writeback: account per-bdi accumulated written pages
[linux-2.6.git] / mm / backing-dev.c
1
2 #include <linux/wait.h>
3 #include <linux/backing-dev.h>
4 #include <linux/kthread.h>
5 #include <linux/freezer.h>
6 #include <linux/fs.h>
7 #include <linux/pagemap.h>
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/module.h>
11 #include <linux/writeback.h>
12 #include <linux/device.h>
13 #include <trace/events/writeback.h>
14
15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
17 struct backing_dev_info default_backing_dev_info = {
18         .name           = "default",
19         .ra_pages       = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
20         .state          = 0,
21         .capabilities   = BDI_CAP_MAP_COPY,
22 };
23 EXPORT_SYMBOL_GPL(default_backing_dev_info);
24
25 struct backing_dev_info noop_backing_dev_info = {
26         .name           = "noop",
27         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
28 };
29 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
30
31 static struct class *bdi_class;
32
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  */
38 DEFINE_SPINLOCK(bdi_lock);
39 LIST_HEAD(bdi_list);
40 LIST_HEAD(bdi_pending_list);
41
42 static struct task_struct *sync_supers_tsk;
43 static struct timer_list sync_supers_timer;
44
45 static int bdi_sync_supers(void *);
46 static void sync_supers_timer_fn(unsigned long);
47
48 void 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
59 #ifdef CONFIG_DEBUG_FS
60 #include <linux/debugfs.h>
61 #include <linux/seq_file.h>
62
63 static struct dentry *bdi_debug_root;
64
65 static void bdi_debug_init(void)
66 {
67         bdi_debug_root = debugfs_create_dir("bdi", NULL);
68 }
69
70 static int bdi_debug_stats_show(struct seq_file *m, void *v)
71 {
72         struct backing_dev_info *bdi = m->private;
73         struct bdi_writeback *wb = &bdi->wb;
74         unsigned long background_thresh;
75         unsigned long dirty_thresh;
76         unsigned long bdi_thresh;
77         unsigned long nr_dirty, nr_io, nr_more_io;
78         struct inode *inode;
79
80         nr_dirty = nr_io = nr_more_io = 0;
81         spin_lock(&wb->list_lock);
82         list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
83                 nr_dirty++;
84         list_for_each_entry(inode, &wb->b_io, i_wb_list)
85                 nr_io++;
86         list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
87                 nr_more_io++;
88         spin_unlock(&wb->list_lock);
89
90         global_dirty_limits(&background_thresh, &dirty_thresh);
91         bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
92
93 #define K(x) ((x) << (PAGE_SHIFT - 10))
94         seq_printf(m,
95                    "BdiWriteback:     %8lu kB\n"
96                    "BdiReclaimable:   %8lu kB\n"
97                    "BdiDirtyThresh:   %8lu kB\n"
98                    "DirtyThresh:      %8lu kB\n"
99                    "BackgroundThresh: %8lu kB\n"
100                    "BdiWritten:       %8lu kB\n"
101                    "b_dirty:          %8lu\n"
102                    "b_io:             %8lu\n"
103                    "b_more_io:        %8lu\n"
104                    "bdi_list:         %8u\n"
105                    "state:            %8lx\n",
106                    (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
107                    (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
108                    K(bdi_thresh),
109                    K(dirty_thresh),
110                    K(background_thresh),
111                    (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
112                    nr_dirty,
113                    nr_io,
114                    nr_more_io,
115                    !list_empty(&bdi->bdi_list), bdi->state);
116 #undef K
117
118         return 0;
119 }
120
121 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
122 {
123         return single_open(file, bdi_debug_stats_show, inode->i_private);
124 }
125
126 static const struct file_operations bdi_debug_stats_fops = {
127         .open           = bdi_debug_stats_open,
128         .read           = seq_read,
129         .llseek         = seq_lseek,
130         .release        = single_release,
131 };
132
133 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
134 {
135         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
136         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
137                                                bdi, &bdi_debug_stats_fops);
138 }
139
140 static void bdi_debug_unregister(struct backing_dev_info *bdi)
141 {
142         debugfs_remove(bdi->debug_stats);
143         debugfs_remove(bdi->debug_dir);
144 }
145 #else
146 static inline void bdi_debug_init(void)
147 {
148 }
149 static inline void bdi_debug_register(struct backing_dev_info *bdi,
150                                       const char *name)
151 {
152 }
153 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
154 {
155 }
156 #endif
157
158 static ssize_t read_ahead_kb_store(struct device *dev,
159                                   struct device_attribute *attr,
160                                   const char *buf, size_t count)
161 {
162         struct backing_dev_info *bdi = dev_get_drvdata(dev);
163         char *end;
164         unsigned long read_ahead_kb;
165         ssize_t ret = -EINVAL;
166
167         read_ahead_kb = simple_strtoul(buf, &end, 10);
168         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
169                 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
170                 ret = count;
171         }
172         return ret;
173 }
174
175 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
176
177 #define BDI_SHOW(name, expr)                                            \
178 static ssize_t name##_show(struct device *dev,                          \
179                            struct device_attribute *attr, char *page)   \
180 {                                                                       \
181         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
182                                                                         \
183         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
184 }
185
186 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
187
188 static ssize_t min_ratio_store(struct device *dev,
189                 struct device_attribute *attr, const char *buf, size_t count)
190 {
191         struct backing_dev_info *bdi = dev_get_drvdata(dev);
192         char *end;
193         unsigned int ratio;
194         ssize_t ret = -EINVAL;
195
196         ratio = simple_strtoul(buf, &end, 10);
197         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
198                 ret = bdi_set_min_ratio(bdi, ratio);
199                 if (!ret)
200                         ret = count;
201         }
202         return ret;
203 }
204 BDI_SHOW(min_ratio, bdi->min_ratio)
205
206 static ssize_t max_ratio_store(struct device *dev,
207                 struct device_attribute *attr, const char *buf, size_t count)
208 {
209         struct backing_dev_info *bdi = dev_get_drvdata(dev);
210         char *end;
211         unsigned int ratio;
212         ssize_t ret = -EINVAL;
213
214         ratio = simple_strtoul(buf, &end, 10);
215         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
216                 ret = bdi_set_max_ratio(bdi, ratio);
217                 if (!ret)
218                         ret = count;
219         }
220         return ret;
221 }
222 BDI_SHOW(max_ratio, bdi->max_ratio)
223
224 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
225
226 static struct device_attribute bdi_dev_attrs[] = {
227         __ATTR_RW(read_ahead_kb),
228         __ATTR_RW(min_ratio),
229         __ATTR_RW(max_ratio),
230         __ATTR_NULL,
231 };
232
233 static __init int bdi_class_init(void)
234 {
235         bdi_class = class_create(THIS_MODULE, "bdi");
236         if (IS_ERR(bdi_class))
237                 return PTR_ERR(bdi_class);
238
239         bdi_class->dev_attrs = bdi_dev_attrs;
240         bdi_debug_init();
241         return 0;
242 }
243 postcore_initcall(bdi_class_init);
244
245 static int __init default_bdi_init(void)
246 {
247         int err;
248
249         sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
250         BUG_ON(IS_ERR(sync_supers_tsk));
251
252         setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
253         bdi_arm_supers_timer();
254
255         err = bdi_init(&default_backing_dev_info);
256         if (!err)
257                 bdi_register(&default_backing_dev_info, NULL, "default");
258         err = bdi_init(&noop_backing_dev_info);
259
260         return err;
261 }
262 subsys_initcall(default_bdi_init);
263
264 int bdi_has_dirty_io(struct backing_dev_info *bdi)
265 {
266         return wb_has_dirty_io(&bdi->wb);
267 }
268
269 /*
270  * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
271  * or we risk deadlocking on ->s_umount. The longer term solution would be
272  * to implement sync_supers_bdi() or similar and simply do it from the
273  * bdi writeback thread individually.
274  */
275 static int bdi_sync_supers(void *unused)
276 {
277         set_user_nice(current, 0);
278
279         while (!kthread_should_stop()) {
280                 set_current_state(TASK_INTERRUPTIBLE);
281                 schedule();
282
283                 /*
284                  * Do this periodically, like kupdated() did before.
285                  */
286                 sync_supers();
287         }
288
289         return 0;
290 }
291
292 void bdi_arm_supers_timer(void)
293 {
294         unsigned long next;
295
296         if (!dirty_writeback_interval)
297                 return;
298
299         next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
300         mod_timer(&sync_supers_timer, round_jiffies_up(next));
301 }
302
303 static void sync_supers_timer_fn(unsigned long unused)
304 {
305         wake_up_process(sync_supers_tsk);
306         bdi_arm_supers_timer();
307 }
308
309 static void wakeup_timer_fn(unsigned long data)
310 {
311         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
312
313         spin_lock_bh(&bdi->wb_lock);
314         if (bdi->wb.task) {
315                 trace_writeback_wake_thread(bdi);
316                 wake_up_process(bdi->wb.task);
317         } else {
318                 /*
319                  * When bdi tasks are inactive for long time, they are killed.
320                  * In this case we have to wake-up the forker thread which
321                  * should create and run the bdi thread.
322                  */
323                 trace_writeback_wake_forker_thread(bdi);
324                 wake_up_process(default_backing_dev_info.wb.task);
325         }
326         spin_unlock_bh(&bdi->wb_lock);
327 }
328
329 /*
330  * This function is used when the first inode for this bdi is marked dirty. It
331  * wakes-up the corresponding bdi thread which should then take care of the
332  * periodic background write-out of dirty inodes. Since the write-out would
333  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
334  * set up a timer which wakes the bdi thread up later.
335  *
336  * Note, we wouldn't bother setting up the timer, but this function is on the
337  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
338  * by delaying the wake-up.
339  */
340 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
341 {
342         unsigned long timeout;
343
344         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
345         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
346 }
347
348 /*
349  * Calculate the longest interval (jiffies) bdi threads are allowed to be
350  * inactive.
351  */
352 static unsigned long bdi_longest_inactive(void)
353 {
354         unsigned long interval;
355
356         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
357         return max(5UL * 60 * HZ, interval);
358 }
359
360 static int bdi_forker_thread(void *ptr)
361 {
362         struct bdi_writeback *me = ptr;
363
364         current->flags |= PF_SWAPWRITE;
365         set_freezable();
366
367         /*
368          * Our parent may run at a different priority, just set us to normal
369          */
370         set_user_nice(current, 0);
371
372         for (;;) {
373                 struct task_struct *task = NULL;
374                 struct backing_dev_info *bdi;
375                 enum {
376                         NO_ACTION,   /* Nothing to do */
377                         FORK_THREAD, /* Fork bdi thread */
378                         KILL_THREAD, /* Kill inactive bdi thread */
379                 } action = NO_ACTION;
380
381                 /*
382                  * Temporary measure, we want to make sure we don't see
383                  * dirty data on the default backing_dev_info
384                  */
385                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
386                         del_timer(&me->wakeup_timer);
387                         wb_do_writeback(me, 0);
388                 }
389
390                 spin_lock_bh(&bdi_lock);
391                 set_current_state(TASK_INTERRUPTIBLE);
392
393                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
394                         bool have_dirty_io;
395
396                         if (!bdi_cap_writeback_dirty(bdi) ||
397                              bdi_cap_flush_forker(bdi))
398                                 continue;
399
400                         WARN(!test_bit(BDI_registered, &bdi->state),
401                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
402
403                         have_dirty_io = !list_empty(&bdi->work_list) ||
404                                         wb_has_dirty_io(&bdi->wb);
405
406                         /*
407                          * If the bdi has work to do, but the thread does not
408                          * exist - create it.
409                          */
410                         if (!bdi->wb.task && have_dirty_io) {
411                                 /*
412                                  * Set the pending bit - if someone will try to
413                                  * unregister this bdi - it'll wait on this bit.
414                                  */
415                                 set_bit(BDI_pending, &bdi->state);
416                                 action = FORK_THREAD;
417                                 break;
418                         }
419
420                         spin_lock(&bdi->wb_lock);
421
422                         /*
423                          * If there is no work to do and the bdi thread was
424                          * inactive long enough - kill it. The wb_lock is taken
425                          * to make sure no-one adds more work to this bdi and
426                          * wakes the bdi thread up.
427                          */
428                         if (bdi->wb.task && !have_dirty_io &&
429                             time_after(jiffies, bdi->wb.last_active +
430                                                 bdi_longest_inactive())) {
431                                 task = bdi->wb.task;
432                                 bdi->wb.task = NULL;
433                                 spin_unlock(&bdi->wb_lock);
434                                 set_bit(BDI_pending, &bdi->state);
435                                 action = KILL_THREAD;
436                                 break;
437                         }
438                         spin_unlock(&bdi->wb_lock);
439                 }
440                 spin_unlock_bh(&bdi_lock);
441
442                 /* Keep working if default bdi still has things to do */
443                 if (!list_empty(&me->bdi->work_list))
444                         __set_current_state(TASK_RUNNING);
445
446                 switch (action) {
447                 case FORK_THREAD:
448                         __set_current_state(TASK_RUNNING);
449                         task = kthread_create(bdi_writeback_thread, &bdi->wb,
450                                               "flush-%s", dev_name(bdi->dev));
451                         if (IS_ERR(task)) {
452                                 /*
453                                  * If thread creation fails, force writeout of
454                                  * the bdi from the thread. Hopefully 1024 is
455                                  * large enough for efficient IO.
456                                  */
457                                 writeback_inodes_wb(&bdi->wb, 1024);
458                         } else {
459                                 /*
460                                  * The spinlock makes sure we do not lose
461                                  * wake-ups when racing with 'bdi_queue_work()'.
462                                  * And as soon as the bdi thread is visible, we
463                                  * can start it.
464                                  */
465                                 spin_lock_bh(&bdi->wb_lock);
466                                 bdi->wb.task = task;
467                                 spin_unlock_bh(&bdi->wb_lock);
468                                 wake_up_process(task);
469                         }
470                         break;
471
472                 case KILL_THREAD:
473                         __set_current_state(TASK_RUNNING);
474                         kthread_stop(task);
475                         break;
476
477                 case NO_ACTION:
478                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
479                                 /*
480                                  * There are no dirty data. The only thing we
481                                  * should now care about is checking for
482                                  * inactive bdi threads and killing them. Thus,
483                                  * let's sleep for longer time, save energy and
484                                  * be friendly for battery-driven devices.
485                                  */
486                                 schedule_timeout(bdi_longest_inactive());
487                         else
488                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
489                         try_to_freeze();
490                         /* Back to the main loop */
491                         continue;
492                 }
493
494                 /*
495                  * Clear pending bit and wakeup anybody waiting to tear us down.
496                  */
497                 clear_bit(BDI_pending, &bdi->state);
498                 smp_mb__after_clear_bit();
499                 wake_up_bit(&bdi->state, BDI_pending);
500         }
501
502         return 0;
503 }
504
505 /*
506  * Remove bdi from bdi_list, and ensure that it is no longer visible
507  */
508 static void bdi_remove_from_list(struct backing_dev_info *bdi)
509 {
510         spin_lock_bh(&bdi_lock);
511         list_del_rcu(&bdi->bdi_list);
512         spin_unlock_bh(&bdi_lock);
513
514         synchronize_rcu();
515 }
516
517 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
518                 const char *fmt, ...)
519 {
520         va_list args;
521         struct device *dev;
522
523         if (bdi->dev)   /* The driver needs to use separate queues per device */
524                 return 0;
525
526         va_start(args, fmt);
527         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
528         va_end(args);
529         if (IS_ERR(dev))
530                 return PTR_ERR(dev);
531
532         bdi->dev = dev;
533
534         /*
535          * Just start the forker thread for our default backing_dev_info,
536          * and add other bdi's to the list. They will get a thread created
537          * on-demand when they need it.
538          */
539         if (bdi_cap_flush_forker(bdi)) {
540                 struct bdi_writeback *wb = &bdi->wb;
541
542                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
543                                                 dev_name(dev));
544                 if (IS_ERR(wb->task))
545                         return PTR_ERR(wb->task);
546         }
547
548         bdi_debug_register(bdi, dev_name(dev));
549         set_bit(BDI_registered, &bdi->state);
550
551         spin_lock_bh(&bdi_lock);
552         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
553         spin_unlock_bh(&bdi_lock);
554
555         trace_writeback_bdi_register(bdi);
556         return 0;
557 }
558 EXPORT_SYMBOL(bdi_register);
559
560 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
561 {
562         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
563 }
564 EXPORT_SYMBOL(bdi_register_dev);
565
566 /*
567  * Remove bdi from the global list and shutdown any threads we have running
568  */
569 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
570 {
571         if (!bdi_cap_writeback_dirty(bdi))
572                 return;
573
574         /*
575          * Make sure nobody finds us on the bdi_list anymore
576          */
577         bdi_remove_from_list(bdi);
578
579         /*
580          * If setup is pending, wait for that to complete first
581          */
582         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
583                         TASK_UNINTERRUPTIBLE);
584
585         /*
586          * Finally, kill the kernel thread. We don't need to be RCU
587          * safe anymore, since the bdi is gone from visibility. Force
588          * unfreeze of the thread before calling kthread_stop(), otherwise
589          * it would never exet if it is currently stuck in the refrigerator.
590          */
591         if (bdi->wb.task) {
592                 thaw_process(bdi->wb.task);
593                 kthread_stop(bdi->wb.task);
594         }
595 }
596
597 /*
598  * This bdi is going away now, make sure that no super_blocks point to it
599  */
600 static void bdi_prune_sb(struct backing_dev_info *bdi)
601 {
602         struct super_block *sb;
603
604         spin_lock(&sb_lock);
605         list_for_each_entry(sb, &super_blocks, s_list) {
606                 if (sb->s_bdi == bdi)
607                         sb->s_bdi = &default_backing_dev_info;
608         }
609         spin_unlock(&sb_lock);
610 }
611
612 void bdi_unregister(struct backing_dev_info *bdi)
613 {
614         if (bdi->dev) {
615                 trace_writeback_bdi_unregister(bdi);
616                 bdi_prune_sb(bdi);
617                 del_timer_sync(&bdi->wb.wakeup_timer);
618
619                 if (!bdi_cap_flush_forker(bdi))
620                         bdi_wb_shutdown(bdi);
621                 bdi_debug_unregister(bdi);
622                 device_unregister(bdi->dev);
623                 bdi->dev = NULL;
624         }
625 }
626 EXPORT_SYMBOL(bdi_unregister);
627
628 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
629 {
630         memset(wb, 0, sizeof(*wb));
631
632         wb->bdi = bdi;
633         wb->last_old_flush = jiffies;
634         INIT_LIST_HEAD(&wb->b_dirty);
635         INIT_LIST_HEAD(&wb->b_io);
636         INIT_LIST_HEAD(&wb->b_more_io);
637         spin_lock_init(&wb->list_lock);
638         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
639 }
640
641 int bdi_init(struct backing_dev_info *bdi)
642 {
643         int i, err;
644
645         bdi->dev = NULL;
646
647         bdi->min_ratio = 0;
648         bdi->max_ratio = 100;
649         bdi->max_prop_frac = PROP_FRAC_BASE;
650         spin_lock_init(&bdi->wb_lock);
651         INIT_LIST_HEAD(&bdi->bdi_list);
652         INIT_LIST_HEAD(&bdi->work_list);
653
654         bdi_wb_init(&bdi->wb, bdi);
655
656         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
657                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
658                 if (err)
659                         goto err;
660         }
661
662         bdi->dirty_exceeded = 0;
663         err = prop_local_init_percpu(&bdi->completions);
664
665         if (err) {
666 err:
667                 while (i--)
668                         percpu_counter_destroy(&bdi->bdi_stat[i]);
669         }
670
671         return err;
672 }
673 EXPORT_SYMBOL(bdi_init);
674
675 void bdi_destroy(struct backing_dev_info *bdi)
676 {
677         int i;
678
679         /*
680          * Splice our entries to the default_backing_dev_info, if this
681          * bdi disappears
682          */
683         if (bdi_has_dirty_io(bdi)) {
684                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
685
686                 bdi_lock_two(&bdi->wb, dst);
687                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
688                 list_splice(&bdi->wb.b_io, &dst->b_io);
689                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
690                 spin_unlock(&bdi->wb.list_lock);
691                 spin_unlock(&dst->list_lock);
692         }
693
694         bdi_unregister(bdi);
695
696         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
697                 percpu_counter_destroy(&bdi->bdi_stat[i]);
698
699         prop_local_destroy_percpu(&bdi->completions);
700 }
701 EXPORT_SYMBOL(bdi_destroy);
702
703 /*
704  * For use from filesystems to quickly init and register a bdi associated
705  * with dirty writeback
706  */
707 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
708                            unsigned int cap)
709 {
710         char tmp[32];
711         int err;
712
713         bdi->name = name;
714         bdi->capabilities = cap;
715         err = bdi_init(bdi);
716         if (err)
717                 return err;
718
719         sprintf(tmp, "%.28s%s", name, "-%d");
720         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
721         if (err) {
722                 bdi_destroy(bdi);
723                 return err;
724         }
725
726         return 0;
727 }
728 EXPORT_SYMBOL(bdi_setup_and_register);
729
730 static wait_queue_head_t congestion_wqh[2] = {
731                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
732                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
733         };
734 static atomic_t nr_bdi_congested[2];
735
736 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
737 {
738         enum bdi_state bit;
739         wait_queue_head_t *wqh = &congestion_wqh[sync];
740
741         bit = sync ? BDI_sync_congested : BDI_async_congested;
742         if (test_and_clear_bit(bit, &bdi->state))
743                 atomic_dec(&nr_bdi_congested[sync]);
744         smp_mb__after_clear_bit();
745         if (waitqueue_active(wqh))
746                 wake_up(wqh);
747 }
748 EXPORT_SYMBOL(clear_bdi_congested);
749
750 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
751 {
752         enum bdi_state bit;
753
754         bit = sync ? BDI_sync_congested : BDI_async_congested;
755         if (!test_and_set_bit(bit, &bdi->state))
756                 atomic_inc(&nr_bdi_congested[sync]);
757 }
758 EXPORT_SYMBOL(set_bdi_congested);
759
760 /**
761  * congestion_wait - wait for a backing_dev to become uncongested
762  * @sync: SYNC or ASYNC IO
763  * @timeout: timeout in jiffies
764  *
765  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
766  * write congestion.  If no backing_devs are congested then just wait for the
767  * next write to be completed.
768  */
769 long congestion_wait(int sync, long timeout)
770 {
771         long ret;
772         unsigned long start = jiffies;
773         DEFINE_WAIT(wait);
774         wait_queue_head_t *wqh = &congestion_wqh[sync];
775
776         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
777         ret = io_schedule_timeout(timeout);
778         finish_wait(wqh, &wait);
779
780         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
781                                         jiffies_to_usecs(jiffies - start));
782
783         return ret;
784 }
785 EXPORT_SYMBOL(congestion_wait);
786
787 /**
788  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
789  * @zone: A zone to check if it is heavily congested
790  * @sync: SYNC or ASYNC IO
791  * @timeout: timeout in jiffies
792  *
793  * In the event of a congested backing_dev (any backing_dev) and the given
794  * @zone has experienced recent congestion, this waits for up to @timeout
795  * jiffies for either a BDI to exit congestion of the given @sync queue
796  * or a write to complete.
797  *
798  * In the absence of zone congestion, cond_resched() is called to yield
799  * the processor if necessary but otherwise does not sleep.
800  *
801  * The return value is 0 if the sleep is for the full timeout. Otherwise,
802  * it is the number of jiffies that were still remaining when the function
803  * returned. return_value == timeout implies the function did not sleep.
804  */
805 long wait_iff_congested(struct zone *zone, int sync, long timeout)
806 {
807         long ret;
808         unsigned long start = jiffies;
809         DEFINE_WAIT(wait);
810         wait_queue_head_t *wqh = &congestion_wqh[sync];
811
812         /*
813          * If there is no congestion, or heavy congestion is not being
814          * encountered in the current zone, yield if necessary instead
815          * of sleeping on the congestion queue
816          */
817         if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
818                         !zone_is_reclaim_congested(zone)) {
819                 cond_resched();
820
821                 /* In case we scheduled, work out time remaining */
822                 ret = timeout - (jiffies - start);
823                 if (ret < 0)
824                         ret = 0;
825
826                 goto out;
827         }
828
829         /* Sleep until uncongested or a write happens */
830         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
831         ret = io_schedule_timeout(timeout);
832         finish_wait(wqh, &wait);
833
834 out:
835         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
836                                         jiffies_to_usecs(jiffies - start));
837
838         return ret;
839 }
840 EXPORT_SYMBOL(wait_iff_congested);