bdi: Shutdown writeback on all cgwbs in cgwb_bdi_destroy()
[sfrench/cifs-2.6.git] / mm / backing-dev.c
1
2 #include <linux/wait.h>
3 #include <linux/backing-dev.h>
4 #include <linux/kthread.h>
5 #include <linux/freezer.h>
6 #include <linux/fs.h>
7 #include <linux/pagemap.h>
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/module.h>
11 #include <linux/writeback.h>
12 #include <linux/device.h>
13 #include <trace/events/writeback.h>
14
15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
17 struct backing_dev_info noop_backing_dev_info = {
18         .name           = "noop",
19         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
20 };
21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
22
23 static struct class *bdi_class;
24
25 /*
26  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
27  * locking.
28  */
29 DEFINE_SPINLOCK(bdi_lock);
30 LIST_HEAD(bdi_list);
31
32 /* bdi_wq serves all asynchronous writeback tasks */
33 struct workqueue_struct *bdi_wq;
34
35 #ifdef CONFIG_DEBUG_FS
36 #include <linux/debugfs.h>
37 #include <linux/seq_file.h>
38
39 static struct dentry *bdi_debug_root;
40
41 static void bdi_debug_init(void)
42 {
43         bdi_debug_root = debugfs_create_dir("bdi", NULL);
44 }
45
46 static int bdi_debug_stats_show(struct seq_file *m, void *v)
47 {
48         struct backing_dev_info *bdi = m->private;
49         struct bdi_writeback *wb = &bdi->wb;
50         unsigned long background_thresh;
51         unsigned long dirty_thresh;
52         unsigned long wb_thresh;
53         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
54         struct inode *inode;
55
56         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
57         spin_lock(&wb->list_lock);
58         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
59                 nr_dirty++;
60         list_for_each_entry(inode, &wb->b_io, i_io_list)
61                 nr_io++;
62         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
63                 nr_more_io++;
64         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
65                 if (inode->i_state & I_DIRTY_TIME)
66                         nr_dirty_time++;
67         spin_unlock(&wb->list_lock);
68
69         global_dirty_limits(&background_thresh, &dirty_thresh);
70         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
71
72 #define K(x) ((x) << (PAGE_SHIFT - 10))
73         seq_printf(m,
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"
79                    "BdiDirtied:         %10lu kB\n"
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"
85                    "b_dirty_time:       %10lu\n"
86                    "bdi_list:           %10u\n"
87                    "state:              %10lx\n",
88                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
89                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
90                    K(wb_thresh),
91                    K(dirty_thresh),
92                    K(background_thresh),
93                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
94                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
95                    (unsigned long) K(wb->write_bandwidth),
96                    nr_dirty,
97                    nr_io,
98                    nr_more_io,
99                    nr_dirty_time,
100                    !list_empty(&bdi->bdi_list), bdi->wb.state);
101 #undef K
102
103         return 0;
104 }
105
106 static 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
111 static 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
118 static 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
125 static 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
131 static inline void bdi_debug_init(void)
132 {
133 }
134 static inline void bdi_debug_register(struct backing_dev_info *bdi,
135                                       const char *name)
136 {
137 }
138 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
139 {
140 }
141 #endif
142
143 static 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);
148         unsigned long read_ahead_kb;
149         ssize_t ret;
150
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;
158 }
159
160 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
161
162 #define BDI_SHOW(name, expr)                                            \
163 static 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);  \
169 }                                                                       \
170 static DEVICE_ATTR_RW(name);
171
172 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
173
174 static 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);
178         unsigned int ratio;
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;
188
189         return ret;
190 }
191 BDI_SHOW(min_ratio, bdi->min_ratio)
192
193 static 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);
197         unsigned int ratio;
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;
207
208         return ret;
209 }
210 BDI_SHOW(max_ratio, bdi->max_ratio)
211
212 static 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 }
221 static DEVICE_ATTR_RO(stable_pages_required);
222
223 static 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,
229 };
230 ATTRIBUTE_GROUPS(bdi_dev);
231
232 static __init int bdi_class_init(void)
233 {
234         bdi_class = class_create(THIS_MODULE, "bdi");
235         if (IS_ERR(bdi_class))
236                 return PTR_ERR(bdi_class);
237
238         bdi_class->dev_groups = bdi_dev_groups;
239         bdi_debug_init();
240
241         return 0;
242 }
243 postcore_initcall(bdi_class_init);
244
245 static int __init default_bdi_init(void)
246 {
247         int err;
248
249         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
250                                               WQ_UNBOUND | WQ_SYSFS, 0);
251         if (!bdi_wq)
252                 return -ENOMEM;
253
254         err = bdi_init(&noop_backing_dev_info);
255
256         return err;
257 }
258 subsys_initcall(default_bdi_init);
259
260 /*
261  * This function is used when the first inode for this wb is marked dirty. It
262  * wakes-up the corresponding bdi thread which should then take care of the
263  * periodic background write-out of dirty inodes. Since the write-out would
264  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
265  * set up a timer which wakes the bdi thread up later.
266  *
267  * Note, we wouldn't bother setting up the timer, but this function is on the
268  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
269  * by delaying the wake-up.
270  *
271  * We have to be careful not to postpone flush work if it is scheduled for
272  * earlier. Thus we use queue_delayed_work().
273  */
274 void wb_wakeup_delayed(struct bdi_writeback *wb)
275 {
276         unsigned long timeout;
277
278         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
279         spin_lock_bh(&wb->work_lock);
280         if (test_bit(WB_registered, &wb->state))
281                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
282         spin_unlock_bh(&wb->work_lock);
283 }
284
285 /*
286  * Initial write bandwidth: 100 MB/s
287  */
288 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
289
290 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
291                    int blkcg_id, gfp_t gfp)
292 {
293         int i, err;
294
295         memset(wb, 0, sizeof(*wb));
296
297         if (wb != &bdi->wb)
298                 bdi_get(bdi);
299         wb->bdi = bdi;
300         wb->last_old_flush = jiffies;
301         INIT_LIST_HEAD(&wb->b_dirty);
302         INIT_LIST_HEAD(&wb->b_io);
303         INIT_LIST_HEAD(&wb->b_more_io);
304         INIT_LIST_HEAD(&wb->b_dirty_time);
305         spin_lock_init(&wb->list_lock);
306
307         wb->bw_time_stamp = jiffies;
308         wb->balanced_dirty_ratelimit = INIT_BW;
309         wb->dirty_ratelimit = INIT_BW;
310         wb->write_bandwidth = INIT_BW;
311         wb->avg_write_bandwidth = INIT_BW;
312
313         spin_lock_init(&wb->work_lock);
314         INIT_LIST_HEAD(&wb->work_list);
315         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
316         wb->dirty_sleep = jiffies;
317
318         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
319         if (!wb->congested) {
320                 err = -ENOMEM;
321                 goto out_put_bdi;
322         }
323
324         err = fprop_local_init_percpu(&wb->completions, gfp);
325         if (err)
326                 goto out_put_cong;
327
328         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
329                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
330                 if (err)
331                         goto out_destroy_stat;
332         }
333
334         return 0;
335
336 out_destroy_stat:
337         while (i--)
338                 percpu_counter_destroy(&wb->stat[i]);
339         fprop_local_destroy_percpu(&wb->completions);
340 out_put_cong:
341         wb_congested_put(wb->congested);
342 out_put_bdi:
343         if (wb != &bdi->wb)
344                 bdi_put(bdi);
345         return err;
346 }
347
348 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
349
350 /*
351  * Remove bdi from the global list and shutdown any threads we have running
352  */
353 static void wb_shutdown(struct bdi_writeback *wb)
354 {
355         /* Make sure nobody queues further work */
356         spin_lock_bh(&wb->work_lock);
357         if (!test_and_clear_bit(WB_registered, &wb->state)) {
358                 spin_unlock_bh(&wb->work_lock);
359                 /*
360                  * Wait for wb shutdown to finish if someone else is just
361                  * running wb_shutdown(). Otherwise we could proceed to wb /
362                  * bdi destruction before wb_shutdown() is finished.
363                  */
364                 wait_on_bit(&wb->state, WB_shutting_down, TASK_UNINTERRUPTIBLE);
365                 return;
366         }
367         set_bit(WB_shutting_down, &wb->state);
368         spin_unlock_bh(&wb->work_lock);
369
370         cgwb_remove_from_bdi_list(wb);
371         /*
372          * Drain work list and shutdown the delayed_work.  !WB_registered
373          * tells wb_workfn() that @wb is dying and its work_list needs to
374          * be drained no matter what.
375          */
376         mod_delayed_work(bdi_wq, &wb->dwork, 0);
377         flush_delayed_work(&wb->dwork);
378         WARN_ON(!list_empty(&wb->work_list));
379         /*
380          * Make sure bit gets cleared after shutdown is finished. Matches with
381          * the barrier provided by test_and_clear_bit() above.
382          */
383         smp_wmb();
384         clear_bit(WB_shutting_down, &wb->state);
385 }
386
387 static void wb_exit(struct bdi_writeback *wb)
388 {
389         int i;
390
391         WARN_ON(delayed_work_pending(&wb->dwork));
392
393         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
394                 percpu_counter_destroy(&wb->stat[i]);
395
396         fprop_local_destroy_percpu(&wb->completions);
397         wb_congested_put(wb->congested);
398         if (wb != &wb->bdi->wb)
399                 bdi_put(wb->bdi);
400 }
401
402 #ifdef CONFIG_CGROUP_WRITEBACK
403
404 #include <linux/memcontrol.h>
405
406 /*
407  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
408  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
409  * protected.  cgwb_release_wait is used to wait for the completion of cgwb
410  * releases from bdi destruction path.
411  */
412 static DEFINE_SPINLOCK(cgwb_lock);
413 static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
414
415 /**
416  * wb_congested_get_create - get or create a wb_congested
417  * @bdi: associated bdi
418  * @blkcg_id: ID of the associated blkcg
419  * @gfp: allocation mask
420  *
421  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
422  * The returned wb_congested has its reference count incremented.  Returns
423  * NULL on failure.
424  */
425 struct bdi_writeback_congested *
426 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
427 {
428         struct bdi_writeback_congested *new_congested = NULL, *congested;
429         struct rb_node **node, *parent;
430         unsigned long flags;
431 retry:
432         spin_lock_irqsave(&cgwb_lock, flags);
433
434         node = &bdi->cgwb_congested_tree.rb_node;
435         parent = NULL;
436
437         while (*node != NULL) {
438                 parent = *node;
439                 congested = rb_entry(parent, struct bdi_writeback_congested,
440                                      rb_node);
441                 if (congested->blkcg_id < blkcg_id)
442                         node = &parent->rb_left;
443                 else if (congested->blkcg_id > blkcg_id)
444                         node = &parent->rb_right;
445                 else
446                         goto found;
447         }
448
449         if (new_congested) {
450                 /* !found and storage for new one already allocated, insert */
451                 congested = new_congested;
452                 new_congested = NULL;
453                 rb_link_node(&congested->rb_node, parent, node);
454                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
455                 goto found;
456         }
457
458         spin_unlock_irqrestore(&cgwb_lock, flags);
459
460         /* allocate storage for new one and retry */
461         new_congested = kzalloc(sizeof(*new_congested), gfp);
462         if (!new_congested)
463                 return NULL;
464
465         atomic_set(&new_congested->refcnt, 0);
466         new_congested->__bdi = bdi;
467         new_congested->blkcg_id = blkcg_id;
468         goto retry;
469
470 found:
471         atomic_inc(&congested->refcnt);
472         spin_unlock_irqrestore(&cgwb_lock, flags);
473         kfree(new_congested);
474         return congested;
475 }
476
477 /**
478  * wb_congested_put - put a wb_congested
479  * @congested: wb_congested to put
480  *
481  * Put @congested and destroy it if the refcnt reaches zero.
482  */
483 void wb_congested_put(struct bdi_writeback_congested *congested)
484 {
485         unsigned long flags;
486
487         local_irq_save(flags);
488         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
489                 local_irq_restore(flags);
490                 return;
491         }
492
493         /* bdi might already have been destroyed leaving @congested unlinked */
494         if (congested->__bdi) {
495                 rb_erase(&congested->rb_node,
496                          &congested->__bdi->cgwb_congested_tree);
497                 congested->__bdi = NULL;
498         }
499
500         spin_unlock_irqrestore(&cgwb_lock, flags);
501         kfree(congested);
502 }
503
504 static void cgwb_release_workfn(struct work_struct *work)
505 {
506         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
507                                                 release_work);
508         struct backing_dev_info *bdi = wb->bdi;
509
510         wb_shutdown(wb);
511
512         css_put(wb->memcg_css);
513         css_put(wb->blkcg_css);
514
515         fprop_local_destroy_percpu(&wb->memcg_completions);
516         percpu_ref_exit(&wb->refcnt);
517         wb_exit(wb);
518         kfree_rcu(wb, rcu);
519
520         if (atomic_dec_and_test(&bdi->usage_cnt))
521                 wake_up_all(&cgwb_release_wait);
522 }
523
524 static void cgwb_release(struct percpu_ref *refcnt)
525 {
526         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
527                                                 refcnt);
528         schedule_work(&wb->release_work);
529 }
530
531 static void cgwb_kill(struct bdi_writeback *wb)
532 {
533         lockdep_assert_held(&cgwb_lock);
534
535         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
536         list_del(&wb->memcg_node);
537         list_del(&wb->blkcg_node);
538         percpu_ref_kill(&wb->refcnt);
539 }
540
541 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
542 {
543         spin_lock_irq(&cgwb_lock);
544         list_del_rcu(&wb->bdi_node);
545         spin_unlock_irq(&cgwb_lock);
546 }
547
548 static int cgwb_create(struct backing_dev_info *bdi,
549                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
550 {
551         struct mem_cgroup *memcg;
552         struct cgroup_subsys_state *blkcg_css;
553         struct blkcg *blkcg;
554         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
555         struct bdi_writeback *wb;
556         unsigned long flags;
557         int ret = 0;
558
559         memcg = mem_cgroup_from_css(memcg_css);
560         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
561         blkcg = css_to_blkcg(blkcg_css);
562         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
563         blkcg_cgwb_list = &blkcg->cgwb_list;
564
565         /* look up again under lock and discard on blkcg mismatch */
566         spin_lock_irqsave(&cgwb_lock, flags);
567         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
568         if (wb && wb->blkcg_css != blkcg_css) {
569                 cgwb_kill(wb);
570                 wb = NULL;
571         }
572         spin_unlock_irqrestore(&cgwb_lock, flags);
573         if (wb)
574                 goto out_put;
575
576         /* need to create a new one */
577         wb = kmalloc(sizeof(*wb), gfp);
578         if (!wb)
579                 return -ENOMEM;
580
581         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
582         if (ret)
583                 goto err_free;
584
585         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
586         if (ret)
587                 goto err_wb_exit;
588
589         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
590         if (ret)
591                 goto err_ref_exit;
592
593         wb->memcg_css = memcg_css;
594         wb->blkcg_css = blkcg_css;
595         INIT_WORK(&wb->release_work, cgwb_release_workfn);
596         set_bit(WB_registered, &wb->state);
597
598         /*
599          * The root wb determines the registered state of the whole bdi and
600          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
601          * whether they're still online.  Don't link @wb if any is dead.
602          * See wb_memcg_offline() and wb_blkcg_offline().
603          */
604         ret = -ENODEV;
605         spin_lock_irqsave(&cgwb_lock, flags);
606         if (test_bit(WB_registered, &bdi->wb.state) &&
607             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
608                 /* we might have raced another instance of this function */
609                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
610                 if (!ret) {
611                         atomic_inc(&bdi->usage_cnt);
612                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
613                         list_add(&wb->memcg_node, memcg_cgwb_list);
614                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
615                         css_get(memcg_css);
616                         css_get(blkcg_css);
617                 }
618         }
619         spin_unlock_irqrestore(&cgwb_lock, flags);
620         if (ret) {
621                 if (ret == -EEXIST)
622                         ret = 0;
623                 goto err_fprop_exit;
624         }
625         goto out_put;
626
627 err_fprop_exit:
628         fprop_local_destroy_percpu(&wb->memcg_completions);
629 err_ref_exit:
630         percpu_ref_exit(&wb->refcnt);
631 err_wb_exit:
632         wb_exit(wb);
633 err_free:
634         kfree(wb);
635 out_put:
636         css_put(blkcg_css);
637         return ret;
638 }
639
640 /**
641  * wb_get_create - get wb for a given memcg, create if necessary
642  * @bdi: target bdi
643  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
644  * @gfp: allocation mask to use
645  *
646  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
647  * create one.  The returned wb has its refcount incremented.
648  *
649  * This function uses css_get() on @memcg_css and thus expects its refcnt
650  * to be positive on invocation.  IOW, rcu_read_lock() protection on
651  * @memcg_css isn't enough.  try_get it before calling this function.
652  *
653  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
654  * memcg on the default hierarchy, memcg association is guaranteed to be
655  * more specific (equal or descendant to the associated blkcg) and thus can
656  * identify both the memcg and blkcg associations.
657  *
658  * Because the blkcg associated with a memcg may change as blkcg is enabled
659  * and disabled closer to root in the hierarchy, each wb keeps track of
660  * both the memcg and blkcg associated with it and verifies the blkcg on
661  * each lookup.  On mismatch, the existing wb is discarded and a new one is
662  * created.
663  */
664 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
665                                     struct cgroup_subsys_state *memcg_css,
666                                     gfp_t gfp)
667 {
668         struct bdi_writeback *wb;
669
670         might_sleep_if(gfpflags_allow_blocking(gfp));
671
672         if (!memcg_css->parent)
673                 return &bdi->wb;
674
675         do {
676                 rcu_read_lock();
677                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
678                 if (wb) {
679                         struct cgroup_subsys_state *blkcg_css;
680
681                         /* see whether the blkcg association has changed */
682                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
683                                                      &io_cgrp_subsys);
684                         if (unlikely(wb->blkcg_css != blkcg_css ||
685                                      !wb_tryget(wb)))
686                                 wb = NULL;
687                         css_put(blkcg_css);
688                 }
689                 rcu_read_unlock();
690         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
691
692         return wb;
693 }
694
695 static int cgwb_bdi_init(struct backing_dev_info *bdi)
696 {
697         int ret;
698
699         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
700         bdi->cgwb_congested_tree = RB_ROOT;
701         atomic_set(&bdi->usage_cnt, 1);
702
703         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
704         if (!ret) {
705                 bdi->wb.memcg_css = &root_mem_cgroup->css;
706                 bdi->wb.blkcg_css = blkcg_root_css;
707         }
708         return ret;
709 }
710
711 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
712 {
713         struct radix_tree_iter iter;
714         void **slot;
715         struct bdi_writeback *wb;
716
717         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
718
719         spin_lock_irq(&cgwb_lock);
720         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
721                 cgwb_kill(*slot);
722
723         while (!list_empty(&bdi->wb_list)) {
724                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
725                                       bdi_node);
726                 spin_unlock_irq(&cgwb_lock);
727                 wb_shutdown(wb);
728                 spin_lock_irq(&cgwb_lock);
729         }
730         spin_unlock_irq(&cgwb_lock);
731
732         /*
733          * All cgwb's must be shutdown and released before returning.  Drain
734          * the usage counter to wait for all cgwb's ever created on @bdi.
735          */
736         atomic_dec(&bdi->usage_cnt);
737         wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
738         /*
739          * Grab back our reference so that we hold it when @bdi gets
740          * re-registered.
741          */
742         atomic_inc(&bdi->usage_cnt);
743 }
744
745 /**
746  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
747  * @memcg: memcg being offlined
748  *
749  * Also prevents creation of any new wb's associated with @memcg.
750  */
751 void wb_memcg_offline(struct mem_cgroup *memcg)
752 {
753         LIST_HEAD(to_destroy);
754         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
755         struct bdi_writeback *wb, *next;
756
757         spin_lock_irq(&cgwb_lock);
758         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
759                 cgwb_kill(wb);
760         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
761         spin_unlock_irq(&cgwb_lock);
762 }
763
764 /**
765  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
766  * @blkcg: blkcg being offlined
767  *
768  * Also prevents creation of any new wb's associated with @blkcg.
769  */
770 void wb_blkcg_offline(struct blkcg *blkcg)
771 {
772         LIST_HEAD(to_destroy);
773         struct bdi_writeback *wb, *next;
774
775         spin_lock_irq(&cgwb_lock);
776         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
777                 cgwb_kill(wb);
778         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
779         spin_unlock_irq(&cgwb_lock);
780 }
781
782 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
783 {
784         struct rb_node *rbn;
785
786         spin_lock_irq(&cgwb_lock);
787         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
788                 struct bdi_writeback_congested *congested =
789                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
790
791                 rb_erase(rbn, &bdi->cgwb_congested_tree);
792                 congested->__bdi = NULL;        /* mark @congested unlinked */
793         }
794         spin_unlock_irq(&cgwb_lock);
795 }
796
797 static void cgwb_bdi_register(struct backing_dev_info *bdi)
798 {
799         spin_lock_irq(&cgwb_lock);
800         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
801         spin_unlock_irq(&cgwb_lock);
802 }
803
804 #else   /* CONFIG_CGROUP_WRITEBACK */
805
806 static int cgwb_bdi_init(struct backing_dev_info *bdi)
807 {
808         int err;
809
810         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
811         if (!bdi->wb_congested)
812                 return -ENOMEM;
813
814         atomic_set(&bdi->wb_congested->refcnt, 1);
815
816         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
817         if (err) {
818                 wb_congested_put(bdi->wb_congested);
819                 return err;
820         }
821         return 0;
822 }
823
824 static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
825
826 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
827 {
828         wb_congested_put(bdi->wb_congested);
829 }
830
831 static void cgwb_bdi_register(struct backing_dev_info *bdi)
832 {
833         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
834 }
835
836 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
837 {
838         list_del_rcu(&wb->bdi_node);
839 }
840
841 #endif  /* CONFIG_CGROUP_WRITEBACK */
842
843 int bdi_init(struct backing_dev_info *bdi)
844 {
845         int ret;
846
847         bdi->dev = NULL;
848
849         kref_init(&bdi->refcnt);
850         bdi->min_ratio = 0;
851         bdi->max_ratio = 100;
852         bdi->max_prop_frac = FPROP_FRAC_BASE;
853         INIT_LIST_HEAD(&bdi->bdi_list);
854         INIT_LIST_HEAD(&bdi->wb_list);
855         init_waitqueue_head(&bdi->wb_waitq);
856
857         ret = cgwb_bdi_init(bdi);
858
859         return ret;
860 }
861 EXPORT_SYMBOL(bdi_init);
862
863 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
864 {
865         struct backing_dev_info *bdi;
866
867         bdi = kmalloc_node(sizeof(struct backing_dev_info),
868                            gfp_mask | __GFP_ZERO, node_id);
869         if (!bdi)
870                 return NULL;
871
872         if (bdi_init(bdi)) {
873                 kfree(bdi);
874                 return NULL;
875         }
876         return bdi;
877 }
878
879 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
880                 const char *fmt, ...)
881 {
882         va_list args;
883         struct device *dev;
884
885         if (bdi->dev)   /* The driver needs to use separate queues per device */
886                 return 0;
887
888         va_start(args, fmt);
889         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
890         va_end(args);
891         if (IS_ERR(dev))
892                 return PTR_ERR(dev);
893
894         cgwb_bdi_register(bdi);
895         bdi->dev = dev;
896
897         bdi_debug_register(bdi, dev_name(dev));
898         set_bit(WB_registered, &bdi->wb.state);
899
900         spin_lock_bh(&bdi_lock);
901         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
902         spin_unlock_bh(&bdi_lock);
903
904         trace_writeback_bdi_register(bdi);
905         return 0;
906 }
907 EXPORT_SYMBOL(bdi_register);
908
909 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
910 {
911         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
912 }
913 EXPORT_SYMBOL(bdi_register_dev);
914
915 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
916 {
917         int rc;
918
919         rc = bdi_register(bdi, NULL, "%u:%u", MAJOR(owner->devt),
920                         MINOR(owner->devt));
921         if (rc)
922                 return rc;
923         /* Leaking owner reference... */
924         WARN_ON(bdi->owner);
925         bdi->owner = owner;
926         get_device(owner);
927         return 0;
928 }
929 EXPORT_SYMBOL(bdi_register_owner);
930
931 /*
932  * Remove bdi from bdi_list, and ensure that it is no longer visible
933  */
934 static void bdi_remove_from_list(struct backing_dev_info *bdi)
935 {
936         spin_lock_bh(&bdi_lock);
937         list_del_rcu(&bdi->bdi_list);
938         spin_unlock_bh(&bdi_lock);
939
940         synchronize_rcu_expedited();
941 }
942
943 void bdi_unregister(struct backing_dev_info *bdi)
944 {
945         /* make sure nobody finds us on the bdi_list anymore */
946         bdi_remove_from_list(bdi);
947         wb_shutdown(&bdi->wb);
948         cgwb_bdi_destroy(bdi);
949
950         if (bdi->dev) {
951                 bdi_debug_unregister(bdi);
952                 device_unregister(bdi->dev);
953                 bdi->dev = NULL;
954         }
955
956         if (bdi->owner) {
957                 put_device(bdi->owner);
958                 bdi->owner = NULL;
959         }
960 }
961
962 static void bdi_exit(struct backing_dev_info *bdi)
963 {
964         WARN_ON_ONCE(bdi->dev);
965         wb_exit(&bdi->wb);
966         cgwb_bdi_exit(bdi);
967 }
968
969 static void release_bdi(struct kref *ref)
970 {
971         struct backing_dev_info *bdi =
972                         container_of(ref, struct backing_dev_info, refcnt);
973
974         bdi_exit(bdi);
975         kfree(bdi);
976 }
977
978 void bdi_put(struct backing_dev_info *bdi)
979 {
980         kref_put(&bdi->refcnt, release_bdi);
981 }
982
983 void bdi_destroy(struct backing_dev_info *bdi)
984 {
985         bdi_unregister(bdi);
986         bdi_exit(bdi);
987 }
988 EXPORT_SYMBOL(bdi_destroy);
989
990 /*
991  * For use from filesystems to quickly init and register a bdi associated
992  * with dirty writeback
993  */
994 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
995 {
996         int err;
997
998         bdi->name = name;
999         bdi->capabilities = 0;
1000         err = bdi_init(bdi);
1001         if (err)
1002                 return err;
1003
1004         err = bdi_register(bdi, NULL, "%.28s-%ld", name,
1005                            atomic_long_inc_return(&bdi_seq));
1006         if (err) {
1007                 bdi_destroy(bdi);
1008                 return err;
1009         }
1010
1011         return 0;
1012 }
1013 EXPORT_SYMBOL(bdi_setup_and_register);
1014
1015 static wait_queue_head_t congestion_wqh[2] = {
1016                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
1017                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
1018         };
1019 static atomic_t nr_wb_congested[2];
1020
1021 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
1022 {
1023         wait_queue_head_t *wqh = &congestion_wqh[sync];
1024         enum wb_congested_state bit;
1025
1026         bit = sync ? WB_sync_congested : WB_async_congested;
1027         if (test_and_clear_bit(bit, &congested->state))
1028                 atomic_dec(&nr_wb_congested[sync]);
1029         smp_mb__after_atomic();
1030         if (waitqueue_active(wqh))
1031                 wake_up(wqh);
1032 }
1033 EXPORT_SYMBOL(clear_wb_congested);
1034
1035 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1036 {
1037         enum wb_congested_state bit;
1038
1039         bit = sync ? WB_sync_congested : WB_async_congested;
1040         if (!test_and_set_bit(bit, &congested->state))
1041                 atomic_inc(&nr_wb_congested[sync]);
1042 }
1043 EXPORT_SYMBOL(set_wb_congested);
1044
1045 /**
1046  * congestion_wait - wait for a backing_dev to become uncongested
1047  * @sync: SYNC or ASYNC IO
1048  * @timeout: timeout in jiffies
1049  *
1050  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1051  * write congestion.  If no backing_devs are congested then just wait for the
1052  * next write to be completed.
1053  */
1054 long congestion_wait(int sync, long timeout)
1055 {
1056         long ret;
1057         unsigned long start = jiffies;
1058         DEFINE_WAIT(wait);
1059         wait_queue_head_t *wqh = &congestion_wqh[sync];
1060
1061         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1062         ret = io_schedule_timeout(timeout);
1063         finish_wait(wqh, &wait);
1064
1065         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1066                                         jiffies_to_usecs(jiffies - start));
1067
1068         return ret;
1069 }
1070 EXPORT_SYMBOL(congestion_wait);
1071
1072 /**
1073  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1074  * @pgdat: A pgdat to check if it is heavily congested
1075  * @sync: SYNC or ASYNC IO
1076  * @timeout: timeout in jiffies
1077  *
1078  * In the event of a congested backing_dev (any backing_dev) and the given
1079  * @pgdat has experienced recent congestion, this waits for up to @timeout
1080  * jiffies for either a BDI to exit congestion of the given @sync queue
1081  * or a write to complete.
1082  *
1083  * In the absence of pgdat congestion, cond_resched() is called to yield
1084  * the processor if necessary but otherwise does not sleep.
1085  *
1086  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1087  * it is the number of jiffies that were still remaining when the function
1088  * returned. return_value == timeout implies the function did not sleep.
1089  */
1090 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1091 {
1092         long ret;
1093         unsigned long start = jiffies;
1094         DEFINE_WAIT(wait);
1095         wait_queue_head_t *wqh = &congestion_wqh[sync];
1096
1097         /*
1098          * If there is no congestion, or heavy congestion is not being
1099          * encountered in the current pgdat, yield if necessary instead
1100          * of sleeping on the congestion queue
1101          */
1102         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1103             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1104                 cond_resched();
1105
1106                 /* In case we scheduled, work out time remaining */
1107                 ret = timeout - (jiffies - start);
1108                 if (ret < 0)
1109                         ret = 0;
1110
1111                 goto out;
1112         }
1113
1114         /* Sleep until uncongested or a write happens */
1115         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1116         ret = io_schedule_timeout(timeout);
1117         finish_wait(wqh, &wait);
1118
1119 out:
1120         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1121                                         jiffies_to_usecs(jiffies - start));
1122
1123         return ret;
1124 }
1125 EXPORT_SYMBOL(wait_iff_congested);
1126
1127 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1128                         void __user *buffer, size_t *lenp, loff_t *ppos)
1129 {
1130         char kbuf[] = "0\n";
1131
1132         if (*ppos || *lenp < sizeof(kbuf)) {
1133                 *lenp = 0;
1134                 return 0;
1135         }
1136
1137         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1138                 return -EFAULT;
1139         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1140                      table->procname);
1141
1142         *lenp = 2;
1143         *ppos += *lenp;
1144         return 2;
1145 }