blob: 3d06ccc953aafaa8f590528015953e2e53a2c0c9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020017#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020023#include <linux/kthread.h>
24#include <linux/freezer.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
28#include <linux/buffer_head.h>
Dave Chinner455b2862010-07-07 13:24:06 +100029#include <linux/tracepoint.h>
David Howells07f3f05c2006-09-30 20:52:18 +020030#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
Jens Axboed0bceac2009-05-18 08:20:32 +020032/*
Jens Axboec4a77a62009-09-16 15:18:25 +020033 * Passed into wb_writeback(), essentially a subset of writeback_control
34 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020035struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020036 long nr_pages;
37 struct super_block *sb;
38 enum writeback_sync_modes sync_mode;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050039 unsigned int for_kupdate:1;
40 unsigned int range_cyclic:1;
41 unsigned int for_background:1;
Jens Axboec4a77a62009-09-16 15:18:25 +020042
Jens Axboe8010c3b2009-09-15 20:04:57 +020043 struct list_head list; /* pending work list */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020044 struct completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020045};
46
Dave Chinner455b2862010-07-07 13:24:06 +100047/*
48 * Include the creation of the trace points after defining the
49 * wb_writeback_work structure so that the definition remains local to this
50 * file.
51 */
52#define CREATE_TRACE_POINTS
53#include <trace/events/writeback.h>
54
Dave Chinner455b2862010-07-07 13:24:06 +100055/*
56 * We don't actually have pdflush, but this one is exported though /proc...
57 */
58int nr_pdflush_threads;
59
Adrian Bunkf11b00f2008-04-29 00:58:56 -070060/**
61 * writeback_in_progress - determine whether there is writeback in progress
62 * @bdi: the device's backing_dev_info structure.
63 *
Jens Axboe03ba3782009-09-09 09:08:54 +020064 * Determine whether there is writeback waiting to be handled against a
65 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070066 */
67int writeback_in_progress(struct backing_dev_info *bdi)
68{
Jan Kara81d73a32010-08-11 14:17:44 -070069 return test_bit(BDI_writeback_running, &bdi->state);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070070}
71
Jan Kara692ebd12010-09-21 11:51:01 +020072static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
73{
74 struct super_block *sb = inode->i_sb;
Jan Kara692ebd12010-09-21 11:51:01 +020075
Christoph Hellwigaaead252010-10-04 14:25:33 +020076 if (strcmp(sb->s_type->name, "bdev") == 0)
77 return inode->i_mapping->backing_dev_info;
78
79 return sb->s_bdi;
Jan Kara692ebd12010-09-21 11:51:01 +020080}
81
Nick Piggin7ccf19a2010-10-21 11:49:30 +110082static inline struct inode *wb_inode(struct list_head *head)
83{
84 return list_entry(head, struct inode, i_wb_list);
85}
86
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020087static void bdi_queue_work(struct backing_dev_info *bdi,
88 struct wb_writeback_work *work)
Nick Piggin4195f732009-05-28 09:01:15 +020089{
Dave Chinner455b2862010-07-07 13:24:06 +100090 trace_writeback_queue(bdi, work);
91
Artem Bityutskiy64677162010-07-25 14:29:22 +030092 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020093 list_add_tail(&work->list, &bdi->work_list);
Artem Bityutskiyfff5b852010-07-25 14:29:20 +030094 if (bdi->wb.task) {
95 wake_up_process(bdi->wb.task);
96 } else {
97 /*
98 * The bdi thread isn't there, wake up the forker thread which
99 * will create and run it.
100 */
Dave Chinner455b2862010-07-07 13:24:06 +1000101 trace_writeback_nothread(bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200102 wake_up_process(default_backing_dev_info.wb.task);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 }
Artem Bityutskiy64677162010-07-25 14:29:22 +0300104 spin_unlock_bh(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105}
106
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200107static void
108__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
109 bool range_cyclic, bool for_background)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200111 struct wb_writeback_work *work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200112
Jens Axboebcddc3f2009-09-13 20:07:36 +0200113 /*
114 * This is WB_SYNC_NONE writeback, so if allocation fails just
115 * wakeup the thread for old dirty data writeback
116 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200117 work = kzalloc(sizeof(*work), GFP_ATOMIC);
118 if (!work) {
Dave Chinner455b2862010-07-07 13:24:06 +1000119 if (bdi->wb.task) {
120 trace_writeback_nowork(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200121 wake_up_process(bdi->wb.task);
Dave Chinner455b2862010-07-07 13:24:06 +1000122 }
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200123 return;
Jens Axboebcddc3f2009-09-13 20:07:36 +0200124 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200125
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200126 work->sync_mode = WB_SYNC_NONE;
127 work->nr_pages = nr_pages;
128 work->range_cyclic = range_cyclic;
129 work->for_background = for_background;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200130
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200131 bdi_queue_work(bdi, work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200132}
133
134/**
135 * bdi_start_writeback - start writeback
136 * @bdi: the backing device to write from
137 * @nr_pages: the number of pages to write
138 *
139 * Description:
140 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
141 * started when this function returns, we make no guarentees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200142 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200143 *
144 */
Christoph Hellwigc5444192010-06-08 18:15:15 +0200145void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200146{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200147 __bdi_start_writeback(bdi, nr_pages, true, false);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200148}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800149
Christoph Hellwigc5444192010-06-08 18:15:15 +0200150/**
151 * bdi_start_background_writeback - start background writeback
152 * @bdi: the backing device to write from
153 *
154 * Description:
155 * This does WB_SYNC_NONE background writeback. The IO is only
156 * started when this function returns, we make no guarentees on
157 * completion. Caller need not hold sb s_umount semaphore.
158 */
159void bdi_start_background_writeback(struct backing_dev_info *bdi)
160{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200161 __bdi_start_writeback(bdi, LONG_MAX, true, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162}
163
164/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700165 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
166 * furthest end of its superblock's dirty-inode list.
167 *
168 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200169 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700170 * the case then the inode must have been redirtied while it was being written
171 * out and we don't reset its dirtied_when.
172 */
173static void redirty_tail(struct inode *inode)
174{
Jens Axboe03ba3782009-09-09 09:08:54 +0200175 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700176
Jens Axboe03ba3782009-09-09 09:08:54 +0200177 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200178 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700179
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100180 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200181 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700182 inode->dirtied_when = jiffies;
183 }
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100184 list_move(&inode->i_wb_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700185}
186
187/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200188 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700189 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700190static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700191{
Jens Axboe03ba3782009-09-09 09:08:54 +0200192 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
193
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100194 list_move(&inode->i_wb_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700195}
196
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700197static void inode_sync_complete(struct inode *inode)
198{
199 /*
200 * Prevent speculative execution through spin_unlock(&inode_lock);
201 */
202 smp_mb();
203 wake_up_bit(&inode->i_state, __I_SYNC);
204}
205
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700206static bool inode_dirtied_after(struct inode *inode, unsigned long t)
207{
208 bool ret = time_after(inode->dirtied_when, t);
209#ifndef CONFIG_64BIT
210 /*
211 * For inodes being constantly redirtied, dirtied_when can get stuck.
212 * It _appears_ to be in the future, but is actually in distant past.
213 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200214 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700215 */
216 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
217#endif
218 return ret;
219}
220
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700221/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700222 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
223 */
224static void move_expired_inodes(struct list_head *delaying_queue,
225 struct list_head *dispatch_queue,
226 unsigned long *older_than_this)
227{
Shaohua Li5c034492009-09-24 14:42:33 +0200228 LIST_HEAD(tmp);
229 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200230 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200231 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200232 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200233
Fengguang Wu2c136572007-10-16 23:30:39 -0700234 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100235 inode = wb_inode(delaying_queue->prev);
Fengguang Wu2c136572007-10-16 23:30:39 -0700236 if (older_than_this &&
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700237 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700238 break;
Jens Axboecf137302009-09-24 15:12:57 +0200239 if (sb && sb != inode->i_sb)
240 do_sb_sort = 1;
241 sb = inode->i_sb;
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100242 list_move(&inode->i_wb_list, &tmp);
Shaohua Li5c034492009-09-24 14:42:33 +0200243 }
244
Jens Axboecf137302009-09-24 15:12:57 +0200245 /* just one sb in list, splice to dispatch_queue and we're done */
246 if (!do_sb_sort) {
247 list_splice(&tmp, dispatch_queue);
248 return;
249 }
250
Shaohua Li5c034492009-09-24 14:42:33 +0200251 /* Move inodes from one superblock together */
252 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100253 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200254 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100255 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +0200256 if (inode->i_sb == sb)
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100257 list_move(&inode->i_wb_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +0200258 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700259 }
260}
261
262/*
263 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700264 * Before
265 * newly dirtied b_dirty b_io b_more_io
266 * =============> gf edc BA
267 * After
268 * newly dirtied b_dirty b_io b_more_io
269 * =============> g fBAedc
270 * |
271 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -0700272 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200273static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700274{
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700275 list_splice_init(&wb->b_more_io, &wb->b_io);
Jens Axboe03ba3782009-09-09 09:08:54 +0200276 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200277}
278
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100279static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200280{
Jens Axboe03ba3782009-09-09 09:08:54 +0200281 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100282 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200283 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700284}
285
286/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200287 * Wait for writeback on an inode to complete.
288 */
289static void inode_wait_for_writeback(struct inode *inode)
290{
291 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
292 wait_queue_head_t *wqh;
293
294 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700295 while (inode->i_state & I_SYNC) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200296 spin_unlock(&inode_lock);
297 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
298 spin_lock(&inode_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -0700299 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200300}
301
302/*
303 * Write out an inode's dirty pages. Called under inode_lock. Either the
304 * caller has ref on the inode (either via __iget or via syscall against an fd)
305 * or the inode has I_WILL_FREE set (via generic_forget_inode)
306 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 * If `wait' is set, wait on the writeout.
308 *
309 * The whole writeout design is quite complex and fragile. We want to avoid
310 * starvation of particular inodes when others are being redirtied, prevent
311 * livelocks, etc.
312 *
313 * Called under inode_lock.
314 */
315static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200316writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200319 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 int ret;
321
Christoph Hellwig01c03192009-06-08 13:35:40 +0200322 if (!atomic_read(&inode->i_count))
323 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
324 else
325 WARN_ON(inode->i_state & I_WILL_FREE);
326
327 if (inode->i_state & I_SYNC) {
328 /*
329 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200330 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200331 * writeback can proceed with the other inodes on s_io.
332 *
333 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200334 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200335 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100336 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200337 requeue_io(inode);
338 return 0;
339 }
340
341 /*
342 * It's a data-integrity sync. We must wait.
343 */
344 inode_wait_for_writeback(inode);
345 }
346
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700347 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400349 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700350 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400351 inode->i_state &= ~I_DIRTY_PAGES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 spin_unlock(&inode_lock);
353
354 ret = do_writepages(mapping, wbc);
355
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100356 /*
357 * Make sure to wait on the data before writing out the metadata.
358 * This is important for filesystems that modify metadata on data
359 * I/O completion.
360 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100361 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100362 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 if (ret == 0)
364 ret = err;
365 }
366
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400367 /*
368 * Some filesystems may redirty the inode during the writeback
369 * due to delalloc, clear dirty metadata flags right before
370 * write_inode()
371 */
372 spin_lock(&inode_lock);
373 dirty = inode->i_state & I_DIRTY;
374 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
375 spin_unlock(&inode_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100376 /* Don't write the inode if only I_DIRTY_PAGES was set */
377 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100378 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 if (ret == 0)
380 ret = err;
381 }
382
383 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700384 inode->i_state &= ~I_SYNC;
Al Viroa4ffdde2010-06-02 17:38:30 -0400385 if (!(inode->i_state & I_FREEING)) {
Wu Fengguang23539af2010-08-11 14:17:41 -0700386 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 /*
388 * We didn't write back all the pages. nfs_writepages()
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700389 * sometimes bales out without doing anything.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700390 */
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700391 inode->i_state |= I_DIRTY_PAGES;
392 if (wbc->nr_to_write <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700394 * slice used up: queue for next turn
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 */
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700396 requeue_io(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 } else {
398 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700399 * Writeback blocked by something other than
400 * congestion. Delay the inode for some time to
401 * avoid spinning on the CPU (100% iowait)
402 * retrying writeback of the dirty page/inode
403 * that cannot be performed immediately.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 */
Andrew Morton1b43ef92007-10-16 23:30:35 -0700405 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 }
Wu Fengguang23539af2010-08-11 14:17:41 -0700407 } else if (inode->i_state & I_DIRTY) {
408 /*
409 * Filesystems can dirty the inode during writeback
410 * operations, such as delayed allocation during
411 * submission or metadata updates after data IO
412 * completion.
413 */
414 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 } else {
416 /*
Nick Piggin9e38d862010-10-23 06:55:17 -0400417 * The inode is clean. At this point we either have
418 * a reference to the inode or it's on it's way out.
419 * No need to add it back to the LRU.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 */
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100421 list_del_init(&inode->i_wb_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 }
423 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700424 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 return ret;
426}
427
Jens Axboe03ba3782009-09-09 09:08:54 +0200428/*
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200429 * For background writeback the caller does not have the sb pinned
Jens Axboe03ba3782009-09-09 09:08:54 +0200430 * before calling writeback. So make sure that we do pin it, so it doesn't
431 * go away while we are writing inodes from it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200432 */
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200433static bool pin_sb_for_writeback(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434{
Jens Axboe03ba3782009-09-09 09:08:54 +0200435 spin_lock(&sb_lock);
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200436 if (list_empty(&sb->s_instances)) {
437 spin_unlock(&sb_lock);
438 return false;
439 }
440
Jens Axboe03ba3782009-09-09 09:08:54 +0200441 sb->s_count++;
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200442 spin_unlock(&sb_lock);
443
Jens Axboe03ba3782009-09-09 09:08:54 +0200444 if (down_read_trylock(&sb->s_umount)) {
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200445 if (sb->s_root)
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200446 return true;
Jens Axboe03ba3782009-09-09 09:08:54 +0200447 up_read(&sb->s_umount);
448 }
Christoph Hellwig29cb48592010-06-09 15:31:01 +0200449
450 put_super(sb);
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200451 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200452}
453
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800454/*
455 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200456 *
457 * If @only_this_sb is true, then find and write all such
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800458 * inodes. Otherwise write only ones which go sequentially
459 * in reverse order.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200460 *
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800461 * Return 1, if the caller writeback routine should be
462 * interrupted. Otherwise return 0.
463 */
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200464static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
465 struct writeback_control *wbc, bool only_this_sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200466{
Jens Axboe03ba3782009-09-09 09:08:54 +0200467 while (!list_empty(&wb->b_io)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 long pages_skipped;
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100469 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200470
471 if (inode->i_sb != sb) {
472 if (only_this_sb) {
473 /*
474 * We only want to write back data for this
475 * superblock, move all inodes not belonging
476 * to it back onto the dirty list.
477 */
478 redirty_tail(inode);
479 continue;
480 }
481
482 /*
483 * The inode belongs to a different superblock.
484 * Bounce back to the caller to unpin this and
485 * pin the next superblock.
486 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800487 return 0;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200488 }
489
Christoph Hellwig9843b762010-10-24 19:40:46 +0200490 /*
491 * Don't bother with new inodes or inodes beeing freed, first
492 * kind does not need peridic writeout yet, and for the latter
493 * kind writeout is handled by the freer.
494 */
495 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700496 requeue_io(inode);
497 continue;
498 }
Christoph Hellwig9843b762010-10-24 19:40:46 +0200499
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700500 /*
501 * Was this inode dirtied after sync_sb_inodes was called?
502 * This keeps sync from extra jobs and livelock.
503 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800504 if (inode_dirtied_after(inode, wbc->wb_start))
505 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 __iget(inode);
508 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200509 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 if (wbc->pages_skipped != pages_skipped) {
511 /*
512 * writeback is not making progress due to locked
513 * buffers. Skip this inode for now.
514 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700515 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 }
517 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800519 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800521 if (wbc->nr_to_write <= 0) {
522 wbc->more_io = 1;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800523 return 1;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800524 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200525 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800526 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800528 /* b_io is empty */
529 return 1;
530}
Nick Piggin38f21972009-01-06 14:40:25 -0800531
Christoph Hellwig9c3a8ee2010-06-10 12:07:27 +0200532void writeback_inodes_wb(struct bdi_writeback *wb,
533 struct writeback_control *wbc)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800534{
535 int ret = 0;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200536
Jan Kara7624ee72010-08-09 17:20:03 -0700537 if (!wbc->wb_start)
538 wbc->wb_start = jiffies; /* livelock avoidance */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800539 spin_lock(&inode_lock);
540 if (!wbc->for_kupdate || list_empty(&wb->b_io))
541 queue_io(wb, wbc->older_than_this);
542
543 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100544 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800545 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800546
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200547 if (!pin_sb_for_writeback(sb)) {
548 requeue_io(inode);
549 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800550 }
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200551 ret = writeback_sb_inodes(sb, wb, wbc, false);
552 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800553
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800554 if (ret)
555 break;
556 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200557 spin_unlock(&inode_lock);
558 /* Leave any unwritten inodes on b_io */
559}
560
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200561static void __writeback_inodes_sb(struct super_block *sb,
562 struct bdi_writeback *wb, struct writeback_control *wbc)
563{
564 WARN_ON(!rwsem_is_locked(&sb->s_umount));
565
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200566 spin_lock(&inode_lock);
567 if (!wbc->for_kupdate || list_empty(&wb->b_io))
568 queue_io(wb, wbc->older_than_this);
569 writeback_sb_inodes(sb, wb, wbc, true);
570 spin_unlock(&inode_lock);
571}
572
Jens Axboe03ba3782009-09-09 09:08:54 +0200573/*
574 * The maximum number of pages to writeout in a single bdi flush/kupdate
575 * operation. We do this so we don't hold I_SYNC against an inode for
576 * enormous amounts of time, which would block a userspace task which has
577 * been forced to throttle against that inode. Also, the code reevaluates
578 * the dirty each time it has written this many pages.
579 */
580#define MAX_WRITEBACK_PAGES 1024
581
582static inline bool over_bground_thresh(void)
583{
584 unsigned long background_thresh, dirty_thresh;
585
Wu Fengguang16c40422010-08-11 14:17:39 -0700586 global_dirty_limits(&background_thresh, &dirty_thresh);
Jens Axboe03ba3782009-09-09 09:08:54 +0200587
588 return (global_page_state(NR_FILE_DIRTY) +
Wu Fengguang4cbec4c2010-10-26 14:21:45 -0700589 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
Jens Axboe03ba3782009-09-09 09:08:54 +0200590}
591
592/*
593 * Explicit flushing or periodic writeback of "old" data.
594 *
595 * Define "old": the first time one of an inode's pages is dirtied, we mark the
596 * dirtying-time in the inode's address_space. So this periodic writeback code
597 * just walks the superblock inode list, writing back any inodes which are
598 * older than a specific point in time.
599 *
600 * Try to run once per dirty_writeback_interval. But if a writeback event
601 * takes longer than a dirty_writeback_interval interval, then leave a
602 * one-second gap.
603 *
604 * older_than_this takes precedence over nr_to_write. So we'll only write back
605 * all dirty pages if they are all attached to "old" mappings.
606 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200607static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200608 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200609{
610 struct writeback_control wbc = {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200611 .sync_mode = work->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200612 .older_than_this = NULL,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200613 .for_kupdate = work->for_kupdate,
614 .for_background = work->for_background,
615 .range_cyclic = work->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200616 };
617 unsigned long oldest_jif;
618 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200619 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200620
621 if (wbc.for_kupdate) {
622 wbc.older_than_this = &oldest_jif;
623 oldest_jif = jiffies -
624 msecs_to_jiffies(dirty_expire_interval * 10);
625 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200626 if (!wbc.range_cyclic) {
627 wbc.range_start = 0;
628 wbc.range_end = LLONG_MAX;
629 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200630
Jan Kara7624ee72010-08-09 17:20:03 -0700631 wbc.wb_start = jiffies; /* livelock avoidance */
Jens Axboe03ba3782009-09-09 09:08:54 +0200632 for (;;) {
633 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800634 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200635 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200636 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200637 break;
638
639 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800640 * For background writeout, stop when we are below the
641 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200642 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200643 if (work->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200644 break;
645
646 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200647 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
648 wbc.pages_skipped = 0;
Dave Chinner028c2dd2010-07-07 13:24:07 +1000649
650 trace_wbc_writeback_start(&wbc, wb->bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200651 if (work->sb)
652 __writeback_inodes_sb(work->sb, wb, &wbc);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200653 else
654 writeback_inodes_wb(wb, &wbc);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000655 trace_wbc_writeback_written(&wbc, wb->bdi);
656
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200657 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200658 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
659
660 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200661 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200662 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200663 if (wbc.nr_to_write <= 0)
664 continue;
665 /*
666 * Didn't write everything and we don't have more IO, bail
667 */
668 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200669 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200670 /*
671 * Did we write something? Try for more
672 */
673 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
674 continue;
675 /*
676 * Nothing written. Wait for some inode to
677 * become available for writeback. Otherwise
678 * we'll just busyloop.
679 */
680 spin_lock(&inode_lock);
681 if (!list_empty(&wb->b_more_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100682 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000683 trace_wbc_writeback_wait(&wbc, wb->bdi);
Jens Axboe71fd05a2009-09-23 19:32:26 +0200684 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200685 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200686 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200687 }
688
689 return wrote;
690}
691
692/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200693 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +0200694 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200695static struct wb_writeback_work *
Minchan Kim08852b62010-08-03 12:51:16 +0200696get_next_work_item(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200697{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200698 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200699
Artem Bityutskiy64677162010-07-25 14:29:22 +0300700 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200701 if (!list_empty(&bdi->work_list)) {
702 work = list_entry(bdi->work_list.next,
703 struct wb_writeback_work, list);
704 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200705 }
Artem Bityutskiy64677162010-07-25 14:29:22 +0300706 spin_unlock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200707 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200708}
709
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700710/*
711 * Add in the number of potentially dirty inodes, because each inode
712 * write can dirty pagecache in the underlying blockdev.
713 */
714static unsigned long get_nr_dirty_pages(void)
715{
716 return global_page_state(NR_FILE_DIRTY) +
717 global_page_state(NR_UNSTABLE_NFS) +
718 get_nr_dirty_inodes();
719}
720
Jens Axboe03ba3782009-09-09 09:08:54 +0200721static long wb_check_old_data_flush(struct bdi_writeback *wb)
722{
723 unsigned long expired;
724 long nr_pages;
725
Jens Axboe69b62d02010-05-17 12:51:03 +0200726 /*
727 * When set to zero, disable periodic writeback
728 */
729 if (!dirty_writeback_interval)
730 return 0;
731
Jens Axboe03ba3782009-09-09 09:08:54 +0200732 expired = wb->last_old_flush +
733 msecs_to_jiffies(dirty_writeback_interval * 10);
734 if (time_before(jiffies, expired))
735 return 0;
736
737 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -0700738 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +0200739
Jens Axboec4a77a62009-09-16 15:18:25 +0200740 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200741 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200742 .nr_pages = nr_pages,
743 .sync_mode = WB_SYNC_NONE,
744 .for_kupdate = 1,
745 .range_cyclic = 1,
746 };
747
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200748 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +0200749 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200750
751 return 0;
752}
753
754/*
755 * Retrieve work items and do the writeback they describe
756 */
757long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
758{
759 struct backing_dev_info *bdi = wb->bdi;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200760 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200761 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200762
Jan Kara81d73a32010-08-11 14:17:44 -0700763 set_bit(BDI_writeback_running, &wb->bdi->state);
Minchan Kim08852b62010-08-03 12:51:16 +0200764 while ((work = get_next_work_item(bdi)) != NULL) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200765 /*
766 * Override sync mode, in case we must wait for completion
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200767 * because this thread is exiting now.
Jens Axboe03ba3782009-09-09 09:08:54 +0200768 */
769 if (force_wait)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200770 work->sync_mode = WB_SYNC_ALL;
771
Dave Chinner455b2862010-07-07 13:24:06 +1000772 trace_writeback_exec(bdi, work);
773
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200774 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200775
776 /*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200777 * Notify the caller of completion if this is a synchronous
778 * work item, otherwise just free it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200779 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200780 if (work->done)
781 complete(work->done);
782 else
783 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200784 }
785
786 /*
787 * Check for periodic writeback, kupdated() style
788 */
789 wrote += wb_check_old_data_flush(wb);
Jan Kara81d73a32010-08-11 14:17:44 -0700790 clear_bit(BDI_writeback_running, &wb->bdi->state);
Jens Axboe03ba3782009-09-09 09:08:54 +0200791
792 return wrote;
793}
794
795/*
796 * Handle writeback of dirty data for the device backed by this bdi. Also
797 * wakes up periodically and does kupdated style flushing.
798 */
Christoph Hellwig08243902010-06-19 23:08:22 +0200799int bdi_writeback_thread(void *data)
Jens Axboe03ba3782009-09-09 09:08:54 +0200800{
Christoph Hellwig08243902010-06-19 23:08:22 +0200801 struct bdi_writeback *wb = data;
802 struct backing_dev_info *bdi = wb->bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200803 long pages_written;
804
Peter Zijlstra766f9162010-10-26 14:22:45 -0700805 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +0200806 set_freezable();
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300807 wb->last_active = jiffies;
Christoph Hellwig08243902010-06-19 23:08:22 +0200808
809 /*
810 * Our parent may run at a different priority, just set us to normal
811 */
812 set_user_nice(current, 0);
813
Dave Chinner455b2862010-07-07 13:24:06 +1000814 trace_writeback_thread_start(bdi);
815
Jens Axboe03ba3782009-09-09 09:08:54 +0200816 while (!kthread_should_stop()) {
Artem Bityutskiy64677162010-07-25 14:29:22 +0300817 /*
818 * Remove own delayed wake-up timer, since we are already awake
819 * and we'll take care of the preriodic write-back.
820 */
821 del_timer(&wb->wakeup_timer);
822
Jens Axboe03ba3782009-09-09 09:08:54 +0200823 pages_written = wb_do_writeback(wb, 0);
824
Dave Chinner455b2862010-07-07 13:24:06 +1000825 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +0200826
Jens Axboe03ba3782009-09-09 09:08:54 +0200827 if (pages_written)
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300828 wb->last_active = jiffies;
Jens Axboe03ba3782009-09-09 09:08:54 +0200829
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300830 set_current_state(TASK_INTERRUPTIBLE);
J. Bruce Fieldsb76b4012010-08-28 08:52:10 +0200831 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300832 __set_current_state(TASK_RUNNING);
833 continue;
Jens Axboe03ba3782009-09-09 09:08:54 +0200834 }
835
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300836 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300837 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300838 else {
839 /*
840 * We have nothing to do, so can go sleep without any
841 * timeout and save power. When a work is queued or
842 * something is made dirty - we will be woken up.
843 */
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300844 schedule();
Jens Axboef9eadbb2010-05-18 14:31:45 +0200845 }
Jens Axboe69b62d02010-05-17 12:51:03 +0200846
Jens Axboe03ba3782009-09-09 09:08:54 +0200847 try_to_freeze();
848 }
849
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300850 /* Flush any work that raced with us exiting */
Christoph Hellwig08243902010-06-19 23:08:22 +0200851 if (!list_empty(&bdi->work_list))
852 wb_do_writeback(wb, 1);
Dave Chinner455b2862010-07-07 13:24:06 +1000853
854 trace_writeback_thread_stop(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +0200855 return 0;
856}
857
Christoph Hellwig08243902010-06-19 23:08:22 +0200858
Jens Axboe03ba3782009-09-09 09:08:54 +0200859/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200860 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
861 * the whole world.
862 */
863void wakeup_flusher_threads(long nr_pages)
864{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200865 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200866
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200867 if (!nr_pages) {
868 nr_pages = global_page_state(NR_FILE_DIRTY) +
Jens Axboe03ba3782009-09-09 09:08:54 +0200869 global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200870 }
871
872 rcu_read_lock();
873 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
874 if (!bdi_has_dirty_io(bdi))
875 continue;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200876 __bdi_start_writeback(bdi, nr_pages, false, false);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200877 }
878 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200879}
880
881static noinline void block_dump___mark_inode_dirty(struct inode *inode)
882{
883 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
884 struct dentry *dentry;
885 const char *name = "?";
886
887 dentry = d_find_alias(inode);
888 if (dentry) {
889 spin_lock(&dentry->d_lock);
890 name = (const char *) dentry->d_name.name;
891 }
892 printk(KERN_DEBUG
893 "%s(%d): dirtied inode %lu (%s) on %s\n",
894 current->comm, task_pid_nr(current), inode->i_ino,
895 name, inode->i_sb->s_id);
896 if (dentry) {
897 spin_unlock(&dentry->d_lock);
898 dput(dentry);
899 }
900 }
901}
902
903/**
904 * __mark_inode_dirty - internal function
905 * @inode: inode to mark
906 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
907 * Mark an inode as dirty. Callers should use mark_inode_dirty or
908 * mark_inode_dirty_sync.
909 *
910 * Put the inode on the super block's dirty list.
911 *
912 * CAREFUL! We mark it dirty unconditionally, but move it onto the
913 * dirty list only if it is hashed or if it refers to a blockdev.
914 * If it was not hashed, it will never be added to the dirty list
915 * even if it is later hashed, as it will have been marked dirty already.
916 *
917 * In short, make sure you hash any inodes _before_ you start marking
918 * them dirty.
919 *
920 * This function *must* be atomic for the I_DIRTY_PAGES case -
921 * set_page_dirty() is called under spinlock in several places.
922 *
923 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
924 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
925 * the kernel-internal blockdev inode represents the dirtying time of the
926 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
927 * page->mapping->host, so the page-dirtying time is recorded in the internal
928 * blockdev inode.
929 */
930void __mark_inode_dirty(struct inode *inode, int flags)
931{
932 struct super_block *sb = inode->i_sb;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300933 struct backing_dev_info *bdi = NULL;
934 bool wakeup_bdi = false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200935
936 /*
937 * Don't do this for I_DIRTY_PAGES - that doesn't actually
938 * dirty the inode itself
939 */
940 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
941 if (sb->s_op->dirty_inode)
942 sb->s_op->dirty_inode(inode);
943 }
944
945 /*
946 * make sure that changes are seen by all cpus before we test i_state
947 * -- mikulas
948 */
949 smp_mb();
950
951 /* avoid the locking if we can */
952 if ((inode->i_state & flags) == flags)
953 return;
954
955 if (unlikely(block_dump))
956 block_dump___mark_inode_dirty(inode);
957
958 spin_lock(&inode_lock);
959 if ((inode->i_state & flags) != flags) {
960 const int was_dirty = inode->i_state & I_DIRTY;
961
962 inode->i_state |= flags;
963
964 /*
965 * If the inode is being synced, just update its dirty state.
966 * The unlocker will place the inode on the appropriate
967 * superblock list, based upon its state.
968 */
969 if (inode->i_state & I_SYNC)
970 goto out;
971
972 /*
973 * Only add valid (hashed) inodes to the superblock's
974 * dirty list. Add blockdev inodes as well.
975 */
976 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -0400977 if (inode_unhashed(inode))
Jens Axboe03ba3782009-09-09 09:08:54 +0200978 goto out;
979 }
Al Viroa4ffdde2010-06-02 17:38:30 -0400980 if (inode->i_state & I_FREEING)
Jens Axboe03ba3782009-09-09 09:08:54 +0200981 goto out;
982
983 /*
984 * If the inode was already on b_dirty/b_io/b_more_io, don't
985 * reposition it (that would break b_dirty time-ordering).
986 */
987 if (!was_dirty) {
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300988 bdi = inode_to_bdi(inode);
Jens Axboe500b0672009-09-09 09:10:25 +0200989
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300990 if (bdi_cap_writeback_dirty(bdi)) {
991 WARN(!test_bit(BDI_registered, &bdi->state),
992 "bdi-%s not registered\n", bdi->name);
993
994 /*
995 * If this is the first dirty inode for this
996 * bdi, we have to wake-up the corresponding
997 * bdi thread to make sure background
998 * write-back happens later.
999 */
1000 if (!wb_has_dirty_io(&bdi->wb))
1001 wakeup_bdi = true;
Jens Axboe500b0672009-09-09 09:10:25 +02001002 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001003
1004 inode->dirtied_when = jiffies;
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001005 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
Jens Axboe03ba3782009-09-09 09:08:54 +02001006 }
1007 }
1008out:
1009 spin_unlock(&inode_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03001010
1011 if (wakeup_bdi)
Artem Bityutskiy64677162010-07-25 14:29:22 +03001012 bdi_wakeup_thread_delayed(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +02001013}
1014EXPORT_SYMBOL(__mark_inode_dirty);
1015
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001016/*
1017 * Write out a superblock's list of dirty inodes. A wait will be performed
1018 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1019 *
1020 * If older_than_this is non-NULL, then only write out inodes which
1021 * had their first dirtying at a time earlier than *older_than_this.
1022 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001023 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1024 * This function assumes that the blockdev superblock's inodes are backed by
1025 * a variety of queues, so all inodes are searched. For other superblocks,
1026 * assume that all inodes are backed by the same queue.
1027 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001028 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1029 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1030 * on the writer throttling path, and we get decent balancing between many
1031 * throttled threads: we don't want them all piling up on inode_sync_wait.
1032 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001033static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001034{
Jens Axboe03ba3782009-09-09 09:08:54 +02001035 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001036
Jens Axboe03ba3782009-09-09 09:08:54 +02001037 /*
1038 * We need to be protected against the filesystem going from
1039 * r/o to r/w or vice versa.
1040 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001041 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001042
Jens Axboe03ba3782009-09-09 09:08:54 +02001043 spin_lock(&inode_lock);
1044
1045 /*
1046 * Data integrity sync. Must wait for all pages under writeback,
1047 * because there may have been pages dirtied before our sync
1048 * call, but which had writeout started before we write it out.
1049 * In which case, the inode may not be on the dirty list, but
1050 * we still have to wait for that writeout.
1051 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001052 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001053 struct address_space *mapping;
1054
Al Viroa4ffdde2010-06-02 17:38:30 -04001055 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW))
Jens Axboe03ba3782009-09-09 09:08:54 +02001056 continue;
1057 mapping = inode->i_mapping;
1058 if (mapping->nrpages == 0)
1059 continue;
1060 __iget(inode);
1061 spin_unlock(&inode_lock);
1062 /*
1063 * We hold a reference to 'inode' so it couldn't have
1064 * been removed from s_inodes list while we dropped the
1065 * inode_lock. We cannot iput the inode now as we can
1066 * be holding the last reference and we cannot iput it
1067 * under inode_lock. So we keep the reference and iput
1068 * it later.
1069 */
1070 iput(old_inode);
1071 old_inode = inode;
1072
1073 filemap_fdatawait(mapping);
1074
1075 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001076
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001077 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001078 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001079 spin_unlock(&inode_lock);
1080 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081}
1082
Jens Axboed8a85592009-09-02 12:34:32 +02001083/**
Chris Mason3259f8b2010-10-29 11:16:17 -04001084 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
Jens Axboed8a85592009-09-02 12:34:32 +02001085 * @sb: the superblock
Chris Mason3259f8b2010-10-29 11:16:17 -04001086 * @nr: the number of pages to write
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 *
Jens Axboed8a85592009-09-02 12:34:32 +02001088 * Start writeback on some inodes on this super_block. No guarantees are made
1089 * on how many (if any) will be written, and this function does not wait
Chris Mason3259f8b2010-10-29 11:16:17 -04001090 * for IO completion of submitted IO.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 */
Chris Mason3259f8b2010-10-29 11:16:17 -04001092void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001094 DECLARE_COMPLETION_ONSTACK(done);
1095 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001096 .sb = sb,
1097 .sync_mode = WB_SYNC_NONE,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001098 .done = &done,
Chris Mason3259f8b2010-10-29 11:16:17 -04001099 .nr_pages = nr,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001100 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001101
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001102 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001103 bdi_queue_work(sb->s_bdi, &work);
1104 wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105}
Chris Mason3259f8b2010-10-29 11:16:17 -04001106EXPORT_SYMBOL(writeback_inodes_sb_nr);
1107
1108/**
1109 * writeback_inodes_sb - writeback dirty inodes from given super_block
1110 * @sb: the superblock
1111 *
1112 * Start writeback on some inodes on this super_block. No guarantees are made
1113 * on how many (if any) will be written, and this function does not wait
1114 * for IO completion of submitted IO.
1115 */
1116void writeback_inodes_sb(struct super_block *sb)
1117{
Linus Torvalds925d1692010-10-30 09:05:48 -07001118 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
Chris Mason3259f8b2010-10-29 11:16:17 -04001119}
Jens Axboed8a85592009-09-02 12:34:32 +02001120EXPORT_SYMBOL(writeback_inodes_sb);
1121
1122/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001123 * writeback_inodes_sb_if_idle - start writeback if none underway
1124 * @sb: the superblock
1125 *
1126 * Invoke writeback_inodes_sb if no writeback is currently underway.
1127 * Returns 1 if writeback was started, 0 if not.
1128 */
1129int writeback_inodes_sb_if_idle(struct super_block *sb)
1130{
1131 if (!writeback_in_progress(sb->s_bdi)) {
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001132 down_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001133 writeback_inodes_sb(sb);
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001134 up_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001135 return 1;
1136 } else
1137 return 0;
1138}
1139EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1140
1141/**
Chris Mason3259f8b2010-10-29 11:16:17 -04001142 * writeback_inodes_sb_if_idle - start writeback if none underway
1143 * @sb: the superblock
1144 * @nr: the number of pages to write
1145 *
1146 * Invoke writeback_inodes_sb if no writeback is currently underway.
1147 * Returns 1 if writeback was started, 0 if not.
1148 */
1149int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1150 unsigned long nr)
1151{
1152 if (!writeback_in_progress(sb->s_bdi)) {
1153 down_read(&sb->s_umount);
1154 writeback_inodes_sb_nr(sb, nr);
1155 up_read(&sb->s_umount);
1156 return 1;
1157 } else
1158 return 0;
1159}
1160EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1161
1162/**
Jens Axboed8a85592009-09-02 12:34:32 +02001163 * sync_inodes_sb - sync sb inode pages
1164 * @sb: the superblock
1165 *
1166 * This function writes and waits on any dirty inode belonging to this
1167 * super_block. The number of pages synced is returned.
1168 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001169void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001170{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001171 DECLARE_COMPLETION_ONSTACK(done);
1172 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001173 .sb = sb,
1174 .sync_mode = WB_SYNC_ALL,
1175 .nr_pages = LONG_MAX,
1176 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001177 .done = &done,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001178 };
1179
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001180 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1181
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001182 bdi_queue_work(sb->s_bdi, &work);
1183 wait_for_completion(&done);
1184
Jens Axboeb6e51312009-09-16 15:13:54 +02001185 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001186}
1187EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001190 * write_inode_now - write an inode to disk
1191 * @inode: inode to write to disk
1192 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001194 * This function commits an inode to disk immediately if it is dirty. This is
1195 * primarily needed by knfsd.
1196 *
1197 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199int write_inode_now(struct inode *inode, int sync)
1200{
1201 int ret;
1202 struct writeback_control wbc = {
1203 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001204 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001205 .range_start = 0,
1206 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 };
1208
1209 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001210 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211
1212 might_sleep();
1213 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001214 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 spin_unlock(&inode_lock);
1216 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001217 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 return ret;
1219}
1220EXPORT_SYMBOL(write_inode_now);
1221
1222/**
1223 * sync_inode - write an inode and its pages to disk.
1224 * @inode: the inode to sync
1225 * @wbc: controls the writeback mode
1226 *
1227 * sync_inode() will write an inode and its pages to disk. It will also
1228 * correctly update the inode on its superblock's dirty inode lists and will
1229 * update inode->i_state.
1230 *
1231 * The caller must have a ref on the inode.
1232 */
1233int sync_inode(struct inode *inode, struct writeback_control *wbc)
1234{
1235 int ret;
1236
1237 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001238 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 spin_unlock(&inode_lock);
1240 return ret;
1241}
1242EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02001243
1244/**
1245 * sync_inode - write an inode to disk
1246 * @inode: the inode to sync
1247 * @wait: wait for I/O to complete.
1248 *
1249 * Write an inode to disk and adjust it's dirty state after completion.
1250 *
1251 * Note: only writes the actual inode, no associated data or other metadata.
1252 */
1253int sync_inode_metadata(struct inode *inode, int wait)
1254{
1255 struct writeback_control wbc = {
1256 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1257 .nr_to_write = 0, /* metadata-only */
1258 };
1259
1260 return sync_inode(inode, &wbc);
1261}
1262EXPORT_SYMBOL(sync_inode_metadata);