blob: 9f5cab75c15770350f897720194cfb85f3956194 [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
55#define inode_to_bdi(inode) ((inode)->i_mapping->backing_dev_info)
56
57/*
58 * We don't actually have pdflush, but this one is exported though /proc...
59 */
60int nr_pdflush_threads;
61
Adrian Bunkf11b00f2008-04-29 00:58:56 -070062/**
63 * writeback_in_progress - determine whether there is writeback in progress
64 * @bdi: the device's backing_dev_info structure.
65 *
Jens Axboe03ba3782009-09-09 09:08:54 +020066 * Determine whether there is writeback waiting to be handled against a
67 * backing device.
Adrian Bunkf11b00f2008-04-29 00:58:56 -070068 */
69int writeback_in_progress(struct backing_dev_info *bdi)
70{
Jens Axboe03ba3782009-09-09 09:08:54 +020071 return !list_empty(&bdi->work_list);
Adrian Bunkf11b00f2008-04-29 00:58:56 -070072}
73
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020074static void bdi_queue_work(struct backing_dev_info *bdi,
75 struct wb_writeback_work *work)
Nick Piggin4195f732009-05-28 09:01:15 +020076{
Dave Chinner455b2862010-07-07 13:24:06 +100077 trace_writeback_queue(bdi, work);
78
Jens Axboebcddc3f2009-09-13 20:07:36 +020079 spin_lock(&bdi->wb_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020080 list_add_tail(&work->list, &bdi->work_list);
Jens Axboebcddc3f2009-09-13 20:07:36 +020081 spin_unlock(&bdi->wb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070082
83 /*
Jens Axboe03ba3782009-09-09 09:08:54 +020084 * If the default thread isn't there, make sure we add it. When
85 * it gets created and wakes up, we'll run this work.
Linus Torvalds1da177e2005-04-16 15:20:36 -070086 */
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020087 if (unlikely(!bdi->wb.task)) {
Dave Chinner455b2862010-07-07 13:24:06 +100088 trace_writeback_nothread(bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +020089 wake_up_process(default_backing_dev_info.wb.task);
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020090 } else {
Jens Axboe03ba3782009-09-09 09:08:54 +020091 struct bdi_writeback *wb = &bdi->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
Nick Piggin1ef7d9a2009-09-15 21:37:55 +020093 if (wb->task)
Jens Axboe03ba3782009-09-09 09:08:54 +020094 wake_up_process(wb->task);
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070096}
97
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020098static void
99__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
100 bool range_cyclic, bool for_background)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200102 struct wb_writeback_work *work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200103
Jens Axboebcddc3f2009-09-13 20:07:36 +0200104 /*
105 * This is WB_SYNC_NONE writeback, so if allocation fails just
106 * wakeup the thread for old dirty data writeback
107 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200108 work = kzalloc(sizeof(*work), GFP_ATOMIC);
109 if (!work) {
Dave Chinner455b2862010-07-07 13:24:06 +1000110 if (bdi->wb.task) {
111 trace_writeback_nowork(bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200112 wake_up_process(bdi->wb.task);
Dave Chinner455b2862010-07-07 13:24:06 +1000113 }
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200114 return;
Jens Axboebcddc3f2009-09-13 20:07:36 +0200115 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200116
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200117 work->sync_mode = WB_SYNC_NONE;
118 work->nr_pages = nr_pages;
119 work->range_cyclic = range_cyclic;
120 work->for_background = for_background;
Christoph Hellwigf0fad8a2009-09-11 09:47:56 +0200121
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200122 bdi_queue_work(bdi, work);
Jens Axboeb6e51312009-09-16 15:13:54 +0200123}
124
125/**
126 * bdi_start_writeback - start writeback
127 * @bdi: the backing device to write from
128 * @nr_pages: the number of pages to write
129 *
130 * Description:
131 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
132 * started when this function returns, we make no guarentees on
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200133 * completion. Caller need not hold sb s_umount semaphore.
Jens Axboeb6e51312009-09-16 15:13:54 +0200134 *
135 */
Christoph Hellwigc5444192010-06-08 18:15:15 +0200136void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
Jens Axboeb6e51312009-09-16 15:13:54 +0200137{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200138 __bdi_start_writeback(bdi, nr_pages, true, false);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200139}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800140
Christoph Hellwigc5444192010-06-08 18:15:15 +0200141/**
142 * bdi_start_background_writeback - start background writeback
143 * @bdi: the backing device to write from
144 *
145 * Description:
146 * This does WB_SYNC_NONE background writeback. The IO is only
147 * started when this function returns, we make no guarentees on
148 * completion. Caller need not hold sb s_umount semaphore.
149 */
150void bdi_start_background_writeback(struct backing_dev_info *bdi)
151{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200152 __bdi_start_writeback(bdi, LONG_MAX, true, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153}
154
155/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700156 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
157 * furthest end of its superblock's dirty-inode list.
158 *
159 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200160 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700161 * the case then the inode must have been redirtied while it was being written
162 * out and we don't reset its dirtied_when.
163 */
164static void redirty_tail(struct inode *inode)
165{
Jens Axboe03ba3782009-09-09 09:08:54 +0200166 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700167
Jens Axboe03ba3782009-09-09 09:08:54 +0200168 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200169 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700170
Jens Axboe03ba3782009-09-09 09:08:54 +0200171 tail = list_entry(wb->b_dirty.next, struct inode, i_list);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200172 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700173 inode->dirtied_when = jiffies;
174 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200175 list_move(&inode->i_list, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700176}
177
178/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200179 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700180 */
Ken Chen0e0f4fc2007-10-16 23:30:38 -0700181static void requeue_io(struct inode *inode)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700182{
Jens Axboe03ba3782009-09-09 09:08:54 +0200183 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
184
185 list_move(&inode->i_list, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700186}
187
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700188static void inode_sync_complete(struct inode *inode)
189{
190 /*
191 * Prevent speculative execution through spin_unlock(&inode_lock);
192 */
193 smp_mb();
194 wake_up_bit(&inode->i_state, __I_SYNC);
195}
196
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700197static bool inode_dirtied_after(struct inode *inode, unsigned long t)
198{
199 bool ret = time_after(inode->dirtied_when, t);
200#ifndef CONFIG_64BIT
201 /*
202 * For inodes being constantly redirtied, dirtied_when can get stuck.
203 * It _appears_ to be in the future, but is actually in distant past.
204 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200205 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700206 */
207 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
208#endif
209 return ret;
210}
211
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700212/*
Fengguang Wu2c136572007-10-16 23:30:39 -0700213 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
214 */
215static void move_expired_inodes(struct list_head *delaying_queue,
216 struct list_head *dispatch_queue,
217 unsigned long *older_than_this)
218{
Shaohua Li5c034492009-09-24 14:42:33 +0200219 LIST_HEAD(tmp);
220 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +0200221 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +0200222 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +0200223 int do_sb_sort = 0;
Shaohua Li5c034492009-09-24 14:42:33 +0200224
Fengguang Wu2c136572007-10-16 23:30:39 -0700225 while (!list_empty(delaying_queue)) {
Shaohua Li5c034492009-09-24 14:42:33 +0200226 inode = list_entry(delaying_queue->prev, struct inode, i_list);
Fengguang Wu2c136572007-10-16 23:30:39 -0700227 if (older_than_this &&
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700228 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -0700229 break;
Jens Axboecf137302009-09-24 15:12:57 +0200230 if (sb && sb != inode->i_sb)
231 do_sb_sort = 1;
232 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +0200233 list_move(&inode->i_list, &tmp);
234 }
235
Jens Axboecf137302009-09-24 15:12:57 +0200236 /* just one sb in list, splice to dispatch_queue and we're done */
237 if (!do_sb_sort) {
238 list_splice(&tmp, dispatch_queue);
239 return;
240 }
241
Shaohua Li5c034492009-09-24 14:42:33 +0200242 /* Move inodes from one superblock together */
243 while (!list_empty(&tmp)) {
244 inode = list_entry(tmp.prev, struct inode, i_list);
245 sb = inode->i_sb;
246 list_for_each_prev_safe(pos, node, &tmp) {
247 inode = list_entry(pos, struct inode, i_list);
248 if (inode->i_sb == sb)
249 list_move(&inode->i_list, dispatch_queue);
250 }
Fengguang Wu2c136572007-10-16 23:30:39 -0700251 }
252}
253
254/*
255 * Queue all expired dirty inodes for io, eldest first.
256 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200257static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
Fengguang Wu2c136572007-10-16 23:30:39 -0700258{
Jens Axboe03ba3782009-09-09 09:08:54 +0200259 list_splice_init(&wb->b_more_io, wb->b_io.prev);
260 move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200261}
262
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100263static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200264{
Jens Axboe03ba3782009-09-09 09:08:54 +0200265 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100266 return inode->i_sb->s_op->write_inode(inode, wbc);
Jens Axboe03ba3782009-09-09 09:08:54 +0200267 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -0700268}
269
270/*
Christoph Hellwig01c03192009-06-08 13:35:40 +0200271 * Wait for writeback on an inode to complete.
272 */
273static void inode_wait_for_writeback(struct inode *inode)
274{
275 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
276 wait_queue_head_t *wqh;
277
278 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Richard Kennedy58a9d3d2010-05-24 14:32:38 -0700279 while (inode->i_state & I_SYNC) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200280 spin_unlock(&inode_lock);
281 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
282 spin_lock(&inode_lock);
Richard Kennedy58a9d3d2010-05-24 14:32:38 -0700283 }
Christoph Hellwig01c03192009-06-08 13:35:40 +0200284}
285
286/*
287 * Write out an inode's dirty pages. Called under inode_lock. Either the
288 * caller has ref on the inode (either via __iget or via syscall against an fd)
289 * or the inode has I_WILL_FREE set (via generic_forget_inode)
290 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 * If `wait' is set, wait on the writeout.
292 *
293 * The whole writeout design is quite complex and fragile. We want to avoid
294 * starvation of particular inodes when others are being redirtied, prevent
295 * livelocks, etc.
296 *
297 * Called under inode_lock.
298 */
299static int
Christoph Hellwig01c03192009-06-08 13:35:40 +0200300writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 struct address_space *mapping = inode->i_mapping;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200303 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 int ret;
305
Christoph Hellwig01c03192009-06-08 13:35:40 +0200306 if (!atomic_read(&inode->i_count))
307 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
308 else
309 WARN_ON(inode->i_state & I_WILL_FREE);
310
311 if (inode->i_state & I_SYNC) {
312 /*
313 * If this inode is locked for writeback and we are not doing
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200314 * writeback-for-data-integrity, move it to b_more_io so that
Christoph Hellwig01c03192009-06-08 13:35:40 +0200315 * writeback can proceed with the other inodes on s_io.
316 *
317 * We'll have another go at writing back this inode when we
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200318 * completed a full scan of b_io.
Christoph Hellwig01c03192009-06-08 13:35:40 +0200319 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100320 if (wbc->sync_mode != WB_SYNC_ALL) {
Christoph Hellwig01c03192009-06-08 13:35:40 +0200321 requeue_io(inode);
322 return 0;
323 }
324
325 /*
326 * It's a data-integrity sync. We must wait.
327 */
328 inode_wait_for_writeback(inode);
329 }
330
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700331 BUG_ON(inode->i_state & I_SYNC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400333 /* Set I_SYNC, reset I_DIRTY_PAGES */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700334 inode->i_state |= I_SYNC;
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400335 inode->i_state &= ~I_DIRTY_PAGES;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 spin_unlock(&inode_lock);
337
338 ret = do_writepages(mapping, wbc);
339
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100340 /*
341 * Make sure to wait on the data before writing out the metadata.
342 * This is important for filesystems that modify metadata on data
343 * I/O completion.
344 */
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100345 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100346 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 if (ret == 0)
348 ret = err;
349 }
350
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +0400351 /*
352 * Some filesystems may redirty the inode during the writeback
353 * due to delalloc, clear dirty metadata flags right before
354 * write_inode()
355 */
356 spin_lock(&inode_lock);
357 dirty = inode->i_state & I_DIRTY;
358 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
359 spin_unlock(&inode_lock);
Christoph Hellwig26821ed2010-03-05 09:21:21 +0100360 /* Don't write the inode if only I_DIRTY_PAGES was set */
361 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +0100362 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 if (ret == 0)
364 ret = err;
365 }
366
367 spin_lock(&inode_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700368 inode->i_state &= ~I_SYNC;
Wu Fengguang84a89242009-06-16 15:33:17 -0700369 if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
Wu Fengguangb3af9462009-09-25 06:04:10 +0200370 if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800371 /*
Wu Fengguangb3af9462009-09-25 06:04:10 +0200372 * More pages get dirtied by a fast dirtier.
373 */
374 goto select_queue;
375 } else if (inode->i_state & I_DIRTY) {
376 /*
377 * At least XFS will redirty the inode during the
378 * writeback (delalloc) and on io completion (isize).
Wu Fengguangae1b7f72009-09-23 20:33:42 +0800379 */
380 redirty_tail(inode);
381 } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 /*
383 * We didn't write back all the pages. nfs_writepages()
384 * sometimes bales out without doing anything. Redirty
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200385 * the inode; Move it from b_io onto b_more_io/b_dirty.
Andrew Morton1b43ef92007-10-16 23:30:35 -0700386 */
387 /*
388 * akpm: if the caller was the kupdate function we put
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200389 * this inode at the head of b_dirty so it gets first
Andrew Morton1b43ef92007-10-16 23:30:35 -0700390 * consideration. Otherwise, move it to the tail, for
391 * the reasons described there. I'm not really sure
392 * how much sense this makes. Presumably I had a good
393 * reasons for doing it this way, and I'd rather not
394 * muck with it at present.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 */
396 if (wbc->for_kupdate) {
397 /*
Fengguang Wu2c136572007-10-16 23:30:39 -0700398 * For the kupdate function we move the inode
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200399 * to b_more_io so it will get more writeout as
Fengguang Wu2c136572007-10-16 23:30:39 -0700400 * soon as the queue becomes uncongested.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 */
402 inode->i_state |= I_DIRTY_PAGES;
Wu Fengguangb3af9462009-09-25 06:04:10 +0200403select_queue:
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800404 if (wbc->nr_to_write <= 0) {
405 /*
406 * slice used up: queue for next turn
407 */
408 requeue_io(inode);
409 } else {
410 /*
411 * somehow blocked: retry later
412 */
413 redirty_tail(inode);
414 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 } else {
416 /*
417 * Otherwise fully redirty the inode so that
418 * other inodes on this superblock will get some
419 * writeout. Otherwise heavy writing to one
420 * file would indefinitely suspend writeout of
421 * all the other files.
422 */
423 inode->i_state |= I_DIRTY_PAGES;
Andrew Morton1b43ef92007-10-16 23:30:35 -0700424 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 } else if (atomic_read(&inode->i_count)) {
427 /*
428 * The inode is clean, inuse
429 */
430 list_move(&inode->i_list, &inode_in_use);
431 } else {
432 /*
433 * The inode is clean, unused
434 */
435 list_move(&inode->i_list, &inode_unused);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 }
437 }
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700438 inode_sync_complete(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 return ret;
440}
441
Jens Axboe03ba3782009-09-09 09:08:54 +0200442/*
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200443 * For background writeback the caller does not have the sb pinned
Jens Axboe03ba3782009-09-09 09:08:54 +0200444 * before calling writeback. So make sure that we do pin it, so it doesn't
445 * go away while we are writing inodes from it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200446 */
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200447static bool pin_sb_for_writeback(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448{
Jens Axboe03ba3782009-09-09 09:08:54 +0200449 spin_lock(&sb_lock);
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200450 if (list_empty(&sb->s_instances)) {
451 spin_unlock(&sb_lock);
452 return false;
453 }
454
Jens Axboe03ba3782009-09-09 09:08:54 +0200455 sb->s_count++;
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200456 spin_unlock(&sb_lock);
457
Jens Axboe03ba3782009-09-09 09:08:54 +0200458 if (down_read_trylock(&sb->s_umount)) {
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200459 if (sb->s_root)
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200460 return true;
Jens Axboe03ba3782009-09-09 09:08:54 +0200461 up_read(&sb->s_umount);
462 }
Christoph Hellwig29cb4852010-06-09 15:31:01 +0200463
464 put_super(sb);
Christoph Hellwigd19de7e2010-06-08 18:14:58 +0200465 return false;
Jens Axboe03ba3782009-09-09 09:08:54 +0200466}
467
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800468/*
469 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200470 *
471 * If @only_this_sb is true, then find and write all such
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800472 * inodes. Otherwise write only ones which go sequentially
473 * in reverse order.
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200474 *
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800475 * Return 1, if the caller writeback routine should be
476 * interrupted. Otherwise return 0.
477 */
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200478static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
479 struct writeback_control *wbc, bool only_this_sb)
Jens Axboe03ba3782009-09-09 09:08:54 +0200480{
Jens Axboe03ba3782009-09-09 09:08:54 +0200481 while (!list_empty(&wb->b_io)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 long pages_skipped;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800483 struct inode *inode = list_entry(wb->b_io.prev,
484 struct inode, i_list);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200485
486 if (inode->i_sb != sb) {
487 if (only_this_sb) {
488 /*
489 * We only want to write back data for this
490 * superblock, move all inodes not belonging
491 * to it back onto the dirty list.
492 */
493 redirty_tail(inode);
494 continue;
495 }
496
497 /*
498 * The inode belongs to a different superblock.
499 * Bounce back to the caller to unpin this and
500 * pin the next superblock.
501 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800502 return 0;
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200503 }
504
Wu Fengguang84a89242009-06-16 15:33:17 -0700505 if (inode->i_state & (I_NEW | I_WILL_FREE)) {
Nick Piggin7ef0d732009-03-12 14:31:38 -0700506 requeue_io(inode);
507 continue;
508 }
Jeff Laytond2caa3c2009-04-02 16:56:37 -0700509 /*
510 * Was this inode dirtied after sync_sb_inodes was called?
511 * This keeps sync from extra jobs and livelock.
512 */
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800513 if (inode_dirtied_after(inode, wbc->wb_start))
514 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Wu Fengguang84a89242009-06-16 15:33:17 -0700516 BUG_ON(inode->i_state & (I_FREEING | I_CLEAR));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 __iget(inode);
518 pages_skipped = wbc->pages_skipped;
Christoph Hellwig01c03192009-06-08 13:35:40 +0200519 writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 if (wbc->pages_skipped != pages_skipped) {
521 /*
522 * writeback is not making progress due to locked
523 * buffers. Skip this inode for now.
524 */
Andrew Mortonf57b9b72007-10-16 23:30:34 -0700525 redirty_tail(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 }
527 spin_unlock(&inode_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 iput(inode);
OGAWA Hirofumi4ffc8442006-03-25 03:07:44 -0800529 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 spin_lock(&inode_lock);
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800531 if (wbc->nr_to_write <= 0) {
532 wbc->more_io = 1;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800533 return 1;
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800534 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200535 if (!list_empty(&wb->b_more_io))
Fengguang Wu8bc3be22008-02-04 22:29:36 -0800536 wbc->more_io = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800538 /* b_io is empty */
539 return 1;
540}
Nick Piggin38f21972009-01-06 14:40:25 -0800541
Christoph Hellwig9c3a8ee2010-06-10 12:07:27 +0200542void writeback_inodes_wb(struct bdi_writeback *wb,
543 struct writeback_control *wbc)
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800544{
545 int ret = 0;
Jens Axboe9ecc2732009-09-24 15:25:11 +0200546
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800547 wbc->wb_start = jiffies; /* livelock avoidance */
548 spin_lock(&inode_lock);
549 if (!wbc->for_kupdate || list_empty(&wb->b_io))
550 queue_io(wb, wbc->older_than_this);
551
552 while (!list_empty(&wb->b_io)) {
553 struct inode *inode = list_entry(wb->b_io.prev,
554 struct inode, i_list);
555 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800556
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200557 if (!pin_sb_for_writeback(sb)) {
558 requeue_io(inode);
559 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800560 }
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200561 ret = writeback_sb_inodes(sb, wb, wbc, false);
562 drop_super(sb);
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800563
Edward Shishkinf11c9c52010-03-11 14:09:47 -0800564 if (ret)
565 break;
566 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200567 spin_unlock(&inode_lock);
568 /* Leave any unwritten inodes on b_io */
569}
570
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200571static void __writeback_inodes_sb(struct super_block *sb,
572 struct bdi_writeback *wb, struct writeback_control *wbc)
573{
574 WARN_ON(!rwsem_is_locked(&sb->s_umount));
575
576 wbc->wb_start = jiffies; /* livelock avoidance */
577 spin_lock(&inode_lock);
578 if (!wbc->for_kupdate || list_empty(&wb->b_io))
579 queue_io(wb, wbc->older_than_this);
580 writeback_sb_inodes(sb, wb, wbc, true);
581 spin_unlock(&inode_lock);
582}
583
Jens Axboe03ba3782009-09-09 09:08:54 +0200584/*
585 * The maximum number of pages to writeout in a single bdi flush/kupdate
586 * operation. We do this so we don't hold I_SYNC against an inode for
587 * enormous amounts of time, which would block a userspace task which has
588 * been forced to throttle against that inode. Also, the code reevaluates
589 * the dirty each time it has written this many pages.
590 */
591#define MAX_WRITEBACK_PAGES 1024
592
593static inline bool over_bground_thresh(void)
594{
595 unsigned long background_thresh, dirty_thresh;
596
597 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
598
599 return (global_page_state(NR_FILE_DIRTY) +
600 global_page_state(NR_UNSTABLE_NFS) >= background_thresh);
601}
602
603/*
604 * Explicit flushing or periodic writeback of "old" data.
605 *
606 * Define "old": the first time one of an inode's pages is dirtied, we mark the
607 * dirtying-time in the inode's address_space. So this periodic writeback code
608 * just walks the superblock inode list, writing back any inodes which are
609 * older than a specific point in time.
610 *
611 * Try to run once per dirty_writeback_interval. But if a writeback event
612 * takes longer than a dirty_writeback_interval interval, then leave a
613 * one-second gap.
614 *
615 * older_than_this takes precedence over nr_to_write. So we'll only write back
616 * all dirty pages if they are all attached to "old" mappings.
617 */
Jens Axboec4a77a62009-09-16 15:18:25 +0200618static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200619 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +0200620{
621 struct writeback_control wbc = {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200622 .sync_mode = work->sync_mode,
Jens Axboe03ba3782009-09-09 09:08:54 +0200623 .older_than_this = NULL,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200624 .for_kupdate = work->for_kupdate,
625 .for_background = work->for_background,
626 .range_cyclic = work->range_cyclic,
Jens Axboe03ba3782009-09-09 09:08:54 +0200627 };
628 unsigned long oldest_jif;
629 long wrote = 0;
Jan Karaa5989bd2009-09-16 19:22:48 +0200630 struct inode *inode;
Jens Axboe03ba3782009-09-09 09:08:54 +0200631
632 if (wbc.for_kupdate) {
633 wbc.older_than_this = &oldest_jif;
634 oldest_jif = jiffies -
635 msecs_to_jiffies(dirty_expire_interval * 10);
636 }
Jens Axboec4a77a62009-09-16 15:18:25 +0200637 if (!wbc.range_cyclic) {
638 wbc.range_start = 0;
639 wbc.range_end = LLONG_MAX;
640 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200641
642 for (;;) {
643 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800644 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +0200645 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200646 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +0200647 break;
648
649 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800650 * For background writeout, stop when we are below the
651 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +0200652 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200653 if (work->for_background && !over_bground_thresh())
Jens Axboe03ba3782009-09-09 09:08:54 +0200654 break;
655
656 wbc.more_io = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200657 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
658 wbc.pages_skipped = 0;
Dave Chinner028c2dd2010-07-07 13:24:07 +1000659
660 trace_wbc_writeback_start(&wbc, wb->bdi);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200661 if (work->sb)
662 __writeback_inodes_sb(work->sb, wb, &wbc);
Christoph Hellwigedadfb12010-06-10 12:07:54 +0200663 else
664 writeback_inodes_wb(wb, &wbc);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000665 trace_wbc_writeback_written(&wbc, wb->bdi);
666
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200667 work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
Jens Axboe03ba3782009-09-09 09:08:54 +0200668 wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
669
670 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +0200671 * If we consumed everything, see if we have more
Jens Axboe03ba3782009-09-09 09:08:54 +0200672 */
Jens Axboe71fd05a2009-09-23 19:32:26 +0200673 if (wbc.nr_to_write <= 0)
674 continue;
675 /*
676 * Didn't write everything and we don't have more IO, bail
677 */
678 if (!wbc.more_io)
Jens Axboe03ba3782009-09-09 09:08:54 +0200679 break;
Jens Axboe71fd05a2009-09-23 19:32:26 +0200680 /*
681 * Did we write something? Try for more
682 */
683 if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
684 continue;
685 /*
686 * Nothing written. Wait for some inode to
687 * become available for writeback. Otherwise
688 * we'll just busyloop.
689 */
690 spin_lock(&inode_lock);
691 if (!list_empty(&wb->b_more_io)) {
692 inode = list_entry(wb->b_more_io.prev,
693 struct inode, i_list);
Dave Chinner028c2dd2010-07-07 13:24:07 +1000694 trace_wbc_writeback_wait(&wbc, wb->bdi);
Jens Axboe71fd05a2009-09-23 19:32:26 +0200695 inode_wait_for_writeback(inode);
Jens Axboe03ba3782009-09-09 09:08:54 +0200696 }
Jens Axboe71fd05a2009-09-23 19:32:26 +0200697 spin_unlock(&inode_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200698 }
699
700 return wrote;
701}
702
703/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200704 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +0200705 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200706static struct wb_writeback_work *
Minchan Kim08852b62010-08-03 12:51:16 +0200707get_next_work_item(struct backing_dev_info *bdi)
Jens Axboe03ba3782009-09-09 09:08:54 +0200708{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200709 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +0200710
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200711 spin_lock(&bdi->wb_lock);
712 if (!list_empty(&bdi->work_list)) {
713 work = list_entry(bdi->work_list.next,
714 struct wb_writeback_work, list);
715 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200716 }
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200717 spin_unlock(&bdi->wb_lock);
718 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +0200719}
720
721static 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;
738 nr_pages = global_page_state(NR_FILE_DIRTY) +
739 global_page_state(NR_UNSTABLE_NFS) +
740 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
741
Jens Axboec4a77a62009-09-16 15:18:25 +0200742 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200743 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +0200744 .nr_pages = nr_pages,
745 .sync_mode = WB_SYNC_NONE,
746 .for_kupdate = 1,
747 .range_cyclic = 1,
748 };
749
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200750 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +0200751 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200752
753 return 0;
754}
755
756/*
757 * Retrieve work items and do the writeback they describe
758 */
759long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
760{
761 struct backing_dev_info *bdi = wb->bdi;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200762 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +0200763 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +0200764
Minchan Kim08852b62010-08-03 12:51:16 +0200765 while ((work = get_next_work_item(bdi)) != NULL) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200766 /*
767 * Override sync mode, in case we must wait for completion
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200768 * because this thread is exiting now.
Jens Axboe03ba3782009-09-09 09:08:54 +0200769 */
770 if (force_wait)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200771 work->sync_mode = WB_SYNC_ALL;
772
Dave Chinner455b2862010-07-07 13:24:06 +1000773 trace_writeback_exec(bdi, work);
774
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200775 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200776
777 /*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200778 * Notify the caller of completion if this is a synchronous
779 * work item, otherwise just free it.
Jens Axboe03ba3782009-09-09 09:08:54 +0200780 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200781 if (work->done)
782 complete(work->done);
783 else
784 kfree(work);
Jens Axboe03ba3782009-09-09 09:08:54 +0200785 }
786
787 /*
788 * Check for periodic writeback, kupdated() style
789 */
790 wrote += wb_check_old_data_flush(wb);
791
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 unsigned long wait_jiffies = -1UL;
804 long pages_written;
805
Christoph Hellwig08243902010-06-19 23:08:22 +0200806 current->flags |= PF_FLUSHER | PF_SWAPWRITE;
807 set_freezable();
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300808 wb->last_active = jiffies;
Christoph Hellwig08243902010-06-19 23:08:22 +0200809
810 /*
811 * Our parent may run at a different priority, just set us to normal
812 */
813 set_user_nice(current, 0);
814
815 /*
816 * Clear pending bit and wakeup anybody waiting to tear us down
817 */
818 clear_bit(BDI_pending, &bdi->state);
819 smp_mb__after_clear_bit();
820 wake_up_bit(&bdi->state, BDI_pending);
821
Dave Chinner455b2862010-07-07 13:24:06 +1000822 trace_writeback_thread_start(bdi);
823
Jens Axboe03ba3782009-09-09 09:08:54 +0200824 while (!kthread_should_stop()) {
825 pages_written = wb_do_writeback(wb, 0);
826
Dave Chinner455b2862010-07-07 13:24:06 +1000827 trace_writeback_pages_written(pages_written);
828
Jens Axboe03ba3782009-09-09 09:08:54 +0200829 if (pages_written)
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300830 wb->last_active = jiffies;
Jens Axboe03ba3782009-09-09 09:08:54 +0200831 else if (wait_jiffies != -1UL) {
832 unsigned long max_idle;
833
834 /*
835 * Longest period of inactivity that we tolerate. If we
Artem Bityutskiy6f904ff2010-07-25 14:29:11 +0300836 * see dirty data again later, the thread will get
Jens Axboe03ba3782009-09-09 09:08:54 +0200837 * recreated automatically.
838 */
839 max_idle = max(5UL * 60 * HZ, wait_jiffies);
Artem Bityutskiyecd58402010-07-25 14:29:18 +0300840 if (time_after(jiffies, max_idle + wb->last_active))
Jens Axboe03ba3782009-09-09 09:08:54 +0200841 break;
842 }
843
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300844 set_current_state(TASK_INTERRUPTIBLE);
845 if (!list_empty(&bdi->work_list)) {
846 __set_current_state(TASK_RUNNING);
847 continue;
848 }
849
Jens Axboe69b62d02010-05-17 12:51:03 +0200850 if (dirty_writeback_interval) {
851 wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
Artem Bityutskiy297252c2010-07-25 14:29:15 +0300852 schedule_timeout(wait_jiffies);
853 } else
854 schedule();
Jens Axboe69b62d02010-05-17 12:51:03 +0200855
Jens Axboe03ba3782009-09-09 09:08:54 +0200856 try_to_freeze();
857 }
858
Christoph Hellwig08243902010-06-19 23:08:22 +0200859 wb->task = NULL;
860
861 /*
862 * Flush any work that raced with us exiting. No new work
863 * will be added, since this bdi isn't discoverable anymore.
864 */
865 if (!list_empty(&bdi->work_list))
866 wb_do_writeback(wb, 1);
Dave Chinner455b2862010-07-07 13:24:06 +1000867
868 trace_writeback_thread_stop(bdi);
Jens Axboe03ba3782009-09-09 09:08:54 +0200869 return 0;
870}
871
Christoph Hellwig08243902010-06-19 23:08:22 +0200872
Jens Axboe03ba3782009-09-09 09:08:54 +0200873/*
Jens Axboe03ba3782009-09-09 09:08:54 +0200874 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
875 * the whole world.
876 */
877void wakeup_flusher_threads(long nr_pages)
878{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200879 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200880
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200881 if (!nr_pages) {
882 nr_pages = global_page_state(NR_FILE_DIRTY) +
Jens Axboe03ba3782009-09-09 09:08:54 +0200883 global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200884 }
885
886 rcu_read_lock();
887 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
888 if (!bdi_has_dirty_io(bdi))
889 continue;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +0200890 __bdi_start_writeback(bdi, nr_pages, false, false);
Christoph Hellwigb8c2f342010-06-08 18:15:07 +0200891 }
892 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +0200893}
894
895static noinline void block_dump___mark_inode_dirty(struct inode *inode)
896{
897 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
898 struct dentry *dentry;
899 const char *name = "?";
900
901 dentry = d_find_alias(inode);
902 if (dentry) {
903 spin_lock(&dentry->d_lock);
904 name = (const char *) dentry->d_name.name;
905 }
906 printk(KERN_DEBUG
907 "%s(%d): dirtied inode %lu (%s) on %s\n",
908 current->comm, task_pid_nr(current), inode->i_ino,
909 name, inode->i_sb->s_id);
910 if (dentry) {
911 spin_unlock(&dentry->d_lock);
912 dput(dentry);
913 }
914 }
915}
916
917/**
918 * __mark_inode_dirty - internal function
919 * @inode: inode to mark
920 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
921 * Mark an inode as dirty. Callers should use mark_inode_dirty or
922 * mark_inode_dirty_sync.
923 *
924 * Put the inode on the super block's dirty list.
925 *
926 * CAREFUL! We mark it dirty unconditionally, but move it onto the
927 * dirty list only if it is hashed or if it refers to a blockdev.
928 * If it was not hashed, it will never be added to the dirty list
929 * even if it is later hashed, as it will have been marked dirty already.
930 *
931 * In short, make sure you hash any inodes _before_ you start marking
932 * them dirty.
933 *
934 * This function *must* be atomic for the I_DIRTY_PAGES case -
935 * set_page_dirty() is called under spinlock in several places.
936 *
937 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
938 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
939 * the kernel-internal blockdev inode represents the dirtying time of the
940 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
941 * page->mapping->host, so the page-dirtying time is recorded in the internal
942 * blockdev inode.
943 */
944void __mark_inode_dirty(struct inode *inode, int flags)
945{
946 struct super_block *sb = inode->i_sb;
947
948 /*
949 * Don't do this for I_DIRTY_PAGES - that doesn't actually
950 * dirty the inode itself
951 */
952 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
953 if (sb->s_op->dirty_inode)
954 sb->s_op->dirty_inode(inode);
955 }
956
957 /*
958 * make sure that changes are seen by all cpus before we test i_state
959 * -- mikulas
960 */
961 smp_mb();
962
963 /* avoid the locking if we can */
964 if ((inode->i_state & flags) == flags)
965 return;
966
967 if (unlikely(block_dump))
968 block_dump___mark_inode_dirty(inode);
969
970 spin_lock(&inode_lock);
971 if ((inode->i_state & flags) != flags) {
972 const int was_dirty = inode->i_state & I_DIRTY;
973
974 inode->i_state |= flags;
975
976 /*
977 * If the inode is being synced, just update its dirty state.
978 * The unlocker will place the inode on the appropriate
979 * superblock list, based upon its state.
980 */
981 if (inode->i_state & I_SYNC)
982 goto out;
983
984 /*
985 * Only add valid (hashed) inodes to the superblock's
986 * dirty list. Add blockdev inodes as well.
987 */
988 if (!S_ISBLK(inode->i_mode)) {
989 if (hlist_unhashed(&inode->i_hash))
990 goto out;
991 }
992 if (inode->i_state & (I_FREEING|I_CLEAR))
993 goto out;
994
995 /*
996 * If the inode was already on b_dirty/b_io/b_more_io, don't
997 * reposition it (that would break b_dirty time-ordering).
998 */
999 if (!was_dirty) {
1000 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Jens Axboe500b0672009-09-09 09:10:25 +02001001 struct backing_dev_info *bdi = wb->bdi;
1002
1003 if (bdi_cap_writeback_dirty(bdi) &&
1004 !test_bit(BDI_registered, &bdi->state)) {
1005 WARN_ON(1);
1006 printk(KERN_ERR "bdi-%s not registered\n",
1007 bdi->name);
1008 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001009
1010 inode->dirtied_when = jiffies;
1011 list_move(&inode->i_list, &wb->b_dirty);
1012 }
1013 }
1014out:
1015 spin_unlock(&inode_lock);
1016}
1017EXPORT_SYMBOL(__mark_inode_dirty);
1018
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001019/*
1020 * Write out a superblock's list of dirty inodes. A wait will be performed
1021 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1022 *
1023 * If older_than_this is non-NULL, then only write out inodes which
1024 * had their first dirtying at a time earlier than *older_than_this.
1025 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001026 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1027 * This function assumes that the blockdev superblock's inodes are backed by
1028 * a variety of queues, so all inodes are searched. For other superblocks,
1029 * assume that all inodes are backed by the same queue.
1030 *
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001031 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1032 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1033 * on the writer throttling path, and we get decent balancing between many
1034 * throttled threads: we don't want them all piling up on inode_sync_wait.
1035 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001036static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001037{
Jens Axboe03ba3782009-09-09 09:08:54 +02001038 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001039
Jens Axboe03ba3782009-09-09 09:08:54 +02001040 /*
1041 * We need to be protected against the filesystem going from
1042 * r/o to r/w or vice versa.
1043 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001044 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001045
Jens Axboe03ba3782009-09-09 09:08:54 +02001046 spin_lock(&inode_lock);
1047
1048 /*
1049 * Data integrity sync. Must wait for all pages under writeback,
1050 * because there may have been pages dirtied before our sync
1051 * call, but which had writeout started before we write it out.
1052 * In which case, the inode may not be on the dirty list, but
1053 * we still have to wait for that writeout.
1054 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001055 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001056 struct address_space *mapping;
1057
1058 if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
1059 continue;
1060 mapping = inode->i_mapping;
1061 if (mapping->nrpages == 0)
1062 continue;
1063 __iget(inode);
1064 spin_unlock(&inode_lock);
1065 /*
1066 * We hold a reference to 'inode' so it couldn't have
1067 * been removed from s_inodes list while we dropped the
1068 * inode_lock. We cannot iput the inode now as we can
1069 * be holding the last reference and we cannot iput it
1070 * under inode_lock. So we keep the reference and iput
1071 * it later.
1072 */
1073 iput(old_inode);
1074 old_inode = inode;
1075
1076 filemap_fdatawait(mapping);
1077
1078 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08001079
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001080 spin_lock(&inode_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001081 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001082 spin_unlock(&inode_lock);
1083 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084}
1085
Jens Axboed8a85592009-09-02 12:34:32 +02001086/**
1087 * writeback_inodes_sb - writeback dirty inodes from given super_block
1088 * @sb: the superblock
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 *
Jens Axboed8a85592009-09-02 12:34:32 +02001090 * Start writeback on some inodes on this super_block. No guarantees are made
1091 * on how many (if any) will be written, and this function does not wait
1092 * for IO completion of submitted IO. The number of pages submitted is
1093 * returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001095void writeback_inodes_sb(struct super_block *sb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096{
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001097 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
1098 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001099 DECLARE_COMPLETION_ONSTACK(done);
1100 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001101 .sb = sb,
1102 .sync_mode = WB_SYNC_NONE,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001103 .done = &done,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001104 };
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001105
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001106 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1107
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001108 work.nr_pages = nr_dirty + nr_unstable +
Jens Axboe0e3c9a22010-06-01 11:08:43 +02001109 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
1110
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001111 bdi_queue_work(sb->s_bdi, &work);
1112 wait_for_completion(&done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113}
Jens Axboed8a85592009-09-02 12:34:32 +02001114EXPORT_SYMBOL(writeback_inodes_sb);
1115
1116/**
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001117 * writeback_inodes_sb_if_idle - start writeback if none underway
1118 * @sb: the superblock
1119 *
1120 * Invoke writeback_inodes_sb if no writeback is currently underway.
1121 * Returns 1 if writeback was started, 0 if not.
1122 */
1123int writeback_inodes_sb_if_idle(struct super_block *sb)
1124{
1125 if (!writeback_in_progress(sb->s_bdi)) {
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001126 down_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001127 writeback_inodes_sb(sb);
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001128 up_read(&sb->s_umount);
Eric Sandeen17bd55d2009-12-23 07:57:07 -05001129 return 1;
1130 } else
1131 return 0;
1132}
1133EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1134
1135/**
Jens Axboed8a85592009-09-02 12:34:32 +02001136 * sync_inodes_sb - sync sb inode pages
1137 * @sb: the superblock
1138 *
1139 * This function writes and waits on any dirty inode belonging to this
1140 * super_block. The number of pages synced is returned.
1141 */
Jens Axboeb6e51312009-09-16 15:13:54 +02001142void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02001143{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001144 DECLARE_COMPLETION_ONSTACK(done);
1145 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001146 .sb = sb,
1147 .sync_mode = WB_SYNC_ALL,
1148 .nr_pages = LONG_MAX,
1149 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001150 .done = &done,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02001151 };
1152
Christoph Hellwigcf37e972010-06-08 18:14:51 +02001153 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1154
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001155 bdi_queue_work(sb->s_bdi, &work);
1156 wait_for_completion(&done);
1157
Jens Axboeb6e51312009-09-16 15:13:54 +02001158 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02001159}
1160EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001163 * write_inode_now - write an inode to disk
1164 * @inode: inode to write to disk
1165 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08001167 * This function commits an inode to disk immediately if it is dirty. This is
1168 * primarily needed by knfsd.
1169 *
1170 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172int write_inode_now(struct inode *inode, int sync)
1173{
1174 int ret;
1175 struct writeback_control wbc = {
1176 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08001177 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07001178 .range_start = 0,
1179 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 };
1181
1182 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08001183 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
1185 might_sleep();
1186 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001187 ret = writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 spin_unlock(&inode_lock);
1189 if (sync)
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001190 inode_sync_wait(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 return ret;
1192}
1193EXPORT_SYMBOL(write_inode_now);
1194
1195/**
1196 * sync_inode - write an inode and its pages to disk.
1197 * @inode: the inode to sync
1198 * @wbc: controls the writeback mode
1199 *
1200 * sync_inode() will write an inode and its pages to disk. It will also
1201 * correctly update the inode on its superblock's dirty inode lists and will
1202 * update inode->i_state.
1203 *
1204 * The caller must have a ref on the inode.
1205 */
1206int sync_inode(struct inode *inode, struct writeback_control *wbc)
1207{
1208 int ret;
1209
1210 spin_lock(&inode_lock);
Christoph Hellwig01c03192009-06-08 13:35:40 +02001211 ret = writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 spin_unlock(&inode_lock);
1213 return ret;
1214}
1215EXPORT_SYMBOL(sync_inode);