writeback: let balance_dirty_pages() work on the matching cgroup bdi_writeback
[linux-2.6-block.git] / mm / backing-dev.c
CommitLineData
3fcfab16
AM
1
2#include <linux/wait.h>
3#include <linux/backing-dev.h>
03ba3782
JA
4#include <linux/kthread.h>
5#include <linux/freezer.h>
3fcfab16 6#include <linux/fs.h>
26160158 7#include <linux/pagemap.h>
03ba3782 8#include <linux/mm.h>
3fcfab16
AM
9#include <linux/sched.h>
10#include <linux/module.h>
cf0ca9fe
PZ
11#include <linux/writeback.h>
12#include <linux/device.h>
455b2864 13#include <trace/events/writeback.h>
cf0ca9fe 14
c3c53206
JA
15static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
5129a469
JE
17struct backing_dev_info noop_backing_dev_info = {
18 .name = "noop",
976e48f8 19 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
5129a469 20};
a212b105 21EXPORT_SYMBOL_GPL(noop_backing_dev_info);
5129a469 22
cf0ca9fe 23static struct class *bdi_class;
cfc4ba53
JA
24
25/*
181387da 26 * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
cfc4ba53
JA
27 * locking.
28 */
03ba3782 29DEFINE_SPINLOCK(bdi_lock);
66f3b8e2 30LIST_HEAD(bdi_list);
03ba3782 31
839a8e86
TH
32/* bdi_wq serves all asynchronous writeback tasks */
33struct workqueue_struct *bdi_wq;
34
76f1418b
MS
35#ifdef CONFIG_DEBUG_FS
36#include <linux/debugfs.h>
37#include <linux/seq_file.h>
38
39static struct dentry *bdi_debug_root;
40
41static void bdi_debug_init(void)
42{
43 bdi_debug_root = debugfs_create_dir("bdi", NULL);
44}
45
46static int bdi_debug_stats_show(struct seq_file *m, void *v)
47{
48 struct backing_dev_info *bdi = m->private;
c1955ce3 49 struct bdi_writeback *wb = &bdi->wb;
364aeb28
DR
50 unsigned long background_thresh;
51 unsigned long dirty_thresh;
52 unsigned long bdi_thresh;
0ae45f63 53 unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
f09b00d3
JA
54 struct inode *inode;
55
0ae45f63 56 nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
f758eeab 57 spin_lock(&wb->list_lock);
7ccf19a8 58 list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
c1955ce3 59 nr_dirty++;
7ccf19a8 60 list_for_each_entry(inode, &wb->b_io, i_wb_list)
c1955ce3 61 nr_io++;
7ccf19a8 62 list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
c1955ce3 63 nr_more_io++;
0ae45f63
TT
64 list_for_each_entry(inode, &wb->b_dirty_time, i_wb_list)
65 if (inode->i_state & I_DIRTY_TIME)
66 nr_dirty_time++;
f758eeab 67 spin_unlock(&wb->list_lock);
76f1418b 68
16c4042f 69 global_dirty_limits(&background_thresh, &dirty_thresh);
a88a341a 70 bdi_thresh = wb_dirty_limit(wb, dirty_thresh);
76f1418b
MS
71
72#define K(x) ((x) << (PAGE_SHIFT - 10))
73 seq_printf(m,
00821b00
WF
74 "BdiWriteback: %10lu kB\n"
75 "BdiReclaimable: %10lu kB\n"
76 "BdiDirtyThresh: %10lu kB\n"
77 "DirtyThresh: %10lu kB\n"
78 "BackgroundThresh: %10lu kB\n"
c8e28ce0 79 "BdiDirtied: %10lu kB\n"
00821b00
WF
80 "BdiWritten: %10lu kB\n"
81 "BdiWriteBandwidth: %10lu kBps\n"
82 "b_dirty: %10lu\n"
83 "b_io: %10lu\n"
84 "b_more_io: %10lu\n"
0ae45f63 85 "b_dirty_time: %10lu\n"
00821b00
WF
86 "bdi_list: %10u\n"
87 "state: %10lx\n",
93f78d88
TH
88 (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
89 (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
f7d2b1ec
JK
90 K(bdi_thresh),
91 K(dirty_thresh),
92 K(background_thresh),
93f78d88
TH
93 (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
94 (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
a88a341a 95 (unsigned long) K(wb->write_bandwidth),
f7d2b1ec
JK
96 nr_dirty,
97 nr_io,
98 nr_more_io,
0ae45f63 99 nr_dirty_time,
4452226e 100 !list_empty(&bdi->bdi_list), bdi->wb.state);
76f1418b
MS
101#undef K
102
103 return 0;
104}
105
106static int bdi_debug_stats_open(struct inode *inode, struct file *file)
107{
108 return single_open(file, bdi_debug_stats_show, inode->i_private);
109}
110
111static const struct file_operations bdi_debug_stats_fops = {
112 .open = bdi_debug_stats_open,
113 .read = seq_read,
114 .llseek = seq_lseek,
115 .release = single_release,
116};
117
118static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
119{
120 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
121 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
122 bdi, &bdi_debug_stats_fops);
123}
124
125static void bdi_debug_unregister(struct backing_dev_info *bdi)
126{
127 debugfs_remove(bdi->debug_stats);
128 debugfs_remove(bdi->debug_dir);
129}
130#else
131static inline void bdi_debug_init(void)
132{
133}
134static inline void bdi_debug_register(struct backing_dev_info *bdi,
135 const char *name)
136{
137}
138static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
139{
140}
141#endif
142
cf0ca9fe
PZ
143static ssize_t read_ahead_kb_store(struct device *dev,
144 struct device_attribute *attr,
145 const char *buf, size_t count)
146{
147 struct backing_dev_info *bdi = dev_get_drvdata(dev);
cf0ca9fe 148 unsigned long read_ahead_kb;
7034ed13 149 ssize_t ret;
cf0ca9fe 150
7034ed13
NJ
151 ret = kstrtoul(buf, 10, &read_ahead_kb);
152 if (ret < 0)
153 return ret;
154
155 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
156
157 return count;
cf0ca9fe
PZ
158}
159
160#define K(pages) ((pages) << (PAGE_SHIFT - 10))
161
162#define BDI_SHOW(name, expr) \
163static ssize_t name##_show(struct device *dev, \
164 struct device_attribute *attr, char *page) \
165{ \
166 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
167 \
168 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
d9e1241e
GKH
169} \
170static DEVICE_ATTR_RW(name);
cf0ca9fe
PZ
171
172BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
173
189d3c4a
PZ
174static ssize_t min_ratio_store(struct device *dev,
175 struct device_attribute *attr, const char *buf, size_t count)
176{
177 struct backing_dev_info *bdi = dev_get_drvdata(dev);
189d3c4a 178 unsigned int ratio;
7034ed13
NJ
179 ssize_t ret;
180
181 ret = kstrtouint(buf, 10, &ratio);
182 if (ret < 0)
183 return ret;
184
185 ret = bdi_set_min_ratio(bdi, ratio);
186 if (!ret)
187 ret = count;
189d3c4a 188
189d3c4a
PZ
189 return ret;
190}
191BDI_SHOW(min_ratio, bdi->min_ratio)
192
a42dde04
PZ
193static ssize_t max_ratio_store(struct device *dev,
194 struct device_attribute *attr, const char *buf, size_t count)
195{
196 struct backing_dev_info *bdi = dev_get_drvdata(dev);
a42dde04 197 unsigned int ratio;
7034ed13
NJ
198 ssize_t ret;
199
200 ret = kstrtouint(buf, 10, &ratio);
201 if (ret < 0)
202 return ret;
203
204 ret = bdi_set_max_ratio(bdi, ratio);
205 if (!ret)
206 ret = count;
a42dde04 207
a42dde04
PZ
208 return ret;
209}
210BDI_SHOW(max_ratio, bdi->max_ratio)
211
7d311cda
DW
212static ssize_t stable_pages_required_show(struct device *dev,
213 struct device_attribute *attr,
214 char *page)
215{
216 struct backing_dev_info *bdi = dev_get_drvdata(dev);
217
218 return snprintf(page, PAGE_SIZE-1, "%d\n",
219 bdi_cap_stable_pages_required(bdi) ? 1 : 0);
220}
d9e1241e
GKH
221static DEVICE_ATTR_RO(stable_pages_required);
222
223static struct attribute *bdi_dev_attrs[] = {
224 &dev_attr_read_ahead_kb.attr,
225 &dev_attr_min_ratio.attr,
226 &dev_attr_max_ratio.attr,
227 &dev_attr_stable_pages_required.attr,
228 NULL,
cf0ca9fe 229};
d9e1241e 230ATTRIBUTE_GROUPS(bdi_dev);
cf0ca9fe
PZ
231
232static __init int bdi_class_init(void)
233{
234 bdi_class = class_create(THIS_MODULE, "bdi");
14421453
AB
235 if (IS_ERR(bdi_class))
236 return PTR_ERR(bdi_class);
237
d9e1241e 238 bdi_class->dev_groups = bdi_dev_groups;
76f1418b 239 bdi_debug_init();
cf0ca9fe
PZ
240 return 0;
241}
76f1418b 242postcore_initcall(bdi_class_init);
cf0ca9fe 243
26160158
JA
244static int __init default_bdi_init(void)
245{
246 int err;
247
839a8e86 248 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
b5c872dd 249 WQ_UNBOUND | WQ_SYSFS, 0);
839a8e86
TH
250 if (!bdi_wq)
251 return -ENOMEM;
252
976e48f8 253 err = bdi_init(&noop_backing_dev_info);
26160158
JA
254
255 return err;
256}
257subsys_initcall(default_bdi_init);
258
03ba3782
JA
259int bdi_has_dirty_io(struct backing_dev_info *bdi)
260{
261 return wb_has_dirty_io(&bdi->wb);
262}
263
6467716a 264/*
f0054bb1 265 * This function is used when the first inode for this wb is marked dirty. It
6467716a
AB
266 * wakes-up the corresponding bdi thread which should then take care of the
267 * periodic background write-out of dirty inodes. Since the write-out would
268 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
269 * set up a timer which wakes the bdi thread up later.
270 *
271 * Note, we wouldn't bother setting up the timer, but this function is on the
272 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
273 * by delaying the wake-up.
6ca738d6
DB
274 *
275 * We have to be careful not to postpone flush work if it is scheduled for
276 * earlier. Thus we use queue_delayed_work().
6467716a 277 */
f0054bb1 278void wb_wakeup_delayed(struct bdi_writeback *wb)
6467716a
AB
279{
280 unsigned long timeout;
281
282 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
f0054bb1
TH
283 spin_lock_bh(&wb->work_lock);
284 if (test_bit(WB_registered, &wb->state))
285 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
286 spin_unlock_bh(&wb->work_lock);
03ba3782
JA
287}
288
a88a341a
TH
289/*
290 * Initial write bandwidth: 100 MB/s
291 */
292#define INIT_BW (100 << (20 - PAGE_SHIFT))
293
8395cd9f
TH
294static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
295 gfp_t gfp)
6467716a 296{
93f78d88
TH
297 int i, err;
298
6467716a
AB
299 memset(wb, 0, sizeof(*wb));
300
301 wb->bdi = bdi;
302 wb->last_old_flush = jiffies;
303 INIT_LIST_HEAD(&wb->b_dirty);
304 INIT_LIST_HEAD(&wb->b_io);
305 INIT_LIST_HEAD(&wb->b_more_io);
0ae45f63 306 INIT_LIST_HEAD(&wb->b_dirty_time);
f758eeab 307 spin_lock_init(&wb->list_lock);
93f78d88 308
a88a341a
TH
309 wb->bw_time_stamp = jiffies;
310 wb->balanced_dirty_ratelimit = INIT_BW;
311 wb->dirty_ratelimit = INIT_BW;
312 wb->write_bandwidth = INIT_BW;
313 wb->avg_write_bandwidth = INIT_BW;
314
f0054bb1
TH
315 spin_lock_init(&wb->work_lock);
316 INIT_LIST_HEAD(&wb->work_list);
317 INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
318
8395cd9f 319 err = fprop_local_init_percpu(&wb->completions, gfp);
a88a341a
TH
320 if (err)
321 return err;
322
93f78d88 323 for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
8395cd9f 324 err = percpu_counter_init(&wb->stat[i], 0, gfp);
93f78d88
TH
325 if (err) {
326 while (--i)
327 percpu_counter_destroy(&wb->stat[i]);
a88a341a 328 fprop_local_destroy_percpu(&wb->completions);
93f78d88
TH
329 return err;
330 }
331 }
332
333 return 0;
334}
335
46100071
TH
336/*
337 * Remove bdi from the global list and shutdown any threads we have running
338 */
339static void wb_shutdown(struct bdi_writeback *wb)
340{
341 /* Make sure nobody queues further work */
342 spin_lock_bh(&wb->work_lock);
343 if (!test_and_clear_bit(WB_registered, &wb->state)) {
344 spin_unlock_bh(&wb->work_lock);
345 return;
346 }
347 spin_unlock_bh(&wb->work_lock);
348
349 /*
350 * Drain work list and shutdown the delayed_work. !WB_registered
351 * tells wb_workfn() that @wb is dying and its work_list needs to
352 * be drained no matter what.
353 */
354 mod_delayed_work(bdi_wq, &wb->dwork, 0);
355 flush_delayed_work(&wb->dwork);
356 WARN_ON(!list_empty(&wb->work_list));
357}
358
f0054bb1 359static void wb_exit(struct bdi_writeback *wb)
93f78d88
TH
360{
361 int i;
362
363 WARN_ON(delayed_work_pending(&wb->dwork));
364
365 for (i = 0; i < NR_WB_STAT_ITEMS; i++)
366 percpu_counter_destroy(&wb->stat[i]);
6467716a 367
a88a341a
TH
368 fprop_local_destroy_percpu(&wb->completions);
369}
e98be2d5 370
52ebea74
TH
371#ifdef CONFIG_CGROUP_WRITEBACK
372
373#include <linux/memcontrol.h>
374
375/*
376 * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
377 * blkcg->cgwb_list, and memcg->cgwb_list. bdi->cgwb_tree is also RCU
378 * protected. cgwb_release_wait is used to wait for the completion of cgwb
379 * releases from bdi destruction path.
380 */
381static DEFINE_SPINLOCK(cgwb_lock);
382static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
383
384/**
385 * wb_congested_get_create - get or create a wb_congested
386 * @bdi: associated bdi
387 * @blkcg_id: ID of the associated blkcg
388 * @gfp: allocation mask
389 *
390 * Look up the wb_congested for @blkcg_id on @bdi. If missing, create one.
391 * The returned wb_congested has its reference count incremented. Returns
392 * NULL on failure.
393 */
394struct bdi_writeback_congested *
395wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
396{
397 struct bdi_writeback_congested *new_congested = NULL, *congested;
398 struct rb_node **node, *parent;
399 unsigned long flags;
400
401 if (blkcg_id == 1)
402 return &bdi->wb_congested;
403retry:
404 spin_lock_irqsave(&cgwb_lock, flags);
405
406 node = &bdi->cgwb_congested_tree.rb_node;
407 parent = NULL;
408
409 while (*node != NULL) {
410 parent = *node;
411 congested = container_of(parent, struct bdi_writeback_congested,
412 rb_node);
413 if (congested->blkcg_id < blkcg_id)
414 node = &parent->rb_left;
415 else if (congested->blkcg_id > blkcg_id)
416 node = &parent->rb_right;
417 else
418 goto found;
419 }
420
421 if (new_congested) {
422 /* !found and storage for new one already allocated, insert */
423 congested = new_congested;
424 new_congested = NULL;
425 rb_link_node(&congested->rb_node, parent, node);
426 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
427 atomic_inc(&bdi->usage_cnt);
428 goto found;
429 }
430
431 spin_unlock_irqrestore(&cgwb_lock, flags);
432
433 /* allocate storage for new one and retry */
434 new_congested = kzalloc(sizeof(*new_congested), gfp);
435 if (!new_congested)
436 return NULL;
437
438 atomic_set(&new_congested->refcnt, 0);
439 new_congested->bdi = bdi;
440 new_congested->blkcg_id = blkcg_id;
441 goto retry;
442
443found:
444 atomic_inc(&congested->refcnt);
445 spin_unlock_irqrestore(&cgwb_lock, flags);
446 kfree(new_congested);
447 return congested;
448}
449
450/**
451 * wb_congested_put - put a wb_congested
452 * @congested: wb_congested to put
453 *
454 * Put @congested and destroy it if the refcnt reaches zero.
455 */
456void wb_congested_put(struct bdi_writeback_congested *congested)
457{
458 struct backing_dev_info *bdi = congested->bdi;
459 unsigned long flags;
460
461 if (congested->blkcg_id == 1)
462 return;
463
464 local_irq_save(flags);
465 if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
466 local_irq_restore(flags);
467 return;
468 }
469
470 rb_erase(&congested->rb_node, &congested->bdi->cgwb_congested_tree);
471 spin_unlock_irqrestore(&cgwb_lock, flags);
472 kfree(congested);
473
474 if (atomic_dec_and_test(&bdi->usage_cnt))
475 wake_up_all(&cgwb_release_wait);
476}
477
478static void cgwb_release_workfn(struct work_struct *work)
479{
480 struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
481 release_work);
482 struct backing_dev_info *bdi = wb->bdi;
483
484 wb_shutdown(wb);
485
486 css_put(wb->memcg_css);
487 css_put(wb->blkcg_css);
488 wb_congested_put(wb->congested);
489
490 percpu_ref_exit(&wb->refcnt);
491 wb_exit(wb);
492 kfree_rcu(wb, rcu);
493
494 if (atomic_dec_and_test(&bdi->usage_cnt))
495 wake_up_all(&cgwb_release_wait);
496}
497
498static void cgwb_release(struct percpu_ref *refcnt)
499{
500 struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
501 refcnt);
502 schedule_work(&wb->release_work);
503}
504
505static void cgwb_kill(struct bdi_writeback *wb)
506{
507 lockdep_assert_held(&cgwb_lock);
508
509 WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
510 list_del(&wb->memcg_node);
511 list_del(&wb->blkcg_node);
512 percpu_ref_kill(&wb->refcnt);
513}
514
515static int cgwb_create(struct backing_dev_info *bdi,
516 struct cgroup_subsys_state *memcg_css, gfp_t gfp)
517{
518 struct mem_cgroup *memcg;
519 struct cgroup_subsys_state *blkcg_css;
520 struct blkcg *blkcg;
521 struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
522 struct bdi_writeback *wb;
523 unsigned long flags;
524 int ret = 0;
525
526 memcg = mem_cgroup_from_css(memcg_css);
527 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &blkio_cgrp_subsys);
528 blkcg = css_to_blkcg(blkcg_css);
529 memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
530 blkcg_cgwb_list = &blkcg->cgwb_list;
531
532 /* look up again under lock and discard on blkcg mismatch */
533 spin_lock_irqsave(&cgwb_lock, flags);
534 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
535 if (wb && wb->blkcg_css != blkcg_css) {
536 cgwb_kill(wb);
537 wb = NULL;
538 }
539 spin_unlock_irqrestore(&cgwb_lock, flags);
540 if (wb)
541 goto out_put;
542
543 /* need to create a new one */
544 wb = kmalloc(sizeof(*wb), gfp);
545 if (!wb)
546 return -ENOMEM;
547
548 ret = wb_init(wb, bdi, gfp);
549 if (ret)
550 goto err_free;
551
552 ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
553 if (ret)
554 goto err_wb_exit;
555
556 wb->congested = wb_congested_get_create(bdi, blkcg_css->id, gfp);
557 if (!wb->congested)
558 goto err_ref_exit;
559
560 wb->memcg_css = memcg_css;
561 wb->blkcg_css = blkcg_css;
562 INIT_WORK(&wb->release_work, cgwb_release_workfn);
563 set_bit(WB_registered, &wb->state);
564
565 /*
566 * The root wb determines the registered state of the whole bdi and
567 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
568 * whether they're still online. Don't link @wb if any is dead.
569 * See wb_memcg_offline() and wb_blkcg_offline().
570 */
571 ret = -ENODEV;
572 spin_lock_irqsave(&cgwb_lock, flags);
573 if (test_bit(WB_registered, &bdi->wb.state) &&
574 blkcg_cgwb_list->next && memcg_cgwb_list->next) {
575 /* we might have raced another instance of this function */
576 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
577 if (!ret) {
578 atomic_inc(&bdi->usage_cnt);
579 list_add(&wb->memcg_node, memcg_cgwb_list);
580 list_add(&wb->blkcg_node, blkcg_cgwb_list);
581 css_get(memcg_css);
582 css_get(blkcg_css);
583 }
584 }
585 spin_unlock_irqrestore(&cgwb_lock, flags);
586 if (ret) {
587 if (ret == -EEXIST)
588 ret = 0;
589 goto err_put_congested;
590 }
591 goto out_put;
592
593err_put_congested:
594 wb_congested_put(wb->congested);
595err_ref_exit:
596 percpu_ref_exit(&wb->refcnt);
597err_wb_exit:
598 wb_exit(wb);
599err_free:
600 kfree(wb);
601out_put:
602 css_put(blkcg_css);
603 return ret;
604}
605
606/**
607 * wb_get_create - get wb for a given memcg, create if necessary
608 * @bdi: target bdi
609 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
610 * @gfp: allocation mask to use
611 *
612 * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to
613 * create one. The returned wb has its refcount incremented.
614 *
615 * This function uses css_get() on @memcg_css and thus expects its refcnt
616 * to be positive on invocation. IOW, rcu_read_lock() protection on
617 * @memcg_css isn't enough. try_get it before calling this function.
618 *
619 * A wb is keyed by its associated memcg. As blkcg implicitly enables
620 * memcg on the default hierarchy, memcg association is guaranteed to be
621 * more specific (equal or descendant to the associated blkcg) and thus can
622 * identify both the memcg and blkcg associations.
623 *
624 * Because the blkcg associated with a memcg may change as blkcg is enabled
625 * and disabled closer to root in the hierarchy, each wb keeps track of
626 * both the memcg and blkcg associated with it and verifies the blkcg on
627 * each lookup. On mismatch, the existing wb is discarded and a new one is
628 * created.
629 */
630struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
631 struct cgroup_subsys_state *memcg_css,
632 gfp_t gfp)
633{
634 struct bdi_writeback *wb;
635
636 might_sleep_if(gfp & __GFP_WAIT);
637
638 if (!memcg_css->parent)
639 return &bdi->wb;
640
641 do {
642 rcu_read_lock();
643 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
644 if (wb) {
645 struct cgroup_subsys_state *blkcg_css;
646
647 /* see whether the blkcg association has changed */
648 blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
649 &blkio_cgrp_subsys);
650 if (unlikely(wb->blkcg_css != blkcg_css ||
651 !wb_tryget(wb)))
652 wb = NULL;
653 css_put(blkcg_css);
654 }
655 rcu_read_unlock();
656 } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
657
658 return wb;
659}
660
661void __inode_attach_wb(struct inode *inode, struct page *page)
662{
663 struct backing_dev_info *bdi = inode_to_bdi(inode);
664 struct bdi_writeback *wb = NULL;
665
666 if (inode_cgwb_enabled(inode)) {
667 struct cgroup_subsys_state *memcg_css;
668
669 if (page) {
670 memcg_css = mem_cgroup_css_from_page(page);
671 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
672 } else {
673 /* must pin memcg_css, see wb_get_create() */
674 memcg_css = task_get_css(current, memory_cgrp_id);
675 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
676 css_put(memcg_css);
677 }
678 }
679
680 if (!wb)
681 wb = &bdi->wb;
682
683 /*
684 * There may be multiple instances of this function racing to
685 * update the same inode. Use cmpxchg() to tell the winner.
686 */
687 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
688 wb_put(wb);
689}
690
691static void cgwb_bdi_init(struct backing_dev_info *bdi)
692{
693 bdi->wb.memcg_css = mem_cgroup_root_css;
694 bdi->wb.blkcg_css = blkcg_root_css;
695 bdi->wb_congested.blkcg_id = 1;
696 INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
697 bdi->cgwb_congested_tree = RB_ROOT;
698 atomic_set(&bdi->usage_cnt, 1);
699}
700
701static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
702{
703 struct radix_tree_iter iter;
704 void **slot;
705
706 WARN_ON(test_bit(WB_registered, &bdi->wb.state));
707
708 spin_lock_irq(&cgwb_lock);
709 radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
710 cgwb_kill(*slot);
711 spin_unlock_irq(&cgwb_lock);
712
713 /*
714 * All cgwb's and their congested states must be shutdown and
715 * released before returning. Drain the usage counter to wait for
716 * all cgwb's and cgwb_congested's ever created on @bdi.
717 */
718 atomic_dec(&bdi->usage_cnt);
719 wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
720}
721
722/**
723 * wb_memcg_offline - kill all wb's associated with a memcg being offlined
724 * @memcg: memcg being offlined
725 *
726 * Also prevents creation of any new wb's associated with @memcg.
727 */
728void wb_memcg_offline(struct mem_cgroup *memcg)
729{
730 LIST_HEAD(to_destroy);
731 struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
732 struct bdi_writeback *wb, *next;
733
734 spin_lock_irq(&cgwb_lock);
735 list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
736 cgwb_kill(wb);
737 memcg_cgwb_list->next = NULL; /* prevent new wb's */
738 spin_unlock_irq(&cgwb_lock);
739}
740
741/**
742 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
743 * @blkcg: blkcg being offlined
744 *
745 * Also prevents creation of any new wb's associated with @blkcg.
746 */
747void wb_blkcg_offline(struct blkcg *blkcg)
748{
749 LIST_HEAD(to_destroy);
750 struct bdi_writeback *wb, *next;
751
752 spin_lock_irq(&cgwb_lock);
753 list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
754 cgwb_kill(wb);
755 blkcg->cgwb_list.next = NULL; /* prevent new wb's */
756 spin_unlock_irq(&cgwb_lock);
757}
758
759#else /* CONFIG_CGROUP_WRITEBACK */
760
761static void cgwb_bdi_init(struct backing_dev_info *bdi) { }
762static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
763
764#endif /* CONFIG_CGROUP_WRITEBACK */
765
b2e8fb6e
PZ
766int bdi_init(struct backing_dev_info *bdi)
767{
93f78d88 768 int err;
b2e8fb6e 769
cf0ca9fe
PZ
770 bdi->dev = NULL;
771
189d3c4a 772 bdi->min_ratio = 0;
a42dde04 773 bdi->max_ratio = 100;
eb608e3a 774 bdi->max_prop_frac = FPROP_FRAC_BASE;
66f3b8e2 775 INIT_LIST_HEAD(&bdi->bdi_list);
03ba3782 776
8395cd9f 777 err = wb_init(&bdi->wb, bdi, GFP_KERNEL);
93f78d88
TH
778 if (err)
779 return err;
04fbfdc1 780
4aa9c692
TH
781 bdi->wb_congested.state = 0;
782 bdi->wb.congested = &bdi->wb_congested;
783
52ebea74 784 cgwb_bdi_init(bdi);
93f78d88 785 return 0;
b2e8fb6e
PZ
786}
787EXPORT_SYMBOL(bdi_init);
788
46100071
TH
789int bdi_register(struct backing_dev_info *bdi, struct device *parent,
790 const char *fmt, ...)
791{
792 va_list args;
793 struct device *dev;
794
795 if (bdi->dev) /* The driver needs to use separate queues per device */
796 return 0;
797
798 va_start(args, fmt);
799 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
800 va_end(args);
801 if (IS_ERR(dev))
802 return PTR_ERR(dev);
803
804 bdi->dev = dev;
805
806 bdi_debug_register(bdi, dev_name(dev));
807 set_bit(WB_registered, &bdi->wb.state);
808
809 spin_lock_bh(&bdi_lock);
810 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
811 spin_unlock_bh(&bdi_lock);
812
813 trace_writeback_bdi_register(bdi);
814 return 0;
815}
816EXPORT_SYMBOL(bdi_register);
817
818int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
819{
820 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
821}
822EXPORT_SYMBOL(bdi_register_dev);
823
824/*
825 * Remove bdi from bdi_list, and ensure that it is no longer visible
826 */
827static void bdi_remove_from_list(struct backing_dev_info *bdi)
828{
829 spin_lock_bh(&bdi_lock);
830 list_del_rcu(&bdi->bdi_list);
831 spin_unlock_bh(&bdi_lock);
832
833 synchronize_rcu_expedited();
834}
835
836/*
837 * Called when the device behind @bdi has been removed or ejected.
838 *
839 * We can't really do much here except for reducing the dirty ratio at
840 * the moment. In the future we should be able to set a flag so that
841 * the filesystem can handle errors at mark_inode_dirty time instead
842 * of only at writeback time.
843 */
844void bdi_unregister(struct backing_dev_info *bdi)
845{
846 if (WARN_ON_ONCE(!bdi->dev))
847 return;
848
849 bdi_set_min_ratio(bdi, 0);
850}
851EXPORT_SYMBOL(bdi_unregister);
852
b2e8fb6e
PZ
853void bdi_destroy(struct backing_dev_info *bdi)
854{
f0054bb1
TH
855 /* make sure nobody finds us on the bdi_list anymore */
856 bdi_remove_from_list(bdi);
857 wb_shutdown(&bdi->wb);
52ebea74 858 cgwb_bdi_destroy(bdi);
7a401a97 859
c4db59d3
CH
860 if (bdi->dev) {
861 bdi_debug_unregister(bdi);
862 device_unregister(bdi->dev);
863 bdi->dev = NULL;
864 }
865
f0054bb1 866 wb_exit(&bdi->wb);
b2e8fb6e
PZ
867}
868EXPORT_SYMBOL(bdi_destroy);
869
c3c53206
JA
870/*
871 * For use from filesystems to quickly init and register a bdi associated
872 * with dirty writeback
873 */
b4caecd4 874int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
c3c53206 875{
c3c53206
JA
876 int err;
877
878 bdi->name = name;
b4caecd4 879 bdi->capabilities = 0;
c3c53206
JA
880 err = bdi_init(bdi);
881 if (err)
882 return err;
883
02aa2a37
KC
884 err = bdi_register(bdi, NULL, "%.28s-%ld", name,
885 atomic_long_inc_return(&bdi_seq));
c3c53206
JA
886 if (err) {
887 bdi_destroy(bdi);
888 return err;
889 }
890
891 return 0;
892}
893EXPORT_SYMBOL(bdi_setup_and_register);
894
3fcfab16
AM
895static wait_queue_head_t congestion_wqh[2] = {
896 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
897 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
898 };
0e093d99 899static atomic_t nr_bdi_congested[2];
3fcfab16 900
1faa16d2 901void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
3fcfab16 902{
4452226e 903 enum wb_state bit;
1faa16d2 904 wait_queue_head_t *wqh = &congestion_wqh[sync];
3fcfab16 905
4452226e 906 bit = sync ? WB_sync_congested : WB_async_congested;
4aa9c692 907 if (test_and_clear_bit(bit, &bdi->wb.congested->state))
0e093d99 908 atomic_dec(&nr_bdi_congested[sync]);
4e857c58 909 smp_mb__after_atomic();
3fcfab16
AM
910 if (waitqueue_active(wqh))
911 wake_up(wqh);
912}
913EXPORT_SYMBOL(clear_bdi_congested);
914
1faa16d2 915void set_bdi_congested(struct backing_dev_info *bdi, int sync)
3fcfab16 916{
4452226e 917 enum wb_state bit;
3fcfab16 918
4452226e 919 bit = sync ? WB_sync_congested : WB_async_congested;
4aa9c692 920 if (!test_and_set_bit(bit, &bdi->wb.congested->state))
0e093d99 921 atomic_inc(&nr_bdi_congested[sync]);
3fcfab16
AM
922}
923EXPORT_SYMBOL(set_bdi_congested);
924
925/**
926 * congestion_wait - wait for a backing_dev to become uncongested
8aa7e847 927 * @sync: SYNC or ASYNC IO
3fcfab16
AM
928 * @timeout: timeout in jiffies
929 *
930 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
931 * write congestion. If no backing_devs are congested then just wait for the
932 * next write to be completed.
933 */
8aa7e847 934long congestion_wait(int sync, long timeout)
3fcfab16
AM
935{
936 long ret;
52bb9198 937 unsigned long start = jiffies;
3fcfab16 938 DEFINE_WAIT(wait);
8aa7e847 939 wait_queue_head_t *wqh = &congestion_wqh[sync];
3fcfab16
AM
940
941 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
942 ret = io_schedule_timeout(timeout);
943 finish_wait(wqh, &wait);
52bb9198
MG
944
945 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
946 jiffies_to_usecs(jiffies - start));
947
3fcfab16
AM
948 return ret;
949}
950EXPORT_SYMBOL(congestion_wait);
04fbfdc1 951
0e093d99
MG
952/**
953 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
954 * @zone: A zone to check if it is heavily congested
955 * @sync: SYNC or ASYNC IO
956 * @timeout: timeout in jiffies
957 *
958 * In the event of a congested backing_dev (any backing_dev) and the given
959 * @zone has experienced recent congestion, this waits for up to @timeout
960 * jiffies for either a BDI to exit congestion of the given @sync queue
961 * or a write to complete.
962 *
25985edc 963 * In the absence of zone congestion, cond_resched() is called to yield
0e093d99
MG
964 * the processor if necessary but otherwise does not sleep.
965 *
966 * The return value is 0 if the sleep is for the full timeout. Otherwise,
967 * it is the number of jiffies that were still remaining when the function
968 * returned. return_value == timeout implies the function did not sleep.
969 */
970long wait_iff_congested(struct zone *zone, int sync, long timeout)
971{
972 long ret;
973 unsigned long start = jiffies;
974 DEFINE_WAIT(wait);
975 wait_queue_head_t *wqh = &congestion_wqh[sync];
976
977 /*
978 * If there is no congestion, or heavy congestion is not being
979 * encountered in the current zone, yield if necessary instead
980 * of sleeping on the congestion queue
981 */
982 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
57054651 983 !test_bit(ZONE_CONGESTED, &zone->flags)) {
0e093d99
MG
984 cond_resched();
985
986 /* In case we scheduled, work out time remaining */
987 ret = timeout - (jiffies - start);
988 if (ret < 0)
989 ret = 0;
990
991 goto out;
992 }
993
994 /* Sleep until uncongested or a write happens */
995 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
996 ret = io_schedule_timeout(timeout);
997 finish_wait(wqh, &wait);
998
999out:
1000 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1001 jiffies_to_usecs(jiffies - start));
1002
1003 return ret;
1004}
1005EXPORT_SYMBOL(wait_iff_congested);
3965c9ae
WL
1006
1007int pdflush_proc_obsolete(struct ctl_table *table, int write,
1008 void __user *buffer, size_t *lenp, loff_t *ppos)
1009{
1010 char kbuf[] = "0\n";
1011
4c3bffc2 1012 if (*ppos || *lenp < sizeof(kbuf)) {
3965c9ae
WL
1013 *lenp = 0;
1014 return 0;
1015 }
1016
1017 if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1018 return -EFAULT;
1019 printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
1020 table->procname);
1021
1022 *lenp = 2;
1023 *ppos += *lenp;
1024 return 2;
1025}