for_writeback(inode);
416 }
417
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700418 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400420 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700421 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400422 inode->i_state &= ~I_DIRTY_PAGES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 spin_unlock(&inode_lock);
424
425 ret = do_writepages(mapping, wbc);
426
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100427 /*
428 * Make sure to wait on the data before writing out the metadata.
429 * This is important for filesystems that modify metadata on data
430 * I/O completion.
431 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100432 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100433 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 if (ret == 0)
435 ret = err;
436 }
437
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400438 /*
439 * Some filesystems may redirty the inode during the writeback
440 * due to delalloc, clear dirty metadata flags right before
441 * write_inode()
442 */
443 spin_lock(&inode_lock);
444 dirty = inode->i_state & I_DIRTY;
445 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
446 spin_unlock(&inode_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100447 /* Don't write the inode if only I_DIRTY_PAGES was set */
448 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100449 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 if (ret == 0)
451 ret = err;
452 }
453
454 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700455 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700456 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangb3af9462009-09-25 06:04:10 +0200457 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800458 /*
Wu Fengguangb3af9462009-09-25 06:04:10 +0200459 * More pages get dirtied by a fast dirtier.
460 */
461 goto select_queue;
462 } else if (inode->i_state & I_DIRTY) {
463 /*
464 * At least XFS will redirty the inode during the
465 * writeback (delalloc) and on io completion (isize).
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800466 */
467 redirty_tail(inode);
468 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 /*
470 * We didn't write back all the pages. nfs_writepages()
471 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200472 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700473 */
474 /*
475 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200476 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700477 * consideration. Otherwise, move it to the tail, for
478 * the reasons described there. I'm not really sure
479 * how much sense this makes. Presumably I had a good
480 * reasons for doing it this way, and I'd rather not
481 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
483 if (wbc->for_kupdate) {
484 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700485 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200486 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700487 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 */
489 inode->i_state |= I_DIRTY_PAGES;
Wu Fengguangb3af9462009-09-25 06:04:10 +0200490select_queue:
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800491 if (wbc->nr_to_write <= 0) {
492 /*
493 * slice used up: queue for next turn
494 */
495 requeue_io(inode);
496 } else {
497 /*
498 * somehow blocked: retry later
499 */
500 redirty_tail(inode);
501 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 } else {
503 /*
504 * Otherwise fully redirty the inode so that
505 * other inodes on this superblock will get some
506 * writeout. Otherwise heavy writing to one
507 * file would indefinitely suspend writeout of
508 * all the other files.
509 */
510 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700511 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 } else if (atomic_read(&inode->i_count)) {
514 /*
515 * The inode is clean, inuse
516 */
517 list_move(&inode->i_list, &inode_in_use);
518 } else {
519 /*
520 * The inode is clean, unused
521 */
522 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 }
524 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700525 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 return ret;
527}
528
Jens Axboe03ba3782009-09-09 09:08:54 +0200529/*
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200530 * For background writeback the caller does not have the sb pinned
Jens Axboe03ba3782009-09-09 09:08:54 +0200531 * before calling writeback. So make sure that we do pin it, so it doesn't
532 * go away while we are writing inodes from it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200533 */
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200534static bool pin_sb_for_writeback(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
Jens Axboe03ba3782009-09-09 09:08:54 +0200536 spin_lock(&sb_lock);
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200537 if (list_empty(&sb->s_instances)) {
538 spin_unlock(&sb_lock);
539 return false;
540 }
541
Jens Axboe03ba3782009-09-09 09:08:54 +0200542 sb->s_count++;
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200543 spin_unlock(&sb_lock);
544
Jens Axboe03ba3782009-09-09 09:08:54 +0200545 if (down_read_trylock(&sb->s_umount)) {
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200546 if (sb->s_root)
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200547 return true;
Jens Axboe03ba3782009-09-09 09:08:54 +0200548 up_read(&sb->s_umount);
549 }
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200550
551 put_super(sb);
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200552 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200553}
554
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800555/*
556 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200557 *
558 * If @only_this_sb is true, then find and write all such
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800559 * inodes. Otherwise write only ones which go sequentially
560 * in reverse order.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200561 *
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800562 * Return 1, if the caller writeback routine should be
563 * interrupted. Otherwise return 0.
564 */
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200565static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
566 struct writeback_control *wbc, bool only_this_sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200567{
Jens Axboe03ba3782009-09-09 09:08:54 +0200568 while (!list_empty(&wb->b_io)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 long pages_skipped;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800570 struct inode *inode = list_entry(wb->b_io.prev,
571 struct inode, i_list);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200572
573 if (inode->i_sb != sb) {
574 if (only_this_sb) {
575 /*
576 * We only want to write back data for this
577 * superblock, move all inodes not belonging
578 * to it back onto the dirty list.
579 */
580 redirty_tail(inode);
581 continue;
582 }
583
584 /*
585 * The inode belongs to a different superblock.
586 * Bounce back to the caller to unpin this and
587 * pin the next superblock.
588 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800589 return 0;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200590 }
591
Wu Fengguang84a89242009-06-16 15:33:17 -0700592 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700593 requeue_io(inode);
594 continue;
595 }
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700596 /*
597 * Was this inode dirtied after sync_sb_inodes was called?
598 * This keeps sync from extra jobs and livelock.
599 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800600 if (inode_dirtied_after(inode, wbc->wb_start))
601 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
Wu Fengguang84a89242009-06-16 15:33:17 -0700603 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 __iget(inode);
605 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200606 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 if (wbc->pages_skipped != pages_skipped) {
608 /*
609 * writeback is not making progress due to locked
610 * buffers. Skip this inode for now.
611 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700612 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 }
614 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800616 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800618 if (wbc->nr_to_write <= 0) {
619 wbc->more_io = 1;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800620 return 1;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800621 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200622 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800623 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800625 /* b_io is empty */
626 return 1;
627}
Nick Piggin38f21972009-01-06 14:40:25 -0800628
Christoph Hellwig9c3a8ee2010-06-10 12:07:27 +0200629void writeback_inodes_wb(struct bdi_writeback *wb,
630 struct writeback_control *wbc)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800631{
632 int ret = 0;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200633
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800634 wbc->wb_start = jiffies; /* livelock avoidance */
635 spin_lock(&inode_lock);
636 if (!wbc->for_kupdate || list_empty(&wb->b_io))
637 queue_io(wb, wbc->older_than_this);
638
639 while (!list_empty(&wb->b_io)) {
640 struct inode *inode = list_entry(wb->b_io.prev,
641 struct inode, i_list);
642 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800643
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200644 if (!pin_sb_for_writeback(sb)) {
645 requeue_io(inode);
646 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800647 }
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200648 ret = writeback_sb_inodes(sb, wb, wbc, false);
649 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800650
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800651 if (ret)
652 break;
653 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200654 spin_unlock(&inode_lock);
655 /* Leave any unwritten inodes on b_io */
656}
657
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200658static void __writeback_inodes_sb(struct super_block *sb,
659 struct bdi_writeback *wb, struct writeback_control *wbc)
660{
661 WARN_ON(!rwsem_is_locked(&sb->s_umount));
662
663 wbc->wb_start = jiffies; /* livelock avoidance */
664 spin_lock(&inode_lock);
665 if (!wbc->for_kupdate || list_empty(&wb->b_io))
666 queue_io(wb, wbc->older_than_this);
667 writeback_sb_inodes(sb, wb, wbc, true);
668 spin_unlock(&inode_lock);
669}
670
Jens Axboe03ba3782009-09-09 09:08:54 +0200671/*
672 * The maximum number of pages to writeout in a single bdi flush/kupdate
673 * operation. We do this so we don't hold I_SYNC against an inode for
674 * enormous amounts of time, which would block a userspace task which has
675 * been forced to throttle against that inode. Also, the code reevaluates
676 * the dirty each time it has written this many pages.
677 */
678#define MAX_WRITEBACK_PAGES 1024
679
680static inline bool over_bground_thresh(void)
681{
682 unsigned long background_thresh, dirty_thresh;
683
684 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
685
686 return (global_page_state(NR_FILE_DIRTY) +
687 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
688}
689
690/*
691 * Explicit flushing or periodic writeback of "old" data.
692 *
693 * Define "old": the first time one of an inode's pages is dirtied, we mark the
694 * dirtying-time in the inode's address_space. So this periodic writeback code
695 * just walks the superblock inode list, writing back any inodes which are
696 * older than a specific point in time.
697 *
698 * Try to run once per dirty_writeback_interval. But if a writeback event
699 * takes longer than a dirty_writeback_interval interval, then leave a
700 * one-second gap.
701 *
702 * older_than_this takes precedence over nr_to_write. So we'll only write back
703 * all dirty pages if they are all attached to "old" mappings.
704 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200705static long wb_writeback(struct bdi_writeback *wb,
706 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200707{
708 struct writeback_control wbc = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200709 .sync_mode = args->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200710 .older_than_this = NULL,
Jens Axboec4a77a62009-09-16 15:18:25 +0200711 .for_kupdate = args->for_kupdate,
Wu Fengguangb17621f2009-12-03 13:54:25 +0100712 .for_background = args->for_background,
Jens Axboec4a77a62009-09-16 15:18:25 +0200713 .range_cyclic = args->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200714 };
715 unsigned long oldest_jif;
716 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200717 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200718
719 if (wbc.for_kupdate) {
720 wbc.older_than_this = &oldest_jif;
721 oldest_jif = jiffies -
722 msecs_to_jiffies(dirty_expire_interval * 10);
723 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200724 if (!wbc.range_cyclic) {
725 wbc.range_start = 0;
726 wbc.range_end = LLONG_MAX;
727 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200728
729 for (;;) {
730 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800731 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200732 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800733 if (args->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200734 break;
735
736 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800737 * For background writeout, stop when we are below the
738 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200739 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800740 if (args->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200741 break;
742
743 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200744 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
745 wbc.pages_skipped = 0;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200746 if (args->sb)
747 __writeback_inodes_sb(args->sb, wb, &wbc);
748 else
749 writeback_inodes_wb(wb, &wbc);
Jens Axboec4a77a62009-09-16 15:18:25 +0200750 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200751 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
752
753 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200754 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200755 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200756 if (wbc.nr_to_write <= 0)
757 continue;
758 /*
759 * Didn't write everything and we don't have more IO, bail
760 */
761 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200762 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200763 /*
764 * Did we write something? Try for more
765 */
766 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
767 continue;
768 /*
769 * Nothing written. Wait for some inode to
770 * become available for writeback. Otherwise
771 * we'll just busyloop.
772 */
773 spin_lock(&inode_lock);
774 if (!list_empty(&wb->b_more_io)) {
775 inode = list_entry(wb->b_more_io.prev,
776 struct inode, i_list);
777 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200778 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200779 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200780 }
781
782 return wrote;
783}
784
785/*
786 * Return the next bdi_work struct that hasn't been processed by this
Jens Axboe8010c3b2009-09-15 20:04:57 +0200787 * wb thread yet. ->seen is initially set for each thread that exists
788 * for this device, when a thread first notices a piece of work it
789 * clears its bit. Depending on writeback type, the thread will notify
790 * completion on either receiving the work (WB_SYNC_NONE) or after
791 * it is done (WB_SYNC_ALL).
Jens Axboe03ba3782009-09-09 09:08:54 +0200792 */
793static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
794 struct bdi_writeback *wb)
795{
796 struct bdi_work *work, *ret = NULL;
797
798 rcu_read_lock();
799
800 list_for_each_entry_rcu(work, &bdi->work_list, list) {
Nick Piggin77fad5e2009-09-15 21:34:12 +0200801 if (!test_bit(wb->nr, &work->seen))
Jens Axboe03ba3782009-09-09 09:08:54 +0200802 continue;
Nick Piggin77fad5e2009-09-15 21:34:12 +0200803 clear_bit(wb->nr, &work->seen);
Jens Axboe03ba3782009-09-09 09:08:54 +0200804
805 ret = work;
806 break;
807 }
808
809 rcu_read_unlock();
810 return ret;
811}
812
813static long wb_check_old_data_flush(struct bdi_writeback *wb)
814{
815 unsigned long expired;
816 long nr_pages;
817
Jens Axboe69b62d02010-05-17 12:51:03 +0200818 /*
819 * When set to zero, disable periodic writeback
820 */
821 if (!dirty_writeback_interval)
822 return 0;
823
Jens Axboe03ba3782009-09-09 09:08:54 +0200824 expired = wb->last_old_flush +
825 msecs_to_jiffies(dirty_writeback_interval * 10);
826 if (time_before(jiffies, expired))
827 return 0;
828
829 wb->last_old_flush = jiffies;
830 nr_pages = global_page_state(NR_FILE_DIRTY) +
831 global_page_state(NR_UNSTABLE_NFS) +
832 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
833
Jens Axboec4a77a62009-09-16 15:18:25 +0200834 if (nr_pages) {
835 struct wb_writeback_args args = {
836 .nr_pages = nr_pages,
837 .sync_mode = WB_SYNC_NONE,
838 .for_kupdate = 1,
839 .range_cyclic = 1,
840 };
841
842 return wb_writeback(wb, &args);
843 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200844
845 return 0;
846}
847
848/*
849 * Retrieve work items and do the writeback they describe
850 */
851long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
852{
853 struct backing_dev_info *bdi = wb->bdi;
854 struct bdi_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200855 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200856
857 while ((work = get_next_work_item(bdi, wb)) != NULL) {
Jens Axboec4a77a62009-09-16 15:18:25 +0200858 struct wb_writeback_args args = work->args;
Jens Axboe03ba3782009-09-09 09:08:54 +0200859
860 /*
861 * Override sync mode, in case we must wait for completion
862 */
863 if (force_wait)
Jens Axboec4a77a62009-09-16 15:18:25 +0200864 work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200865
866 /*
867 * If this isn't a data integrity operation, just notify
868 * that we have seen this work and we are now starting it.
869 */
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200870 if (!test_bit(WS_ONSTACK, &work->state))
Jens Axboe03ba3782009-09-09 09:08:54 +0200871 wb_clear_pending(wb, work);
872
Jens Axboec4a77a62009-09-16 15:18:25 +0200873 wrote += wb_writeback(wb, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200874
875 /*
876 * This is a data integrity writeback, so only do the
877 * notification when we have completed the work.
878 */
Christoph Hellwig7f0e7be2010-06-08 18:14:34 +0200879 if (test_bit(WS_ONSTACK, &work->state))
Jens Axboe03ba3782009-09-09 09:08:54 +0200880 wb_clear_pending(wb, work);
881 }
882
883 /*
884 * Check for periodic writeback, kupdated() style
885 */
886 wrote += wb_check_old_data_flush(wb);
887
888 return wrote;
889}
890
891/*
892 * Handle writeback of dirty data for the device backed by this bdi. Also
893 * wakes up periodically and does kupdated style flushing.
894 */
895int bdi_writeback_task(struct bdi_writeback *wb)
896{
897 unsigned long last_active = jiffies;
898 unsigned long wait_jiffies = -1UL;
899 long pages_written;
900
901 while (!kthread_should_stop()) {
902 pages_written = wb_do_writeback(wb, 0);
903
904 if (pages_written)
905 last_active = jiffies;
906 else if (wait_jiffies != -1UL) {
907 unsigned long max_idle;
908
909 /*
910 * Longest period of inactivity that we tolerate. If we
911 * see dirty data again later, the task will get
912 * recreated automatically.
913 */
914 max_idle = max(5UL * 60 * HZ, wait_jiffies);
915 if (time_after(jiffies, max_idle + last_active))
916 break;
917 }
918
Jens Axboe69b62d02010-05-17 12:51:03 +0200919 if (dirty_writeback_interval) {
920 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
921 schedule_timeout_interruptible(wait_jiffies);
Jens Axboef9eadbb2010-05-18 14:31:45 +0200922 } else {
923 set_current_state(TASK_INTERRUPTIBLE);
924 if (list_empty_careful(&wb->bdi->work_list) &&
925 !kthread_should_stop())
926 schedule();
927 __set_current_state(TASK_RUNNING);
928 }
Jens Axboe69b62d02010-05-17 12:51:03 +0200929
Jens Axboe03ba3782009-09-09 09:08:54 +0200930 try_to_freeze();
931 }
932
933 return 0;
934}
935
936/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200937 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
938 * the whole world.
939 */
940void wakeup_flusher_threads(long nr_pages)
941{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200942 struct backing_dev_info *bdi;
943 struct wb_writeback_args args = {
944 .sync_mode = WB_SYNC_NONE,
945 };
946
947 if (nr_pages) {
948 args.nr_pages = nr_pages;
949 } else {
950 args.nr_pages = global_page_state(NR_FILE_DIRTY) +
Jens Axboe03ba3782009-09-09 09:08:54 +0200951 global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200952 }
953
954 rcu_read_lock();
955 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
956 if (!bdi_has_dirty_io(bdi))
957 continue;
958 bdi_alloc_queue_work(bdi, &args);
959 }
960 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200961}
962
963static noinline void block_dump___mark_inode_dirty(struct inode *inode)
964{
965 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
966 struct dentry *dentry;
967 const char *name = "?";
968
969 dentry = d_find_alias(inode);
970 if (dentry) {
971 spin_lock(&dentry->d_lock);
972 name = (const char *) dentry->d_name.name;
973 }
974 printk(KERN_DEBUG
975 "%s(%d): dirtied inode %lu (%s) on %s\n",
976 current->comm, task_pid_nr(current), inode->i_ino,
977 name, inode->i_sb->s_id);
978 if (dentry) {
979 spin_unlock(&dentry->d_lock);
980 dput(dentry);
981 }
982 }
983}
984
985/**
986 * __mark_inode_dirty - internal function
987 * @inode: inode to mark
988 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
989 * Mark an inode as dirty. Callers should use mark_inode_dirty or
990 * mark_inode_dirty_sync.
991 *
992 * Put the inode on the super block's dirty list.
993 *
994 * CAREFUL! We mark it dirty unconditionally, but move it onto the
995 * dirty list only if it is hashed or if it refers to a blockdev.
996 * If it was not hashed, it will never be added to the dirty list
997 * even if it is later hashed, as it will have been marked dirty already.
998 *
999 * In short, make sure you hash any inodes _before_ you start marking
1000 * them dirty.
1001 *
1002 * This function *must* be atomic for the I_DIRTY_PAGES case -
1003 * set_page_dirty() is called under spinlock in several places.
1004 *
1005 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1006 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1007 * the kernel-internal blockdev inode represents the dirtying time of the
1008 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1009 * page->mapping->host, so the page-dirtying time is recorded in the internal
1010 * blockdev inode.
1011 */
1012void __mark_inode_dirty(struct inode *inode, int flags)
1013{
1014 struct super_block *sb = inode->i_sb;
1015
1016 /*
1017 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1018 * dirty the inode itself
1019 */
1020 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1021 if (sb->s_op->dirty_inode)
1022 sb->s_op->dirty_inode(inode);
1023 }
1024
1025 /*
1026 * make sure that changes are seen by all cpus before we test i_state
1027 * -- mikulas
1028 */
1029 smp_mb();
1030
1031 /* avoid the locking if we can */
1032 if ((inode->i_state & flags) == flags)
1033 return;
1034
1035 if (unlikely(block_dump))
1036 block_dump___mark_inode_dirty(inode);
1037
1038 spin_lock(&inode_lock);
1039 if ((inode->i_state & flags) != flags) {
1040 const int was_dirty = inode->i_state & I_DIRTY;
1041
1042 inode->i_state |= flags;
1043
1044 /*
1045 * If the inode is being synced, just update its dirty state.
1046 * The unlocker will place the inode on the appropriate
1047 * superblock list, based upon its state.
1048 */
1049 if (inode->i_state & I_SYNC)
1050 goto out;
1051
1052 /*
1053 * Only add valid (hashed) inodes to the superblock's
1054 * dirty list. Add blockdev inodes as well.
1055 */
1056 if (!S_ISBLK(inode->i_mode)) {
1057 if (hlist_unhashed(&inode->i_hash))
1058 goto out;
1059 }
1060 if (inode->i_state & (I_FREEING|I_CLEAR))
1061 goto out;
1062
1063 /*
1064 * If the inode was already on b_dirty/b_io/b_more_io, don't
1065 * reposition it (that would break b_dirty time-ordering).
1066 */
1067 if (!was_dirty) {
1068 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001069 struct backing_dev_info *bdi = wb->bdi;
1070
1071 if (bdi_cap_writeback_dirty(bdi) &&
1072 !test_bit(BDI_registered, &bdi->state)) {
1073 WARN_ON(1);
1074 printk(KERN_ERR "bdi-%s not registered\n",
1075 bdi->name);
1076 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001077
1078 inode->dirtied_when = jiffies;
1079 list_move(&inode->i_list, &wb->b_dirty);
1080 }
1081 }
1082out:
1083 spin_unlock(&inode_lock);
1084}
1085EXPORT_SYMBOL(__mark_inode_dirty);
1086
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001087/*
1088 * Write out a superblock's list of dirty inodes. A wait will be performed
1089 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1090 *
1091 * If older_than_this is non-NULL, then only write out inodes which
1092 * had their first dirtying at a time earlier than *older_than_this.
1093 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001094 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1095 * This function assumes that the blockdev superblock's inodes are backed by
1096 * a variety of queues, so all inodes are searched. For other superblocks,
1097 * assume that all inodes are backed by the same queue.
1098 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001099 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1100 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1101 * on the writer throttling path, and we get decent balancing between many
1102 * throttled threads: we don't want them all piling up on inode_sync_wait.
1103 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001104static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001105{
Jens Axboe03ba3782009-09-09 09:08:54 +02001106 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001107
Jens Axboe03ba3782009-09-09 09:08:54 +02001108 /*
1109 * We need to be protected against the filesystem going from
1110 * r/o to r/w or vice versa.
1111 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001112 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001113
Jens Axboe03ba3782009-09-09 09:08:54 +02001114 spin_lock(&inode_lock);
1115
1116 /*
1117 * Data integrity sync. Must wait for all pages under writeback,
1118 * because there may have been pages dirtied before our sync
1119 * call, but which had writeout started before we write it out.
1120 * In which case, the inode may not be on the dirty list, but
1121 * we still have to wait for that writeout.
1122 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001123 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001124 struct address_space *mapping;
1125
1126 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1127 continue;
1128 mapping = inode->i_mapping;
1129 if (mapping->nrpages == 0)
1130 continue;
1131 __iget(inode);
1132 spin_unlock(&inode_lock);
1133 /*
1134 * We hold a reference to 'inode' so it couldn't have
1135 * been removed from s_inodes list while we dropped the
1136 * inode_lock. We cannot iput the inode now as we can
1137 * be holding the last reference and we cannot iput it
1138 * under inode_lock. So we keep the reference and iput
1139 * it later.
1140 */
1141 iput(old_inode);
1142 old_inode = inode;
1143
1144 filemap_fdatawait(mapping);
1145
1146 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001147
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001148 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001149 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001150 spin_unlock(&inode_lock);
1151 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152}
1153
Jens Axboed8a85592009-09-02 12:34:32 +02001154/**
1155 * writeback_inodes_sb - writeback dirty inodes from given super_block
1156 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 *
Jens Axboed8a85592009-09-02 12:34:32 +02001158 * Start writeback on some inodes on this super_block. No guarantees are made
1159 * on how many (if any) will be written, and this function does not wait
1160 * for IO completion of submitted IO. The number of pages submitted is
1161 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001163void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164{
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001165 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1166 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001167 struct wb_writeback_args args = {
1168 .sb = sb,
1169 .sync_mode = WB_SYNC_NONE,
1170 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001171
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001172 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1173
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001174 args.nr_pages = nr_dirty + nr_unstable +
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001175 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1176
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001177 bdi_queue_work_onstack(&args);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178}
Jens Axboed8a85592009-09-02 12:34:32 +02001179EXPORT_SYMBOL(writeback_inodes_sb);
1180
1181/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001182 * writeback_inodes_sb_if_idle - start writeback if none underway
1183 * @sb: the superblock
1184 *
1185 * Invoke writeback_inodes_sb if no writeback is currently underway.
1186 * Returns 1 if writeback was started, 0 if not.
1187 */
1188int writeback_inodes_sb_if_idle(struct super_block *sb)
1189{
1190 if (!writeback_in_progress(sb->s_bdi)) {
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001191 down_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001192 writeback_inodes_sb(sb);
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001193 up_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001194 return 1;
1195 } else
1196 return 0;
1197}
1198EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1199
1200/**
Jens Axboed8a85592009-09-02 12:34:32 +02001201 * sync_inodes_sb - sync sb inode pages
1202 * @sb: the superblock
1203 *
1204 * This function writes and waits on any dirty inode belonging to this
1205 * super_block. The number of pages synced is returned.
1206 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001207void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001208{
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001209 struct wb_writeback_args args = {
1210 .sb = sb,
1211 .sync_mode = WB_SYNC_ALL,
1212 .nr_pages = LONG_MAX,
1213 .range_cyclic = 0,
1214 };
1215
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001216 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1217
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001218 bdi_queue_work_onstack(&args);
Jens Axboeb6e51312009-09-16 15:13:54 +02001219 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001220}
1221EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001224 * write_inode_now - write an inode to disk
1225 * @inode: inode to write to disk
1226 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001228 * This function commits an inode to disk immediately if it is dirty. This is
1229 * primarily needed by knfsd.
1230 *
1231 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233int write_inode_now(struct inode *inode, int sync)
1234{
1235 int ret;
1236 struct writeback_control wbc = {
1237 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001238 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001239 .range_start = 0,
1240 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 };
1242
1243 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001244 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
1246 might_sleep();
1247 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001248 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 spin_unlock(&inode_lock);
1250 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001251 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 return ret;
1253}
1254EXPORT_SYMBOL(write_inode_now);
1255
1256/**
1257 * sync_inode - write an inode and its pages to disk.
1258 * @inode: the inode to sync
1259 * @wbc: controls the writeback mode
1260 *
1261 * sync_inode() will write an inode and its pages to disk. It will also
1262 * correctly update the inode on its superblock's dirty inode lists and will
1263 * update inode->i_state.
1264 *
1265 * The caller must have a ref on the inode.
1266 */
1267int sync_inode(struct inode *inode, struct writeback_control *wbc)
1268{
1269 int ret;
1270
1271 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001272 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 spin_unlock(&inode_lock);
1274 return ret;
1275}
1276EXPORT_SYMBOL(sync_inode);