2 #include <linux/wait.h>
3 #include <linux/backing-dev.h>
4 #include <linux/kthread.h>
5 #include <linux/freezer.h>
7 #include <linux/pagemap.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>
15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
17 struct backing_dev_info default_backing_dev_info = {
19 .ra_pages = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
21 EXPORT_SYMBOL_GPL(default_backing_dev_info);
23 struct backing_dev_info noop_backing_dev_info = {
25 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK,
27 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
29 static struct class *bdi_class;
32 * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
35 DEFINE_SPINLOCK(bdi_lock);
38 /* bdi_wq serves all asynchronous writeback tasks */
39 struct workqueue_struct *bdi_wq;
41 static void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
44 spin_lock(&wb1->list_lock);
45 spin_lock_nested(&wb2->list_lock, 1);
47 spin_lock(&wb2->list_lock);
48 spin_lock_nested(&wb1->list_lock, 1);
52 #ifdef CONFIG_DEBUG_FS
53 #include <linux/debugfs.h>
54 #include <linux/seq_file.h>
56 static struct dentry *bdi_debug_root;
58 static void bdi_debug_init(void)
60 bdi_debug_root = debugfs_create_dir("bdi", NULL);
63 static int bdi_debug_stats_show(struct seq_file *m, void *v)
65 struct backing_dev_info *bdi = m->private;
66 struct bdi_writeback *wb = &bdi->wb;
67 unsigned long background_thresh;
68 unsigned long dirty_thresh;
69 unsigned long bdi_thresh;
70 unsigned long nr_dirty, nr_io, nr_more_io;
73 nr_dirty = nr_io = nr_more_io = 0;
74 spin_lock(&wb->list_lock);
75 list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
77 list_for_each_entry(inode, &wb->b_io, i_wb_list)
79 list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
81 spin_unlock(&wb->list_lock);
83 global_dirty_limits(&background_thresh, &dirty_thresh);
84 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
86 #define K(x) ((x) << (PAGE_SHIFT - 10))
88 "BdiWriteback: %10lu kB\n"
89 "BdiReclaimable: %10lu kB\n"
90 "BdiDirtyThresh: %10lu kB\n"
91 "DirtyThresh: %10lu kB\n"
92 "BackgroundThresh: %10lu kB\n"
93 "BdiDirtied: %10lu kB\n"
94 "BdiWritten: %10lu kB\n"
95 "BdiWriteBandwidth: %10lu kBps\n"
101 (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
102 (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
105 K(background_thresh),
106 (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
107 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
108 (unsigned long) K(bdi->write_bandwidth),
112 !list_empty(&bdi->bdi_list), bdi->state);
118 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
120 return single_open(file, bdi_debug_stats_show, inode->i_private);
123 static const struct file_operations bdi_debug_stats_fops = {
124 .open = bdi_debug_stats_open,
127 .release = single_release,
130 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
132 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
133 bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
134 bdi, &bdi_debug_stats_fops);
137 static void bdi_debug_unregister(struct backing_dev_info *bdi)
139 debugfs_remove(bdi->debug_stats);
140 debugfs_remove(bdi->debug_dir);
143 static inline void bdi_debug_init(void)
146 static inline void bdi_debug_register(struct backing_dev_info *bdi,
150 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
155 static ssize_t read_ahead_kb_store(struct device *dev,
156 struct device_attribute *attr,
157 const char *buf, size_t count)
159 struct backing_dev_info *bdi = dev_get_drvdata(dev);
160 unsigned long read_ahead_kb;
163 ret = kstrtoul(buf, 10, &read_ahead_kb);
167 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
172 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
174 #define BDI_SHOW(name, expr) \
175 static ssize_t name##_show(struct device *dev, \
176 struct device_attribute *attr, char *page) \
178 struct backing_dev_info *bdi = dev_get_drvdata(dev); \
180 return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr); \
182 static DEVICE_ATTR_RW(name);
184 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
186 static ssize_t min_ratio_store(struct device *dev,
187 struct device_attribute *attr, const char *buf, size_t count)
189 struct backing_dev_info *bdi = dev_get_drvdata(dev);
193 ret = kstrtouint(buf, 10, &ratio);
197 ret = bdi_set_min_ratio(bdi, ratio);
203 BDI_SHOW(min_ratio, bdi->min_ratio)
205 static ssize_t max_ratio_store(struct device *dev,
206 struct device_attribute *attr, const char *buf, size_t count)
208 struct backing_dev_info *bdi = dev_get_drvdata(dev);
212 ret = kstrtouint(buf, 10, &ratio);
216 ret = bdi_set_max_ratio(bdi, ratio);
222 BDI_SHOW(max_ratio, bdi->max_ratio)
224 static ssize_t stable_pages_required_show(struct device *dev,
225 struct device_attribute *attr,
228 struct backing_dev_info *bdi = dev_get_drvdata(dev);
230 return snprintf(page, PAGE_SIZE-1, "%d\n",
231 bdi_cap_stable_pages_required(bdi) ? 1 : 0);
233 static DEVICE_ATTR_RO(stable_pages_required);
235 static struct attribute *bdi_dev_attrs[] = {
236 &dev_attr_read_ahead_kb.attr,
237 &dev_attr_min_ratio.attr,
238 &dev_attr_max_ratio.attr,
239 &dev_attr_stable_pages_required.attr,
242 ATTRIBUTE_GROUPS(bdi_dev);
244 static __init int bdi_class_init(void)
246 bdi_class = class_create(THIS_MODULE, "bdi");
247 if (IS_ERR(bdi_class))
248 return PTR_ERR(bdi_class);
250 bdi_class->dev_groups = bdi_dev_groups;
254 postcore_initcall(bdi_class_init);
256 static int __init default_bdi_init(void)
260 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
261 WQ_UNBOUND | WQ_SYSFS, 0);
265 err = bdi_init(&default_backing_dev_info);
267 bdi_register(&default_backing_dev_info, NULL, "default");
268 err = bdi_init(&noop_backing_dev_info);
272 subsys_initcall(default_bdi_init);
274 int bdi_has_dirty_io(struct backing_dev_info *bdi)
276 return wb_has_dirty_io(&bdi->wb);
280 * This function is used when the first inode for this bdi is marked dirty. It
281 * wakes-up the corresponding bdi thread which should then take care of the
282 * periodic background write-out of dirty inodes. Since the write-out would
283 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
284 * set up a timer which wakes the bdi thread up later.
286 * Note, we wouldn't bother setting up the timer, but this function is on the
287 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
288 * by delaying the wake-up.
290 * We have to be careful not to postpone flush work if it is scheduled for
291 * earlier. Thus we use queue_delayed_work().
293 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
295 unsigned long timeout;
297 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
298 spin_lock_bh(&bdi->wb_lock);
299 if (test_bit(BDI_registered, &bdi->state))
300 queue_delayed_work(bdi_wq, &bdi->wb.dwork, timeout);
301 spin_unlock_bh(&bdi->wb_lock);
305 * Remove bdi from bdi_list, and ensure that it is no longer visible
307 static void bdi_remove_from_list(struct backing_dev_info *bdi)
309 spin_lock_bh(&bdi_lock);
310 list_del_rcu(&bdi->bdi_list);
311 spin_unlock_bh(&bdi_lock);
313 synchronize_rcu_expedited();
316 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
317 const char *fmt, ...)
322 if (bdi->dev) /* The driver needs to use separate queues per device */
326 dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
333 bdi_debug_register(bdi, dev_name(dev));
334 set_bit(BDI_registered, &bdi->state);
336 spin_lock_bh(&bdi_lock);
337 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
338 spin_unlock_bh(&bdi_lock);
340 trace_writeback_bdi_register(bdi);
343 EXPORT_SYMBOL(bdi_register);
345 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
347 return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
349 EXPORT_SYMBOL(bdi_register_dev);
352 * Remove bdi from the global list and shutdown any threads we have running
354 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
356 if (!bdi_cap_writeback_dirty(bdi))
360 * Make sure nobody finds us on the bdi_list anymore
362 bdi_remove_from_list(bdi);
364 /* Make sure nobody queues further work */
365 spin_lock_bh(&bdi->wb_lock);
366 clear_bit(BDI_registered, &bdi->state);
367 spin_unlock_bh(&bdi->wb_lock);
370 * Drain work list and shutdown the delayed_work. At this point,
371 * @bdi->bdi_list is empty telling bdi_Writeback_workfn() that @bdi
372 * is dying and its work_list needs to be drained no matter what.
374 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
375 flush_delayed_work(&bdi->wb.dwork);
376 WARN_ON(!list_empty(&bdi->work_list));
377 WARN_ON(delayed_work_pending(&bdi->wb.dwork));
381 * This bdi is going away now, make sure that no super_blocks point to it
383 static void bdi_prune_sb(struct backing_dev_info *bdi)
385 struct super_block *sb;
388 list_for_each_entry(sb, &super_blocks, s_list) {
389 if (sb->s_bdi == bdi)
390 sb->s_bdi = &default_backing_dev_info;
392 spin_unlock(&sb_lock);
395 void bdi_unregister(struct backing_dev_info *bdi)
398 bdi_set_min_ratio(bdi, 0);
399 trace_writeback_bdi_unregister(bdi);
402 bdi_wb_shutdown(bdi);
403 bdi_debug_unregister(bdi);
404 device_unregister(bdi->dev);
408 EXPORT_SYMBOL(bdi_unregister);
410 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
412 memset(wb, 0, sizeof(*wb));
415 wb->last_old_flush = jiffies;
416 INIT_LIST_HEAD(&wb->b_dirty);
417 INIT_LIST_HEAD(&wb->b_io);
418 INIT_LIST_HEAD(&wb->b_more_io);
419 spin_lock_init(&wb->list_lock);
420 INIT_DELAYED_WORK(&wb->dwork, bdi_writeback_workfn);
424 * Initial write bandwidth: 100 MB/s
426 #define INIT_BW (100 << (20 - PAGE_SHIFT))
428 int bdi_init(struct backing_dev_info *bdi)
435 bdi->max_ratio = 100;
436 bdi->max_prop_frac = FPROP_FRAC_BASE;
437 spin_lock_init(&bdi->wb_lock);
438 INIT_LIST_HEAD(&bdi->bdi_list);
439 INIT_LIST_HEAD(&bdi->work_list);
441 bdi_wb_init(&bdi->wb, bdi);
443 for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
444 err = percpu_counter_init(&bdi->bdi_stat[i], 0, GFP_KERNEL);
449 bdi->dirty_exceeded = 0;
451 bdi->bw_time_stamp = jiffies;
452 bdi->written_stamp = 0;
454 bdi->balanced_dirty_ratelimit = INIT_BW;
455 bdi->dirty_ratelimit = INIT_BW;
456 bdi->write_bandwidth = INIT_BW;
457 bdi->avg_write_bandwidth = INIT_BW;
459 err = fprop_local_init_percpu(&bdi->completions, GFP_KERNEL);
464 percpu_counter_destroy(&bdi->bdi_stat[i]);
469 EXPORT_SYMBOL(bdi_init);
471 void bdi_destroy(struct backing_dev_info *bdi)
476 * Splice our entries to the default_backing_dev_info. This
477 * condition shouldn't happen. @wb must be empty at this point and
478 * dirty inodes on it might cause other issues. This workaround is
479 * added by ce5f8e779519 ("writeback: splice dirty inode entries to
480 * default bdi on bdi_destroy()") without root-causing the issue.
482 * http://lkml.kernel.org/g/1253038617-30204-11-git-send-email-jens.axboe@oracle.com
483 * http://thread.gmane.org/gmane.linux.file-systems/35341/focus=35350
485 * We should probably add WARN_ON() to find out whether it still
486 * happens and track it down if so.
488 if (bdi_has_dirty_io(bdi)) {
489 struct bdi_writeback *dst = &default_backing_dev_info.wb;
491 bdi_lock_two(&bdi->wb, dst);
492 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
493 list_splice(&bdi->wb.b_io, &dst->b_io);
494 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
495 spin_unlock(&bdi->wb.list_lock);
496 spin_unlock(&dst->list_lock);
501 WARN_ON(delayed_work_pending(&bdi->wb.dwork));
503 for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
504 percpu_counter_destroy(&bdi->bdi_stat[i]);
506 fprop_local_destroy_percpu(&bdi->completions);
508 EXPORT_SYMBOL(bdi_destroy);
511 * For use from filesystems to quickly init and register a bdi associated
512 * with dirty writeback
514 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
519 bdi->capabilities = 0;
524 err = bdi_register(bdi, NULL, "%.28s-%ld", name,
525 atomic_long_inc_return(&bdi_seq));
533 EXPORT_SYMBOL(bdi_setup_and_register);
535 static wait_queue_head_t congestion_wqh[2] = {
536 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
537 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
539 static atomic_t nr_bdi_congested[2];
541 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
544 wait_queue_head_t *wqh = &congestion_wqh[sync];
546 bit = sync ? BDI_sync_congested : BDI_async_congested;
547 if (test_and_clear_bit(bit, &bdi->state))
548 atomic_dec(&nr_bdi_congested[sync]);
549 smp_mb__after_atomic();
550 if (waitqueue_active(wqh))
553 EXPORT_SYMBOL(clear_bdi_congested);
555 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
559 bit = sync ? BDI_sync_congested : BDI_async_congested;
560 if (!test_and_set_bit(bit, &bdi->state))
561 atomic_inc(&nr_bdi_congested[sync]);
563 EXPORT_SYMBOL(set_bdi_congested);
566 * congestion_wait - wait for a backing_dev to become uncongested
567 * @sync: SYNC or ASYNC IO
568 * @timeout: timeout in jiffies
570 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
571 * write congestion. If no backing_devs are congested then just wait for the
572 * next write to be completed.
574 long congestion_wait(int sync, long timeout)
577 unsigned long start = jiffies;
579 wait_queue_head_t *wqh = &congestion_wqh[sync];
581 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
582 ret = io_schedule_timeout(timeout);
583 finish_wait(wqh, &wait);
585 trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
586 jiffies_to_usecs(jiffies - start));
590 EXPORT_SYMBOL(congestion_wait);
593 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
594 * @zone: A zone to check if it is heavily congested
595 * @sync: SYNC or ASYNC IO
596 * @timeout: timeout in jiffies
598 * In the event of a congested backing_dev (any backing_dev) and the given
599 * @zone has experienced recent congestion, this waits for up to @timeout
600 * jiffies for either a BDI to exit congestion of the given @sync queue
601 * or a write to complete.
603 * In the absence of zone congestion, cond_resched() is called to yield
604 * the processor if necessary but otherwise does not sleep.
606 * The return value is 0 if the sleep is for the full timeout. Otherwise,
607 * it is the number of jiffies that were still remaining when the function
608 * returned. return_value == timeout implies the function did not sleep.
610 long wait_iff_congested(struct zone *zone, int sync, long timeout)
613 unsigned long start = jiffies;
615 wait_queue_head_t *wqh = &congestion_wqh[sync];
618 * If there is no congestion, or heavy congestion is not being
619 * encountered in the current zone, yield if necessary instead
620 * of sleeping on the congestion queue
622 if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
623 !test_bit(ZONE_CONGESTED, &zone->flags)) {
626 /* In case we scheduled, work out time remaining */
627 ret = timeout - (jiffies - start);
634 /* Sleep until uncongested or a write happens */
635 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
636 ret = io_schedule_timeout(timeout);
637 finish_wait(wqh, &wait);
640 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
641 jiffies_to_usecs(jiffies - start));
645 EXPORT_SYMBOL(wait_iff_congested);
647 int pdflush_proc_obsolete(struct ctl_table *table, int write,
648 void __user *buffer, size_t *lenp, loff_t *ppos)
652 if (*ppos || *lenp < sizeof(kbuf)) {
657 if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
659 printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",