2 #include <linux/wait.h>
3 #include <linux/backing-dev.h>
4 #include <linux/kthread.h>
5 #include <linux/freezer.h>
7 #include <linux/pagemap.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>
15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
17 struct backing_dev_info default_backing_dev_info = {
19 .ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
21 .capabilities = BDI_CAP_MAP_COPY,
23 EXPORT_SYMBOL_GPL(default_backing_dev_info);
25 struct backing_dev_info noop_backing_dev_info = {
27 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
29 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
31 static struct class *bdi_class;
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
38 DEFINE_SPINLOCK(bdi_lock);
40 LIST_HEAD(bdi_pending_list);
42 static struct task_struct *sync_supers_tsk;
43 static struct timer_list sync_supers_timer;
45 static int bdi_sync_supers(void *);
46 static void sync_supers_timer_fn(unsigned long);
48 void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
51 spin_lock(&wb1->list_lock);
52 spin_lock_nested(&wb2->list_lock, 1);
54 spin_lock(&wb2->list_lock);
55 spin_lock_nested(&wb1->list_lock, 1);
59 #ifdef CONFIG_DEBUG_FS
60 #include <linux/debugfs.h>
61 #include <linux/seq_file.h>
63 static struct dentry *bdi_debug_root;
65 static void bdi_debug_init(void)
67 bdi_debug_root = debugfs_create_dir("bdi", NULL);
70 static int bdi_debug_stats_show(struct seq_file *m, void *v)
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;
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)
84 list_for_each_entry(inode, &wb->b_io, i_wb_list)
86 list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
88 spin_unlock(&wb->list_lock);
90 global_dirty_limits(&background_thresh, &dirty_thresh);
91 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
93 #define K(x) ((x) << (PAGE_SHIFT - 10))
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"
105 (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
106 (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
107 K(bdi_thresh), K(dirty_thresh),
108 K(background_thresh), nr_dirty, nr_io, nr_more_io,
109 !list_empty(&bdi->bdi_list), bdi->state);
115 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
117 return single_open(file, bdi_debug_stats_show, inode->i_private);
120 static const struct file_operations bdi_debug_stats_fops = {
121 .open = bdi_debug_stats_open,
124 .release = single_release,
127 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
129 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
130 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
131 bdi, &bdi_debug_stats_fops);
134 static void bdi_debug_unregister(struct backing_dev_info *bdi)
136 debugfs_remove(bdi->debug_stats);
137 debugfs_remove(bdi->debug_dir);
140 static inline void bdi_debug_init(void)
143 static inline void bdi_debug_register(struct backing_dev_info *bdi,
147 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
152 static ssize_t read_ahead_kb_store(struct device *dev,
153 struct device_attribute *attr,
154 const char *buf, size_t count)
156 struct backing_dev_info *bdi = dev_get_drvdata(dev);
158 unsigned long read_ahead_kb;
159 ssize_t ret = -EINVAL;
161 read_ahead_kb = simple_strtoul(buf, &end, 10);
162 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
163 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
169 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
171 #define BDI_SHOW(name, expr) \
172 static ssize_t name##_show(struct device *dev, \
173 struct device_attribute *attr, char *page) \
175 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
177 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
180 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
182 static ssize_t min_ratio_store(struct device *dev,
183 struct device_attribute *attr, const char *buf, size_t count)
185 struct backing_dev_info *bdi = dev_get_drvdata(dev);
188 ssize_t ret = -EINVAL;
190 ratio = simple_strtoul(buf, &end, 10);
191 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
192 ret = bdi_set_min_ratio(bdi, ratio);
198 BDI_SHOW(min_ratio, bdi->min_ratio)
200 static ssize_t max_ratio_store(struct device *dev,
201 struct device_attribute *attr, const char *buf, size_t count)
203 struct backing_dev_info *bdi = dev_get_drvdata(dev);
206 ssize_t ret = -EINVAL;
208 ratio = simple_strtoul(buf, &end, 10);
209 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
210 ret = bdi_set_max_ratio(bdi, ratio);
216 BDI_SHOW(max_ratio, bdi->max_ratio)
218 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
220 static struct device_attribute bdi_dev_attrs[] = {
221 __ATTR_RW(read_ahead_kb),
222 __ATTR_RW(min_ratio),
223 __ATTR_RW(max_ratio),
227 static __init int bdi_class_init(void)
229 bdi_class = class_create(THIS_MODULE, "bdi");
230 if (IS_ERR(bdi_class))
231 return PTR_ERR(bdi_class);
233 bdi_class->dev_attrs = bdi_dev_attrs;
237 postcore_initcall(bdi_class_init);
239 static int __init default_bdi_init(void)
243 sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
244 BUG_ON(IS_ERR(sync_supers_tsk));
246 setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
247 bdi_arm_supers_timer();
249 err = bdi_init(&default_backing_dev_info);
251 bdi_register(&default_backing_dev_info, NULL, "default");
252 err = bdi_init(&noop_backing_dev_info);
256 subsys_initcall(default_bdi_init);
258 int bdi_has_dirty_io(struct backing_dev_info *bdi)
260 return wb_has_dirty_io(&bdi->wb);
264 * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
265 * or we risk deadlocking on ->s_umount. The longer term solution would be
266 * to implement sync_supers_bdi() or similar and simply do it from the
267 * bdi writeback thread individually.
269 static int bdi_sync_supers(void *unused)
271 set_user_nice(current, 0);
273 while (!kthread_should_stop()) {
274 set_current_state(TASK_INTERRUPTIBLE);
278 * Do this periodically, like kupdated() did before.
286 void bdi_arm_supers_timer(void)
290 if (!dirty_writeback_interval)
293 next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
294 mod_timer(&sync_supers_timer, round_jiffies_up(next));
297 static void sync_supers_timer_fn(unsigned long unused)
299 wake_up_process(sync_supers_tsk);
300 bdi_arm_supers_timer();
303 static void wakeup_timer_fn(unsigned long data)
305 struct backing_dev_info *bdi = (struct backing_dev_info *)data;
307 spin_lock_bh(&bdi->wb_lock);
309 trace_writeback_wake_thread(bdi);
310 wake_up_process(bdi->wb.task);
313 * When bdi tasks are inactive for long time, they are killed.
314 * In this case we have to wake-up the forker thread which
315 * should create and run the bdi thread.
317 trace_writeback_wake_forker_thread(bdi);
318 wake_up_process(default_backing_dev_info.wb.task);
320 spin_unlock_bh(&bdi->wb_lock);
324 * This function is used when the first inode for this bdi is marked dirty. It
325 * wakes-up the corresponding bdi thread which should then take care of the
326 * periodic background write-out of dirty inodes. Since the write-out would
327 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
328 * set up a timer which wakes the bdi thread up later.
330 * Note, we wouldn't bother setting up the timer, but this function is on the
331 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
332 * by delaying the wake-up.
334 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
336 unsigned long timeout;
338 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
339 mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
343 * Calculate the longest interval (jiffies) bdi threads are allowed to be
346 static unsigned long bdi_longest_inactive(void)
348 unsigned long interval;
350 interval = msecs_to_jiffies(dirty_writeback_interval * 10);
351 return max(5UL * 60 * HZ, interval);
354 static int bdi_forker_thread(void *ptr)
356 struct bdi_writeback *me = ptr;
358 current->flags |= PF_SWAPWRITE;
362 * Our parent may run at a different priority, just set us to normal
364 set_user_nice(current, 0);
367 struct task_struct *task = NULL;
368 struct backing_dev_info *bdi;
370 NO_ACTION, /* Nothing to do */
371 FORK_THREAD, /* Fork bdi thread */
372 KILL_THREAD, /* Kill inactive bdi thread */
373 } action = NO_ACTION;
376 * Temporary measure, we want to make sure we don't see
377 * dirty data on the default backing_dev_info
379 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
380 del_timer(&me->wakeup_timer);
381 wb_do_writeback(me, 0);
384 spin_lock_bh(&bdi_lock);
385 set_current_state(TASK_INTERRUPTIBLE);
387 list_for_each_entry(bdi, &bdi_list, bdi_list) {
390 if (!bdi_cap_writeback_dirty(bdi) ||
391 bdi_cap_flush_forker(bdi))
394 WARN(!test_bit(BDI_registered, &bdi->state),
395 "bdi %p/%s is not registered!\n", bdi, bdi->name);
397 have_dirty_io = !list_empty(&bdi->work_list) ||
398 wb_has_dirty_io(&bdi->wb);
401 * If the bdi has work to do, but the thread does not
404 if (!bdi->wb.task && have_dirty_io) {
406 * Set the pending bit - if someone will try to
407 * unregister this bdi - it'll wait on this bit.
409 set_bit(BDI_pending, &bdi->state);
410 action = FORK_THREAD;
414 spin_lock(&bdi->wb_lock);
417 * If there is no work to do and the bdi thread was
418 * inactive long enough - kill it. The wb_lock is taken
419 * to make sure no-one adds more work to this bdi and
420 * wakes the bdi thread up.
422 if (bdi->wb.task && !have_dirty_io &&
423 time_after(jiffies, bdi->wb.last_active +
424 bdi_longest_inactive())) {
427 spin_unlock(&bdi->wb_lock);
428 set_bit(BDI_pending, &bdi->state);
429 action = KILL_THREAD;
432 spin_unlock(&bdi->wb_lock);
434 spin_unlock_bh(&bdi_lock);
436 /* Keep working if default bdi still has things to do */
437 if (!list_empty(&me->bdi->work_list))
438 __set_current_state(TASK_RUNNING);
442 __set_current_state(TASK_RUNNING);
443 task = kthread_create(bdi_writeback_thread, &bdi->wb,
444 "flush-%s", dev_name(bdi->dev));
447 * If thread creation fails, force writeout of
448 * the bdi from the thread. Hopefully 1024 is
449 * large enough for efficient IO.
451 writeback_inodes_wb(&bdi->wb, 1024);
454 * The spinlock makes sure we do not lose
455 * wake-ups when racing with 'bdi_queue_work()'.
456 * And as soon as the bdi thread is visible, we
459 spin_lock_bh(&bdi->wb_lock);
461 spin_unlock_bh(&bdi->wb_lock);
462 wake_up_process(task);
467 __set_current_state(TASK_RUNNING);
472 if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
474 * There are no dirty data. The only thing we
475 * should now care about is checking for
476 * inactive bdi threads and killing them. Thus,
477 * let's sleep for longer time, save energy and
478 * be friendly for battery-driven devices.
480 schedule_timeout(bdi_longest_inactive());
482 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
484 /* Back to the main loop */
489 * Clear pending bit and wakeup anybody waiting to tear us down.
491 clear_bit(BDI_pending, &bdi->state);
492 smp_mb__after_clear_bit();
493 wake_up_bit(&bdi->state, BDI_pending);
500 * Remove bdi from bdi_list, and ensure that it is no longer visible
502 static void bdi_remove_from_list(struct backing_dev_info *bdi)
504 spin_lock_bh(&bdi_lock);
505 list_del_rcu(&bdi->bdi_list);
506 spin_unlock_bh(&bdi_lock);
511 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
512 const char *fmt, ...)
517 if (bdi->dev) /* The driver needs to use separate queues per device */
521 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
529 * Just start the forker thread for our default backing_dev_info,
530 * and add other bdi's to the list. They will get a thread created
531 * on-demand when they need it.
533 if (bdi_cap_flush_forker(bdi)) {
534 struct bdi_writeback *wb = &bdi->wb;
536 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
538 if (IS_ERR(wb->task))
539 return PTR_ERR(wb->task);
542 bdi_debug_register(bdi, dev_name(dev));
543 set_bit(BDI_registered, &bdi->state);
545 spin_lock_bh(&bdi_lock);
546 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
547 spin_unlock_bh(&bdi_lock);
549 trace_writeback_bdi_register(bdi);
552 EXPORT_SYMBOL(bdi_register);
554 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
556 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
558 EXPORT_SYMBOL(bdi_register_dev);
561 * Remove bdi from the global list and shutdown any threads we have running
563 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
565 if (!bdi_cap_writeback_dirty(bdi))
569 * Make sure nobody finds us on the bdi_list anymore
571 bdi_remove_from_list(bdi);
574 * If setup is pending, wait for that to complete first
576 wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
577 TASK_UNINTERRUPTIBLE);
580 * Finally, kill the kernel thread. We don't need to be RCU
581 * safe anymore, since the bdi is gone from visibility. Force
582 * unfreeze of the thread before calling kthread_stop(), otherwise
583 * it would never exet if it is currently stuck in the refrigerator.
586 thaw_process(bdi->wb.task);
587 kthread_stop(bdi->wb.task);
592 * This bdi is going away now, make sure that no super_blocks point to it
594 static void bdi_prune_sb(struct backing_dev_info *bdi)
596 struct super_block *sb;
599 list_for_each_entry(sb, &super_blocks, s_list) {
600 if (sb->s_bdi == bdi)
601 sb->s_bdi = &default_backing_dev_info;
603 spin_unlock(&sb_lock);
606 void bdi_unregister(struct backing_dev_info *bdi)
609 trace_writeback_bdi_unregister(bdi);
611 del_timer_sync(&bdi->wb.wakeup_timer);
613 if (!bdi_cap_flush_forker(bdi))
614 bdi_wb_shutdown(bdi);
615 bdi_debug_unregister(bdi);
616 device_unregister(bdi->dev);
620 EXPORT_SYMBOL(bdi_unregister);
622 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
624 memset(wb, 0, sizeof(*wb));
627 wb->last_old_flush = jiffies;
628 INIT_LIST_HEAD(&wb->b_dirty);
629 INIT_LIST_HEAD(&wb->b_io);
630 INIT_LIST_HEAD(&wb->b_more_io);
631 spin_lock_init(&wb->list_lock);
632 setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
635 int bdi_init(struct backing_dev_info *bdi)
642 bdi->max_ratio = 100;
643 bdi->max_prop_frac = PROP_FRAC_BASE;
644 spin_lock_init(&bdi->wb_lock);
645 INIT_LIST_HEAD(&bdi->bdi_list);
646 INIT_LIST_HEAD(&bdi->work_list);
648 bdi_wb_init(&bdi->wb, bdi);
650 for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
651 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
656 bdi->dirty_exceeded = 0;
657 err = prop_local_init_percpu(&bdi->completions);
662 percpu_counter_destroy(&bdi->bdi_stat[i]);
667 EXPORT_SYMBOL(bdi_init);
669 void bdi_destroy(struct backing_dev_info *bdi)
674 * Splice our entries to the default_backing_dev_info, if this
677 if (bdi_has_dirty_io(bdi)) {
678 struct bdi_writeback *dst = &default_backing_dev_info.wb;
680 bdi_lock_two(&bdi->wb, dst);
681 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
682 list_splice(&bdi->wb.b_io, &dst->b_io);
683 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
684 spin_unlock(&bdi->wb.list_lock);
685 spin_unlock(&dst->list_lock);
690 for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
691 percpu_counter_destroy(&bdi->bdi_stat[i]);
693 prop_local_destroy_percpu(&bdi->completions);
695 EXPORT_SYMBOL(bdi_destroy);
698 * For use from filesystems to quickly init and register a bdi associated
699 * with dirty writeback
701 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
708 bdi->capabilities = cap;
713 sprintf(tmp, "%.28s%s", name, "-%d");
714 err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
722 EXPORT_SYMBOL(bdi_setup_and_register);
724 static wait_queue_head_t congestion_wqh[2] = {
725 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
726 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
728 static atomic_t nr_bdi_congested[2];
730 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
733 wait_queue_head_t *wqh = &congestion_wqh[sync];
735 bit = sync ? BDI_sync_congested : BDI_async_congested;
736 if (test_and_clear_bit(bit, &bdi->state))
737 atomic_dec(&nr_bdi_congested[sync]);
738 smp_mb__after_clear_bit();
739 if (waitqueue_active(wqh))
742 EXPORT_SYMBOL(clear_bdi_congested);
744 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
748 bit = sync ? BDI_sync_congested : BDI_async_congested;
749 if (!test_and_set_bit(bit, &bdi->state))
750 atomic_inc(&nr_bdi_congested[sync]);
752 EXPORT_SYMBOL(set_bdi_congested);
755 * congestion_wait - wait for a backing_dev to become uncongested
756 * @sync: SYNC or ASYNC IO
757 * @timeout: timeout in jiffies
759 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
760 * write congestion. If no backing_devs are congested then just wait for the
761 * next write to be completed.
763 long congestion_wait(int sync, long timeout)
766 unsigned long start = jiffies;
768 wait_queue_head_t *wqh = &congestion_wqh[sync];
770 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
771 ret = io_schedule_timeout(timeout);
772 finish_wait(wqh, &wait);
774 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
775 jiffies_to_usecs(jiffies - start));
779 EXPORT_SYMBOL(congestion_wait);
782 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
783 * @zone: A zone to check if it is heavily congested
784 * @sync: SYNC or ASYNC IO
785 * @timeout: timeout in jiffies
787 * In the event of a congested backing_dev (any backing_dev) and the given
788 * @zone has experienced recent congestion, this waits for up to @timeout
789 * jiffies for either a BDI to exit congestion of the given @sync queue
790 * or a write to complete.
792 * In the absence of zone congestion, cond_resched() is called to yield
793 * the processor if necessary but otherwise does not sleep.
795 * The return value is 0 if the sleep is for the full timeout. Otherwise,
796 * it is the number of jiffies that were still remaining when the function
797 * returned. return_value == timeout implies the function did not sleep.
799 long wait_iff_congested(struct zone *zone, int sync, long timeout)
802 unsigned long start = jiffies;
804 wait_queue_head_t *wqh = &congestion_wqh[sync];
807 * If there is no congestion, or heavy congestion is not being
808 * encountered in the current zone, yield if necessary instead
809 * of sleeping on the congestion queue
811 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
812 !zone_is_reclaim_congested(zone)) {
815 /* In case we scheduled, work out time remaining */
816 ret = timeout - (jiffies - start);
823 /* Sleep until uncongested or a write happens */
824 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
825 ret = io_schedule_timeout(timeout);
826 finish_wait(wqh, &wait);
829 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
830 jiffies_to_usecs(jiffies - start));
834 EXPORT_SYMBOL(wait_iff_congested);