blob: 76fc4d594acb4a3c8091531df7ee37f974f192b6 [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>
19#include <linux/sched.h>
20#include <linux/fs.h>
21#include <linux/mm.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020022#include <linux/kthread.h>
23#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/backing-dev.h>
27#include <linux/buffer_head.h>
David Howells07f3f052006-09-30 20:52:18 +020028#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
Jens Axboe66f3b8e2009-09-02 09:19:46 +020030#define inode_to_bdi(inode) ((inode)->i_mapping->backing_dev_info)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070031
Jens Axboe03ba3782009-09-09 09:08:54 +020032/*
Jens Axboed0bceac2009-05-18 08:20:32 +020033 * We don't actually have pdflush, but this one is exported though /proc...
34 */
35int nr_pdflush_threads;
36
37/*
Jens Axboec4a77a62009-09-16 15:18:25 +020038 * Passed into wb_writeback(), essentially a subset of writeback_control
39 */
40struct wb_writeback_args {
41 long nr_pages;
42 struct super_block *sb;
43 enum writeback_sync_modes sync_mode;
Wu Fengguangd3ddec72009-09-23 20:33:40 +080044 int for_kupdate:1;
45 int range_cyclic:1;
46 int for_background:1;
Jens Axboec4a77a62009-09-16 15:18:25 +020047};
48
49/*
Jens Axboe03ba3782009-09-09 09:08:54 +020050 * Work items for the bdi_writeback threads
Adrian Bunkf11b00f2008-04-29 00:58:56 -070051 */
Jens Axboe03ba3782009-09-09 09:08:54 +020052struct bdi_work {
Jens Axboe8010c3b2009-09-15 20:04:57 +020053 struct list_head list; /* pending work list */
54 struct rcu_head rcu_head; /* for RCU free/clear of work */
Jens Axboe03ba3782009-09-09 09:08:54 +020055
Jens Axboe8010c3b2009-09-15 20:04:57 +020056 unsigned long seen; /* threads that have seen this work */
57 atomic_t pending; /* number of threads still to do work */
Jens Axboe03ba3782009-09-09 09:08:54 +020058
Jens Axboe8010c3b2009-09-15 20:04:57 +020059 struct wb_writeback_args args; /* writeback arguments */
Jens Axboe03ba3782009-09-09 09:08:54 +020060
Jens Axboe8010c3b2009-09-15 20:04:57 +020061 unsigned long state; /* flag bits, see WS_* */
Jens Axboe03ba3782009-09-09 09:08:54 +020062};
63
64enum {
65 WS_USED_B = 0,
66 WS_ONSTACK_B,
67};
68
69#define WS_USED (1 << WS_USED_B)
70#define WS_ONSTACK (1 << WS_ONSTACK_B)
71
72static inline bool bdi_work_on_stack(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070073{
Jens Axboe03ba3782009-09-09 09:08:54 +020074 return test_bit(WS_ONSTACK_B, &work->state);
75}
76
77static inline void bdi_work_init(struct bdi_work *work,
Jens Axboeb6e51312009-09-16 15:13:54 +020078 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +020079{
80 INIT_RCU_HEAD(&work->rcu_head);
Jens Axboeb6e51312009-09-16 15:13:54 +020081 work->args = *args;
Jens Axboe03ba3782009-09-09 09:08:54 +020082 work->state = WS_USED;
83}
84
Adrian Bunkf11b00f2008-04-29 00:58:56 -070085/**
86 * writeback_in_progress - determine whether there is writeback in progress
87 * @bdi: the device's backing_dev_info structure.
88 *
Jens Axboe03ba3782009-09-09 09:08:54 +020089 * Determine whether there is writeback waiting to be handled against a
90 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070091 */
92int writeback_in_progress(struct backing_dev_info *bdi)
93{
Jens Axboe03ba3782009-09-09 09:08:54 +020094 return !list_empty(&bdi->work_list);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070095}
96
Jens Axboe03ba3782009-09-09 09:08:54 +020097static void bdi_work_clear(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070098{
Jens Axboe03ba3782009-09-09 09:08:54 +020099 clear_bit(WS_USED_B, &work->state);
100 smp_mb__after_clear_bit();
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200101 /*
102 * work can have disappeared at this point. bit waitq functions
103 * should be able to tolerate this, provided bdi_sched_wait does
104 * not dereference it's pointer argument.
105 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200106 wake_up_bit(&work->state, WS_USED_B);
Adrian Bunkf11b00f2008-04-29 00:58:56 -0700107}
108
Jens Axboe03ba3782009-09-09 09:08:54 +0200109static void bdi_work_free(struct rcu_head *head)
Nick Piggin4195f732009-05-28 09:01:15 +0200110{
Jens Axboe03ba3782009-09-09 09:08:54 +0200111 struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
Nick Piggin4195f732009-05-28 09:01:15 +0200112
Jens Axboe03ba3782009-09-09 09:08:54 +0200113 if (!bdi_work_on_stack(work))
114 kfree(work);
115 else
116 bdi_work_clear(work);
117}
118
119static void wb_work_complete(struct bdi_work *work)
120{
Jens Axboec4a77a62009-09-16 15:18:25 +0200121 const enum writeback_sync_modes sync_mode = work->args.sync_mode;
Nick Piggin77b9d052009-09-15 21:34:51 +0200122 int onstack = bdi_work_on_stack(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200123
124 /*
125 * For allocated work, we can clear the done/seen bit right here.
126 * For on-stack work, we need to postpone both the clear and free
127 * to after the RCU grace period, since the stack could be invalidated
128 * as soon as bdi_work_clear() has done the wakeup.
129 */
Nick Piggin77b9d052009-09-15 21:34:51 +0200130 if (!onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200131 bdi_work_clear(work);
Nick Piggin77b9d052009-09-15 21:34:51 +0200132 if (sync_mode == WB_SYNC_NONE || onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200133 call_rcu(&work->rcu_head, bdi_work_free);
134}
135
136static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
137{
138 /*
139 * The caller has retrieved the work arguments from this work,
140 * drop our reference. If this is the last ref, delete and free it
141 */
142 if (atomic_dec_and_test(&work->pending)) {
143 struct backing_dev_info *bdi = wb->bdi;
144
145 spin_lock(&bdi->wb_lock);
146 list_del_rcu(&work->list);
147 spin_unlock(&bdi->wb_lock);
148
149 wb_work_complete(work);
Nick Piggin4195f732009-05-28 09:01:15 +0200150 }
151}
152
Jens Axboe03ba3782009-09-09 09:08:54 +0200153static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154{
Jens Axboebcddc3f2009-09-13 20:07:36 +0200155 work->seen = bdi->wb_mask;
156 BUG_ON(!work->seen);
157 atomic_set(&work->pending, bdi->wb_cnt);
158 BUG_ON(!bdi->wb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159
Jens Axboebcddc3f2009-09-13 20:07:36 +0200160 /*
Nick Piggindeed62e2009-09-15 21:32:58 +0200161 * list_add_tail_rcu() contains the necessary barriers to
162 * make sure the above stores are seen before the item is
163 * noticed on the list
Jens Axboebcddc3f2009-09-13 20:07:36 +0200164 */
Jens Axboebcddc3f2009-09-13 20:07:36 +0200165 spin_lock(&bdi->wb_lock);
166 list_add_tail_rcu(&work->list, &bdi->work_list);
167 spin_unlock(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200170 * If the default thread isn't there, make sure we add it. When
171 * it gets created and wakes up, we'll run this work.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200173 if (unlikely(list_empty_careful(&bdi->wb_list)))
174 wake_up_process(default_backing_dev_info.wb.task);
175 else {
176 struct bdi_writeback *wb = &bdi->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200178 if (wb->task)
Jens Axboe03ba3782009-09-09 09:08:54 +0200179 wake_up_process(wb->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181}
182
Jens Axboe03ba3782009-09-09 09:08:54 +0200183/*
184 * Used for on-stack allocated work items. The caller needs to wait until
185 * the wb threads have acked the work before it's safe to continue.
186 */
187static void bdi_wait_on_work_clear(struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188{
Jens Axboe03ba3782009-09-09 09:08:54 +0200189 wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
190 TASK_UNINTERRUPTIBLE);
191}
192
Jens Axboef11fcae2009-09-15 09:53:35 +0200193static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
Jens Axboeb6e51312009-09-16 15:13:54 +0200194 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200195{
196 struct bdi_work *work;
197
Jens Axboebcddc3f2009-09-13 20:07:36 +0200198 /*
199 * This is WB_SYNC_NONE writeback, so if allocation fails just
200 * wakeup the thread for old dirty data writeback
201 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200202 work = kmalloc(sizeof(*work), GFP_ATOMIC);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200203 if (work) {
Jens Axboeb6e51312009-09-16 15:13:54 +0200204 bdi_work_init(work, args);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200205 bdi_queue_work(bdi, work);
206 } else {
207 struct bdi_writeback *wb = &bdi->wb;
Jens Axboe03ba3782009-09-09 09:08:54 +0200208
Jens Axboebcddc3f2009-09-13 20:07:36 +0200209 if (wb->task)
210 wake_up_process(wb->task);
211 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200212}
213
Jens Axboeb6e51312009-09-16 15:13:54 +0200214/**
215 * bdi_sync_writeback - start and wait for writeback
216 * @bdi: the backing device to write from
217 * @sb: write inodes from this super_block
218 *
219 * Description:
220 * This does WB_SYNC_ALL data integrity writeback and waits for the
221 * IO to complete. Callers must hold the sb s_umount semaphore for
222 * reading, to avoid having the super disappear before we are done.
223 */
224static void bdi_sync_writeback(struct backing_dev_info *bdi,
225 struct super_block *sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200226{
Jens Axboeb6e51312009-09-16 15:13:54 +0200227 struct wb_writeback_args args = {
228 .sb = sb,
229 .sync_mode = WB_SYNC_ALL,
230 .nr_pages = LONG_MAX,
231 .range_cyclic = 0,
232 };
233 struct bdi_work work;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200234
Jens Axboeb6e51312009-09-16 15:13:54 +0200235 bdi_work_init(&work, &args);
236 work.state |= WS_ONSTACK;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200237
Jens Axboeb6e51312009-09-16 15:13:54 +0200238 bdi_queue_work(bdi, &work);
239 bdi_wait_on_work_clear(&work);
240}
241
242/**
243 * bdi_start_writeback - start writeback
244 * @bdi: the backing device to write from
Jaswinder Singh Rajput4b6764f2010-01-01 20:35:23 -0800245 * @sb: write inodes from this super_block
Jens Axboeb6e51312009-09-16 15:13:54 +0200246 * @nr_pages: the number of pages to write
247 *
248 * Description:
249 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
250 * started when this function returns, we make no guarentees on
251 * completion. Caller need not hold sb s_umount semaphore.
252 *
253 */
Jens Axboea72bfd42009-09-26 00:07:46 +0200254void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
255 long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200256{
257 struct wb_writeback_args args = {
Jens Axboea72bfd42009-09-26 00:07:46 +0200258 .sb = sb,
Jens Axboeb6e51312009-09-16 15:13:54 +0200259 .sync_mode = WB_SYNC_NONE,
260 .nr_pages = nr_pages,
261 .range_cyclic = 1,
262 };
263
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800264 /*
265 * We treat @nr_pages=0 as the special case to do background writeback,
266 * ie. to sync pages until the background dirty threshold is reached.
267 */
268 if (!nr_pages) {
269 args.nr_pages = LONG_MAX;
270 args.for_background = 1;
271 }
272
Jens Axboeb6e51312009-09-16 15:13:54 +0200273 bdi_alloc_queue_work(bdi, &args);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274}
275
276/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700277 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
278 * furthest end of its superblock's dirty-inode list.
279 *
280 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200281 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700282 * the case then the inode must have been redirtied while it was being written
283 * out and we don't reset its dirtied_when.
284 */
285static void redirty_tail(struct inode *inode)
286{
Jens Axboe03ba3782009-09-09 09:08:54 +0200287 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700288
Jens Axboe03ba3782009-09-09 09:08:54 +0200289 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200290 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700291
Jens Axboe03ba3782009-09-09 09:08:54 +0200292 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200293 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700294 inode->dirtied_when = jiffies;
295 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200296 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700297}
298
299/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200300 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700301 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700302static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700303{
Jens Axboe03ba3782009-09-09 09:08:54 +0200304 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
305
306 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700307}
308
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700309static void inode_sync_complete(struct inode *inode)
310{
311 /*
312 * Prevent speculative execution through spin_unlock(&inode_lock);
313 */
314 smp_mb();
315 wake_up_bit(&inode->i_state, __I_SYNC);
316}
317
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700318static bool inode_dirtied_after(struct inode *inode, unsigned long t)
319{
320 bool ret = time_after(inode->dirtied_when, t);
321#ifndef CONFIG_64BIT
322 /*
323 * For inodes being constantly redirtied, dirtied_when can get stuck.
324 * It _appears_ to be in the future, but is actually in distant past.
325 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200326 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700327 */
328 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
329#endif
330 return ret;
331}
332
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700333/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700334 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
335 */
336static void move_expired_inodes(struct list_head *delaying_queue,
337 struct list_head *dispatch_queue,
338 unsigned long *older_than_this)
339{
Shaohua Li5c034492009-09-24 14:42:33 +0200340 LIST_HEAD(tmp);
341 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200342 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200343 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200344 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200345
Fengguang Wu2c136572007-10-16 23:30:39 -0700346 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200347 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700348 if (older_than_this &&
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700349 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700350 break;
Jens Axboecf137302009-09-24 15:12:57 +0200351 if (sb && sb != inode->i_sb)
352 do_sb_sort = 1;
353 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200354 list_move(&inode->i_list, &tmp);
355 }
356
Jens Axboecf137302009-09-24 15:12:57 +0200357 /* just one sb in list, splice to dispatch_queue and we're done */
358 if (!do_sb_sort) {
359 list_splice(&tmp, dispatch_queue);
360 return;
361 }
362
Shaohua Li5c034492009-09-24 14:42:33 +0200363 /* Move inodes from one superblock together */
364 while (!list_empty(&tmp)) {
365 inode = list_entry(tmp.prev, struct inode, i_list);
366 sb = inode->i_sb;
367 list_for_each_prev_safe(pos, node, &tmp) {
368 inode = list_entry(pos, struct inode, i_list);
369 if (inode->i_sb == sb)
370 list_move(&inode->i_list, dispatch_queue);
371 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700372 }
373}
374
375/*
376 * Queue all expired dirty inodes for io, eldest first.
377 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200378static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700379{
Jens Axboe03ba3782009-09-09 09:08:54 +0200380 list_splice_init(&wb->b_more_io, wb->b_io.prev);
381 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200382}
383
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100384static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200385{
Jens Axboe03ba3782009-09-09 09:08:54 +0200386 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100387 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200388 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700389}
390
391/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200392 * Wait for writeback on an inode to complete.
393 */
394static void inode_wait_for_writeback(struct inode *inode)
395{
396 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
397 wait_queue_head_t *wqh;
398
399 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
400 do {
401 spin_unlock(&inode_lock);
402 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
403 spin_lock(&inode_lock);
404 } while (inode->i_state & I_SYNC);
405}
406
407/*
408 * Write out an inode's dirty pages. Called under inode_lock. Either the
409 * caller has ref on the inode (either via __iget or via syscall against an fd)
410 * or the inode has I_WILL_FREE set (via generic_forget_inode)
411 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 * If `wait' is set, wait on the writeout.
413 *
414 * The whole writeout design is quite complex and fragile. We want to avoid
415 * starvation of particular inodes when others are being redirtied, prevent
416 * livelocks, etc.
417 *
418 * Called under inode_lock.
419 */
420static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200421writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200424 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 int ret;
426
Christoph Hellwig01c03192009-06-08 13:35:40 +0200427 if (!atomic_read(&inode->i_count))
428 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
429 else
430 WARN_ON(inode->i_state & I_WILL_FREE);
431
432 if (inode->i_state & I_SYNC) {
433 /*
434 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200435 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200436 * writeback can proceed with the other inodes on s_io.
437 *
438 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200439 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200440 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100441 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200442 requeue_io(inode);
443 return 0;
444 }
445
446 /*
447 * It's a data-integrity sync. We must wait.
448 */
449 inode_wait_for_writeback(inode);
450 }
451
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700452 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700454 /* Set I_SYNC, reset I_DIRTY */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 dirty = inode->i_state & I_DIRTY;
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700456 inode->i_state |= I_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 inode->i_state &= ~I_DIRTY;
458
459 spin_unlock(&inode_lock);
460
461 ret = do_writepages(mapping, wbc);
462
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100463 /*
464 * Make sure to wait on the data before writing out the metadata.
465 * This is important for filesystems that modify metadata on data
466 * I/O completion.
467 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100468 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100469 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 if (ret == 0)
471 ret = err;
472 }
473
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100474 /* Don't write the inode if only I_DIRTY_PAGES was set */
475 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100476 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 if (ret == 0)
478 ret = err;
479 }
480
481 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700482 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700483 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangb3af9462009-09-25 06:04:10 +0200484 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800485 /*
Wu Fengguangb3af9462009-09-25 06:04:10 +0200486 * More pages get dirtied by a fast dirtier.
487 */
488 goto select_queue;
489 } else if (inode->i_state & I_DIRTY) {
490 /*
491 * At least XFS will redirty the inode during the
492 * writeback (delalloc) and on io completion (isize).
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800493 */
494 redirty_tail(inode);
495 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 /*
497 * We didn't write back all the pages. nfs_writepages()
498 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200499 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700500 */
501 /*
502 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200503 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700504 * consideration. Otherwise, move it to the tail, for
505 * the reasons described there. I'm not really sure
506 * how much sense this makes. Presumably I had a good
507 * reasons for doing it this way, and I'd rather not
508 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 */
510 if (wbc->for_kupdate) {
511 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700512 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200513 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700514 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 */
516 inode->i_state |= I_DIRTY_PAGES;
Wu Fengguangb3af9462009-09-25 06:04:10 +0200517select_queue:
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800518 if (wbc->nr_to_write <= 0) {
519 /*
520 * slice used up: queue for next turn
521 */
522 requeue_io(inode);
523 } else {
524 /*
525 * somehow blocked: retry later
526 */
527 redirty_tail(inode);
528 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 } else {
530 /*
531 * Otherwise fully redirty the inode so that
532 * other inodes on this superblock will get some
533 * writeout. Otherwise heavy writing to one
534 * file would indefinitely suspend writeout of
535 * all the other files.
536 */
537 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700538 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 } else if (atomic_read(&inode->i_count)) {
541 /*
542 * The inode is clean, inuse
543 */
544 list_move(&inode->i_list, &inode_in_use);
545 } else {
546 /*
547 * The inode is clean, unused
548 */
549 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 }
551 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700552 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 return ret;
554}
555
Jens Axboe9ecc2732009-09-24 15:25:11 +0200556static void unpin_sb_for_writeback(struct super_block **psb)
557{
558 struct super_block *sb = *psb;
559
560 if (sb) {
561 up_read(&sb->s_umount);
562 put_super(sb);
563 *psb = NULL;
564 }
565}
566
Jens Axboe03ba3782009-09-09 09:08:54 +0200567/*
568 * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
569 * before calling writeback. So make sure that we do pin it, so it doesn't
570 * go away while we are writing inodes from it.
571 *
572 * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
573 * 1 if we failed.
574 */
575static int pin_sb_for_writeback(struct writeback_control *wbc,
Jens Axboe9ecc2732009-09-24 15:25:11 +0200576 struct inode *inode, struct super_block **psb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577{
Jens Axboe03ba3782009-09-09 09:08:54 +0200578 struct super_block *sb = inode->i_sb;
579
580 /*
Jens Axboe9ecc2732009-09-24 15:25:11 +0200581 * If this sb is already pinned, nothing more to do. If not and
582 * *psb is non-NULL, unpin the old one first
583 */
584 if (sb == *psb)
585 return 0;
586 else if (*psb)
587 unpin_sb_for_writeback(psb);
588
589 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200590 * Caller must already hold the ref for this
591 */
592 if (wbc->sync_mode == WB_SYNC_ALL) {
593 WARN_ON(!rwsem_is_locked(&sb->s_umount));
594 return 0;
595 }
596
597 spin_lock(&sb_lock);
598 sb->s_count++;
599 if (down_read_trylock(&sb->s_umount)) {
600 if (sb->s_root) {
601 spin_unlock(&sb_lock);
Jens Axboe9ecc2732009-09-24 15:25:11 +0200602 goto pinned;
Jens Axboe03ba3782009-09-09 09:08:54 +0200603 }
604 /*
605 * umounted, drop rwsem again and fall through to failure
606 */
607 up_read(&sb->s_umount);
608 }
609
610 sb->s_count--;
611 spin_unlock(&sb_lock);
612 return 1;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200613pinned:
614 *psb = sb;
615 return 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200616}
617
618static void writeback_inodes_wb(struct bdi_writeback *wb,
619 struct writeback_control *wbc)
620{
Jens Axboe9ecc2732009-09-24 15:25:11 +0200621 struct super_block *sb = wbc->sb, *pin_sb = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 const unsigned long start = jiffies; /* livelock avoidance */
623
Hans Reiserae8547b2008-05-07 15:48:57 +0300624 spin_lock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
Jens Axboe03ba3782009-09-09 09:08:54 +0200626 if (!wbc->for_kupdate || list_empty(&wb->b_io))
627 queue_io(wb, wbc->older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200628
Jens Axboe03ba3782009-09-09 09:08:54 +0200629 while (!list_empty(&wb->b_io)) {
630 struct inode *inode = list_entry(wb->b_io.prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 struct inode, i_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 long pages_skipped;
633
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200634 /*
635 * super block given and doesn't match, skip this inode
636 */
637 if (sb && sb != inode->i_sb) {
638 redirty_tail(inode);
639 continue;
640 }
641
Wu Fengguang84a89242009-06-16 15:33:17 -0700642 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700643 requeue_io(inode);
644 continue;
645 }
646
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700647 /*
648 * Was this inode dirtied after sync_sb_inodes was called?
649 * This keeps sync from extra jobs and livelock.
650 */
651 if (inode_dirtied_after(inode, start))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 break;
653
Jens Axboe9ecc2732009-09-24 15:25:11 +0200654 if (pin_sb_for_writeback(wbc, inode, &pin_sb)) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200655 requeue_io(inode);
656 continue;
657 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
Wu Fengguang84a89242009-06-16 15:33:17 -0700659 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 __iget(inode);
661 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200662 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 if (wbc->pages_skipped != pages_skipped) {
664 /*
665 * writeback is not making progress due to locked
666 * buffers. Skip this inode for now.
667 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700668 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 }
670 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800672 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800674 if (wbc->nr_to_write <= 0) {
675 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800677 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200678 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800679 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 }
Nick Piggin38f21972009-01-06 14:40:25 -0800681
Jens Axboe9ecc2732009-09-24 15:25:11 +0200682 unpin_sb_for_writeback(&pin_sb);
683
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200684 spin_unlock(&inode_lock);
685 /* Leave any unwritten inodes on b_io */
686}
687
Jens Axboe03ba3782009-09-09 09:08:54 +0200688void writeback_inodes_wbc(struct writeback_control *wbc)
689{
690 struct backing_dev_info *bdi = wbc->bdi;
691
692 writeback_inodes_wb(&bdi->wb, wbc);
693}
694
695/*
696 * The maximum number of pages to writeout in a single bdi flush/kupdate
697 * operation. We do this so we don't hold I_SYNC against an inode for
698 * enormous amounts of time, which would block a userspace task which has
699 * been forced to throttle against that inode. Also, the code reevaluates
700 * the dirty each time it has written this many pages.
701 */
702#define MAX_WRITEBACK_PAGES 1024
703
704static inline bool over_bground_thresh(void)
705{
706 unsigned long background_thresh, dirty_thresh;
707
708 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
709
710 return (global_page_state(NR_FILE_DIRTY) +
711 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
712}
713
714/*
715 * Explicit flushing or periodic writeback of "old" data.
716 *
717 * Define "old": the first time one of an inode's pages is dirtied, we mark the
718 * dirtying-time in the inode's address_space. So this periodic writeback code
719 * just walks the superblock inode list, writing back any inodes which are
720 * older than a specific point in time.
721 *
722 * Try to run once per dirty_writeback_interval. But if a writeback event
723 * takes longer than a dirty_writeback_interval interval, then leave a
724 * one-second gap.
725 *
726 * older_than_this takes precedence over nr_to_write. So we'll only write back
727 * all dirty pages if they are all attached to "old" mappings.
728 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200729static long wb_writeback(struct bdi_writeback *wb,
730 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200731{
732 struct writeback_control wbc = {
733 .bdi = wb->bdi,
Jens Axboec4a77a62009-09-16 15:18:25 +0200734 .sb = args->sb,
735 .sync_mode = args->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200736 .older_than_this = NULL,
Jens Axboec4a77a62009-09-16 15:18:25 +0200737 .for_kupdate = args->for_kupdate,
Wu Fengguangb17621f2009-12-03 13:54:25 +0100738 .for_background = args->for_background,
Jens Axboec4a77a62009-09-16 15:18:25 +0200739 .range_cyclic = args->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200740 };
741 unsigned long oldest_jif;
742 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200743 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200744
745 if (wbc.for_kupdate) {
746 wbc.older_than_this = &oldest_jif;
747 oldest_jif = jiffies -
748 msecs_to_jiffies(dirty_expire_interval * 10);
749 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200750 if (!wbc.range_cyclic) {
751 wbc.range_start = 0;
752 wbc.range_end = LLONG_MAX;
753 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200754
755 for (;;) {
756 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800757 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200758 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800759 if (args->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200760 break;
761
762 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800763 * For background writeout, stop when we are below the
764 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200765 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800766 if (args->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200767 break;
768
769 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200770 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
771 wbc.pages_skipped = 0;
772 writeback_inodes_wb(wb, &wbc);
Jens Axboec4a77a62009-09-16 15:18:25 +0200773 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200774 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
775
776 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200777 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200778 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200779 if (wbc.nr_to_write <= 0)
780 continue;
781 /*
782 * Didn't write everything and we don't have more IO, bail
783 */
784 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200785 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200786 /*
787 * Did we write something? Try for more
788 */
789 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
790 continue;
791 /*
792 * Nothing written. Wait for some inode to
793 * become available for writeback. Otherwise
794 * we'll just busyloop.
795 */
796 spin_lock(&inode_lock);
797 if (!list_empty(&wb->b_more_io)) {
798 inode = list_entry(wb->b_more_io.prev,
799 struct inode, i_list);
800 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200801 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200802 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200803 }
804
805 return wrote;
806}
807
808/*
809 * Return the next bdi_work struct that hasn't been processed by this
Jens Axboe8010c3b2009-09-15 20:04:57 +0200810 * wb thread yet. ->seen is initially set for each thread that exists
811 * for this device, when a thread first notices a piece of work it
812 * clears its bit. Depending on writeback type, the thread will notify
813 * completion on either receiving the work (WB_SYNC_NONE) or after
814 * it is done (WB_SYNC_ALL).
Jens Axboe03ba3782009-09-09 09:08:54 +0200815 */
816static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
817 struct bdi_writeback *wb)
818{
819 struct bdi_work *work, *ret = NULL;
820
821 rcu_read_lock();
822
823 list_for_each_entry_rcu(work, &bdi->work_list, list) {
Nick Piggin77fad5e2009-09-15 21:34:12 +0200824 if (!test_bit(wb->nr, &work->seen))
Jens Axboe03ba3782009-09-09 09:08:54 +0200825 continue;
Nick Piggin77fad5e2009-09-15 21:34:12 +0200826 clear_bit(wb->nr, &work->seen);
Jens Axboe03ba3782009-09-09 09:08:54 +0200827
828 ret = work;
829 break;
830 }
831
832 rcu_read_unlock();
833 return ret;
834}
835
836static long wb_check_old_data_flush(struct bdi_writeback *wb)
837{
838 unsigned long expired;
839 long nr_pages;
840
841 expired = wb->last_old_flush +
842 msecs_to_jiffies(dirty_writeback_interval * 10);
843 if (time_before(jiffies, expired))
844 return 0;
845
846 wb->last_old_flush = jiffies;
847 nr_pages = global_page_state(NR_FILE_DIRTY) +
848 global_page_state(NR_UNSTABLE_NFS) +
849 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
850
Jens Axboec4a77a62009-09-16 15:18:25 +0200851 if (nr_pages) {
852 struct wb_writeback_args args = {
853 .nr_pages = nr_pages,
854 .sync_mode = WB_SYNC_NONE,
855 .for_kupdate = 1,
856 .range_cyclic = 1,
857 };
858
859 return wb_writeback(wb, &args);
860 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200861
862 return 0;
863}
864
865/*
866 * Retrieve work items and do the writeback they describe
867 */
868long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
869{
870 struct backing_dev_info *bdi = wb->bdi;
871 struct bdi_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200872 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200873
874 while ((work = get_next_work_item(bdi, wb)) != NULL) {
Jens Axboec4a77a62009-09-16 15:18:25 +0200875 struct wb_writeback_args args = work->args;
Jens Axboe03ba3782009-09-09 09:08:54 +0200876
877 /*
878 * Override sync mode, in case we must wait for completion
879 */
880 if (force_wait)
Jens Axboec4a77a62009-09-16 15:18:25 +0200881 work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200882
883 /*
884 * If this isn't a data integrity operation, just notify
885 * that we have seen this work and we are now starting it.
886 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200887 if (args.sync_mode == WB_SYNC_NONE)
Jens Axboe03ba3782009-09-09 09:08:54 +0200888 wb_clear_pending(wb, work);
889
Jens Axboec4a77a62009-09-16 15:18:25 +0200890 wrote += wb_writeback(wb, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200891
892 /*
893 * This is a data integrity writeback, so only do the
894 * notification when we have completed the work.
895 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200896 if (args.sync_mode == WB_SYNC_ALL)
Jens Axboe03ba3782009-09-09 09:08:54 +0200897 wb_clear_pending(wb, work);
898 }
899
900 /*
901 * Check for periodic writeback, kupdated() style
902 */
903 wrote += wb_check_old_data_flush(wb);
904
905 return wrote;
906}
907
908/*
909 * Handle writeback of dirty data for the device backed by this bdi. Also
910 * wakes up periodically and does kupdated style flushing.
911 */
912int bdi_writeback_task(struct bdi_writeback *wb)
913{
914 unsigned long last_active = jiffies;
915 unsigned long wait_jiffies = -1UL;
916 long pages_written;
917
918 while (!kthread_should_stop()) {
919 pages_written = wb_do_writeback(wb, 0);
920
921 if (pages_written)
922 last_active = jiffies;
923 else if (wait_jiffies != -1UL) {
924 unsigned long max_idle;
925
926 /*
927 * Longest period of inactivity that we tolerate. If we
928 * see dirty data again later, the task will get
929 * recreated automatically.
930 */
931 max_idle = max(5UL * 60 * HZ, wait_jiffies);
932 if (time_after(jiffies, max_idle + last_active))
933 break;
934 }
935
936 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
Jens Axboe49db0412009-09-15 21:27:40 +0200937 schedule_timeout_interruptible(wait_jiffies);
Jens Axboe03ba3782009-09-09 09:08:54 +0200938 try_to_freeze();
939 }
940
941 return 0;
942}
943
944/*
Jens Axboeb6e51312009-09-16 15:13:54 +0200945 * Schedule writeback for all backing devices. This does WB_SYNC_NONE
946 * writeback, for integrity writeback see bdi_sync_writeback().
Jens Axboe03ba3782009-09-09 09:08:54 +0200947 */
Jens Axboeb6e51312009-09-16 15:13:54 +0200948static void bdi_writeback_all(struct super_block *sb, long nr_pages)
Jens Axboe03ba3782009-09-09 09:08:54 +0200949{
Jens Axboeb6e51312009-09-16 15:13:54 +0200950 struct wb_writeback_args args = {
951 .sb = sb,
952 .nr_pages = nr_pages,
953 .sync_mode = WB_SYNC_NONE,
954 };
Jens Axboe03ba3782009-09-09 09:08:54 +0200955 struct backing_dev_info *bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200956
Jens Axboecfc4ba52009-09-14 13:12:40 +0200957 rcu_read_lock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200958
Jens Axboecfc4ba52009-09-14 13:12:40 +0200959 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200960 if (!bdi_has_dirty_io(bdi))
961 continue;
962
Jens Axboeb6e51312009-09-16 15:13:54 +0200963 bdi_alloc_queue_work(bdi, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200964 }
965
Jens Axboecfc4ba52009-09-14 13:12:40 +0200966 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200967}
968
969/*
970 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
971 * the whole world.
972 */
973void wakeup_flusher_threads(long nr_pages)
974{
Jens Axboe03ba3782009-09-09 09:08:54 +0200975 if (nr_pages == 0)
976 nr_pages = global_page_state(NR_FILE_DIRTY) +
977 global_page_state(NR_UNSTABLE_NFS);
Jens Axboeb6e51312009-09-16 15:13:54 +0200978 bdi_writeback_all(NULL, nr_pages);
Jens Axboe03ba3782009-09-09 09:08:54 +0200979}
980
981static noinline void block_dump___mark_inode_dirty(struct inode *inode)
982{
983 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
984 struct dentry *dentry;
985 const char *name = "?";
986
987 dentry = d_find_alias(inode);
988 if (dentry) {
989 spin_lock(&dentry->d_lock);
990 name = (const char *) dentry->d_name.name;
991 }
992 printk(KERN_DEBUG
993 "%s(%d): dirtied inode %lu (%s) on %s\n",
994 current->comm, task_pid_nr(current), inode->i_ino,
995 name, inode->i_sb->s_id);
996 if (dentry) {
997 spin_unlock(&dentry->d_lock);
998 dput(dentry);
999 }
1000 }
1001}
1002
1003/**
1004 * __mark_inode_dirty - internal function
1005 * @inode: inode to mark
1006 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1007 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1008 * mark_inode_dirty_sync.
1009 *
1010 * Put the inode on the super block's dirty list.
1011 *
1012 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1013 * dirty list only if it is hashed or if it refers to a blockdev.
1014 * If it was not hashed, it will never be added to the dirty list
1015 * even if it is later hashed, as it will have been marked dirty already.
1016 *
1017 * In short, make sure you hash any inodes _before_ you start marking
1018 * them dirty.
1019 *
1020 * This function *must* be atomic for the I_DIRTY_PAGES case -
1021 * set_page_dirty() is called under spinlock in several places.
1022 *
1023 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1024 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1025 * the kernel-internal blockdev inode represents the dirtying time of the
1026 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1027 * page->mapping->host, so the page-dirtying time is recorded in the internal
1028 * blockdev inode.
1029 */
1030void __mark_inode_dirty(struct inode *inode, int flags)
1031{
1032 struct super_block *sb = inode->i_sb;
1033
1034 /*
1035 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1036 * dirty the inode itself
1037 */
1038 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1039 if (sb->s_op->dirty_inode)
1040 sb->s_op->dirty_inode(inode);
1041 }
1042
1043 /*
1044 * make sure that changes are seen by all cpus before we test i_state
1045 * -- mikulas
1046 */
1047 smp_mb();
1048
1049 /* avoid the locking if we can */
1050 if ((inode->i_state & flags) == flags)
1051 return;
1052
1053 if (unlikely(block_dump))
1054 block_dump___mark_inode_dirty(inode);
1055
1056 spin_lock(&inode_lock);
1057 if ((inode->i_state & flags) != flags) {
1058 const int was_dirty = inode->i_state & I_DIRTY;
1059
1060 inode->i_state |= flags;
1061
1062 /*
1063 * If the inode is being synced, just update its dirty state.
1064 * The unlocker will place the inode on the appropriate
1065 * superblock list, based upon its state.
1066 */
1067 if (inode->i_state & I_SYNC)
1068 goto out;
1069
1070 /*
1071 * Only add valid (hashed) inodes to the superblock's
1072 * dirty list. Add blockdev inodes as well.
1073 */
1074 if (!S_ISBLK(inode->i_mode)) {
1075 if (hlist_unhashed(&inode->i_hash))
1076 goto out;
1077 }
1078 if (inode->i_state & (I_FREEING|I_CLEAR))
1079 goto out;
1080
1081 /*
1082 * If the inode was already on b_dirty/b_io/b_more_io, don't
1083 * reposition it (that would break b_dirty time-ordering).
1084 */
1085 if (!was_dirty) {
1086 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001087 struct backing_dev_info *bdi = wb->bdi;
1088
1089 if (bdi_cap_writeback_dirty(bdi) &&
1090 !test_bit(BDI_registered, &bdi->state)) {
1091 WARN_ON(1);
1092 printk(KERN_ERR "bdi-%s not registered\n",
1093 bdi->name);
1094 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001095
1096 inode->dirtied_when = jiffies;
1097 list_move(&inode->i_list, &wb->b_dirty);
1098 }
1099 }
1100out:
1101 spin_unlock(&inode_lock);
1102}
1103EXPORT_SYMBOL(__mark_inode_dirty);
1104
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001105/*
1106 * Write out a superblock's list of dirty inodes. A wait will be performed
1107 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1108 *
1109 * If older_than_this is non-NULL, then only write out inodes which
1110 * had their first dirtying at a time earlier than *older_than_this.
1111 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001112 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1113 * This function assumes that the blockdev superblock's inodes are backed by
1114 * a variety of queues, so all inodes are searched. For other superblocks,
1115 * assume that all inodes are backed by the same queue.
1116 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001117 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1118 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1119 * on the writer throttling path, and we get decent balancing between many
1120 * throttled threads: we don't want them all piling up on inode_sync_wait.
1121 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001122static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001123{
Jens Axboe03ba3782009-09-09 09:08:54 +02001124 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001125
Jens Axboe03ba3782009-09-09 09:08:54 +02001126 /*
1127 * We need to be protected against the filesystem going from
1128 * r/o to r/w or vice versa.
1129 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001130 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001131
Jens Axboe03ba3782009-09-09 09:08:54 +02001132 spin_lock(&inode_lock);
1133
1134 /*
1135 * Data integrity sync. Must wait for all pages under writeback,
1136 * because there may have been pages dirtied before our sync
1137 * call, but which had writeout started before we write it out.
1138 * In which case, the inode may not be on the dirty list, but
1139 * we still have to wait for that writeout.
1140 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001141 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001142 struct address_space *mapping;
1143
1144 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1145 continue;
1146 mapping = inode->i_mapping;
1147 if (mapping->nrpages == 0)
1148 continue;
1149 __iget(inode);
1150 spin_unlock(&inode_lock);
1151 /*
1152 * We hold a reference to 'inode' so it couldn't have
1153 * been removed from s_inodes list while we dropped the
1154 * inode_lock. We cannot iput the inode now as we can
1155 * be holding the last reference and we cannot iput it
1156 * under inode_lock. So we keep the reference and iput
1157 * it later.
1158 */
1159 iput(old_inode);
1160 old_inode = inode;
1161
1162 filemap_fdatawait(mapping);
1163
1164 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001165
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001166 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001167 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001168 spin_unlock(&inode_lock);
1169 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170}
1171
Jens Axboed8a85592009-09-02 12:34:32 +02001172/**
1173 * writeback_inodes_sb - writeback dirty inodes from given super_block
1174 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 *
Jens Axboed8a85592009-09-02 12:34:32 +02001176 * Start writeback on some inodes on this super_block. No guarantees are made
1177 * on how many (if any) will be written, and this function does not wait
1178 * for IO completion of submitted IO. The number of pages submitted is
1179 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001181void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182{
Jens Axboed8a85592009-09-02 12:34:32 +02001183 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1184 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1185 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Jens Axboed8a85592009-09-02 12:34:32 +02001187 nr_to_write = nr_dirty + nr_unstable +
Nick Piggin38f21972009-01-06 14:40:25 -08001188 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
Nick Piggin38f21972009-01-06 14:40:25 -08001189
Jens Axboea72bfd42009-09-26 00:07:46 +02001190 bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191}
Jens Axboed8a85592009-09-02 12:34:32 +02001192EXPORT_SYMBOL(writeback_inodes_sb);
1193
1194/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001195 * writeback_inodes_sb_if_idle - start writeback if none underway
1196 * @sb: the superblock
1197 *
1198 * Invoke writeback_inodes_sb if no writeback is currently underway.
1199 * Returns 1 if writeback was started, 0 if not.
1200 */
1201int writeback_inodes_sb_if_idle(struct super_block *sb)
1202{
1203 if (!writeback_in_progress(sb->s_bdi)) {
1204 writeback_inodes_sb(sb);
1205 return 1;
1206 } else
1207 return 0;
1208}
1209EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1210
1211/**
Jens Axboed8a85592009-09-02 12:34:32 +02001212 * sync_inodes_sb - sync sb inode pages
1213 * @sb: the superblock
1214 *
1215 * This function writes and waits on any dirty inode belonging to this
1216 * super_block. The number of pages synced is returned.
1217 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001218void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001219{
Jens Axboeb6e51312009-09-16 15:13:54 +02001220 bdi_sync_writeback(sb->s_bdi, sb);
1221 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001222}
1223EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001226 * write_inode_now - write an inode to disk
1227 * @inode: inode to write to disk
1228 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001230 * This function commits an inode to disk immediately if it is dirty. This is
1231 * primarily needed by knfsd.
1232 *
1233 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235int write_inode_now(struct inode *inode, int sync)
1236{
1237 int ret;
1238 struct writeback_control wbc = {
1239 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001240 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001241 .range_start = 0,
1242 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 };
1244
1245 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001246 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
1248 might_sleep();
1249 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001250 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 spin_unlock(&inode_lock);
1252 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001253 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 return ret;
1255}
1256EXPORT_SYMBOL(write_inode_now);
1257
1258/**
1259 * sync_inode - write an inode and its pages to disk.
1260 * @inode: the inode to sync
1261 * @wbc: controls the writeback mode
1262 *
1263 * sync_inode() will write an inode and its pages to disk. It will also
1264 * correctly update the inode on its superblock's dirty inode lists and will
1265 * update inode->i_state.
1266 *
1267 * The caller must have a ref on the inode.
1268 */
1269int sync_inode(struct inode *inode, struct writeback_control *wbc)
1270{
1271 int ret;
1272
1273 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001274 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 spin_unlock(&inode_lock);
1276 return ret;
1277}
1278EXPORT_SYMBOL(sync_inode);