Loading...
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
15static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
17struct 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};
23EXPORT_SYMBOL_GPL(default_backing_dev_info);
24
25struct backing_dev_info noop_backing_dev_info = {
26 .name = "noop",
27 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
28};
29EXPORT_SYMBOL_GPL(noop_backing_dev_info);
30
31static 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 */
38DEFINE_SPINLOCK(bdi_lock);
39LIST_HEAD(bdi_list);
40LIST_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);
47
48void 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
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;
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: %10lu kB\n"
96 "BdiReclaimable: %10lu kB\n"
97 "BdiDirtyThresh: %10lu kB\n"
98 "DirtyThresh: %10lu kB\n"
99 "BackgroundThresh: %10lu kB\n"
100 "BdiWritten: %10lu kB\n"
101 "BdiWriteBandwidth: %10lu kBps\n"
102 "b_dirty: %10lu\n"
103 "b_io: %10lu\n"
104 "b_more_io: %10lu\n"
105 "bdi_list: %10u\n"
106 "state: %10lx\n",
107 (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
108 (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
109 K(bdi_thresh),
110 K(dirty_thresh),
111 K(background_thresh),
112 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
113 (unsigned long) K(bdi->write_bandwidth),
114 nr_dirty,
115 nr_io,
116 nr_more_io,
117 !list_empty(&bdi->bdi_list), bdi->state);
118#undef K
119
120 return 0;
121}
122
123static int bdi_debug_stats_open(struct inode *inode, struct file *file)
124{
125 return single_open(file, bdi_debug_stats_show, inode->i_private);
126}
127
128static const struct file_operations bdi_debug_stats_fops = {
129 .open = bdi_debug_stats_open,
130 .read = seq_read,
131 .llseek = seq_lseek,
132 .release = single_release,
133};
134
135static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
136{
137 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
138 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
139 bdi, &bdi_debug_stats_fops);
140}
141
142static void bdi_debug_unregister(struct backing_dev_info *bdi)
143{
144 debugfs_remove(bdi->debug_stats);
145 debugfs_remove(bdi->debug_dir);
146}
147#else
148static inline void bdi_debug_init(void)
149{
150}
151static inline void bdi_debug_register(struct backing_dev_info *bdi,
152 const char *name)
153{
154}
155static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
156{
157}
158#endif
159
160static ssize_t read_ahead_kb_store(struct device *dev,
161 struct device_attribute *attr,
162 const char *buf, size_t count)
163{
164 struct backing_dev_info *bdi = dev_get_drvdata(dev);
165 char *end;
166 unsigned long read_ahead_kb;
167 ssize_t ret = -EINVAL;
168
169 read_ahead_kb = simple_strtoul(buf, &end, 10);
170 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
171 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
172 ret = count;
173 }
174 return ret;
175}
176
177#define K(pages) ((pages) << (PAGE_SHIFT - 10))
178
179#define BDI_SHOW(name, expr) \
180static ssize_t name##_show(struct device *dev, \
181 struct device_attribute *attr, char *page) \
182{ \
183 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
184 \
185 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
186}
187
188BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
189
190static ssize_t min_ratio_store(struct device *dev,
191 struct device_attribute *attr, const char *buf, size_t count)
192{
193 struct backing_dev_info *bdi = dev_get_drvdata(dev);
194 char *end;
195 unsigned int ratio;
196 ssize_t ret = -EINVAL;
197
198 ratio = simple_strtoul(buf, &end, 10);
199 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
200 ret = bdi_set_min_ratio(bdi, ratio);
201 if (!ret)
202 ret = count;
203 }
204 return ret;
205}
206BDI_SHOW(min_ratio, bdi->min_ratio)
207
208static ssize_t max_ratio_store(struct device *dev,
209 struct device_attribute *attr, const char *buf, size_t count)
210{
211 struct backing_dev_info *bdi = dev_get_drvdata(dev);
212 char *end;
213 unsigned int ratio;
214 ssize_t ret = -EINVAL;
215
216 ratio = simple_strtoul(buf, &end, 10);
217 if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
218 ret = bdi_set_max_ratio(bdi, ratio);
219 if (!ret)
220 ret = count;
221 }
222 return ret;
223}
224BDI_SHOW(max_ratio, bdi->max_ratio)
225
226#define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
227
228static struct device_attribute bdi_dev_attrs[] = {
229 __ATTR_RW(read_ahead_kb),
230 __ATTR_RW(min_ratio),
231 __ATTR_RW(max_ratio),
232 __ATTR_NULL,
233};
234
235static __init int bdi_class_init(void)
236{
237 bdi_class = class_create(THIS_MODULE, "bdi");
238 if (IS_ERR(bdi_class))
239 return PTR_ERR(bdi_class);
240
241 bdi_class->dev_attrs = bdi_dev_attrs;
242 bdi_debug_init();
243 return 0;
244}
245postcore_initcall(bdi_class_init);
246
247static int __init default_bdi_init(void)
248{
249 int err;
250
251 sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
252 BUG_ON(IS_ERR(sync_supers_tsk));
253
254 setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
255 bdi_arm_supers_timer();
256
257 err = bdi_init(&default_backing_dev_info);
258 if (!err)
259 bdi_register(&default_backing_dev_info, NULL, "default");
260 err = bdi_init(&noop_backing_dev_info);
261
262 return err;
263}
264subsys_initcall(default_bdi_init);
265
266int bdi_has_dirty_io(struct backing_dev_info *bdi)
267{
268 return wb_has_dirty_io(&bdi->wb);
269}
270
271/*
272 * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
273 * or we risk deadlocking on ->s_umount. The longer term solution would be
274 * to implement sync_supers_bdi() or similar and simply do it from the
275 * bdi writeback thread individually.
276 */
277static int bdi_sync_supers(void *unused)
278{
279 set_user_nice(current, 0);
280
281 while (!kthread_should_stop()) {
282 set_current_state(TASK_INTERRUPTIBLE);
283 schedule();
284
285 /*
286 * Do this periodically, like kupdated() did before.
287 */
288 sync_supers();
289 }
290
291 return 0;
292}
293
294void bdi_arm_supers_timer(void)
295{
296 unsigned long next;
297
298 if (!dirty_writeback_interval)
299 return;
300
301 next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
302 mod_timer(&sync_supers_timer, round_jiffies_up(next));
303}
304
305static void sync_supers_timer_fn(unsigned long unused)
306{
307 wake_up_process(sync_supers_tsk);
308 bdi_arm_supers_timer();
309}
310
311static void wakeup_timer_fn(unsigned long data)
312{
313 struct backing_dev_info *bdi = (struct backing_dev_info *)data;
314
315 spin_lock_bh(&bdi->wb_lock);
316 if (bdi->wb.task) {
317 trace_writeback_wake_thread(bdi);
318 wake_up_process(bdi->wb.task);
319 } else {
320 /*
321 * When bdi tasks are inactive for long time, they are killed.
322 * In this case we have to wake-up the forker thread which
323 * should create and run the bdi thread.
324 */
325 trace_writeback_wake_forker_thread(bdi);
326 wake_up_process(default_backing_dev_info.wb.task);
327 }
328 spin_unlock_bh(&bdi->wb_lock);
329}
330
331/*
332 * This function is used when the first inode for this bdi is marked dirty. It
333 * wakes-up the corresponding bdi thread which should then take care of the
334 * periodic background write-out of dirty inodes. Since the write-out would
335 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
336 * set up a timer which wakes the bdi thread up later.
337 *
338 * Note, we wouldn't bother setting up the timer, but this function is on the
339 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
340 * by delaying the wake-up.
341 */
342void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
343{
344 unsigned long timeout;
345
346 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
347 mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
348}
349
350/*
351 * Calculate the longest interval (jiffies) bdi threads are allowed to be
352 * inactive.
353 */
354static unsigned long bdi_longest_inactive(void)
355{
356 unsigned long interval;
357
358 interval = msecs_to_jiffies(dirty_writeback_interval * 10);
359 return max(5UL * 60 * HZ, interval);
360}
361
362/*
363 * Clear pending bit and wakeup anybody waiting for flusher thread creation or
364 * shutdown
365 */
366static void bdi_clear_pending(struct backing_dev_info *bdi)
367{
368 clear_bit(BDI_pending, &bdi->state);
369 smp_mb__after_clear_bit();
370 wake_up_bit(&bdi->state, BDI_pending);
371}
372
373static int bdi_forker_thread(void *ptr)
374{
375 struct bdi_writeback *me = ptr;
376
377 current->flags |= PF_SWAPWRITE;
378 set_freezable();
379
380 /*
381 * Our parent may run at a different priority, just set us to normal
382 */
383 set_user_nice(current, 0);
384
385 for (;;) {
386 struct task_struct *task = NULL;
387 struct backing_dev_info *bdi;
388 enum {
389 NO_ACTION, /* Nothing to do */
390 FORK_THREAD, /* Fork bdi thread */
391 KILL_THREAD, /* Kill inactive bdi thread */
392 } action = NO_ACTION;
393
394 /*
395 * Temporary measure, we want to make sure we don't see
396 * dirty data on the default backing_dev_info
397 */
398 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
399 del_timer(&me->wakeup_timer);
400 wb_do_writeback(me, 0);
401 }
402
403 spin_lock_bh(&bdi_lock);
404 /*
405 * In the following loop we are going to check whether we have
406 * some work to do without any synchronization with tasks
407 * waking us up to do work for them. So we have to set task
408 * state already here so that we don't miss wakeups coming
409 * after we verify some condition.
410 */
411 set_current_state(TASK_INTERRUPTIBLE);
412
413 list_for_each_entry(bdi, &bdi_list, bdi_list) {
414 bool have_dirty_io;
415
416 if (!bdi_cap_writeback_dirty(bdi) ||
417 bdi_cap_flush_forker(bdi))
418 continue;
419
420 WARN(!test_bit(BDI_registered, &bdi->state),
421 "bdi %p/%s is not registered!\n", bdi, bdi->name);
422
423 have_dirty_io = !list_empty(&bdi->work_list) ||
424 wb_has_dirty_io(&bdi->wb);
425
426 /*
427 * If the bdi has work to do, but the thread does not
428 * exist - create it.
429 */
430 if (!bdi->wb.task && have_dirty_io) {
431 /*
432 * Set the pending bit - if someone will try to
433 * unregister this bdi - it'll wait on this bit.
434 */
435 set_bit(BDI_pending, &bdi->state);
436 action = FORK_THREAD;
437 break;
438 }
439
440 spin_lock(&bdi->wb_lock);
441
442 /*
443 * If there is no work to do and the bdi thread was
444 * inactive long enough - kill it. The wb_lock is taken
445 * to make sure no-one adds more work to this bdi and
446 * wakes the bdi thread up.
447 */
448 if (bdi->wb.task && !have_dirty_io &&
449 time_after(jiffies, bdi->wb.last_active +
450 bdi_longest_inactive())) {
451 task = bdi->wb.task;
452 bdi->wb.task = NULL;
453 spin_unlock(&bdi->wb_lock);
454 set_bit(BDI_pending, &bdi->state);
455 action = KILL_THREAD;
456 break;
457 }
458 spin_unlock(&bdi->wb_lock);
459 }
460 spin_unlock_bh(&bdi_lock);
461
462 /* Keep working if default bdi still has things to do */
463 if (!list_empty(&me->bdi->work_list))
464 __set_current_state(TASK_RUNNING);
465
466 switch (action) {
467 case FORK_THREAD:
468 __set_current_state(TASK_RUNNING);
469 task = kthread_create(bdi_writeback_thread, &bdi->wb,
470 "flush-%s", dev_name(bdi->dev));
471 if (IS_ERR(task)) {
472 /*
473 * If thread creation fails, force writeout of
474 * the bdi from the thread. Hopefully 1024 is
475 * large enough for efficient IO.
476 */
477 writeback_inodes_wb(&bdi->wb, 1024);
478 } else {
479 /*
480 * The spinlock makes sure we do not lose
481 * wake-ups when racing with 'bdi_queue_work()'.
482 * And as soon as the bdi thread is visible, we
483 * can start it.
484 */
485 spin_lock_bh(&bdi->wb_lock);
486 bdi->wb.task = task;
487 spin_unlock_bh(&bdi->wb_lock);
488 wake_up_process(task);
489 }
490 bdi_clear_pending(bdi);
491 break;
492
493 case KILL_THREAD:
494 __set_current_state(TASK_RUNNING);
495 kthread_stop(task);
496 bdi_clear_pending(bdi);
497 break;
498
499 case NO_ACTION:
500 if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
501 /*
502 * There are no dirty data. The only thing we
503 * should now care about is checking for
504 * inactive bdi threads and killing them. Thus,
505 * let's sleep for longer time, save energy and
506 * be friendly for battery-driven devices.
507 */
508 schedule_timeout(bdi_longest_inactive());
509 else
510 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
511 try_to_freeze();
512 break;
513 }
514 }
515
516 return 0;
517}
518
519/*
520 * Remove bdi from bdi_list, and ensure that it is no longer visible
521 */
522static void bdi_remove_from_list(struct backing_dev_info *bdi)
523{
524 spin_lock_bh(&bdi_lock);
525 list_del_rcu(&bdi->bdi_list);
526 spin_unlock_bh(&bdi_lock);
527
528 synchronize_rcu_expedited();
529}
530
531int bdi_register(struct backing_dev_info *bdi, struct device *parent,
532 const char *fmt, ...)
533{
534 va_list args;
535 struct device *dev;
536
537 if (bdi->dev) /* The driver needs to use separate queues per device */
538 return 0;
539
540 va_start(args, fmt);
541 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
542 va_end(args);
543 if (IS_ERR(dev))
544 return PTR_ERR(dev);
545
546 bdi->dev = dev;
547
548 /*
549 * Just start the forker thread for our default backing_dev_info,
550 * and add other bdi's to the list. They will get a thread created
551 * on-demand when they need it.
552 */
553 if (bdi_cap_flush_forker(bdi)) {
554 struct bdi_writeback *wb = &bdi->wb;
555
556 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
557 dev_name(dev));
558 if (IS_ERR(wb->task))
559 return PTR_ERR(wb->task);
560 }
561
562 bdi_debug_register(bdi, dev_name(dev));
563 set_bit(BDI_registered, &bdi->state);
564
565 spin_lock_bh(&bdi_lock);
566 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
567 spin_unlock_bh(&bdi_lock);
568
569 trace_writeback_bdi_register(bdi);
570 return 0;
571}
572EXPORT_SYMBOL(bdi_register);
573
574int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
575{
576 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
577}
578EXPORT_SYMBOL(bdi_register_dev);
579
580/*
581 * Remove bdi from the global list and shutdown any threads we have running
582 */
583static void bdi_wb_shutdown(struct backing_dev_info *bdi)
584{
585 if (!bdi_cap_writeback_dirty(bdi))
586 return;
587
588 /*
589 * Make sure nobody finds us on the bdi_list anymore
590 */
591 bdi_remove_from_list(bdi);
592
593 /*
594 * If setup is pending, wait for that to complete first
595 */
596 wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
597 TASK_UNINTERRUPTIBLE);
598
599 /*
600 * Finally, kill the kernel thread. We don't need to be RCU
601 * safe anymore, since the bdi is gone from visibility. Force
602 * unfreeze of the thread before calling kthread_stop(), otherwise
603 * it would never exet if it is currently stuck in the refrigerator.
604 */
605 if (bdi->wb.task) {
606 thaw_process(bdi->wb.task);
607 kthread_stop(bdi->wb.task);
608 }
609}
610
611/*
612 * This bdi is going away now, make sure that no super_blocks point to it
613 */
614static void bdi_prune_sb(struct backing_dev_info *bdi)
615{
616 struct super_block *sb;
617
618 spin_lock(&sb_lock);
619 list_for_each_entry(sb, &super_blocks, s_list) {
620 if (sb->s_bdi == bdi)
621 sb->s_bdi = &default_backing_dev_info;
622 }
623 spin_unlock(&sb_lock);
624}
625
626void bdi_unregister(struct backing_dev_info *bdi)
627{
628 if (bdi->dev) {
629 bdi_set_min_ratio(bdi, 0);
630 trace_writeback_bdi_unregister(bdi);
631 bdi_prune_sb(bdi);
632 del_timer_sync(&bdi->wb.wakeup_timer);
633
634 if (!bdi_cap_flush_forker(bdi))
635 bdi_wb_shutdown(bdi);
636 bdi_debug_unregister(bdi);
637 device_unregister(bdi->dev);
638 bdi->dev = NULL;
639 }
640}
641EXPORT_SYMBOL(bdi_unregister);
642
643static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
644{
645 memset(wb, 0, sizeof(*wb));
646
647 wb->bdi = bdi;
648 wb->last_old_flush = jiffies;
649 INIT_LIST_HEAD(&wb->b_dirty);
650 INIT_LIST_HEAD(&wb->b_io);
651 INIT_LIST_HEAD(&wb->b_more_io);
652 spin_lock_init(&wb->list_lock);
653 setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
654}
655
656/*
657 * Initial write bandwidth: 100 MB/s
658 */
659#define INIT_BW (100 << (20 - PAGE_SHIFT))
660
661int bdi_init(struct backing_dev_info *bdi)
662{
663 int i, err;
664
665 bdi->dev = NULL;
666
667 bdi->min_ratio = 0;
668 bdi->max_ratio = 100;
669 bdi->max_prop_frac = PROP_FRAC_BASE;
670 spin_lock_init(&bdi->wb_lock);
671 INIT_LIST_HEAD(&bdi->bdi_list);
672 INIT_LIST_HEAD(&bdi->work_list);
673
674 bdi_wb_init(&bdi->wb, bdi);
675
676 for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
677 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
678 if (err)
679 goto err;
680 }
681
682 bdi->dirty_exceeded = 0;
683
684 bdi->bw_time_stamp = jiffies;
685 bdi->written_stamp = 0;
686
687 bdi->write_bandwidth = INIT_BW;
688 bdi->avg_write_bandwidth = INIT_BW;
689
690 err = prop_local_init_percpu(&bdi->completions);
691
692 if (err) {
693err:
694 while (i--)
695 percpu_counter_destroy(&bdi->bdi_stat[i]);
696 }
697
698 return err;
699}
700EXPORT_SYMBOL(bdi_init);
701
702void bdi_destroy(struct backing_dev_info *bdi)
703{
704 int i;
705
706 /*
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
713 bdi_lock_two(&bdi->wb, dst);
714 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);
717 spin_unlock(&bdi->wb.list_lock);
718 spin_unlock(&dst->list_lock);
719 }
720
721 bdi_unregister(bdi);
722
723 for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
724 percpu_counter_destroy(&bdi->bdi_stat[i]);
725
726 prop_local_destroy_percpu(&bdi->completions);
727}
728EXPORT_SYMBOL(bdi_destroy);
729
730/*
731 * For use from filesystems to quickly init and register a bdi associated
732 * with dirty writeback
733 */
734int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
735 unsigned int cap)
736{
737 char tmp[32];
738 int err;
739
740 bdi->name = name;
741 bdi->capabilities = cap;
742 err = bdi_init(bdi);
743 if (err)
744 return err;
745
746 sprintf(tmp, "%.28s%s", name, "-%d");
747 err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
748 if (err) {
749 bdi_destroy(bdi);
750 return err;
751 }
752
753 return 0;
754}
755EXPORT_SYMBOL(bdi_setup_and_register);
756
757static wait_queue_head_t congestion_wqh[2] = {
758 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
759 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
760 };
761static atomic_t nr_bdi_congested[2];
762
763void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
764{
765 enum bdi_state bit;
766 wait_queue_head_t *wqh = &congestion_wqh[sync];
767
768 bit = sync ? BDI_sync_congested : BDI_async_congested;
769 if (test_and_clear_bit(bit, &bdi->state))
770 atomic_dec(&nr_bdi_congested[sync]);
771 smp_mb__after_clear_bit();
772 if (waitqueue_active(wqh))
773 wake_up(wqh);
774}
775EXPORT_SYMBOL(clear_bdi_congested);
776
777void set_bdi_congested(struct backing_dev_info *bdi, int sync)
778{
779 enum bdi_state bit;
780
781 bit = sync ? BDI_sync_congested : BDI_async_congested;
782 if (!test_and_set_bit(bit, &bdi->state))
783 atomic_inc(&nr_bdi_congested[sync]);
784}
785EXPORT_SYMBOL(set_bdi_congested);
786
787/**
788 * congestion_wait - wait for a backing_dev to become uncongested
789 * @sync: SYNC or ASYNC IO
790 * @timeout: timeout in jiffies
791 *
792 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
793 * write congestion. If no backing_devs are congested then just wait for the
794 * next write to be completed.
795 */
796long congestion_wait(int sync, long timeout)
797{
798 long ret;
799 unsigned long start = jiffies;
800 DEFINE_WAIT(wait);
801 wait_queue_head_t *wqh = &congestion_wqh[sync];
802
803 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
804 ret = io_schedule_timeout(timeout);
805 finish_wait(wqh, &wait);
806
807 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
808 jiffies_to_usecs(jiffies - start));
809
810 return ret;
811}
812EXPORT_SYMBOL(congestion_wait);
813
814/**
815 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
816 * @zone: A zone to check if it is heavily congested
817 * @sync: SYNC or ASYNC IO
818 * @timeout: timeout in jiffies
819 *
820 * In the event of a congested backing_dev (any backing_dev) and the given
821 * @zone has experienced recent congestion, this waits for up to @timeout
822 * jiffies for either a BDI to exit congestion of the given @sync queue
823 * or a write to complete.
824 *
825 * In the absence of zone congestion, cond_resched() is called to yield
826 * the processor if necessary but otherwise does not sleep.
827 *
828 * The return value is 0 if the sleep is for the full timeout. Otherwise,
829 * it is the number of jiffies that were still remaining when the function
830 * returned. return_value == timeout implies the function did not sleep.
831 */
832long wait_iff_congested(struct zone *zone, int sync, long timeout)
833{
834 long ret;
835 unsigned long start = jiffies;
836 DEFINE_WAIT(wait);
837 wait_queue_head_t *wqh = &congestion_wqh[sync];
838
839 /*
840 * If there is no congestion, or heavy congestion is not being
841 * encountered in the current zone, yield if necessary instead
842 * of sleeping on the congestion queue
843 */
844 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
845 !zone_is_reclaim_congested(zone)) {
846 cond_resched();
847
848 /* In case we scheduled, work out time remaining */
849 ret = timeout - (jiffies - start);
850 if (ret < 0)
851 ret = 0;
852
853 goto out;
854 }
855
856 /* Sleep until uncongested or a write happens */
857 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
858 ret = io_schedule_timeout(timeout);
859 finish_wait(wqh, &wait);
860
861out:
862 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
863 jiffies_to_usecs(jiffies - start));
864
865 return ret;
866}
867EXPORT_SYMBOL(wait_iff_congested);
1// SPDX-License-Identifier: GPL-2.0-only
2
3#include <linux/blkdev.h>
4#include <linux/wait.h>
5#include <linux/rbtree.h>
6#include <linux/kthread.h>
7#include <linux/backing-dev.h>
8#include <linux/blk-cgroup.h>
9#include <linux/freezer.h>
10#include <linux/fs.h>
11#include <linux/pagemap.h>
12#include <linux/mm.h>
13#include <linux/sched/mm.h>
14#include <linux/sched.h>
15#include <linux/module.h>
16#include <linux/writeback.h>
17#include <linux/device.h>
18#include <trace/events/writeback.h>
19
20struct backing_dev_info noop_backing_dev_info;
21EXPORT_SYMBOL_GPL(noop_backing_dev_info);
22
23static struct class *bdi_class;
24static const char *bdi_unknown_name = "(unknown)";
25
26/*
27 * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
28 * reader side locking.
29 */
30DEFINE_SPINLOCK(bdi_lock);
31static u64 bdi_id_cursor;
32static struct rb_root bdi_tree = RB_ROOT;
33LIST_HEAD(bdi_list);
34
35/* bdi_wq serves all asynchronous writeback tasks */
36struct workqueue_struct *bdi_wq;
37
38#define K(x) ((x) << (PAGE_SHIFT - 10))
39
40#ifdef CONFIG_DEBUG_FS
41#include <linux/debugfs.h>
42#include <linux/seq_file.h>
43
44static struct dentry *bdi_debug_root;
45
46static void bdi_debug_init(void)
47{
48 bdi_debug_root = debugfs_create_dir("bdi", NULL);
49}
50
51static int bdi_debug_stats_show(struct seq_file *m, void *v)
52{
53 struct backing_dev_info *bdi = m->private;
54 struct bdi_writeback *wb = &bdi->wb;
55 unsigned long background_thresh;
56 unsigned long dirty_thresh;
57 unsigned long wb_thresh;
58 unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
59 struct inode *inode;
60
61 nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
62 spin_lock(&wb->list_lock);
63 list_for_each_entry(inode, &wb->b_dirty, i_io_list)
64 nr_dirty++;
65 list_for_each_entry(inode, &wb->b_io, i_io_list)
66 nr_io++;
67 list_for_each_entry(inode, &wb->b_more_io, i_io_list)
68 nr_more_io++;
69 list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
70 if (inode->i_state & I_DIRTY_TIME)
71 nr_dirty_time++;
72 spin_unlock(&wb->list_lock);
73
74 global_dirty_limits(&background_thresh, &dirty_thresh);
75 wb_thresh = wb_calc_thresh(wb, dirty_thresh);
76
77 seq_printf(m,
78 "BdiWriteback: %10lu kB\n"
79 "BdiReclaimable: %10lu kB\n"
80 "BdiDirtyThresh: %10lu kB\n"
81 "DirtyThresh: %10lu kB\n"
82 "BackgroundThresh: %10lu kB\n"
83 "BdiDirtied: %10lu kB\n"
84 "BdiWritten: %10lu kB\n"
85 "BdiWriteBandwidth: %10lu kBps\n"
86 "b_dirty: %10lu\n"
87 "b_io: %10lu\n"
88 "b_more_io: %10lu\n"
89 "b_dirty_time: %10lu\n"
90 "bdi_list: %10u\n"
91 "state: %10lx\n",
92 (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
93 (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
94 K(wb_thresh),
95 K(dirty_thresh),
96 K(background_thresh),
97 (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
98 (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
99 (unsigned long) K(wb->write_bandwidth),
100 nr_dirty,
101 nr_io,
102 nr_more_io,
103 nr_dirty_time,
104 !list_empty(&bdi->bdi_list), bdi->wb.state);
105
106 return 0;
107}
108DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
109
110static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
111{
112 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
113
114 debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
115 &bdi_debug_stats_fops);
116}
117
118static void bdi_debug_unregister(struct backing_dev_info *bdi)
119{
120 debugfs_remove_recursive(bdi->debug_dir);
121}
122#else
123static inline void bdi_debug_init(void)
124{
125}
126static inline void bdi_debug_register(struct backing_dev_info *bdi,
127 const char *name)
128{
129}
130static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
131{
132}
133#endif
134
135static ssize_t read_ahead_kb_store(struct device *dev,
136 struct device_attribute *attr,
137 const char *buf, size_t count)
138{
139 struct backing_dev_info *bdi = dev_get_drvdata(dev);
140 unsigned long read_ahead_kb;
141 ssize_t ret;
142
143 ret = kstrtoul(buf, 10, &read_ahead_kb);
144 if (ret < 0)
145 return ret;
146
147 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
148
149 return count;
150}
151
152#define BDI_SHOW(name, expr) \
153static ssize_t name##_show(struct device *dev, \
154 struct device_attribute *attr, char *buf) \
155{ \
156 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
157 \
158 return sysfs_emit(buf, "%lld\n", (long long)expr); \
159} \
160static DEVICE_ATTR_RW(name);
161
162BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
163
164static ssize_t min_ratio_store(struct device *dev,
165 struct device_attribute *attr, const char *buf, size_t count)
166{
167 struct backing_dev_info *bdi = dev_get_drvdata(dev);
168 unsigned int ratio;
169 ssize_t ret;
170
171 ret = kstrtouint(buf, 10, &ratio);
172 if (ret < 0)
173 return ret;
174
175 ret = bdi_set_min_ratio(bdi, ratio);
176 if (!ret)
177 ret = count;
178
179 return ret;
180}
181BDI_SHOW(min_ratio, bdi->min_ratio / BDI_RATIO_SCALE)
182
183static ssize_t min_ratio_fine_store(struct device *dev,
184 struct device_attribute *attr, const char *buf, size_t count)
185{
186 struct backing_dev_info *bdi = dev_get_drvdata(dev);
187 unsigned int ratio;
188 ssize_t ret;
189
190 ret = kstrtouint(buf, 10, &ratio);
191 if (ret < 0)
192 return ret;
193
194 ret = bdi_set_min_ratio_no_scale(bdi, ratio);
195 if (!ret)
196 ret = count;
197
198 return ret;
199}
200BDI_SHOW(min_ratio_fine, bdi->min_ratio)
201
202static ssize_t max_ratio_store(struct device *dev,
203 struct device_attribute *attr, const char *buf, size_t count)
204{
205 struct backing_dev_info *bdi = dev_get_drvdata(dev);
206 unsigned int ratio;
207 ssize_t ret;
208
209 ret = kstrtouint(buf, 10, &ratio);
210 if (ret < 0)
211 return ret;
212
213 ret = bdi_set_max_ratio(bdi, ratio);
214 if (!ret)
215 ret = count;
216
217 return ret;
218}
219BDI_SHOW(max_ratio, bdi->max_ratio / BDI_RATIO_SCALE)
220
221static ssize_t max_ratio_fine_store(struct device *dev,
222 struct device_attribute *attr, const char *buf, size_t count)
223{
224 struct backing_dev_info *bdi = dev_get_drvdata(dev);
225 unsigned int ratio;
226 ssize_t ret;
227
228 ret = kstrtouint(buf, 10, &ratio);
229 if (ret < 0)
230 return ret;
231
232 ret = bdi_set_max_ratio_no_scale(bdi, ratio);
233 if (!ret)
234 ret = count;
235
236 return ret;
237}
238BDI_SHOW(max_ratio_fine, bdi->max_ratio)
239
240static ssize_t min_bytes_show(struct device *dev,
241 struct device_attribute *attr,
242 char *buf)
243{
244 struct backing_dev_info *bdi = dev_get_drvdata(dev);
245
246 return sysfs_emit(buf, "%llu\n", bdi_get_min_bytes(bdi));
247}
248
249static ssize_t min_bytes_store(struct device *dev,
250 struct device_attribute *attr, const char *buf, size_t count)
251{
252 struct backing_dev_info *bdi = dev_get_drvdata(dev);
253 u64 bytes;
254 ssize_t ret;
255
256 ret = kstrtoull(buf, 10, &bytes);
257 if (ret < 0)
258 return ret;
259
260 ret = bdi_set_min_bytes(bdi, bytes);
261 if (!ret)
262 ret = count;
263
264 return ret;
265}
266DEVICE_ATTR_RW(min_bytes);
267
268static ssize_t max_bytes_show(struct device *dev,
269 struct device_attribute *attr,
270 char *buf)
271{
272 struct backing_dev_info *bdi = dev_get_drvdata(dev);
273
274 return sysfs_emit(buf, "%llu\n", bdi_get_max_bytes(bdi));
275}
276
277static ssize_t max_bytes_store(struct device *dev,
278 struct device_attribute *attr, const char *buf, size_t count)
279{
280 struct backing_dev_info *bdi = dev_get_drvdata(dev);
281 u64 bytes;
282 ssize_t ret;
283
284 ret = kstrtoull(buf, 10, &bytes);
285 if (ret < 0)
286 return ret;
287
288 ret = bdi_set_max_bytes(bdi, bytes);
289 if (!ret)
290 ret = count;
291
292 return ret;
293}
294DEVICE_ATTR_RW(max_bytes);
295
296static ssize_t stable_pages_required_show(struct device *dev,
297 struct device_attribute *attr,
298 char *buf)
299{
300 dev_warn_once(dev,
301 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
302 return sysfs_emit(buf, "%d\n", 0);
303}
304static DEVICE_ATTR_RO(stable_pages_required);
305
306static ssize_t strict_limit_store(struct device *dev,
307 struct device_attribute *attr, const char *buf, size_t count)
308{
309 struct backing_dev_info *bdi = dev_get_drvdata(dev);
310 unsigned int strict_limit;
311 ssize_t ret;
312
313 ret = kstrtouint(buf, 10, &strict_limit);
314 if (ret < 0)
315 return ret;
316
317 ret = bdi_set_strict_limit(bdi, strict_limit);
318 if (!ret)
319 ret = count;
320
321 return ret;
322}
323
324static ssize_t strict_limit_show(struct device *dev,
325 struct device_attribute *attr, char *buf)
326{
327 struct backing_dev_info *bdi = dev_get_drvdata(dev);
328
329 return sysfs_emit(buf, "%d\n",
330 !!(bdi->capabilities & BDI_CAP_STRICTLIMIT));
331}
332static DEVICE_ATTR_RW(strict_limit);
333
334static struct attribute *bdi_dev_attrs[] = {
335 &dev_attr_read_ahead_kb.attr,
336 &dev_attr_min_ratio.attr,
337 &dev_attr_min_ratio_fine.attr,
338 &dev_attr_max_ratio.attr,
339 &dev_attr_max_ratio_fine.attr,
340 &dev_attr_min_bytes.attr,
341 &dev_attr_max_bytes.attr,
342 &dev_attr_stable_pages_required.attr,
343 &dev_attr_strict_limit.attr,
344 NULL,
345};
346ATTRIBUTE_GROUPS(bdi_dev);
347
348static __init int bdi_class_init(void)
349{
350 bdi_class = class_create(THIS_MODULE, "bdi");
351 if (IS_ERR(bdi_class))
352 return PTR_ERR(bdi_class);
353
354 bdi_class->dev_groups = bdi_dev_groups;
355 bdi_debug_init();
356
357 return 0;
358}
359postcore_initcall(bdi_class_init);
360
361static int __init default_bdi_init(void)
362{
363 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
364 WQ_SYSFS, 0);
365 if (!bdi_wq)
366 return -ENOMEM;
367 return 0;
368}
369subsys_initcall(default_bdi_init);
370
371/*
372 * This function is used when the first inode for this wb is marked dirty. It
373 * wakes-up the corresponding bdi thread which should then take care of the
374 * periodic background write-out of dirty inodes. Since the write-out would
375 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
376 * set up a timer which wakes the bdi thread up later.
377 *
378 * Note, we wouldn't bother setting up the timer, but this function is on the
379 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
380 * by delaying the wake-up.
381 *
382 * We have to be careful not to postpone flush work if it is scheduled for
383 * earlier. Thus we use queue_delayed_work().
384 */
385void wb_wakeup_delayed(struct bdi_writeback *wb)
386{
387 unsigned long timeout;
388
389 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
390 spin_lock_irq(&wb->work_lock);
391 if (test_bit(WB_registered, &wb->state))
392 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
393 spin_unlock_irq(&wb->work_lock);
394}
395
396static void wb_update_bandwidth_workfn(struct work_struct *work)
397{
398 struct bdi_writeback *wb = container_of(to_delayed_work(work),
399 struct bdi_writeback, bw_dwork);
400
401 wb_update_bandwidth(wb);
402}
403
404/*
405 * Initial write bandwidth: 100 MB/s
406 */
407#define INIT_BW (100 << (20 - PAGE_SHIFT))
408
409static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
410 gfp_t gfp)
411{
412 int i, err;
413
414 memset(wb, 0, sizeof(*wb));
415
416 wb->bdi = bdi;
417 wb->last_old_flush = jiffies;
418 INIT_LIST_HEAD(&wb->b_dirty);
419 INIT_LIST_HEAD(&wb->b_io);
420 INIT_LIST_HEAD(&wb->b_more_io);
421 INIT_LIST_HEAD(&wb->b_dirty_time);
422 spin_lock_init(&wb->list_lock);
423
424 atomic_set(&wb->writeback_inodes, 0);
425 wb->bw_time_stamp = jiffies;
426 wb->balanced_dirty_ratelimit = INIT_BW;
427 wb->dirty_ratelimit = INIT_BW;
428 wb->write_bandwidth = INIT_BW;
429 wb->avg_write_bandwidth = INIT_BW;
430
431 spin_lock_init(&wb->work_lock);
432 INIT_LIST_HEAD(&wb->work_list);
433 INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
434 INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
435 wb->dirty_sleep = jiffies;
436
437 err = fprop_local_init_percpu(&wb->completions, gfp);
438 if (err)
439 return err;
440
441 for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
442 err = percpu_counter_init(&wb->stat[i], 0, gfp);
443 if (err)
444 goto out_destroy_stat;
445 }
446
447 return 0;
448
449out_destroy_stat:
450 while (i--)
451 percpu_counter_destroy(&wb->stat[i]);
452 fprop_local_destroy_percpu(&wb->completions);
453 return err;
454}
455
456static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
457
458/*
459 * Remove bdi from the global list and shutdown any threads we have running
460 */
461static void wb_shutdown(struct bdi_writeback *wb)
462{
463 /* Make sure nobody queues further work */
464 spin_lock_irq(&wb->work_lock);
465 if (!test_and_clear_bit(WB_registered, &wb->state)) {
466 spin_unlock_irq(&wb->work_lock);
467 return;
468 }
469 spin_unlock_irq(&wb->work_lock);
470
471 cgwb_remove_from_bdi_list(wb);
472 /*
473 * Drain work list and shutdown the delayed_work. !WB_registered
474 * tells wb_workfn() that @wb is dying and its work_list needs to
475 * be drained no matter what.
476 */
477 mod_delayed_work(bdi_wq, &wb->dwork, 0);
478 flush_delayed_work(&wb->dwork);
479 WARN_ON(!list_empty(&wb->work_list));
480 flush_delayed_work(&wb->bw_dwork);
481}
482
483static void wb_exit(struct bdi_writeback *wb)
484{
485 int i;
486
487 WARN_ON(delayed_work_pending(&wb->dwork));
488
489 for (i = 0; i < NR_WB_STAT_ITEMS; i++)
490 percpu_counter_destroy(&wb->stat[i]);
491
492 fprop_local_destroy_percpu(&wb->completions);
493}
494
495#ifdef CONFIG_CGROUP_WRITEBACK
496
497#include <linux/memcontrol.h>
498
499/*
500 * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
501 * memcg->cgwb_list. bdi->cgwb_tree is also RCU protected.
502 */
503static DEFINE_SPINLOCK(cgwb_lock);
504static struct workqueue_struct *cgwb_release_wq;
505
506static LIST_HEAD(offline_cgwbs);
507static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
508static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
509
510static void cgwb_release_workfn(struct work_struct *work)
511{
512 struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
513 release_work);
514 struct backing_dev_info *bdi = wb->bdi;
515
516 mutex_lock(&wb->bdi->cgwb_release_mutex);
517 wb_shutdown(wb);
518
519 css_put(wb->memcg_css);
520 css_put(wb->blkcg_css);
521 mutex_unlock(&wb->bdi->cgwb_release_mutex);
522
523 /* triggers blkg destruction if no online users left */
524 blkcg_unpin_online(wb->blkcg_css);
525
526 fprop_local_destroy_percpu(&wb->memcg_completions);
527
528 spin_lock_irq(&cgwb_lock);
529 list_del(&wb->offline_node);
530 spin_unlock_irq(&cgwb_lock);
531
532 percpu_ref_exit(&wb->refcnt);
533 wb_exit(wb);
534 bdi_put(bdi);
535 WARN_ON_ONCE(!list_empty(&wb->b_attached));
536 kfree_rcu(wb, rcu);
537}
538
539static void cgwb_release(struct percpu_ref *refcnt)
540{
541 struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
542 refcnt);
543 queue_work(cgwb_release_wq, &wb->release_work);
544}
545
546static void cgwb_kill(struct bdi_writeback *wb)
547{
548 lockdep_assert_held(&cgwb_lock);
549
550 WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
551 list_del(&wb->memcg_node);
552 list_del(&wb->blkcg_node);
553 list_add(&wb->offline_node, &offline_cgwbs);
554 percpu_ref_kill(&wb->refcnt);
555}
556
557static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
558{
559 spin_lock_irq(&cgwb_lock);
560 list_del_rcu(&wb->bdi_node);
561 spin_unlock_irq(&cgwb_lock);
562}
563
564static int cgwb_create(struct backing_dev_info *bdi,
565 struct cgroup_subsys_state *memcg_css, gfp_t gfp)
566{
567 struct mem_cgroup *memcg;
568 struct cgroup_subsys_state *blkcg_css;
569 struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
570 struct bdi_writeback *wb;
571 unsigned long flags;
572 int ret = 0;
573
574 memcg = mem_cgroup_from_css(memcg_css);
575 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
576 memcg_cgwb_list = &memcg->cgwb_list;
577 blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
578
579 /* look up again under lock and discard on blkcg mismatch */
580 spin_lock_irqsave(&cgwb_lock, flags);
581 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
582 if (wb && wb->blkcg_css != blkcg_css) {
583 cgwb_kill(wb);
584 wb = NULL;
585 }
586 spin_unlock_irqrestore(&cgwb_lock, flags);
587 if (wb)
588 goto out_put;
589
590 /* need to create a new one */
591 wb = kmalloc(sizeof(*wb), gfp);
592 if (!wb) {
593 ret = -ENOMEM;
594 goto out_put;
595 }
596
597 ret = wb_init(wb, bdi, gfp);
598 if (ret)
599 goto err_free;
600
601 ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
602 if (ret)
603 goto err_wb_exit;
604
605 ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
606 if (ret)
607 goto err_ref_exit;
608
609 wb->memcg_css = memcg_css;
610 wb->blkcg_css = blkcg_css;
611 INIT_LIST_HEAD(&wb->b_attached);
612 INIT_WORK(&wb->release_work, cgwb_release_workfn);
613 set_bit(WB_registered, &wb->state);
614 bdi_get(bdi);
615
616 /*
617 * The root wb determines the registered state of the whole bdi and
618 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
619 * whether they're still online. Don't link @wb if any is dead.
620 * See wb_memcg_offline() and wb_blkcg_offline().
621 */
622 ret = -ENODEV;
623 spin_lock_irqsave(&cgwb_lock, flags);
624 if (test_bit(WB_registered, &bdi->wb.state) &&
625 blkcg_cgwb_list->next && memcg_cgwb_list->next) {
626 /* we might have raced another instance of this function */
627 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
628 if (!ret) {
629 list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
630 list_add(&wb->memcg_node, memcg_cgwb_list);
631 list_add(&wb->blkcg_node, blkcg_cgwb_list);
632 blkcg_pin_online(blkcg_css);
633 css_get(memcg_css);
634 css_get(blkcg_css);
635 }
636 }
637 spin_unlock_irqrestore(&cgwb_lock, flags);
638 if (ret) {
639 if (ret == -EEXIST)
640 ret = 0;
641 goto err_fprop_exit;
642 }
643 goto out_put;
644
645err_fprop_exit:
646 bdi_put(bdi);
647 fprop_local_destroy_percpu(&wb->memcg_completions);
648err_ref_exit:
649 percpu_ref_exit(&wb->refcnt);
650err_wb_exit:
651 wb_exit(wb);
652err_free:
653 kfree(wb);
654out_put:
655 css_put(blkcg_css);
656 return ret;
657}
658
659/**
660 * wb_get_lookup - get wb for a given memcg
661 * @bdi: target bdi
662 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
663 *
664 * Try to get the wb for @memcg_css on @bdi. The returned wb has its
665 * refcount incremented.
666 *
667 * This function uses css_get() on @memcg_css and thus expects its refcnt
668 * to be positive on invocation. IOW, rcu_read_lock() protection on
669 * @memcg_css isn't enough. try_get it before calling this function.
670 *
671 * A wb is keyed by its associated memcg. As blkcg implicitly enables
672 * memcg on the default hierarchy, memcg association is guaranteed to be
673 * more specific (equal or descendant to the associated blkcg) and thus can
674 * identify both the memcg and blkcg associations.
675 *
676 * Because the blkcg associated with a memcg may change as blkcg is enabled
677 * and disabled closer to root in the hierarchy, each wb keeps track of
678 * both the memcg and blkcg associated with it and verifies the blkcg on
679 * each lookup. On mismatch, the existing wb is discarded and a new one is
680 * created.
681 */
682struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
683 struct cgroup_subsys_state *memcg_css)
684{
685 struct bdi_writeback *wb;
686
687 if (!memcg_css->parent)
688 return &bdi->wb;
689
690 rcu_read_lock();
691 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
692 if (wb) {
693 struct cgroup_subsys_state *blkcg_css;
694
695 /* see whether the blkcg association has changed */
696 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
697 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
698 wb = NULL;
699 css_put(blkcg_css);
700 }
701 rcu_read_unlock();
702
703 return wb;
704}
705
706/**
707 * wb_get_create - get wb for a given memcg, create if necessary
708 * @bdi: target bdi
709 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
710 * @gfp: allocation mask to use
711 *
712 * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to
713 * create one. See wb_get_lookup() for more details.
714 */
715struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
716 struct cgroup_subsys_state *memcg_css,
717 gfp_t gfp)
718{
719 struct bdi_writeback *wb;
720
721 might_alloc(gfp);
722
723 if (!memcg_css->parent)
724 return &bdi->wb;
725
726 do {
727 wb = wb_get_lookup(bdi, memcg_css);
728 } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
729
730 return wb;
731}
732
733static int cgwb_bdi_init(struct backing_dev_info *bdi)
734{
735 int ret;
736
737 INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
738 mutex_init(&bdi->cgwb_release_mutex);
739 init_rwsem(&bdi->wb_switch_rwsem);
740
741 ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
742 if (!ret) {
743 bdi->wb.memcg_css = &root_mem_cgroup->css;
744 bdi->wb.blkcg_css = blkcg_root_css;
745 }
746 return ret;
747}
748
749static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
750{
751 struct radix_tree_iter iter;
752 void **slot;
753 struct bdi_writeback *wb;
754
755 WARN_ON(test_bit(WB_registered, &bdi->wb.state));
756
757 spin_lock_irq(&cgwb_lock);
758 radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
759 cgwb_kill(*slot);
760 spin_unlock_irq(&cgwb_lock);
761
762 mutex_lock(&bdi->cgwb_release_mutex);
763 spin_lock_irq(&cgwb_lock);
764 while (!list_empty(&bdi->wb_list)) {
765 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
766 bdi_node);
767 spin_unlock_irq(&cgwb_lock);
768 wb_shutdown(wb);
769 spin_lock_irq(&cgwb_lock);
770 }
771 spin_unlock_irq(&cgwb_lock);
772 mutex_unlock(&bdi->cgwb_release_mutex);
773}
774
775/*
776 * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
777 *
778 * Try to release dying cgwbs by switching attached inodes to the nearest
779 * living ancestor's writeback. Processed wbs are placed at the end
780 * of the list to guarantee the forward progress.
781 */
782static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
783{
784 struct bdi_writeback *wb;
785 LIST_HEAD(processed);
786
787 spin_lock_irq(&cgwb_lock);
788
789 while (!list_empty(&offline_cgwbs)) {
790 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
791 offline_node);
792 list_move(&wb->offline_node, &processed);
793
794 /*
795 * If wb is dirty, cleaning up the writeback by switching
796 * attached inodes will result in an effective removal of any
797 * bandwidth restrictions, which isn't the goal. Instead,
798 * it can be postponed until the next time, when all io
799 * will be likely completed. If in the meantime some inodes
800 * will get re-dirtied, they should be eventually switched to
801 * a new cgwb.
802 */
803 if (wb_has_dirty_io(wb))
804 continue;
805
806 if (!wb_tryget(wb))
807 continue;
808
809 spin_unlock_irq(&cgwb_lock);
810 while (cleanup_offline_cgwb(wb))
811 cond_resched();
812 spin_lock_irq(&cgwb_lock);
813
814 wb_put(wb);
815 }
816
817 if (!list_empty(&processed))
818 list_splice_tail(&processed, &offline_cgwbs);
819
820 spin_unlock_irq(&cgwb_lock);
821}
822
823/**
824 * wb_memcg_offline - kill all wb's associated with a memcg being offlined
825 * @memcg: memcg being offlined
826 *
827 * Also prevents creation of any new wb's associated with @memcg.
828 */
829void wb_memcg_offline(struct mem_cgroup *memcg)
830{
831 struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
832 struct bdi_writeback *wb, *next;
833
834 spin_lock_irq(&cgwb_lock);
835 list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
836 cgwb_kill(wb);
837 memcg_cgwb_list->next = NULL; /* prevent new wb's */
838 spin_unlock_irq(&cgwb_lock);
839
840 queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
841}
842
843/**
844 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
845 * @css: blkcg being offlined
846 *
847 * Also prevents creation of any new wb's associated with @blkcg.
848 */
849void wb_blkcg_offline(struct cgroup_subsys_state *css)
850{
851 struct bdi_writeback *wb, *next;
852 struct list_head *list = blkcg_get_cgwb_list(css);
853
854 spin_lock_irq(&cgwb_lock);
855 list_for_each_entry_safe(wb, next, list, blkcg_node)
856 cgwb_kill(wb);
857 list->next = NULL; /* prevent new wb's */
858 spin_unlock_irq(&cgwb_lock);
859}
860
861static void cgwb_bdi_register(struct backing_dev_info *bdi)
862{
863 spin_lock_irq(&cgwb_lock);
864 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
865 spin_unlock_irq(&cgwb_lock);
866}
867
868static int __init cgwb_init(void)
869{
870 /*
871 * There can be many concurrent release work items overwhelming
872 * system_wq. Put them in a separate wq and limit concurrency.
873 * There's no point in executing many of these in parallel.
874 */
875 cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
876 if (!cgwb_release_wq)
877 return -ENOMEM;
878
879 return 0;
880}
881subsys_initcall(cgwb_init);
882
883#else /* CONFIG_CGROUP_WRITEBACK */
884
885static int cgwb_bdi_init(struct backing_dev_info *bdi)
886{
887 return wb_init(&bdi->wb, bdi, GFP_KERNEL);
888}
889
890static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
891
892static void cgwb_bdi_register(struct backing_dev_info *bdi)
893{
894 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
895}
896
897static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
898{
899 list_del_rcu(&wb->bdi_node);
900}
901
902#endif /* CONFIG_CGROUP_WRITEBACK */
903
904int bdi_init(struct backing_dev_info *bdi)
905{
906 bdi->dev = NULL;
907
908 kref_init(&bdi->refcnt);
909 bdi->min_ratio = 0;
910 bdi->max_ratio = 100 * BDI_RATIO_SCALE;
911 bdi->max_prop_frac = FPROP_FRAC_BASE;
912 INIT_LIST_HEAD(&bdi->bdi_list);
913 INIT_LIST_HEAD(&bdi->wb_list);
914 init_waitqueue_head(&bdi->wb_waitq);
915
916 return cgwb_bdi_init(bdi);
917}
918
919struct backing_dev_info *bdi_alloc(int node_id)
920{
921 struct backing_dev_info *bdi;
922
923 bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
924 if (!bdi)
925 return NULL;
926
927 if (bdi_init(bdi)) {
928 kfree(bdi);
929 return NULL;
930 }
931 bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
932 bdi->ra_pages = VM_READAHEAD_PAGES;
933 bdi->io_pages = VM_READAHEAD_PAGES;
934 timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
935 return bdi;
936}
937EXPORT_SYMBOL(bdi_alloc);
938
939static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
940{
941 struct rb_node **p = &bdi_tree.rb_node;
942 struct rb_node *parent = NULL;
943 struct backing_dev_info *bdi;
944
945 lockdep_assert_held(&bdi_lock);
946
947 while (*p) {
948 parent = *p;
949 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
950
951 if (bdi->id > id)
952 p = &(*p)->rb_left;
953 else if (bdi->id < id)
954 p = &(*p)->rb_right;
955 else
956 break;
957 }
958
959 if (parentp)
960 *parentp = parent;
961 return p;
962}
963
964/**
965 * bdi_get_by_id - lookup and get bdi from its id
966 * @id: bdi id to lookup
967 *
968 * Find bdi matching @id and get it. Returns NULL if the matching bdi
969 * doesn't exist or is already unregistered.
970 */
971struct backing_dev_info *bdi_get_by_id(u64 id)
972{
973 struct backing_dev_info *bdi = NULL;
974 struct rb_node **p;
975
976 spin_lock_bh(&bdi_lock);
977 p = bdi_lookup_rb_node(id, NULL);
978 if (*p) {
979 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
980 bdi_get(bdi);
981 }
982 spin_unlock_bh(&bdi_lock);
983
984 return bdi;
985}
986
987int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
988{
989 struct device *dev;
990 struct rb_node *parent, **p;
991
992 if (bdi->dev) /* The driver needs to use separate queues per device */
993 return 0;
994
995 vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
996 dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
997 if (IS_ERR(dev))
998 return PTR_ERR(dev);
999
1000 cgwb_bdi_register(bdi);
1001 bdi->dev = dev;
1002
1003 bdi_debug_register(bdi, dev_name(dev));
1004 set_bit(WB_registered, &bdi->wb.state);
1005
1006 spin_lock_bh(&bdi_lock);
1007
1008 bdi->id = ++bdi_id_cursor;
1009
1010 p = bdi_lookup_rb_node(bdi->id, &parent);
1011 rb_link_node(&bdi->rb_node, parent, p);
1012 rb_insert_color(&bdi->rb_node, &bdi_tree);
1013
1014 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
1015
1016 spin_unlock_bh(&bdi_lock);
1017
1018 trace_writeback_bdi_register(bdi);
1019 return 0;
1020}
1021
1022int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1023{
1024 va_list args;
1025 int ret;
1026
1027 va_start(args, fmt);
1028 ret = bdi_register_va(bdi, fmt, args);
1029 va_end(args);
1030 return ret;
1031}
1032EXPORT_SYMBOL(bdi_register);
1033
1034void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1035{
1036 WARN_ON_ONCE(bdi->owner);
1037 bdi->owner = owner;
1038 get_device(owner);
1039}
1040
1041/*
1042 * Remove bdi from bdi_list, and ensure that it is no longer visible
1043 */
1044static void bdi_remove_from_list(struct backing_dev_info *bdi)
1045{
1046 spin_lock_bh(&bdi_lock);
1047 rb_erase(&bdi->rb_node, &bdi_tree);
1048 list_del_rcu(&bdi->bdi_list);
1049 spin_unlock_bh(&bdi_lock);
1050
1051 synchronize_rcu_expedited();
1052}
1053
1054void bdi_unregister(struct backing_dev_info *bdi)
1055{
1056 del_timer_sync(&bdi->laptop_mode_wb_timer);
1057
1058 /* make sure nobody finds us on the bdi_list anymore */
1059 bdi_remove_from_list(bdi);
1060 wb_shutdown(&bdi->wb);
1061 cgwb_bdi_unregister(bdi);
1062
1063 /*
1064 * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1065 * update the global bdi_min_ratio.
1066 */
1067 if (bdi->min_ratio)
1068 bdi_set_min_ratio(bdi, 0);
1069
1070 if (bdi->dev) {
1071 bdi_debug_unregister(bdi);
1072 device_unregister(bdi->dev);
1073 bdi->dev = NULL;
1074 }
1075
1076 if (bdi->owner) {
1077 put_device(bdi->owner);
1078 bdi->owner = NULL;
1079 }
1080}
1081EXPORT_SYMBOL(bdi_unregister);
1082
1083static void release_bdi(struct kref *ref)
1084{
1085 struct backing_dev_info *bdi =
1086 container_of(ref, struct backing_dev_info, refcnt);
1087
1088 WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1089 WARN_ON_ONCE(bdi->dev);
1090 wb_exit(&bdi->wb);
1091 kfree(bdi);
1092}
1093
1094void bdi_put(struct backing_dev_info *bdi)
1095{
1096 kref_put(&bdi->refcnt, release_bdi);
1097}
1098EXPORT_SYMBOL(bdi_put);
1099
1100struct backing_dev_info *inode_to_bdi(struct inode *inode)
1101{
1102 struct super_block *sb;
1103
1104 if (!inode)
1105 return &noop_backing_dev_info;
1106
1107 sb = inode->i_sb;
1108#ifdef CONFIG_BLOCK
1109 if (sb_is_blkdev_sb(sb))
1110 return I_BDEV(inode)->bd_disk->bdi;
1111#endif
1112 return sb->s_bdi;
1113}
1114EXPORT_SYMBOL(inode_to_bdi);
1115
1116const char *bdi_dev_name(struct backing_dev_info *bdi)
1117{
1118 if (!bdi || !bdi->dev)
1119 return bdi_unknown_name;
1120 return bdi->dev_name;
1121}
1122EXPORT_SYMBOL_GPL(bdi_dev_name);