xref: /linux/mm/backing-dev.c (revision 34efe1c3b688944d9817a5faaab7aad870182c59)
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 static void wb_update_bandwidth_workfn(struct work_struct *work)
376 {
377 	struct bdi_writeback *wb = container_of(to_delayed_work(work),
378 						struct bdi_writeback, bw_dwork);
379 
380 	wb_update_bandwidth(wb);
381 }
382 
383 /*
384  * Initial write bandwidth: 100 MB/s
385  */
386 #define INIT_BW		(100 << (20 - PAGE_SHIFT))
387 
388 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
389 		   gfp_t gfp)
390 {
391 	int err;
392 
393 	memset(wb, 0, sizeof(*wb));
394 
395 	wb->bdi = bdi;
396 	wb->last_old_flush = jiffies;
397 	INIT_LIST_HEAD(&wb->b_dirty);
398 	INIT_LIST_HEAD(&wb->b_io);
399 	INIT_LIST_HEAD(&wb->b_more_io);
400 	INIT_LIST_HEAD(&wb->b_dirty_time);
401 	spin_lock_init(&wb->list_lock);
402 
403 	atomic_set(&wb->writeback_inodes, 0);
404 	wb->bw_time_stamp = jiffies;
405 	wb->balanced_dirty_ratelimit = INIT_BW;
406 	wb->dirty_ratelimit = INIT_BW;
407 	wb->write_bandwidth = INIT_BW;
408 	wb->avg_write_bandwidth = INIT_BW;
409 
410 	spin_lock_init(&wb->work_lock);
411 	INIT_LIST_HEAD(&wb->work_list);
412 	INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
413 	INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
414 
415 	err = fprop_local_init_percpu(&wb->completions, gfp);
416 	if (err)
417 		return err;
418 
419 	err = percpu_counter_init_many(wb->stat, 0, gfp, NR_WB_STAT_ITEMS);
420 	if (err)
421 		fprop_local_destroy_percpu(&wb->completions);
422 
423 	return err;
424 }
425 
426 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
427 
428 /*
429  * Remove bdi from the global list and shutdown any threads we have running
430  */
431 static void wb_shutdown(struct bdi_writeback *wb)
432 {
433 	/* Make sure nobody queues further work */
434 	spin_lock_irq(&wb->work_lock);
435 	if (!test_and_clear_bit(WB_registered, &wb->state)) {
436 		spin_unlock_irq(&wb->work_lock);
437 		return;
438 	}
439 	spin_unlock_irq(&wb->work_lock);
440 
441 	cgwb_remove_from_bdi_list(wb);
442 	/*
443 	 * Drain work list and shutdown the delayed_work.  !WB_registered
444 	 * tells wb_workfn() that @wb is dying and its work_list needs to
445 	 * be drained no matter what.
446 	 */
447 	mod_delayed_work(bdi_wq, &wb->dwork, 0);
448 	flush_delayed_work(&wb->dwork);
449 	WARN_ON(!list_empty(&wb->work_list));
450 	flush_delayed_work(&wb->bw_dwork);
451 }
452 
453 static void wb_exit(struct bdi_writeback *wb)
454 {
455 	WARN_ON(delayed_work_pending(&wb->dwork));
456 	percpu_counter_destroy_many(wb->stat, NR_WB_STAT_ITEMS);
457 	fprop_local_destroy_percpu(&wb->completions);
458 }
459 
460 #ifdef CONFIG_CGROUP_WRITEBACK
461 
462 #include <linux/memcontrol.h>
463 
464 /*
465  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
466  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
467  */
468 static DEFINE_SPINLOCK(cgwb_lock);
469 static struct workqueue_struct *cgwb_release_wq;
470 
471 static LIST_HEAD(offline_cgwbs);
472 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
473 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
474 
475 static void cgwb_free_rcu(struct rcu_head *rcu_head)
476 {
477 	struct bdi_writeback *wb = container_of(rcu_head,
478 			struct bdi_writeback, rcu);
479 
480 	percpu_ref_exit(&wb->refcnt);
481 	kfree(wb);
482 }
483 
484 static void cgwb_release_workfn(struct work_struct *work)
485 {
486 	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
487 						release_work);
488 	struct backing_dev_info *bdi = wb->bdi;
489 
490 	mutex_lock(&wb->bdi->cgwb_release_mutex);
491 	wb_shutdown(wb);
492 
493 	css_put(wb->memcg_css);
494 	css_put(wb->blkcg_css);
495 	mutex_unlock(&wb->bdi->cgwb_release_mutex);
496 
497 	/* triggers blkg destruction if no online users left */
498 	blkcg_unpin_online(wb->blkcg_css);
499 
500 	fprop_local_destroy_percpu(&wb->memcg_completions);
501 
502 	spin_lock_irq(&cgwb_lock);
503 	list_del(&wb->offline_node);
504 	spin_unlock_irq(&cgwb_lock);
505 
506 	wb_exit(wb);
507 	bdi_put(bdi);
508 	WARN_ON_ONCE(!list_empty(&wb->b_attached));
509 	call_rcu(&wb->rcu, cgwb_free_rcu);
510 }
511 
512 static void cgwb_release(struct percpu_ref *refcnt)
513 {
514 	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
515 						refcnt);
516 	queue_work(cgwb_release_wq, &wb->release_work);
517 }
518 
519 static void cgwb_kill(struct bdi_writeback *wb)
520 {
521 	lockdep_assert_held(&cgwb_lock);
522 
523 	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
524 	list_del(&wb->memcg_node);
525 	list_del(&wb->blkcg_node);
526 	list_add(&wb->offline_node, &offline_cgwbs);
527 	percpu_ref_kill(&wb->refcnt);
528 }
529 
530 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
531 {
532 	spin_lock_irq(&cgwb_lock);
533 	list_del_rcu(&wb->bdi_node);
534 	spin_unlock_irq(&cgwb_lock);
535 }
536 
537 static int cgwb_create(struct backing_dev_info *bdi,
538 		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
539 {
540 	struct mem_cgroup *memcg;
541 	struct cgroup_subsys_state *blkcg_css;
542 	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
543 	struct bdi_writeback *wb;
544 	unsigned long flags;
545 	int ret = 0;
546 
547 	memcg = mem_cgroup_from_css(memcg_css);
548 	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
549 	memcg_cgwb_list = &memcg->cgwb_list;
550 	blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
551 
552 	/* look up again under lock and discard on blkcg mismatch */
553 	spin_lock_irqsave(&cgwb_lock, flags);
554 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
555 	if (wb && wb->blkcg_css != blkcg_css) {
556 		cgwb_kill(wb);
557 		wb = NULL;
558 	}
559 	spin_unlock_irqrestore(&cgwb_lock, flags);
560 	if (wb)
561 		goto out_put;
562 
563 	/* need to create a new one */
564 	wb = kmalloc(sizeof(*wb), gfp);
565 	if (!wb) {
566 		ret = -ENOMEM;
567 		goto out_put;
568 	}
569 
570 	ret = wb_init(wb, bdi, gfp);
571 	if (ret)
572 		goto err_free;
573 
574 	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
575 	if (ret)
576 		goto err_wb_exit;
577 
578 	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
579 	if (ret)
580 		goto err_ref_exit;
581 
582 	wb->memcg_css = memcg_css;
583 	wb->blkcg_css = blkcg_css;
584 	INIT_LIST_HEAD(&wb->b_attached);
585 	INIT_WORK(&wb->release_work, cgwb_release_workfn);
586 	set_bit(WB_registered, &wb->state);
587 	bdi_get(bdi);
588 
589 	/*
590 	 * The root wb determines the registered state of the whole bdi and
591 	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
592 	 * whether they're still online.  Don't link @wb if any is dead.
593 	 * See wb_memcg_offline() and wb_blkcg_offline().
594 	 */
595 	ret = -ENODEV;
596 	spin_lock_irqsave(&cgwb_lock, flags);
597 	if (test_bit(WB_registered, &bdi->wb.state) &&
598 	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
599 		/* we might have raced another instance of this function */
600 		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
601 		if (!ret) {
602 			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
603 			list_add(&wb->memcg_node, memcg_cgwb_list);
604 			list_add(&wb->blkcg_node, blkcg_cgwb_list);
605 			blkcg_pin_online(blkcg_css);
606 			css_get(memcg_css);
607 			css_get(blkcg_css);
608 		}
609 	}
610 	spin_unlock_irqrestore(&cgwb_lock, flags);
611 	if (ret) {
612 		if (ret == -EEXIST)
613 			ret = 0;
614 		goto err_fprop_exit;
615 	}
616 	goto out_put;
617 
618 err_fprop_exit:
619 	bdi_put(bdi);
620 	fprop_local_destroy_percpu(&wb->memcg_completions);
621 err_ref_exit:
622 	percpu_ref_exit(&wb->refcnt);
623 err_wb_exit:
624 	wb_exit(wb);
625 err_free:
626 	kfree(wb);
627 out_put:
628 	css_put(blkcg_css);
629 	return ret;
630 }
631 
632 /**
633  * wb_get_lookup - get wb for a given memcg
634  * @bdi: target bdi
635  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
636  *
637  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
638  * refcount incremented.
639  *
640  * This function uses css_get() on @memcg_css and thus expects its refcnt
641  * to be positive on invocation.  IOW, rcu_read_lock() protection on
642  * @memcg_css isn't enough.  try_get it before calling this function.
643  *
644  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
645  * memcg on the default hierarchy, memcg association is guaranteed to be
646  * more specific (equal or descendant to the associated blkcg) and thus can
647  * identify both the memcg and blkcg associations.
648  *
649  * Because the blkcg associated with a memcg may change as blkcg is enabled
650  * and disabled closer to root in the hierarchy, each wb keeps track of
651  * both the memcg and blkcg associated with it and verifies the blkcg on
652  * each lookup.  On mismatch, the existing wb is discarded and a new one is
653  * created.
654  */
655 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
656 				    struct cgroup_subsys_state *memcg_css)
657 {
658 	struct bdi_writeback *wb;
659 
660 	if (!memcg_css->parent)
661 		return &bdi->wb;
662 
663 	rcu_read_lock();
664 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
665 	if (wb) {
666 		struct cgroup_subsys_state *blkcg_css;
667 
668 		/* see whether the blkcg association has changed */
669 		blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
670 		if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
671 			wb = NULL;
672 		css_put(blkcg_css);
673 	}
674 	rcu_read_unlock();
675 
676 	return wb;
677 }
678 
679 /**
680  * wb_get_create - get wb for a given memcg, create if necessary
681  * @bdi: target bdi
682  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
683  * @gfp: allocation mask to use
684  *
685  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
686  * create one.  See wb_get_lookup() for more details.
687  */
688 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
689 				    struct cgroup_subsys_state *memcg_css,
690 				    gfp_t gfp)
691 {
692 	struct bdi_writeback *wb;
693 
694 	might_alloc(gfp);
695 
696 	do {
697 		wb = wb_get_lookup(bdi, memcg_css);
698 	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
699 
700 	return wb;
701 }
702 
703 static int cgwb_bdi_init(struct backing_dev_info *bdi)
704 {
705 	int ret;
706 
707 	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
708 	mutex_init(&bdi->cgwb_release_mutex);
709 	init_rwsem(&bdi->wb_switch_rwsem);
710 
711 	ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
712 	if (!ret) {
713 		bdi->wb.memcg_css = &root_mem_cgroup->css;
714 		bdi->wb.blkcg_css = blkcg_root_css;
715 	}
716 	return ret;
717 }
718 
719 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
720 {
721 	struct radix_tree_iter iter;
722 	void **slot;
723 	struct bdi_writeback *wb;
724 
725 	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
726 
727 	spin_lock_irq(&cgwb_lock);
728 	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
729 		cgwb_kill(*slot);
730 	spin_unlock_irq(&cgwb_lock);
731 
732 	mutex_lock(&bdi->cgwb_release_mutex);
733 	spin_lock_irq(&cgwb_lock);
734 	while (!list_empty(&bdi->wb_list)) {
735 		wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
736 				      bdi_node);
737 		spin_unlock_irq(&cgwb_lock);
738 		wb_shutdown(wb);
739 		spin_lock_irq(&cgwb_lock);
740 	}
741 	spin_unlock_irq(&cgwb_lock);
742 	mutex_unlock(&bdi->cgwb_release_mutex);
743 }
744 
745 /*
746  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
747  *
748  * Try to release dying cgwbs by switching attached inodes to the nearest
749  * living ancestor's writeback. Processed wbs are placed at the end
750  * of the list to guarantee the forward progress.
751  */
752 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
753 {
754 	struct bdi_writeback *wb;
755 	LIST_HEAD(processed);
756 
757 	spin_lock_irq(&cgwb_lock);
758 
759 	while (!list_empty(&offline_cgwbs)) {
760 		wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
761 				      offline_node);
762 		list_move(&wb->offline_node, &processed);
763 
764 		/*
765 		 * If wb is dirty, cleaning up the writeback by switching
766 		 * attached inodes will result in an effective removal of any
767 		 * bandwidth restrictions, which isn't the goal.  Instead,
768 		 * it can be postponed until the next time, when all io
769 		 * will be likely completed.  If in the meantime some inodes
770 		 * will get re-dirtied, they should be eventually switched to
771 		 * a new cgwb.
772 		 */
773 		if (wb_has_dirty_io(wb))
774 			continue;
775 
776 		if (!wb_tryget(wb))
777 			continue;
778 
779 		spin_unlock_irq(&cgwb_lock);
780 		while (cleanup_offline_cgwb(wb))
781 			cond_resched();
782 		spin_lock_irq(&cgwb_lock);
783 
784 		wb_put(wb);
785 	}
786 
787 	if (!list_empty(&processed))
788 		list_splice_tail(&processed, &offline_cgwbs);
789 
790 	spin_unlock_irq(&cgwb_lock);
791 }
792 
793 /**
794  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
795  * @memcg: memcg being offlined
796  *
797  * Also prevents creation of any new wb's associated with @memcg.
798  */
799 void wb_memcg_offline(struct mem_cgroup *memcg)
800 {
801 	struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
802 	struct bdi_writeback *wb, *next;
803 
804 	spin_lock_irq(&cgwb_lock);
805 	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
806 		cgwb_kill(wb);
807 	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
808 	spin_unlock_irq(&cgwb_lock);
809 
810 	queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
811 }
812 
813 /**
814  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
815  * @css: blkcg being offlined
816  *
817  * Also prevents creation of any new wb's associated with @blkcg.
818  */
819 void wb_blkcg_offline(struct cgroup_subsys_state *css)
820 {
821 	struct bdi_writeback *wb, *next;
822 	struct list_head *list = blkcg_get_cgwb_list(css);
823 
824 	spin_lock_irq(&cgwb_lock);
825 	list_for_each_entry_safe(wb, next, list, blkcg_node)
826 		cgwb_kill(wb);
827 	list->next = NULL;	/* prevent new wb's */
828 	spin_unlock_irq(&cgwb_lock);
829 }
830 
831 static void cgwb_bdi_register(struct backing_dev_info *bdi)
832 {
833 	spin_lock_irq(&cgwb_lock);
834 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
835 	spin_unlock_irq(&cgwb_lock);
836 }
837 
838 static int __init cgwb_init(void)
839 {
840 	/*
841 	 * There can be many concurrent release work items overwhelming
842 	 * system_wq.  Put them in a separate wq and limit concurrency.
843 	 * There's no point in executing many of these in parallel.
844 	 */
845 	cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
846 	if (!cgwb_release_wq)
847 		return -ENOMEM;
848 
849 	return 0;
850 }
851 subsys_initcall(cgwb_init);
852 
853 #else	/* CONFIG_CGROUP_WRITEBACK */
854 
855 static int cgwb_bdi_init(struct backing_dev_info *bdi)
856 {
857 	return wb_init(&bdi->wb, bdi, GFP_KERNEL);
858 }
859 
860 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
861 
862 static void cgwb_bdi_register(struct backing_dev_info *bdi)
863 {
864 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
865 }
866 
867 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
868 {
869 	list_del_rcu(&wb->bdi_node);
870 }
871 
872 #endif	/* CONFIG_CGROUP_WRITEBACK */
873 
874 int bdi_init(struct backing_dev_info *bdi)
875 {
876 	bdi->dev = NULL;
877 
878 	kref_init(&bdi->refcnt);
879 	bdi->min_ratio = 0;
880 	bdi->max_ratio = 100 * BDI_RATIO_SCALE;
881 	bdi->max_prop_frac = FPROP_FRAC_BASE;
882 	INIT_LIST_HEAD(&bdi->bdi_list);
883 	INIT_LIST_HEAD(&bdi->wb_list);
884 	init_waitqueue_head(&bdi->wb_waitq);
885 	bdi->last_bdp_sleep = jiffies;
886 
887 	return cgwb_bdi_init(bdi);
888 }
889 
890 struct backing_dev_info *bdi_alloc(int node_id)
891 {
892 	struct backing_dev_info *bdi;
893 
894 	bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
895 	if (!bdi)
896 		return NULL;
897 
898 	if (bdi_init(bdi)) {
899 		kfree(bdi);
900 		return NULL;
901 	}
902 	bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
903 	bdi->ra_pages = VM_READAHEAD_PAGES;
904 	bdi->io_pages = VM_READAHEAD_PAGES;
905 	timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
906 	return bdi;
907 }
908 EXPORT_SYMBOL(bdi_alloc);
909 
910 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
911 {
912 	struct rb_node **p = &bdi_tree.rb_node;
913 	struct rb_node *parent = NULL;
914 	struct backing_dev_info *bdi;
915 
916 	lockdep_assert_held(&bdi_lock);
917 
918 	while (*p) {
919 		parent = *p;
920 		bdi = rb_entry(parent, struct backing_dev_info, rb_node);
921 
922 		if (bdi->id > id)
923 			p = &(*p)->rb_left;
924 		else if (bdi->id < id)
925 			p = &(*p)->rb_right;
926 		else
927 			break;
928 	}
929 
930 	if (parentp)
931 		*parentp = parent;
932 	return p;
933 }
934 
935 /**
936  * bdi_get_by_id - lookup and get bdi from its id
937  * @id: bdi id to lookup
938  *
939  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
940  * doesn't exist or is already unregistered.
941  */
942 struct backing_dev_info *bdi_get_by_id(u64 id)
943 {
944 	struct backing_dev_info *bdi = NULL;
945 	struct rb_node **p;
946 
947 	spin_lock_bh(&bdi_lock);
948 	p = bdi_lookup_rb_node(id, NULL);
949 	if (*p) {
950 		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
951 		bdi_get(bdi);
952 	}
953 	spin_unlock_bh(&bdi_lock);
954 
955 	return bdi;
956 }
957 
958 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
959 {
960 	struct device *dev;
961 	struct rb_node *parent, **p;
962 
963 	if (bdi->dev)	/* The driver needs to use separate queues per device */
964 		return 0;
965 
966 	vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
967 	dev = device_create(&bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
968 	if (IS_ERR(dev))
969 		return PTR_ERR(dev);
970 
971 	cgwb_bdi_register(bdi);
972 	bdi->dev = dev;
973 
974 	bdi_debug_register(bdi, dev_name(dev));
975 	set_bit(WB_registered, &bdi->wb.state);
976 
977 	spin_lock_bh(&bdi_lock);
978 
979 	bdi->id = ++bdi_id_cursor;
980 
981 	p = bdi_lookup_rb_node(bdi->id, &parent);
982 	rb_link_node(&bdi->rb_node, parent, p);
983 	rb_insert_color(&bdi->rb_node, &bdi_tree);
984 
985 	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
986 
987 	spin_unlock_bh(&bdi_lock);
988 
989 	trace_writeback_bdi_register(bdi);
990 	return 0;
991 }
992 
993 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
994 {
995 	va_list args;
996 	int ret;
997 
998 	va_start(args, fmt);
999 	ret = bdi_register_va(bdi, fmt, args);
1000 	va_end(args);
1001 	return ret;
1002 }
1003 EXPORT_SYMBOL(bdi_register);
1004 
1005 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1006 {
1007 	WARN_ON_ONCE(bdi->owner);
1008 	bdi->owner = owner;
1009 	get_device(owner);
1010 }
1011 
1012 /*
1013  * Remove bdi from bdi_list, and ensure that it is no longer visible
1014  */
1015 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1016 {
1017 	spin_lock_bh(&bdi_lock);
1018 	rb_erase(&bdi->rb_node, &bdi_tree);
1019 	list_del_rcu(&bdi->bdi_list);
1020 	spin_unlock_bh(&bdi_lock);
1021 
1022 	synchronize_rcu_expedited();
1023 }
1024 
1025 void bdi_unregister(struct backing_dev_info *bdi)
1026 {
1027 	del_timer_sync(&bdi->laptop_mode_wb_timer);
1028 
1029 	/* make sure nobody finds us on the bdi_list anymore */
1030 	bdi_remove_from_list(bdi);
1031 	wb_shutdown(&bdi->wb);
1032 	cgwb_bdi_unregister(bdi);
1033 
1034 	/*
1035 	 * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1036 	 * update the global bdi_min_ratio.
1037 	 */
1038 	if (bdi->min_ratio)
1039 		bdi_set_min_ratio(bdi, 0);
1040 
1041 	if (bdi->dev) {
1042 		bdi_debug_unregister(bdi);
1043 		device_unregister(bdi->dev);
1044 		bdi->dev = NULL;
1045 	}
1046 
1047 	if (bdi->owner) {
1048 		put_device(bdi->owner);
1049 		bdi->owner = NULL;
1050 	}
1051 }
1052 EXPORT_SYMBOL(bdi_unregister);
1053 
1054 static void release_bdi(struct kref *ref)
1055 {
1056 	struct backing_dev_info *bdi =
1057 			container_of(ref, struct backing_dev_info, refcnt);
1058 
1059 	WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1060 	WARN_ON_ONCE(bdi->dev);
1061 	wb_exit(&bdi->wb);
1062 	kfree(bdi);
1063 }
1064 
1065 void bdi_put(struct backing_dev_info *bdi)
1066 {
1067 	kref_put(&bdi->refcnt, release_bdi);
1068 }
1069 EXPORT_SYMBOL(bdi_put);
1070 
1071 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1072 {
1073 	struct super_block *sb;
1074 
1075 	if (!inode)
1076 		return &noop_backing_dev_info;
1077 
1078 	sb = inode->i_sb;
1079 #ifdef CONFIG_BLOCK
1080 	if (sb_is_blkdev_sb(sb))
1081 		return I_BDEV(inode)->bd_disk->bdi;
1082 #endif
1083 	return sb->s_bdi;
1084 }
1085 EXPORT_SYMBOL(inode_to_bdi);
1086 
1087 const char *bdi_dev_name(struct backing_dev_info *bdi)
1088 {
1089 	if (!bdi || !bdi->dev)
1090 		return bdi_unknown_name;
1091 	return bdi->dev_name;
1092 }
1093 EXPORT_SYMBOL_GPL(bdi_dev_name);
1094