xref: /linux/mm/backing-dev.c (revision 6fbd6cf85a3be127454a1ad58525a3adcf8612ab)
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.h>
12 #include <linux/module.h>
13 #include <linux/writeback.h>
14 #include <linux/device.h>
15 #include <trace/events/writeback.h>
16 
17 struct backing_dev_info noop_backing_dev_info;
18 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
19 
20 static struct class *bdi_class;
21 static const char *bdi_unknown_name = "(unknown)";
22 
23 /*
24  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
25  * reader side locking.
26  */
27 DEFINE_SPINLOCK(bdi_lock);
28 static u64 bdi_id_cursor;
29 static struct rb_root bdi_tree = RB_ROOT;
30 LIST_HEAD(bdi_list);
31 
32 /* bdi_wq serves all asynchronous writeback tasks */
33 struct workqueue_struct *bdi_wq;
34 
35 #define K(x) ((x) << (PAGE_SHIFT - 10))
36 
37 #ifdef CONFIG_DEBUG_FS
38 #include <linux/debugfs.h>
39 #include <linux/seq_file.h>
40 
41 static struct dentry *bdi_debug_root;
42 
43 static void bdi_debug_init(void)
44 {
45 	bdi_debug_root = debugfs_create_dir("bdi", NULL);
46 }
47 
48 static int bdi_debug_stats_show(struct seq_file *m, void *v)
49 {
50 	struct backing_dev_info *bdi = m->private;
51 	struct bdi_writeback *wb = &bdi->wb;
52 	unsigned long background_thresh;
53 	unsigned long dirty_thresh;
54 	unsigned long wb_thresh;
55 	unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
56 	struct inode *inode;
57 
58 	nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
59 	spin_lock(&wb->list_lock);
60 	list_for_each_entry(inode, &wb->b_dirty, i_io_list)
61 		nr_dirty++;
62 	list_for_each_entry(inode, &wb->b_io, i_io_list)
63 		nr_io++;
64 	list_for_each_entry(inode, &wb->b_more_io, i_io_list)
65 		nr_more_io++;
66 	list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
67 		if (inode->i_state & I_DIRTY_TIME)
68 			nr_dirty_time++;
69 	spin_unlock(&wb->list_lock);
70 
71 	global_dirty_limits(&background_thresh, &dirty_thresh);
72 	wb_thresh = wb_calc_thresh(wb, dirty_thresh);
73 
74 	seq_printf(m,
75 		   "BdiWriteback:       %10lu kB\n"
76 		   "BdiReclaimable:     %10lu kB\n"
77 		   "BdiDirtyThresh:     %10lu kB\n"
78 		   "DirtyThresh:        %10lu kB\n"
79 		   "BackgroundThresh:   %10lu kB\n"
80 		   "BdiDirtied:         %10lu kB\n"
81 		   "BdiWritten:         %10lu kB\n"
82 		   "BdiWriteBandwidth:  %10lu kBps\n"
83 		   "b_dirty:            %10lu\n"
84 		   "b_io:               %10lu\n"
85 		   "b_more_io:          %10lu\n"
86 		   "b_dirty_time:       %10lu\n"
87 		   "bdi_list:           %10u\n"
88 		   "state:              %10lx\n",
89 		   (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
90 		   (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
91 		   K(wb_thresh),
92 		   K(dirty_thresh),
93 		   K(background_thresh),
94 		   (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
95 		   (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
96 		   (unsigned long) K(wb->write_bandwidth),
97 		   nr_dirty,
98 		   nr_io,
99 		   nr_more_io,
100 		   nr_dirty_time,
101 		   !list_empty(&bdi->bdi_list), bdi->wb.state);
102 
103 	return 0;
104 }
105 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
106 
107 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
108 {
109 	bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
110 
111 	debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
112 			    &bdi_debug_stats_fops);
113 }
114 
115 static void bdi_debug_unregister(struct backing_dev_info *bdi)
116 {
117 	debugfs_remove_recursive(bdi->debug_dir);
118 }
119 #else
120 static inline void bdi_debug_init(void)
121 {
122 }
123 static inline void bdi_debug_register(struct backing_dev_info *bdi,
124 				      const char *name)
125 {
126 }
127 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
128 {
129 }
130 #endif
131 
132 static ssize_t read_ahead_kb_store(struct device *dev,
133 				  struct device_attribute *attr,
134 				  const char *buf, size_t count)
135 {
136 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
137 	unsigned long read_ahead_kb;
138 	ssize_t ret;
139 
140 	ret = kstrtoul(buf, 10, &read_ahead_kb);
141 	if (ret < 0)
142 		return ret;
143 
144 	bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
145 
146 	return count;
147 }
148 
149 #define BDI_SHOW(name, expr)						\
150 static ssize_t name##_show(struct device *dev,				\
151 			   struct device_attribute *attr, char *buf)	\
152 {									\
153 	struct backing_dev_info *bdi = dev_get_drvdata(dev);		\
154 									\
155 	return sysfs_emit(buf, "%lld\n", (long long)expr);		\
156 }									\
157 static DEVICE_ATTR_RW(name);
158 
159 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
160 
161 static ssize_t min_ratio_store(struct device *dev,
162 		struct device_attribute *attr, const char *buf, size_t count)
163 {
164 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
165 	unsigned int ratio;
166 	ssize_t ret;
167 
168 	ret = kstrtouint(buf, 10, &ratio);
169 	if (ret < 0)
170 		return ret;
171 
172 	ret = bdi_set_min_ratio(bdi, ratio);
173 	if (!ret)
174 		ret = count;
175 
176 	return ret;
177 }
178 BDI_SHOW(min_ratio, bdi->min_ratio)
179 
180 static ssize_t max_ratio_store(struct device *dev,
181 		struct device_attribute *attr, const char *buf, size_t count)
182 {
183 	struct backing_dev_info *bdi = dev_get_drvdata(dev);
184 	unsigned int ratio;
185 	ssize_t ret;
186 
187 	ret = kstrtouint(buf, 10, &ratio);
188 	if (ret < 0)
189 		return ret;
190 
191 	ret = bdi_set_max_ratio(bdi, ratio);
192 	if (!ret)
193 		ret = count;
194 
195 	return ret;
196 }
197 BDI_SHOW(max_ratio, bdi->max_ratio)
198 
199 static ssize_t stable_pages_required_show(struct device *dev,
200 					  struct device_attribute *attr,
201 					  char *buf)
202 {
203 	dev_warn_once(dev,
204 		"the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
205 	return sysfs_emit(buf, "%d\n", 0);
206 }
207 static DEVICE_ATTR_RO(stable_pages_required);
208 
209 static struct attribute *bdi_dev_attrs[] = {
210 	&dev_attr_read_ahead_kb.attr,
211 	&dev_attr_min_ratio.attr,
212 	&dev_attr_max_ratio.attr,
213 	&dev_attr_stable_pages_required.attr,
214 	NULL,
215 };
216 ATTRIBUTE_GROUPS(bdi_dev);
217 
218 static __init int bdi_class_init(void)
219 {
220 	bdi_class = class_create(THIS_MODULE, "bdi");
221 	if (IS_ERR(bdi_class))
222 		return PTR_ERR(bdi_class);
223 
224 	bdi_class->dev_groups = bdi_dev_groups;
225 	bdi_debug_init();
226 
227 	return 0;
228 }
229 postcore_initcall(bdi_class_init);
230 
231 static int bdi_init(struct backing_dev_info *bdi);
232 
233 static int __init default_bdi_init(void)
234 {
235 	int err;
236 
237 	bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
238 				 WQ_SYSFS, 0);
239 	if (!bdi_wq)
240 		return -ENOMEM;
241 
242 	err = bdi_init(&noop_backing_dev_info);
243 
244 	return err;
245 }
246 subsys_initcall(default_bdi_init);
247 
248 /*
249  * This function is used when the first inode for this wb is marked dirty. It
250  * wakes-up the corresponding bdi thread which should then take care of the
251  * periodic background write-out of dirty inodes. Since the write-out would
252  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
253  * set up a timer which wakes the bdi thread up later.
254  *
255  * Note, we wouldn't bother setting up the timer, but this function is on the
256  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
257  * by delaying the wake-up.
258  *
259  * We have to be careful not to postpone flush work if it is scheduled for
260  * earlier. Thus we use queue_delayed_work().
261  */
262 void wb_wakeup_delayed(struct bdi_writeback *wb)
263 {
264 	unsigned long timeout;
265 
266 	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
267 	spin_lock_bh(&wb->work_lock);
268 	if (test_bit(WB_registered, &wb->state))
269 		queue_delayed_work(bdi_wq, &wb->dwork, timeout);
270 	spin_unlock_bh(&wb->work_lock);
271 }
272 
273 /*
274  * Initial write bandwidth: 100 MB/s
275  */
276 #define INIT_BW		(100 << (20 - PAGE_SHIFT))
277 
278 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
279 		   gfp_t gfp)
280 {
281 	int i, err;
282 
283 	memset(wb, 0, sizeof(*wb));
284 
285 	if (wb != &bdi->wb)
286 		bdi_get(bdi);
287 	wb->bdi = bdi;
288 	wb->last_old_flush = jiffies;
289 	INIT_LIST_HEAD(&wb->b_dirty);
290 	INIT_LIST_HEAD(&wb->b_io);
291 	INIT_LIST_HEAD(&wb->b_more_io);
292 	INIT_LIST_HEAD(&wb->b_dirty_time);
293 	spin_lock_init(&wb->list_lock);
294 
295 	wb->bw_time_stamp = jiffies;
296 	wb->balanced_dirty_ratelimit = INIT_BW;
297 	wb->dirty_ratelimit = INIT_BW;
298 	wb->write_bandwidth = INIT_BW;
299 	wb->avg_write_bandwidth = INIT_BW;
300 
301 	spin_lock_init(&wb->work_lock);
302 	INIT_LIST_HEAD(&wb->work_list);
303 	INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
304 	wb->dirty_sleep = jiffies;
305 
306 	err = fprop_local_init_percpu(&wb->completions, gfp);
307 	if (err)
308 		goto out_put_bdi;
309 
310 	for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
311 		err = percpu_counter_init(&wb->stat[i], 0, gfp);
312 		if (err)
313 			goto out_destroy_stat;
314 	}
315 
316 	return 0;
317 
318 out_destroy_stat:
319 	while (i--)
320 		percpu_counter_destroy(&wb->stat[i]);
321 	fprop_local_destroy_percpu(&wb->completions);
322 out_put_bdi:
323 	if (wb != &bdi->wb)
324 		bdi_put(bdi);
325 	return err;
326 }
327 
328 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
329 
330 /*
331  * Remove bdi from the global list and shutdown any threads we have running
332  */
333 static void wb_shutdown(struct bdi_writeback *wb)
334 {
335 	/* Make sure nobody queues further work */
336 	spin_lock_bh(&wb->work_lock);
337 	if (!test_and_clear_bit(WB_registered, &wb->state)) {
338 		spin_unlock_bh(&wb->work_lock);
339 		return;
340 	}
341 	spin_unlock_bh(&wb->work_lock);
342 
343 	cgwb_remove_from_bdi_list(wb);
344 	/*
345 	 * Drain work list and shutdown the delayed_work.  !WB_registered
346 	 * tells wb_workfn() that @wb is dying and its work_list needs to
347 	 * be drained no matter what.
348 	 */
349 	mod_delayed_work(bdi_wq, &wb->dwork, 0);
350 	flush_delayed_work(&wb->dwork);
351 	WARN_ON(!list_empty(&wb->work_list));
352 }
353 
354 static void wb_exit(struct bdi_writeback *wb)
355 {
356 	int i;
357 
358 	WARN_ON(delayed_work_pending(&wb->dwork));
359 
360 	for (i = 0; i < NR_WB_STAT_ITEMS; i++)
361 		percpu_counter_destroy(&wb->stat[i]);
362 
363 	fprop_local_destroy_percpu(&wb->completions);
364 	if (wb != &wb->bdi->wb)
365 		bdi_put(wb->bdi);
366 }
367 
368 #ifdef CONFIG_CGROUP_WRITEBACK
369 
370 #include <linux/memcontrol.h>
371 
372 /*
373  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, and memcg->cgwb_list.
374  * bdi->cgwb_tree is also RCU protected.
375  */
376 static DEFINE_SPINLOCK(cgwb_lock);
377 static struct workqueue_struct *cgwb_release_wq;
378 
379 static void cgwb_release_workfn(struct work_struct *work)
380 {
381 	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
382 						release_work);
383 	struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
384 
385 	mutex_lock(&wb->bdi->cgwb_release_mutex);
386 	wb_shutdown(wb);
387 
388 	css_put(wb->memcg_css);
389 	css_put(wb->blkcg_css);
390 	mutex_unlock(&wb->bdi->cgwb_release_mutex);
391 
392 	/* triggers blkg destruction if no online users left */
393 	blkcg_unpin_online(blkcg);
394 
395 	fprop_local_destroy_percpu(&wb->memcg_completions);
396 	percpu_ref_exit(&wb->refcnt);
397 	wb_exit(wb);
398 	kfree_rcu(wb, rcu);
399 }
400 
401 static void cgwb_release(struct percpu_ref *refcnt)
402 {
403 	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
404 						refcnt);
405 	queue_work(cgwb_release_wq, &wb->release_work);
406 }
407 
408 static void cgwb_kill(struct bdi_writeback *wb)
409 {
410 	lockdep_assert_held(&cgwb_lock);
411 
412 	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
413 	list_del(&wb->memcg_node);
414 	list_del(&wb->blkcg_node);
415 	percpu_ref_kill(&wb->refcnt);
416 }
417 
418 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
419 {
420 	spin_lock_irq(&cgwb_lock);
421 	list_del_rcu(&wb->bdi_node);
422 	spin_unlock_irq(&cgwb_lock);
423 }
424 
425 static int cgwb_create(struct backing_dev_info *bdi,
426 		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
427 {
428 	struct mem_cgroup *memcg;
429 	struct cgroup_subsys_state *blkcg_css;
430 	struct blkcg *blkcg;
431 	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
432 	struct bdi_writeback *wb;
433 	unsigned long flags;
434 	int ret = 0;
435 
436 	memcg = mem_cgroup_from_css(memcg_css);
437 	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
438 	blkcg = css_to_blkcg(blkcg_css);
439 	memcg_cgwb_list = &memcg->cgwb_list;
440 	blkcg_cgwb_list = &blkcg->cgwb_list;
441 
442 	/* look up again under lock and discard on blkcg mismatch */
443 	spin_lock_irqsave(&cgwb_lock, flags);
444 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
445 	if (wb && wb->blkcg_css != blkcg_css) {
446 		cgwb_kill(wb);
447 		wb = NULL;
448 	}
449 	spin_unlock_irqrestore(&cgwb_lock, flags);
450 	if (wb)
451 		goto out_put;
452 
453 	/* need to create a new one */
454 	wb = kmalloc(sizeof(*wb), gfp);
455 	if (!wb) {
456 		ret = -ENOMEM;
457 		goto out_put;
458 	}
459 
460 	ret = wb_init(wb, bdi, gfp);
461 	if (ret)
462 		goto err_free;
463 
464 	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
465 	if (ret)
466 		goto err_wb_exit;
467 
468 	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
469 	if (ret)
470 		goto err_ref_exit;
471 
472 	wb->memcg_css = memcg_css;
473 	wb->blkcg_css = blkcg_css;
474 	INIT_WORK(&wb->release_work, cgwb_release_workfn);
475 	set_bit(WB_registered, &wb->state);
476 
477 	/*
478 	 * The root wb determines the registered state of the whole bdi and
479 	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
480 	 * whether they're still online.  Don't link @wb if any is dead.
481 	 * See wb_memcg_offline() and wb_blkcg_offline().
482 	 */
483 	ret = -ENODEV;
484 	spin_lock_irqsave(&cgwb_lock, flags);
485 	if (test_bit(WB_registered, &bdi->wb.state) &&
486 	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
487 		/* we might have raced another instance of this function */
488 		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
489 		if (!ret) {
490 			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
491 			list_add(&wb->memcg_node, memcg_cgwb_list);
492 			list_add(&wb->blkcg_node, blkcg_cgwb_list);
493 			blkcg_pin_online(blkcg);
494 			css_get(memcg_css);
495 			css_get(blkcg_css);
496 		}
497 	}
498 	spin_unlock_irqrestore(&cgwb_lock, flags);
499 	if (ret) {
500 		if (ret == -EEXIST)
501 			ret = 0;
502 		goto err_fprop_exit;
503 	}
504 	goto out_put;
505 
506 err_fprop_exit:
507 	fprop_local_destroy_percpu(&wb->memcg_completions);
508 err_ref_exit:
509 	percpu_ref_exit(&wb->refcnt);
510 err_wb_exit:
511 	wb_exit(wb);
512 err_free:
513 	kfree(wb);
514 out_put:
515 	css_put(blkcg_css);
516 	return ret;
517 }
518 
519 /**
520  * wb_get_lookup - get wb for a given memcg
521  * @bdi: target bdi
522  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
523  *
524  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
525  * refcount incremented.
526  *
527  * This function uses css_get() on @memcg_css and thus expects its refcnt
528  * to be positive on invocation.  IOW, rcu_read_lock() protection on
529  * @memcg_css isn't enough.  try_get it before calling this function.
530  *
531  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
532  * memcg on the default hierarchy, memcg association is guaranteed to be
533  * more specific (equal or descendant to the associated blkcg) and thus can
534  * identify both the memcg and blkcg associations.
535  *
536  * Because the blkcg associated with a memcg may change as blkcg is enabled
537  * and disabled closer to root in the hierarchy, each wb keeps track of
538  * both the memcg and blkcg associated with it and verifies the blkcg on
539  * each lookup.  On mismatch, the existing wb is discarded and a new one is
540  * created.
541  */
542 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
543 				    struct cgroup_subsys_state *memcg_css)
544 {
545 	struct bdi_writeback *wb;
546 
547 	if (!memcg_css->parent)
548 		return &bdi->wb;
549 
550 	rcu_read_lock();
551 	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
552 	if (wb) {
553 		struct cgroup_subsys_state *blkcg_css;
554 
555 		/* see whether the blkcg association has changed */
556 		blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
557 		if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
558 			wb = NULL;
559 		css_put(blkcg_css);
560 	}
561 	rcu_read_unlock();
562 
563 	return wb;
564 }
565 
566 /**
567  * wb_get_create - get wb for a given memcg, create if necessary
568  * @bdi: target bdi
569  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
570  * @gfp: allocation mask to use
571  *
572  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
573  * create one.  See wb_get_lookup() for more details.
574  */
575 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
576 				    struct cgroup_subsys_state *memcg_css,
577 				    gfp_t gfp)
578 {
579 	struct bdi_writeback *wb;
580 
581 	might_sleep_if(gfpflags_allow_blocking(gfp));
582 
583 	if (!memcg_css->parent)
584 		return &bdi->wb;
585 
586 	do {
587 		wb = wb_get_lookup(bdi, memcg_css);
588 	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
589 
590 	return wb;
591 }
592 
593 static int cgwb_bdi_init(struct backing_dev_info *bdi)
594 {
595 	int ret;
596 
597 	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
598 	mutex_init(&bdi->cgwb_release_mutex);
599 	init_rwsem(&bdi->wb_switch_rwsem);
600 
601 	ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
602 	if (!ret) {
603 		bdi->wb.memcg_css = &root_mem_cgroup->css;
604 		bdi->wb.blkcg_css = blkcg_root_css;
605 	}
606 	return ret;
607 }
608 
609 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
610 {
611 	struct radix_tree_iter iter;
612 	void **slot;
613 	struct bdi_writeback *wb;
614 
615 	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
616 
617 	spin_lock_irq(&cgwb_lock);
618 	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
619 		cgwb_kill(*slot);
620 	spin_unlock_irq(&cgwb_lock);
621 
622 	mutex_lock(&bdi->cgwb_release_mutex);
623 	spin_lock_irq(&cgwb_lock);
624 	while (!list_empty(&bdi->wb_list)) {
625 		wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
626 				      bdi_node);
627 		spin_unlock_irq(&cgwb_lock);
628 		wb_shutdown(wb);
629 		spin_lock_irq(&cgwb_lock);
630 	}
631 	spin_unlock_irq(&cgwb_lock);
632 	mutex_unlock(&bdi->cgwb_release_mutex);
633 }
634 
635 /**
636  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
637  * @memcg: memcg being offlined
638  *
639  * Also prevents creation of any new wb's associated with @memcg.
640  */
641 void wb_memcg_offline(struct mem_cgroup *memcg)
642 {
643 	struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
644 	struct bdi_writeback *wb, *next;
645 
646 	spin_lock_irq(&cgwb_lock);
647 	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
648 		cgwb_kill(wb);
649 	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
650 	spin_unlock_irq(&cgwb_lock);
651 }
652 
653 /**
654  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
655  * @blkcg: blkcg being offlined
656  *
657  * Also prevents creation of any new wb's associated with @blkcg.
658  */
659 void wb_blkcg_offline(struct blkcg *blkcg)
660 {
661 	struct bdi_writeback *wb, *next;
662 
663 	spin_lock_irq(&cgwb_lock);
664 	list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
665 		cgwb_kill(wb);
666 	blkcg->cgwb_list.next = NULL;	/* prevent new wb's */
667 	spin_unlock_irq(&cgwb_lock);
668 }
669 
670 static void cgwb_bdi_register(struct backing_dev_info *bdi)
671 {
672 	spin_lock_irq(&cgwb_lock);
673 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
674 	spin_unlock_irq(&cgwb_lock);
675 }
676 
677 static int __init cgwb_init(void)
678 {
679 	/*
680 	 * There can be many concurrent release work items overwhelming
681 	 * system_wq.  Put them in a separate wq and limit concurrency.
682 	 * There's no point in executing many of these in parallel.
683 	 */
684 	cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
685 	if (!cgwb_release_wq)
686 		return -ENOMEM;
687 
688 	return 0;
689 }
690 subsys_initcall(cgwb_init);
691 
692 #else	/* CONFIG_CGROUP_WRITEBACK */
693 
694 static int cgwb_bdi_init(struct backing_dev_info *bdi)
695 {
696 	return wb_init(&bdi->wb, bdi, GFP_KERNEL);
697 }
698 
699 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
700 
701 static void cgwb_bdi_register(struct backing_dev_info *bdi)
702 {
703 	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
704 }
705 
706 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
707 {
708 	list_del_rcu(&wb->bdi_node);
709 }
710 
711 #endif	/* CONFIG_CGROUP_WRITEBACK */
712 
713 static int bdi_init(struct backing_dev_info *bdi)
714 {
715 	int ret;
716 
717 	bdi->dev = NULL;
718 
719 	kref_init(&bdi->refcnt);
720 	bdi->min_ratio = 0;
721 	bdi->max_ratio = 100;
722 	bdi->max_prop_frac = FPROP_FRAC_BASE;
723 	INIT_LIST_HEAD(&bdi->bdi_list);
724 	INIT_LIST_HEAD(&bdi->wb_list);
725 	init_waitqueue_head(&bdi->wb_waitq);
726 
727 	ret = cgwb_bdi_init(bdi);
728 
729 	return ret;
730 }
731 
732 struct backing_dev_info *bdi_alloc(int node_id)
733 {
734 	struct backing_dev_info *bdi;
735 
736 	bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
737 	if (!bdi)
738 		return NULL;
739 
740 	if (bdi_init(bdi)) {
741 		kfree(bdi);
742 		return NULL;
743 	}
744 	bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
745 	bdi->ra_pages = VM_READAHEAD_PAGES;
746 	bdi->io_pages = VM_READAHEAD_PAGES;
747 	return bdi;
748 }
749 EXPORT_SYMBOL(bdi_alloc);
750 
751 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
752 {
753 	struct rb_node **p = &bdi_tree.rb_node;
754 	struct rb_node *parent = NULL;
755 	struct backing_dev_info *bdi;
756 
757 	lockdep_assert_held(&bdi_lock);
758 
759 	while (*p) {
760 		parent = *p;
761 		bdi = rb_entry(parent, struct backing_dev_info, rb_node);
762 
763 		if (bdi->id > id)
764 			p = &(*p)->rb_left;
765 		else if (bdi->id < id)
766 			p = &(*p)->rb_right;
767 		else
768 			break;
769 	}
770 
771 	if (parentp)
772 		*parentp = parent;
773 	return p;
774 }
775 
776 /**
777  * bdi_get_by_id - lookup and get bdi from its id
778  * @id: bdi id to lookup
779  *
780  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
781  * doesn't exist or is already unregistered.
782  */
783 struct backing_dev_info *bdi_get_by_id(u64 id)
784 {
785 	struct backing_dev_info *bdi = NULL;
786 	struct rb_node **p;
787 
788 	spin_lock_bh(&bdi_lock);
789 	p = bdi_lookup_rb_node(id, NULL);
790 	if (*p) {
791 		bdi = rb_entry(*p, struct backing_dev_info, rb_node);
792 		bdi_get(bdi);
793 	}
794 	spin_unlock_bh(&bdi_lock);
795 
796 	return bdi;
797 }
798 
799 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
800 {
801 	struct device *dev;
802 	struct rb_node *parent, **p;
803 
804 	if (bdi->dev)	/* The driver needs to use separate queues per device */
805 		return 0;
806 
807 	vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
808 	dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
809 	if (IS_ERR(dev))
810 		return PTR_ERR(dev);
811 
812 	cgwb_bdi_register(bdi);
813 	bdi->dev = dev;
814 
815 	bdi_debug_register(bdi, dev_name(dev));
816 	set_bit(WB_registered, &bdi->wb.state);
817 
818 	spin_lock_bh(&bdi_lock);
819 
820 	bdi->id = ++bdi_id_cursor;
821 
822 	p = bdi_lookup_rb_node(bdi->id, &parent);
823 	rb_link_node(&bdi->rb_node, parent, p);
824 	rb_insert_color(&bdi->rb_node, &bdi_tree);
825 
826 	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
827 
828 	spin_unlock_bh(&bdi_lock);
829 
830 	trace_writeback_bdi_register(bdi);
831 	return 0;
832 }
833 
834 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
835 {
836 	va_list args;
837 	int ret;
838 
839 	va_start(args, fmt);
840 	ret = bdi_register_va(bdi, fmt, args);
841 	va_end(args);
842 	return ret;
843 }
844 EXPORT_SYMBOL(bdi_register);
845 
846 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
847 {
848 	WARN_ON_ONCE(bdi->owner);
849 	bdi->owner = owner;
850 	get_device(owner);
851 }
852 
853 /*
854  * Remove bdi from bdi_list, and ensure that it is no longer visible
855  */
856 static void bdi_remove_from_list(struct backing_dev_info *bdi)
857 {
858 	spin_lock_bh(&bdi_lock);
859 	rb_erase(&bdi->rb_node, &bdi_tree);
860 	list_del_rcu(&bdi->bdi_list);
861 	spin_unlock_bh(&bdi_lock);
862 
863 	synchronize_rcu_expedited();
864 }
865 
866 void bdi_unregister(struct backing_dev_info *bdi)
867 {
868 	/* make sure nobody finds us on the bdi_list anymore */
869 	bdi_remove_from_list(bdi);
870 	wb_shutdown(&bdi->wb);
871 	cgwb_bdi_unregister(bdi);
872 
873 	if (bdi->dev) {
874 		bdi_debug_unregister(bdi);
875 		device_unregister(bdi->dev);
876 		bdi->dev = NULL;
877 	}
878 
879 	if (bdi->owner) {
880 		put_device(bdi->owner);
881 		bdi->owner = NULL;
882 	}
883 }
884 
885 static void release_bdi(struct kref *ref)
886 {
887 	struct backing_dev_info *bdi =
888 			container_of(ref, struct backing_dev_info, refcnt);
889 
890 	if (test_bit(WB_registered, &bdi->wb.state))
891 		bdi_unregister(bdi);
892 	WARN_ON_ONCE(bdi->dev);
893 	wb_exit(&bdi->wb);
894 	kfree(bdi);
895 }
896 
897 void bdi_put(struct backing_dev_info *bdi)
898 {
899 	kref_put(&bdi->refcnt, release_bdi);
900 }
901 EXPORT_SYMBOL(bdi_put);
902 
903 const char *bdi_dev_name(struct backing_dev_info *bdi)
904 {
905 	if (!bdi || !bdi->dev)
906 		return bdi_unknown_name;
907 	return bdi->dev_name;
908 }
909 EXPORT_SYMBOL_GPL(bdi_dev_name);
910 
911 static wait_queue_head_t congestion_wqh[2] = {
912 		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
913 		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
914 	};
915 static atomic_t nr_wb_congested[2];
916 
917 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
918 {
919 	wait_queue_head_t *wqh = &congestion_wqh[sync];
920 	enum wb_congested_state bit;
921 
922 	bit = sync ? WB_sync_congested : WB_async_congested;
923 	if (test_and_clear_bit(bit, &bdi->wb.congested))
924 		atomic_dec(&nr_wb_congested[sync]);
925 	smp_mb__after_atomic();
926 	if (waitqueue_active(wqh))
927 		wake_up(wqh);
928 }
929 EXPORT_SYMBOL(clear_bdi_congested);
930 
931 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
932 {
933 	enum wb_congested_state bit;
934 
935 	bit = sync ? WB_sync_congested : WB_async_congested;
936 	if (!test_and_set_bit(bit, &bdi->wb.congested))
937 		atomic_inc(&nr_wb_congested[sync]);
938 }
939 EXPORT_SYMBOL(set_bdi_congested);
940 
941 /**
942  * congestion_wait - wait for a backing_dev to become uncongested
943  * @sync: SYNC or ASYNC IO
944  * @timeout: timeout in jiffies
945  *
946  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
947  * write congestion.  If no backing_devs are congested then just wait for the
948  * next write to be completed.
949  */
950 long congestion_wait(int sync, long timeout)
951 {
952 	long ret;
953 	unsigned long start = jiffies;
954 	DEFINE_WAIT(wait);
955 	wait_queue_head_t *wqh = &congestion_wqh[sync];
956 
957 	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
958 	ret = io_schedule_timeout(timeout);
959 	finish_wait(wqh, &wait);
960 
961 	trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
962 					jiffies_to_usecs(jiffies - start));
963 
964 	return ret;
965 }
966 EXPORT_SYMBOL(congestion_wait);
967 
968 /**
969  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
970  * @sync: SYNC or ASYNC IO
971  * @timeout: timeout in jiffies
972  *
973  * In the event of a congested backing_dev (any backing_dev) this waits
974  * for up to @timeout jiffies for either a BDI to exit congestion of the
975  * given @sync queue or a write to complete.
976  *
977  * The return value is 0 if the sleep is for the full timeout. Otherwise,
978  * it is the number of jiffies that were still remaining when the function
979  * returned. return_value == timeout implies the function did not sleep.
980  */
981 long wait_iff_congested(int sync, long timeout)
982 {
983 	long ret;
984 	unsigned long start = jiffies;
985 	DEFINE_WAIT(wait);
986 	wait_queue_head_t *wqh = &congestion_wqh[sync];
987 
988 	/*
989 	 * If there is no congestion, yield if necessary instead
990 	 * of sleeping on the congestion queue
991 	 */
992 	if (atomic_read(&nr_wb_congested[sync]) == 0) {
993 		cond_resched();
994 
995 		/* In case we scheduled, work out time remaining */
996 		ret = timeout - (jiffies - start);
997 		if (ret < 0)
998 			ret = 0;
999 
1000 		goto out;
1001 	}
1002 
1003 	/* Sleep until uncongested or a write happens */
1004 	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1005 	ret = io_schedule_timeout(timeout);
1006 	finish_wait(wqh, &wait);
1007 
1008 out:
1009 	trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1010 					jiffies_to_usecs(jiffies - start));
1011 
1012 	return ret;
1013 }
1014 EXPORT_SYMBOL(wait_iff_congested);
1015