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