blob: 1d1088f48bc2dfe713d53a9af0fa4d51e1161428 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020017#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020023#include <linux/kthread.h>
24#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
28#include <linux/buffer_head.h>
David Howells07f3f052006-09-30 20:52:18 +020029#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070030
Jens Axboe66f3b8e2009-09-02 09:19:46 +020031#define inode_to_bdi(inode) ((inode)->i_mapping->backing_dev_info)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070032
Jens Axboe03ba3782009-09-09 09:08:54 +020033/*
Jens Axboed0bceac2009-05-18 08:20:32 +020034 * We don't actually have pdflush, but this one is exported though /proc...
35 */
36int nr_pdflush_threads;
37
38/*
Jens Axboec4a77a62009-09-16 15:18:25 +020039 * Passed into wb_writeback(), essentially a subset of writeback_control
40 */
41struct wb_writeback_args {
42 long nr_pages;
43 struct super_block *sb;
44 enum writeback_sync_modes sync_mode;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050045 unsigned int for_kupdate:1;
46 unsigned int range_cyclic:1;
47 unsigned int for_background:1;
Jens Axboec4a77a62009-09-16 15:18:25 +020048};
49
50/*
Jens Axboe03ba3782009-09-09 09:08:54 +020051 * Work items for the bdi_writeback threads
Adrian Bunkf11b00f2008-04-29 00:58:56 -070052 */
Jens Axboe03ba3782009-09-09 09:08:54 +020053struct bdi_work {
Jens Axboe8010c3b2009-09-15 20:04:57 +020054 struct list_head list; /* pending work list */
55 struct rcu_head rcu_head; /* for RCU free/clear of work */
Jens Axboe03ba3782009-09-09 09:08:54 +020056
Jens Axboe8010c3b2009-09-15 20:04:57 +020057 unsigned long seen; /* threads that have seen this work */
58 atomic_t pending; /* number of threads still to do work */
Jens Axboe03ba3782009-09-09 09:08:54 +020059
Jens Axboe8010c3b2009-09-15 20:04:57 +020060 struct wb_writeback_args args; /* writeback arguments */
Jens Axboe03ba3782009-09-09 09:08:54 +020061
Jens Axboe8010c3b2009-09-15 20:04:57 +020062 unsigned long state; /* flag bits, see WS_* */
Jens Axboe03ba3782009-09-09 09:08:54 +020063};
64
65enum {
66 WS_USED_B = 0,
67 WS_ONSTACK_B,
68};
69
70#define WS_USED (1 << WS_USED_B)
71#define WS_ONSTACK (1 << WS_ONSTACK_B)
72
73static inline bool bdi_work_on_stack(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070074{
Jens Axboe03ba3782009-09-09 09:08:54 +020075 return test_bit(WS_ONSTACK_B, &work->state);
76}
77
78static inline void bdi_work_init(struct bdi_work *work,
Jens Axboeb6e51312009-09-16 15:13:54 +020079 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +020080{
81 INIT_RCU_HEAD(&work->rcu_head);
Jens Axboeb6e51312009-09-16 15:13:54 +020082 work->args = *args;
Jens Axboe03ba3782009-09-09 09:08:54 +020083 work->state = WS_USED;
84}
85
Adrian Bunkf11b00f2008-04-29 00:58:56 -070086/**
87 * writeback_in_progress - determine whether there is writeback in progress
88 * @bdi: the device's backing_dev_info structure.
89 *
Jens Axboe03ba3782009-09-09 09:08:54 +020090 * Determine whether there is writeback waiting to be handled against a
91 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070092 */
93int writeback_in_progress(struct backing_dev_info *bdi)
94{
Jens Axboe03ba3782009-09-09 09:08:54 +020095 return !list_empty(&bdi->work_list);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070096}
97
Jens Axboe03ba3782009-09-09 09:08:54 +020098static void bdi_work_clear(struct bdi_work *work)
Adrian Bunkf11b00f2008-04-29 00:58:56 -070099{
Jens Axboe03ba3782009-09-09 09:08:54 +0200100 clear_bit(WS_USED_B, &work->state);
101 smp_mb__after_clear_bit();
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200102 /*
103 * work can have disappeared at this point. bit waitq functions
104 * should be able to tolerate this, provided bdi_sched_wait does
105 * not dereference it's pointer argument.
106 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200107 wake_up_bit(&work->state, WS_USED_B);
Adrian Bunkf11b00f2008-04-29 00:58:56 -0700108}
109
Jens Axboe03ba3782009-09-09 09:08:54 +0200110static void bdi_work_free(struct rcu_head *head)
Nick Piggin4195f732009-05-28 09:01:15 +0200111{
Jens Axboe03ba3782009-09-09 09:08:54 +0200112 struct bdi_work *work = container_of(head, struct bdi_work, rcu_head);
Nick Piggin4195f732009-05-28 09:01:15 +0200113
Jens Axboe03ba3782009-09-09 09:08:54 +0200114 if (!bdi_work_on_stack(work))
115 kfree(work);
116 else
117 bdi_work_clear(work);
118}
119
120static void wb_work_complete(struct bdi_work *work)
121{
Jens Axboec4a77a62009-09-16 15:18:25 +0200122 const enum writeback_sync_modes sync_mode = work->args.sync_mode;
Nick Piggin77b9d052009-09-15 21:34:51 +0200123 int onstack = bdi_work_on_stack(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200124
125 /*
126 * For allocated work, we can clear the done/seen bit right here.
127 * For on-stack work, we need to postpone both the clear and free
128 * to after the RCU grace period, since the stack could be invalidated
129 * as soon as bdi_work_clear() has done the wakeup.
130 */
Nick Piggin77b9d052009-09-15 21:34:51 +0200131 if (!onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200132 bdi_work_clear(work);
Nick Piggin77b9d052009-09-15 21:34:51 +0200133 if (sync_mode == WB_SYNC_NONE || onstack)
Jens Axboe03ba3782009-09-09 09:08:54 +0200134 call_rcu(&work->rcu_head, bdi_work_free);
135}
136
137static void wb_clear_pending(struct bdi_writeback *wb, struct bdi_work *work)
138{
139 /*
140 * The caller has retrieved the work arguments from this work,
141 * drop our reference. If this is the last ref, delete and free it
142 */
143 if (atomic_dec_and_test(&work->pending)) {
144 struct backing_dev_info *bdi = wb->bdi;
145
146 spin_lock(&bdi->wb_lock);
147 list_del_rcu(&work->list);
148 spin_unlock(&bdi->wb_lock);
149
150 wb_work_complete(work);
Nick Piggin4195f732009-05-28 09:01:15 +0200151 }
152}
153
Jens Axboe03ba3782009-09-09 09:08:54 +0200154static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155{
Jens Axboebcddc3f2009-09-13 20:07:36 +0200156 work->seen = bdi->wb_mask;
157 BUG_ON(!work->seen);
158 atomic_set(&work->pending, bdi->wb_cnt);
159 BUG_ON(!bdi->wb_cnt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
Jens Axboebcddc3f2009-09-13 20:07:36 +0200161 /*
Nick Piggindeed62e2009-09-15 21:32:58 +0200162 * list_add_tail_rcu() contains the necessary barriers to
163 * make sure the above stores are seen before the item is
164 * noticed on the list
Jens Axboebcddc3f2009-09-13 20:07:36 +0200165 */
Jens Axboebcddc3f2009-09-13 20:07:36 +0200166 spin_lock(&bdi->wb_lock);
167 list_add_tail_rcu(&work->list, &bdi->work_list);
168 spin_unlock(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
170 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200171 * If the default thread isn't there, make sure we add it. When
172 * it gets created and wakes up, we'll run this work.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200174 if (unlikely(list_empty_careful(&bdi->wb_list)))
175 wake_up_process(default_backing_dev_info.wb.task);
176 else {
177 struct bdi_writeback *wb = &bdi->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
Nick Piggin1ef7d9a2009-09-15 21:37:55 +0200179 if (wb->task)
Jens Axboe03ba3782009-09-09 09:08:54 +0200180 wake_up_process(wb->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183
Jens Axboe03ba3782009-09-09 09:08:54 +0200184/*
185 * Used for on-stack allocated work items. The caller needs to wait until
186 * the wb threads have acked the work before it's safe to continue.
187 */
188static void bdi_wait_on_work_clear(struct bdi_work *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189{
Jens Axboe03ba3782009-09-09 09:08:54 +0200190 wait_on_bit(&work->state, WS_USED_B, bdi_sched_wait,
191 TASK_UNINTERRUPTIBLE);
192}
193
Jens Axboef11fcae2009-09-15 09:53:35 +0200194static void bdi_alloc_queue_work(struct backing_dev_info *bdi,
Jens Axboef17625b2010-06-01 11:05:22 +0200195 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200196{
197 struct bdi_work *work;
198
Jens Axboebcddc3f2009-09-13 20:07:36 +0200199 /*
200 * This is WB_SYNC_NONE writeback, so if allocation fails just
201 * wakeup the thread for old dirty data writeback
202 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200203 work = kmalloc(sizeof(*work), GFP_ATOMIC);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200204 if (work) {
Jens Axboeb6e51312009-09-16 15:13:54 +0200205 bdi_work_init(work, args);
Jens Axboebcddc3f2009-09-13 20:07:36 +0200206 bdi_queue_work(bdi, work);
207 } else {
208 struct bdi_writeback *wb = &bdi->wb;
Jens Axboe03ba3782009-09-09 09:08:54 +0200209
Jens Axboebcddc3f2009-09-13 20:07:36 +0200210 if (wb->task)
211 wake_up_process(wb->task);
212 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200213}
214
Jens Axboeb6e51312009-09-16 15:13:54 +0200215/**
216 * bdi_sync_writeback - start and wait for writeback
217 * @bdi: the backing device to write from
218 * @sb: write inodes from this super_block
219 *
220 * Description:
221 * This does WB_SYNC_ALL data integrity writeback and waits for the
222 * IO to complete. Callers must hold the sb s_umount semaphore for
223 * reading, to avoid having the super disappear before we are done.
224 */
225static void bdi_sync_writeback(struct backing_dev_info *bdi,
226 struct super_block *sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200227{
Jens Axboeb6e51312009-09-16 15:13:54 +0200228 struct wb_writeback_args args = {
229 .sb = sb,
230 .sync_mode = WB_SYNC_ALL,
231 .nr_pages = LONG_MAX,
232 .range_cyclic = 0,
233 };
234 struct bdi_work work;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200235
Jens Axboeb6e51312009-09-16 15:13:54 +0200236 bdi_work_init(&work, &args);
237 work.state |= WS_ONSTACK;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200238
Jens Axboeb6e51312009-09-16 15:13:54 +0200239 bdi_queue_work(bdi, &work);
240 bdi_wait_on_work_clear(&work);
241}
242
243/**
244 * bdi_start_writeback - start writeback
245 * @bdi: the backing device to write from
Jaswinder Singh Rajput4b6764f2010-01-01 20:35:23 -0800246 * @sb: write inodes from this super_block
Jens Axboeb6e51312009-09-16 15:13:54 +0200247 * @nr_pages: the number of pages to write
248 *
249 * Description:
250 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
251 * started when this function returns, we make no guarentees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200252 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200253 *
254 */
Jens Axboea72bfd42009-09-26 00:07:46 +0200255void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200256 long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200257{
258 struct wb_writeback_args args = {
Jens Axboea72bfd42009-09-26 00:07:46 +0200259 .sb = sb,
Jens Axboeb6e51312009-09-16 15:13:54 +0200260 .sync_mode = WB_SYNC_NONE,
261 .nr_pages = nr_pages,
262 .range_cyclic = 1,
263 };
264
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800265 /*
266 * We treat @nr_pages=0 as the special case to do background writeback,
267 * ie. to sync pages until the background dirty threshold is reached.
268 */
269 if (!nr_pages) {
270 args.nr_pages = LONG_MAX;
271 args.for_background = 1;
272 }
273
Jens Axboef17625b2010-06-01 11:05:22 +0200274 bdi_alloc_queue_work(bdi, &args);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275}
276
277/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700278 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
279 * furthest end of its superblock's dirty-inode list.
280 *
281 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200282 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700283 * the case then the inode must have been redirtied while it was being written
284 * out and we don't reset its dirtied_when.
285 */
286static void redirty_tail(struct inode *inode)
287{
Jens Axboe03ba3782009-09-09 09:08:54 +0200288 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700289
Jens Axboe03ba3782009-09-09 09:08:54 +0200290 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200291 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700292
Jens Axboe03ba3782009-09-09 09:08:54 +0200293 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200294 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700295 inode->dirtied_when = jiffies;
296 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200297 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700298}
299
300/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200301 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700302 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700303static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700304{
Jens Axboe03ba3782009-09-09 09:08:54 +0200305 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
306
307 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700308}
309
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700310static void inode_sync_complete(struct inode *inode)
311{
312 /*
313 * Prevent speculative execution through spin_unlock(&inode_lock);
314 */
315 smp_mb();
316 wake_up_bit(&inode->i_state, __I_SYNC);
317}
318
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700319static bool inode_dirtied_after(struct inode *inode, unsigned long t)
320{
321 bool ret = time_after(inode->dirtied_when, t);
322#ifndef CONFIG_64BIT
323 /*
324 * For inodes being constantly redirtied, dirtied_when can get stuck.
325 * It _appears_ to be in the future, but is actually in distant past.
326 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200327 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700328 */
329 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
330#endif
331 return ret;
332}
333
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700334/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700335 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
336 */
337static void move_expired_inodes(struct list_head *delaying_queue,
338 struct list_head *dispatch_queue,
339 unsigned long *older_than_this)
340{
Shaohua Li5c034492009-09-24 14:42:33 +0200341 LIST_HEAD(tmp);
342 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200343 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200344 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200345 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200346
Fengguang Wu2c136572007-10-16 23:30:39 -0700347 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200348 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700349 if (older_than_this &&
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700350 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700351 break;
Jens Axboecf137302009-09-24 15:12:57 +0200352 if (sb && sb != inode->i_sb)
353 do_sb_sort = 1;
354 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200355 list_move(&inode->i_list, &tmp);
356 }
357
Jens Axboecf137302009-09-24 15:12:57 +0200358 /* just one sb in list, splice to dispatch_queue and we're done */
359 if (!do_sb_sort) {
360 list_splice(&tmp, dispatch_queue);
361 return;
362 }
363
Shaohua Li5c034492009-09-24 14:42:33 +0200364 /* Move inodes from one superblock together */
365 while (!list_empty(&tmp)) {
366 inode = list_entry(tmp.prev, struct inode, i_list);
367 sb = inode->i_sb;
368 list_for_each_prev_safe(pos, node, &tmp) {
369 inode = list_entry(pos, struct inode, i_list);
370 if (inode->i_sb == sb)
371 list_move(&inode->i_list, dispatch_queue);
372 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700373 }
374}
375
376/*
377 * Queue all expired dirty inodes for io, eldest first.
378 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200379static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700380{
Jens Axboe03ba3782009-09-09 09:08:54 +0200381 list_splice_init(&wb->b_more_io, wb->b_io.prev);
382 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200383}
384
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100385static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200386{
Jens Axboe03ba3782009-09-09 09:08:54 +0200387 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100388 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200389 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700390}
391
392/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200393 * Wait for writeback on an inode to complete.
394 */
395static void inode_wait_for_writeback(struct inode *inode)
396{
397 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
398 wait_queue_head_t *wqh;
399
400 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700401 while (inode->i_state & I_SYNC) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200402 spin_unlock(&inode_lock);
403 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
404 spin_lock(&inode_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700405 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200406}
407
408/*
409 * Write out an inode's dirty pages. Called under inode_lock. Either the
410 * caller has ref on the inode (either via __iget or via syscall against an fd)
411 * or the inode has I_WILL_FREE set (via generic_forget_inode)
412 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 * If `wait' is set, wait on the writeout.
414 *
415 * The whole writeout design is quite complex and fragile. We want to avoid
416 * starvation of particular inodes when others are being redirtied, prevent
417 * livelocks, etc.
418 *
419 * Called under inode_lock.
420 */
421static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200422writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200425 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 int ret;
427
Christoph Hellwig01c03192009-06-08 13:35:40 +0200428 if (!atomic_read(&inode->i_count))
429 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
430 else
431 WARN_ON(inode->i_state & I_WILL_FREE);
432
433 if (inode->i_state & I_SYNC) {
434 /*
435 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200436 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200437 * writeback can proceed with the other inodes on s_io.
438 *
439 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200440 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200441 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100442 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200443 requeue_io(inode);
444 return 0;
445 }
446
447 /*
448 * It's a data-integrity sync. We must wait.
449 */
450 inode_wait_for_writeback(inode);
451 }
452
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700453 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400455 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700456 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400457 inode->i_state &= ~I_DIRTY_PAGES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 spin_unlock(&inode_lock);
459
460 ret = do_writepages(mapping, wbc);
461
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100462 /*
463 * Make sure to wait on the data before writing out the metadata.
464 * This is important for filesystems that modify metadata on data
465 * I/O completion.
466 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100467 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100468 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 if (ret == 0)
470 ret = err;
471 }
472
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400473 /*
474 * Some filesystems may redirty the inode during the writeback
475 * due to delalloc, clear dirty metadata flags right before
476 * write_inode()
477 */
478 spin_lock(&inode_lock);
479 dirty = inode->i_state & I_DIRTY;
480 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
481 spin_unlock(&inode_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100482 /* Don't write the inode if only I_DIRTY_PAGES was set */
483 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100484 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 if (ret == 0)
486 ret = err;
487 }
488
489 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700490 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700491 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangb3af9462009-09-25 06:04:10 +0200492 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800493 /*
Wu Fengguangb3af9462009-09-25 06:04:10 +0200494 * More pages get dirtied by a fast dirtier.
495 */
496 goto select_queue;
497 } else if (inode->i_state & I_DIRTY) {
498 /*
499 * At least XFS will redirty the inode during the
500 * writeback (delalloc) and on io completion (isize).
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800501 */
502 redirty_tail(inode);
503 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 /*
505 * We didn't write back all the pages. nfs_writepages()
506 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200507 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700508 */
509 /*
510 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200511 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700512 * consideration. Otherwise, move it to the tail, for
513 * the reasons described there. I'm not really sure
514 * how much sense this makes. Presumably I had a good
515 * reasons for doing it this way, and I'd rather not
516 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 */
518 if (wbc->for_kupdate) {
519 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700520 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200521 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700522 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 */
524 inode->i_state |= I_DIRTY_PAGES;
Wu Fengguangb3af9462009-09-25 06:04:10 +0200525select_queue:
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800526 if (wbc->nr_to_write <= 0) {
527 /*
528 * slice used up: queue for next turn
529 */
530 requeue_io(inode);
531 } else {
532 /*
533 * somehow blocked: retry later
534 */
535 redirty_tail(inode);
536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 } else {
538 /*
539 * Otherwise fully redirty the inode so that
540 * other inodes on this superblock will get some
541 * writeout. Otherwise heavy writing to one
542 * file would indefinitely suspend writeout of
543 * all the other files.
544 */
545 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700546 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 } else if (atomic_read(&inode->i_count)) {
549 /*
550 * The inode is clean, inuse
551 */
552 list_move(&inode->i_list, &inode_in_use);
553 } else {
554 /*
555 * The inode is clean, unused
556 */
557 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 }
559 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700560 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 return ret;
562}
563
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800564static void unpin_sb_for_writeback(struct super_block *sb)
Jens Axboe9ecc2732009-09-24 15:25:11 +0200565{
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800566 up_read(&sb->s_umount);
567 put_super(sb);
Jens Axboe9ecc2732009-09-24 15:25:11 +0200568}
569
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800570enum sb_pin_state {
571 SB_PINNED,
572 SB_NOT_PINNED,
573 SB_PIN_FAILED
574};
575
Jens Axboe03ba3782009-09-09 09:08:54 +0200576/*
577 * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
578 * before calling writeback. So make sure that we do pin it, so it doesn't
579 * go away while we are writing inodes from it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200580 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800581static enum sb_pin_state pin_sb_for_writeback(struct writeback_control *wbc,
582 struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583{
Jens Axboe9ecc2732009-09-24 15:25:11 +0200584 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200585 * Caller must already hold the ref for this
586 */
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200587 if (wbc->sync_mode == WB_SYNC_ALL) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200588 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800589 return SB_NOT_PINNED;
Jens Axboe03ba3782009-09-09 09:08:54 +0200590 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200591 spin_lock(&sb_lock);
592 sb->s_count++;
593 if (down_read_trylock(&sb->s_umount)) {
594 if (sb->s_root) {
595 spin_unlock(&sb_lock);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800596 return SB_PINNED;
Jens Axboe03ba3782009-09-09 09:08:54 +0200597 }
598 /*
599 * umounted, drop rwsem again and fall through to failure
600 */
601 up_read(&sb->s_umount);
602 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200603 sb->s_count--;
604 spin_unlock(&sb_lock);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800605 return SB_PIN_FAILED;
Jens Axboe03ba3782009-09-09 09:08:54 +0200606}
607
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800608/*
609 * Write a portion of b_io inodes which belong to @sb.
610 * If @wbc->sb != NULL, then find and write all such
611 * inodes. Otherwise write only ones which go sequentially
612 * in reverse order.
613 * Return 1, if the caller writeback routine should be
614 * interrupted. Otherwise return 0.
615 */
616static int writeback_sb_inodes(struct super_block *sb,
617 struct bdi_writeback *wb,
618 struct writeback_control *wbc)
Jens Axboe03ba3782009-09-09 09:08:54 +0200619{
Jens Axboe03ba3782009-09-09 09:08:54 +0200620 while (!list_empty(&wb->b_io)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 long pages_skipped;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800622 struct inode *inode = list_entry(wb->b_io.prev,
623 struct inode, i_list);
624 if (wbc->sb && sb != inode->i_sb) {
625 /* super block given and doesn't
626 match, skip this inode */
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200627 redirty_tail(inode);
628 continue;
629 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800630 if (sb != inode->i_sb)
631 /* finish with this superblock */
632 return 0;
Wu Fengguang84a89242009-06-16 15:33:17 -0700633 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700634 requeue_io(inode);
635 continue;
636 }
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700637 /*
638 * Was this inode dirtied after sync_sb_inodes was called?
639 * This keeps sync from extra jobs and livelock.
640 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800641 if (inode_dirtied_after(inode, wbc->wb_start))
642 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
Wu Fengguang84a89242009-06-16 15:33:17 -0700644 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 __iget(inode);
646 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200647 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648 if (wbc->pages_skipped != pages_skipped) {
649 /*
650 * writeback is not making progress due to locked
651 * buffers. Skip this inode for now.
652 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700653 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 }
655 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800657 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800659 if (wbc->nr_to_write <= 0) {
660 wbc->more_io = 1;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800661 return 1;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800662 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200663 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800664 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800666 /* b_io is empty */
667 return 1;
668}
Nick Piggin38f21972009-01-06 14:40:25 -0800669
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800670static void writeback_inodes_wb(struct bdi_writeback *wb,
671 struct writeback_control *wbc)
672{
673 int ret = 0;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200674
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800675 wbc->wb_start = jiffies; /* livelock avoidance */
676 spin_lock(&inode_lock);
677 if (!wbc->for_kupdate || list_empty(&wb->b_io))
678 queue_io(wb, wbc->older_than_this);
679
680 while (!list_empty(&wb->b_io)) {
681 struct inode *inode = list_entry(wb->b_io.prev,
682 struct inode, i_list);
683 struct super_block *sb = inode->i_sb;
684 enum sb_pin_state state;
685
686 if (wbc->sb && sb != wbc->sb) {
687 /* super block given and doesn't
688 match, skip this inode */
689 redirty_tail(inode);
690 continue;
691 }
692 state = pin_sb_for_writeback(wbc, sb);
693
694 if (state == SB_PIN_FAILED) {
695 requeue_io(inode);
696 continue;
697 }
698 ret = writeback_sb_inodes(sb, wb, wbc);
699
700 if (state == SB_PINNED)
701 unpin_sb_for_writeback(sb);
702 if (ret)
703 break;
704 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200705 spin_unlock(&inode_lock);
706 /* Leave any unwritten inodes on b_io */
707}
708
Jens Axboe03ba3782009-09-09 09:08:54 +0200709void writeback_inodes_wbc(struct writeback_control *wbc)
710{
711 struct backing_dev_info *bdi = wbc->bdi;
712
713 writeback_inodes_wb(&bdi->wb, wbc);
714}
715
716/*
717 * The maximum number of pages to writeout in a single bdi flush/kupdate
718 * operation. We do this so we don't hold I_SYNC against an inode for
719 * enormous amounts of time, which would block a userspace task which has
720 * been forced to throttle against that inode. Also, the code reevaluates
721 * the dirty each time it has written this many pages.
722 */
723#define MAX_WRITEBACK_PAGES 1024
724
725static inline bool over_bground_thresh(void)
726{
727 unsigned long background_thresh, dirty_thresh;
728
729 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
730
731 return (global_page_state(NR_FILE_DIRTY) +
732 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
733}
734
735/*
736 * Explicit flushing or periodic writeback of "old" data.
737 *
738 * Define "old": the first time one of an inode's pages is dirtied, we mark the
739 * dirtying-time in the inode's address_space. So this periodic writeback code
740 * just walks the superblock inode list, writing back any inodes which are
741 * older than a specific point in time.
742 *
743 * Try to run once per dirty_writeback_interval. But if a writeback event
744 * takes longer than a dirty_writeback_interval interval, then leave a
745 * one-second gap.
746 *
747 * older_than_this takes precedence over nr_to_write. So we'll only write back
748 * all dirty pages if they are all attached to "old" mappings.
749 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200750static long wb_writeback(struct bdi_writeback *wb,
751 struct wb_writeback_args *args)
Jens Axboe03ba3782009-09-09 09:08:54 +0200752{
753 struct writeback_control wbc = {
754 .bdi = wb->bdi,
Jens Axboec4a77a62009-09-16 15:18:25 +0200755 .sb = args->sb,
756 .sync_mode = args->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200757 .older_than_this = NULL,
Jens Axboec4a77a62009-09-16 15:18:25 +0200758 .for_kupdate = args->for_kupdate,
Wu Fengguangb17621f2009-12-03 13:54:25 +0100759 .for_background = args->for_background,
Jens Axboec4a77a62009-09-16 15:18:25 +0200760 .range_cyclic = args->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200761 };
762 unsigned long oldest_jif;
763 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200764 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200765
766 if (wbc.for_kupdate) {
767 wbc.older_than_this = &oldest_jif;
768 oldest_jif = jiffies -
769 msecs_to_jiffies(dirty_expire_interval * 10);
770 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200771 if (!wbc.range_cyclic) {
772 wbc.range_start = 0;
773 wbc.range_end = LLONG_MAX;
774 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200775
776 for (;;) {
777 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800778 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200779 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800780 if (args->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200781 break;
782
783 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800784 * For background writeout, stop when we are below the
785 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200786 */
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800787 if (args->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200788 break;
789
790 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200791 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
792 wbc.pages_skipped = 0;
793 writeback_inodes_wb(wb, &wbc);
Jens Axboec4a77a62009-09-16 15:18:25 +0200794 args->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200795 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
796
797 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200798 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200799 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200800 if (wbc.nr_to_write <= 0)
801 continue;
802 /*
803 * Didn't write everything and we don't have more IO, bail
804 */
805 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200806 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200807 /*
808 * Did we write something? Try for more
809 */
810 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
811 continue;
812 /*
813 * Nothing written. Wait for some inode to
814 * become available for writeback. Otherwise
815 * we'll just busyloop.
816 */
817 spin_lock(&inode_lock);
818 if (!list_empty(&wb->b_more_io)) {
819 inode = list_entry(wb->b_more_io.prev,
820 struct inode, i_list);
821 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200822 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200823 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200824 }
825
826 return wrote;
827}
828
829/*
830 * Return the next bdi_work struct that hasn't been processed by this
Jens Axboe8010c3b2009-09-15 20:04:57 +0200831 * wb thread yet. ->seen is initially set for each thread that exists
832 * for this device, when a thread first notices a piece of work it
833 * clears its bit. Depending on writeback type, the thread will notify
834 * completion on either receiving the work (WB_SYNC_NONE) or after
835 * it is done (WB_SYNC_ALL).
Jens Axboe03ba3782009-09-09 09:08:54 +0200836 */
837static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
838 struct bdi_writeback *wb)
839{
840 struct bdi_work *work, *ret = NULL;
841
842 rcu_read_lock();
843
844 list_for_each_entry_rcu(work, &bdi->work_list, list) {
Nick Piggin77fad5e2009-09-15 21:34:12 +0200845 if (!test_bit(wb->nr, &work->seen))
Jens Axboe03ba3782009-09-09 09:08:54 +0200846 continue;
Nick Piggin77fad5e2009-09-15 21:34:12 +0200847 clear_bit(wb->nr, &work->seen);
Jens Axboe03ba3782009-09-09 09:08:54 +0200848
849 ret = work;
850 break;
851 }
852
853 rcu_read_unlock();
854 return ret;
855}
856
857static long wb_check_old_data_flush(struct bdi_writeback *wb)
858{
859 unsigned long expired;
860 long nr_pages;
861
Jens Axboe69b62d02010-05-17 12:51:03 +0200862 /*
863 * When set to zero, disable periodic writeback
864 */
865 if (!dirty_writeback_interval)
866 return 0;
867
Jens Axboe03ba3782009-09-09 09:08:54 +0200868 expired = wb->last_old_flush +
869 msecs_to_jiffies(dirty_writeback_interval * 10);
870 if (time_before(jiffies, expired))
871 return 0;
872
873 wb->last_old_flush = jiffies;
874 nr_pages = global_page_state(NR_FILE_DIRTY) +
875 global_page_state(NR_UNSTABLE_NFS) +
876 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
877
Jens Axboec4a77a62009-09-16 15:18:25 +0200878 if (nr_pages) {
879 struct wb_writeback_args args = {
880 .nr_pages = nr_pages,
881 .sync_mode = WB_SYNC_NONE,
882 .for_kupdate = 1,
883 .range_cyclic = 1,
884 };
885
886 return wb_writeback(wb, &args);
887 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200888
889 return 0;
890}
891
892/*
893 * Retrieve work items and do the writeback they describe
894 */
895long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
896{
897 struct backing_dev_info *bdi = wb->bdi;
898 struct bdi_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200899 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200900
901 while ((work = get_next_work_item(bdi, wb)) != NULL) {
Jens Axboec4a77a62009-09-16 15:18:25 +0200902 struct wb_writeback_args args = work->args;
Jens Axboe03ba3782009-09-09 09:08:54 +0200903
904 /*
905 * Override sync mode, in case we must wait for completion
906 */
907 if (force_wait)
Jens Axboec4a77a62009-09-16 15:18:25 +0200908 work->args.sync_mode = args.sync_mode = WB_SYNC_ALL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200909
910 /*
911 * If this isn't a data integrity operation, just notify
912 * that we have seen this work and we are now starting it.
913 */
Jens Axboef17625b2010-06-01 11:05:22 +0200914 if (args.sync_mode == WB_SYNC_NONE)
Jens Axboe03ba3782009-09-09 09:08:54 +0200915 wb_clear_pending(wb, work);
916
Jens Axboec4a77a62009-09-16 15:18:25 +0200917 wrote += wb_writeback(wb, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +0200918
919 /*
920 * This is a data integrity writeback, so only do the
921 * notification when we have completed the work.
922 */
Jens Axboef17625b2010-06-01 11:05:22 +0200923 if (args.sync_mode == WB_SYNC_ALL)
Jens Axboe03ba3782009-09-09 09:08:54 +0200924 wb_clear_pending(wb, work);
925 }
926
927 /*
928 * Check for periodic writeback, kupdated() style
929 */
930 wrote += wb_check_old_data_flush(wb);
931
932 return wrote;
933}
934
935/*
936 * Handle writeback of dirty data for the device backed by this bdi. Also
937 * wakes up periodically and does kupdated style flushing.
938 */
939int bdi_writeback_task(struct bdi_writeback *wb)
940{
941 unsigned long last_active = jiffies;
942 unsigned long wait_jiffies = -1UL;
943 long pages_written;
944
945 while (!kthread_should_stop()) {
946 pages_written = wb_do_writeback(wb, 0);
947
948 if (pages_written)
949 last_active = jiffies;
950 else if (wait_jiffies != -1UL) {
951 unsigned long max_idle;
952
953 /*
954 * Longest period of inactivity that we tolerate. If we
955 * see dirty data again later, the task will get
956 * recreated automatically.
957 */
958 max_idle = max(5UL * 60 * HZ, wait_jiffies);
959 if (time_after(jiffies, max_idle + last_active))
960 break;
961 }
962
Jens Axboe69b62d02010-05-17 12:51:03 +0200963 if (dirty_writeback_interval) {
964 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
965 schedule_timeout_interruptible(wait_jiffies);
Jens Axboef9eadbb2010-05-18 14:31:45 +0200966 } else {
967 set_current_state(TASK_INTERRUPTIBLE);
968 if (list_empty_careful(&wb->bdi->work_list) &&
969 !kthread_should_stop())
970 schedule();
971 __set_current_state(TASK_RUNNING);
972 }
Jens Axboe69b62d02010-05-17 12:51:03 +0200973
Jens Axboe03ba3782009-09-09 09:08:54 +0200974 try_to_freeze();
975 }
976
977 return 0;
978}
979
980/*
Jens Axboeb6e51312009-09-16 15:13:54 +0200981 * Schedule writeback for all backing devices. This does WB_SYNC_NONE
982 * writeback, for integrity writeback see bdi_sync_writeback().
Jens Axboe03ba3782009-09-09 09:08:54 +0200983 */
Jens Axboeb6e51312009-09-16 15:13:54 +0200984static void bdi_writeback_all(struct super_block *sb, long nr_pages)
Jens Axboe03ba3782009-09-09 09:08:54 +0200985{
Jens Axboeb6e51312009-09-16 15:13:54 +0200986 struct wb_writeback_args args = {
987 .sb = sb,
988 .nr_pages = nr_pages,
989 .sync_mode = WB_SYNC_NONE,
990 };
Jens Axboe03ba3782009-09-09 09:08:54 +0200991 struct backing_dev_info *bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200992
Jens Axboecfc4ba52009-09-14 13:12:40 +0200993 rcu_read_lock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200994
Jens Axboecfc4ba52009-09-14 13:12:40 +0200995 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200996 if (!bdi_has_dirty_io(bdi))
997 continue;
998
Jens Axboef17625b2010-06-01 11:05:22 +0200999 bdi_alloc_queue_work(bdi, &args);
Jens Axboe03ba3782009-09-09 09:08:54 +02001000 }
1001
Jens Axboecfc4ba52009-09-14 13:12:40 +02001002 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001003}
1004
1005/*
1006 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1007 * the whole world.
1008 */
1009void wakeup_flusher_threads(long nr_pages)
1010{
Jens Axboe03ba3782009-09-09 09:08:54 +02001011 if (nr_pages == 0)
1012 nr_pages = global_page_state(NR_FILE_DIRTY) +
1013 global_page_state(NR_UNSTABLE_NFS);
Jens Axboeb6e51312009-09-16 15:13:54 +02001014 bdi_writeback_all(NULL, nr_pages);
Jens Axboe03ba3782009-09-09 09:08:54 +02001015}
1016
1017static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1018{
1019 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1020 struct dentry *dentry;
1021 const char *name = "?";
1022
1023 dentry = d_find_alias(inode);
1024 if (dentry) {
1025 spin_lock(&dentry->d_lock);
1026 name = (const char *) dentry->d_name.name;
1027 }
1028 printk(KERN_DEBUG
1029 "%s(%d): dirtied inode %lu (%s) on %s\n",
1030 current->comm, task_pid_nr(current), inode->i_ino,
1031 name, inode->i_sb->s_id);
1032 if (dentry) {
1033 spin_unlock(&dentry->d_lock);
1034 dput(dentry);
1035 }
1036 }
1037}
1038
1039/**
1040 * __mark_inode_dirty - internal function
1041 * @inode: inode to mark
1042 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1043 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1044 * mark_inode_dirty_sync.
1045 *
1046 * Put the inode on the super block's dirty list.
1047 *
1048 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1049 * dirty list only if it is hashed or if it refers to a blockdev.
1050 * If it was not hashed, it will never be added to the dirty list
1051 * even if it is later hashed, as it will have been marked dirty already.
1052 *
1053 * In short, make sure you hash any inodes _before_ you start marking
1054 * them dirty.
1055 *
1056 * This function *must* be atomic for the I_DIRTY_PAGES case -
1057 * set_page_dirty() is called under spinlock in several places.
1058 *
1059 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1060 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1061 * the kernel-internal blockdev inode represents the dirtying time of the
1062 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1063 * page->mapping->host, so the page-dirtying time is recorded in the internal
1064 * blockdev inode.
1065 */
1066void __mark_inode_dirty(struct inode *inode, int flags)
1067{
1068 struct super_block *sb = inode->i_sb;
1069
1070 /*
1071 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1072 * dirty the inode itself
1073 */
1074 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1075 if (sb->s_op->dirty_inode)
1076 sb->s_op->dirty_inode(inode);
1077 }
1078
1079 /*
1080 * make sure that changes are seen by all cpus before we test i_state
1081 * -- mikulas
1082 */
1083 smp_mb();
1084
1085 /* avoid the locking if we can */
1086 if ((inode->i_state & flags) == flags)
1087 return;
1088
1089 if (unlikely(block_dump))
1090 block_dump___mark_inode_dirty(inode);
1091
1092 spin_lock(&inode_lock);
1093 if ((inode->i_state & flags) != flags) {
1094 const int was_dirty = inode->i_state & I_DIRTY;
1095
1096 inode->i_state |= flags;
1097
1098 /*
1099 * If the inode is being synced, just update its dirty state.
1100 * The unlocker will place the inode on the appropriate
1101 * superblock list, based upon its state.
1102 */
1103 if (inode->i_state & I_SYNC)
1104 goto out;
1105
1106 /*
1107 * Only add valid (hashed) inodes to the superblock's
1108 * dirty list. Add blockdev inodes as well.
1109 */
1110 if (!S_ISBLK(inode->i_mode)) {
1111 if (hlist_unhashed(&inode->i_hash))
1112 goto out;
1113 }
1114 if (inode->i_state & (I_FREEING|I_CLEAR))
1115 goto out;
1116
1117 /*
1118 * If the inode was already on b_dirty/b_io/b_more_io, don't
1119 * reposition it (that would break b_dirty time-ordering).
1120 */
1121 if (!was_dirty) {
1122 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001123 struct backing_dev_info *bdi = wb->bdi;
1124
1125 if (bdi_cap_writeback_dirty(bdi) &&
1126 !test_bit(BDI_registered, &bdi->state)) {
1127 WARN_ON(1);
1128 printk(KERN_ERR "bdi-%s not registered\n",
1129 bdi->name);
1130 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001131
1132 inode->dirtied_when = jiffies;
1133 list_move(&inode->i_list, &wb->b_dirty);
1134 }
1135 }
1136out:
1137 spin_unlock(&inode_lock);
1138}
1139EXPORT_SYMBOL(__mark_inode_dirty);
1140
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001141/*
1142 * Write out a superblock's list of dirty inodes. A wait will be performed
1143 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1144 *
1145 * If older_than_this is non-NULL, then only write out inodes which
1146 * had their first dirtying at a time earlier than *older_than_this.
1147 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001148 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1149 * This function assumes that the blockdev superblock's inodes are backed by
1150 * a variety of queues, so all inodes are searched. For other superblocks,
1151 * assume that all inodes are backed by the same queue.
1152 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001153 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1154 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1155 * on the writer throttling path, and we get decent balancing between many
1156 * throttled threads: we don't want them all piling up on inode_sync_wait.
1157 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001158static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001159{
Jens Axboe03ba3782009-09-09 09:08:54 +02001160 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001161
Jens Axboe03ba3782009-09-09 09:08:54 +02001162 /*
1163 * We need to be protected against the filesystem going from
1164 * r/o to r/w or vice versa.
1165 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001166 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001167
Jens Axboe03ba3782009-09-09 09:08:54 +02001168 spin_lock(&inode_lock);
1169
1170 /*
1171 * Data integrity sync. Must wait for all pages under writeback,
1172 * because there may have been pages dirtied before our sync
1173 * call, but which had writeout started before we write it out.
1174 * In which case, the inode may not be on the dirty list, but
1175 * we still have to wait for that writeout.
1176 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001177 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001178 struct address_space *mapping;
1179
1180 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1181 continue;
1182 mapping = inode->i_mapping;
1183 if (mapping->nrpages == 0)
1184 continue;
1185 __iget(inode);
1186 spin_unlock(&inode_lock);
1187 /*
1188 * We hold a reference to 'inode' so it couldn't have
1189 * been removed from s_inodes list while we dropped the
1190 * inode_lock. We cannot iput the inode now as we can
1191 * be holding the last reference and we cannot iput it
1192 * under inode_lock. So we keep the reference and iput
1193 * it later.
1194 */
1195 iput(old_inode);
1196 old_inode = inode;
1197
1198 filemap_fdatawait(mapping);
1199
1200 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001201
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001202 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001203 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001204 spin_unlock(&inode_lock);
1205 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206}
1207
Jens Axboed8a85592009-09-02 12:34:32 +02001208/**
1209 * writeback_inodes_sb - writeback dirty inodes from given super_block
1210 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 *
Jens Axboed8a85592009-09-02 12:34:32 +02001212 * Start writeback on some inodes on this super_block. No guarantees are made
1213 * on how many (if any) will be written, and this function does not wait
1214 * for IO completion of submitted IO. The number of pages submitted is
1215 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001217void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218{
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001219 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1220 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
1221 long nr_to_write;
1222
1223 nr_to_write = nr_dirty + nr_unstable +
1224 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1225
1226 bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227}
Jens Axboed8a85592009-09-02 12:34:32 +02001228EXPORT_SYMBOL(writeback_inodes_sb);
1229
1230/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001231 * writeback_inodes_sb_if_idle - start writeback if none underway
1232 * @sb: the superblock
1233 *
1234 * Invoke writeback_inodes_sb if no writeback is currently underway.
1235 * Returns 1 if writeback was started, 0 if not.
1236 */
1237int writeback_inodes_sb_if_idle(struct super_block *sb)
1238{
1239 if (!writeback_in_progress(sb->s_bdi)) {
1240 writeback_inodes_sb(sb);
1241 return 1;
1242 } else
1243 return 0;
1244}
1245EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1246
1247/**
Jens Axboed8a85592009-09-02 12:34:32 +02001248 * sync_inodes_sb - sync sb inode pages
1249 * @sb: the superblock
1250 *
1251 * This function writes and waits on any dirty inode belonging to this
1252 * super_block. The number of pages synced is returned.
1253 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001254void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001255{
Jens Axboeb6e51312009-09-16 15:13:54 +02001256 bdi_sync_writeback(sb->s_bdi, sb);
1257 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001258}
1259EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001262 * write_inode_now - write an inode to disk
1263 * @inode: inode to write to disk
1264 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001266 * This function commits an inode to disk immediately if it is dirty. This is
1267 * primarily needed by knfsd.
1268 *
1269 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271int write_inode_now(struct inode *inode, int sync)
1272{
1273 int ret;
1274 struct writeback_control wbc = {
1275 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001276 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001277 .range_start = 0,
1278 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 };
1280
1281 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001282 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283
1284 might_sleep();
1285 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001286 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 spin_unlock(&inode_lock);
1288 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001289 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 return ret;
1291}
1292EXPORT_SYMBOL(write_inode_now);
1293
1294/**
1295 * sync_inode - write an inode and its pages to disk.
1296 * @inode: the inode to sync
1297 * @wbc: controls the writeback mode
1298 *
1299 * sync_inode() will write an inode and its pages to disk. It will also
1300 * correctly update the inode on its superblock's dirty inode lists and will
1301 * update inode->i_state.
1302 *
1303 * The caller must have a ref on the inode.
1304 */
1305int sync_inode(struct inode *inode, struct writeback_control *wbc)
1306{
1307 int ret;
1308
1309 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001310 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 spin_unlock(&inode_lock);
1312 return ret;
1313}
1314EXPORT_SYMBOL(sync_inode);