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 EXPORT_SYMBOL_GPL(bdi_list);
35
36 /* bdi_wq serves all asynchronous writeback tasks */
37 struct workqueue_struct *bdi_wq;
38
39 #ifdef CONFIG_DEBUG_FS
40 #include <linux/debugfs.h>
41 #include <linux/seq_file.h>
42
43 struct wb_stats {
44 unsigned long nr_dirty;
45 unsigned long nr_io;
46 unsigned long nr_more_io;
47 unsigned long nr_dirty_time;
48 unsigned long nr_writeback;
49 unsigned long nr_reclaimable;
50 unsigned long nr_dirtied;
51 unsigned long nr_written;
52 unsigned long dirty_thresh;
53 unsigned long wb_thresh;
54 };
55
56 static struct dentry *bdi_debug_root;
57
bdi_debug_init(void)58 static void bdi_debug_init(void)
59 {
60 bdi_debug_root = debugfs_create_dir("bdi", NULL);
61 }
62
collect_wb_stats(struct wb_stats * stats,struct bdi_writeback * wb)63 static void collect_wb_stats(struct wb_stats *stats,
64 struct bdi_writeback *wb)
65 {
66 struct inode *inode;
67
68 spin_lock(&wb->list_lock);
69 list_for_each_entry(inode, &wb->b_dirty, i_io_list)
70 stats->nr_dirty++;
71 list_for_each_entry(inode, &wb->b_io, i_io_list)
72 stats->nr_io++;
73 list_for_each_entry(inode, &wb->b_more_io, i_io_list)
74 stats->nr_more_io++;
75 list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
76 if (inode->i_state & I_DIRTY_TIME)
77 stats->nr_dirty_time++;
78 spin_unlock(&wb->list_lock);
79
80 stats->nr_writeback += wb_stat(wb, WB_WRITEBACK);
81 stats->nr_reclaimable += wb_stat(wb, WB_RECLAIMABLE);
82 stats->nr_dirtied += wb_stat(wb, WB_DIRTIED);
83 stats->nr_written += wb_stat(wb, WB_WRITTEN);
84 stats->wb_thresh += wb_calc_thresh(wb, stats->dirty_thresh);
85 }
86
87 #ifdef CONFIG_CGROUP_WRITEBACK
bdi_collect_stats(struct backing_dev_info * bdi,struct wb_stats * stats)88 static void bdi_collect_stats(struct backing_dev_info *bdi,
89 struct wb_stats *stats)
90 {
91 struct bdi_writeback *wb;
92
93 rcu_read_lock();
94 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) {
95 if (!wb_tryget(wb))
96 continue;
97
98 collect_wb_stats(stats, wb);
99 wb_put(wb);
100 }
101 rcu_read_unlock();
102 }
103 #else
bdi_collect_stats(struct backing_dev_info * bdi,struct wb_stats * stats)104 static void bdi_collect_stats(struct backing_dev_info *bdi,
105 struct wb_stats *stats)
106 {
107 collect_wb_stats(stats, &bdi->wb);
108 }
109 #endif
110
bdi_debug_stats_show(struct seq_file * m,void * v)111 static int bdi_debug_stats_show(struct seq_file *m, void *v)
112 {
113 struct backing_dev_info *bdi = m->private;
114 unsigned long background_thresh;
115 unsigned long dirty_thresh;
116 struct wb_stats stats;
117 unsigned long tot_bw;
118
119 global_dirty_limits(&background_thresh, &dirty_thresh);
120
121 memset(&stats, 0, sizeof(stats));
122 stats.dirty_thresh = dirty_thresh;
123 bdi_collect_stats(bdi, &stats);
124 tot_bw = atomic_long_read(&bdi->tot_write_bandwidth);
125
126 seq_printf(m,
127 "BdiWriteback: %10lu kB\n"
128 "BdiReclaimable: %10lu kB\n"
129 "BdiDirtyThresh: %10lu kB\n"
130 "DirtyThresh: %10lu kB\n"
131 "BackgroundThresh: %10lu kB\n"
132 "BdiDirtied: %10lu kB\n"
133 "BdiWritten: %10lu kB\n"
134 "BdiWriteBandwidth: %10lu kBps\n"
135 "b_dirty: %10lu\n"
136 "b_io: %10lu\n"
137 "b_more_io: %10lu\n"
138 "b_dirty_time: %10lu\n"
139 "bdi_list: %10u\n"
140 "state: %10lx\n",
141 K(stats.nr_writeback),
142 K(stats.nr_reclaimable),
143 K(stats.wb_thresh),
144 K(dirty_thresh),
145 K(background_thresh),
146 K(stats.nr_dirtied),
147 K(stats.nr_written),
148 K(tot_bw),
149 stats.nr_dirty,
150 stats.nr_io,
151 stats.nr_more_io,
152 stats.nr_dirty_time,
153 !list_empty(&bdi->bdi_list), bdi->wb.state);
154
155 return 0;
156 }
157 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
158
wb_stats_show(struct seq_file * m,struct bdi_writeback * wb,struct wb_stats * stats)159 static void wb_stats_show(struct seq_file *m, struct bdi_writeback *wb,
160 struct wb_stats *stats)
161 {
162
163 seq_printf(m,
164 "WbCgIno: %10lu\n"
165 "WbWriteback: %10lu kB\n"
166 "WbReclaimable: %10lu kB\n"
167 "WbDirtyThresh: %10lu kB\n"
168 "WbDirtied: %10lu kB\n"
169 "WbWritten: %10lu kB\n"
170 "WbWriteBandwidth: %10lu kBps\n"
171 "b_dirty: %10lu\n"
172 "b_io: %10lu\n"
173 "b_more_io: %10lu\n"
174 "b_dirty_time: %10lu\n"
175 "state: %10lx\n\n",
176 #ifdef CONFIG_CGROUP_WRITEBACK
177 cgroup_ino(wb->memcg_css->cgroup),
178 #else
179 1ul,
180 #endif
181 K(stats->nr_writeback),
182 K(stats->nr_reclaimable),
183 K(stats->wb_thresh),
184 K(stats->nr_dirtied),
185 K(stats->nr_written),
186 K(wb->avg_write_bandwidth),
187 stats->nr_dirty,
188 stats->nr_io,
189 stats->nr_more_io,
190 stats->nr_dirty_time,
191 wb->state);
192 }
193
cgwb_debug_stats_show(struct seq_file * m,void * v)194 static int cgwb_debug_stats_show(struct seq_file *m, void *v)
195 {
196 struct backing_dev_info *bdi = m->private;
197 unsigned long background_thresh;
198 unsigned long dirty_thresh;
199 struct bdi_writeback *wb;
200
201 global_dirty_limits(&background_thresh, &dirty_thresh);
202
203 rcu_read_lock();
204 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) {
205 struct wb_stats stats = { .dirty_thresh = dirty_thresh };
206
207 if (!wb_tryget(wb))
208 continue;
209
210 collect_wb_stats(&stats, wb);
211
212 /*
213 * Calculate thresh of wb in writeback cgroup which is min of
214 * thresh in global domain and thresh in cgroup domain. Drop
215 * rcu lock because cgwb_calc_thresh may sleep in
216 * cgroup_rstat_flush. We can do so here because we have a ref.
217 */
218 if (mem_cgroup_wb_domain(wb)) {
219 rcu_read_unlock();
220 stats.wb_thresh = min(stats.wb_thresh, cgwb_calc_thresh(wb));
221 rcu_read_lock();
222 }
223
224 wb_stats_show(m, wb, &stats);
225
226 wb_put(wb);
227 }
228 rcu_read_unlock();
229
230 return 0;
231 }
232 DEFINE_SHOW_ATTRIBUTE(cgwb_debug_stats);
233
bdi_debug_register(struct backing_dev_info * bdi,const char * name)234 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
235 {
236 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
237
238 debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
239 &bdi_debug_stats_fops);
240 debugfs_create_file("wb_stats", 0444, bdi->debug_dir, bdi,
241 &cgwb_debug_stats_fops);
242 }
243
bdi_debug_unregister(struct backing_dev_info * bdi)244 static void bdi_debug_unregister(struct backing_dev_info *bdi)
245 {
246 debugfs_remove_recursive(bdi->debug_dir);
247 }
248 #else /* CONFIG_DEBUG_FS */
bdi_debug_init(void)249 static inline void bdi_debug_init(void)
250 {
251 }
bdi_debug_register(struct backing_dev_info * bdi,const char * name)252 static inline void bdi_debug_register(struct backing_dev_info *bdi,
253 const char *name)
254 {
255 }
bdi_debug_unregister(struct backing_dev_info * bdi)256 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
257 {
258 }
259 #endif /* CONFIG_DEBUG_FS */
260
read_ahead_kb_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)261 static ssize_t read_ahead_kb_store(struct device *dev,
262 struct device_attribute *attr,
263 const char *buf, size_t count)
264 {
265 struct backing_dev_info *bdi = dev_get_drvdata(dev);
266 unsigned long read_ahead_kb;
267 ssize_t ret;
268
269 ret = kstrtoul(buf, 10, &read_ahead_kb);
270 if (ret < 0)
271 return ret;
272
273 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
274
275 return count;
276 }
277
278 #define BDI_SHOW(name, expr) \
279 static ssize_t name##_show(struct device *dev, \
280 struct device_attribute *attr, char *buf) \
281 { \
282 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
283 \
284 return sysfs_emit(buf, "%lld\n", (long long)expr); \
285 } \
286 static DEVICE_ATTR_RW(name);
287
288 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
289
min_ratio_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)290 static ssize_t min_ratio_store(struct device *dev,
291 struct device_attribute *attr, const char *buf, size_t count)
292 {
293 struct backing_dev_info *bdi = dev_get_drvdata(dev);
294 unsigned int ratio;
295 ssize_t ret;
296
297 ret = kstrtouint(buf, 10, &ratio);
298 if (ret < 0)
299 return ret;
300
301 ret = bdi_set_min_ratio(bdi, ratio);
302 if (!ret)
303 ret = count;
304
305 return ret;
306 }
307 BDI_SHOW(min_ratio, bdi->min_ratio / BDI_RATIO_SCALE)
308
min_ratio_fine_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)309 static ssize_t min_ratio_fine_store(struct device *dev,
310 struct device_attribute *attr, const char *buf, size_t count)
311 {
312 struct backing_dev_info *bdi = dev_get_drvdata(dev);
313 unsigned int ratio;
314 ssize_t ret;
315
316 ret = kstrtouint(buf, 10, &ratio);
317 if (ret < 0)
318 return ret;
319
320 ret = bdi_set_min_ratio_no_scale(bdi, ratio);
321 if (!ret)
322 ret = count;
323
324 return ret;
325 }
326 BDI_SHOW(min_ratio_fine, bdi->min_ratio)
327
max_ratio_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)328 static ssize_t max_ratio_store(struct device *dev,
329 struct device_attribute *attr, const char *buf, size_t count)
330 {
331 struct backing_dev_info *bdi = dev_get_drvdata(dev);
332 unsigned int ratio;
333 ssize_t ret;
334
335 ret = kstrtouint(buf, 10, &ratio);
336 if (ret < 0)
337 return ret;
338
339 ret = bdi_set_max_ratio(bdi, ratio);
340 if (!ret)
341 ret = count;
342
343 return ret;
344 }
345 BDI_SHOW(max_ratio, bdi->max_ratio / BDI_RATIO_SCALE)
346
max_ratio_fine_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)347 static ssize_t max_ratio_fine_store(struct device *dev,
348 struct device_attribute *attr, const char *buf, size_t count)
349 {
350 struct backing_dev_info *bdi = dev_get_drvdata(dev);
351 unsigned int ratio;
352 ssize_t ret;
353
354 ret = kstrtouint(buf, 10, &ratio);
355 if (ret < 0)
356 return ret;
357
358 ret = bdi_set_max_ratio_no_scale(bdi, ratio);
359 if (!ret)
360 ret = count;
361
362 return ret;
363 }
364 BDI_SHOW(max_ratio_fine, bdi->max_ratio)
365
min_bytes_show(struct device * dev,struct device_attribute * attr,char * buf)366 static ssize_t min_bytes_show(struct device *dev,
367 struct device_attribute *attr,
368 char *buf)
369 {
370 struct backing_dev_info *bdi = dev_get_drvdata(dev);
371
372 return sysfs_emit(buf, "%llu\n", bdi_get_min_bytes(bdi));
373 }
374
min_bytes_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)375 static ssize_t min_bytes_store(struct device *dev,
376 struct device_attribute *attr, const char *buf, size_t count)
377 {
378 struct backing_dev_info *bdi = dev_get_drvdata(dev);
379 u64 bytes;
380 ssize_t ret;
381
382 ret = kstrtoull(buf, 10, &bytes);
383 if (ret < 0)
384 return ret;
385
386 ret = bdi_set_min_bytes(bdi, bytes);
387 if (!ret)
388 ret = count;
389
390 return ret;
391 }
392 static DEVICE_ATTR_RW(min_bytes);
393
max_bytes_show(struct device * dev,struct device_attribute * attr,char * buf)394 static ssize_t max_bytes_show(struct device *dev,
395 struct device_attribute *attr,
396 char *buf)
397 {
398 struct backing_dev_info *bdi = dev_get_drvdata(dev);
399
400 return sysfs_emit(buf, "%llu\n", bdi_get_max_bytes(bdi));
401 }
402
max_bytes_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)403 static ssize_t max_bytes_store(struct device *dev,
404 struct device_attribute *attr, const char *buf, size_t count)
405 {
406 struct backing_dev_info *bdi = dev_get_drvdata(dev);
407 u64 bytes;
408 ssize_t ret;
409
410 ret = kstrtoull(buf, 10, &bytes);
411 if (ret < 0)
412 return ret;
413
414 ret = bdi_set_max_bytes(bdi, bytes);
415 if (!ret)
416 ret = count;
417
418 return ret;
419 }
420 static DEVICE_ATTR_RW(max_bytes);
421
stable_pages_required_show(struct device * dev,struct device_attribute * attr,char * buf)422 static ssize_t stable_pages_required_show(struct device *dev,
423 struct device_attribute *attr,
424 char *buf)
425 {
426 dev_warn_once(dev,
427 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
428 return sysfs_emit(buf, "%d\n", 0);
429 }
430 static DEVICE_ATTR_RO(stable_pages_required);
431
strict_limit_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)432 static ssize_t strict_limit_store(struct device *dev,
433 struct device_attribute *attr, const char *buf, size_t count)
434 {
435 struct backing_dev_info *bdi = dev_get_drvdata(dev);
436 unsigned int strict_limit;
437 ssize_t ret;
438
439 ret = kstrtouint(buf, 10, &strict_limit);
440 if (ret < 0)
441 return ret;
442
443 ret = bdi_set_strict_limit(bdi, strict_limit);
444 if (!ret)
445 ret = count;
446
447 return ret;
448 }
449
strict_limit_show(struct device * dev,struct device_attribute * attr,char * buf)450 static ssize_t strict_limit_show(struct device *dev,
451 struct device_attribute *attr, char *buf)
452 {
453 struct backing_dev_info *bdi = dev_get_drvdata(dev);
454
455 return sysfs_emit(buf, "%d\n",
456 !!(bdi->capabilities & BDI_CAP_STRICTLIMIT));
457 }
458 static DEVICE_ATTR_RW(strict_limit);
459
460 static struct attribute *bdi_dev_attrs[] = {
461 &dev_attr_read_ahead_kb.attr,
462 &dev_attr_min_ratio.attr,
463 &dev_attr_min_ratio_fine.attr,
464 &dev_attr_max_ratio.attr,
465 &dev_attr_max_ratio_fine.attr,
466 &dev_attr_min_bytes.attr,
467 &dev_attr_max_bytes.attr,
468 &dev_attr_stable_pages_required.attr,
469 &dev_attr_strict_limit.attr,
470 NULL,
471 };
472 ATTRIBUTE_GROUPS(bdi_dev);
473
474 static const struct class bdi_class = {
475 .name = "bdi",
476 .dev_groups = bdi_dev_groups,
477 };
478
bdi_class_init(void)479 static __init int bdi_class_init(void)
480 {
481 int ret;
482
483 ret = class_register(&bdi_class);
484 if (ret)
485 return ret;
486
487 bdi_debug_init();
488
489 return 0;
490 }
491 postcore_initcall(bdi_class_init);
492
default_bdi_init(void)493 static int __init default_bdi_init(void)
494 {
495 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
496 WQ_SYSFS, 0);
497 if (!bdi_wq)
498 return -ENOMEM;
499 return 0;
500 }
501 subsys_initcall(default_bdi_init);
502
wb_update_bandwidth_workfn(struct work_struct * work)503 static void wb_update_bandwidth_workfn(struct work_struct *work)
504 {
505 struct bdi_writeback *wb = container_of(to_delayed_work(work),
506 struct bdi_writeback, bw_dwork);
507
508 wb_update_bandwidth(wb);
509 }
510
511 /*
512 * Initial write bandwidth: 100 MB/s
513 */
514 #define INIT_BW (100 << (20 - PAGE_SHIFT))
515
wb_init(struct bdi_writeback * wb,struct backing_dev_info * bdi,gfp_t gfp)516 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
517 gfp_t gfp)
518 {
519 int err;
520
521 memset(wb, 0, sizeof(*wb));
522
523 wb->bdi = bdi;
524 wb->last_old_flush = jiffies;
525 INIT_LIST_HEAD(&wb->b_dirty);
526 INIT_LIST_HEAD(&wb->b_io);
527 INIT_LIST_HEAD(&wb->b_more_io);
528 INIT_LIST_HEAD(&wb->b_dirty_time);
529 spin_lock_init(&wb->list_lock);
530
531 atomic_set(&wb->writeback_inodes, 0);
532 wb->bw_time_stamp = jiffies;
533 wb->balanced_dirty_ratelimit = INIT_BW;
534 wb->dirty_ratelimit = INIT_BW;
535 wb->write_bandwidth = INIT_BW;
536 wb->avg_write_bandwidth = INIT_BW;
537
538 spin_lock_init(&wb->work_lock);
539 INIT_LIST_HEAD(&wb->work_list);
540 INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
541 INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
542
543 err = fprop_local_init_percpu(&wb->completions, gfp);
544 if (err)
545 return err;
546
547 err = percpu_counter_init_many(wb->stat, 0, gfp, NR_WB_STAT_ITEMS);
548 if (err)
549 fprop_local_destroy_percpu(&wb->completions);
550
551 return err;
552 }
553
554 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
555
556 /*
557 * Remove bdi from the global list and shutdown any threads we have running
558 */
wb_shutdown(struct bdi_writeback * wb)559 static void wb_shutdown(struct bdi_writeback *wb)
560 {
561 /* Make sure nobody queues further work */
562 spin_lock_irq(&wb->work_lock);
563 if (!test_and_clear_bit(WB_registered, &wb->state)) {
564 spin_unlock_irq(&wb->work_lock);
565 return;
566 }
567 spin_unlock_irq(&wb->work_lock);
568
569 cgwb_remove_from_bdi_list(wb);
570 /*
571 * Drain work list and shutdown the delayed_work. !WB_registered
572 * tells wb_workfn() that @wb is dying and its work_list needs to
573 * be drained no matter what.
574 */
575 mod_delayed_work(bdi_wq, &wb->dwork, 0);
576 flush_delayed_work(&wb->dwork);
577 WARN_ON(!list_empty(&wb->work_list));
578 flush_delayed_work(&wb->bw_dwork);
579 }
580
wb_exit(struct bdi_writeback * wb)581 static void wb_exit(struct bdi_writeback *wb)
582 {
583 WARN_ON(delayed_work_pending(&wb->dwork));
584 percpu_counter_destroy_many(wb->stat, NR_WB_STAT_ITEMS);
585 fprop_local_destroy_percpu(&wb->completions);
586 }
587
588 #ifdef CONFIG_CGROUP_WRITEBACK
589
590 #include <linux/memcontrol.h>
591
592 /*
593 * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
594 * memcg->cgwb_list. bdi->cgwb_tree is also RCU protected.
595 */
596 static DEFINE_SPINLOCK(cgwb_lock);
597 static struct workqueue_struct *cgwb_release_wq;
598
599 static LIST_HEAD(offline_cgwbs);
600 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
601 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
602
cgwb_free_rcu(struct rcu_head * rcu_head)603 static void cgwb_free_rcu(struct rcu_head *rcu_head)
604 {
605 struct bdi_writeback *wb = container_of(rcu_head,
606 struct bdi_writeback, rcu);
607
608 percpu_ref_exit(&wb->refcnt);
609 kfree(wb);
610 }
611
cgwb_release_workfn(struct work_struct * work)612 static void cgwb_release_workfn(struct work_struct *work)
613 {
614 struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
615 release_work);
616 struct backing_dev_info *bdi = wb->bdi;
617
618 mutex_lock(&wb->bdi->cgwb_release_mutex);
619 wb_shutdown(wb);
620
621 css_put(wb->memcg_css);
622 css_put(wb->blkcg_css);
623 mutex_unlock(&wb->bdi->cgwb_release_mutex);
624
625 /* triggers blkg destruction if no online users left */
626 blkcg_unpin_online(wb->blkcg_css);
627
628 fprop_local_destroy_percpu(&wb->memcg_completions);
629
630 spin_lock_irq(&cgwb_lock);
631 list_del(&wb->offline_node);
632 spin_unlock_irq(&cgwb_lock);
633
634 wb_exit(wb);
635 bdi_put(bdi);
636 WARN_ON_ONCE(!list_empty(&wb->b_attached));
637 call_rcu(&wb->rcu, cgwb_free_rcu);
638 }
639
cgwb_release(struct percpu_ref * refcnt)640 static void cgwb_release(struct percpu_ref *refcnt)
641 {
642 struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
643 refcnt);
644 queue_work(cgwb_release_wq, &wb->release_work);
645 }
646
cgwb_kill(struct bdi_writeback * wb)647 static void cgwb_kill(struct bdi_writeback *wb)
648 {
649 lockdep_assert_held(&cgwb_lock);
650
651 WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
652 list_del(&wb->memcg_node);
653 list_del(&wb->blkcg_node);
654 list_add(&wb->offline_node, &offline_cgwbs);
655 percpu_ref_kill(&wb->refcnt);
656 }
657
cgwb_remove_from_bdi_list(struct bdi_writeback * wb)658 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
659 {
660 spin_lock_irq(&cgwb_lock);
661 list_del_rcu(&wb->bdi_node);
662 spin_unlock_irq(&cgwb_lock);
663 }
664
cgwb_create(struct backing_dev_info * bdi,struct cgroup_subsys_state * memcg_css,gfp_t gfp)665 static int cgwb_create(struct backing_dev_info *bdi,
666 struct cgroup_subsys_state *memcg_css, gfp_t gfp)
667 {
668 struct mem_cgroup *memcg;
669 struct cgroup_subsys_state *blkcg_css;
670 struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
671 struct bdi_writeback *wb;
672 unsigned long flags;
673 int ret = 0;
674
675 memcg = mem_cgroup_from_css(memcg_css);
676 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
677 memcg_cgwb_list = &memcg->cgwb_list;
678 blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
679
680 /* look up again under lock and discard on blkcg mismatch */
681 spin_lock_irqsave(&cgwb_lock, flags);
682 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
683 if (wb && wb->blkcg_css != blkcg_css) {
684 cgwb_kill(wb);
685 wb = NULL;
686 }
687 spin_unlock_irqrestore(&cgwb_lock, flags);
688 if (wb)
689 goto out_put;
690
691 /* need to create a new one */
692 wb = kmalloc(sizeof(*wb), gfp);
693 if (!wb) {
694 ret = -ENOMEM;
695 goto out_put;
696 }
697
698 ret = wb_init(wb, bdi, gfp);
699 if (ret)
700 goto err_free;
701
702 ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
703 if (ret)
704 goto err_wb_exit;
705
706 ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
707 if (ret)
708 goto err_ref_exit;
709
710 wb->memcg_css = memcg_css;
711 wb->blkcg_css = blkcg_css;
712 INIT_LIST_HEAD(&wb->b_attached);
713 INIT_WORK(&wb->release_work, cgwb_release_workfn);
714 set_bit(WB_registered, &wb->state);
715 bdi_get(bdi);
716
717 /*
718 * The root wb determines the registered state of the whole bdi and
719 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
720 * whether they're still online. Don't link @wb if any is dead.
721 * See wb_memcg_offline() and wb_blkcg_offline().
722 */
723 ret = -ENODEV;
724 spin_lock_irqsave(&cgwb_lock, flags);
725 if (test_bit(WB_registered, &bdi->wb.state) &&
726 blkcg_cgwb_list->next && memcg_cgwb_list->next) {
727 /* we might have raced another instance of this function */
728 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
729 if (!ret) {
730 list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
731 list_add(&wb->memcg_node, memcg_cgwb_list);
732 list_add(&wb->blkcg_node, blkcg_cgwb_list);
733 blkcg_pin_online(blkcg_css);
734 css_get(memcg_css);
735 css_get(blkcg_css);
736 }
737 }
738 spin_unlock_irqrestore(&cgwb_lock, flags);
739 if (ret) {
740 if (ret == -EEXIST)
741 ret = 0;
742 goto err_fprop_exit;
743 }
744 goto out_put;
745
746 err_fprop_exit:
747 bdi_put(bdi);
748 fprop_local_destroy_percpu(&wb->memcg_completions);
749 err_ref_exit:
750 percpu_ref_exit(&wb->refcnt);
751 err_wb_exit:
752 wb_exit(wb);
753 err_free:
754 kfree(wb);
755 out_put:
756 css_put(blkcg_css);
757 return ret;
758 }
759
760 /**
761 * wb_get_lookup - get wb for a given memcg
762 * @bdi: target bdi
763 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
764 *
765 * Try to get the wb for @memcg_css on @bdi. The returned wb has its
766 * refcount incremented.
767 *
768 * This function uses css_get() on @memcg_css and thus expects its refcnt
769 * to be positive on invocation. IOW, rcu_read_lock() protection on
770 * @memcg_css isn't enough. try_get it before calling this function.
771 *
772 * A wb is keyed by its associated memcg. As blkcg implicitly enables
773 * memcg on the default hierarchy, memcg association is guaranteed to be
774 * more specific (equal or descendant to the associated blkcg) and thus can
775 * identify both the memcg and blkcg associations.
776 *
777 * Because the blkcg associated with a memcg may change as blkcg is enabled
778 * and disabled closer to root in the hierarchy, each wb keeps track of
779 * both the memcg and blkcg associated with it and verifies the blkcg on
780 * each lookup. On mismatch, the existing wb is discarded and a new one is
781 * created.
782 */
wb_get_lookup(struct backing_dev_info * bdi,struct cgroup_subsys_state * memcg_css)783 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
784 struct cgroup_subsys_state *memcg_css)
785 {
786 struct bdi_writeback *wb;
787
788 if (!memcg_css->parent)
789 return &bdi->wb;
790
791 rcu_read_lock();
792 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
793 if (wb) {
794 struct cgroup_subsys_state *blkcg_css;
795
796 /* see whether the blkcg association has changed */
797 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
798 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
799 wb = NULL;
800 css_put(blkcg_css);
801 }
802 rcu_read_unlock();
803
804 return wb;
805 }
806
807 /**
808 * wb_get_create - get wb for a given memcg, create if necessary
809 * @bdi: target bdi
810 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
811 * @gfp: allocation mask to use
812 *
813 * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to
814 * create one. See wb_get_lookup() for more details.
815 */
wb_get_create(struct backing_dev_info * bdi,struct cgroup_subsys_state * memcg_css,gfp_t gfp)816 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
817 struct cgroup_subsys_state *memcg_css,
818 gfp_t gfp)
819 {
820 struct bdi_writeback *wb;
821
822 might_alloc(gfp);
823
824 do {
825 wb = wb_get_lookup(bdi, memcg_css);
826 } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
827
828 return wb;
829 }
830
cgwb_bdi_init(struct backing_dev_info * bdi)831 static int cgwb_bdi_init(struct backing_dev_info *bdi)
832 {
833 int ret;
834
835 INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
836 mutex_init(&bdi->cgwb_release_mutex);
837 init_rwsem(&bdi->wb_switch_rwsem);
838
839 ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
840 if (!ret) {
841 bdi->wb.memcg_css = &root_mem_cgroup->css;
842 bdi->wb.blkcg_css = blkcg_root_css;
843 }
844 return ret;
845 }
846
cgwb_bdi_unregister(struct backing_dev_info * bdi)847 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
848 {
849 struct radix_tree_iter iter;
850 void **slot;
851 struct bdi_writeback *wb;
852
853 WARN_ON(test_bit(WB_registered, &bdi->wb.state));
854
855 spin_lock_irq(&cgwb_lock);
856 radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
857 cgwb_kill(*slot);
858 spin_unlock_irq(&cgwb_lock);
859
860 mutex_lock(&bdi->cgwb_release_mutex);
861 spin_lock_irq(&cgwb_lock);
862 while (!list_empty(&bdi->wb_list)) {
863 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
864 bdi_node);
865 spin_unlock_irq(&cgwb_lock);
866 wb_shutdown(wb);
867 spin_lock_irq(&cgwb_lock);
868 }
869 spin_unlock_irq(&cgwb_lock);
870 mutex_unlock(&bdi->cgwb_release_mutex);
871 }
872
873 /*
874 * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
875 *
876 * Try to release dying cgwbs by switching attached inodes to the nearest
877 * living ancestor's writeback. Processed wbs are placed at the end
878 * of the list to guarantee the forward progress.
879 */
cleanup_offline_cgwbs_workfn(struct work_struct * work)880 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
881 {
882 struct bdi_writeback *wb;
883 LIST_HEAD(processed);
884
885 spin_lock_irq(&cgwb_lock);
886
887 while (!list_empty(&offline_cgwbs)) {
888 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
889 offline_node);
890 list_move(&wb->offline_node, &processed);
891
892 /*
893 * If wb is dirty, cleaning up the writeback by switching
894 * attached inodes will result in an effective removal of any
895 * bandwidth restrictions, which isn't the goal. Instead,
896 * it can be postponed until the next time, when all io
897 * will be likely completed. If in the meantime some inodes
898 * will get re-dirtied, they should be eventually switched to
899 * a new cgwb.
900 */
901 if (wb_has_dirty_io(wb))
902 continue;
903
904 if (!wb_tryget(wb))
905 continue;
906
907 spin_unlock_irq(&cgwb_lock);
908 while (cleanup_offline_cgwb(wb))
909 cond_resched();
910 spin_lock_irq(&cgwb_lock);
911
912 wb_put(wb);
913 }
914
915 if (!list_empty(&processed))
916 list_splice_tail(&processed, &offline_cgwbs);
917
918 spin_unlock_irq(&cgwb_lock);
919 }
920
921 /**
922 * wb_memcg_offline - kill all wb's associated with a memcg being offlined
923 * @memcg: memcg being offlined
924 *
925 * Also prevents creation of any new wb's associated with @memcg.
926 */
wb_memcg_offline(struct mem_cgroup * memcg)927 void wb_memcg_offline(struct mem_cgroup *memcg)
928 {
929 struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
930 struct bdi_writeback *wb, *next;
931
932 spin_lock_irq(&cgwb_lock);
933 list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
934 cgwb_kill(wb);
935 memcg_cgwb_list->next = NULL; /* prevent new wb's */
936 spin_unlock_irq(&cgwb_lock);
937
938 queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
939 }
940
941 /**
942 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
943 * @css: blkcg being offlined
944 *
945 * Also prevents creation of any new wb's associated with @blkcg.
946 */
wb_blkcg_offline(struct cgroup_subsys_state * css)947 void wb_blkcg_offline(struct cgroup_subsys_state *css)
948 {
949 struct bdi_writeback *wb, *next;
950 struct list_head *list = blkcg_get_cgwb_list(css);
951
952 spin_lock_irq(&cgwb_lock);
953 list_for_each_entry_safe(wb, next, list, blkcg_node)
954 cgwb_kill(wb);
955 list->next = NULL; /* prevent new wb's */
956 spin_unlock_irq(&cgwb_lock);
957 }
958
cgwb_bdi_register(struct backing_dev_info * bdi)959 static void cgwb_bdi_register(struct backing_dev_info *bdi)
960 {
961 spin_lock_irq(&cgwb_lock);
962 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
963 spin_unlock_irq(&cgwb_lock);
964 }
965
cgwb_init(void)966 static int __init cgwb_init(void)
967 {
968 /*
969 * There can be many concurrent release work items overwhelming
970 * system_wq. Put them in a separate wq and limit concurrency.
971 * There's no point in executing many of these in parallel.
972 */
973 cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
974 if (!cgwb_release_wq)
975 return -ENOMEM;
976
977 return 0;
978 }
979 subsys_initcall(cgwb_init);
980
981 #else /* CONFIG_CGROUP_WRITEBACK */
982
cgwb_bdi_init(struct backing_dev_info * bdi)983 static int cgwb_bdi_init(struct backing_dev_info *bdi)
984 {
985 return wb_init(&bdi->wb, bdi, GFP_KERNEL);
986 }
987
cgwb_bdi_unregister(struct backing_dev_info * bdi)988 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
989
cgwb_bdi_register(struct backing_dev_info * bdi)990 static void cgwb_bdi_register(struct backing_dev_info *bdi)
991 {
992 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
993 }
994
cgwb_remove_from_bdi_list(struct bdi_writeback * wb)995 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
996 {
997 list_del_rcu(&wb->bdi_node);
998 }
999
1000 #endif /* CONFIG_CGROUP_WRITEBACK */
1001
bdi_init(struct backing_dev_info * bdi)1002 int bdi_init(struct backing_dev_info *bdi)
1003 {
1004 bdi->dev = NULL;
1005
1006 kref_init(&bdi->refcnt);
1007 bdi->min_ratio = 0;
1008 bdi->max_ratio = 100 * BDI_RATIO_SCALE;
1009 bdi->max_prop_frac = FPROP_FRAC_BASE;
1010 INIT_LIST_HEAD(&bdi->bdi_list);
1011 INIT_LIST_HEAD(&bdi->wb_list);
1012 init_waitqueue_head(&bdi->wb_waitq);
1013 bdi->last_bdp_sleep = jiffies;
1014
1015 return cgwb_bdi_init(bdi);
1016 }
1017
bdi_alloc(int node_id)1018 struct backing_dev_info *bdi_alloc(int node_id)
1019 {
1020 struct backing_dev_info *bdi;
1021
1022 bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
1023 if (!bdi)
1024 return NULL;
1025
1026 if (bdi_init(bdi)) {
1027 kfree(bdi);
1028 return NULL;
1029 }
1030 bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
1031 bdi->ra_pages = VM_READAHEAD_PAGES;
1032 bdi->io_pages = VM_READAHEAD_PAGES;
1033 timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
1034 return bdi;
1035 }
1036 EXPORT_SYMBOL(bdi_alloc);
1037
bdi_lookup_rb_node(u64 id,struct rb_node ** parentp)1038 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
1039 {
1040 struct rb_node **p = &bdi_tree.rb_node;
1041 struct rb_node *parent = NULL;
1042 struct backing_dev_info *bdi;
1043
1044 lockdep_assert_held(&bdi_lock);
1045
1046 while (*p) {
1047 parent = *p;
1048 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
1049
1050 if (bdi->id > id)
1051 p = &(*p)->rb_left;
1052 else if (bdi->id < id)
1053 p = &(*p)->rb_right;
1054 else
1055 break;
1056 }
1057
1058 if (parentp)
1059 *parentp = parent;
1060 return p;
1061 }
1062
1063 /**
1064 * bdi_get_by_id - lookup and get bdi from its id
1065 * @id: bdi id to lookup
1066 *
1067 * Find bdi matching @id and get it. Returns NULL if the matching bdi
1068 * doesn't exist or is already unregistered.
1069 */
bdi_get_by_id(u64 id)1070 struct backing_dev_info *bdi_get_by_id(u64 id)
1071 {
1072 struct backing_dev_info *bdi = NULL;
1073 struct rb_node **p;
1074
1075 spin_lock_bh(&bdi_lock);
1076 p = bdi_lookup_rb_node(id, NULL);
1077 if (*p) {
1078 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
1079 bdi_get(bdi);
1080 }
1081 spin_unlock_bh(&bdi_lock);
1082
1083 return bdi;
1084 }
1085
bdi_register_va(struct backing_dev_info * bdi,const char * fmt,va_list args)1086 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
1087 {
1088 struct device *dev;
1089 struct rb_node *parent, **p;
1090
1091 if (bdi->dev) /* The driver needs to use separate queues per device */
1092 return 0;
1093
1094 vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
1095 dev = device_create(&bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
1096 if (IS_ERR(dev))
1097 return PTR_ERR(dev);
1098
1099 cgwb_bdi_register(bdi);
1100 bdi->dev = dev;
1101
1102 bdi_debug_register(bdi, dev_name(dev));
1103 set_bit(WB_registered, &bdi->wb.state);
1104
1105 spin_lock_bh(&bdi_lock);
1106
1107 bdi->id = ++bdi_id_cursor;
1108
1109 p = bdi_lookup_rb_node(bdi->id, &parent);
1110 rb_link_node(&bdi->rb_node, parent, p);
1111 rb_insert_color(&bdi->rb_node, &bdi_tree);
1112
1113 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
1114
1115 spin_unlock_bh(&bdi_lock);
1116
1117 trace_writeback_bdi_register(bdi);
1118 return 0;
1119 }
1120
bdi_register(struct backing_dev_info * bdi,const char * fmt,...)1121 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1122 {
1123 va_list args;
1124 int ret;
1125
1126 va_start(args, fmt);
1127 ret = bdi_register_va(bdi, fmt, args);
1128 va_end(args);
1129 return ret;
1130 }
1131 EXPORT_SYMBOL(bdi_register);
1132
bdi_set_owner(struct backing_dev_info * bdi,struct device * owner)1133 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1134 {
1135 WARN_ON_ONCE(bdi->owner);
1136 bdi->owner = owner;
1137 get_device(owner);
1138 }
1139
1140 /*
1141 * Remove bdi from bdi_list, and ensure that it is no longer visible
1142 */
bdi_remove_from_list(struct backing_dev_info * bdi)1143 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1144 {
1145 spin_lock_bh(&bdi_lock);
1146 rb_erase(&bdi->rb_node, &bdi_tree);
1147 list_del_rcu(&bdi->bdi_list);
1148 spin_unlock_bh(&bdi_lock);
1149
1150 synchronize_rcu_expedited();
1151 }
1152
bdi_unregister(struct backing_dev_info * bdi)1153 void bdi_unregister(struct backing_dev_info *bdi)
1154 {
1155 del_timer_sync(&bdi->laptop_mode_wb_timer);
1156
1157 /* make sure nobody finds us on the bdi_list anymore */
1158 bdi_remove_from_list(bdi);
1159 wb_shutdown(&bdi->wb);
1160 cgwb_bdi_unregister(bdi);
1161
1162 /*
1163 * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1164 * update the global bdi_min_ratio.
1165 */
1166 if (bdi->min_ratio)
1167 bdi_set_min_ratio(bdi, 0);
1168
1169 if (bdi->dev) {
1170 bdi_debug_unregister(bdi);
1171 device_unregister(bdi->dev);
1172 bdi->dev = NULL;
1173 }
1174
1175 if (bdi->owner) {
1176 put_device(bdi->owner);
1177 bdi->owner = NULL;
1178 }
1179 }
1180 EXPORT_SYMBOL(bdi_unregister);
1181
release_bdi(struct kref * ref)1182 static void release_bdi(struct kref *ref)
1183 {
1184 struct backing_dev_info *bdi =
1185 container_of(ref, struct backing_dev_info, refcnt);
1186
1187 WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1188 WARN_ON_ONCE(bdi->dev);
1189 wb_exit(&bdi->wb);
1190 kfree(bdi);
1191 }
1192
bdi_put(struct backing_dev_info * bdi)1193 void bdi_put(struct backing_dev_info *bdi)
1194 {
1195 kref_put(&bdi->refcnt, release_bdi);
1196 }
1197 EXPORT_SYMBOL(bdi_put);
1198
inode_to_bdi(struct inode * inode)1199 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1200 {
1201 struct super_block *sb;
1202
1203 if (!inode)
1204 return &noop_backing_dev_info;
1205
1206 sb = inode->i_sb;
1207 #ifdef CONFIG_BLOCK
1208 if (sb_is_blkdev_sb(sb))
1209 return I_BDEV(inode)->bd_disk->bdi;
1210 #endif
1211 return sb->s_bdi;
1212 }
1213 EXPORT_SYMBOL(inode_to_bdi);
1214
bdi_dev_name(struct backing_dev_info * bdi)1215 const char *bdi_dev_name(struct backing_dev_info *bdi)
1216 {
1217 if (!bdi || !bdi->dev)
1218 return bdi_unknown_name;
1219 return bdi->dev_name;
1220 }
1221 EXPORT_SYMBOL_GPL(bdi_dev_name);
1222