xref: /openbmc/linux/mm/backing-dev.c (revision f0931824)
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 #include "internal.h"
20 
21 struct backing_dev_info noop_backing_dev_info;
22 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
23 
24 static 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  */
30 DEFINE_SPINLOCK(bdi_lock);
31 static u64 bdi_id_cursor;
32 static struct rb_root bdi_tree = RB_ROOT;
33 LIST_HEAD(bdi_list);
34 
35 /* bdi_wq serves all asynchronous writeback tasks */
36 struct workqueue_struct *bdi_wq;
37 
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/debugfs.h>
40 #include <linux/seq_file.h>
41 
42 static struct dentry *bdi_debug_root;
43 
44 static void bdi_debug_init(void)
45 {
46 	bdi_debug_root = debugfs_create_dir("bdi", NULL);
47 }
48 
49 static 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 }
106 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
107 
108 static 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 
116 static void bdi_debug_unregister(struct backing_dev_info *bdi)
117 {
118 	debugfs_remove_recursive(bdi->debug_dir);
119 }
120 #else
121 static inline void bdi_debug_init(void)
122 {
123 }
124 static inline void bdi_debug_register(struct backing_dev_info *bdi,
125 				      const char *name)
126 {
127 }
128 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
129 {
130 }
131 #endif
132 
133 static 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)						\
151 static 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 }									\
158 static DEVICE_ATTR_RW(name);
159 
160 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
161 
162 static 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 }
179 BDI_SHOW(min_ratio, bdi->min_ratio / BDI_RATIO_SCALE)
180 
181 static ssize_t min_ratio_fine_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_min_ratio_no_scale(bdi, ratio);
193 	if (!ret)
194 		ret = count;
195 
196 	return ret;
197 }
198 BDI_SHOW(min_ratio_fine, bdi->min_ratio)
199 
200 static ssize_t max_ratio_store(struct device *dev,
201 		struct device_attribute *attr, const char *buf, size_t count)
202 {
203 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
204 	unsigned int ratio;
205 	ssize_t ret;
206 
207 	ret = kstrtouint(buf, 10, &ratio);
208 	if (ret < 0)
209 		return ret;
210 
211 	ret = bdi_set_max_ratio(bdi, ratio);
212 	if (!ret)
213 		ret = count;
214 
215 	return ret;
216 }
217 BDI_SHOW(max_ratio, bdi->max_ratio / BDI_RATIO_SCALE)
218 
219 static ssize_t max_ratio_fine_store(struct device *dev,
220 		struct device_attribute *attr, const char *buf, size_t count)
221 {
222 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
223 	unsigned int ratio;
224 	ssize_t ret;
225 
226 	ret = kstrtouint(buf, 10, &ratio);
227 	if (ret < 0)
228 		return ret;
229 
230 	ret = bdi_set_max_ratio_no_scale(bdi, ratio);
231 	if (!ret)
232 		ret = count;
233 
234 	return ret;
235 }
236 BDI_SHOW(max_ratio_fine, bdi->max_ratio)
237 
238 static ssize_t min_bytes_show(struct device *dev,
239 			      struct device_attribute *attr,
240 			      char *buf)
241 {
242 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
243 
244 	return sysfs_emit(buf, "%llu\n", bdi_get_min_bytes(bdi));
245 }
246 
247 static ssize_t min_bytes_store(struct device *dev,
248 		struct device_attribute *attr, const char *buf, size_t count)
249 {
250 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
251 	u64 bytes;
252 	ssize_t ret;
253 
254 	ret = kstrtoull(buf, 10, &bytes);
255 	if (ret < 0)
256 		return ret;
257 
258 	ret = bdi_set_min_bytes(bdi, bytes);
259 	if (!ret)
260 		ret = count;
261 
262 	return ret;
263 }
264 static DEVICE_ATTR_RW(min_bytes);
265 
266 static ssize_t max_bytes_show(struct device *dev,
267 			      struct device_attribute *attr,
268 			      char *buf)
269 {
270 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
271 
272 	return sysfs_emit(buf, "%llu\n", bdi_get_max_bytes(bdi));
273 }
274 
275 static ssize_t max_bytes_store(struct device *dev,
276 		struct device_attribute *attr, const char *buf, size_t count)
277 {
278 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
279 	u64 bytes;
280 	ssize_t ret;
281 
282 	ret = kstrtoull(buf, 10, &bytes);
283 	if (ret < 0)
284 		return ret;
285 
286 	ret = bdi_set_max_bytes(bdi, bytes);
287 	if (!ret)
288 		ret = count;
289 
290 	return ret;
291 }
292 static DEVICE_ATTR_RW(max_bytes);
293 
294 static ssize_t stable_pages_required_show(struct device *dev,
295 					  struct device_attribute *attr,
296 					  char *buf)
297 {
298 	dev_warn_once(dev,
299 		"the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
300 	return sysfs_emit(buf, "%d\n", 0);
301 }
302 static DEVICE_ATTR_RO(stable_pages_required);
303 
304 static ssize_t strict_limit_store(struct device *dev,
305 		struct device_attribute *attr, const char *buf, size_t count)
306 {
307 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
308 	unsigned int strict_limit;
309 	ssize_t ret;
310 
311 	ret = kstrtouint(buf, 10, &strict_limit);
312 	if (ret < 0)
313 		return ret;
314 
315 	ret = bdi_set_strict_limit(bdi, strict_limit);
316 	if (!ret)
317 		ret = count;
318 
319 	return ret;
320 }
321 
322 static ssize_t strict_limit_show(struct device *dev,
323 		struct device_attribute *attr, char *buf)
324 {
325 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
326 
327 	return sysfs_emit(buf, "%d\n",
328 			!!(bdi->capabilities & BDI_CAP_STRICTLIMIT));
329 }
330 static DEVICE_ATTR_RW(strict_limit);
331 
332 static struct attribute *bdi_dev_attrs[] = {
333 	&dev_attr_read_ahead_kb.attr,
334 	&dev_attr_min_ratio.attr,
335 	&dev_attr_min_ratio_fine.attr,
336 	&dev_attr_max_ratio.attr,
337 	&dev_attr_max_ratio_fine.attr,
338 	&dev_attr_min_bytes.attr,
339 	&dev_attr_max_bytes.attr,
340 	&dev_attr_stable_pages_required.attr,
341 	&dev_attr_strict_limit.attr,
342 	NULL,
343 };
344 ATTRIBUTE_GROUPS(bdi_dev);
345 
346 static const struct class bdi_class = {
347 	.name		= "bdi",
348 	.dev_groups	= bdi_dev_groups,
349 };
350 
351 static __init int bdi_class_init(void)
352 {
353 	int ret;
354 
355 	ret = class_register(&bdi_class);
356 	if (ret)
357 		return ret;
358 
359 	bdi_debug_init();
360 
361 	return 0;
362 }
363 postcore_initcall(bdi_class_init);
364 
365 static int __init default_bdi_init(void)
366 {
367 	bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
368 				 WQ_SYSFS, 0);
369 	if (!bdi_wq)
370 		return -ENOMEM;
371 	return 0;
372 }
373 subsys_initcall(default_bdi_init);
374 
375 /*
376  * This function is used when the first inode for this wb is marked dirty. It
377  * wakes-up the corresponding bdi thread which should then take care of the
378  * periodic background write-out of dirty inodes. Since the write-out would
379  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
380  * set up a timer which wakes the bdi thread up later.
381  *
382  * Note, we wouldn't bother setting up the timer, but this function is on the
383  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
384  * by delaying the wake-up.
385  *
386  * We have to be careful not to postpone flush work if it is scheduled for
387  * earlier. Thus we use queue_delayed_work().
388  */
389 void wb_wakeup_delayed(struct bdi_writeback *wb)
390 {
391 	unsigned long timeout;
392 
393 	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
394 	spin_lock_irq(&wb->work_lock);
395 	if (test_bit(WB_registered, &wb->state))
396 		queue_delayed_work(bdi_wq, &wb->dwork, timeout);
397 	spin_unlock_irq(&wb->work_lock);
398 }
399 
400 static void wb_update_bandwidth_workfn(struct work_struct *work)
401 {
402 	struct bdi_writeback *wb = container_of(to_delayed_work(work),
403 						struct bdi_writeback, bw_dwork);
404 
405 	wb_update_bandwidth(wb);
406 }
407 
408 /*
409  * Initial write bandwidth: 100 MB/s
410  */
411 #define INIT_BW		(100 << (20 - PAGE_SHIFT))
412 
413 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
414 		   gfp_t gfp)
415 {
416 	int i, err;
417 
418 	memset(wb, 0, sizeof(*wb));
419 
420 	wb->bdi = bdi;
421 	wb->last_old_flush = jiffies;
422 	INIT_LIST_HEAD(&wb->b_dirty);
423 	INIT_LIST_HEAD(&wb->b_io);
424 	INIT_LIST_HEAD(&wb->b_more_io);
425 	INIT_LIST_HEAD(&wb->b_dirty_time);
426 	spin_lock_init(&wb->list_lock);
427 
428 	atomic_set(&wb->writeback_inodes, 0);
429 	wb->bw_time_stamp = jiffies;
430 	wb->balanced_dirty_ratelimit = INIT_BW;
431 	wb->dirty_ratelimit = INIT_BW;
432 	wb->write_bandwidth = INIT_BW;
433 	wb->avg_write_bandwidth = INIT_BW;
434 
435 	spin_lock_init(&wb->work_lock);
436 	INIT_LIST_HEAD(&wb->work_list);
437 	INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
438 	INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
439 
440 	err = fprop_local_init_percpu(&wb->completions, gfp);
441 	if (err)
442 		return err;
443 
444 	for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
445 		err = percpu_counter_init(&wb->stat[i], 0, gfp);
446 		if (err)
447 			goto out_destroy_stat;
448 	}
449 
450 	return 0;
451 
452 out_destroy_stat:
453 	while (i--)
454 		percpu_counter_destroy(&wb->stat[i]);
455 	fprop_local_destroy_percpu(&wb->completions);
456 	return err;
457 }
458 
459 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
460 
461 /*
462  * Remove bdi from the global list and shutdown any threads we have running
463  */
464 static void wb_shutdown(struct bdi_writeback *wb)
465 {
466 	/* Make sure nobody queues further work */
467 	spin_lock_irq(&wb->work_lock);
468 	if (!test_and_clear_bit(WB_registered, &wb->state)) {
469 		spin_unlock_irq(&wb->work_lock);
470 		return;
471 	}
472 	spin_unlock_irq(&wb->work_lock);
473 
474 	cgwb_remove_from_bdi_list(wb);
475 	/*
476 	 * Drain work list and shutdown the delayed_work.  !WB_registered
477 	 * tells wb_workfn() that @wb is dying and its work_list needs to
478 	 * be drained no matter what.
479 	 */
480 	mod_delayed_work(bdi_wq, &wb->dwork, 0);
481 	flush_delayed_work(&wb->dwork);
482 	WARN_ON(!list_empty(&wb->work_list));
483 	flush_delayed_work(&wb->bw_dwork);
484 }
485 
486 static void wb_exit(struct bdi_writeback *wb)
487 {
488 	int i;
489 
490 	WARN_ON(delayed_work_pending(&wb->dwork));
491 
492 	for (i = 0; i < NR_WB_STAT_ITEMS; i++)
493 		percpu_counter_destroy(&wb->stat[i]);
494 
495 	fprop_local_destroy_percpu(&wb->completions);
496 }
497 
498 #ifdef CONFIG_CGROUP_WRITEBACK
499 
500 #include <linux/memcontrol.h>
501 
502 /*
503  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
504  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
505  */
506 static DEFINE_SPINLOCK(cgwb_lock);
507 static struct workqueue_struct *cgwb_release_wq;
508 
509 static LIST_HEAD(offline_cgwbs);
510 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
511 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
512 
513 static void cgwb_free_rcu(struct rcu_head *rcu_head)
514 {
515 	struct bdi_writeback *wb = container_of(rcu_head,
516 			struct bdi_writeback, rcu);
517 
518 	percpu_ref_exit(&wb->refcnt);
519 	kfree(wb);
520 }
521 
522 static void cgwb_release_workfn(struct work_struct *work)
523 {
524 	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
525 						release_work);
526 	struct backing_dev_info *bdi = wb->bdi;
527 
528 	mutex_lock(&wb->bdi->cgwb_release_mutex);
529 	wb_shutdown(wb);
530 
531 	css_put(wb->memcg_css);
532 	css_put(wb->blkcg_css);
533 	mutex_unlock(&wb->bdi->cgwb_release_mutex);
534 
535 	/* triggers blkg destruction if no online users left */
536 	blkcg_unpin_online(wb->blkcg_css);
537 
538 	fprop_local_destroy_percpu(&wb->memcg_completions);
539 
540 	spin_lock_irq(&cgwb_lock);
541 	list_del(&wb->offline_node);
542 	spin_unlock_irq(&cgwb_lock);
543 
544 	wb_exit(wb);
545 	bdi_put(bdi);
546 	WARN_ON_ONCE(!list_empty(&wb->b_attached));
547 	call_rcu(&wb->rcu, cgwb_free_rcu);
548 }
549 
550 static void cgwb_release(struct percpu_ref *refcnt)
551 {
552 	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
553 						refcnt);
554 	queue_work(cgwb_release_wq, &wb->release_work);
555 }
556 
557 static void cgwb_kill(struct bdi_writeback *wb)
558 {
559 	lockdep_assert_held(&cgwb_lock);
560 
561 	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
562 	list_del(&wb->memcg_node);
563 	list_del(&wb->blkcg_node);
564 	list_add(&wb->offline_node, &offline_cgwbs);
565 	percpu_ref_kill(&wb->refcnt);
566 }
567 
568 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
569 {
570 	spin_lock_irq(&cgwb_lock);
571 	list_del_rcu(&wb->bdi_node);
572 	spin_unlock_irq(&cgwb_lock);
573 }
574 
575 static int cgwb_create(struct backing_dev_info *bdi,
576 		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
577 {
578 	struct mem_cgroup *memcg;
579 	struct cgroup_subsys_state *blkcg_css;
580 	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
581 	struct bdi_writeback *wb;
582 	unsigned long flags;
583 	int ret = 0;
584 
585 	memcg = mem_cgroup_from_css(memcg_css);
586 	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
587 	memcg_cgwb_list = &memcg->cgwb_list;
588 	blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
589 
590 	/* look up again under lock and discard on blkcg mismatch */
591 	spin_lock_irqsave(&cgwb_lock, flags);
592 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
593 	if (wb && wb->blkcg_css != blkcg_css) {
594 		cgwb_kill(wb);
595 		wb = NULL;
596 	}
597 	spin_unlock_irqrestore(&cgwb_lock, flags);
598 	if (wb)
599 		goto out_put;
600 
601 	/* need to create a new one */
602 	wb = kmalloc(sizeof(*wb), gfp);
603 	if (!wb) {
604 		ret = -ENOMEM;
605 		goto out_put;
606 	}
607 
608 	ret = wb_init(wb, bdi, gfp);
609 	if (ret)
610 		goto err_free;
611 
612 	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
613 	if (ret)
614 		goto err_wb_exit;
615 
616 	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
617 	if (ret)
618 		goto err_ref_exit;
619 
620 	wb->memcg_css = memcg_css;
621 	wb->blkcg_css = blkcg_css;
622 	INIT_LIST_HEAD(&wb->b_attached);
623 	INIT_WORK(&wb->release_work, cgwb_release_workfn);
624 	set_bit(WB_registered, &wb->state);
625 	bdi_get(bdi);
626 
627 	/*
628 	 * The root wb determines the registered state of the whole bdi and
629 	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
630 	 * whether they're still online.  Don't link @wb if any is dead.
631 	 * See wb_memcg_offline() and wb_blkcg_offline().
632 	 */
633 	ret = -ENODEV;
634 	spin_lock_irqsave(&cgwb_lock, flags);
635 	if (test_bit(WB_registered, &bdi->wb.state) &&
636 	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
637 		/* we might have raced another instance of this function */
638 		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
639 		if (!ret) {
640 			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
641 			list_add(&wb->memcg_node, memcg_cgwb_list);
642 			list_add(&wb->blkcg_node, blkcg_cgwb_list);
643 			blkcg_pin_online(blkcg_css);
644 			css_get(memcg_css);
645 			css_get(blkcg_css);
646 		}
647 	}
648 	spin_unlock_irqrestore(&cgwb_lock, flags);
649 	if (ret) {
650 		if (ret == -EEXIST)
651 			ret = 0;
652 		goto err_fprop_exit;
653 	}
654 	goto out_put;
655 
656 err_fprop_exit:
657 	bdi_put(bdi);
658 	fprop_local_destroy_percpu(&wb->memcg_completions);
659 err_ref_exit:
660 	percpu_ref_exit(&wb->refcnt);
661 err_wb_exit:
662 	wb_exit(wb);
663 err_free:
664 	kfree(wb);
665 out_put:
666 	css_put(blkcg_css);
667 	return ret;
668 }
669 
670 /**
671  * wb_get_lookup - get wb for a given memcg
672  * @bdi: target bdi
673  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
674  *
675  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
676  * refcount incremented.
677  *
678  * This function uses css_get() on @memcg_css and thus expects its refcnt
679  * to be positive on invocation.  IOW, rcu_read_lock() protection on
680  * @memcg_css isn't enough.  try_get it before calling this function.
681  *
682  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
683  * memcg on the default hierarchy, memcg association is guaranteed to be
684  * more specific (equal or descendant to the associated blkcg) and thus can
685  * identify both the memcg and blkcg associations.
686  *
687  * Because the blkcg associated with a memcg may change as blkcg is enabled
688  * and disabled closer to root in the hierarchy, each wb keeps track of
689  * both the memcg and blkcg associated with it and verifies the blkcg on
690  * each lookup.  On mismatch, the existing wb is discarded and a new one is
691  * created.
692  */
693 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
694 				    struct cgroup_subsys_state *memcg_css)
695 {
696 	struct bdi_writeback *wb;
697 
698 	if (!memcg_css->parent)
699 		return &bdi->wb;
700 
701 	rcu_read_lock();
702 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
703 	if (wb) {
704 		struct cgroup_subsys_state *blkcg_css;
705 
706 		/* see whether the blkcg association has changed */
707 		blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
708 		if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
709 			wb = NULL;
710 		css_put(blkcg_css);
711 	}
712 	rcu_read_unlock();
713 
714 	return wb;
715 }
716 
717 /**
718  * wb_get_create - get wb for a given memcg, create if necessary
719  * @bdi: target bdi
720  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
721  * @gfp: allocation mask to use
722  *
723  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
724  * create one.  See wb_get_lookup() for more details.
725  */
726 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
727 				    struct cgroup_subsys_state *memcg_css,
728 				    gfp_t gfp)
729 {
730 	struct bdi_writeback *wb;
731 
732 	might_alloc(gfp);
733 
734 	do {
735 		wb = wb_get_lookup(bdi, memcg_css);
736 	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
737 
738 	return wb;
739 }
740 
741 static int cgwb_bdi_init(struct backing_dev_info *bdi)
742 {
743 	int ret;
744 
745 	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
746 	mutex_init(&bdi->cgwb_release_mutex);
747 	init_rwsem(&bdi->wb_switch_rwsem);
748 
749 	ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
750 	if (!ret) {
751 		bdi->wb.memcg_css = &root_mem_cgroup->css;
752 		bdi->wb.blkcg_css = blkcg_root_css;
753 	}
754 	return ret;
755 }
756 
757 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
758 {
759 	struct radix_tree_iter iter;
760 	void **slot;
761 	struct bdi_writeback *wb;
762 
763 	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
764 
765 	spin_lock_irq(&cgwb_lock);
766 	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
767 		cgwb_kill(*slot);
768 	spin_unlock_irq(&cgwb_lock);
769 
770 	mutex_lock(&bdi->cgwb_release_mutex);
771 	spin_lock_irq(&cgwb_lock);
772 	while (!list_empty(&bdi->wb_list)) {
773 		wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
774 				      bdi_node);
775 		spin_unlock_irq(&cgwb_lock);
776 		wb_shutdown(wb);
777 		spin_lock_irq(&cgwb_lock);
778 	}
779 	spin_unlock_irq(&cgwb_lock);
780 	mutex_unlock(&bdi->cgwb_release_mutex);
781 }
782 
783 /*
784  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
785  *
786  * Try to release dying cgwbs by switching attached inodes to the nearest
787  * living ancestor's writeback. Processed wbs are placed at the end
788  * of the list to guarantee the forward progress.
789  */
790 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
791 {
792 	struct bdi_writeback *wb;
793 	LIST_HEAD(processed);
794 
795 	spin_lock_irq(&cgwb_lock);
796 
797 	while (!list_empty(&offline_cgwbs)) {
798 		wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
799 				      offline_node);
800 		list_move(&wb->offline_node, &processed);
801 
802 		/*
803 		 * If wb is dirty, cleaning up the writeback by switching
804 		 * attached inodes will result in an effective removal of any
805 		 * bandwidth restrictions, which isn't the goal.  Instead,
806 		 * it can be postponed until the next time, when all io
807 		 * will be likely completed.  If in the meantime some inodes
808 		 * will get re-dirtied, they should be eventually switched to
809 		 * a new cgwb.
810 		 */
811 		if (wb_has_dirty_io(wb))
812 			continue;
813 
814 		if (!wb_tryget(wb))
815 			continue;
816 
817 		spin_unlock_irq(&cgwb_lock);
818 		while (cleanup_offline_cgwb(wb))
819 			cond_resched();
820 		spin_lock_irq(&cgwb_lock);
821 
822 		wb_put(wb);
823 	}
824 
825 	if (!list_empty(&processed))
826 		list_splice_tail(&processed, &offline_cgwbs);
827 
828 	spin_unlock_irq(&cgwb_lock);
829 }
830 
831 /**
832  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
833  * @memcg: memcg being offlined
834  *
835  * Also prevents creation of any new wb's associated with @memcg.
836  */
837 void wb_memcg_offline(struct mem_cgroup *memcg)
838 {
839 	struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
840 	struct bdi_writeback *wb, *next;
841 
842 	spin_lock_irq(&cgwb_lock);
843 	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
844 		cgwb_kill(wb);
845 	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
846 	spin_unlock_irq(&cgwb_lock);
847 
848 	queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
849 }
850 
851 /**
852  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
853  * @css: blkcg being offlined
854  *
855  * Also prevents creation of any new wb's associated with @blkcg.
856  */
857 void wb_blkcg_offline(struct cgroup_subsys_state *css)
858 {
859 	struct bdi_writeback *wb, *next;
860 	struct list_head *list = blkcg_get_cgwb_list(css);
861 
862 	spin_lock_irq(&cgwb_lock);
863 	list_for_each_entry_safe(wb, next, list, blkcg_node)
864 		cgwb_kill(wb);
865 	list->next = NULL;	/* prevent new wb's */
866 	spin_unlock_irq(&cgwb_lock);
867 }
868 
869 static void cgwb_bdi_register(struct backing_dev_info *bdi)
870 {
871 	spin_lock_irq(&cgwb_lock);
872 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
873 	spin_unlock_irq(&cgwb_lock);
874 }
875 
876 static int __init cgwb_init(void)
877 {
878 	/*
879 	 * There can be many concurrent release work items overwhelming
880 	 * system_wq.  Put them in a separate wq and limit concurrency.
881 	 * There's no point in executing many of these in parallel.
882 	 */
883 	cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
884 	if (!cgwb_release_wq)
885 		return -ENOMEM;
886 
887 	return 0;
888 }
889 subsys_initcall(cgwb_init);
890 
891 #else	/* CONFIG_CGROUP_WRITEBACK */
892 
893 static int cgwb_bdi_init(struct backing_dev_info *bdi)
894 {
895 	return wb_init(&bdi->wb, bdi, GFP_KERNEL);
896 }
897 
898 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
899 
900 static void cgwb_bdi_register(struct backing_dev_info *bdi)
901 {
902 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
903 }
904 
905 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
906 {
907 	list_del_rcu(&wb->bdi_node);
908 }
909 
910 #endif	/* CONFIG_CGROUP_WRITEBACK */
911 
912 int bdi_init(struct backing_dev_info *bdi)
913 {
914 	bdi->dev = NULL;
915 
916 	kref_init(&bdi->refcnt);
917 	bdi->min_ratio = 0;
918 	bdi->max_ratio = 100 * BDI_RATIO_SCALE;
919 	bdi->max_prop_frac = FPROP_FRAC_BASE;
920 	INIT_LIST_HEAD(&bdi->bdi_list);
921 	INIT_LIST_HEAD(&bdi->wb_list);
922 	init_waitqueue_head(&bdi->wb_waitq);
923 	bdi->last_bdp_sleep = jiffies;
924 
925 	return cgwb_bdi_init(bdi);
926 }
927 
928 struct backing_dev_info *bdi_alloc(int node_id)
929 {
930 	struct backing_dev_info *bdi;
931 
932 	bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
933 	if (!bdi)
934 		return NULL;
935 
936 	if (bdi_init(bdi)) {
937 		kfree(bdi);
938 		return NULL;
939 	}
940 	bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
941 	bdi->ra_pages = VM_READAHEAD_PAGES;
942 	bdi->io_pages = VM_READAHEAD_PAGES;
943 	timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
944 	return bdi;
945 }
946 EXPORT_SYMBOL(bdi_alloc);
947 
948 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
949 {
950 	struct rb_node **p = &bdi_tree.rb_node;
951 	struct rb_node *parent = NULL;
952 	struct backing_dev_info *bdi;
953 
954 	lockdep_assert_held(&bdi_lock);
955 
956 	while (*p) {
957 		parent = *p;
958 		bdi = rb_entry(parent, struct backing_dev_info, rb_node);
959 
960 		if (bdi->id > id)
961 			p = &(*p)->rb_left;
962 		else if (bdi->id < id)
963 			p = &(*p)->rb_right;
964 		else
965 			break;
966 	}
967 
968 	if (parentp)
969 		*parentp = parent;
970 	return p;
971 }
972 
973 /**
974  * bdi_get_by_id - lookup and get bdi from its id
975  * @id: bdi id to lookup
976  *
977  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
978  * doesn't exist or is already unregistered.
979  */
980 struct backing_dev_info *bdi_get_by_id(u64 id)
981 {
982 	struct backing_dev_info *bdi = NULL;
983 	struct rb_node **p;
984 
985 	spin_lock_bh(&bdi_lock);
986 	p = bdi_lookup_rb_node(id, NULL);
987 	if (*p) {
988 		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
989 		bdi_get(bdi);
990 	}
991 	spin_unlock_bh(&bdi_lock);
992 
993 	return bdi;
994 }
995 
996 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
997 {
998 	struct device *dev;
999 	struct rb_node *parent, **p;
1000 
1001 	if (bdi->dev)	/* The driver needs to use separate queues per device */
1002 		return 0;
1003 
1004 	vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
1005 	dev = device_create(&bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
1006 	if (IS_ERR(dev))
1007 		return PTR_ERR(dev);
1008 
1009 	cgwb_bdi_register(bdi);
1010 	bdi->dev = dev;
1011 
1012 	bdi_debug_register(bdi, dev_name(dev));
1013 	set_bit(WB_registered, &bdi->wb.state);
1014 
1015 	spin_lock_bh(&bdi_lock);
1016 
1017 	bdi->id = ++bdi_id_cursor;
1018 
1019 	p = bdi_lookup_rb_node(bdi->id, &parent);
1020 	rb_link_node(&bdi->rb_node, parent, p);
1021 	rb_insert_color(&bdi->rb_node, &bdi_tree);
1022 
1023 	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
1024 
1025 	spin_unlock_bh(&bdi_lock);
1026 
1027 	trace_writeback_bdi_register(bdi);
1028 	return 0;
1029 }
1030 
1031 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1032 {
1033 	va_list args;
1034 	int ret;
1035 
1036 	va_start(args, fmt);
1037 	ret = bdi_register_va(bdi, fmt, args);
1038 	va_end(args);
1039 	return ret;
1040 }
1041 EXPORT_SYMBOL(bdi_register);
1042 
1043 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1044 {
1045 	WARN_ON_ONCE(bdi->owner);
1046 	bdi->owner = owner;
1047 	get_device(owner);
1048 }
1049 
1050 /*
1051  * Remove bdi from bdi_list, and ensure that it is no longer visible
1052  */
1053 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1054 {
1055 	spin_lock_bh(&bdi_lock);
1056 	rb_erase(&bdi->rb_node, &bdi_tree);
1057 	list_del_rcu(&bdi->bdi_list);
1058 	spin_unlock_bh(&bdi_lock);
1059 
1060 	synchronize_rcu_expedited();
1061 }
1062 
1063 void bdi_unregister(struct backing_dev_info *bdi)
1064 {
1065 	del_timer_sync(&bdi->laptop_mode_wb_timer);
1066 
1067 	/* make sure nobody finds us on the bdi_list anymore */
1068 	bdi_remove_from_list(bdi);
1069 	wb_shutdown(&bdi->wb);
1070 	cgwb_bdi_unregister(bdi);
1071 
1072 	/*
1073 	 * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1074 	 * update the global bdi_min_ratio.
1075 	 */
1076 	if (bdi->min_ratio)
1077 		bdi_set_min_ratio(bdi, 0);
1078 
1079 	if (bdi->dev) {
1080 		bdi_debug_unregister(bdi);
1081 		device_unregister(bdi->dev);
1082 		bdi->dev = NULL;
1083 	}
1084 
1085 	if (bdi->owner) {
1086 		put_device(bdi->owner);
1087 		bdi->owner = NULL;
1088 	}
1089 }
1090 EXPORT_SYMBOL(bdi_unregister);
1091 
1092 static void release_bdi(struct kref *ref)
1093 {
1094 	struct backing_dev_info *bdi =
1095 			container_of(ref, struct backing_dev_info, refcnt);
1096 
1097 	WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1098 	WARN_ON_ONCE(bdi->dev);
1099 	wb_exit(&bdi->wb);
1100 	kfree(bdi);
1101 }
1102 
1103 void bdi_put(struct backing_dev_info *bdi)
1104 {
1105 	kref_put(&bdi->refcnt, release_bdi);
1106 }
1107 EXPORT_SYMBOL(bdi_put);
1108 
1109 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1110 {
1111 	struct super_block *sb;
1112 
1113 	if (!inode)
1114 		return &noop_backing_dev_info;
1115 
1116 	sb = inode->i_sb;
1117 #ifdef CONFIG_BLOCK
1118 	if (sb_is_blkdev_sb(sb))
1119 		return I_BDEV(inode)->bd_disk->bdi;
1120 #endif
1121 	return sb->s_bdi;
1122 }
1123 EXPORT_SYMBOL(inode_to_bdi);
1124 
1125 const char *bdi_dev_name(struct backing_dev_info *bdi)
1126 {
1127 	if (!bdi || !bdi->dev)
1128 		return bdi_unknown_name;
1129 	return bdi->dev_name;
1130 }
1131 EXPORT_SYMBOL_GPL(bdi_dev_name);
1132