Linux Audio

Check our new training course

Loading...
v3.1
  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 and bdi_pending_list, as well as
 35 * reader side protection for bdi_pending_list. bdi_list has RCU reader side
 36 * locking.
 37 */
 38DEFINE_SPINLOCK(bdi_lock);
 39LIST_HEAD(bdi_list);
 40LIST_HEAD(bdi_pending_list);
 41
 42static struct task_struct *sync_supers_tsk;
 43static struct timer_list sync_supers_timer;
 44
 45static int bdi_sync_supers(void *);
 46static void sync_supers_timer_fn(unsigned long);
 47
 48void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
 49{
 50	if (wb1 < wb2) {
 51		spin_lock(&wb1->list_lock);
 52		spin_lock_nested(&wb2->list_lock, 1);
 53	} else {
 54		spin_lock(&wb2->list_lock);
 55		spin_lock_nested(&wb1->list_lock, 1);
 56	}
 57}
 58
 59#ifdef CONFIG_DEBUG_FS
 60#include <linux/debugfs.h>
 61#include <linux/seq_file.h>
 62
 63static struct dentry *bdi_debug_root;
 64
 65static void bdi_debug_init(void)
 66{
 67	bdi_debug_root = debugfs_create_dir("bdi", NULL);
 68}
 69
 70static int bdi_debug_stats_show(struct seq_file *m, void *v)
 71{
 72	struct backing_dev_info *bdi = m->private;
 73	struct bdi_writeback *wb = &bdi->wb;
 74	unsigned long background_thresh;
 75	unsigned long dirty_thresh;
 76	unsigned long bdi_thresh;
 77	unsigned long nr_dirty, nr_io, nr_more_io;
 78	struct inode *inode;
 79
 80	nr_dirty = nr_io = nr_more_io = 0;
 81	spin_lock(&wb->list_lock);
 82	list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
 83		nr_dirty++;
 84	list_for_each_entry(inode, &wb->b_io, i_wb_list)
 85		nr_io++;
 86	list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
 87		nr_more_io++;
 
 
 
 88	spin_unlock(&wb->list_lock);
 89
 90	global_dirty_limits(&background_thresh, &dirty_thresh);
 91	bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
 92
 93#define K(x) ((x) << (PAGE_SHIFT - 10))
 94	seq_printf(m,
 95		   "BdiWriteback:       %10lu kB\n"
 96		   "BdiReclaimable:     %10lu kB\n"
 97		   "BdiDirtyThresh:     %10lu kB\n"
 98		   "DirtyThresh:        %10lu kB\n"
 99		   "BackgroundThresh:   %10lu kB\n"
 
100		   "BdiWritten:         %10lu kB\n"
101		   "BdiWriteBandwidth:  %10lu kBps\n"
102		   "b_dirty:            %10lu\n"
103		   "b_io:               %10lu\n"
104		   "b_more_io:          %10lu\n"
 
105		   "bdi_list:           %10u\n"
106		   "state:              %10lx\n",
107		   (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
108		   (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
109		   K(bdi_thresh),
110		   K(dirty_thresh),
111		   K(background_thresh),
112		   (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
113		   (unsigned long) K(bdi->write_bandwidth),
 
114		   nr_dirty,
115		   nr_io,
116		   nr_more_io,
117		   !list_empty(&bdi->bdi_list), bdi->state);
 
118#undef K
119
120	return 0;
121}
122
123static int bdi_debug_stats_open(struct inode *inode, struct file *file)
124{
125	return single_open(file, bdi_debug_stats_show, inode->i_private);
126}
127
128static const struct file_operations bdi_debug_stats_fops = {
129	.open		= bdi_debug_stats_open,
130	.read		= seq_read,
131	.llseek		= seq_lseek,
132	.release	= single_release,
133};
134
135static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
136{
137	bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
138	bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
139					       bdi, &bdi_debug_stats_fops);
140}
141
142static void bdi_debug_unregister(struct backing_dev_info *bdi)
143{
144	debugfs_remove(bdi->debug_stats);
145	debugfs_remove(bdi->debug_dir);
146}
147#else
148static inline void bdi_debug_init(void)
149{
150}
151static inline void bdi_debug_register(struct backing_dev_info *bdi,
152				      const char *name)
153{
154}
155static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
156{
157}
158#endif
159
160static ssize_t read_ahead_kb_store(struct device *dev,
161				  struct device_attribute *attr,
162				  const char *buf, size_t count)
163{
164	struct backing_dev_info *bdi = dev_get_drvdata(dev);
165	char *end;
166	unsigned long read_ahead_kb;
167	ssize_t ret = -EINVAL;
168
169	read_ahead_kb = simple_strtoul(buf, &end, 10);
170	if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
171		bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
172		ret = count;
173	}
174	return ret;
 
175}
176
177#define K(pages) ((pages) << (PAGE_SHIFT - 10))
178
179#define BDI_SHOW(name, expr)						\
180static ssize_t name##_show(struct device *dev,				\
181			   struct device_attribute *attr, char *page)	\
182{									\
183	struct backing_dev_info *bdi = dev_get_drvdata(dev);		\
184									\
185	return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);	\
186}
 
187
188BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
189
190static ssize_t min_ratio_store(struct device *dev,
191		struct device_attribute *attr, const char *buf, size_t count)
192{
193	struct backing_dev_info *bdi = dev_get_drvdata(dev);
194	char *end;
195	unsigned int ratio;
196	ssize_t ret = -EINVAL;
 
 
 
 
 
 
 
 
197
198	ratio = simple_strtoul(buf, &end, 10);
199	if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
200		ret = bdi_set_min_ratio(bdi, ratio);
201		if (!ret)
202			ret = count;
203	}
204	return ret;
205}
206BDI_SHOW(min_ratio, bdi->min_ratio)
207
208static ssize_t max_ratio_store(struct device *dev,
209		struct device_attribute *attr, const char *buf, size_t count)
210{
211	struct backing_dev_info *bdi = dev_get_drvdata(dev);
212	char *end;
213	unsigned int ratio;
214	ssize_t ret = -EINVAL;
 
 
 
 
 
 
 
 
215
216	ratio = simple_strtoul(buf, &end, 10);
217	if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
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
226#define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 
 
 
 
227
228static struct device_attribute bdi_dev_attrs[] = {
229	__ATTR_RW(read_ahead_kb),
230	__ATTR_RW(min_ratio),
231	__ATTR_RW(max_ratio),
232	__ATTR_NULL,
 
 
 
 
 
 
233};
 
234
235static __init int bdi_class_init(void)
236{
237	bdi_class = class_create(THIS_MODULE, "bdi");
238	if (IS_ERR(bdi_class))
239		return PTR_ERR(bdi_class);
240
241	bdi_class->dev_attrs = bdi_dev_attrs;
242	bdi_debug_init();
243	return 0;
244}
245postcore_initcall(bdi_class_init);
246
247static int __init default_bdi_init(void)
248{
249	int err;
250
251	sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
252	BUG_ON(IS_ERR(sync_supers_tsk));
253
254	setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
255	bdi_arm_supers_timer();
256
257	err = bdi_init(&default_backing_dev_info);
258	if (!err)
259		bdi_register(&default_backing_dev_info, NULL, "default");
260	err = bdi_init(&noop_backing_dev_info);
261
262	return err;
263}
264subsys_initcall(default_bdi_init);
265
266int bdi_has_dirty_io(struct backing_dev_info *bdi)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
267{
268	return wb_has_dirty_io(&bdi->wb);
 
 
 
 
 
 
269}
270
271/*
272 * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
273 * or we risk deadlocking on ->s_umount. The longer term solution would be
274 * to implement sync_supers_bdi() or similar and simply do it from the
275 * bdi writeback thread individually.
276 */
277static int bdi_sync_supers(void *unused)
278{
279	set_user_nice(current, 0);
280
281	while (!kthread_should_stop()) {
282		set_current_state(TASK_INTERRUPTIBLE);
283		schedule();
284
285		/*
286		 * Do this periodically, like kupdated() did before.
287		 */
288		sync_supers();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
289	}
290
291	return 0;
 
 
 
 
 
 
 
 
292}
293
294void bdi_arm_supers_timer(void)
 
 
 
295{
296	unsigned long next;
297
298	if (!dirty_writeback_interval)
 
299		return;
 
 
300
301	next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
302	mod_timer(&sync_supers_timer, round_jiffies_up(next));
 
 
 
 
 
 
303}
304
305static void sync_supers_timer_fn(unsigned long unused)
306{
307	wake_up_process(sync_supers_tsk);
308	bdi_arm_supers_timer();
 
 
 
 
 
 
 
309}
310
311static void wakeup_timer_fn(unsigned long data)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
312{
313	struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
314
315	spin_lock_bh(&bdi->wb_lock);
316	if (bdi->wb.task) {
317		trace_writeback_wake_thread(bdi);
318		wake_up_process(bdi->wb.task);
319	} else {
320		/*
321		 * When bdi tasks are inactive for long time, they are killed.
322		 * In this case we have to wake-up the forker thread which
323		 * should create and run the bdi thread.
324		 */
325		trace_writeback_wake_forker_thread(bdi);
326		wake_up_process(default_backing_dev_info.wb.task);
327	}
328	spin_unlock_bh(&bdi->wb_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
329}
330
331/*
332 * This function is used when the first inode for this bdi is marked dirty. It
333 * wakes-up the corresponding bdi thread which should then take care of the
334 * periodic background write-out of dirty inodes. Since the write-out would
335 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
336 * set up a timer which wakes the bdi thread up later.
337 *
338 * Note, we wouldn't bother setting up the timer, but this function is on the
339 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
340 * by delaying the wake-up.
341 */
342void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
343{
344	unsigned long timeout;
345
346	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
347	mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 
 
 
 
 
 
 
 
 
 
 
 
 
348}
349
350/*
351 * Calculate the longest interval (jiffies) bdi threads are allowed to be
352 * inactive.
353 */
354static unsigned long bdi_longest_inactive(void)
355{
356	unsigned long interval;
 
 
 
 
 
 
 
 
 
 
 
357
358	interval = msecs_to_jiffies(dirty_writeback_interval * 10);
359	return max(5UL * 60 * HZ, interval);
 
 
 
 
 
360}
361
362/*
363 * Clear pending bit and wakeup anybody waiting for flusher thread creation or
364 * shutdown
365 */
366static void bdi_clear_pending(struct backing_dev_info *bdi)
367{
368	clear_bit(BDI_pending, &bdi->state);
369	smp_mb__after_clear_bit();
370	wake_up_bit(&bdi->state, BDI_pending);
371}
372
373static int bdi_forker_thread(void *ptr)
374{
375	struct bdi_writeback *me = ptr;
 
 
 
 
 
 
376
377	current->flags |= PF_SWAPWRITE;
378	set_freezable();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
379
380	/*
381	 * Our parent may run at a different priority, just set us to normal
 
 
 
382	 */
383	set_user_nice(current, 0);
384
385	for (;;) {
386		struct task_struct *task = NULL;
387		struct backing_dev_info *bdi;
388		enum {
389			NO_ACTION,   /* Nothing to do */
390			FORK_THREAD, /* Fork bdi thread */
391			KILL_THREAD, /* Kill inactive bdi thread */
392		} action = NO_ACTION;
393
394		/*
395		 * Temporary measure, we want to make sure we don't see
396		 * dirty data on the default backing_dev_info
397		 */
398		if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
399			del_timer(&me->wakeup_timer);
400			wb_do_writeback(me, 0);
401		}
 
 
 
 
 
 
 
 
402
403		spin_lock_bh(&bdi_lock);
404		/*
405		 * In the following loop we are going to check whether we have
406		 * some work to do without any synchronization with tasks
407		 * waking us up to do work for them. So we have to set task
408		 * state already here so that we don't miss wakeups coming
409		 * after we verify some condition.
410		 */
411		set_current_state(TASK_INTERRUPTIBLE);
412
413		list_for_each_entry(bdi, &bdi_list, bdi_list) {
414			bool have_dirty_io;
415
416			if (!bdi_cap_writeback_dirty(bdi) ||
417			     bdi_cap_flush_forker(bdi))
418				continue;
419
420			WARN(!test_bit(BDI_registered, &bdi->state),
421			     "bdi %p/%s is not registered!\n", bdi, bdi->name);
422
423			have_dirty_io = !list_empty(&bdi->work_list) ||
424					wb_has_dirty_io(&bdi->wb);
425
426			/*
427			 * If the bdi has work to do, but the thread does not
428			 * exist - create it.
429			 */
430			if (!bdi->wb.task && have_dirty_io) {
431				/*
432				 * Set the pending bit - if someone will try to
433				 * unregister this bdi - it'll wait on this bit.
434				 */
435				set_bit(BDI_pending, &bdi->state);
436				action = FORK_THREAD;
437				break;
438			}
439
440			spin_lock(&bdi->wb_lock);
441
442			/*
443			 * If there is no work to do and the bdi thread was
444			 * inactive long enough - kill it. The wb_lock is taken
445			 * to make sure no-one adds more work to this bdi and
446			 * wakes the bdi thread up.
447			 */
448			if (bdi->wb.task && !have_dirty_io &&
449			    time_after(jiffies, bdi->wb.last_active +
450						bdi_longest_inactive())) {
451				task = bdi->wb.task;
452				bdi->wb.task = NULL;
453				spin_unlock(&bdi->wb_lock);
454				set_bit(BDI_pending, &bdi->state);
455				action = KILL_THREAD;
456				break;
457			}
458			spin_unlock(&bdi->wb_lock);
459		}
460		spin_unlock_bh(&bdi_lock);
461
462		/* Keep working if default bdi still has things to do */
463		if (!list_empty(&me->bdi->work_list))
464			__set_current_state(TASK_RUNNING);
465
466		switch (action) {
467		case FORK_THREAD:
468			__set_current_state(TASK_RUNNING);
469			task = kthread_create(bdi_writeback_thread, &bdi->wb,
470					      "flush-%s", dev_name(bdi->dev));
471			if (IS_ERR(task)) {
472				/*
473				 * If thread creation fails, force writeout of
474				 * the bdi from the thread. Hopefully 1024 is
475				 * large enough for efficient IO.
476				 */
477				writeback_inodes_wb(&bdi->wb, 1024);
478			} else {
479				/*
480				 * The spinlock makes sure we do not lose
481				 * wake-ups when racing with 'bdi_queue_work()'.
482				 * And as soon as the bdi thread is visible, we
483				 * can start it.
484				 */
485				spin_lock_bh(&bdi->wb_lock);
486				bdi->wb.task = task;
487				spin_unlock_bh(&bdi->wb_lock);
488				wake_up_process(task);
489			}
490			bdi_clear_pending(bdi);
491			break;
492
493		case KILL_THREAD:
494			__set_current_state(TASK_RUNNING);
495			kthread_stop(task);
496			bdi_clear_pending(bdi);
497			break;
498
499		case NO_ACTION:
500			if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
501				/*
502				 * There are no dirty data. The only thing we
503				 * should now care about is checking for
504				 * inactive bdi threads and killing them. Thus,
505				 * let's sleep for longer time, save energy and
506				 * be friendly for battery-driven devices.
507				 */
508				schedule_timeout(bdi_longest_inactive());
509			else
510				schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
511			try_to_freeze();
512			break;
513		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
514	}
 
 
515
516	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
517}
518
519/*
520 * Remove bdi from bdi_list, and ensure that it is no longer visible
 
 
 
521 */
522static void bdi_remove_from_list(struct backing_dev_info *bdi)
523{
524	spin_lock_bh(&bdi_lock);
525	list_del_rcu(&bdi->bdi_list);
526	spin_unlock_bh(&bdi_lock);
 
 
 
 
 
 
 
527
528	synchronize_rcu_expedited();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
529}
530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
531int bdi_register(struct backing_dev_info *bdi, struct device *parent,
532		const char *fmt, ...)
533{
534	va_list args;
535	struct device *dev;
536
537	if (bdi->dev)	/* The driver needs to use separate queues per device */
538		return 0;
539
540	va_start(args, fmt);
541	dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
542	va_end(args);
543	if (IS_ERR(dev))
544		return PTR_ERR(dev);
545
546	bdi->dev = dev;
547
548	/*
549	 * Just start the forker thread for our default backing_dev_info,
550	 * and add other bdi's to the list. They will get a thread created
551	 * on-demand when they need it.
552	 */
553	if (bdi_cap_flush_forker(bdi)) {
554		struct bdi_writeback *wb = &bdi->wb;
555
556		wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
557						dev_name(dev));
558		if (IS_ERR(wb->task))
559			return PTR_ERR(wb->task);
560	}
561
562	bdi_debug_register(bdi, dev_name(dev));
563	set_bit(BDI_registered, &bdi->state);
564
565	spin_lock_bh(&bdi_lock);
566	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
567	spin_unlock_bh(&bdi_lock);
568
569	trace_writeback_bdi_register(bdi);
570	return 0;
571}
572EXPORT_SYMBOL(bdi_register);
573
574int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
575{
576	return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
577}
578EXPORT_SYMBOL(bdi_register_dev);
579
580/*
581 * Remove bdi from the global list and shutdown any threads we have running
582 */
583static void bdi_wb_shutdown(struct backing_dev_info *bdi)
584{
585	if (!bdi_cap_writeback_dirty(bdi))
586		return;
587
588	/*
589	 * Make sure nobody finds us on the bdi_list anymore
590	 */
591	bdi_remove_from_list(bdi);
592
593	/*
594	 * If setup is pending, wait for that to complete first
595	 */
596	wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
597			TASK_UNINTERRUPTIBLE);
598
599	/*
600	 * Finally, kill the kernel thread. We don't need to be RCU
601	 * safe anymore, since the bdi is gone from visibility. Force
602	 * unfreeze of the thread before calling kthread_stop(), otherwise
603	 * it would never exet if it is currently stuck in the refrigerator.
604	 */
605	if (bdi->wb.task) {
606		thaw_process(bdi->wb.task);
607		kthread_stop(bdi->wb.task);
608	}
609}
 
610
611/*
612 * This bdi is going away now, make sure that no super_blocks point to it
613 */
614static void bdi_prune_sb(struct backing_dev_info *bdi)
615{
616	struct super_block *sb;
 
 
617
618	spin_lock(&sb_lock);
619	list_for_each_entry(sb, &super_blocks, s_list) {
620		if (sb->s_bdi == bdi)
621			sb->s_bdi = &default_backing_dev_info;
622	}
623	spin_unlock(&sb_lock);
624}
625
626void bdi_unregister(struct backing_dev_info *bdi)
627{
628	if (bdi->dev) {
629		bdi_set_min_ratio(bdi, 0);
630		trace_writeback_bdi_unregister(bdi);
631		bdi_prune_sb(bdi);
632		del_timer_sync(&bdi->wb.wakeup_timer);
633
634		if (!bdi_cap_flush_forker(bdi))
635			bdi_wb_shutdown(bdi);
636		bdi_debug_unregister(bdi);
637		device_unregister(bdi->dev);
638		bdi->dev = NULL;
639	}
640}
641EXPORT_SYMBOL(bdi_unregister);
642
643static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
644{
645	memset(wb, 0, sizeof(*wb));
646
647	wb->bdi = bdi;
648	wb->last_old_flush = jiffies;
649	INIT_LIST_HEAD(&wb->b_dirty);
650	INIT_LIST_HEAD(&wb->b_io);
651	INIT_LIST_HEAD(&wb->b_more_io);
652	spin_lock_init(&wb->list_lock);
653	setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
654}
655
656/*
657 * Initial write bandwidth: 100 MB/s
658 */
659#define INIT_BW		(100 << (20 - PAGE_SHIFT))
660
661int bdi_init(struct backing_dev_info *bdi)
662{
663	int i, err;
664
665	bdi->dev = NULL;
666
667	bdi->min_ratio = 0;
668	bdi->max_ratio = 100;
669	bdi->max_prop_frac = PROP_FRAC_BASE;
670	spin_lock_init(&bdi->wb_lock);
671	INIT_LIST_HEAD(&bdi->bdi_list);
672	INIT_LIST_HEAD(&bdi->work_list);
673
674	bdi_wb_init(&bdi->wb, bdi);
675
676	for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
677		err = percpu_counter_init(&bdi->bdi_stat[i], 0);
678		if (err)
679			goto err;
680	}
681
682	bdi->dirty_exceeded = 0;
683
684	bdi->bw_time_stamp = jiffies;
685	bdi->written_stamp = 0;
686
687	bdi->write_bandwidth = INIT_BW;
688	bdi->avg_write_bandwidth = INIT_BW;
689
690	err = prop_local_init_percpu(&bdi->completions);
691
692	if (err) {
693err:
694		while (i--)
695			percpu_counter_destroy(&bdi->bdi_stat[i]);
696	}
697
698	return err;
699}
700EXPORT_SYMBOL(bdi_init);
701
702void bdi_destroy(struct backing_dev_info *bdi)
703{
704	int i;
705
706	/*
707	 * Splice our entries to the default_backing_dev_info, if this
708	 * bdi disappears
709	 */
710	if (bdi_has_dirty_io(bdi)) {
711		struct bdi_writeback *dst = &default_backing_dev_info.wb;
712
713		bdi_lock_two(&bdi->wb, dst);
714		list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
715		list_splice(&bdi->wb.b_io, &dst->b_io);
716		list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
717		spin_unlock(&bdi->wb.list_lock);
718		spin_unlock(&dst->list_lock);
719	}
720
721	bdi_unregister(bdi);
722
723	for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
724		percpu_counter_destroy(&bdi->bdi_stat[i]);
725
726	prop_local_destroy_percpu(&bdi->completions);
727}
728EXPORT_SYMBOL(bdi_destroy);
729
730/*
731 * For use from filesystems to quickly init and register a bdi associated
732 * with dirty writeback
733 */
734int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
735			   unsigned int cap)
736{
737	char tmp[32];
738	int err;
739
740	bdi->name = name;
741	bdi->capabilities = cap;
742	err = bdi_init(bdi);
743	if (err)
744		return err;
745
746	sprintf(tmp, "%.28s%s", name, "-%d");
747	err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
748	if (err) {
749		bdi_destroy(bdi);
750		return err;
751	}
752
753	return 0;
754}
755EXPORT_SYMBOL(bdi_setup_and_register);
756
757static wait_queue_head_t congestion_wqh[2] = {
758		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
759		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
760	};
761static atomic_t nr_bdi_congested[2];
762
763void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
764{
765	enum bdi_state bit;
766	wait_queue_head_t *wqh = &congestion_wqh[sync];
 
767
768	bit = sync ? BDI_sync_congested : BDI_async_congested;
769	if (test_and_clear_bit(bit, &bdi->state))
770		atomic_dec(&nr_bdi_congested[sync]);
771	smp_mb__after_clear_bit();
772	if (waitqueue_active(wqh))
773		wake_up(wqh);
774}
775EXPORT_SYMBOL(clear_bdi_congested);
776
777void set_bdi_congested(struct backing_dev_info *bdi, int sync)
778{
779	enum bdi_state bit;
780
781	bit = sync ? BDI_sync_congested : BDI_async_congested;
782	if (!test_and_set_bit(bit, &bdi->state))
783		atomic_inc(&nr_bdi_congested[sync]);
784}
785EXPORT_SYMBOL(set_bdi_congested);
786
787/**
788 * congestion_wait - wait for a backing_dev to become uncongested
789 * @sync: SYNC or ASYNC IO
790 * @timeout: timeout in jiffies
791 *
792 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
793 * write congestion.  If no backing_devs are congested then just wait for the
794 * next write to be completed.
795 */
796long congestion_wait(int sync, long timeout)
797{
798	long ret;
799	unsigned long start = jiffies;
800	DEFINE_WAIT(wait);
801	wait_queue_head_t *wqh = &congestion_wqh[sync];
802
803	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
804	ret = io_schedule_timeout(timeout);
805	finish_wait(wqh, &wait);
806
807	trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
808					jiffies_to_usecs(jiffies - start));
809
810	return ret;
811}
812EXPORT_SYMBOL(congestion_wait);
813
814/**
815 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
816 * @zone: A zone to check if it is heavily congested
817 * @sync: SYNC or ASYNC IO
818 * @timeout: timeout in jiffies
819 *
820 * In the event of a congested backing_dev (any backing_dev) and the given
821 * @zone has experienced recent congestion, this waits for up to @timeout
822 * jiffies for either a BDI to exit congestion of the given @sync queue
823 * or a write to complete.
824 *
825 * In the absence of zone congestion, cond_resched() is called to yield
826 * the processor if necessary but otherwise does not sleep.
827 *
828 * The return value is 0 if the sleep is for the full timeout. Otherwise,
829 * it is the number of jiffies that were still remaining when the function
830 * returned. return_value == timeout implies the function did not sleep.
831 */
832long wait_iff_congested(struct zone *zone, int sync, long timeout)
833{
834	long ret;
835	unsigned long start = jiffies;
836	DEFINE_WAIT(wait);
837	wait_queue_head_t *wqh = &congestion_wqh[sync];
838
839	/*
840	 * If there is no congestion, or heavy congestion is not being
841	 * encountered in the current zone, yield if necessary instead
842	 * of sleeping on the congestion queue
843	 */
844	if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
845			!zone_is_reclaim_congested(zone)) {
846		cond_resched();
847
848		/* In case we scheduled, work out time remaining */
849		ret = timeout - (jiffies - start);
850		if (ret < 0)
851			ret = 0;
852
853		goto out;
854	}
855
856	/* Sleep until uncongested or a write happens */
857	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
858	ret = io_schedule_timeout(timeout);
859	finish_wait(wqh, &wait);
860
861out:
862	trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
863					jiffies_to_usecs(jiffies - start));
864
865	return ret;
866}
867EXPORT_SYMBOL(wait_iff_congested);
v4.10.11
   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 noop_backing_dev_info = {
  18	.name		= "noop",
  19	.capabilities	= BDI_CAP_NO_ACCT_AND_WRITEBACK,
  20};
  21EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23static struct class *bdi_class;
  24
  25/*
  26 * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
 
  27 * locking.
  28 */
  29DEFINE_SPINLOCK(bdi_lock);
  30LIST_HEAD(bdi_list);
 
 
 
 
  31
  32/* bdi_wq serves all asynchronous writeback tasks */
  33struct workqueue_struct *bdi_wq;
 
 
 
 
 
 
 
 
 
 
 
  34
  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;
  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
 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
 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);
 
 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)						\
 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);	\
 169}									\
 170static DEVICE_ATTR_RW(name);
 171
 172BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 173
 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);
 
 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}
 191BDI_SHOW(min_ratio, bdi->min_ratio)
 192
 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);
 
 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}
 210BDI_SHOW(max_ratio, bdi->max_ratio)
 211
 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}
 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,
 229};
 230ATTRIBUTE_GROUPS(bdi_dev);
 231
 232static __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	return 0;
 241}
 242postcore_initcall(bdi_class_init);
 243
 244static int __init default_bdi_init(void)
 245{
 246	int err;
 247
 248	bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 249					      WQ_UNBOUND | WQ_SYSFS, 0);
 250	if (!bdi_wq)
 251		return -ENOMEM;
 
 252
 
 
 
 253	err = bdi_init(&noop_backing_dev_info);
 254
 255	return err;
 256}
 257subsys_initcall(default_bdi_init);
 258
 259/*
 260 * This function is used when the first inode for this wb is marked dirty. It
 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.
 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().
 272 */
 273void wb_wakeup_delayed(struct bdi_writeback *wb)
 274{
 275	unsigned long timeout;
 276
 277	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 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);
 282}
 283
 284/*
 285 * Initial write bandwidth: 100 MB/s
 286 */
 287#define INIT_BW		(100 << (20 - PAGE_SHIFT))
 288
 289static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 290		   int blkcg_id, gfp_t gfp)
 291{
 292	int i, err;
 293
 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);
 301	INIT_LIST_HEAD(&wb->b_dirty_time);
 302	spin_lock_init(&wb->list_lock);
 303
 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
 310	spin_lock_init(&wb->work_lock);
 311	INIT_LIST_HEAD(&wb->work_list);
 312	INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 313	wb->dirty_sleep = jiffies;
 314
 315	wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 316	if (!wb->congested)
 317		return -ENOMEM;
 318
 319	err = fprop_local_init_percpu(&wb->completions, gfp);
 320	if (err)
 321		goto out_put_cong;
 322
 323	for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 324		err = percpu_counter_init(&wb->stat[i], 0, gfp);
 325		if (err)
 326			goto out_destroy_stat;
 327	}
 328
 329	return 0;
 330
 331out_destroy_stat:
 332	while (i--)
 333		percpu_counter_destroy(&wb->stat[i]);
 334	fprop_local_destroy_percpu(&wb->completions);
 335out_put_cong:
 336	wb_congested_put(wb->congested);
 337	return err;
 338}
 339
 340/*
 341 * Remove bdi from the global list and shutdown any threads we have running
 342 */
 343static void wb_shutdown(struct bdi_writeback *wb)
 344{
 345	/* Make sure nobody queues further work */
 346	spin_lock_bh(&wb->work_lock);
 347	if (!test_and_clear_bit(WB_registered, &wb->state)) {
 348		spin_unlock_bh(&wb->work_lock);
 349		return;
 350	}
 351	spin_unlock_bh(&wb->work_lock);
 352
 353	/*
 354	 * Drain work list and shutdown the delayed_work.  !WB_registered
 355	 * tells wb_workfn() that @wb is dying and its work_list needs to
 356	 * be drained no matter what.
 357	 */
 358	mod_delayed_work(bdi_wq, &wb->dwork, 0);
 359	flush_delayed_work(&wb->dwork);
 360	WARN_ON(!list_empty(&wb->work_list));
 361}
 362
 363static void wb_exit(struct bdi_writeback *wb)
 364{
 365	int i;
 366
 367	WARN_ON(delayed_work_pending(&wb->dwork));
 368
 369	for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 370		percpu_counter_destroy(&wb->stat[i]);
 371
 372	fprop_local_destroy_percpu(&wb->completions);
 373	wb_congested_put(wb->congested);
 374}
 375
 376#ifdef CONFIG_CGROUP_WRITEBACK
 377
 378#include <linux/memcontrol.h>
 379
 380/*
 381 * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 382 * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 383 * protected.  cgwb_release_wait is used to wait for the completion of cgwb
 384 * releases from bdi destruction path.
 385 */
 386static DEFINE_SPINLOCK(cgwb_lock);
 387static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
 388
 389/**
 390 * wb_congested_get_create - get or create a wb_congested
 391 * @bdi: associated bdi
 392 * @blkcg_id: ID of the associated blkcg
 393 * @gfp: allocation mask
 394 *
 395 * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 396 * The returned wb_congested has its reference count incremented.  Returns
 397 * NULL on failure.
 398 */
 399struct bdi_writeback_congested *
 400wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 401{
 402	struct bdi_writeback_congested *new_congested = NULL, *congested;
 403	struct rb_node **node, *parent;
 404	unsigned long flags;
 405retry:
 406	spin_lock_irqsave(&cgwb_lock, flags);
 407
 408	node = &bdi->cgwb_congested_tree.rb_node;
 409	parent = NULL;
 410
 411	while (*node != NULL) {
 412		parent = *node;
 413		congested = container_of(parent, struct bdi_writeback_congested,
 414					 rb_node);
 415		if (congested->blkcg_id < blkcg_id)
 416			node = &parent->rb_left;
 417		else if (congested->blkcg_id > blkcg_id)
 418			node = &parent->rb_right;
 419		else
 420			goto found;
 421	}
 422
 423	if (new_congested) {
 424		/* !found and storage for new one already allocated, insert */
 425		congested = new_congested;
 426		new_congested = NULL;
 427		rb_link_node(&congested->rb_node, parent, node);
 428		rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 429		goto found;
 
 
 
 
 
 430	}
 431
 432	spin_unlock_irqrestore(&cgwb_lock, flags);
 433
 434	/* allocate storage for new one and retry */
 435	new_congested = kzalloc(sizeof(*new_congested), gfp);
 436	if (!new_congested)
 437		return NULL;
 438
 439	atomic_set(&new_congested->refcnt, 0);
 440	new_congested->bdi = bdi;
 441	new_congested->blkcg_id = blkcg_id;
 442	goto retry;
 443
 444found:
 445	atomic_inc(&congested->refcnt);
 446	spin_unlock_irqrestore(&cgwb_lock, flags);
 447	kfree(new_congested);
 448	return congested;
 449}
 450
 451/**
 452 * wb_congested_put - put a wb_congested
 453 * @congested: wb_congested to put
 
 
 
 454 *
 455 * Put @congested and destroy it if the refcnt reaches zero.
 
 
 456 */
 457void wb_congested_put(struct bdi_writeback_congested *congested)
 458{
 459	unsigned long flags;
 460
 461	local_irq_save(flags);
 462	if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 463		local_irq_restore(flags);
 464		return;
 465	}
 466
 467	/* bdi might already have been destroyed leaving @congested unlinked */
 468	if (congested->bdi) {
 469		rb_erase(&congested->rb_node,
 470			 &congested->bdi->cgwb_congested_tree);
 471		congested->bdi = NULL;
 472	}
 473
 474	spin_unlock_irqrestore(&cgwb_lock, flags);
 475	kfree(congested);
 476}
 477
 478static void cgwb_release_workfn(struct work_struct *work)
 
 
 
 
 479{
 480	struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 481						release_work);
 482	struct backing_dev_info *bdi = wb->bdi;
 483
 484	spin_lock_irq(&cgwb_lock);
 485	list_del_rcu(&wb->bdi_node);
 486	spin_unlock_irq(&cgwb_lock);
 487
 488	wb_shutdown(wb);
 489
 490	css_put(wb->memcg_css);
 491	css_put(wb->blkcg_css);
 492
 493	fprop_local_destroy_percpu(&wb->memcg_completions);
 494	percpu_ref_exit(&wb->refcnt);
 495	wb_exit(wb);
 496	kfree_rcu(wb, rcu);
 497
 498	if (atomic_dec_and_test(&bdi->usage_cnt))
 499		wake_up_all(&cgwb_release_wait);
 500}
 501
 502static void cgwb_release(struct percpu_ref *refcnt)
 
 
 
 
 503{
 504	struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 505						refcnt);
 506	schedule_work(&wb->release_work);
 507}
 508
 509static void cgwb_kill(struct bdi_writeback *wb)
 510{
 511	lockdep_assert_held(&cgwb_lock);
 512
 513	WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 514	list_del(&wb->memcg_node);
 515	list_del(&wb->blkcg_node);
 516	percpu_ref_kill(&wb->refcnt);
 517}
 518
 519static int cgwb_create(struct backing_dev_info *bdi,
 520		       struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 521{
 522	struct mem_cgroup *memcg;
 523	struct cgroup_subsys_state *blkcg_css;
 524	struct blkcg *blkcg;
 525	struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 526	struct bdi_writeback *wb;
 527	unsigned long flags;
 528	int ret = 0;
 529
 530	memcg = mem_cgroup_from_css(memcg_css);
 531	blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 532	blkcg = css_to_blkcg(blkcg_css);
 533	memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 534	blkcg_cgwb_list = &blkcg->cgwb_list;
 535
 536	/* look up again under lock and discard on blkcg mismatch */
 537	spin_lock_irqsave(&cgwb_lock, flags);
 538	wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 539	if (wb && wb->blkcg_css != blkcg_css) {
 540		cgwb_kill(wb);
 541		wb = NULL;
 542	}
 543	spin_unlock_irqrestore(&cgwb_lock, flags);
 544	if (wb)
 545		goto out_put;
 546
 547	/* need to create a new one */
 548	wb = kmalloc(sizeof(*wb), gfp);
 549	if (!wb)
 550		return -ENOMEM;
 551
 552	ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 553	if (ret)
 554		goto err_free;
 555
 556	ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 557	if (ret)
 558		goto err_wb_exit;
 559
 560	ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 561	if (ret)
 562		goto err_ref_exit;
 563
 564	wb->memcg_css = memcg_css;
 565	wb->blkcg_css = blkcg_css;
 566	INIT_WORK(&wb->release_work, cgwb_release_workfn);
 567	set_bit(WB_registered, &wb->state);
 568
 569	/*
 570	 * The root wb determines the registered state of the whole bdi and
 571	 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 572	 * whether they're still online.  Don't link @wb if any is dead.
 573	 * See wb_memcg_offline() and wb_blkcg_offline().
 574	 */
 575	ret = -ENODEV;
 576	spin_lock_irqsave(&cgwb_lock, flags);
 577	if (test_bit(WB_registered, &bdi->wb.state) &&
 578	    blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 579		/* we might have raced another instance of this function */
 580		ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 581		if (!ret) {
 582			atomic_inc(&bdi->usage_cnt);
 583			list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 584			list_add(&wb->memcg_node, memcg_cgwb_list);
 585			list_add(&wb->blkcg_node, blkcg_cgwb_list);
 586			css_get(memcg_css);
 587			css_get(blkcg_css);
 
 
 
 
 
 588		}
 589	}
 590	spin_unlock_irqrestore(&cgwb_lock, flags);
 591	if (ret) {
 592		if (ret == -EEXIST)
 593			ret = 0;
 594		goto err_fprop_exit;
 595	}
 596	goto out_put;
 597
 598err_fprop_exit:
 599	fprop_local_destroy_percpu(&wb->memcg_completions);
 600err_ref_exit:
 601	percpu_ref_exit(&wb->refcnt);
 602err_wb_exit:
 603	wb_exit(wb);
 604err_free:
 605	kfree(wb);
 606out_put:
 607	css_put(blkcg_css);
 608	return ret;
 609}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 610
 611/**
 612 * wb_get_create - get wb for a given memcg, create if necessary
 613 * @bdi: target bdi
 614 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 615 * @gfp: allocation mask to use
 616 *
 617 * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 618 * create one.  The returned wb has its refcount incremented.
 619 *
 620 * This function uses css_get() on @memcg_css and thus expects its refcnt
 621 * to be positive on invocation.  IOW, rcu_read_lock() protection on
 622 * @memcg_css isn't enough.  try_get it before calling this function.
 623 *
 624 * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 625 * memcg on the default hierarchy, memcg association is guaranteed to be
 626 * more specific (equal or descendant to the associated blkcg) and thus can
 627 * identify both the memcg and blkcg associations.
 628 *
 629 * Because the blkcg associated with a memcg may change as blkcg is enabled
 630 * and disabled closer to root in the hierarchy, each wb keeps track of
 631 * both the memcg and blkcg associated with it and verifies the blkcg on
 632 * each lookup.  On mismatch, the existing wb is discarded and a new one is
 633 * created.
 634 */
 635struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 636				    struct cgroup_subsys_state *memcg_css,
 637				    gfp_t gfp)
 638{
 639	struct bdi_writeback *wb;
 640
 641	might_sleep_if(gfpflags_allow_blocking(gfp));
 642
 643	if (!memcg_css->parent)
 644		return &bdi->wb;
 645
 646	do {
 647		rcu_read_lock();
 648		wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 649		if (wb) {
 650			struct cgroup_subsys_state *blkcg_css;
 651
 652			/* see whether the blkcg association has changed */
 653			blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 654						     &io_cgrp_subsys);
 655			if (unlikely(wb->blkcg_css != blkcg_css ||
 656				     !wb_tryget(wb)))
 657				wb = NULL;
 658			css_put(blkcg_css);
 
 
 
 659		}
 660		rcu_read_unlock();
 661	} while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 662
 663	return wb;
 664}
 665
 666static int cgwb_bdi_init(struct backing_dev_info *bdi)
 667{
 668	int ret;
 669
 670	INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 671	bdi->cgwb_congested_tree = RB_ROOT;
 672	atomic_set(&bdi->usage_cnt, 1);
 673
 674	ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 675	if (!ret) {
 676		bdi->wb.memcg_css = &root_mem_cgroup->css;
 677		bdi->wb.blkcg_css = blkcg_root_css;
 678	}
 679	return ret;
 680}
 681
 682static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 683{
 684	struct radix_tree_iter iter;
 685	struct rb_node *rbn;
 686	void **slot;
 687
 688	WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 689
 690	spin_lock_irq(&cgwb_lock);
 691
 692	radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 693		cgwb_kill(*slot);
 694
 695	while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 696		struct bdi_writeback_congested *congested =
 697			rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 698
 699		rb_erase(rbn, &bdi->cgwb_congested_tree);
 700		congested->bdi = NULL;	/* mark @congested unlinked */
 701	}
 702
 703	spin_unlock_irq(&cgwb_lock);
 704
 705	/*
 706	 * All cgwb's and their congested states must be shutdown and
 707	 * released before returning.  Drain the usage counter to wait for
 708	 * all cgwb's and cgwb_congested's ever created on @bdi.
 709	 */
 710	atomic_dec(&bdi->usage_cnt);
 711	wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
 712}
 713
 714/**
 715 * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 716 * @memcg: memcg being offlined
 717 *
 718 * Also prevents creation of any new wb's associated with @memcg.
 719 */
 720void wb_memcg_offline(struct mem_cgroup *memcg)
 721{
 722	LIST_HEAD(to_destroy);
 723	struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 724	struct bdi_writeback *wb, *next;
 725
 726	spin_lock_irq(&cgwb_lock);
 727	list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 728		cgwb_kill(wb);
 729	memcg_cgwb_list->next = NULL;	/* prevent new wb's */
 730	spin_unlock_irq(&cgwb_lock);
 731}
 732
 733/**
 734 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 735 * @blkcg: blkcg being offlined
 736 *
 737 * Also prevents creation of any new wb's associated with @blkcg.
 738 */
 739void wb_blkcg_offline(struct blkcg *blkcg)
 740{
 741	LIST_HEAD(to_destroy);
 742	struct bdi_writeback *wb, *next;
 743
 744	spin_lock_irq(&cgwb_lock);
 745	list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 746		cgwb_kill(wb);
 747	blkcg->cgwb_list.next = NULL;	/* prevent new wb's */
 748	spin_unlock_irq(&cgwb_lock);
 749}
 750
 751#else	/* CONFIG_CGROUP_WRITEBACK */
 752
 753static int cgwb_bdi_init(struct backing_dev_info *bdi)
 754{
 755	int err;
 756
 757	bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 758	if (!bdi->wb_congested)
 759		return -ENOMEM;
 760
 761	atomic_set(&bdi->wb_congested->refcnt, 1);
 762
 763	err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 764	if (err) {
 765		wb_congested_put(bdi->wb_congested);
 766		return err;
 767	}
 768	return 0;
 769}
 770
 771static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 772{
 773	wb_congested_put(bdi->wb_congested);
 774}
 775
 776#endif	/* CONFIG_CGROUP_WRITEBACK */
 777
 778int bdi_init(struct backing_dev_info *bdi)
 779{
 780	int ret;
 781
 782	bdi->dev = NULL;
 783
 784	bdi->min_ratio = 0;
 785	bdi->max_ratio = 100;
 786	bdi->max_prop_frac = FPROP_FRAC_BASE;
 787	INIT_LIST_HEAD(&bdi->bdi_list);
 788	INIT_LIST_HEAD(&bdi->wb_list);
 789	init_waitqueue_head(&bdi->wb_waitq);
 790
 791	ret = cgwb_bdi_init(bdi);
 792
 793	list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 794
 795	return ret;
 796}
 797EXPORT_SYMBOL(bdi_init);
 798
 799int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 800		const char *fmt, ...)
 801{
 802	va_list args;
 803	struct device *dev;
 804
 805	if (bdi->dev)	/* The driver needs to use separate queues per device */
 806		return 0;
 807
 808	va_start(args, fmt);
 809	dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 810	va_end(args);
 811	if (IS_ERR(dev))
 812		return PTR_ERR(dev);
 813
 814	bdi->dev = dev;
 815
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 816	bdi_debug_register(bdi, dev_name(dev));
 817	set_bit(WB_registered, &bdi->wb.state);
 818
 819	spin_lock_bh(&bdi_lock);
 820	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 821	spin_unlock_bh(&bdi_lock);
 822
 823	trace_writeback_bdi_register(bdi);
 824	return 0;
 825}
 826EXPORT_SYMBOL(bdi_register);
 827
 828int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 829{
 830	return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 831}
 832EXPORT_SYMBOL(bdi_register_dev);
 833
 834int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 
 
 
 835{
 836	int rc;
 
 
 
 
 
 
 
 
 
 
 
 
 837
 838	rc = bdi_register(bdi, NULL, "%u:%u", MAJOR(owner->devt),
 839			MINOR(owner->devt));
 840	if (rc)
 841		return rc;
 842	bdi->owner = owner;
 843	get_device(owner);
 844	return 0;
 
 
 
 845}
 846EXPORT_SYMBOL(bdi_register_owner);
 847
 848/*
 849 * Remove bdi from bdi_list, and ensure that it is no longer visible
 850 */
 851static void bdi_remove_from_list(struct backing_dev_info *bdi)
 852{
 853	spin_lock_bh(&bdi_lock);
 854	list_del_rcu(&bdi->bdi_list);
 855	spin_unlock_bh(&bdi_lock);
 856
 857	synchronize_rcu_expedited();
 
 
 
 
 
 858}
 859
 860void bdi_unregister(struct backing_dev_info *bdi)
 861{
 862	/* make sure nobody finds us on the bdi_list anymore */
 863	bdi_remove_from_list(bdi);
 864	wb_shutdown(&bdi->wb);
 865	cgwb_bdi_destroy(bdi);
 
 866
 867	if (bdi->dev) {
 
 868		bdi_debug_unregister(bdi);
 869		device_unregister(bdi->dev);
 870		bdi->dev = NULL;
 871	}
 
 
 872
 873	if (bdi->owner) {
 874		put_device(bdi->owner);
 875		bdi->owner = NULL;
 876	}
 
 
 
 
 
 
 
 877}
 878
 879void bdi_exit(struct backing_dev_info *bdi)
 
 
 
 
 
 880{
 881	WARN_ON_ONCE(bdi->dev);
 882	wb_exit(&bdi->wb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 883}
 
 884
 885void bdi_destroy(struct backing_dev_info *bdi)
 886{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 887	bdi_unregister(bdi);
 888	bdi_exit(bdi);
 
 
 
 
 889}
 890EXPORT_SYMBOL(bdi_destroy);
 891
 892/*
 893 * For use from filesystems to quickly init and register a bdi associated
 894 * with dirty writeback
 895 */
 896int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
 
 897{
 
 898	int err;
 899
 900	bdi->name = name;
 901	bdi->capabilities = 0;
 902	err = bdi_init(bdi);
 903	if (err)
 904		return err;
 905
 906	err = bdi_register(bdi, NULL, "%.28s-%ld", name,
 907			   atomic_long_inc_return(&bdi_seq));
 908	if (err) {
 909		bdi_destroy(bdi);
 910		return err;
 911	}
 912
 913	return 0;
 914}
 915EXPORT_SYMBOL(bdi_setup_and_register);
 916
 917static wait_queue_head_t congestion_wqh[2] = {
 918		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 919		__WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 920	};
 921static atomic_t nr_wb_congested[2];
 922
 923void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 924{
 
 925	wait_queue_head_t *wqh = &congestion_wqh[sync];
 926	enum wb_congested_state bit;
 927
 928	bit = sync ? WB_sync_congested : WB_async_congested;
 929	if (test_and_clear_bit(bit, &congested->state))
 930		atomic_dec(&nr_wb_congested[sync]);
 931	smp_mb__after_atomic();
 932	if (waitqueue_active(wqh))
 933		wake_up(wqh);
 934}
 935EXPORT_SYMBOL(clear_wb_congested);
 936
 937void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 938{
 939	enum wb_congested_state bit;
 940
 941	bit = sync ? WB_sync_congested : WB_async_congested;
 942	if (!test_and_set_bit(bit, &congested->state))
 943		atomic_inc(&nr_wb_congested[sync]);
 944}
 945EXPORT_SYMBOL(set_wb_congested);
 946
 947/**
 948 * congestion_wait - wait for a backing_dev to become uncongested
 949 * @sync: SYNC or ASYNC IO
 950 * @timeout: timeout in jiffies
 951 *
 952 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 953 * write congestion.  If no backing_devs are congested then just wait for the
 954 * next write to be completed.
 955 */
 956long congestion_wait(int sync, long timeout)
 957{
 958	long ret;
 959	unsigned long start = jiffies;
 960	DEFINE_WAIT(wait);
 961	wait_queue_head_t *wqh = &congestion_wqh[sync];
 962
 963	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 964	ret = io_schedule_timeout(timeout);
 965	finish_wait(wqh, &wait);
 966
 967	trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 968					jiffies_to_usecs(jiffies - start));
 969
 970	return ret;
 971}
 972EXPORT_SYMBOL(congestion_wait);
 973
 974/**
 975 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
 976 * @pgdat: A pgdat to check if it is heavily congested
 977 * @sync: SYNC or ASYNC IO
 978 * @timeout: timeout in jiffies
 979 *
 980 * In the event of a congested backing_dev (any backing_dev) and the given
 981 * @pgdat has experienced recent congestion, this waits for up to @timeout
 982 * jiffies for either a BDI to exit congestion of the given @sync queue
 983 * or a write to complete.
 984 *
 985 * In the absence of pgdat congestion, cond_resched() is called to yield
 986 * the processor if necessary but otherwise does not sleep.
 987 *
 988 * The return value is 0 if the sleep is for the full timeout. Otherwise,
 989 * it is the number of jiffies that were still remaining when the function
 990 * returned. return_value == timeout implies the function did not sleep.
 991 */
 992long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
 993{
 994	long ret;
 995	unsigned long start = jiffies;
 996	DEFINE_WAIT(wait);
 997	wait_queue_head_t *wqh = &congestion_wqh[sync];
 998
 999	/*
1000	 * If there is no congestion, or heavy congestion is not being
1001	 * encountered in the current pgdat, yield if necessary instead
1002	 * of sleeping on the congestion queue
1003	 */
1004	if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1005	    !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1006		cond_resched();
1007
1008		/* In case we scheduled, work out time remaining */
1009		ret = timeout - (jiffies - start);
1010		if (ret < 0)
1011			ret = 0;
1012
1013		goto out;
1014	}
1015
1016	/* Sleep until uncongested or a write happens */
1017	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1018	ret = io_schedule_timeout(timeout);
1019	finish_wait(wqh, &wait);
1020
1021out:
1022	trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1023					jiffies_to_usecs(jiffies - start));
1024
1025	return ret;
1026}
1027EXPORT_SYMBOL(wait_iff_congested);
1028
1029int pdflush_proc_obsolete(struct ctl_table *table, int write,
1030			void __user *buffer, size_t *lenp, loff_t *ppos)
1031{
1032	char kbuf[] = "0\n";
1033
1034	if (*ppos || *lenp < sizeof(kbuf)) {
1035		*lenp = 0;
1036		return 0;
1037	}
1038
1039	if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1040		return -EFAULT;
1041	pr_warn_once("%s exported in /proc is scheduled for removal\n",
1042		     table->procname);
1043
1044	*lenp = 2;
1045	*ppos += *lenp;
1046	return 2;
1047}