xref: /openbmc/linux/mm/backing-dev.c (revision ee7da21a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 #include <linux/wait.h>
4 #include <linux/rbtree.h>
5 #include <linux/backing-dev.h>
6 #include <linux/kthread.h>
7 #include <linux/freezer.h>
8 #include <linux/fs.h>
9 #include <linux/pagemap.h>
10 #include <linux/mm.h>
11 #include <linux/sched/mm.h>
12 #include <linux/sched.h>
13 #include <linux/module.h>
14 #include <linux/writeback.h>
15 #include <linux/device.h>
16 #include <trace/events/writeback.h>
17 
18 struct backing_dev_info noop_backing_dev_info;
19 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
20 
21 static struct class *bdi_class;
22 static const char *bdi_unknown_name = "(unknown)";
23 
24 /*
25  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
26  * reader side locking.
27  */
28 DEFINE_SPINLOCK(bdi_lock);
29 static u64 bdi_id_cursor;
30 static struct rb_root bdi_tree = RB_ROOT;
31 LIST_HEAD(bdi_list);
32 
33 /* bdi_wq serves all asynchronous writeback tasks */
34 struct workqueue_struct *bdi_wq;
35 
36 #define K(x) ((x) << (PAGE_SHIFT - 10))
37 
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/debugfs.h>
40 #include <linux/seq_file.h>
41 
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)
180 
181 static ssize_t max_ratio_store(struct device *dev,
182 		struct device_attribute *attr, const char *buf, size_t count)
183 {
184 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
185 	unsigned int ratio;
186 	ssize_t ret;
187 
188 	ret = kstrtouint(buf, 10, &ratio);
189 	if (ret < 0)
190 		return ret;
191 
192 	ret = bdi_set_max_ratio(bdi, ratio);
193 	if (!ret)
194 		ret = count;
195 
196 	return ret;
197 }
198 BDI_SHOW(max_ratio, bdi->max_ratio)
199 
200 static ssize_t stable_pages_required_show(struct device *dev,
201 					  struct device_attribute *attr,
202 					  char *buf)
203 {
204 	dev_warn_once(dev,
205 		"the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
206 	return sysfs_emit(buf, "%d\n", 0);
207 }
208 static DEVICE_ATTR_RO(stable_pages_required);
209 
210 static struct attribute *bdi_dev_attrs[] = {
211 	&dev_attr_read_ahead_kb.attr,
212 	&dev_attr_min_ratio.attr,
213 	&dev_attr_max_ratio.attr,
214 	&dev_attr_stable_pages_required.attr,
215 	NULL,
216 };
217 ATTRIBUTE_GROUPS(bdi_dev);
218 
219 static __init int bdi_class_init(void)
220 {
221 	bdi_class = class_create(THIS_MODULE, "bdi");
222 	if (IS_ERR(bdi_class))
223 		return PTR_ERR(bdi_class);
224 
225 	bdi_class->dev_groups = bdi_dev_groups;
226 	bdi_debug_init();
227 
228 	return 0;
229 }
230 postcore_initcall(bdi_class_init);
231 
232 static int bdi_init(struct backing_dev_info *bdi);
233 
234 static int __init default_bdi_init(void)
235 {
236 	int err;
237 
238 	bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
239 				 WQ_SYSFS, 0);
240 	if (!bdi_wq)
241 		return -ENOMEM;
242 
243 	err = bdi_init(&noop_backing_dev_info);
244 
245 	return err;
246 }
247 subsys_initcall(default_bdi_init);
248 
249 /*
250  * This function is used when the first inode for this wb is marked dirty. It
251  * wakes-up the corresponding bdi thread which should then take care of the
252  * periodic background write-out of dirty inodes. Since the write-out would
253  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
254  * set up a timer which wakes the bdi thread up later.
255  *
256  * Note, we wouldn't bother setting up the timer, but this function is on the
257  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
258  * by delaying the wake-up.
259  *
260  * We have to be careful not to postpone flush work if it is scheduled for
261  * earlier. Thus we use queue_delayed_work().
262  */
263 void wb_wakeup_delayed(struct bdi_writeback *wb)
264 {
265 	unsigned long timeout;
266 
267 	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
268 	spin_lock_bh(&wb->work_lock);
269 	if (test_bit(WB_registered, &wb->state))
270 		queue_delayed_work(bdi_wq, &wb->dwork, timeout);
271 	spin_unlock_bh(&wb->work_lock);
272 }
273 
274 /*
275  * Initial write bandwidth: 100 MB/s
276  */
277 #define INIT_BW		(100 << (20 - PAGE_SHIFT))
278 
279 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
280 		   gfp_t gfp)
281 {
282 	int i, err;
283 
284 	memset(wb, 0, sizeof(*wb));
285 
286 	if (wb != &bdi->wb)
287 		bdi_get(bdi);
288 	wb->bdi = bdi;
289 	wb->last_old_flush = jiffies;
290 	INIT_LIST_HEAD(&wb->b_dirty);
291 	INIT_LIST_HEAD(&wb->b_io);
292 	INIT_LIST_HEAD(&wb->b_more_io);
293 	INIT_LIST_HEAD(&wb->b_dirty_time);
294 	spin_lock_init(&wb->list_lock);
295 
296 	wb->bw_time_stamp = jiffies;
297 	wb->balanced_dirty_ratelimit = INIT_BW;
298 	wb->dirty_ratelimit = INIT_BW;
299 	wb->write_bandwidth = INIT_BW;
300 	wb->avg_write_bandwidth = INIT_BW;
301 
302 	spin_lock_init(&wb->work_lock);
303 	INIT_LIST_HEAD(&wb->work_list);
304 	INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
305 	wb->dirty_sleep = jiffies;
306 
307 	err = fprop_local_init_percpu(&wb->completions, gfp);
308 	if (err)
309 		goto out_put_bdi;
310 
311 	for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
312 		err = percpu_counter_init(&wb->stat[i], 0, gfp);
313 		if (err)
314 			goto out_destroy_stat;
315 	}
316 
317 	return 0;
318 
319 out_destroy_stat:
320 	while (i--)
321 		percpu_counter_destroy(&wb->stat[i]);
322 	fprop_local_destroy_percpu(&wb->completions);
323 out_put_bdi:
324 	if (wb != &bdi->wb)
325 		bdi_put(bdi);
326 	return err;
327 }
328 
329 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
330 
331 /*
332  * Remove bdi from the global list and shutdown any threads we have running
333  */
334 static void wb_shutdown(struct bdi_writeback *wb)
335 {
336 	/* Make sure nobody queues further work */
337 	spin_lock_bh(&wb->work_lock);
338 	if (!test_and_clear_bit(WB_registered, &wb->state)) {
339 		spin_unlock_bh(&wb->work_lock);
340 		return;
341 	}
342 	spin_unlock_bh(&wb->work_lock);
343 
344 	cgwb_remove_from_bdi_list(wb);
345 	/*
346 	 * Drain work list and shutdown the delayed_work.  !WB_registered
347 	 * tells wb_workfn() that @wb is dying and its work_list needs to
348 	 * be drained no matter what.
349 	 */
350 	mod_delayed_work(bdi_wq, &wb->dwork, 0);
351 	flush_delayed_work(&wb->dwork);
352 	WARN_ON(!list_empty(&wb->work_list));
353 }
354 
355 static void wb_exit(struct bdi_writeback *wb)
356 {
357 	int i;
358 
359 	WARN_ON(delayed_work_pending(&wb->dwork));
360 
361 	for (i = 0; i < NR_WB_STAT_ITEMS; i++)
362 		percpu_counter_destroy(&wb->stat[i]);
363 
364 	fprop_local_destroy_percpu(&wb->completions);
365 	if (wb != &wb->bdi->wb)
366 		bdi_put(wb->bdi);
367 }
368 
369 #ifdef CONFIG_CGROUP_WRITEBACK
370 
371 #include <linux/memcontrol.h>
372 
373 /*
374  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
375  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
376  */
377 static DEFINE_SPINLOCK(cgwb_lock);
378 static struct workqueue_struct *cgwb_release_wq;
379 
380 static LIST_HEAD(offline_cgwbs);
381 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
382 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
383 
384 static void cgwb_release_workfn(struct work_struct *work)
385 {
386 	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
387 						release_work);
388 	struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
389 
390 	mutex_lock(&wb->bdi->cgwb_release_mutex);
391 	wb_shutdown(wb);
392 
393 	css_put(wb->memcg_css);
394 	css_put(wb->blkcg_css);
395 	mutex_unlock(&wb->bdi->cgwb_release_mutex);
396 
397 	/* triggers blkg destruction if no online users left */
398 	blkcg_unpin_online(blkcg);
399 
400 	fprop_local_destroy_percpu(&wb->memcg_completions);
401 
402 	spin_lock_irq(&cgwb_lock);
403 	list_del(&wb->offline_node);
404 	spin_unlock_irq(&cgwb_lock);
405 
406 	percpu_ref_exit(&wb->refcnt);
407 	wb_exit(wb);
408 	WARN_ON_ONCE(!list_empty(&wb->b_attached));
409 	kfree_rcu(wb, rcu);
410 }
411 
412 static void cgwb_release(struct percpu_ref *refcnt)
413 {
414 	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
415 						refcnt);
416 	queue_work(cgwb_release_wq, &wb->release_work);
417 }
418 
419 static void cgwb_kill(struct bdi_writeback *wb)
420 {
421 	lockdep_assert_held(&cgwb_lock);
422 
423 	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
424 	list_del(&wb->memcg_node);
425 	list_del(&wb->blkcg_node);
426 	list_add(&wb->offline_node, &offline_cgwbs);
427 	percpu_ref_kill(&wb->refcnt);
428 }
429 
430 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
431 {
432 	spin_lock_irq(&cgwb_lock);
433 	list_del_rcu(&wb->bdi_node);
434 	spin_unlock_irq(&cgwb_lock);
435 }
436 
437 static int cgwb_create(struct backing_dev_info *bdi,
438 		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
439 {
440 	struct mem_cgroup *memcg;
441 	struct cgroup_subsys_state *blkcg_css;
442 	struct blkcg *blkcg;
443 	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
444 	struct bdi_writeback *wb;
445 	unsigned long flags;
446 	int ret = 0;
447 
448 	memcg = mem_cgroup_from_css(memcg_css);
449 	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
450 	blkcg = css_to_blkcg(blkcg_css);
451 	memcg_cgwb_list = &memcg->cgwb_list;
452 	blkcg_cgwb_list = &blkcg->cgwb_list;
453 
454 	/* look up again under lock and discard on blkcg mismatch */
455 	spin_lock_irqsave(&cgwb_lock, flags);
456 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
457 	if (wb && wb->blkcg_css != blkcg_css) {
458 		cgwb_kill(wb);
459 		wb = NULL;
460 	}
461 	spin_unlock_irqrestore(&cgwb_lock, flags);
462 	if (wb)
463 		goto out_put;
464 
465 	/* need to create a new one */
466 	wb = kmalloc(sizeof(*wb), gfp);
467 	if (!wb) {
468 		ret = -ENOMEM;
469 		goto out_put;
470 	}
471 
472 	ret = wb_init(wb, bdi, gfp);
473 	if (ret)
474 		goto err_free;
475 
476 	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
477 	if (ret)
478 		goto err_wb_exit;
479 
480 	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
481 	if (ret)
482 		goto err_ref_exit;
483 
484 	wb->memcg_css = memcg_css;
485 	wb->blkcg_css = blkcg_css;
486 	INIT_LIST_HEAD(&wb->b_attached);
487 	INIT_WORK(&wb->release_work, cgwb_release_workfn);
488 	set_bit(WB_registered, &wb->state);
489 
490 	/*
491 	 * The root wb determines the registered state of the whole bdi and
492 	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
493 	 * whether they're still online.  Don't link @wb if any is dead.
494 	 * See wb_memcg_offline() and wb_blkcg_offline().
495 	 */
496 	ret = -ENODEV;
497 	spin_lock_irqsave(&cgwb_lock, flags);
498 	if (test_bit(WB_registered, &bdi->wb.state) &&
499 	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
500 		/* we might have raced another instance of this function */
501 		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
502 		if (!ret) {
503 			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
504 			list_add(&wb->memcg_node, memcg_cgwb_list);
505 			list_add(&wb->blkcg_node, blkcg_cgwb_list);
506 			blkcg_pin_online(blkcg);
507 			css_get(memcg_css);
508 			css_get(blkcg_css);
509 		}
510 	}
511 	spin_unlock_irqrestore(&cgwb_lock, flags);
512 	if (ret) {
513 		if (ret == -EEXIST)
514 			ret = 0;
515 		goto err_fprop_exit;
516 	}
517 	goto out_put;
518 
519 err_fprop_exit:
520 	fprop_local_destroy_percpu(&wb->memcg_completions);
521 err_ref_exit:
522 	percpu_ref_exit(&wb->refcnt);
523 err_wb_exit:
524 	wb_exit(wb);
525 err_free:
526 	kfree(wb);
527 out_put:
528 	css_put(blkcg_css);
529 	return ret;
530 }
531 
532 /**
533  * wb_get_lookup - get wb for a given memcg
534  * @bdi: target bdi
535  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
536  *
537  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
538  * refcount incremented.
539  *
540  * This function uses css_get() on @memcg_css and thus expects its refcnt
541  * to be positive on invocation.  IOW, rcu_read_lock() protection on
542  * @memcg_css isn't enough.  try_get it before calling this function.
543  *
544  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
545  * memcg on the default hierarchy, memcg association is guaranteed to be
546  * more specific (equal or descendant to the associated blkcg) and thus can
547  * identify both the memcg and blkcg associations.
548  *
549  * Because the blkcg associated with a memcg may change as blkcg is enabled
550  * and disabled closer to root in the hierarchy, each wb keeps track of
551  * both the memcg and blkcg associated with it and verifies the blkcg on
552  * each lookup.  On mismatch, the existing wb is discarded and a new one is
553  * created.
554  */
555 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
556 				    struct cgroup_subsys_state *memcg_css)
557 {
558 	struct bdi_writeback *wb;
559 
560 	if (!memcg_css->parent)
561 		return &bdi->wb;
562 
563 	rcu_read_lock();
564 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
565 	if (wb) {
566 		struct cgroup_subsys_state *blkcg_css;
567 
568 		/* see whether the blkcg association has changed */
569 		blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
570 		if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
571 			wb = NULL;
572 		css_put(blkcg_css);
573 	}
574 	rcu_read_unlock();
575 
576 	return wb;
577 }
578 
579 /**
580  * wb_get_create - get wb for a given memcg, create if necessary
581  * @bdi: target bdi
582  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
583  * @gfp: allocation mask to use
584  *
585  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
586  * create one.  See wb_get_lookup() for more details.
587  */
588 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
589 				    struct cgroup_subsys_state *memcg_css,
590 				    gfp_t gfp)
591 {
592 	struct bdi_writeback *wb;
593 
594 	might_alloc(gfp);
595 
596 	if (!memcg_css->parent)
597 		return &bdi->wb;
598 
599 	do {
600 		wb = wb_get_lookup(bdi, memcg_css);
601 	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
602 
603 	return wb;
604 }
605 
606 static int cgwb_bdi_init(struct backing_dev_info *bdi)
607 {
608 	int ret;
609 
610 	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
611 	mutex_init(&bdi->cgwb_release_mutex);
612 	init_rwsem(&bdi->wb_switch_rwsem);
613 
614 	ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
615 	if (!ret) {
616 		bdi->wb.memcg_css = &root_mem_cgroup->css;
617 		bdi->wb.blkcg_css = blkcg_root_css;
618 	}
619 	return ret;
620 }
621 
622 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
623 {
624 	struct radix_tree_iter iter;
625 	void **slot;
626 	struct bdi_writeback *wb;
627 
628 	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
629 
630 	spin_lock_irq(&cgwb_lock);
631 	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
632 		cgwb_kill(*slot);
633 	spin_unlock_irq(&cgwb_lock);
634 
635 	mutex_lock(&bdi->cgwb_release_mutex);
636 	spin_lock_irq(&cgwb_lock);
637 	while (!list_empty(&bdi->wb_list)) {
638 		wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
639 				      bdi_node);
640 		spin_unlock_irq(&cgwb_lock);
641 		wb_shutdown(wb);
642 		spin_lock_irq(&cgwb_lock);
643 	}
644 	spin_unlock_irq(&cgwb_lock);
645 	mutex_unlock(&bdi->cgwb_release_mutex);
646 }
647 
648 /*
649  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
650  *
651  * Try to release dying cgwbs by switching attached inodes to the nearest
652  * living ancestor's writeback. Processed wbs are placed at the end
653  * of the list to guarantee the forward progress.
654  */
655 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
656 {
657 	struct bdi_writeback *wb;
658 	LIST_HEAD(processed);
659 
660 	spin_lock_irq(&cgwb_lock);
661 
662 	while (!list_empty(&offline_cgwbs)) {
663 		wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
664 				      offline_node);
665 		list_move(&wb->offline_node, &processed);
666 
667 		/*
668 		 * If wb is dirty, cleaning up the writeback by switching
669 		 * attached inodes will result in an effective removal of any
670 		 * bandwidth restrictions, which isn't the goal.  Instead,
671 		 * it can be postponed until the next time, when all io
672 		 * will be likely completed.  If in the meantime some inodes
673 		 * will get re-dirtied, they should be eventually switched to
674 		 * a new cgwb.
675 		 */
676 		if (wb_has_dirty_io(wb))
677 			continue;
678 
679 		if (!wb_tryget(wb))
680 			continue;
681 
682 		spin_unlock_irq(&cgwb_lock);
683 		while (cleanup_offline_cgwb(wb))
684 			cond_resched();
685 		spin_lock_irq(&cgwb_lock);
686 
687 		wb_put(wb);
688 	}
689 
690 	if (!list_empty(&processed))
691 		list_splice_tail(&processed, &offline_cgwbs);
692 
693 	spin_unlock_irq(&cgwb_lock);
694 }
695 
696 /**
697  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
698  * @memcg: memcg being offlined
699  *
700  * Also prevents creation of any new wb's associated with @memcg.
701  */
702 void wb_memcg_offline(struct mem_cgroup *memcg)
703 {
704 	struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
705 	struct bdi_writeback *wb, *next;
706 
707 	spin_lock_irq(&cgwb_lock);
708 	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
709 		cgwb_kill(wb);
710 	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
711 	spin_unlock_irq(&cgwb_lock);
712 
713 	queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
714 }
715 
716 /**
717  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
718  * @blkcg: blkcg being offlined
719  *
720  * Also prevents creation of any new wb's associated with @blkcg.
721  */
722 void wb_blkcg_offline(struct blkcg *blkcg)
723 {
724 	struct bdi_writeback *wb, *next;
725 
726 	spin_lock_irq(&cgwb_lock);
727 	list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
728 		cgwb_kill(wb);
729 	blkcg->cgwb_list.next = NULL;	/* prevent new wb's */
730 	spin_unlock_irq(&cgwb_lock);
731 }
732 
733 static void cgwb_bdi_register(struct backing_dev_info *bdi)
734 {
735 	spin_lock_irq(&cgwb_lock);
736 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
737 	spin_unlock_irq(&cgwb_lock);
738 }
739 
740 static int __init cgwb_init(void)
741 {
742 	/*
743 	 * There can be many concurrent release work items overwhelming
744 	 * system_wq.  Put them in a separate wq and limit concurrency.
745 	 * There's no point in executing many of these in parallel.
746 	 */
747 	cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
748 	if (!cgwb_release_wq)
749 		return -ENOMEM;
750 
751 	return 0;
752 }
753 subsys_initcall(cgwb_init);
754 
755 #else	/* CONFIG_CGROUP_WRITEBACK */
756 
757 static int cgwb_bdi_init(struct backing_dev_info *bdi)
758 {
759 	return wb_init(&bdi->wb, bdi, GFP_KERNEL);
760 }
761 
762 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
763 
764 static void cgwb_bdi_register(struct backing_dev_info *bdi)
765 {
766 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
767 }
768 
769 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
770 {
771 	list_del_rcu(&wb->bdi_node);
772 }
773 
774 #endif	/* CONFIG_CGROUP_WRITEBACK */
775 
776 static int bdi_init(struct backing_dev_info *bdi)
777 {
778 	int ret;
779 
780 	bdi->dev = NULL;
781 
782 	kref_init(&bdi->refcnt);
783 	bdi->min_ratio = 0;
784 	bdi->max_ratio = 100;
785 	bdi->max_prop_frac = FPROP_FRAC_BASE;
786 	INIT_LIST_HEAD(&bdi->bdi_list);
787 	INIT_LIST_HEAD(&bdi->wb_list);
788 	init_waitqueue_head(&bdi->wb_waitq);
789 
790 	ret = cgwb_bdi_init(bdi);
791 
792 	return ret;
793 }
794 
795 struct backing_dev_info *bdi_alloc(int node_id)
796 {
797 	struct backing_dev_info *bdi;
798 
799 	bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
800 	if (!bdi)
801 		return NULL;
802 
803 	if (bdi_init(bdi)) {
804 		kfree(bdi);
805 		return NULL;
806 	}
807 	bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
808 	bdi->ra_pages = VM_READAHEAD_PAGES;
809 	bdi->io_pages = VM_READAHEAD_PAGES;
810 	return bdi;
811 }
812 EXPORT_SYMBOL(bdi_alloc);
813 
814 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
815 {
816 	struct rb_node **p = &bdi_tree.rb_node;
817 	struct rb_node *parent = NULL;
818 	struct backing_dev_info *bdi;
819 
820 	lockdep_assert_held(&bdi_lock);
821 
822 	while (*p) {
823 		parent = *p;
824 		bdi = rb_entry(parent, struct backing_dev_info, rb_node);
825 
826 		if (bdi->id > id)
827 			p = &(*p)->rb_left;
828 		else if (bdi->id < id)
829 			p = &(*p)->rb_right;
830 		else
831 			break;
832 	}
833 
834 	if (parentp)
835 		*parentp = parent;
836 	return p;
837 }
838 
839 /**
840  * bdi_get_by_id - lookup and get bdi from its id
841  * @id: bdi id to lookup
842  *
843  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
844  * doesn't exist or is already unregistered.
845  */
846 struct backing_dev_info *bdi_get_by_id(u64 id)
847 {
848 	struct backing_dev_info *bdi = NULL;
849 	struct rb_node **p;
850 
851 	spin_lock_bh(&bdi_lock);
852 	p = bdi_lookup_rb_node(id, NULL);
853 	if (*p) {
854 		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
855 		bdi_get(bdi);
856 	}
857 	spin_unlock_bh(&bdi_lock);
858 
859 	return bdi;
860 }
861 
862 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
863 {
864 	struct device *dev;
865 	struct rb_node *parent, **p;
866 
867 	if (bdi->dev)	/* The driver needs to use separate queues per device */
868 		return 0;
869 
870 	vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
871 	dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
872 	if (IS_ERR(dev))
873 		return PTR_ERR(dev);
874 
875 	cgwb_bdi_register(bdi);
876 	bdi->dev = dev;
877 
878 	bdi_debug_register(bdi, dev_name(dev));
879 	set_bit(WB_registered, &bdi->wb.state);
880 
881 	spin_lock_bh(&bdi_lock);
882 
883 	bdi->id = ++bdi_id_cursor;
884 
885 	p = bdi_lookup_rb_node(bdi->id, &parent);
886 	rb_link_node(&bdi->rb_node, parent, p);
887 	rb_insert_color(&bdi->rb_node, &bdi_tree);
888 
889 	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
890 
891 	spin_unlock_bh(&bdi_lock);
892 
893 	trace_writeback_bdi_register(bdi);
894 	return 0;
895 }
896 
897 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
898 {
899 	va_list args;
900 	int ret;
901 
902 	va_start(args, fmt);
903 	ret = bdi_register_va(bdi, fmt, args);
904 	va_end(args);
905 	return ret;
906 }
907 EXPORT_SYMBOL(bdi_register);
908 
909 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
910 {
911 	WARN_ON_ONCE(bdi->owner);
912 	bdi->owner = owner;
913 	get_device(owner);
914 }
915 
916 /*
917  * Remove bdi from bdi_list, and ensure that it is no longer visible
918  */
919 static void bdi_remove_from_list(struct backing_dev_info *bdi)
920 {
921 	spin_lock_bh(&bdi_lock);
922 	rb_erase(&bdi->rb_node, &bdi_tree);
923 	list_del_rcu(&bdi->bdi_list);
924 	spin_unlock_bh(&bdi_lock);
925 
926 	synchronize_rcu_expedited();
927 }
928 
929 void bdi_unregister(struct backing_dev_info *bdi)
930 {
931 	/* make sure nobody finds us on the bdi_list anymore */
932 	bdi_remove_from_list(bdi);
933 	wb_shutdown(&bdi->wb);
934 	cgwb_bdi_unregister(bdi);
935 
936 	if (bdi->dev) {
937 		bdi_debug_unregister(bdi);
938 		device_unregister(bdi->dev);
939 		bdi->dev = NULL;
940 	}
941 
942 	if (bdi->owner) {
943 		put_device(bdi->owner);
944 		bdi->owner = NULL;
945 	}
946 }
947 
948 static void release_bdi(struct kref *ref)
949 {
950 	struct backing_dev_info *bdi =
951 			container_of(ref, struct backing_dev_info, refcnt);
952 
953 	if (test_bit(WB_registered, &bdi->wb.state))
954 		bdi_unregister(bdi);
955 	WARN_ON_ONCE(bdi->dev);
956 	wb_exit(&bdi->wb);
957 	kfree(bdi);
958 }
959 
960 void bdi_put(struct backing_dev_info *bdi)
961 {
962 	kref_put(&bdi->refcnt, release_bdi);
963 }
964 EXPORT_SYMBOL(bdi_put);
965 
966 const char *bdi_dev_name(struct backing_dev_info *bdi)
967 {
968 	if (!bdi || !bdi->dev)
969 		return bdi_unknown_name;
970 	return bdi->dev_name;
971 }
972 EXPORT_SYMBOL_GPL(bdi_dev_name);
973 
974 static wait_queue_head_t congestion_wqh[2] = {
975 		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
976 		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
977 	};
978 static atomic_t nr_wb_congested[2];
979 
980 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
981 {
982 	wait_queue_head_t *wqh = &congestion_wqh[sync];
983 	enum wb_congested_state bit;
984 
985 	bit = sync ? WB_sync_congested : WB_async_congested;
986 	if (test_and_clear_bit(bit, &bdi->wb.congested))
987 		atomic_dec(&nr_wb_congested[sync]);
988 	smp_mb__after_atomic();
989 	if (waitqueue_active(wqh))
990 		wake_up(wqh);
991 }
992 EXPORT_SYMBOL(clear_bdi_congested);
993 
994 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
995 {
996 	enum wb_congested_state bit;
997 
998 	bit = sync ? WB_sync_congested : WB_async_congested;
999 	if (!test_and_set_bit(bit, &bdi->wb.congested))
1000 		atomic_inc(&nr_wb_congested[sync]);
1001 }
1002 EXPORT_SYMBOL(set_bdi_congested);
1003 
1004 /**
1005  * congestion_wait - wait for a backing_dev to become uncongested
1006  * @sync: SYNC or ASYNC IO
1007  * @timeout: timeout in jiffies
1008  *
1009  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1010  * write congestion.  If no backing_devs are congested then just wait for the
1011  * next write to be completed.
1012  */
1013 long congestion_wait(int sync, long timeout)
1014 {
1015 	long ret;
1016 	unsigned long start = jiffies;
1017 	DEFINE_WAIT(wait);
1018 	wait_queue_head_t *wqh = &congestion_wqh[sync];
1019 
1020 	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1021 	ret = io_schedule_timeout(timeout);
1022 	finish_wait(wqh, &wait);
1023 
1024 	trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1025 					jiffies_to_usecs(jiffies - start));
1026 
1027 	return ret;
1028 }
1029 EXPORT_SYMBOL(congestion_wait);
1030 
1031 /**
1032  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1033  * @sync: SYNC or ASYNC IO
1034  * @timeout: timeout in jiffies
1035  *
1036  * In the event of a congested backing_dev (any backing_dev) this waits
1037  * for up to @timeout jiffies for either a BDI to exit congestion of the
1038  * given @sync queue or a write to complete.
1039  *
1040  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1041  * it is the number of jiffies that were still remaining when the function
1042  * returned. return_value == timeout implies the function did not sleep.
1043  */
1044 long wait_iff_congested(int sync, long timeout)
1045 {
1046 	long ret;
1047 	unsigned long start = jiffies;
1048 	DEFINE_WAIT(wait);
1049 	wait_queue_head_t *wqh = &congestion_wqh[sync];
1050 
1051 	/*
1052 	 * If there is no congestion, yield if necessary instead
1053 	 * of sleeping on the congestion queue
1054 	 */
1055 	if (atomic_read(&nr_wb_congested[sync]) == 0) {
1056 		cond_resched();
1057 
1058 		/* In case we scheduled, work out time remaining */
1059 		ret = timeout - (jiffies - start);
1060 		if (ret < 0)
1061 			ret = 0;
1062 
1063 		goto out;
1064 	}
1065 
1066 	/* Sleep until uncongested or a write happens */
1067 	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1068 	ret = io_schedule_timeout(timeout);
1069 	finish_wait(wqh, &wait);
1070 
1071 out:
1072 	trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1073 					jiffies_to_usecs(jiffies - start));
1074 
1075 	return ret;
1076 }
1077 EXPORT_SYMBOL(wait_iff_congested);
1078