blob: 49bc1b8e8f1945f1a282f162cd00037a5cefcbb9 [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
245 * @nr_pages: the number of pages to write
246 *
247 * Description:
248 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
249 * started when this function returns, we make no guarentees on
250 * completion. Caller need not hold sb s_umount semaphore.
251 *
252 */
Jens Axboea72bfd42009-09-26 00:07:46 +0200253void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
254 long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200255{
256 struct wb_writeback_args args = {
Jens Axboea72bfd42009-09-26 00:07:46 +0200257 .sb = sb,
Jens Axboeb6e51312009-09-16 15:13:54 +0200258 .sync_mode = WB_SYNC_NONE,
259 .nr_pages = nr_pages,
260 .range_cyclic = 1,
261 };
262
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800263 /*
264 * We treat @nr_pages=0 as the special case to do background writeback,
265 * ie. to sync pages until the background dirty threshold is reached.
266 */
267 if (!nr_pages) {
268 args.nr_pages = LONG_MAX;
269 args.for_background = 1;
270 }
271
Jens Axboeb6e51312009-09-16 15:13:54 +0200272 bdi_alloc_queue_work(bdi, &args);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
274
275/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700276 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
277 * furthest end of its superblock's dirty-inode list.
278 *
279 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200280 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700281 * the case then the inode must have been redirtied while it was being written
282 * out and we don't reset its dirtied_when.
283 */
284static void redirty_tail(struct inode *inode)
285{
Jens Axboe03ba3782009-09-09 09:08:54 +0200286 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700287
Jens Axboe03ba3782009-09-09 09:08:54 +0200288 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200289 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700290
Jens Axboe03ba3782009-09-09 09:08:54 +0200291 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200292 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700293 inode->dirtied_when = jiffies;
294 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200295 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700296}
297
298/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200299 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700300 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700301static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700302{
Jens Axboe03ba3782009-09-09 09:08:54 +0200303 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
304
305 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700306}
307
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700308static void inode_sync_complete(struct inode *inode)
309{
310 /*
311 * Prevent speculative execution through spin_unlock(&inode_lock);
312 */
313 smp_mb();
314 wake_up_bit(&inode->i_state, __I_SYNC);
315}
316
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700317static bool inode_dirtied_after(struct inode *inode, unsigned long t)
318{
319 bool ret = time_after(inode->dirtied_when, t);
320#ifndef CONFIG_64BIT
321 /*
322 * For inodes being constantly redirtied, dirtied_when can get stuck.
323 * It _appears_ to be in the future, but is actually in distant past.
324 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200325 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700326 */
327 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
328#endif
329 return ret;
330}
331
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700332/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700333 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
334 */
335static void move_expired_inodes(struct list_head *delaying_queue,
336 struct list_head *dispatch_queue,
337 unsigned long *older_than_this)
338{
Shaohua Li5c034492009-09-24 14:42:33 +0200339 LIST_HEAD(tmp);
340 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200341 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200342 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200343 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200344
Fengguang Wu2c136572007-10-16 23:30:39 -0700345 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200346 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700347 if (older_than_this &&
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700348 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700349 break;
Jens Axboecf137302009-09-24 15:12:57 +0200350 if (sb && sb != inode->i_sb)
351 do_sb_sort = 1;
352 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200353 list_move(&inode->i_list, &tmp);
354 }
355
Jens Axboecf137302009-09-24 15:12:57 +0200356 /* just one sb in list, splice to dispatch_queue and we're done */
357 if (!do_sb_sort) {
358 list_splice(&tmp, dispatch_queue);
359 return;
360 }
361
Shaohua Li5c034492009-09-24 14:42:33 +0200362 /* Move inodes from one superblock together */
363 while (!list_empty(&tmp)) {
364 inode = list_entry(tmp.prev, struct inode, i_list);
365 sb = inode->i_sb;
366 list_for_each_prev_safe(pos, node, &tmp) {
367 inode = list_entry(pos, struct inode, i_list);
368 if (inode->i_sb == sb)
369 list_move(&inode->i_list, dispatch_queue);
370 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700371 }
372}
373
374/*
375 * Queue all expired dirty inodes for io, eldest first.
376 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200377static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700378{
Jens Axboe03ba3782009-09-09 09:08:54 +0200379 list_splice_init(&wb->b_more_io, wb->b_io.prev);
380 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200381}
382
Jens Axboe03ba3782009-09-09 09:08:54 +0200383static int write_inode(struct inode *inode, int sync)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200384{
Jens Axboe03ba3782009-09-09 09:08:54 +0200385 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
386 return inode->i_sb->s_op->write_inode(inode, sync);
387 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700388}
389
390/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200391 * Wait for writeback on an inode to complete.
392 */
393static void inode_wait_for_writeback(struct inode *inode)
394{
395 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
396 wait_queue_head_t *wqh;
397
398 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
399 do {
400 spin_unlock(&inode_lock);
401 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
402 spin_lock(&inode_lock);
403 } while (inode->i_state & I_SYNC);
404}
405
406/*
407 * Write out an inode's dirty pages. Called under inode_lock. Either the
408 * caller has ref on the inode (either via __iget or via syscall against an fd)
409 * or the inode has I_WILL_FREE set (via generic_forget_inode)
410 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 * If `wait' is set, wait on the writeout.
412 *
413 * The whole writeout design is quite complex and fragile. We want to avoid
414 * starvation of particular inodes when others are being redirtied, prevent
415 * livelocks, etc.
416 *
417 * Called under inode_lock.
418 */
419static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200420writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 struct address_space *mapping = inode->i_mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 int wait = wbc->sync_mode == WB_SYNC_ALL;
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 */
441 if (!wait) {
442 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
463 /* Don't write the inode if only I_DIRTY_PAGES was set */
464 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
465 int err = write_inode(inode, wait);
466 if (ret == 0)
467 ret = err;
468 }
469
470 if (wait) {
471 int err = filemap_fdatawait(mapping);
472 if (ret == 0)
473 ret = err;
474 }
475
476 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700477 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700478 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangb3af9462009-09-25 06:04:10 +0200479 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800480 /*
Wu Fengguangb3af9462009-09-25 06:04:10 +0200481 * More pages get dirtied by a fast dirtier.
482 */
483 goto select_queue;
484 } else if (inode->i_state & I_DIRTY) {
485 /*
486 * At least XFS will redirty the inode during the
487 * writeback (delalloc) and on io completion (isize).
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800488 */
489 redirty_tail(inode);
490 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 /*
492 * We didn't write back all the pages. nfs_writepages()
493 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200494 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700495 */
496 /*
497 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200498 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700499 * consideration. Otherwise, move it to the tail, for
500 * the reasons described there. I'm not really sure
501 * how much sense this makes. Presumably I had a good
502 * reasons for doing it this way, and I'd rather not
503 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 */
505 if (wbc->for_kupdate) {
506 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700507 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200508 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700509 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 */
511 inode->i_state |= I_DIRTY_PAGES;
Wu Fengguangb3af9462009-09-25 06:04:10 +0200512select_queue:
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800513 if (wbc->nr_to_write <= 0) {
514 /*
515 * slice used up: queue for next turn
516 */
517 requeue_io(inode);
518 } else {
519 /*
520 * somehow blocked: retry later
521 */
522 redirty_tail(inode);
523 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 } else {
525 /*
526 * Otherwise fully redirty the inode so that
527 * other inodes on this superblock will get some
528 * writeout. Otherwise heavy writing to one
529 * file would indefinitely suspend writeout of
530 * all the other files.
531 */
532 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700533 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 } else if (atomic_read(&inode->i_count)) {
536 /*
537 * The inode is clean, inuse
538 */
539 list_move(&inode->i_list, &inode_in_use);
540 } else {
541 /*
542 * The inode is clean, unused
543 */
544 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 }
546 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700547 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 return ret;
549}
550
Jens Axboe9ecc2732009-09-24 15:25:11 +0200551static void unpin_sb_for_writeback(struct super_block **psb)
552{
553 struct super_block *sb = *psb;
554
555 if (sb) {
556 up_read(&sb->s_umount);
557 put_super(sb);
558 *psb = NULL;
559 }
560}
561
Jens Axboe03ba3782009-09-09 09:08:54 +0200562/*
563 * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
564 * before calling writeback. So make sure that we do pin it, so it doesn't
565 * go away while we are writing inodes from it.
566 *
567 * Returns 0 if the super was successfully pinned (or pinning wasn't needed),
568 * 1 if we failed.
569 */
570static int pin_sb_for_writeback(struct writeback_control *wbc,
Jens Axboe9ecc2732009-09-24 15:25:11 +0200571 struct inode *inode, struct super_block **psb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572{
Jens Axboe03ba3782009-09-09 09:08:54 +0200573 struct super_block *sb = inode->i_sb;
574
575 /*
Jens Axboe9ecc2732009-09-24 15:25:11 +0200576 * If this sb is already pinned, nothing more to do. If not and
577 * *psb is non-NULL, unpin the old one first
578 */
579 if (sb == *psb)
580 return 0;
581 else if (*psb)
582 unpin_sb_for_writeback(psb);
583
584 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200585 * Caller must already hold the ref for this
586 */
587 if (wbc->sync_mode == WB_SYNC_ALL) {
588 WARN_ON(!rwsem_is_locked(&sb->s_umount));
589 return 0;
590 }
591
592 spin_lock(&sb_lock);
593 sb->s_count++;
594 if (down_read_trylock(&sb->s_umount)) {
595 if (sb->s_root) {
596 spin_unlock(&sb_lock);
Jens Axboe9ecc2732009-09-24 15:25:11 +0200597 goto pinned;
Jens Axboe03ba3782009-09-09 09:08:54 +0200598 }
599 /*
600 * umounted, drop rwsem again and fall through to failure
601 */
602 up_read(&sb->s_umount);
603 }
604
605 sb->s_count--;
606 spin_unlock(&sb_lock);
607 return 1;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200608pinned:
609 *psb = sb;
610 return 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200611}
612
613static void writeback_inodes_wb(struct bdi_writeback *wb,
614 struct writeback_control *wbc)
615{
Jens Axboe9ecc2732009-09-24 15:25:11 +0200616 struct super_block *sb = wbc->sb, *pin_sb = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 const unsigned long start = jiffies; /* livelock avoidance */
618
Hans Reiserae8547b2008-05-07 15:48:57 +0300619 spin_lock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
Jens Axboe03ba3782009-09-09 09:08:54 +0200621 if (!wbc->for_kupdate || list_empty(&wb->b_io))
622 queue_io(wb, wbc->older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200623
Jens Axboe03ba3782009-09-09 09:08:54 +0200624 while (!list_empty(&wb->b_io)) {
625 struct inode *inode = list_entry(wb->b_io.prev,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 struct inode, i_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 long pages_skipped;
628
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200629 /*
630 * super block given and doesn't match, skip this inode
631 */
632 if (sb && sb != inode->i_sb) {
633 redirty_tail(inode);
634 continue;
635 }
636
Wu Fengguang84a89242009-06-16 15:33:17 -0700637 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700638 requeue_io(inode);
639 continue;
640 }
641
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700642 /*
643 * Was this inode dirtied after sync_sb_inodes was called?
644 * This keeps sync from extra jobs and livelock.
645 */
646 if (inode_dirtied_after(inode, start))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 break;
648
Jens Axboe9ecc2732009-09-24 15:25:11 +0200649 if (pin_sb_for_writeback(wbc, inode, &pin_sb)) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200650 requeue_io(inode);
651 continue;
652 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Wu Fengguang84a89242009-06-16 15:33:17 -0700654 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 __iget(inode);
656 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200657 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 if (wbc->pages_skipped != pages_skipped) {
659 /*
660 * writeback is not making progress due to locked
661 * buffers. Skip this inode for now.
662 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700663 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 }
665 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800667 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800669 if (wbc->nr_to_write <= 0) {
670 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800672 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200673 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800674 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
Nick Piggin38f21972009-01-06 14:40:25 -0800676
Jens Axboe9ecc2732009-09-24 15:25:11 +0200677 unpin_sb_for_writeback(&pin_sb);
678
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200679 spin_unlock(&inode_lock);
680 /* Leave any unwritten inodes on b_io */
681}
682
Jens Axboe03ba3782009-09-09 09:08:54 +0200683void writeback_inodes_wbc(struct writeback_control *wbc)
684{
685 struct backing_dev_info *bdi = wbc->bdi;
686
687 writeback_inodes_wb(&bdi->wb, wbc);
688}
689
690/*
691 * The maximum number of pages to writeout in a single bdi flush/kupdate
692 * operation. We do this so we don't hold I_SYNC against an inode for
693 * enormous amounts of time, which would block a userspace task which has
694 * been forced to throttle against that inode. Also, the code reevaluates
695 * the dirty each time it has written this many pages.
696 */
697#define MAX_WRITEBACK_PAGES 1024
698
699static inline bool over_bground_thresh(void)
700{
701 unsigned long background_thresh, dirty_thresh;
702
703 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
704
705 return (global_page_state(NR_FILE_DIRTY) +
706 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
707}
708
709/*
710 * Explicit flushing or periodic writeback of "old" data.
711 *
712 * Define "old": the first time one of an inode's pages is dirtied, we mark the
713 * dirtying-time in the inode's address_space. So this periodic writeback code
714 * just walks the superblock inode list, writing back any inodes which are
715 * older than a specific point in time.
716 *
717 * Try to run once per dirty_writeback_interval. But if a writeback event
718 * takes longer than a dirty_writeback_interval interval, then leave a
719 * one-second gap.
720 *
721 * older_than_this takes precedence over nr_to_write. So we'll only write back
722 * all dirty pages if they are all attached to "old" mappings.
723 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200724static long wb_writeback(struct bdi_writeback *wb,
725 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200726{
727 struct writeback_control wbc = {
728 .bdi = wb->bdi,
Jens Axboec4a77a62009-09-16 15:18:25 +0200729 .sb = args->sb,
730 .sync_mode = args->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200731 .older_than_this = NULL,
Jens Axboec4a77a62009-09-16 15:18:25 +0200732 .for_kupdate = args->for_kupdate,
Wu Fengguangb17621f2009-12-03 13:54:25 +0100733 .for_background = args->for_background,
Jens Axboec4a77a62009-09-16 15:18:25 +0200734 .range_cyclic = args->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200735 };
736 unsigned long oldest_jif;
737 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200738 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200739
740 if (wbc.for_kupdate) {
741 wbc.older_than_this = &oldest_jif;
742 oldest_jif = jiffies -
743 msecs_to_jiffies(dirty_expire_interval * 10);
744 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200745 if (!wbc.range_cyclic) {
746 wbc.range_start = 0;
747 wbc.range_end = LLONG_MAX;
748 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200749
750 for (;;) {
751 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800752 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200753 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800754 if (args->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200755 break;
756
757 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800758 * For background writeout, stop when we are below the
759 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200760 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800761 if (args->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200762 break;
763
764 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200765 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
766 wbc.pages_skipped = 0;
767 writeback_inodes_wb(wb, &wbc);
Jens Axboec4a77a62009-09-16 15:18:25 +0200768 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200769 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
770
771 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200772 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200773 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200774 if (wbc.nr_to_write <= 0)
775 continue;
776 /*
777 * Didn't write everything and we don't have more IO, bail
778 */
779 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200780 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200781 /*
782 * Did we write something? Try for more
783 */
784 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
785 continue;
786 /*
787 * Nothing written. Wait for some inode to
788 * become available for writeback. Otherwise
789 * we'll just busyloop.
790 */
791 spin_lock(&inode_lock);
792 if (!list_empty(&wb->b_more_io)) {
793 inode = list_entry(wb->b_more_io.prev,
794 struct inode, i_list);
795 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200796 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200797 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200798 }
799
800 return wrote;
801}
802
803/*
804 * Return the next bdi_work struct that hasn't been processed by this
Jens Axboe8010c3b2009-09-15 20:04:57 +0200805 * wb thread yet. ->seen is initially set for each thread that exists
806 * for this device, when a thread first notices a piece of work it
807 * clears its bit. Depending on writeback type, the thread will notify
808 * completion on either receiving the work (WB_SYNC_NONE) or after
809 * it is done (WB_SYNC_ALL).
Jens Axboe03ba3782009-09-09 09:08:54 +0200810 */
811static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
812 struct bdi_writeback *wb)
813{
814 struct bdi_work *work, *ret = NULL;
815
816 rcu_read_lock();
817
818 list_for_each_entry_rcu(work, &bdi->work_list, list) {
Nick Piggin77fad5e2009-09-15 21:34:12 +0200819 if (!test_bit(wb->nr, &work->seen))
Jens Axboe03ba3782009-09-09 09:08:54 +0200820 continue;
Nick Piggin77fad5e2009-09-15 21:34:12 +0200821 clear_bit(wb->nr, &work->seen);
Jens Axboe03ba3782009-09-09 09:08:54 +0200822
823 ret = work;
824 break;
825 }
826
827 rcu_read_unlock();
828 return ret;
829}
830
831static long wb_check_old_data_flush(struct bdi_writeback *wb)
832{
833 unsigned long expired;
834 long nr_pages;
835
836 expired = wb->last_old_flush +
837 msecs_to_jiffies(dirty_writeback_interval * 10);
838 if (time_before(jiffies, expired))
839 return 0;
840
841 wb->last_old_flush = jiffies;
842 nr_pages = global_page_state(NR_FILE_DIRTY) +
843 global_page_state(NR_UNSTABLE_NFS) +
844 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
845
Jens Axboec4a77a62009-09-16 15:18:25 +0200846 if (nr_pages) {
847 struct wb_writeback_args args = {
848 .nr_pages = nr_pages,
849 .sync_mode = WB_SYNC_NONE,
850 .for_kupdate = 1,
851 .range_cyclic = 1,
852 };
853
854 return wb_writeback(wb, &args);
855 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200856
857 return 0;
858}
859
860/*
861 * Retrieve work items and do the writeback they describe
862 */
863long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
864{
865 struct backing_dev_info *bdi = wb->bdi;
866 struct bdi_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200867 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200868
869 while ((work = get_next_work_item(bdi, wb)) != NULL) {
Jens Axboec4a77a62009-09-16 15:18:25 +0200870 struct wb_writeback_args args = work->args;
Jens Axboe03ba3782009-09-09 09:08:54 +0200871
872 /*
873 * Override sync mode, in case we must wait for completion
874 */
875 if (force_wait)
Jens Axboec4a77a62009-09-16 15:18:25 +0200876 work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200877
878 /*
879 * If this isn't a data integrity operation, just notify
880 * that we have seen this work and we are now starting it.
881 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200882 if (args.sync_mode == WB_SYNC_NONE)
Jens Axboe03ba3782009-09-09 09:08:54 +0200883 wb_clear_pending(wb, work);
884
Jens Axboec4a77a62009-09-16 15:18:25 +0200885 wrote += wb_writeback(wb, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200886
887 /*
888 * This is a data integrity writeback, so only do the
889 * notification when we have completed the work.
890 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200891 if (args.sync_mode == WB_SYNC_ALL)
Jens Axboe03ba3782009-09-09 09:08:54 +0200892 wb_clear_pending(wb, work);
893 }
894
895 /*
896 * Check for periodic writeback, kupdated() style
897 */
898 wrote += wb_check_old_data_flush(wb);
899
900 return wrote;
901}
902
903/*
904 * Handle writeback of dirty data for the device backed by this bdi. Also
905 * wakes up periodically and does kupdated style flushing.
906 */
907int bdi_writeback_task(struct bdi_writeback *wb)
908{
909 unsigned long last_active = jiffies;
910 unsigned long wait_jiffies = -1UL;
911 long pages_written;
912
913 while (!kthread_should_stop()) {
914 pages_written = wb_do_writeback(wb, 0);
915
916 if (pages_written)
917 last_active = jiffies;
918 else if (wait_jiffies != -1UL) {
919 unsigned long max_idle;
920
921 /*
922 * Longest period of inactivity that we tolerate. If we
923 * see dirty data again later, the task will get
924 * recreated automatically.
925 */
926 max_idle = max(5UL * 60 * HZ, wait_jiffies);
927 if (time_after(jiffies, max_idle + last_active))
928 break;
929 }
930
931 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
Jens Axboe49db0412009-09-15 21:27:40 +0200932 schedule_timeout_interruptible(wait_jiffies);
Jens Axboe03ba3782009-09-09 09:08:54 +0200933 try_to_freeze();
934 }
935
936 return 0;
937}
938
939/*
Jens Axboeb6e51312009-09-16 15:13:54 +0200940 * Schedule writeback for all backing devices. This does WB_SYNC_NONE
941 * writeback, for integrity writeback see bdi_sync_writeback().
Jens Axboe03ba3782009-09-09 09:08:54 +0200942 */
Jens Axboeb6e51312009-09-16 15:13:54 +0200943static void bdi_writeback_all(struct super_block *sb, long nr_pages)
Jens Axboe03ba3782009-09-09 09:08:54 +0200944{
Jens Axboeb6e51312009-09-16 15:13:54 +0200945 struct wb_writeback_args args = {
946 .sb = sb,
947 .nr_pages = nr_pages,
948 .sync_mode = WB_SYNC_NONE,
949 };
Jens Axboe03ba3782009-09-09 09:08:54 +0200950 struct backing_dev_info *bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200951
Jens Axboecfc4ba52009-09-14 13:12:40 +0200952 rcu_read_lock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200953
Jens Axboecfc4ba52009-09-14 13:12:40 +0200954 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200955 if (!bdi_has_dirty_io(bdi))
956 continue;
957
Jens Axboeb6e51312009-09-16 15:13:54 +0200958 bdi_alloc_queue_work(bdi, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200959 }
960
Jens Axboecfc4ba52009-09-14 13:12:40 +0200961 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200962}
963
964/*
965 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
966 * the whole world.
967 */
968void wakeup_flusher_threads(long nr_pages)
969{
Jens Axboe03ba3782009-09-09 09:08:54 +0200970 if (nr_pages == 0)
971 nr_pages = global_page_state(NR_FILE_DIRTY) +
972 global_page_state(NR_UNSTABLE_NFS);
Jens Axboeb6e51312009-09-16 15:13:54 +0200973 bdi_writeback_all(NULL, nr_pages);
Jens Axboe03ba3782009-09-09 09:08:54 +0200974}
975
976static noinline void block_dump___mark_inode_dirty(struct inode *inode)
977{
978 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
979 struct dentry *dentry;
980 const char *name = "?";
981
982 dentry = d_find_alias(inode);
983 if (dentry) {
984 spin_lock(&dentry->d_lock);
985 name = (const char *) dentry->d_name.name;
986 }
987 printk(KERN_DEBUG
988 "%s(%d): dirtied inode %lu (%s) on %s\n",
989 current->comm, task_pid_nr(current), inode->i_ino,
990 name, inode->i_sb->s_id);
991 if (dentry) {
992 spin_unlock(&dentry->d_lock);
993 dput(dentry);
994 }
995 }
996}
997
998/**
999 * __mark_inode_dirty - internal function
1000 * @inode: inode to mark
1001 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1002 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1003 * mark_inode_dirty_sync.
1004 *
1005 * Put the inode on the super block's dirty list.
1006 *
1007 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1008 * dirty list only if it is hashed or if it refers to a blockdev.
1009 * If it was not hashed, it will never be added to the dirty list
1010 * even if it is later hashed, as it will have been marked dirty already.
1011 *
1012 * In short, make sure you hash any inodes _before_ you start marking
1013 * them dirty.
1014 *
1015 * This function *must* be atomic for the I_DIRTY_PAGES case -
1016 * set_page_dirty() is called under spinlock in several places.
1017 *
1018 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1019 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1020 * the kernel-internal blockdev inode represents the dirtying time of the
1021 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1022 * page->mapping->host, so the page-dirtying time is recorded in the internal
1023 * blockdev inode.
1024 */
1025void __mark_inode_dirty(struct inode *inode, int flags)
1026{
1027 struct super_block *sb = inode->i_sb;
1028
1029 /*
1030 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1031 * dirty the inode itself
1032 */
1033 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1034 if (sb->s_op->dirty_inode)
1035 sb->s_op->dirty_inode(inode);
1036 }
1037
1038 /*
1039 * make sure that changes are seen by all cpus before we test i_state
1040 * -- mikulas
1041 */
1042 smp_mb();
1043
1044 /* avoid the locking if we can */
1045 if ((inode->i_state & flags) == flags)
1046 return;
1047
1048 if (unlikely(block_dump))
1049 block_dump___mark_inode_dirty(inode);
1050
1051 spin_lock(&inode_lock);
1052 if ((inode->i_state & flags) != flags) {
1053 const int was_dirty = inode->i_state & I_DIRTY;
1054
1055 inode->i_state |= flags;
1056
1057 /*
1058 * If the inode is being synced, just update its dirty state.
1059 * The unlocker will place the inode on the appropriate
1060 * superblock list, based upon its state.
1061 */
1062 if (inode->i_state & I_SYNC)
1063 goto out;
1064
1065 /*
1066 * Only add valid (hashed) inodes to the superblock's
1067 * dirty list. Add blockdev inodes as well.
1068 */
1069 if (!S_ISBLK(inode->i_mode)) {
1070 if (hlist_unhashed(&inode->i_hash))
1071 goto out;
1072 }
1073 if (inode->i_state & (I_FREEING|I_CLEAR))
1074 goto out;
1075
1076 /*
1077 * If the inode was already on b_dirty/b_io/b_more_io, don't
1078 * reposition it (that would break b_dirty time-ordering).
1079 */
1080 if (!was_dirty) {
1081 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001082 struct backing_dev_info *bdi = wb->bdi;
1083
1084 if (bdi_cap_writeback_dirty(bdi) &&
1085 !test_bit(BDI_registered, &bdi->state)) {
1086 WARN_ON(1);
1087 printk(KERN_ERR "bdi-%s not registered\n",
1088 bdi->name);
1089 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001090
1091 inode->dirtied_when = jiffies;
1092 list_move(&inode->i_list, &wb->b_dirty);
1093 }
1094 }
1095out:
1096 spin_unlock(&inode_lock);
1097}
1098EXPORT_SYMBOL(__mark_inode_dirty);
1099
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001100/*
1101 * Write out a superblock's list of dirty inodes. A wait will be performed
1102 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1103 *
1104 * If older_than_this is non-NULL, then only write out inodes which
1105 * had their first dirtying at a time earlier than *older_than_this.
1106 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001107 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1108 * This function assumes that the blockdev superblock's inodes are backed by
1109 * a variety of queues, so all inodes are searched. For other superblocks,
1110 * assume that all inodes are backed by the same queue.
1111 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001112 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1113 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1114 * on the writer throttling path, and we get decent balancing between many
1115 * throttled threads: we don't want them all piling up on inode_sync_wait.
1116 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001117static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001118{
Jens Axboe03ba3782009-09-09 09:08:54 +02001119 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001120
Jens Axboe03ba3782009-09-09 09:08:54 +02001121 /*
1122 * We need to be protected against the filesystem going from
1123 * r/o to r/w or vice versa.
1124 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001125 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001126
Jens Axboe03ba3782009-09-09 09:08:54 +02001127 spin_lock(&inode_lock);
1128
1129 /*
1130 * Data integrity sync. Must wait for all pages under writeback,
1131 * because there may have been pages dirtied before our sync
1132 * call, but which had writeout started before we write it out.
1133 * In which case, the inode may not be on the dirty list, but
1134 * we still have to wait for that writeout.
1135 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001136 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001137 struct address_space *mapping;
1138
1139 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1140 continue;
1141 mapping = inode->i_mapping;
1142 if (mapping->nrpages == 0)
1143 continue;
1144 __iget(inode);
1145 spin_unlock(&inode_lock);
1146 /*
1147 * We hold a reference to 'inode' so it couldn't have
1148 * been removed from s_inodes list while we dropped the
1149 * inode_lock. We cannot iput the inode now as we can
1150 * be holding the last reference and we cannot iput it
1151 * under inode_lock. So we keep the reference and iput
1152 * it later.
1153 */
1154 iput(old_inode);
1155 old_inode = inode;
1156
1157 filemap_fdatawait(mapping);
1158
1159 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001160
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001161 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001162 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001163 spin_unlock(&inode_lock);
1164 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165}
1166
Jens Axboed8a85592009-09-02 12:34:32 +02001167/**
1168 * writeback_inodes_sb - writeback dirty inodes from given super_block
1169 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 *
Jens Axboed8a85592009-09-02 12:34:32 +02001171 * Start writeback on some inodes on this super_block. No guarantees are made
1172 * on how many (if any) will be written, and this function does not wait
1173 * for IO completion of submitted IO. The number of pages submitted is
1174 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001176void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177{
Jens Axboed8a85592009-09-02 12:34:32 +02001178 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1179 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1180 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Jens Axboed8a85592009-09-02 12:34:32 +02001182 nr_to_write = nr_dirty + nr_unstable +
Nick Piggin38f21972009-01-06 14:40:25 -08001183 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
Nick Piggin38f21972009-01-06 14:40:25 -08001184
Jens Axboea72bfd42009-09-26 00:07:46 +02001185 bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186}
Jens Axboed8a85592009-09-02 12:34:32 +02001187EXPORT_SYMBOL(writeback_inodes_sb);
1188
1189/**
1190 * sync_inodes_sb - sync sb inode pages
1191 * @sb: the superblock
1192 *
1193 * This function writes and waits on any dirty inode belonging to this
1194 * super_block. The number of pages synced is returned.
1195 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001196void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001197{
Jens Axboeb6e51312009-09-16 15:13:54 +02001198 bdi_sync_writeback(sb->s_bdi, sb);
1199 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001200}
1201EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001204 * write_inode_now - write an inode to disk
1205 * @inode: inode to write to disk
1206 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001208 * This function commits an inode to disk immediately if it is dirty. This is
1209 * primarily needed by knfsd.
1210 *
1211 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213int write_inode_now(struct inode *inode, int sync)
1214{
1215 int ret;
1216 struct writeback_control wbc = {
1217 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001218 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001219 .range_start = 0,
1220 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 };
1222
1223 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001224 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225
1226 might_sleep();
1227 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001228 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 spin_unlock(&inode_lock);
1230 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001231 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 return ret;
1233}
1234EXPORT_SYMBOL(write_inode_now);
1235
1236/**
1237 * sync_inode - write an inode and its pages to disk.
1238 * @inode: the inode to sync
1239 * @wbc: controls the writeback mode
1240 *
1241 * sync_inode() will write an inode and its pages to disk. It will also
1242 * correctly update the inode on its superblock's dirty inode lists and will
1243 * update inode->i_state.
1244 *
1245 * The caller must have a ref on the inode.
1246 */
1247int sync_inode(struct inode *inode, struct writeback_control *wbc)
1248{
1249 int ret;
1250
1251 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001252 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 spin_unlock(&inode_lock);
1254 return ret;
1255}
1256EXPORT_SYMBOL(sync_inode);