blob: 7ba8feae11b8a5e1a33f002a6efc0ca9f3a82dde [file] [log] [blame]
Andrew Morton3fcfab12006-10-19 23:28:16 -07001
2#include <linux/wait.h>
3#include <linux/backing-dev.h>
Jens Axboe03ba3782009-09-09 09:08:54 +02004#include <linux/kthread.h>
5#include <linux/freezer.h>
Andrew Morton3fcfab12006-10-19 23:28:16 -07006#include <linux/fs.h>
Jens Axboe26160152009-03-17 09:35:06 +01007#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +02008#include <linux/mm.h>
Andrew Morton3fcfab12006-10-19 23:28:16 -07009#include <linux/sched.h>
10#include <linux/module.h>
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -070011#include <linux/writeback.h>
12#include <linux/device.h>
Dave Chinner455b2862010-07-07 13:24:06 +100013#include <trace/events/writeback.h>
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -070014
Jens Axboec3c53202010-04-22 11:37:01 +020015static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
Jens Axboe26160152009-03-17 09:35:06 +010017struct backing_dev_info default_backing_dev_info = {
Jens Axboed9938312009-06-12 14:45:52 +020018 .name = "default",
Jens Axboe26160152009-03-17 09:35:06 +010019 .ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
20 .state = 0,
21 .capabilities = BDI_CAP_MAP_COPY,
Jens Axboe26160152009-03-17 09:35:06 +010022};
23EXPORT_SYMBOL_GPL(default_backing_dev_info);
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -070024
Jörn Engel5129a462010-04-25 08:54:42 +020025struct backing_dev_info noop_backing_dev_info = {
26 .name = "noop",
Jan Kara976e48f2010-09-21 11:48:55 +020027 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
Jörn Engel5129a462010-04-25 08:54:42 +020028};
29EXPORT_SYMBOL_GPL(noop_backing_dev_info);
30
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -070031static struct class *bdi_class;
Jens Axboecfc4ba52009-09-14 13:12:40 +020032
33/*
34 * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
35 * reader side protection for bdi_pending_list. bdi_list has RCU reader side
36 * locking.
37 */
Jens Axboe03ba3782009-09-09 09:08:54 +020038DEFINE_SPINLOCK(bdi_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +020039LIST_HEAD(bdi_list);
Jens Axboe03ba3782009-09-09 09:08:54 +020040LIST_HEAD(bdi_pending_list);
41
42static struct task_struct *sync_supers_tsk;
43static struct timer_list sync_supers_timer;
44
45static int bdi_sync_supers(void *);
46static void sync_supers_timer_fn(unsigned long);
Jens Axboe03ba3782009-09-09 09:08:54 +020047
Christoph Hellwigf758eea2011-04-21 18:19:44 -060048void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
49{
50 if (wb1 < wb2) {
51 spin_lock(&wb1->list_lock);
52 spin_lock_nested(&wb2->list_lock, 1);
53 } else {
54 spin_lock(&wb2->list_lock);
55 spin_lock_nested(&wb1->list_lock, 1);
56 }
57}
58
Miklos Szeredi76f14182008-04-30 00:54:36 -070059#ifdef CONFIG_DEBUG_FS
60#include <linux/debugfs.h>
61#include <linux/seq_file.h>
62
63static struct dentry *bdi_debug_root;
64
65static void bdi_debug_init(void)
66{
67 bdi_debug_root = debugfs_create_dir("bdi", NULL);
68}
69
70static int bdi_debug_stats_show(struct seq_file *m, void *v)
71{
72 struct backing_dev_info *bdi = m->private;
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020073 struct bdi_writeback *wb = &bdi->wb;
David Rientjes364aeb22009-01-06 14:39:29 -080074 unsigned long background_thresh;
75 unsigned long dirty_thresh;
76 unsigned long bdi_thresh;
Gustavo F. Padovan345227d2011-05-20 21:23:37 +020077 unsigned long nr_dirty, nr_io, nr_more_io;
Jens Axboef09b00d2009-05-25 09:08:21 +020078 struct inode *inode;
79
Gustavo F. Padovan345227d2011-05-20 21:23:37 +020080 nr_dirty = nr_io = nr_more_io = 0;
Christoph Hellwigf758eea2011-04-21 18:19:44 -060081 spin_lock(&wb->list_lock);
Nick Piggin7ccf19a2010-10-21 11:49:30 +110082 list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020083 nr_dirty++;
Nick Piggin7ccf19a2010-10-21 11:49:30 +110084 list_for_each_entry(inode, &wb->b_io, i_wb_list)
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020085 nr_io++;
Nick Piggin7ccf19a2010-10-21 11:49:30 +110086 list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
Christoph Hellwigc1955ce2010-06-19 23:08:06 +020087 nr_more_io++;
Christoph Hellwigf758eea2011-04-21 18:19:44 -060088 spin_unlock(&wb->list_lock);
Miklos Szeredi76f14182008-04-30 00:54:36 -070089
Wu Fengguang16c40422010-08-11 14:17:39 -070090 global_dirty_limits(&background_thresh, &dirty_thresh);
91 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Miklos Szeredi76f14182008-04-30 00:54:36 -070092
93#define K(x) ((x) << (PAGE_SHIFT - 10))
94 seq_printf(m,
Wu Fengguang00821b02010-08-29 11:28:45 -060095 "BdiWriteback: %10lu kB\n"
96 "BdiReclaimable: %10lu kB\n"
97 "BdiDirtyThresh: %10lu kB\n"
98 "DirtyThresh: %10lu kB\n"
99 "BackgroundThresh: %10lu kB\n"
Wu Fengguangc8e28ce2011-01-23 10:07:47 -0600100 "BdiDirtied: %10lu kB\n"
Wu Fengguang00821b02010-08-29 11:28:45 -0600101 "BdiWritten: %10lu kB\n"
102 "BdiWriteBandwidth: %10lu kBps\n"
103 "b_dirty: %10lu\n"
104 "b_io: %10lu\n"
105 "b_more_io: %10lu\n"
106 "bdi_list: %10u\n"
107 "state: %10lx\n",
Miklos Szeredi76f14182008-04-30 00:54:36 -0700108 (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
109 (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600110 K(bdi_thresh),
111 K(dirty_thresh),
112 K(background_thresh),
Wu Fengguangc8e28ce2011-01-23 10:07:47 -0600113 (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600114 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
Wu Fengguang00821b02010-08-29 11:28:45 -0600115 (unsigned long) K(bdi->write_bandwidth),
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600116 nr_dirty,
117 nr_io,
118 nr_more_io,
Christoph Hellwigc1955ce2010-06-19 23:08:06 +0200119 !list_empty(&bdi->bdi_list), bdi->state);
Miklos Szeredi76f14182008-04-30 00:54:36 -0700120#undef K
121
122 return 0;
123}
124
125static int bdi_debug_stats_open(struct inode *inode, struct file *file)
126{
127 return single_open(file, bdi_debug_stats_show, inode->i_private);
128}
129
130static const struct file_operations bdi_debug_stats_fops = {
131 .open = bdi_debug_stats_open,
132 .read = seq_read,
133 .llseek = seq_lseek,
134 .release = single_release,
135};
136
137static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
138{
139 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
140 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
141 bdi, &bdi_debug_stats_fops);
142}
143
144static void bdi_debug_unregister(struct backing_dev_info *bdi)
145{
146 debugfs_remove(bdi->debug_stats);
147 debugfs_remove(bdi->debug_dir);
148}
149#else
150static inline void bdi_debug_init(void)
151{
152}
153static inline void bdi_debug_register(struct backing_dev_info *bdi,
154 const char *name)
155{
156}
157static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
158{
159}
160#endif
161
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700162static ssize_t read_ahead_kb_store(struct device *dev,
163 struct device_attribute *attr,
164 const char *buf, size_t count)
165{
166 struct backing_dev_info *bdi = dev_get_drvdata(dev);
167 char *end;
168 unsigned long read_ahead_kb;
169 ssize_t ret = -EINVAL;
170
171 read_ahead_kb = simple_strtoul(buf, &end, 10);
172 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
173 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
174 ret = count;
175 }
176 return ret;
177}
178
179#define K(pages) ((pages) << (PAGE_SHIFT - 10))
180
181#define BDI_SHOW(name, expr) \
182static ssize_t name##_show(struct device *dev, \
183 struct device_attribute *attr, char *page) \
184{ \
185 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
186 \
187 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
188}
189
190BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
191
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700192static ssize_t min_ratio_store(struct device *dev,
193 struct device_attribute *attr, const char *buf, size_t count)
194{
195 struct backing_dev_info *bdi = dev_get_drvdata(dev);
196 char *end;
197 unsigned int ratio;
198 ssize_t ret = -EINVAL;
199
200 ratio = simple_strtoul(buf, &end, 10);
201 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
202 ret = bdi_set_min_ratio(bdi, ratio);
203 if (!ret)
204 ret = count;
205 }
206 return ret;
207}
208BDI_SHOW(min_ratio, bdi->min_ratio)
209
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700210static ssize_t max_ratio_store(struct device *dev,
211 struct device_attribute *attr, const char *buf, size_t count)
212{
213 struct backing_dev_info *bdi = dev_get_drvdata(dev);
214 char *end;
215 unsigned int ratio;
216 ssize_t ret = -EINVAL;
217
218 ratio = simple_strtoul(buf, &end, 10);
219 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
220 ret = bdi_set_max_ratio(bdi, ratio);
221 if (!ret)
222 ret = count;
223 }
224 return ret;
225}
226BDI_SHOW(max_ratio, bdi->max_ratio)
227
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700228#define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
229
230static struct device_attribute bdi_dev_attrs[] = {
231 __ATTR_RW(read_ahead_kb),
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700232 __ATTR_RW(min_ratio),
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700233 __ATTR_RW(max_ratio),
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700234 __ATTR_NULL,
235};
236
237static __init int bdi_class_init(void)
238{
239 bdi_class = class_create(THIS_MODULE, "bdi");
Anton Blanchard14421452010-04-02 09:46:55 +0200240 if (IS_ERR(bdi_class))
241 return PTR_ERR(bdi_class);
242
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700243 bdi_class->dev_attrs = bdi_dev_attrs;
Miklos Szeredi76f14182008-04-30 00:54:36 -0700244 bdi_debug_init();
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700245 return 0;
246}
Miklos Szeredi76f14182008-04-30 00:54:36 -0700247postcore_initcall(bdi_class_init);
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700248
Jens Axboe26160152009-03-17 09:35:06 +0100249static int __init default_bdi_init(void)
250{
251 int err;
252
Jens Axboe03ba3782009-09-09 09:08:54 +0200253 sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
254 BUG_ON(IS_ERR(sync_supers_tsk));
255
Jens Axboe03ba3782009-09-09 09:08:54 +0200256 setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
Jens Axboe64231042010-05-21 20:00:35 +0200257 bdi_arm_supers_timer();
Jens Axboe03ba3782009-09-09 09:08:54 +0200258
Jens Axboe26160152009-03-17 09:35:06 +0100259 err = bdi_init(&default_backing_dev_info);
260 if (!err)
261 bdi_register(&default_backing_dev_info, NULL, "default");
Jan Kara976e48f2010-09-21 11:48:55 +0200262 err = bdi_init(&noop_backing_dev_info);
Jens Axboe26160152009-03-17 09:35:06 +0100263
264 return err;
265}
266subsys_initcall(default_bdi_init);
267
Jens Axboe03ba3782009-09-09 09:08:54 +0200268int bdi_has_dirty_io(struct backing_dev_info *bdi)
269{
270 return wb_has_dirty_io(&bdi->wb);
271}
272
Jens Axboe03ba3782009-09-09 09:08:54 +0200273/*
Artem Bityutskiy6f904ff2010-07-25 14:29:11 +0300274 * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
Jens Axboe03ba3782009-09-09 09:08:54 +0200275 * or we risk deadlocking on ->s_umount. The longer term solution would be
276 * to implement sync_supers_bdi() or similar and simply do it from the
Artem Bityutskiy6f904ff2010-07-25 14:29:11 +0300277 * bdi writeback thread individually.
Jens Axboe03ba3782009-09-09 09:08:54 +0200278 */
279static int bdi_sync_supers(void *unused)
280{
281 set_user_nice(current, 0);
282
283 while (!kthread_should_stop()) {
284 set_current_state(TASK_INTERRUPTIBLE);
285 schedule();
286
287 /*
288 * Do this periodically, like kupdated() did before.
289 */
290 sync_supers();
291 }
292
293 return 0;
294}
295
Jens Axboe64231042010-05-21 20:00:35 +0200296void bdi_arm_supers_timer(void)
Jens Axboe03ba3782009-09-09 09:08:54 +0200297{
298 unsigned long next;
299
Jens Axboe64231042010-05-21 20:00:35 +0200300 if (!dirty_writeback_interval)
301 return;
302
Jens Axboe03ba3782009-09-09 09:08:54 +0200303 next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
304 mod_timer(&sync_supers_timer, round_jiffies_up(next));
305}
306
307static void sync_supers_timer_fn(unsigned long unused)
308{
309 wake_up_process(sync_supers_tsk);
Jens Axboe64231042010-05-21 20:00:35 +0200310 bdi_arm_supers_timer();
Jens Axboe03ba3782009-09-09 09:08:54 +0200311}
312
Artem Bityutskiy64677162010-07-25 14:29:22 +0300313static void wakeup_timer_fn(unsigned long data)
314{
315 struct backing_dev_info *bdi = (struct backing_dev_info *)data;
316
317 spin_lock_bh(&bdi->wb_lock);
318 if (bdi->wb.task) {
Artem Bityutskiy60332022010-07-25 14:29:24 +0300319 trace_writeback_wake_thread(bdi);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300320 wake_up_process(bdi->wb.task);
321 } else {
322 /*
323 * When bdi tasks are inactive for long time, they are killed.
324 * In this case we have to wake-up the forker thread which
325 * should create and run the bdi thread.
326 */
Artem Bityutskiy60332022010-07-25 14:29:24 +0300327 trace_writeback_wake_forker_thread(bdi);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300328 wake_up_process(default_backing_dev_info.wb.task);
329 }
330 spin_unlock_bh(&bdi->wb_lock);
331}
332
333/*
334 * This function is used when the first inode for this bdi is marked dirty. It
335 * wakes-up the corresponding bdi thread which should then take care of the
336 * periodic background write-out of dirty inodes. Since the write-out would
337 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
338 * set up a timer which wakes the bdi thread up later.
339 *
340 * Note, we wouldn't bother setting up the timer, but this function is on the
341 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
342 * by delaying the wake-up.
343 */
344void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
345{
346 unsigned long timeout;
347
348 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
349 mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
350}
351
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300352/*
353 * Calculate the longest interval (jiffies) bdi threads are allowed to be
354 * inactive.
355 */
356static unsigned long bdi_longest_inactive(void)
357{
358 unsigned long interval;
359
360 interval = msecs_to_jiffies(dirty_writeback_interval * 10);
361 return max(5UL * 60 * HZ, interval);
362}
363
Jan Kara5a042aa2011-09-02 17:04:09 -0600364/*
365 * Clear pending bit and wakeup anybody waiting for flusher thread creation or
366 * shutdown
367 */
368static void bdi_clear_pending(struct backing_dev_info *bdi)
369{
370 clear_bit(BDI_pending, &bdi->state);
371 smp_mb__after_clear_bit();
372 wake_up_bit(&bdi->state, BDI_pending);
373}
374
Artem Bityutskiy6f904ff2010-07-25 14:29:11 +0300375static int bdi_forker_thread(void *ptr)
Jens Axboe03ba3782009-09-09 09:08:54 +0200376{
377 struct bdi_writeback *me = ptr;
378
Peter Zijlstra766f9162010-10-26 14:22:45 -0700379 current->flags |= PF_SWAPWRITE;
Christoph Hellwigc1955ce2010-06-19 23:08:06 +0200380 set_freezable();
381
382 /*
383 * Our parent may run at a different priority, just set us to normal
384 */
385 set_user_nice(current, 0);
Jens Axboe03ba3782009-09-09 09:08:54 +0200386
387 for (;;) {
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300388 struct task_struct *task = NULL;
Artem Bityutskiy78c40cb2010-07-25 14:29:17 +0300389 struct backing_dev_info *bdi;
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300390 enum {
391 NO_ACTION, /* Nothing to do */
392 FORK_THREAD, /* Fork bdi thread */
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300393 KILL_THREAD, /* Kill inactive bdi thread */
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300394 } action = NO_ACTION;
Jens Axboe03ba3782009-09-09 09:08:54 +0200395
396 /*
397 * Temporary measure, we want to make sure we don't see
398 * dirty data on the default backing_dev_info
399 */
Artem Bityutskiy64677162010-07-25 14:29:22 +0300400 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
401 del_timer(&me->wakeup_timer);
Jens Axboe03ba3782009-09-09 09:08:54 +0200402 wb_do_writeback(me, 0);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300403 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200404
Jens Axboecfc4ba52009-09-14 13:12:40 +0200405 spin_lock_bh(&bdi_lock);
Jan Kara09f40f92011-09-02 17:04:10 -0600406 /*
407 * In the following loop we are going to check whether we have
408 * some work to do without any synchronization with tasks
Andrew Morton20c8c622011-10-31 17:08:54 -0700409 * waking us up to do work for them. Set the task state here
410 * so that we don't miss wakeups after verifying conditions.
Jan Kara09f40f92011-09-02 17:04:10 -0600411 */
Artem Bityutskiyc5f7ad22010-07-25 14:29:13 +0300412 set_current_state(TASK_INTERRUPTIBLE);
Jens Axboe03ba3782009-09-09 09:08:54 +0200413
Artem Bityutskiy78c40cb2010-07-25 14:29:17 +0300414 list_for_each_entry(bdi, &bdi_list, bdi_list) {
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300415 bool have_dirty_io;
416
417 if (!bdi_cap_writeback_dirty(bdi) ||
418 bdi_cap_flush_forker(bdi))
Jens Axboe03ba3782009-09-09 09:08:54 +0200419 continue;
420
Artem Bityutskiy080dcec2010-07-25 14:29:16 +0300421 WARN(!test_bit(BDI_registered, &bdi->state),
422 "bdi %p/%s is not registered!\n", bdi, bdi->name);
423
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300424 have_dirty_io = !list_empty(&bdi->work_list) ||
425 wb_has_dirty_io(&bdi->wb);
Artem Bityutskiy78c40cb2010-07-25 14:29:17 +0300426
427 /*
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300428 * If the bdi has work to do, but the thread does not
429 * exist - create it.
Artem Bityutskiy78c40cb2010-07-25 14:29:17 +0300430 */
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300431 if (!bdi->wb.task && have_dirty_io) {
432 /*
433 * Set the pending bit - if someone will try to
434 * unregister this bdi - it'll wait on this bit.
435 */
436 set_bit(BDI_pending, &bdi->state);
437 action = FORK_THREAD;
438 break;
439 }
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300440
Jens Axboe6bf05d02010-08-04 13:34:31 +0200441 spin_lock(&bdi->wb_lock);
442
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300443 /*
444 * If there is no work to do and the bdi thread was
445 * inactive long enough - kill it. The wb_lock is taken
446 * to make sure no-one adds more work to this bdi and
447 * wakes the bdi thread up.
448 */
449 if (bdi->wb.task && !have_dirty_io &&
450 time_after(jiffies, bdi->wb.last_active +
451 bdi_longest_inactive())) {
452 task = bdi->wb.task;
453 bdi->wb.task = NULL;
454 spin_unlock(&bdi->wb_lock);
455 set_bit(BDI_pending, &bdi->state);
456 action = KILL_THREAD;
457 break;
458 }
Jens Axboe6bf05d02010-08-04 13:34:31 +0200459 spin_unlock(&bdi->wb_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200460 }
Artem Bityutskiy080dcec2010-07-25 14:29:16 +0300461 spin_unlock_bh(&bdi_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +0200462
Artem Bityutskiyc4ec7902010-07-25 14:29:14 +0300463 /* Keep working if default bdi still has things to do */
464 if (!list_empty(&me->bdi->work_list))
465 __set_current_state(TASK_RUNNING);
466
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300467 switch (action) {
468 case FORK_THREAD:
469 __set_current_state(TASK_RUNNING);
Artem Bityutskiy6628bc72010-08-27 09:15:09 +0200470 task = kthread_create(bdi_writeback_thread, &bdi->wb,
471 "flush-%s", dev_name(bdi->dev));
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300472 if (IS_ERR(task)) {
473 /*
474 * If thread creation fails, force writeout of
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600475 * the bdi from the thread. Hopefully 1024 is
476 * large enough for efficient IO.
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300477 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -0600478 writeback_inodes_wb(&bdi->wb, 1024,
479 WB_REASON_FORKER_THREAD);
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300480 } else {
481 /*
482 * The spinlock makes sure we do not lose
483 * wake-ups when racing with 'bdi_queue_work()'.
Artem Bityutskiy6628bc72010-08-27 09:15:09 +0200484 * And as soon as the bdi thread is visible, we
485 * can start it.
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300486 */
Artem Bityutskiy64677162010-07-25 14:29:22 +0300487 spin_lock_bh(&bdi->wb_lock);
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300488 bdi->wb.task = task;
Artem Bityutskiy64677162010-07-25 14:29:22 +0300489 spin_unlock_bh(&bdi->wb_lock);
Artem Bityutskiy6628bc72010-08-27 09:15:09 +0200490 wake_up_process(task);
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300491 }
Jan Kara5a042aa2011-09-02 17:04:09 -0600492 bdi_clear_pending(bdi);
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300493 break;
494
495 case KILL_THREAD:
496 __set_current_state(TASK_RUNNING);
497 kthread_stop(task);
Jan Kara5a042aa2011-09-02 17:04:09 -0600498 bdi_clear_pending(bdi);
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300499 break;
Jens Axboe03ba3782009-09-09 09:08:54 +0200500
Artem Bityutskiyadf39242010-07-25 14:29:19 +0300501 case NO_ACTION:
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300502 if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
503 /*
504 * There are no dirty data. The only thing we
505 * should now care about is checking for
506 * inactive bdi threads and killing them. Thus,
507 * let's sleep for longer time, save energy and
508 * be friendly for battery-driven devices.
509 */
510 schedule_timeout(bdi_longest_inactive());
Jens Axboe64231042010-05-21 20:00:35 +0200511 else
Artem Bityutskiy253c34e2010-07-25 14:29:21 +0300512 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
Jens Axboe03ba3782009-09-09 09:08:54 +0200513 try_to_freeze();
Jan Kara5a042aa2011-09-02 17:04:09 -0600514 break;
Jens Axboe03ba3782009-09-09 09:08:54 +0200515 }
Jens Axboe03ba3782009-09-09 09:08:54 +0200516 }
517
518 return 0;
519}
520
Jens Axboecfc4ba52009-09-14 13:12:40 +0200521/*
522 * Remove bdi from bdi_list, and ensure that it is no longer visible
523 */
524static void bdi_remove_from_list(struct backing_dev_info *bdi)
525{
526 spin_lock_bh(&bdi_lock);
527 list_del_rcu(&bdi->bdi_list);
528 spin_unlock_bh(&bdi_lock);
529
Mikulas Patockaef323082011-07-23 20:44:24 +0200530 synchronize_rcu_expedited();
Jens Axboecfc4ba52009-09-14 13:12:40 +0200531}
532
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700533int bdi_register(struct backing_dev_info *bdi, struct device *parent,
534 const char *fmt, ...)
535{
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700536 va_list args;
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700537 struct device *dev;
538
Andrew Morton69fc2082008-12-09 13:14:06 -0800539 if (bdi->dev) /* The driver needs to use separate queues per device */
Artem Bityutskiyc284de62010-07-25 14:29:25 +0300540 return 0;
Kay Sieversf1d0b062008-12-02 10:31:50 -0800541
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700542 va_start(args, fmt);
Greg Kroah-Hartman19051c52008-05-15 13:44:08 -0700543 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700544 va_end(args);
Artem Bityutskiyc284de62010-07-25 14:29:25 +0300545 if (IS_ERR(dev))
546 return PTR_ERR(dev);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200547
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700548 bdi->dev = dev;
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700549
Jens Axboe03ba3782009-09-09 09:08:54 +0200550 /*
551 * Just start the forker thread for our default backing_dev_info,
552 * and add other bdi's to the list. They will get a thread created
553 * on-demand when they need it.
554 */
555 if (bdi_cap_flush_forker(bdi)) {
556 struct bdi_writeback *wb = &bdi->wb;
557
Artem Bityutskiy6f904ff2010-07-25 14:29:11 +0300558 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
Jens Axboe03ba3782009-09-09 09:08:54 +0200559 dev_name(dev));
Artem Bityutskiyc284de62010-07-25 14:29:25 +0300560 if (IS_ERR(wb->task))
561 return PTR_ERR(wb->task);
Jens Axboe03ba3782009-09-09 09:08:54 +0200562 }
563
564 bdi_debug_register(bdi, dev_name(dev));
Jens Axboe500b0672009-09-09 09:10:25 +0200565 set_bit(BDI_registered, &bdi->state);
Artem Bityutskiyc284de62010-07-25 14:29:25 +0300566
567 spin_lock_bh(&bdi_lock);
568 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
569 spin_unlock_bh(&bdi_lock);
570
Dave Chinner455b2862010-07-07 13:24:06 +1000571 trace_writeback_bdi_register(bdi);
Artem Bityutskiyc284de62010-07-25 14:29:25 +0300572 return 0;
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700573}
574EXPORT_SYMBOL(bdi_register);
575
576int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
577{
578 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
579}
580EXPORT_SYMBOL(bdi_register_dev);
581
Jens Axboe03ba3782009-09-09 09:08:54 +0200582/*
583 * Remove bdi from the global list and shutdown any threads we have running
584 */
585static void bdi_wb_shutdown(struct backing_dev_info *bdi)
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200586{
Jens Axboe03ba3782009-09-09 09:08:54 +0200587 if (!bdi_cap_writeback_dirty(bdi))
588 return;
589
590 /*
Artem Bityutskiyfff5b852010-07-25 14:29:20 +0300591 * Make sure nobody finds us on the bdi_list anymore
592 */
593 bdi_remove_from_list(bdi);
594
595 /*
Jens Axboe03ba3782009-09-09 09:08:54 +0200596 * If setup is pending, wait for that to complete first
597 */
598 wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
599 TASK_UNINTERRUPTIBLE);
600
601 /*
Christoph Hellwigc1955ce2010-06-19 23:08:06 +0200602 * Finally, kill the kernel thread. We don't need to be RCU
Tejun Heo8a32c442011-11-21 12:32:23 -0800603 * safe anymore, since the bdi is gone from visibility.
Jens Axboe03ba3782009-09-09 09:08:54 +0200604 */
Tejun Heo8a32c442011-11-21 12:32:23 -0800605 if (bdi->wb.task)
Christoph Hellwigc1955ce2010-06-19 23:08:06 +0200606 kthread_stop(bdi->wb.task);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200607}
608
Jens Axboe592b09a2009-10-29 11:46:12 +0100609/*
610 * This bdi is going away now, make sure that no super_blocks point to it
611 */
612static void bdi_prune_sb(struct backing_dev_info *bdi)
613{
614 struct super_block *sb;
615
616 spin_lock(&sb_lock);
617 list_for_each_entry(sb, &super_blocks, s_list) {
618 if (sb->s_bdi == bdi)
Jens Axboe95f28602011-03-17 11:13:12 +0100619 sb->s_bdi = &default_backing_dev_info;
Jens Axboe592b09a2009-10-29 11:46:12 +0100620 }
621 spin_unlock(&sb_lock);
622}
623
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700624void bdi_unregister(struct backing_dev_info *bdi)
625{
626 if (bdi->dev) {
Peter Zijlstraccb61082011-07-25 17:11:57 -0700627 bdi_set_min_ratio(bdi, 0);
Dave Chinner455b2862010-07-07 13:24:06 +1000628 trace_writeback_bdi_unregister(bdi);
Jens Axboe8c4db332009-11-03 20:18:44 +0100629 bdi_prune_sb(bdi);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300630 del_timer_sync(&bdi->wb.wakeup_timer);
Jens Axboe8c4db332009-11-03 20:18:44 +0100631
Jens Axboe03ba3782009-09-09 09:08:54 +0200632 if (!bdi_cap_flush_forker(bdi))
633 bdi_wb_shutdown(bdi);
Miklos Szeredi76f14182008-04-30 00:54:36 -0700634 bdi_debug_unregister(bdi);
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700635 device_unregister(bdi->dev);
636 bdi->dev = NULL;
637 }
638}
639EXPORT_SYMBOL(bdi_unregister);
Andrew Morton3fcfab12006-10-19 23:28:16 -0700640
Artem Bityutskiy64677162010-07-25 14:29:22 +0300641static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
642{
643 memset(wb, 0, sizeof(*wb));
644
645 wb->bdi = bdi;
646 wb->last_old_flush = jiffies;
647 INIT_LIST_HEAD(&wb->b_dirty);
648 INIT_LIST_HEAD(&wb->b_io);
649 INIT_LIST_HEAD(&wb->b_more_io);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600650 spin_lock_init(&wb->list_lock);
Artem Bityutskiy64677162010-07-25 14:29:22 +0300651 setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
652}
653
Wu Fengguange98be2d2010-08-29 11:22:30 -0600654/*
655 * Initial write bandwidth: 100 MB/s
656 */
657#define INIT_BW (100 << (20 - PAGE_SHIFT))
658
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700659int bdi_init(struct backing_dev_info *bdi)
660{
Jens Axboe03ba3782009-09-09 09:08:54 +0200661 int i, err;
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700662
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700663 bdi->dev = NULL;
664
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700665 bdi->min_ratio = 0;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700666 bdi->max_ratio = 100;
667 bdi->max_prop_frac = PROP_FRAC_BASE;
Jens Axboe03ba3782009-09-09 09:08:54 +0200668 spin_lock_init(&bdi->wb_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200669 INIT_LIST_HEAD(&bdi->bdi_list);
Jens Axboe03ba3782009-09-09 09:08:54 +0200670 INIT_LIST_HEAD(&bdi->work_list);
671
672 bdi_wb_init(&bdi->wb, bdi);
673
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700674 for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
Peter Zijlstraea319512008-12-26 15:08:55 +0100675 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700676 if (err)
677 goto err;
678 }
679
680 bdi->dirty_exceeded = 0;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600681
682 bdi->bw_time_stamp = jiffies;
683 bdi->written_stamp = 0;
684
Wu Fengguang73811312011-08-26 15:53:24 -0600685 bdi->balanced_dirty_ratelimit = INIT_BW;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600686 bdi->dirty_ratelimit = INIT_BW;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600687 bdi->write_bandwidth = INIT_BW;
688 bdi->avg_write_bandwidth = INIT_BW;
689
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700690 err = prop_local_init_percpu(&bdi->completions);
691
692 if (err) {
693err:
Denis Cheng4b01a0b2007-12-04 23:45:07 -0800694 while (i--)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700695 percpu_counter_destroy(&bdi->bdi_stat[i]);
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700696 }
697
698 return err;
699}
700EXPORT_SYMBOL(bdi_init);
701
702void bdi_destroy(struct backing_dev_info *bdi)
703{
704 int i;
705
Jens Axboece5f8e72009-09-14 12:57:56 +0200706 /*
707 * Splice our entries to the default_backing_dev_info, if this
708 * bdi disappears
709 */
710 if (bdi_has_dirty_io(bdi)) {
711 struct bdi_writeback *dst = &default_backing_dev_info.wb;
712
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600713 bdi_lock_two(&bdi->wb, dst);
Jens Axboece5f8e72009-09-14 12:57:56 +0200714 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
715 list_splice(&bdi->wb.b_io, &dst->b_io);
716 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600717 spin_unlock(&bdi->wb.list_lock);
718 spin_unlock(&dst->list_lock);
Jens Axboece5f8e72009-09-14 12:57:56 +0200719 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200720
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700721 bdi_unregister(bdi);
722
Rabin Vincent7a401a92011-11-11 13:29:04 +0100723 /*
724 * If bdi_unregister() had already been called earlier, the
725 * wakeup_timer could still be armed because bdi_prune_sb()
726 * can race with the bdi_wakeup_thread_delayed() calls from
727 * __mark_inode_dirty().
728 */
729 del_timer_sync(&bdi->wb.wakeup_timer);
730
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700731 for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
732 percpu_counter_destroy(&bdi->bdi_stat[i]);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700733
734 prop_local_destroy_percpu(&bdi->completions);
Peter Zijlstrab2e8fb62007-10-16 23:25:47 -0700735}
736EXPORT_SYMBOL(bdi_destroy);
737
Jens Axboec3c53202010-04-22 11:37:01 +0200738/*
739 * For use from filesystems to quickly init and register a bdi associated
740 * with dirty writeback
741 */
742int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
743 unsigned int cap)
744{
745 char tmp[32];
746 int err;
747
748 bdi->name = name;
749 bdi->capabilities = cap;
750 err = bdi_init(bdi);
751 if (err)
752 return err;
753
754 sprintf(tmp, "%.28s%s", name, "-%d");
755 err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
756 if (err) {
757 bdi_destroy(bdi);
758 return err;
759 }
760
761 return 0;
762}
763EXPORT_SYMBOL(bdi_setup_and_register);
764
Andrew Morton3fcfab12006-10-19 23:28:16 -0700765static wait_queue_head_t congestion_wqh[2] = {
766 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
767 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
768 };
Mel Gorman0e093d992010-10-26 14:21:45 -0700769static atomic_t nr_bdi_congested[2];
Andrew Morton3fcfab12006-10-19 23:28:16 -0700770
Jens Axboe1faa16d2009-04-06 14:48:01 +0200771void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700772{
773 enum bdi_state bit;
Jens Axboe1faa16d2009-04-06 14:48:01 +0200774 wait_queue_head_t *wqh = &congestion_wqh[sync];
Andrew Morton3fcfab12006-10-19 23:28:16 -0700775
Jens Axboe1faa16d2009-04-06 14:48:01 +0200776 bit = sync ? BDI_sync_congested : BDI_async_congested;
Mel Gorman0e093d992010-10-26 14:21:45 -0700777 if (test_and_clear_bit(bit, &bdi->state))
778 atomic_dec(&nr_bdi_congested[sync]);
Andrew Morton3fcfab12006-10-19 23:28:16 -0700779 smp_mb__after_clear_bit();
780 if (waitqueue_active(wqh))
781 wake_up(wqh);
782}
783EXPORT_SYMBOL(clear_bdi_congested);
784
Jens Axboe1faa16d2009-04-06 14:48:01 +0200785void set_bdi_congested(struct backing_dev_info *bdi, int sync)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700786{
787 enum bdi_state bit;
788
Jens Axboe1faa16d2009-04-06 14:48:01 +0200789 bit = sync ? BDI_sync_congested : BDI_async_congested;
Mel Gorman0e093d992010-10-26 14:21:45 -0700790 if (!test_and_set_bit(bit, &bdi->state))
791 atomic_inc(&nr_bdi_congested[sync]);
Andrew Morton3fcfab12006-10-19 23:28:16 -0700792}
793EXPORT_SYMBOL(set_bdi_congested);
794
795/**
796 * congestion_wait - wait for a backing_dev to become uncongested
Jens Axboe8aa7e842009-07-09 14:52:32 +0200797 * @sync: SYNC or ASYNC IO
Andrew Morton3fcfab12006-10-19 23:28:16 -0700798 * @timeout: timeout in jiffies
799 *
800 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
801 * write congestion. If no backing_devs are congested then just wait for the
802 * next write to be completed.
803 */
Jens Axboe8aa7e842009-07-09 14:52:32 +0200804long congestion_wait(int sync, long timeout)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700805{
806 long ret;
Mel Gorman52bb9192010-10-26 14:21:41 -0700807 unsigned long start = jiffies;
Andrew Morton3fcfab12006-10-19 23:28:16 -0700808 DEFINE_WAIT(wait);
Jens Axboe8aa7e842009-07-09 14:52:32 +0200809 wait_queue_head_t *wqh = &congestion_wqh[sync];
Andrew Morton3fcfab12006-10-19 23:28:16 -0700810
811 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
812 ret = io_schedule_timeout(timeout);
813 finish_wait(wqh, &wait);
Mel Gorman52bb9192010-10-26 14:21:41 -0700814
815 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
816 jiffies_to_usecs(jiffies - start));
817
Andrew Morton3fcfab12006-10-19 23:28:16 -0700818 return ret;
819}
820EXPORT_SYMBOL(congestion_wait);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700821
Mel Gorman0e093d992010-10-26 14:21:45 -0700822/**
823 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
824 * @zone: A zone to check if it is heavily congested
825 * @sync: SYNC or ASYNC IO
826 * @timeout: timeout in jiffies
827 *
828 * In the event of a congested backing_dev (any backing_dev) and the given
829 * @zone has experienced recent congestion, this waits for up to @timeout
830 * jiffies for either a BDI to exit congestion of the given @sync queue
831 * or a write to complete.
832 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300833 * In the absence of zone congestion, cond_resched() is called to yield
Mel Gorman0e093d992010-10-26 14:21:45 -0700834 * the processor if necessary but otherwise does not sleep.
835 *
836 * The return value is 0 if the sleep is for the full timeout. Otherwise,
837 * it is the number of jiffies that were still remaining when the function
838 * returned. return_value == timeout implies the function did not sleep.
839 */
840long wait_iff_congested(struct zone *zone, int sync, long timeout)
841{
842 long ret;
843 unsigned long start = jiffies;
844 DEFINE_WAIT(wait);
845 wait_queue_head_t *wqh = &congestion_wqh[sync];
846
847 /*
848 * If there is no congestion, or heavy congestion is not being
849 * encountered in the current zone, yield if necessary instead
850 * of sleeping on the congestion queue
851 */
852 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
853 !zone_is_reclaim_congested(zone)) {
854 cond_resched();
855
856 /* In case we scheduled, work out time remaining */
857 ret = timeout - (jiffies - start);
858 if (ret < 0)
859 ret = 0;
860
861 goto out;
862 }
863
864 /* Sleep until uncongested or a write happens */
865 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
866 ret = io_schedule_timeout(timeout);
867 finish_wait(wqh, &wait);
868
869out:
870 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
871 jiffies_to_usecs(jiffies - start));
872
873 return ret;
874}
875EXPORT_SYMBOL(wait_iff_congested);