projects/linux·blob v3.13 · refs

668 lines · 16.5 KiB

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
15static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
16
17struct backing_dev_info default_backing_dev_info = {
18 .name = "default",
19 .ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
20 .state = 0,
21 .capabilities = BDI_CAP_MAP_COPY,
22};
23EXPORT_SYMBOL_GPL(default_backing_dev_info);
24
25struct backing_dev_info noop_backing_dev_info = {
26 .name = "noop",
27 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
28};
29EXPORT_SYMBOL_GPL(noop_backing_dev_info);
30
31static struct class *bdi_class;
32
33/*
34 * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
35 * locking.
36 */
37DEFINE_SPINLOCK(bdi_lock);
38LIST_HEAD(bdi_list);
39
40/* bdi_wq serves all asynchronous writeback tasks */
41struct workqueue_struct *bdi_wq;
42
43void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
44{
45 if (wb1 < wb2) {
46 spin_lock(&wb1->list_lock);
47 spin_lock_nested(&wb2->list_lock, 1);
48 } else {
49 spin_lock(&wb2->list_lock);
50 spin_lock_nested(&wb1->list_lock, 1);
51 }
52}
53
54#ifdef CONFIG_DEBUG_FS
55#include <linux/debugfs.h>
56#include <linux/seq_file.h>
57
58static struct dentry *bdi_debug_root;
59
60static void bdi_debug_init(void)
61{
62 bdi_debug_root = debugfs_create_dir("bdi", NULL);
63}
64
65static int bdi_debug_stats_show(struct seq_file *m, void *v)
66{
67 struct backing_dev_info *bdi = m->private;
68 struct bdi_writeback *wb = &bdi->wb;
69 unsigned long background_thresh;
70 unsigned long dirty_thresh;
71 unsigned long bdi_thresh;
72 unsigned long nr_dirty, nr_io, nr_more_io;
73 struct inode *inode;
74
75 nr_dirty = nr_io = nr_more_io = 0;
76 spin_lock(&wb->list_lock);
77 list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
78 nr_dirty++;
79 list_for_each_entry(inode, &wb->b_io, i_wb_list)
80 nr_io++;
81 list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
82 nr_more_io++;
83 spin_unlock(&wb->list_lock);
84
85 global_dirty_limits(&background_thresh, &dirty_thresh);
86 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
87
88#define K(x) ((x) << (PAGE_SHIFT - 10))
89 seq_printf(m,
90 "BdiWriteback: %10lu kB\n"
91 "BdiReclaimable: %10lu kB\n"
92 "BdiDirtyThresh: %10lu kB\n"
93 "DirtyThresh: %10lu kB\n"
94 "BackgroundThresh: %10lu kB\n"
95 "BdiDirtied: %10lu kB\n"
96 "BdiWritten: %10lu kB\n"
97 "BdiWriteBandwidth: %10lu kBps\n"
98 "b_dirty: %10lu\n"
99 "b_io: %10lu\n"
100 "b_more_io: %10lu\n"
101 "bdi_list: %10u\n"
102 "state: %10lx\n",
103 (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
104 (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
105 K(bdi_thresh),
106 K(dirty_thresh),
107 K(background_thresh),
108 (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
109 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
110 (unsigned long) K(bdi->write_bandwidth),
111 nr_dirty,
112 nr_io,
113 nr_more_io,
114 !list_empty(&bdi->bdi_list), bdi->state);
115#undef K
116
117 return 0;
118}
119
120static int bdi_debug_stats_open(struct inode *inode, struct file *file)
121{
122 return single_open(file, bdi_debug_stats_show, inode->i_private);
123}
124
125static const struct file_operations bdi_debug_stats_fops = {
126 .open = bdi_debug_stats_open,
127 .read = seq_read,
128 .llseek = seq_lseek,
129 .release = single_release,
130};
131
132static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
133{
134 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
135 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
136 bdi, &bdi_debug_stats_fops);
137}
138
139static void bdi_debug_unregister(struct backing_dev_info *bdi)
140{
141 debugfs_remove(bdi->debug_stats);
142 debugfs_remove(bdi->debug_dir);
143}
144#else
145static inline void bdi_debug_init(void)
146{
147}
148static inline void bdi_debug_register(struct backing_dev_info *bdi,
149 const char *name)
150{
151}
152static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
153{
154}
155#endif
156
157static ssize_t read_ahead_kb_store(struct device *dev,
158 struct device_attribute *attr,
159 const char *buf, size_t count)
160{
161 struct backing_dev_info *bdi = dev_get_drvdata(dev);
162 unsigned long read_ahead_kb;
163 ssize_t ret;
164
165 ret = kstrtoul(buf, 10, &read_ahead_kb);
166 if (ret < 0)
167 return ret;
168
169 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
170
171 return count;
172}
173
174#define K(pages) ((pages) << (PAGE_SHIFT - 10))
175
176#define BDI_SHOW(name, expr) \
177static ssize_t name##_show(struct device *dev, \
178 struct device_attribute *attr, char *page) \
179{ \
180 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
181 \
182 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
183} \
184static DEVICE_ATTR_RW(name);
185
186BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
187
188static ssize_t min_ratio_store(struct device *dev,
189 struct device_attribute *attr, const char *buf, size_t count)
190{
191 struct backing_dev_info *bdi = dev_get_drvdata(dev);
192 unsigned int ratio;
193 ssize_t ret;
194
195 ret = kstrtouint(buf, 10, &ratio);
196 if (ret < 0)
197 return ret;
198
199 ret = bdi_set_min_ratio(bdi, ratio);
200 if (!ret)
201 ret = count;
202
203 return ret;
204}
205BDI_SHOW(min_ratio, bdi->min_ratio)
206
207static ssize_t max_ratio_store(struct device *dev,
208 struct device_attribute *attr, const char *buf, size_t count)
209{
210 struct backing_dev_info *bdi = dev_get_drvdata(dev);
211 unsigned int ratio;
212 ssize_t ret;
213
214 ret = kstrtouint(buf, 10, &ratio);
215 if (ret < 0)
216 return ret;
217
218 ret = bdi_set_max_ratio(bdi, ratio);
219 if (!ret)
220 ret = count;
221
222 return ret;
223}
224BDI_SHOW(max_ratio, bdi->max_ratio)
225
226static ssize_t stable_pages_required_show(struct device *dev,
227 struct device_attribute *attr,
228 char *page)
229{
230 struct backing_dev_info *bdi = dev_get_drvdata(dev);
231
232 return snprintf(page, PAGE_SIZE-1, "%d\n",
233 bdi_cap_stable_pages_required(bdi) ? 1 : 0);
234}
235static DEVICE_ATTR_RO(stable_pages_required);
236
237static struct attribute *bdi_dev_attrs[] = {
238 &dev_attr_read_ahead_kb.attr,
239 &dev_attr_min_ratio.attr,
240 &dev_attr_max_ratio.attr,
241 &dev_attr_stable_pages_required.attr,
242 NULL,
243};
244ATTRIBUTE_GROUPS(bdi_dev);
245
246static __init int bdi_class_init(void)
247{
248 bdi_class = class_create(THIS_MODULE, "bdi");
249 if (IS_ERR(bdi_class))
250 return PTR_ERR(bdi_class);
251
252 bdi_class->dev_groups = bdi_dev_groups;
253 bdi_debug_init();
254 return 0;
255}
256postcore_initcall(bdi_class_init);
257
258static int __init default_bdi_init(void)
259{
260 int err;
261
262 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
263 WQ_UNBOUND | WQ_SYSFS, 0);
264 if (!bdi_wq)
265 return -ENOMEM;
266
267 err = bdi_init(&default_backing_dev_info);
268 if (!err)
269 bdi_register(&default_backing_dev_info, NULL, "default");
270 err = bdi_init(&noop_backing_dev_info);
271
272 return err;
273}
274subsys_initcall(default_bdi_init);
275
276int bdi_has_dirty_io(struct backing_dev_info *bdi)
277{
278 return wb_has_dirty_io(&bdi->wb);
279}
280
281/*
282 * This function is used when the first inode for this bdi is marked dirty. It
283 * wakes-up the corresponding bdi thread which should then take care of the
284 * periodic background write-out of dirty inodes. Since the write-out would
285 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
286 * set up a timer which wakes the bdi thread up later.
287 *
288 * Note, we wouldn't bother setting up the timer, but this function is on the
289 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
290 * by delaying the wake-up.
291 */
292void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
293{
294 unsigned long timeout;
295
296 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
297 mod_delayed_work(bdi_wq, &bdi->wb.dwork, timeout);
298}
299
300/*
301 * Remove bdi from bdi_list, and ensure that it is no longer visible
302 */
303static void bdi_remove_from_list(struct backing_dev_info *bdi)
304{
305 spin_lock_bh(&bdi_lock);
306 list_del_rcu(&bdi->bdi_list);
307 spin_unlock_bh(&bdi_lock);
308
309 synchronize_rcu_expedited();
310
311 /* bdi_list is now unused, clear it to mark @bdi dying */
312 INIT_LIST_HEAD(&bdi->bdi_list);
313}
314
315int bdi_register(struct backing_dev_info *bdi, struct device *parent,
316 const char *fmt, ...)
317{
318 va_list args;
319 struct device *dev;
320
321 if (bdi->dev) /* The driver needs to use separate queues per device */
322 return 0;
323
324 va_start(args, fmt);
325 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
326 va_end(args);
327 if (IS_ERR(dev))
328 return PTR_ERR(dev);
329
330 bdi->dev = dev;
331
332 bdi_debug_register(bdi, dev_name(dev));
333 set_bit(BDI_registered, &bdi->state);
334
335 spin_lock_bh(&bdi_lock);
336 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
337 spin_unlock_bh(&bdi_lock);
338
339 trace_writeback_bdi_register(bdi);
340 return 0;
341}
342EXPORT_SYMBOL(bdi_register);
343
344int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
345{
346 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
347}
348EXPORT_SYMBOL(bdi_register_dev);
349
350/*
351 * Remove bdi from the global list and shutdown any threads we have running
352 */
353static void bdi_wb_shutdown(struct backing_dev_info *bdi)
354{
355 if (!bdi_cap_writeback_dirty(bdi))
356 return;
357
358 /*
359 * Make sure nobody finds us on the bdi_list anymore
360 */
361 bdi_remove_from_list(bdi);
362
363 /*
364 * Drain work list and shutdown the delayed_work. At this point,
365 * @bdi->bdi_list is empty telling bdi_Writeback_workfn() that @bdi
366 * is dying and its work_list needs to be drained no matter what.
367 */
368 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
369 flush_delayed_work(&bdi->wb.dwork);
370 WARN_ON(!list_empty(&bdi->work_list));
371
372 /*
373 * This shouldn't be necessary unless @bdi for some reason has
374 * unflushed dirty IO after work_list is drained. Do it anyway
375 * just in case.
376 */
377 cancel_delayed_work_sync(&bdi->wb.dwork);
378}
379
380/*
381 * This bdi is going away now, make sure that no super_blocks point to it
382 */
383static void bdi_prune_sb(struct backing_dev_info *bdi)
384{
385 struct super_block *sb;
386
387 spin_lock(&sb_lock);
388 list_for_each_entry(sb, &super_blocks, s_list) {
389 if (sb->s_bdi == bdi)
390 sb->s_bdi = &default_backing_dev_info;
391 }
392 spin_unlock(&sb_lock);
393}
394
395void bdi_unregister(struct backing_dev_info *bdi)
396{
397 struct device *dev = bdi->dev;
398
399 if (dev) {
400 bdi_set_min_ratio(bdi, 0);
401 trace_writeback_bdi_unregister(bdi);
402 bdi_prune_sb(bdi);
403
404 bdi_wb_shutdown(bdi);
405 bdi_debug_unregister(bdi);
406
407 spin_lock_bh(&bdi->wb_lock);
408 bdi->dev = NULL;
409 spin_unlock_bh(&bdi->wb_lock);
410
411 device_unregister(dev);
412 }
413}
414EXPORT_SYMBOL(bdi_unregister);
415
416static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
417{
418 memset(wb, 0, sizeof(*wb));
419
420 wb->bdi = bdi;
421 wb->last_old_flush = jiffies;
422 INIT_LIST_HEAD(&wb->b_dirty);
423 INIT_LIST_HEAD(&wb->b_io);
424 INIT_LIST_HEAD(&wb->b_more_io);
425 spin_lock_init(&wb->list_lock);
426 INIT_DELAYED_WORK(&wb->dwork, bdi_writeback_workfn);
427}
428
429/*
430 * Initial write bandwidth: 100 MB/s
431 */
432#define INIT_BW (100 << (20 - PAGE_SHIFT))
433
434int bdi_init(struct backing_dev_info *bdi)
435{
436 int i, err;
437
438 bdi->dev = NULL;
439
440 bdi->min_ratio = 0;
441 bdi->max_ratio = 100;
442 bdi->max_prop_frac = FPROP_FRAC_BASE;
443 spin_lock_init(&bdi->wb_lock);
444 INIT_LIST_HEAD(&bdi->bdi_list);
445 INIT_LIST_HEAD(&bdi->work_list);
446
447 bdi_wb_init(&bdi->wb, bdi);
448
449 for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
450 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
451 if (err)
452 goto err;
453 }
454
455 bdi->dirty_exceeded = 0;
456
457 bdi->bw_time_stamp = jiffies;
458 bdi->written_stamp = 0;
459
460 bdi->balanced_dirty_ratelimit = INIT_BW;
461 bdi->dirty_ratelimit = INIT_BW;
462 bdi->write_bandwidth = INIT_BW;
463 bdi->avg_write_bandwidth = INIT_BW;
464
465 err = fprop_local_init_percpu(&bdi->completions);
466
467 if (err) {
468err:
469 while (i--)
470 percpu_counter_destroy(&bdi->bdi_stat[i]);
471 }
472
473 return err;
474}
475EXPORT_SYMBOL(bdi_init);
476
477void bdi_destroy(struct backing_dev_info *bdi)
478{
479 int i;
480
481 /*
482 * Splice our entries to the default_backing_dev_info, if this
483 * bdi disappears
484 */
485 if (bdi_has_dirty_io(bdi)) {
486 struct bdi_writeback *dst = &default_backing_dev_info.wb;
487
488 bdi_lock_two(&bdi->wb, dst);
489 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
490 list_splice(&bdi->wb.b_io, &dst->b_io);
491 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
492 spin_unlock(&bdi->wb.list_lock);
493 spin_unlock(&dst->list_lock);
494 }
495
496 bdi_unregister(bdi);
497
498 /*
499 * If bdi_unregister() had already been called earlier, the dwork
500 * could still be pending because bdi_prune_sb() can race with the
501 * bdi_wakeup_thread_delayed() calls from __mark_inode_dirty().
502 */
503 cancel_delayed_work_sync(&bdi->wb.dwork);
504
505 for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
506 percpu_counter_destroy(&bdi->bdi_stat[i]);
507
508 fprop_local_destroy_percpu(&bdi->completions);
509}
510EXPORT_SYMBOL(bdi_destroy);
511
512/*
513 * For use from filesystems to quickly init and register a bdi associated
514 * with dirty writeback
515 */
516int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
517 unsigned int cap)
518{
519 int err;
520
521 bdi->name = name;
522 bdi->capabilities = cap;
523 err = bdi_init(bdi);
524 if (err)
525 return err;
526
527 err = bdi_register(bdi, NULL, "%.28s-%ld", name,
528 atomic_long_inc_return(&bdi_seq));
529 if (err) {
530 bdi_destroy(bdi);
531 return err;
532 }
533
534 return 0;
535}
536EXPORT_SYMBOL(bdi_setup_and_register);
537
538static wait_queue_head_t congestion_wqh[2] = {
539 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
540 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
541 };
542static atomic_t nr_bdi_congested[2];
543
544void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
545{
546 enum bdi_state bit;
547 wait_queue_head_t *wqh = &congestion_wqh[sync];
548
549 bit = sync ? BDI_sync_congested : BDI_async_congested;
550 if (test_and_clear_bit(bit, &bdi->state))
551 atomic_dec(&nr_bdi_congested[sync]);
552 smp_mb__after_clear_bit();
553 if (waitqueue_active(wqh))
554 wake_up(wqh);
555}
556EXPORT_SYMBOL(clear_bdi_congested);
557
558void set_bdi_congested(struct backing_dev_info *bdi, int sync)
559{
560 enum bdi_state bit;
561
562 bit = sync ? BDI_sync_congested : BDI_async_congested;
563 if (!test_and_set_bit(bit, &bdi->state))
564 atomic_inc(&nr_bdi_congested[sync]);
565}
566EXPORT_SYMBOL(set_bdi_congested);
567
568/**
569 * congestion_wait - wait for a backing_dev to become uncongested
570 * @sync: SYNC or ASYNC IO
571 * @timeout: timeout in jiffies
572 *
573 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
574 * write congestion. If no backing_devs are congested then just wait for the
575 * next write to be completed.
576 */
577long congestion_wait(int sync, long timeout)
578{
579 long ret;
580 unsigned long start = jiffies;
581 DEFINE_WAIT(wait);
582 wait_queue_head_t *wqh = &congestion_wqh[sync];
583
584 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
585 ret = io_schedule_timeout(timeout);
586 finish_wait(wqh, &wait);
587
588 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
589 jiffies_to_usecs(jiffies - start));
590
591 return ret;
592}
593EXPORT_SYMBOL(congestion_wait);
594
595/**
596 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
597 * @zone: A zone to check if it is heavily congested
598 * @sync: SYNC or ASYNC IO
599 * @timeout: timeout in jiffies
600 *
601 * In the event of a congested backing_dev (any backing_dev) and the given
602 * @zone has experienced recent congestion, this waits for up to @timeout
603 * jiffies for either a BDI to exit congestion of the given @sync queue
604 * or a write to complete.
605 *
606 * In the absence of zone congestion, cond_resched() is called to yield
607 * the processor if necessary but otherwise does not sleep.
608 *
609 * The return value is 0 if the sleep is for the full timeout. Otherwise,
610 * it is the number of jiffies that were still remaining when the function
611 * returned. return_value == timeout implies the function did not sleep.
612 */
613long wait_iff_congested(struct zone *zone, int sync, long timeout)
614{
615 long ret;
616 unsigned long start = jiffies;
617 DEFINE_WAIT(wait);
618 wait_queue_head_t *wqh = &congestion_wqh[sync];
619
620 /*
621 * If there is no congestion, or heavy congestion is not being
622 * encountered in the current zone, yield if necessary instead
623 * of sleeping on the congestion queue
624 */
625 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
626 !zone_is_reclaim_congested(zone)) {
627 cond_resched();
628
629 /* In case we scheduled, work out time remaining */
630 ret = timeout - (jiffies - start);
631 if (ret < 0)
632 ret = 0;
633
634 goto out;
635 }
636
637 /* Sleep until uncongested or a write happens */
638 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
639 ret = io_schedule_timeout(timeout);
640 finish_wait(wqh, &wait);
641
642out:
643 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
644 jiffies_to_usecs(jiffies - start));
645
646 return ret;
647}
648EXPORT_SYMBOL(wait_iff_congested);
649
650int pdflush_proc_obsolete(struct ctl_table *table, int write,
651 void __user *buffer, size_t *lenp, loff_t *ppos)
652{
653 char kbuf[] = "0\n";
654
655 if (*ppos || *lenp < sizeof(kbuf)) {
656 *lenp = 0;
657 return 0;
658 }
659
660 if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
661 return -EFAULT;
662 printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
663 table->procname);
664
665 *lenp = 2;
666 *ppos += *lenp;
667 return 2;
668}