blob: e8f65290e83666bd1a32c49880d502a698a22904 [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 Howells07f3f052006-09-30 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
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020082static void bdi_queue_work(struct backing_dev_info *bdi,
83 struct wb_writeback_work *work)
Nick Piggin4195f732009-05-28 09:01:15 +020084{
Dave Chinner455b2862010-07-07 13:24:06 +100085 trace_writeback_queue(bdi, work);
86
Artem Bityutskiy64677162010-07-25 14:29:22 +030087 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020088 list_add_tail(&work->list, &bdi->work_list);
Artem Bityutskiyfff5b852010-07-25 14:29:20 +030089 if (bdi->wb.task) {
90 wake_up_process(bdi->wb.task);
91 } else {
92 /*
93 * The bdi thread isn't there, wake up the forker thread which
94 * will create and run it.
95 */
Dave Chinner455b2862010-07-07 13:24:06 +100096 trace_writeback_nothread(bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +020097 wake_up_process(default_backing_dev_info.wb.task);
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 }
Artem Bityutskiy64677162010-07-25 14:29:22 +030099 spin_unlock_bh(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100}
101
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200102static void
103__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
104 bool range_cyclic, bool for_background)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200106 struct wb_writeback_work *work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200107
Jens Axboebcddc3f2009-09-13 20:07:36 +0200108 /*
109 * This is WB_SYNC_NONE writeback, so if allocation fails just
110 * wakeup the thread for old dirty data writeback
111 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200112 work = kzalloc(sizeof(*work), GFP_ATOMIC);
113 if (!work) {
Dave Chinner455b2862010-07-07 13:24:06 +1000114 if (bdi->wb.task) {
115 trace_writeback_nowork(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200116 wake_up_process(bdi->wb.task);
Dave Chinner455b2862010-07-07 13:24:06 +1000117 }
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200118 return;
Jens Axboebcddc3f2009-09-13 20:07:36 +0200119 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200120
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200121 work->sync_mode = WB_SYNC_NONE;
122 work->nr_pages = nr_pages;
123 work->range_cyclic = range_cyclic;
124 work->for_background = for_background;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200125
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200126 bdi_queue_work(bdi, work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200127}
128
129/**
130 * bdi_start_writeback - start writeback
131 * @bdi: the backing device to write from
132 * @nr_pages: the number of pages to write
133 *
134 * Description:
135 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
136 * started when this function returns, we make no guarentees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200137 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200138 *
139 */
Christoph Hellwigc5444192010-06-08 18:15:15 +0200140void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200141{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200142 __bdi_start_writeback(bdi, nr_pages, true, false);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200143}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800144
Christoph Hellwigc5444192010-06-08 18:15:15 +0200145/**
146 * bdi_start_background_writeback - start background writeback
147 * @bdi: the backing device to write from
148 *
149 * Description:
150 * This does WB_SYNC_NONE background writeback. The IO is only
151 * started when this function returns, we make no guarentees on
152 * completion. Caller need not hold sb s_umount semaphore.
153 */
154void bdi_start_background_writeback(struct backing_dev_info *bdi)
155{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200156 __bdi_start_writeback(bdi, LONG_MAX, true, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157}
158
159/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700160 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
161 * furthest end of its superblock's dirty-inode list.
162 *
163 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200164 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700165 * the case then the inode must have been redirtied while it was being written
166 * out and we don't reset its dirtied_when.
167 */
168static void redirty_tail(struct inode *inode)
169{
Jens Axboe03ba3782009-09-09 09:08:54 +0200170 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700171
Jens Axboe03ba3782009-09-09 09:08:54 +0200172 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200173 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700174
Jens Axboe03ba3782009-09-09 09:08:54 +0200175 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200176 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700177 inode->dirtied_when = jiffies;
178 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200179 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700180}
181
182/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200183 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700184 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700185static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700186{
Jens Axboe03ba3782009-09-09 09:08:54 +0200187 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
188
189 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700190}
191
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700192static void inode_sync_complete(struct inode *inode)
193{
194 /*
195 * Prevent speculative execution through spin_unlock(&inode_lock);
196 */
197 smp_mb();
198 wake_up_bit(&inode->i_state, __I_SYNC);
199}
200
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700201static bool inode_dirtied_after(struct inode *inode, unsigned long t)
202{
203 bool ret = time_after(inode->dirtied_when, t);
204#ifndef CONFIG_64BIT
205 /*
206 * For inodes being constantly redirtied, dirtied_when can get stuck.
207 * It _appears_ to be in the future, but is actually in distant past.
208 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200209 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700210 */
211 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
212#endif
213 return ret;
214}
215
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700216/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700217 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
218 */
219static void move_expired_inodes(struct list_head *delaying_queue,
220 struct list_head *dispatch_queue,
221 unsigned long *older_than_this)
222{
Shaohua Li5c034492009-09-24 14:42:33 +0200223 LIST_HEAD(tmp);
224 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200225 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200226 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200227 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200228
Fengguang Wu2c136572007-10-16 23:30:39 -0700229 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200230 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700231 if (older_than_this &&
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700232 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700233 break;
Jens Axboecf137302009-09-24 15:12:57 +0200234 if (sb && sb != inode->i_sb)
235 do_sb_sort = 1;
236 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200237 list_move(&inode->i_list, &tmp);
238 }
239
Jens Axboecf137302009-09-24 15:12:57 +0200240 /* just one sb in list, splice to dispatch_queue and we're done */
241 if (!do_sb_sort) {
242 list_splice(&tmp, dispatch_queue);
243 return;
244 }
245
Shaohua Li5c034492009-09-24 14:42:33 +0200246 /* Move inodes from one superblock together */
247 while (!list_empty(&tmp)) {
248 inode = list_entry(tmp.prev, struct inode, i_list);
249 sb = inode->i_sb;
250 list_for_each_prev_safe(pos, node, &tmp) {
251 inode = list_entry(pos, struct inode, i_list);
252 if (inode->i_sb == sb)
253 list_move(&inode->i_list, dispatch_queue);
254 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700255 }
256}
257
258/*
259 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700260 * Before
261 * newly dirtied b_dirty b_io b_more_io
262 * =============> gf edc BA
263 * After
264 * newly dirtied b_dirty b_io b_more_io
265 * =============> g fBAedc
266 * |
267 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -0700268 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200269static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700270{
Wu Fengguang4ea879b2010-08-11 14:17:42 -0700271 list_splice_init(&wb->b_more_io, &wb->b_io);
Jens Axboe03ba3782009-09-09 09:08:54 +0200272 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200273}
274
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100275static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200276{
Jens Axboe03ba3782009-09-09 09:08:54 +0200277 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100278 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200279 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700280}
281
282/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200283 * Wait for writeback on an inode to complete.
284 */
285static void inode_wait_for_writeback(struct inode *inode)
286{
287 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
288 wait_queue_head_t *wqh;
289
290 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Richard Kennedy58a9d3d2010-05-24 14:32:38 -0700291 while (inode->i_state & I_SYNC) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200292 spin_unlock(&inode_lock);
293 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
294 spin_lock(&inode_lock);
Richard Kennedy58a9d3d2010-05-24 14:32:38 -0700295 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200296}
297
298/*
299 * Write out an inode's dirty pages. Called under inode_lock. Either the
300 * caller has ref on the inode (either via __iget or via syscall against an fd)
301 * or the inode has I_WILL_FREE set (via generic_forget_inode)
302 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 * If `wait' is set, wait on the writeout.
304 *
305 * The whole writeout design is quite complex and fragile. We want to avoid
306 * starvation of particular inodes when others are being redirtied, prevent
307 * livelocks, etc.
308 *
309 * Called under inode_lock.
310 */
311static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200312writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200315 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 int ret;
317
Christoph Hellwig01c03192009-06-08 13:35:40 +0200318 if (!atomic_read(&inode->i_count))
319 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
320 else
321 WARN_ON(inode->i_state & I_WILL_FREE);
322
323 if (inode->i_state & I_SYNC) {
324 /*
325 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200326 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200327 * writeback can proceed with the other inodes on s_io.
328 *
329 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200330 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200331 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100332 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200333 requeue_io(inode);
334 return 0;
335 }
336
337 /*
338 * It's a data-integrity sync. We must wait.
339 */
340 inode_wait_for_writeback(inode);
341 }
342
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700343 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400345 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700346 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400347 inode->i_state &= ~I_DIRTY_PAGES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 spin_unlock(&inode_lock);
349
350 ret = do_writepages(mapping, wbc);
351
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100352 /*
353 * Make sure to wait on the data before writing out the metadata.
354 * This is important for filesystems that modify metadata on data
355 * I/O completion.
356 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100357 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100358 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 if (ret == 0)
360 ret = err;
361 }
362
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400363 /*
364 * Some filesystems may redirty the inode during the writeback
365 * due to delalloc, clear dirty metadata flags right before
366 * write_inode()
367 */
368 spin_lock(&inode_lock);
369 dirty = inode->i_state & I_DIRTY;
370 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
371 spin_unlock(&inode_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100372 /* Don't write the inode if only I_DIRTY_PAGES was set */
373 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100374 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375 if (ret == 0)
376 ret = err;
377 }
378
379 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700380 inode->i_state &= ~I_SYNC;
Al Viroa4ffdde2010-06-02 17:38:30 -0400381 if (!(inode->i_state & I_FREEING)) {
Wu Fengguang23539af2010-08-11 14:17:41 -0700382 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 /*
384 * We didn't write back all the pages. nfs_writepages()
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700385 * sometimes bales out without doing anything.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700386 */
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700387 inode->i_state |= I_DIRTY_PAGES;
388 if (wbc->nr_to_write <= 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700390 * slice used up: queue for next turn
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 */
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700392 requeue_io(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 } else {
394 /*
Wu Fengguanga50aeb42010-08-11 14:17:43 -0700395 * Writeback blocked by something other than
396 * congestion. Delay the inode for some time to
397 * avoid spinning on the CPU (100% iowait)
398 * retrying writeback of the dirty page/inode
399 * that cannot be performed immediately.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 */
Andrew Morton1b43ef92007-10-16 23:30:35 -0700401 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 }
Wu Fengguang23539af2010-08-11 14:17:41 -0700403 } else if (inode->i_state & I_DIRTY) {
404 /*
405 * Filesystems can dirty the inode during writeback
406 * operations, such as delayed allocation during
407 * submission or metadata updates after data IO
408 * completion.
409 */
410 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 } else {
412 /*
Nick Piggin9e38d862010-10-23 06:55:17 -0400413 * The inode is clean. At this point we either have
414 * a reference to the inode or it's on it's way out.
415 * No need to add it back to the LRU.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 */
Nick Piggin9e38d862010-10-23 06:55:17 -0400417 list_del_init(&inode->i_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 }
419 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700420 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 return ret;
422}
423
Jens Axboe03ba3782009-09-09 09:08:54 +0200424/*
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200425 * For background writeback the caller does not have the sb pinned
Jens Axboe03ba3782009-09-09 09:08:54 +0200426 * before calling writeback. So make sure that we do pin it, so it doesn't
427 * go away while we are writing inodes from it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200428 */
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200429static bool pin_sb_for_writeback(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
Jens Axboe03ba3782009-09-09 09:08:54 +0200431 spin_lock(&sb_lock);
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200432 if (list_empty(&sb->s_instances)) {
433 spin_unlock(&sb_lock);
434 return false;
435 }
436
Jens Axboe03ba3782009-09-09 09:08:54 +0200437 sb->s_count++;
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200438 spin_unlock(&sb_lock);
439
Jens Axboe03ba3782009-09-09 09:08:54 +0200440 if (down_read_trylock(&sb->s_umount)) {
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200441 if (sb->s_root)
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200442 return true;
Jens Axboe03ba3782009-09-09 09:08:54 +0200443 up_read(&sb->s_umount);
444 }
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200445
446 put_super(sb);
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200447 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200448}
449
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800450/*
451 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200452 *
453 * If @only_this_sb is true, then find and write all such
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800454 * inodes. Otherwise write only ones which go sequentially
455 * in reverse order.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200456 *
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800457 * Return 1, if the caller writeback routine should be
458 * interrupted. Otherwise return 0.
459 */
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200460static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
461 struct writeback_control *wbc, bool only_this_sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200462{
Jens Axboe03ba3782009-09-09 09:08:54 +0200463 while (!list_empty(&wb->b_io)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 long pages_skipped;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800465 struct inode *inode = list_entry(wb->b_io.prev,
466 struct inode, i_list);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200467
468 if (inode->i_sb != sb) {
469 if (only_this_sb) {
470 /*
471 * We only want to write back data for this
472 * superblock, move all inodes not belonging
473 * to it back onto the dirty list.
474 */
475 redirty_tail(inode);
476 continue;
477 }
478
479 /*
480 * The inode belongs to a different superblock.
481 * Bounce back to the caller to unpin this and
482 * pin the next superblock.
483 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800484 return 0;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200485 }
486
Wu Fengguang84a89242009-06-16 15:33:17 -0700487 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700488 requeue_io(inode);
489 continue;
490 }
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700491 /*
492 * Was this inode dirtied after sync_sb_inodes was called?
493 * This keeps sync from extra jobs and livelock.
494 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800495 if (inode_dirtied_after(inode, wbc->wb_start))
496 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497
Al Viroa4ffdde2010-06-02 17:38:30 -0400498 BUG_ON(inode->i_state & I_FREEING);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 __iget(inode);
500 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200501 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 if (wbc->pages_skipped != pages_skipped) {
503 /*
504 * writeback is not making progress due to locked
505 * buffers. Skip this inode for now.
506 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700507 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 }
509 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800511 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800513 if (wbc->nr_to_write <= 0) {
514 wbc->more_io = 1;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800515 return 1;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800516 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200517 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800518 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800520 /* b_io is empty */
521 return 1;
522}
Nick Piggin38f21972009-01-06 14:40:25 -0800523
Christoph Hellwig9c3a8ee2010-06-10 12:07:27 +0200524void writeback_inodes_wb(struct bdi_writeback *wb,
525 struct writeback_control *wbc)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800526{
527 int ret = 0;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200528
Jan Kara7624ee72010-08-09 17:20:03 -0700529 if (!wbc->wb_start)
530 wbc->wb_start = jiffies; /* livelock avoidance */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800531 spin_lock(&inode_lock);
532 if (!wbc->for_kupdate || list_empty(&wb->b_io))
533 queue_io(wb, wbc->older_than_this);
534
535 while (!list_empty(&wb->b_io)) {
536 struct inode *inode = list_entry(wb->b_io.prev,
537 struct inode, i_list);
538 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800539
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200540 if (!pin_sb_for_writeback(sb)) {
541 requeue_io(inode);
542 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800543 }
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200544 ret = writeback_sb_inodes(sb, wb, wbc, false);
545 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800546
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800547 if (ret)
548 break;
549 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200550 spin_unlock(&inode_lock);
551 /* Leave any unwritten inodes on b_io */
552}
553
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200554static void __writeback_inodes_sb(struct super_block *sb,
555 struct bdi_writeback *wb, struct writeback_control *wbc)
556{
557 WARN_ON(!rwsem_is_locked(&sb->s_umount));
558
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200559 spin_lock(&inode_lock);
560 if (!wbc->for_kupdate || list_empty(&wb->b_io))
561 queue_io(wb, wbc->older_than_this);
562 writeback_sb_inodes(sb, wb, wbc, true);
563 spin_unlock(&inode_lock);
564}
565
Jens Axboe03ba3782009-09-09 09:08:54 +0200566/*
567 * The maximum number of pages to writeout in a single bdi flush/kupdate
568 * operation. We do this so we don't hold I_SYNC against an inode for
569 * enormous amounts of time, which would block a userspace task which has
570 * been forced to throttle against that inode. Also, the code reevaluates
571 * the dirty each time it has written this many pages.
572 */
573#define MAX_WRITEBACK_PAGES 1024
574
575static inline bool over_bground_thresh(void)
576{
577 unsigned long background_thresh, dirty_thresh;
578
Wu Fengguang16c40422010-08-11 14:17:39 -0700579 global_dirty_limits(&background_thresh, &dirty_thresh);
Jens Axboe03ba3782009-09-09 09:08:54 +0200580
581 return (global_page_state(NR_FILE_DIRTY) +
582 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
583}
584
585/*
586 * Explicit flushing or periodic writeback of "old" data.
587 *
588 * Define "old": the first time one of an inode's pages is dirtied, we mark the
589 * dirtying-time in the inode's address_space. So this periodic writeback code
590 * just walks the superblock inode list, writing back any inodes which are
591 * older than a specific point in time.
592 *
593 * Try to run once per dirty_writeback_interval. But if a writeback event
594 * takes longer than a dirty_writeback_interval interval, then leave a
595 * one-second gap.
596 *
597 * older_than_this takes precedence over nr_to_write. So we'll only write back
598 * all dirty pages if they are all attached to "old" mappings.
599 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200600static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200601 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200602{
603 struct writeback_control wbc = {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200604 .sync_mode = work->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200605 .older_than_this = NULL,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200606 .for_kupdate = work->for_kupdate,
607 .for_background = work->for_background,
608 .range_cyclic = work->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200609 };
610 unsigned long oldest_jif;
611 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200612 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200613
614 if (wbc.for_kupdate) {
615 wbc.older_than_this = &oldest_jif;
616 oldest_jif = jiffies -
617 msecs_to_jiffies(dirty_expire_interval * 10);
618 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200619 if (!wbc.range_cyclic) {
620 wbc.range_start = 0;
621 wbc.range_end = LLONG_MAX;
622 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200623
Jan Kara7624ee72010-08-09 17:20:03 -0700624 wbc.wb_start = jiffies; /* livelock avoidance */
Jens Axboe03ba3782009-09-09 09:08:54 +0200625 for (;;) {
626 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800627 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200628 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200629 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200630 break;
631
632 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800633 * For background writeout, stop when we are below the
634 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200635 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200636 if (work->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200637 break;
638
639 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200640 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
641 wbc.pages_skipped = 0;
Dave Chinner028c2dd2010-07-07 13:24:07 +1000642
643 trace_wbc_writeback_start(&wbc, wb->bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200644 if (work->sb)
645 __writeback_inodes_sb(work->sb, wb, &wbc);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200646 else
647 writeback_inodes_wb(wb, &wbc);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000648 trace_wbc_writeback_written(&wbc, wb->bdi);
649
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200650 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200651 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
652
653 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200654 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200655 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200656 if (wbc.nr_to_write <= 0)
657 continue;
658 /*
659 * Didn't write everything and we don't have more IO, bail
660 */
661 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200662 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200663 /*
664 * Did we write something? Try for more
665 */
666 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
667 continue;
668 /*
669 * Nothing written. Wait for some inode to
670 * become available for writeback. Otherwise
671 * we'll just busyloop.
672 */
673 spin_lock(&inode_lock);
674 if (!list_empty(&wb->b_more_io)) {
675 inode = list_entry(wb->b_more_io.prev,
676 struct inode, i_list);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000677 trace_wbc_writeback_wait(&wbc, wb->bdi);
Jens Axboe71fd05a2009-09-23 19:32:26 +0200678 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200679 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200680 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200681 }
682
683 return wrote;
684}
685
686/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200687 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +0200688 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200689static struct wb_writeback_work *
Minchan Kim08852b62010-08-03 12:51:16 +0200690get_next_work_item(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200691{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200692 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200693
Artem Bityutskiy64677162010-07-25 14:29:22 +0300694 spin_lock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200695 if (!list_empty(&bdi->work_list)) {
696 work = list_entry(bdi->work_list.next,
697 struct wb_writeback_work, list);
698 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200699 }
Artem Bityutskiy64677162010-07-25 14:29:22 +0300700 spin_unlock_bh(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200701 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200702}
703
704static long wb_check_old_data_flush(struct bdi_writeback *wb)
705{
706 unsigned long expired;
707 long nr_pages;
708
Jens Axboe69b62d02010-05-17 12:51:03 +0200709 /*
710 * When set to zero, disable periodic writeback
711 */
712 if (!dirty_writeback_interval)
713 return 0;
714
Jens Axboe03ba3782009-09-09 09:08:54 +0200715 expired = wb->last_old_flush +
716 msecs_to_jiffies(dirty_writeback_interval * 10);
717 if (time_before(jiffies, expired))
718 return 0;
719
720 wb->last_old_flush = jiffies;
721 nr_pages = global_page_state(NR_FILE_DIRTY) +
722 global_page_state(NR_UNSTABLE_NFS) +
Dave Chinnercffbc8a2010-10-23 05:03:02 -0400723 get_nr_dirty_inodes();
Jens Axboe03ba3782009-09-09 09:08:54 +0200724
Jens Axboec4a77a62009-09-16 15:18:25 +0200725 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200726 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200727 .nr_pages = nr_pages,
728 .sync_mode = WB_SYNC_NONE,
729 .for_kupdate = 1,
730 .range_cyclic = 1,
731 };
732
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200733 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +0200734 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200735
736 return 0;
737}
738
739/*
740 * Retrieve work items and do the writeback they describe
741 */
742long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
743{
744 struct backing_dev_info *bdi = wb->bdi;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200745 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200746 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200747
Jan Kara81d73a32010-08-11 14:17:44 -0700748 set_bit(BDI_writeback_running, &wb->bdi->state);
Minchan Kim08852b62010-08-03 12:51:16 +0200749 while ((work = get_next_work_item(bdi)) != NULL) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200750 /*
751 * Override sync mode, in case we must wait for completion
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200752 * because this thread is exiting now.
Jens Axboe03ba3782009-09-09 09:08:54 +0200753 */
754 if (force_wait)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200755 work->sync_mode = WB_SYNC_ALL;
756
Dave Chinner455b2862010-07-07 13:24:06 +1000757 trace_writeback_exec(bdi, work);
758
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200759 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200760
761 /*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200762 * Notify the caller of completion if this is a synchronous
763 * work item, otherwise just free it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200764 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200765 if (work->done)
766 complete(work->done);
767 else
768 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200769 }
770
771 /*
772 * Check for periodic writeback, kupdated() style
773 */
774 wrote += wb_check_old_data_flush(wb);
Jan Kara81d73a32010-08-11 14:17:44 -0700775 clear_bit(BDI_writeback_running, &wb->bdi->state);
Jens Axboe03ba3782009-09-09 09:08:54 +0200776
777 return wrote;
778}
779
780/*
781 * Handle writeback of dirty data for the device backed by this bdi. Also
782 * wakes up periodically and does kupdated style flushing.
783 */
Christoph Hellwig08243902010-06-19 23:08:22 +0200784int bdi_writeback_thread(void *data)
Jens Axboe03ba3782009-09-09 09:08:54 +0200785{
Christoph Hellwig08243902010-06-19 23:08:22 +0200786 struct bdi_writeback *wb = data;
787 struct backing_dev_info *bdi = wb->bdi;
Jens Axboe03ba3782009-09-09 09:08:54 +0200788 long pages_written;
789
Christoph Hellwig08243902010-06-19 23:08:22 +0200790 current->flags |= PF_FLUSHER | PF_SWAPWRITE;
791 set_freezable();
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300792 wb->last_active = jiffies;
Christoph Hellwig08243902010-06-19 23:08:22 +0200793
794 /*
795 * Our parent may run at a different priority, just set us to normal
796 */
797 set_user_nice(current, 0);
798
Dave Chinner455b2862010-07-07 13:24:06 +1000799 trace_writeback_thread_start(bdi);
800
Jens Axboe03ba3782009-09-09 09:08:54 +0200801 while (!kthread_should_stop()) {
Artem Bityutskiy64677162010-07-25 14:29:22 +0300802 /*
803 * Remove own delayed wake-up timer, since we are already awake
804 * and we'll take care of the preriodic write-back.
805 */
806 del_timer(&wb->wakeup_timer);
807
Jens Axboe03ba3782009-09-09 09:08:54 +0200808 pages_written = wb_do_writeback(wb, 0);
809
Dave Chinner455b2862010-07-07 13:24:06 +1000810 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +0200811
Jens Axboe03ba3782009-09-09 09:08:54 +0200812 if (pages_written)
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300813 wb->last_active = jiffies;
Jens Axboe03ba3782009-09-09 09:08:54 +0200814
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300815 set_current_state(TASK_INTERRUPTIBLE);
J. Bruce Fieldsb76b4012010-08-28 08:52:10 +0200816 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300817 __set_current_state(TASK_RUNNING);
818 continue;
Jens Axboe03ba3782009-09-09 09:08:54 +0200819 }
820
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300821 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300822 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300823 else {
824 /*
825 * We have nothing to do, so can go sleep without any
826 * timeout and save power. When a work is queued or
827 * something is made dirty - we will be woken up.
828 */
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300829 schedule();
Jens Axboef9eadbb2010-05-18 14:31:45 +0200830 }
Jens Axboe69b62d02010-05-17 12:51:03 +0200831
Jens Axboe03ba3782009-09-09 09:08:54 +0200832 try_to_freeze();
833 }
834
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300835 /* Flush any work that raced with us exiting */
Christoph Hellwig08243902010-06-19 23:08:22 +0200836 if (!list_empty(&bdi->work_list))
837 wb_do_writeback(wb, 1);
Dave Chinner455b2862010-07-07 13:24:06 +1000838
839 trace_writeback_thread_stop(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +0200840 return 0;
841}
842
Christoph Hellwig08243902010-06-19 23:08:22 +0200843
Jens Axboe03ba3782009-09-09 09:08:54 +0200844/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200845 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
846 * the whole world.
847 */
848void wakeup_flusher_threads(long nr_pages)
849{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200850 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200851
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200852 if (!nr_pages) {
853 nr_pages = global_page_state(NR_FILE_DIRTY) +
Jens Axboe03ba3782009-09-09 09:08:54 +0200854 global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200855 }
856
857 rcu_read_lock();
858 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
859 if (!bdi_has_dirty_io(bdi))
860 continue;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200861 __bdi_start_writeback(bdi, nr_pages, false, false);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200862 }
863 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200864}
865
866static noinline void block_dump___mark_inode_dirty(struct inode *inode)
867{
868 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
869 struct dentry *dentry;
870 const char *name = "?";
871
872 dentry = d_find_alias(inode);
873 if (dentry) {
874 spin_lock(&dentry->d_lock);
875 name = (const char *) dentry->d_name.name;
876 }
877 printk(KERN_DEBUG
878 "%s(%d): dirtied inode %lu (%s) on %s\n",
879 current->comm, task_pid_nr(current), inode->i_ino,
880 name, inode->i_sb->s_id);
881 if (dentry) {
882 spin_unlock(&dentry->d_lock);
883 dput(dentry);
884 }
885 }
886}
887
888/**
889 * __mark_inode_dirty - internal function
890 * @inode: inode to mark
891 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
892 * Mark an inode as dirty. Callers should use mark_inode_dirty or
893 * mark_inode_dirty_sync.
894 *
895 * Put the inode on the super block's dirty list.
896 *
897 * CAREFUL! We mark it dirty unconditionally, but move it onto the
898 * dirty list only if it is hashed or if it refers to a blockdev.
899 * If it was not hashed, it will never be added to the dirty list
900 * even if it is later hashed, as it will have been marked dirty already.
901 *
902 * In short, make sure you hash any inodes _before_ you start marking
903 * them dirty.
904 *
905 * This function *must* be atomic for the I_DIRTY_PAGES case -
906 * set_page_dirty() is called under spinlock in several places.
907 *
908 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
909 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
910 * the kernel-internal blockdev inode represents the dirtying time of the
911 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
912 * page->mapping->host, so the page-dirtying time is recorded in the internal
913 * blockdev inode.
914 */
915void __mark_inode_dirty(struct inode *inode, int flags)
916{
917 struct super_block *sb = inode->i_sb;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300918 struct backing_dev_info *bdi = NULL;
919 bool wakeup_bdi = false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200920
921 /*
922 * Don't do this for I_DIRTY_PAGES - that doesn't actually
923 * dirty the inode itself
924 */
925 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
926 if (sb->s_op->dirty_inode)
927 sb->s_op->dirty_inode(inode);
928 }
929
930 /*
931 * make sure that changes are seen by all cpus before we test i_state
932 * -- mikulas
933 */
934 smp_mb();
935
936 /* avoid the locking if we can */
937 if ((inode->i_state & flags) == flags)
938 return;
939
940 if (unlikely(block_dump))
941 block_dump___mark_inode_dirty(inode);
942
943 spin_lock(&inode_lock);
944 if ((inode->i_state & flags) != flags) {
945 const int was_dirty = inode->i_state & I_DIRTY;
946
947 inode->i_state |= flags;
948
949 /*
950 * If the inode is being synced, just update its dirty state.
951 * The unlocker will place the inode on the appropriate
952 * superblock list, based upon its state.
953 */
954 if (inode->i_state & I_SYNC)
955 goto out;
956
957 /*
958 * Only add valid (hashed) inodes to the superblock's
959 * dirty list. Add blockdev inodes as well.
960 */
961 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382c2010-10-23 15:19:20 -0400962 if (inode_unhashed(inode))
Jens Axboe03ba3782009-09-09 09:08:54 +0200963 goto out;
964 }
Al Viroa4ffdde2010-06-02 17:38:30 -0400965 if (inode->i_state & I_FREEING)
Jens Axboe03ba3782009-09-09 09:08:54 +0200966 goto out;
967
968 /*
969 * If the inode was already on b_dirty/b_io/b_more_io, don't
970 * reposition it (that would break b_dirty time-ordering).
971 */
972 if (!was_dirty) {
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300973 bdi = inode_to_bdi(inode);
Jens Axboe500b0672009-09-09 09:10:25 +0200974
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300975 if (bdi_cap_writeback_dirty(bdi)) {
976 WARN(!test_bit(BDI_registered, &bdi->state),
977 "bdi-%s not registered\n", bdi->name);
978
979 /*
980 * If this is the first dirty inode for this
981 * bdi, we have to wake-up the corresponding
982 * bdi thread to make sure background
983 * write-back happens later.
984 */
985 if (!wb_has_dirty_io(&bdi->wb))
986 wakeup_bdi = true;
Jens Axboe500b0672009-09-09 09:10:25 +0200987 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200988
989 inode->dirtied_when = jiffies;
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300990 list_move(&inode->i_list, &bdi->wb.b_dirty);
Jens Axboe03ba3782009-09-09 09:08:54 +0200991 }
992 }
993out:
994 spin_unlock(&inode_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300995
996 if (wakeup_bdi)
Artem Bityutskiy64677162010-07-25 14:29:22 +0300997 bdi_wakeup_thread_delayed(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +0200998}
999EXPORT_SYMBOL(__mark_inode_dirty);
1000
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001001/*
1002 * Write out a superblock's list of dirty inodes. A wait will be performed
1003 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1004 *
1005 * If older_than_this is non-NULL, then only write out inodes which
1006 * had their first dirtying at a time earlier than *older_than_this.
1007 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001008 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1009 * This function assumes that the blockdev superblock's inodes are backed by
1010 * a variety of queues, so all inodes are searched. For other superblocks,
1011 * assume that all inodes are backed by the same queue.
1012 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001013 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1014 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1015 * on the writer throttling path, and we get decent balancing between many
1016 * throttled threads: we don't want them all piling up on inode_sync_wait.
1017 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001018static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001019{
Jens Axboe03ba3782009-09-09 09:08:54 +02001020 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001021
Jens Axboe03ba3782009-09-09 09:08:54 +02001022 /*
1023 * We need to be protected against the filesystem going from
1024 * r/o to r/w or vice versa.
1025 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001026 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001027
Jens Axboe03ba3782009-09-09 09:08:54 +02001028 spin_lock(&inode_lock);
1029
1030 /*
1031 * Data integrity sync. Must wait for all pages under writeback,
1032 * because there may have been pages dirtied before our sync
1033 * call, but which had writeout started before we write it out.
1034 * In which case, the inode may not be on the dirty list, but
1035 * we still have to wait for that writeout.
1036 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001037 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001038 struct address_space *mapping;
1039
Al Viroa4ffdde2010-06-02 17:38:30 -04001040 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW))
Jens Axboe03ba3782009-09-09 09:08:54 +02001041 continue;
1042 mapping = inode->i_mapping;
1043 if (mapping->nrpages == 0)
1044 continue;
1045 __iget(inode);
1046 spin_unlock(&inode_lock);
1047 /*
1048 * We hold a reference to 'inode' so it couldn't have
1049 * been removed from s_inodes list while we dropped the
1050 * inode_lock. We cannot iput the inode now as we can
1051 * be holding the last reference and we cannot iput it
1052 * under inode_lock. So we keep the reference and iput
1053 * it later.
1054 */
1055 iput(old_inode);
1056 old_inode = inode;
1057
1058 filemap_fdatawait(mapping);
1059
1060 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001061
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001062 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001063 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001064 spin_unlock(&inode_lock);
1065 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066}
1067
Jens Axboed8a85592009-09-02 12:34:32 +02001068/**
1069 * writeback_inodes_sb - writeback dirty inodes from given super_block
1070 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 *
Jens Axboed8a85592009-09-02 12:34:32 +02001072 * Start writeback on some inodes on this super_block. No guarantees are made
1073 * on how many (if any) will be written, and this function does not wait
1074 * for IO completion of submitted IO. The number of pages submitted is
1075 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001077void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078{
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001079 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1080 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001081 DECLARE_COMPLETION_ONSTACK(done);
1082 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001083 .sb = sb,
1084 .sync_mode = WB_SYNC_NONE,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001085 .done = &done,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001086 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001087
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001088 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1089
Dave Chinnercffbc8a2010-10-23 05:03:02 -04001090 work.nr_pages = nr_dirty + nr_unstable + get_nr_dirty_inodes();
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001091
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001092 bdi_queue_work(sb->s_bdi, &work);
1093 wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094}
Jens Axboed8a85592009-09-02 12:34:32 +02001095EXPORT_SYMBOL(writeback_inodes_sb);
1096
1097/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001098 * writeback_inodes_sb_if_idle - start writeback if none underway
1099 * @sb: the superblock
1100 *
1101 * Invoke writeback_inodes_sb if no writeback is currently underway.
1102 * Returns 1 if writeback was started, 0 if not.
1103 */
1104int writeback_inodes_sb_if_idle(struct super_block *sb)
1105{
1106 if (!writeback_in_progress(sb->s_bdi)) {
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001107 down_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001108 writeback_inodes_sb(sb);
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001109 up_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001110 return 1;
1111 } else
1112 return 0;
1113}
1114EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1115
1116/**
Jens Axboed8a85592009-09-02 12:34:32 +02001117 * sync_inodes_sb - sync sb inode pages
1118 * @sb: the superblock
1119 *
1120 * This function writes and waits on any dirty inode belonging to this
1121 * super_block. The number of pages synced is returned.
1122 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001123void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001124{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001125 DECLARE_COMPLETION_ONSTACK(done);
1126 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001127 .sb = sb,
1128 .sync_mode = WB_SYNC_ALL,
1129 .nr_pages = LONG_MAX,
1130 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001131 .done = &done,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001132 };
1133
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001134 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1135
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001136 bdi_queue_work(sb->s_bdi, &work);
1137 wait_for_completion(&done);
1138
Jens Axboeb6e51312009-09-16 15:13:54 +02001139 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001140}
1141EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001144 * write_inode_now - write an inode to disk
1145 * @inode: inode to write to disk
1146 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001148 * This function commits an inode to disk immediately if it is dirty. This is
1149 * primarily needed by knfsd.
1150 *
1151 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153int write_inode_now(struct inode *inode, int sync)
1154{
1155 int ret;
1156 struct writeback_control wbc = {
1157 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001158 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001159 .range_start = 0,
1160 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 };
1162
1163 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001164 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
1166 might_sleep();
1167 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001168 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 spin_unlock(&inode_lock);
1170 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001171 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 return ret;
1173}
1174EXPORT_SYMBOL(write_inode_now);
1175
1176/**
1177 * sync_inode - write an inode and its pages to disk.
1178 * @inode: the inode to sync
1179 * @wbc: controls the writeback mode
1180 *
1181 * sync_inode() will write an inode and its pages to disk. It will also
1182 * correctly update the inode on its superblock's dirty inode lists and will
1183 * update inode->i_state.
1184 *
1185 * The caller must have a ref on the inode.
1186 */
1187int sync_inode(struct inode *inode, struct writeback_control *wbc)
1188{
1189 int ret;
1190
1191 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001192 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 spin_unlock(&inode_lock);
1194 return ret;
1195}
1196EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc3765012010-10-06 10:48:20 +02001197
1198/**
1199 * sync_inode - write an inode to disk
1200 * @inode: the inode to sync
1201 * @wait: wait for I/O to complete.
1202 *
1203 * Write an inode to disk and adjust it's dirty state after completion.
1204 *
1205 * Note: only writes the actual inode, no associated data or other metadata.
1206 */
1207int sync_inode_metadata(struct inode *inode, int wait)
1208{
1209 struct writeback_control wbc = {
1210 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1211 .nr_to_write = 0, /* metadata-only */
1212 };
1213
1214 return sync_inode(inode, &wbc);
1215}
1216EXPORT_SYMBOL(sync_inode_metadata);