xref: /linux/mm/backing-dev.c (revision 0340dc4c82590d8735c58cf904a8aa1173273ab5)
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 i, 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 	for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
420 		err = percpu_counter_init(&wb->stat[i], 0, gfp);
421 		if (err)
422 			goto out_destroy_stat;
423 	}
424 
425 	return 0;
426 
427 out_destroy_stat:
428 	while (i--)
429 		percpu_counter_destroy(&wb->stat[i]);
430 	fprop_local_destroy_percpu(&wb->completions);
431 	return err;
432 }
433 
434 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
435 
436 /*
437  * Remove bdi from the global list and shutdown any threads we have running
438  */
439 static void wb_shutdown(struct bdi_writeback *wb)
440 {
441 	/* Make sure nobody queues further work */
442 	spin_lock_irq(&wb->work_lock);
443 	if (!test_and_clear_bit(WB_registered, &wb->state)) {
444 		spin_unlock_irq(&wb->work_lock);
445 		return;
446 	}
447 	spin_unlock_irq(&wb->work_lock);
448 
449 	cgwb_remove_from_bdi_list(wb);
450 	/*
451 	 * Drain work list and shutdown the delayed_work.  !WB_registered
452 	 * tells wb_workfn() that @wb is dying and its work_list needs to
453 	 * be drained no matter what.
454 	 */
455 	mod_delayed_work(bdi_wq, &wb->dwork, 0);
456 	flush_delayed_work(&wb->dwork);
457 	WARN_ON(!list_empty(&wb->work_list));
458 	flush_delayed_work(&wb->bw_dwork);
459 }
460 
461 static void wb_exit(struct bdi_writeback *wb)
462 {
463 	int i;
464 
465 	WARN_ON(delayed_work_pending(&wb->dwork));
466 
467 	for (i = 0; i < NR_WB_STAT_ITEMS; i++)
468 		percpu_counter_destroy(&wb->stat[i]);
469 
470 	fprop_local_destroy_percpu(&wb->completions);
471 }
472 
473 #ifdef CONFIG_CGROUP_WRITEBACK
474 
475 #include <linux/memcontrol.h>
476 
477 /*
478  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
479  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
480  */
481 static DEFINE_SPINLOCK(cgwb_lock);
482 static struct workqueue_struct *cgwb_release_wq;
483 
484 static LIST_HEAD(offline_cgwbs);
485 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
486 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
487 
488 static void cgwb_free_rcu(struct rcu_head *rcu_head)
489 {
490 	struct bdi_writeback *wb = container_of(rcu_head,
491 			struct bdi_writeback, rcu);
492 
493 	percpu_ref_exit(&wb->refcnt);
494 	kfree(wb);
495 }
496 
497 static void cgwb_release_workfn(struct work_struct *work)
498 {
499 	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
500 						release_work);
501 	struct backing_dev_info *bdi = wb->bdi;
502 
503 	mutex_lock(&wb->bdi->cgwb_release_mutex);
504 	wb_shutdown(wb);
505 
506 	css_put(wb->memcg_css);
507 	css_put(wb->blkcg_css);
508 	mutex_unlock(&wb->bdi->cgwb_release_mutex);
509 
510 	/* triggers blkg destruction if no online users left */
511 	blkcg_unpin_online(wb->blkcg_css);
512 
513 	fprop_local_destroy_percpu(&wb->memcg_completions);
514 
515 	spin_lock_irq(&cgwb_lock);
516 	list_del(&wb->offline_node);
517 	spin_unlock_irq(&cgwb_lock);
518 
519 	wb_exit(wb);
520 	bdi_put(bdi);
521 	WARN_ON_ONCE(!list_empty(&wb->b_attached));
522 	call_rcu(&wb->rcu, cgwb_free_rcu);
523 }
524 
525 static void cgwb_release(struct percpu_ref *refcnt)
526 {
527 	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
528 						refcnt);
529 	queue_work(cgwb_release_wq, &wb->release_work);
530 }
531 
532 static void cgwb_kill(struct bdi_writeback *wb)
533 {
534 	lockdep_assert_held(&cgwb_lock);
535 
536 	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
537 	list_del(&wb->memcg_node);
538 	list_del(&wb->blkcg_node);
539 	list_add(&wb->offline_node, &offline_cgwbs);
540 	percpu_ref_kill(&wb->refcnt);
541 }
542 
543 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
544 {
545 	spin_lock_irq(&cgwb_lock);
546 	list_del_rcu(&wb->bdi_node);
547 	spin_unlock_irq(&cgwb_lock);
548 }
549 
550 static int cgwb_create(struct backing_dev_info *bdi,
551 		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
552 {
553 	struct mem_cgroup *memcg;
554 	struct cgroup_subsys_state *blkcg_css;
555 	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
556 	struct bdi_writeback *wb;
557 	unsigned long flags;
558 	int ret = 0;
559 
560 	memcg = mem_cgroup_from_css(memcg_css);
561 	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
562 	memcg_cgwb_list = &memcg->cgwb_list;
563 	blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
564 
565 	/* look up again under lock and discard on blkcg mismatch */
566 	spin_lock_irqsave(&cgwb_lock, flags);
567 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
568 	if (wb && wb->blkcg_css != blkcg_css) {
569 		cgwb_kill(wb);
570 		wb = NULL;
571 	}
572 	spin_unlock_irqrestore(&cgwb_lock, flags);
573 	if (wb)
574 		goto out_put;
575 
576 	/* need to create a new one */
577 	wb = kmalloc(sizeof(*wb), gfp);
578 	if (!wb) {
579 		ret = -ENOMEM;
580 		goto out_put;
581 	}
582 
583 	ret = wb_init(wb, bdi, gfp);
584 	if (ret)
585 		goto err_free;
586 
587 	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
588 	if (ret)
589 		goto err_wb_exit;
590 
591 	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
592 	if (ret)
593 		goto err_ref_exit;
594 
595 	wb->memcg_css = memcg_css;
596 	wb->blkcg_css = blkcg_css;
597 	INIT_LIST_HEAD(&wb->b_attached);
598 	INIT_WORK(&wb->release_work, cgwb_release_workfn);
599 	set_bit(WB_registered, &wb->state);
600 	bdi_get(bdi);
601 
602 	/*
603 	 * The root wb determines the registered state of the whole bdi and
604 	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
605 	 * whether they're still online.  Don't link @wb if any is dead.
606 	 * See wb_memcg_offline() and wb_blkcg_offline().
607 	 */
608 	ret = -ENODEV;
609 	spin_lock_irqsave(&cgwb_lock, flags);
610 	if (test_bit(WB_registered, &bdi->wb.state) &&
611 	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
612 		/* we might have raced another instance of this function */
613 		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
614 		if (!ret) {
615 			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
616 			list_add(&wb->memcg_node, memcg_cgwb_list);
617 			list_add(&wb->blkcg_node, blkcg_cgwb_list);
618 			blkcg_pin_online(blkcg_css);
619 			css_get(memcg_css);
620 			css_get(blkcg_css);
621 		}
622 	}
623 	spin_unlock_irqrestore(&cgwb_lock, flags);
624 	if (ret) {
625 		if (ret == -EEXIST)
626 			ret = 0;
627 		goto err_fprop_exit;
628 	}
629 	goto out_put;
630 
631 err_fprop_exit:
632 	bdi_put(bdi);
633 	fprop_local_destroy_percpu(&wb->memcg_completions);
634 err_ref_exit:
635 	percpu_ref_exit(&wb->refcnt);
636 err_wb_exit:
637 	wb_exit(wb);
638 err_free:
639 	kfree(wb);
640 out_put:
641 	css_put(blkcg_css);
642 	return ret;
643 }
644 
645 /**
646  * wb_get_lookup - get wb for a given memcg
647  * @bdi: target bdi
648  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
649  *
650  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
651  * refcount incremented.
652  *
653  * This function uses css_get() on @memcg_css and thus expects its refcnt
654  * to be positive on invocation.  IOW, rcu_read_lock() protection on
655  * @memcg_css isn't enough.  try_get it before calling this function.
656  *
657  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
658  * memcg on the default hierarchy, memcg association is guaranteed to be
659  * more specific (equal or descendant to the associated blkcg) and thus can
660  * identify both the memcg and blkcg associations.
661  *
662  * Because the blkcg associated with a memcg may change as blkcg is enabled
663  * and disabled closer to root in the hierarchy, each wb keeps track of
664  * both the memcg and blkcg associated with it and verifies the blkcg on
665  * each lookup.  On mismatch, the existing wb is discarded and a new one is
666  * created.
667  */
668 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
669 				    struct cgroup_subsys_state *memcg_css)
670 {
671 	struct bdi_writeback *wb;
672 
673 	if (!memcg_css->parent)
674 		return &bdi->wb;
675 
676 	rcu_read_lock();
677 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
678 	if (wb) {
679 		struct cgroup_subsys_state *blkcg_css;
680 
681 		/* see whether the blkcg association has changed */
682 		blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
683 		if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
684 			wb = NULL;
685 		css_put(blkcg_css);
686 	}
687 	rcu_read_unlock();
688 
689 	return wb;
690 }
691 
692 /**
693  * wb_get_create - get wb for a given memcg, create if necessary
694  * @bdi: target bdi
695  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
696  * @gfp: allocation mask to use
697  *
698  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
699  * create one.  See wb_get_lookup() for more details.
700  */
701 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
702 				    struct cgroup_subsys_state *memcg_css,
703 				    gfp_t gfp)
704 {
705 	struct bdi_writeback *wb;
706 
707 	might_alloc(gfp);
708 
709 	do {
710 		wb = wb_get_lookup(bdi, memcg_css);
711 	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
712 
713 	return wb;
714 }
715 
716 static int cgwb_bdi_init(struct backing_dev_info *bdi)
717 {
718 	int ret;
719 
720 	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
721 	mutex_init(&bdi->cgwb_release_mutex);
722 	init_rwsem(&bdi->wb_switch_rwsem);
723 
724 	ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
725 	if (!ret) {
726 		bdi->wb.memcg_css = &root_mem_cgroup->css;
727 		bdi->wb.blkcg_css = blkcg_root_css;
728 	}
729 	return ret;
730 }
731 
732 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
733 {
734 	struct radix_tree_iter iter;
735 	void **slot;
736 	struct bdi_writeback *wb;
737 
738 	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
739 
740 	spin_lock_irq(&cgwb_lock);
741 	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
742 		cgwb_kill(*slot);
743 	spin_unlock_irq(&cgwb_lock);
744 
745 	mutex_lock(&bdi->cgwb_release_mutex);
746 	spin_lock_irq(&cgwb_lock);
747 	while (!list_empty(&bdi->wb_list)) {
748 		wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
749 				      bdi_node);
750 		spin_unlock_irq(&cgwb_lock);
751 		wb_shutdown(wb);
752 		spin_lock_irq(&cgwb_lock);
753 	}
754 	spin_unlock_irq(&cgwb_lock);
755 	mutex_unlock(&bdi->cgwb_release_mutex);
756 }
757 
758 /*
759  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
760  *
761  * Try to release dying cgwbs by switching attached inodes to the nearest
762  * living ancestor's writeback. Processed wbs are placed at the end
763  * of the list to guarantee the forward progress.
764  */
765 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
766 {
767 	struct bdi_writeback *wb;
768 	LIST_HEAD(processed);
769 
770 	spin_lock_irq(&cgwb_lock);
771 
772 	while (!list_empty(&offline_cgwbs)) {
773 		wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
774 				      offline_node);
775 		list_move(&wb->offline_node, &processed);
776 
777 		/*
778 		 * If wb is dirty, cleaning up the writeback by switching
779 		 * attached inodes will result in an effective removal of any
780 		 * bandwidth restrictions, which isn't the goal.  Instead,
781 		 * it can be postponed until the next time, when all io
782 		 * will be likely completed.  If in the meantime some inodes
783 		 * will get re-dirtied, they should be eventually switched to
784 		 * a new cgwb.
785 		 */
786 		if (wb_has_dirty_io(wb))
787 			continue;
788 
789 		if (!wb_tryget(wb))
790 			continue;
791 
792 		spin_unlock_irq(&cgwb_lock);
793 		while (cleanup_offline_cgwb(wb))
794 			cond_resched();
795 		spin_lock_irq(&cgwb_lock);
796 
797 		wb_put(wb);
798 	}
799 
800 	if (!list_empty(&processed))
801 		list_splice_tail(&processed, &offline_cgwbs);
802 
803 	spin_unlock_irq(&cgwb_lock);
804 }
805 
806 /**
807  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
808  * @memcg: memcg being offlined
809  *
810  * Also prevents creation of any new wb's associated with @memcg.
811  */
812 void wb_memcg_offline(struct mem_cgroup *memcg)
813 {
814 	struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
815 	struct bdi_writeback *wb, *next;
816 
817 	spin_lock_irq(&cgwb_lock);
818 	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
819 		cgwb_kill(wb);
820 	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
821 	spin_unlock_irq(&cgwb_lock);
822 
823 	queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
824 }
825 
826 /**
827  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
828  * @css: blkcg being offlined
829  *
830  * Also prevents creation of any new wb's associated with @blkcg.
831  */
832 void wb_blkcg_offline(struct cgroup_subsys_state *css)
833 {
834 	struct bdi_writeback *wb, *next;
835 	struct list_head *list = blkcg_get_cgwb_list(css);
836 
837 	spin_lock_irq(&cgwb_lock);
838 	list_for_each_entry_safe(wb, next, list, blkcg_node)
839 		cgwb_kill(wb);
840 	list->next = NULL;	/* prevent new wb's */
841 	spin_unlock_irq(&cgwb_lock);
842 }
843 
844 static void cgwb_bdi_register(struct backing_dev_info *bdi)
845 {
846 	spin_lock_irq(&cgwb_lock);
847 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
848 	spin_unlock_irq(&cgwb_lock);
849 }
850 
851 static int __init cgwb_init(void)
852 {
853 	/*
854 	 * There can be many concurrent release work items overwhelming
855 	 * system_wq.  Put them in a separate wq and limit concurrency.
856 	 * There's no point in executing many of these in parallel.
857 	 */
858 	cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
859 	if (!cgwb_release_wq)
860 		return -ENOMEM;
861 
862 	return 0;
863 }
864 subsys_initcall(cgwb_init);
865 
866 #else	/* CONFIG_CGROUP_WRITEBACK */
867 
868 static int cgwb_bdi_init(struct backing_dev_info *bdi)
869 {
870 	return wb_init(&bdi->wb, bdi, GFP_KERNEL);
871 }
872 
873 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
874 
875 static void cgwb_bdi_register(struct backing_dev_info *bdi)
876 {
877 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
878 }
879 
880 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
881 {
882 	list_del_rcu(&wb->bdi_node);
883 }
884 
885 #endif	/* CONFIG_CGROUP_WRITEBACK */
886 
887 int bdi_init(struct backing_dev_info *bdi)
888 {
889 	bdi->dev = NULL;
890 
891 	kref_init(&bdi->refcnt);
892 	bdi->min_ratio = 0;
893 	bdi->max_ratio = 100 * BDI_RATIO_SCALE;
894 	bdi->max_prop_frac = FPROP_FRAC_BASE;
895 	INIT_LIST_HEAD(&bdi->bdi_list);
896 	INIT_LIST_HEAD(&bdi->wb_list);
897 	init_waitqueue_head(&bdi->wb_waitq);
898 	bdi->last_bdp_sleep = jiffies;
899 
900 	return cgwb_bdi_init(bdi);
901 }
902 
903 struct backing_dev_info *bdi_alloc(int node_id)
904 {
905 	struct backing_dev_info *bdi;
906 
907 	bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
908 	if (!bdi)
909 		return NULL;
910 
911 	if (bdi_init(bdi)) {
912 		kfree(bdi);
913 		return NULL;
914 	}
915 	bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
916 	bdi->ra_pages = VM_READAHEAD_PAGES;
917 	bdi->io_pages = VM_READAHEAD_PAGES;
918 	timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
919 	return bdi;
920 }
921 EXPORT_SYMBOL(bdi_alloc);
922 
923 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
924 {
925 	struct rb_node **p = &bdi_tree.rb_node;
926 	struct rb_node *parent = NULL;
927 	struct backing_dev_info *bdi;
928 
929 	lockdep_assert_held(&bdi_lock);
930 
931 	while (*p) {
932 		parent = *p;
933 		bdi = rb_entry(parent, struct backing_dev_info, rb_node);
934 
935 		if (bdi->id > id)
936 			p = &(*p)->rb_left;
937 		else if (bdi->id < id)
938 			p = &(*p)->rb_right;
939 		else
940 			break;
941 	}
942 
943 	if (parentp)
944 		*parentp = parent;
945 	return p;
946 }
947 
948 /**
949  * bdi_get_by_id - lookup and get bdi from its id
950  * @id: bdi id to lookup
951  *
952  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
953  * doesn't exist or is already unregistered.
954  */
955 struct backing_dev_info *bdi_get_by_id(u64 id)
956 {
957 	struct backing_dev_info *bdi = NULL;
958 	struct rb_node **p;
959 
960 	spin_lock_bh(&bdi_lock);
961 	p = bdi_lookup_rb_node(id, NULL);
962 	if (*p) {
963 		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
964 		bdi_get(bdi);
965 	}
966 	spin_unlock_bh(&bdi_lock);
967 
968 	return bdi;
969 }
970 
971 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
972 {
973 	struct device *dev;
974 	struct rb_node *parent, **p;
975 
976 	if (bdi->dev)	/* The driver needs to use separate queues per device */
977 		return 0;
978 
979 	vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
980 	dev = device_create(&bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
981 	if (IS_ERR(dev))
982 		return PTR_ERR(dev);
983 
984 	cgwb_bdi_register(bdi);
985 	bdi->dev = dev;
986 
987 	bdi_debug_register(bdi, dev_name(dev));
988 	set_bit(WB_registered, &bdi->wb.state);
989 
990 	spin_lock_bh(&bdi_lock);
991 
992 	bdi->id = ++bdi_id_cursor;
993 
994 	p = bdi_lookup_rb_node(bdi->id, &parent);
995 	rb_link_node(&bdi->rb_node, parent, p);
996 	rb_insert_color(&bdi->rb_node, &bdi_tree);
997 
998 	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
999 
1000 	spin_unlock_bh(&bdi_lock);
1001 
1002 	trace_writeback_bdi_register(bdi);
1003 	return 0;
1004 }
1005 
1006 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1007 {
1008 	va_list args;
1009 	int ret;
1010 
1011 	va_start(args, fmt);
1012 	ret = bdi_register_va(bdi, fmt, args);
1013 	va_end(args);
1014 	return ret;
1015 }
1016 EXPORT_SYMBOL(bdi_register);
1017 
1018 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1019 {
1020 	WARN_ON_ONCE(bdi->owner);
1021 	bdi->owner = owner;
1022 	get_device(owner);
1023 }
1024 
1025 /*
1026  * Remove bdi from bdi_list, and ensure that it is no longer visible
1027  */
1028 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1029 {
1030 	spin_lock_bh(&bdi_lock);
1031 	rb_erase(&bdi->rb_node, &bdi_tree);
1032 	list_del_rcu(&bdi->bdi_list);
1033 	spin_unlock_bh(&bdi_lock);
1034 
1035 	synchronize_rcu_expedited();
1036 }
1037 
1038 void bdi_unregister(struct backing_dev_info *bdi)
1039 {
1040 	del_timer_sync(&bdi->laptop_mode_wb_timer);
1041 
1042 	/* make sure nobody finds us on the bdi_list anymore */
1043 	bdi_remove_from_list(bdi);
1044 	wb_shutdown(&bdi->wb);
1045 	cgwb_bdi_unregister(bdi);
1046 
1047 	/*
1048 	 * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1049 	 * update the global bdi_min_ratio.
1050 	 */
1051 	if (bdi->min_ratio)
1052 		bdi_set_min_ratio(bdi, 0);
1053 
1054 	if (bdi->dev) {
1055 		bdi_debug_unregister(bdi);
1056 		device_unregister(bdi->dev);
1057 		bdi->dev = NULL;
1058 	}
1059 
1060 	if (bdi->owner) {
1061 		put_device(bdi->owner);
1062 		bdi->owner = NULL;
1063 	}
1064 }
1065 EXPORT_SYMBOL(bdi_unregister);
1066 
1067 static void release_bdi(struct kref *ref)
1068 {
1069 	struct backing_dev_info *bdi =
1070 			container_of(ref, struct backing_dev_info, refcnt);
1071 
1072 	WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1073 	WARN_ON_ONCE(bdi->dev);
1074 	wb_exit(&bdi->wb);
1075 	kfree(bdi);
1076 }
1077 
1078 void bdi_put(struct backing_dev_info *bdi)
1079 {
1080 	kref_put(&bdi->refcnt, release_bdi);
1081 }
1082 EXPORT_SYMBOL(bdi_put);
1083 
1084 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1085 {
1086 	struct super_block *sb;
1087 
1088 	if (!inode)
1089 		return &noop_backing_dev_info;
1090 
1091 	sb = inode->i_sb;
1092 #ifdef CONFIG_BLOCK
1093 	if (sb_is_blkdev_sb(sb))
1094 		return I_BDEV(inode)->bd_disk->bdi;
1095 #endif
1096 	return sb->s_bdi;
1097 }
1098 EXPORT_SYMBOL(inode_to_bdi);
1099 
1100 const char *bdi_dev_name(struct backing_dev_info *bdi)
1101 {
1102 	if (!bdi || !bdi->dev)
1103 		return bdi_unknown_name;
1104 	return bdi->dev_name;
1105 }
1106 EXPORT_SYMBOL_GPL(bdi_dev_name);
1107