memcg: make mem_cgroup_read_{stat|event}() iterate possible cpus instead of online
[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
6467716a 259/*
f0054bb1 260 * This function is used when the first inode for this wb is marked dirty. It
6467716a
AB
261 * wakes-up the corresponding bdi thread which should then take care of the
262 * periodic background write-out of dirty inodes. Since the write-out would
263 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
264 * set up a timer which wakes the bdi thread up later.
265 *
266 * Note, we wouldn't bother setting up the timer, but this function is on the
267 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
268 * by delaying the wake-up.
6ca738d6
DB
269 *
270 * We have to be careful not to postpone flush work if it is scheduled for
271 * earlier. Thus we use queue_delayed_work().
6467716a 272 */
f0054bb1 273void wb_wakeup_delayed(struct bdi_writeback *wb)
6467716a
AB
274{
275 unsigned long timeout;
276
277 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
f0054bb1
TH
278 spin_lock_bh(&wb->work_lock);
279 if (test_bit(WB_registered, &wb->state))
280 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
281 spin_unlock_bh(&wb->work_lock);
03ba3782
JA
282}
283
a88a341a
TH
284/*
285 * Initial write bandwidth: 100 MB/s
286 */
287#define INIT_BW (100 << (20 - PAGE_SHIFT))
288
8395cd9f
TH
289static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
290 gfp_t gfp)
6467716a 291{
93f78d88
TH
292 int i, err;
293
6467716a
AB
294 memset(wb, 0, sizeof(*wb));
295
296 wb->bdi = bdi;
297 wb->last_old_flush = jiffies;
298 INIT_LIST_HEAD(&wb->b_dirty);
299 INIT_LIST_HEAD(&wb->b_io);
300 INIT_LIST_HEAD(&wb->b_more_io);
0ae45f63 301 INIT_LIST_HEAD(&wb->b_dirty_time);
f758eeab 302 spin_lock_init(&wb->list_lock);
93f78d88 303
a88a341a
TH
304 wb->bw_time_stamp = jiffies;
305 wb->balanced_dirty_ratelimit = INIT_BW;
306 wb->dirty_ratelimit = INIT_BW;
307 wb->write_bandwidth = INIT_BW;
308 wb->avg_write_bandwidth = INIT_BW;
309
f0054bb1
TH
310 spin_lock_init(&wb->work_lock);
311 INIT_LIST_HEAD(&wb->work_list);
312 INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
313
8395cd9f 314 err = fprop_local_init_percpu(&wb->completions, gfp);
a88a341a
TH
315 if (err)
316 return err;
317
93f78d88 318 for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
8395cd9f 319 err = percpu_counter_init(&wb->stat[i], 0, gfp);
93f78d88
TH
320 if (err) {
321 while (--i)
322 percpu_counter_destroy(&wb->stat[i]);
a88a341a 323 fprop_local_destroy_percpu(&wb->completions);
93f78d88
TH
324 return err;
325 }
326 }
327
328 return 0;
329}
330
46100071
TH
331/*
332 * Remove bdi from the global list and shutdown any threads we have running
333 */
334static void wb_shutdown(struct bdi_writeback *wb)
335{
336 /* Make sure nobody queues further work */
337 spin_lock_bh(&wb->work_lock);
338 if (!test_and_clear_bit(WB_registered, &wb->state)) {
339 spin_unlock_bh(&wb->work_lock);
340 return;
341 }
342 spin_unlock_bh(&wb->work_lock);
343
344 /*
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
f0054bb1 354static void wb_exit(struct bdi_writeback *wb)
93f78d88
TH
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]);
6467716a 362
a88a341a
TH
363 fprop_local_destroy_percpu(&wb->completions);
364}
e98be2d5 365
52ebea74
TH
366#ifdef CONFIG_CGROUP_WRITEBACK
367
368#include <linux/memcontrol.h>
369
370/*
371 * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
372 * blkcg->cgwb_list, and memcg->cgwb_list. bdi->cgwb_tree is also RCU
373 * protected. cgwb_release_wait is used to wait for the completion of cgwb
374 * releases from bdi destruction path.
375 */
376static DEFINE_SPINLOCK(cgwb_lock);
377static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
378
379/**
380 * wb_congested_get_create - get or create a wb_congested
381 * @bdi: associated bdi
382 * @blkcg_id: ID of the associated blkcg
383 * @gfp: allocation mask
384 *
385 * Look up the wb_congested for @blkcg_id on @bdi. If missing, create one.
386 * The returned wb_congested has its reference count incremented. Returns
387 * NULL on failure.
388 */
389struct bdi_writeback_congested *
390wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
391{
392 struct bdi_writeback_congested *new_congested = NULL, *congested;
393 struct rb_node **node, *parent;
394 unsigned long flags;
395
396 if (blkcg_id == 1)
397 return &bdi->wb_congested;
398retry:
399 spin_lock_irqsave(&cgwb_lock, flags);
400
401 node = &bdi->cgwb_congested_tree.rb_node;
402 parent = NULL;
403
404 while (*node != NULL) {
405 parent = *node;
406 congested = container_of(parent, struct bdi_writeback_congested,
407 rb_node);
408 if (congested->blkcg_id < blkcg_id)
409 node = &parent->rb_left;
410 else if (congested->blkcg_id > blkcg_id)
411 node = &parent->rb_right;
412 else
413 goto found;
414 }
415
416 if (new_congested) {
417 /* !found and storage for new one already allocated, insert */
418 congested = new_congested;
419 new_congested = NULL;
420 rb_link_node(&congested->rb_node, parent, node);
421 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
422 atomic_inc(&bdi->usage_cnt);
423 goto found;
424 }
425
426 spin_unlock_irqrestore(&cgwb_lock, flags);
427
428 /* allocate storage for new one and retry */
429 new_congested = kzalloc(sizeof(*new_congested), gfp);
430 if (!new_congested)
431 return NULL;
432
433 atomic_set(&new_congested->refcnt, 0);
434 new_congested->bdi = bdi;
435 new_congested->blkcg_id = blkcg_id;
436 goto retry;
437
438found:
439 atomic_inc(&congested->refcnt);
440 spin_unlock_irqrestore(&cgwb_lock, flags);
441 kfree(new_congested);
442 return congested;
443}
444
445/**
446 * wb_congested_put - put a wb_congested
447 * @congested: wb_congested to put
448 *
449 * Put @congested and destroy it if the refcnt reaches zero.
450 */
451void wb_congested_put(struct bdi_writeback_congested *congested)
452{
453 struct backing_dev_info *bdi = congested->bdi;
454 unsigned long flags;
455
456 if (congested->blkcg_id == 1)
457 return;
458
459 local_irq_save(flags);
460 if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
461 local_irq_restore(flags);
462 return;
463 }
464
465 rb_erase(&congested->rb_node, &congested->bdi->cgwb_congested_tree);
466 spin_unlock_irqrestore(&cgwb_lock, flags);
467 kfree(congested);
468
469 if (atomic_dec_and_test(&bdi->usage_cnt))
470 wake_up_all(&cgwb_release_wait);
471}
472
473static void cgwb_release_workfn(struct work_struct *work)
474{
475 struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
476 release_work);
477 struct backing_dev_info *bdi = wb->bdi;
478
479 wb_shutdown(wb);
480
481 css_put(wb->memcg_css);
482 css_put(wb->blkcg_css);
483 wb_congested_put(wb->congested);
484
485 percpu_ref_exit(&wb->refcnt);
486 wb_exit(wb);
487 kfree_rcu(wb, rcu);
488
489 if (atomic_dec_and_test(&bdi->usage_cnt))
490 wake_up_all(&cgwb_release_wait);
491}
492
493static void cgwb_release(struct percpu_ref *refcnt)
494{
495 struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
496 refcnt);
497 schedule_work(&wb->release_work);
498}
499
500static void cgwb_kill(struct bdi_writeback *wb)
501{
502 lockdep_assert_held(&cgwb_lock);
503
504 WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
505 list_del(&wb->memcg_node);
506 list_del(&wb->blkcg_node);
507 percpu_ref_kill(&wb->refcnt);
508}
509
510static int cgwb_create(struct backing_dev_info *bdi,
511 struct cgroup_subsys_state *memcg_css, gfp_t gfp)
512{
513 struct mem_cgroup *memcg;
514 struct cgroup_subsys_state *blkcg_css;
515 struct blkcg *blkcg;
516 struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
517 struct bdi_writeback *wb;
518 unsigned long flags;
519 int ret = 0;
520
521 memcg = mem_cgroup_from_css(memcg_css);
522 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &blkio_cgrp_subsys);
523 blkcg = css_to_blkcg(blkcg_css);
524 memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
525 blkcg_cgwb_list = &blkcg->cgwb_list;
526
527 /* look up again under lock and discard on blkcg mismatch */
528 spin_lock_irqsave(&cgwb_lock, flags);
529 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
530 if (wb && wb->blkcg_css != blkcg_css) {
531 cgwb_kill(wb);
532 wb = NULL;
533 }
534 spin_unlock_irqrestore(&cgwb_lock, flags);
535 if (wb)
536 goto out_put;
537
538 /* need to create a new one */
539 wb = kmalloc(sizeof(*wb), gfp);
540 if (!wb)
541 return -ENOMEM;
542
543 ret = wb_init(wb, bdi, gfp);
544 if (ret)
545 goto err_free;
546
547 ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
548 if (ret)
549 goto err_wb_exit;
550
551 wb->congested = wb_congested_get_create(bdi, blkcg_css->id, gfp);
552 if (!wb->congested)
553 goto err_ref_exit;
554
555 wb->memcg_css = memcg_css;
556 wb->blkcg_css = blkcg_css;
557 INIT_WORK(&wb->release_work, cgwb_release_workfn);
558 set_bit(WB_registered, &wb->state);
559
560 /*
561 * The root wb determines the registered state of the whole bdi and
562 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
563 * whether they're still online. Don't link @wb if any is dead.
564 * See wb_memcg_offline() and wb_blkcg_offline().
565 */
566 ret = -ENODEV;
567 spin_lock_irqsave(&cgwb_lock, flags);
568 if (test_bit(WB_registered, &bdi->wb.state) &&
569 blkcg_cgwb_list->next && memcg_cgwb_list->next) {
570 /* we might have raced another instance of this function */
571 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
572 if (!ret) {
573 atomic_inc(&bdi->usage_cnt);
574 list_add(&wb->memcg_node, memcg_cgwb_list);
575 list_add(&wb->blkcg_node, blkcg_cgwb_list);
576 css_get(memcg_css);
577 css_get(blkcg_css);
578 }
579 }
580 spin_unlock_irqrestore(&cgwb_lock, flags);
581 if (ret) {
582 if (ret == -EEXIST)
583 ret = 0;
584 goto err_put_congested;
585 }
586 goto out_put;
587
588err_put_congested:
589 wb_congested_put(wb->congested);
590err_ref_exit:
591 percpu_ref_exit(&wb->refcnt);
592err_wb_exit:
593 wb_exit(wb);
594err_free:
595 kfree(wb);
596out_put:
597 css_put(blkcg_css);
598 return ret;
599}
600
601/**
602 * wb_get_create - get wb for a given memcg, create if necessary
603 * @bdi: target bdi
604 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
605 * @gfp: allocation mask to use
606 *
607 * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to
608 * create one. The returned wb has its refcount incremented.
609 *
610 * This function uses css_get() on @memcg_css and thus expects its refcnt
611 * to be positive on invocation. IOW, rcu_read_lock() protection on
612 * @memcg_css isn't enough. try_get it before calling this function.
613 *
614 * A wb is keyed by its associated memcg. As blkcg implicitly enables
615 * memcg on the default hierarchy, memcg association is guaranteed to be
616 * more specific (equal or descendant to the associated blkcg) and thus can
617 * identify both the memcg and blkcg associations.
618 *
619 * Because the blkcg associated with a memcg may change as blkcg is enabled
620 * and disabled closer to root in the hierarchy, each wb keeps track of
621 * both the memcg and blkcg associated with it and verifies the blkcg on
622 * each lookup. On mismatch, the existing wb is discarded and a new one is
623 * created.
624 */
625struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
626 struct cgroup_subsys_state *memcg_css,
627 gfp_t gfp)
628{
629 struct bdi_writeback *wb;
630
631 might_sleep_if(gfp & __GFP_WAIT);
632
633 if (!memcg_css->parent)
634 return &bdi->wb;
635
636 do {
637 rcu_read_lock();
638 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
639 if (wb) {
640 struct cgroup_subsys_state *blkcg_css;
641
642 /* see whether the blkcg association has changed */
643 blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
644 &blkio_cgrp_subsys);
645 if (unlikely(wb->blkcg_css != blkcg_css ||
646 !wb_tryget(wb)))
647 wb = NULL;
648 css_put(blkcg_css);
649 }
650 rcu_read_unlock();
651 } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
652
653 return wb;
654}
655
656void __inode_attach_wb(struct inode *inode, struct page *page)
657{
658 struct backing_dev_info *bdi = inode_to_bdi(inode);
659 struct bdi_writeback *wb = NULL;
660
661 if (inode_cgwb_enabled(inode)) {
662 struct cgroup_subsys_state *memcg_css;
663
664 if (page) {
665 memcg_css = mem_cgroup_css_from_page(page);
666 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
667 } else {
668 /* must pin memcg_css, see wb_get_create() */
669 memcg_css = task_get_css(current, memory_cgrp_id);
670 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
671 css_put(memcg_css);
672 }
673 }
674
675 if (!wb)
676 wb = &bdi->wb;
677
678 /*
679 * There may be multiple instances of this function racing to
680 * update the same inode. Use cmpxchg() to tell the winner.
681 */
682 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
683 wb_put(wb);
684}
685
686static void cgwb_bdi_init(struct backing_dev_info *bdi)
687{
688 bdi->wb.memcg_css = mem_cgroup_root_css;
689 bdi->wb.blkcg_css = blkcg_root_css;
690 bdi->wb_congested.blkcg_id = 1;
691 INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
692 bdi->cgwb_congested_tree = RB_ROOT;
693 atomic_set(&bdi->usage_cnt, 1);
694}
695
696static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
697{
698 struct radix_tree_iter iter;
699 void **slot;
700
701 WARN_ON(test_bit(WB_registered, &bdi->wb.state));
702
703 spin_lock_irq(&cgwb_lock);
704 radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
705 cgwb_kill(*slot);
706 spin_unlock_irq(&cgwb_lock);
707
708 /*
709 * All cgwb's and their congested states must be shutdown and
710 * released before returning. Drain the usage counter to wait for
711 * all cgwb's and cgwb_congested's ever created on @bdi.
712 */
713 atomic_dec(&bdi->usage_cnt);
714 wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
715}
716
717/**
718 * wb_memcg_offline - kill all wb's associated with a memcg being offlined
719 * @memcg: memcg being offlined
720 *
721 * Also prevents creation of any new wb's associated with @memcg.
722 */
723void wb_memcg_offline(struct mem_cgroup *memcg)
724{
725 LIST_HEAD(to_destroy);
726 struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
727 struct bdi_writeback *wb, *next;
728
729 spin_lock_irq(&cgwb_lock);
730 list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
731 cgwb_kill(wb);
732 memcg_cgwb_list->next = NULL; /* prevent new wb's */
733 spin_unlock_irq(&cgwb_lock);
734}
735
736/**
737 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
738 * @blkcg: blkcg being offlined
739 *
740 * Also prevents creation of any new wb's associated with @blkcg.
741 */
742void wb_blkcg_offline(struct blkcg *blkcg)
743{
744 LIST_HEAD(to_destroy);
745 struct bdi_writeback *wb, *next;
746
747 spin_lock_irq(&cgwb_lock);
748 list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
749 cgwb_kill(wb);
750 blkcg->cgwb_list.next = NULL; /* prevent new wb's */
751 spin_unlock_irq(&cgwb_lock);
752}
753
754#else /* CONFIG_CGROUP_WRITEBACK */
755
756static void cgwb_bdi_init(struct backing_dev_info *bdi) { }
757static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
758
759#endif /* CONFIG_CGROUP_WRITEBACK */
760
b2e8fb6e
PZ
761int bdi_init(struct backing_dev_info *bdi)
762{
93f78d88 763 int err;
b2e8fb6e 764
cf0ca9fe
PZ
765 bdi->dev = NULL;
766
189d3c4a 767 bdi->min_ratio = 0;
a42dde04 768 bdi->max_ratio = 100;
eb608e3a 769 bdi->max_prop_frac = FPROP_FRAC_BASE;
66f3b8e2 770 INIT_LIST_HEAD(&bdi->bdi_list);
cc395d7f 771 init_waitqueue_head(&bdi->wb_waitq);
03ba3782 772
8395cd9f 773 err = wb_init(&bdi->wb, bdi, GFP_KERNEL);
93f78d88
TH
774 if (err)
775 return err;
04fbfdc1 776
4aa9c692
TH
777 bdi->wb_congested.state = 0;
778 bdi->wb.congested = &bdi->wb_congested;
779
52ebea74 780 cgwb_bdi_init(bdi);
93f78d88 781 return 0;
b2e8fb6e
PZ
782}
783EXPORT_SYMBOL(bdi_init);
784
46100071
TH
785int bdi_register(struct backing_dev_info *bdi, struct device *parent,
786 const char *fmt, ...)
787{
788 va_list args;
789 struct device *dev;
790
791 if (bdi->dev) /* The driver needs to use separate queues per device */
792 return 0;
793
794 va_start(args, fmt);
795 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
796 va_end(args);
797 if (IS_ERR(dev))
798 return PTR_ERR(dev);
799
800 bdi->dev = dev;
801
802 bdi_debug_register(bdi, dev_name(dev));
803 set_bit(WB_registered, &bdi->wb.state);
804
805 spin_lock_bh(&bdi_lock);
806 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
807 spin_unlock_bh(&bdi_lock);
808
809 trace_writeback_bdi_register(bdi);
810 return 0;
811}
812EXPORT_SYMBOL(bdi_register);
813
814int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
815{
816 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
817}
818EXPORT_SYMBOL(bdi_register_dev);
819
820/*
821 * Remove bdi from bdi_list, and ensure that it is no longer visible
822 */
823static void bdi_remove_from_list(struct backing_dev_info *bdi)
824{
825 spin_lock_bh(&bdi_lock);
826 list_del_rcu(&bdi->bdi_list);
827 spin_unlock_bh(&bdi_lock);
828
829 synchronize_rcu_expedited();
830}
831
832/*
833 * Called when the device behind @bdi has been removed or ejected.
834 *
835 * We can't really do much here except for reducing the dirty ratio at
836 * the moment. In the future we should be able to set a flag so that
837 * the filesystem can handle errors at mark_inode_dirty time instead
838 * of only at writeback time.
839 */
840void bdi_unregister(struct backing_dev_info *bdi)
841{
842 if (WARN_ON_ONCE(!bdi->dev))
843 return;
844
845 bdi_set_min_ratio(bdi, 0);
846}
847EXPORT_SYMBOL(bdi_unregister);
848
b2e8fb6e
PZ
849void bdi_destroy(struct backing_dev_info *bdi)
850{
f0054bb1
TH
851 /* make sure nobody finds us on the bdi_list anymore */
852 bdi_remove_from_list(bdi);
853 wb_shutdown(&bdi->wb);
52ebea74 854 cgwb_bdi_destroy(bdi);
7a401a97 855
c4db59d3
CH
856 if (bdi->dev) {
857 bdi_debug_unregister(bdi);
858 device_unregister(bdi->dev);
859 bdi->dev = NULL;
860 }
861
f0054bb1 862 wb_exit(&bdi->wb);
b2e8fb6e
PZ
863}
864EXPORT_SYMBOL(bdi_destroy);
865
c3c53206
JA
866/*
867 * For use from filesystems to quickly init and register a bdi associated
868 * with dirty writeback
869 */
b4caecd4 870int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
c3c53206 871{
c3c53206
JA
872 int err;
873
874 bdi->name = name;
b4caecd4 875 bdi->capabilities = 0;
c3c53206
JA
876 err = bdi_init(bdi);
877 if (err)
878 return err;
879
02aa2a37
KC
880 err = bdi_register(bdi, NULL, "%.28s-%ld", name,
881 atomic_long_inc_return(&bdi_seq));
c3c53206
JA
882 if (err) {
883 bdi_destroy(bdi);
884 return err;
885 }
886
887 return 0;
888}
889EXPORT_SYMBOL(bdi_setup_and_register);
890
3fcfab16
AM
891static wait_queue_head_t congestion_wqh[2] = {
892 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
893 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
894 };
ec8a6f26 895static atomic_t nr_wb_congested[2];
3fcfab16 896
ec8a6f26 897void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
3fcfab16 898{
1faa16d2 899 wait_queue_head_t *wqh = &congestion_wqh[sync];
ec8a6f26 900 enum wb_state bit;
3fcfab16 901
4452226e 902 bit = sync ? WB_sync_congested : WB_async_congested;
ec8a6f26
TH
903 if (test_and_clear_bit(bit, &congested->state))
904 atomic_dec(&nr_wb_congested[sync]);
4e857c58 905 smp_mb__after_atomic();
3fcfab16
AM
906 if (waitqueue_active(wqh))
907 wake_up(wqh);
908}
ec8a6f26 909EXPORT_SYMBOL(clear_wb_congested);
3fcfab16 910
ec8a6f26 911void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
3fcfab16 912{
4452226e 913 enum wb_state bit;
3fcfab16 914
4452226e 915 bit = sync ? WB_sync_congested : WB_async_congested;
ec8a6f26
TH
916 if (!test_and_set_bit(bit, &congested->state))
917 atomic_inc(&nr_wb_congested[sync]);
3fcfab16 918}
ec8a6f26 919EXPORT_SYMBOL(set_wb_congested);
3fcfab16
AM
920
921/**
922 * congestion_wait - wait for a backing_dev to become uncongested
8aa7e847 923 * @sync: SYNC or ASYNC IO
3fcfab16
AM
924 * @timeout: timeout in jiffies
925 *
926 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
927 * write congestion. If no backing_devs are congested then just wait for the
928 * next write to be completed.
929 */
8aa7e847 930long congestion_wait(int sync, long timeout)
3fcfab16
AM
931{
932 long ret;
52bb9198 933 unsigned long start = jiffies;
3fcfab16 934 DEFINE_WAIT(wait);
8aa7e847 935 wait_queue_head_t *wqh = &congestion_wqh[sync];
3fcfab16
AM
936
937 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
938 ret = io_schedule_timeout(timeout);
939 finish_wait(wqh, &wait);
52bb9198
MG
940
941 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
942 jiffies_to_usecs(jiffies - start));
943
3fcfab16
AM
944 return ret;
945}
946EXPORT_SYMBOL(congestion_wait);
04fbfdc1 947
0e093d99
MG
948/**
949 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
950 * @zone: A zone to check if it is heavily congested
951 * @sync: SYNC or ASYNC IO
952 * @timeout: timeout in jiffies
953 *
954 * In the event of a congested backing_dev (any backing_dev) and the given
955 * @zone has experienced recent congestion, this waits for up to @timeout
956 * jiffies for either a BDI to exit congestion of the given @sync queue
957 * or a write to complete.
958 *
25985edc 959 * In the absence of zone congestion, cond_resched() is called to yield
0e093d99
MG
960 * the processor if necessary but otherwise does not sleep.
961 *
962 * The return value is 0 if the sleep is for the full timeout. Otherwise,
963 * it is the number of jiffies that were still remaining when the function
964 * returned. return_value == timeout implies the function did not sleep.
965 */
966long wait_iff_congested(struct zone *zone, int sync, long timeout)
967{
968 long ret;
969 unsigned long start = jiffies;
970 DEFINE_WAIT(wait);
971 wait_queue_head_t *wqh = &congestion_wqh[sync];
972
973 /*
974 * If there is no congestion, or heavy congestion is not being
975 * encountered in the current zone, yield if necessary instead
976 * of sleeping on the congestion queue
977 */
ec8a6f26 978 if (atomic_read(&nr_wb_congested[sync]) == 0 ||
57054651 979 !test_bit(ZONE_CONGESTED, &zone->flags)) {
0e093d99
MG
980 cond_resched();
981
982 /* In case we scheduled, work out time remaining */
983 ret = timeout - (jiffies - start);
984 if (ret < 0)
985 ret = 0;
986
987 goto out;
988 }
989
990 /* Sleep until uncongested or a write happens */
991 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
992 ret = io_schedule_timeout(timeout);
993 finish_wait(wqh, &wait);
994
995out:
996 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
997 jiffies_to_usecs(jiffies - start));
998
999 return ret;
1000}
1001EXPORT_SYMBOL(wait_iff_congested);
3965c9ae
WL
1002
1003int pdflush_proc_obsolete(struct ctl_table *table, int write,
1004 void __user *buffer, size_t *lenp, loff_t *ppos)
1005{
1006 char kbuf[] = "0\n";
1007
4c3bffc2 1008 if (*ppos || *lenp < sizeof(kbuf)) {
3965c9ae
WL
1009 *lenp = 0;
1010 return 0;
1011 }
1012
1013 if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1014 return -EFAULT;
1015 printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
1016 table->procname);
1017
1018 *lenp = 2;
1019 *ppos += *lenp;
1020 return 2;
1021}