blk-cgroup.c revision 1adaf3dde37a8b9b59ea59c5f58fed7761178383
1/*
2 * Common Block IO controller cgroup interface
3 *
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 *		      Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * 	              Nauman Rafique <nauman@google.com>
12 */
13#include <linux/ioprio.h>
14#include <linux/seq_file.h>
15#include <linux/kdev_t.h>
16#include <linux/module.h>
17#include <linux/err.h>
18#include <linux/blkdev.h>
19#include <linux/slab.h>
20#include <linux/genhd.h>
21#include <linux/delay.h>
22#include "blk-cgroup.h"
23#include "blk.h"
24
25#define MAX_KEY_LEN 100
26
27static DEFINE_SPINLOCK(blkio_list_lock);
28static LIST_HEAD(blkio_list);
29
30static DEFINE_MUTEX(all_q_mutex);
31static LIST_HEAD(all_q_list);
32
33struct blkio_cgroup blkio_root_cgroup = { .weight = 2*BLKIO_WEIGHT_DEFAULT };
34EXPORT_SYMBOL_GPL(blkio_root_cgroup);
35
36static struct blkio_policy_type *blkio_policy[BLKIO_NR_POLICIES];
37
38static struct cgroup_subsys_state *blkiocg_create(struct cgroup_subsys *,
39						  struct cgroup *);
40static int blkiocg_can_attach(struct cgroup_subsys *, struct cgroup *,
41			      struct cgroup_taskset *);
42static void blkiocg_attach(struct cgroup_subsys *, struct cgroup *,
43			   struct cgroup_taskset *);
44static int blkiocg_pre_destroy(struct cgroup_subsys *, struct cgroup *);
45static void blkiocg_destroy(struct cgroup_subsys *, struct cgroup *);
46static int blkiocg_populate(struct cgroup_subsys *, struct cgroup *);
47
48/* for encoding cft->private value on file */
49#define BLKIOFILE_PRIVATE(x, val)	(((x) << 16) | (val))
50/* What policy owns the file, proportional or throttle */
51#define BLKIOFILE_POLICY(val)		(((val) >> 16) & 0xffff)
52#define BLKIOFILE_ATTR(val)		((val) & 0xffff)
53
54struct cgroup_subsys blkio_subsys = {
55	.name = "blkio",
56	.create = blkiocg_create,
57	.can_attach = blkiocg_can_attach,
58	.attach = blkiocg_attach,
59	.pre_destroy = blkiocg_pre_destroy,
60	.destroy = blkiocg_destroy,
61	.populate = blkiocg_populate,
62	.subsys_id = blkio_subsys_id,
63	.module = THIS_MODULE,
64};
65EXPORT_SYMBOL_GPL(blkio_subsys);
66
67struct blkio_cgroup *cgroup_to_blkio_cgroup(struct cgroup *cgroup)
68{
69	return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
70			    struct blkio_cgroup, css);
71}
72EXPORT_SYMBOL_GPL(cgroup_to_blkio_cgroup);
73
74struct blkio_cgroup *task_blkio_cgroup(struct task_struct *tsk)
75{
76	return container_of(task_subsys_state(tsk, blkio_subsys_id),
77			    struct blkio_cgroup, css);
78}
79EXPORT_SYMBOL_GPL(task_blkio_cgroup);
80
81static inline void
82blkio_update_group_weight(struct blkio_group *blkg, unsigned int weight)
83{
84	struct blkio_policy_type *blkiop;
85
86	list_for_each_entry(blkiop, &blkio_list, list) {
87		/* If this policy does not own the blkg, do not send updates */
88		if (blkiop->plid != blkg->plid)
89			continue;
90		if (blkiop->ops.blkio_update_group_weight_fn)
91			blkiop->ops.blkio_update_group_weight_fn(blkg->q,
92							blkg, weight);
93	}
94}
95
96static inline void blkio_update_group_bps(struct blkio_group *blkg, u64 bps,
97				int fileid)
98{
99	struct blkio_policy_type *blkiop;
100
101	list_for_each_entry(blkiop, &blkio_list, list) {
102
103		/* If this policy does not own the blkg, do not send updates */
104		if (blkiop->plid != blkg->plid)
105			continue;
106
107		if (fileid == BLKIO_THROTL_read_bps_device
108		    && blkiop->ops.blkio_update_group_read_bps_fn)
109			blkiop->ops.blkio_update_group_read_bps_fn(blkg->q,
110								blkg, bps);
111
112		if (fileid == BLKIO_THROTL_write_bps_device
113		    && blkiop->ops.blkio_update_group_write_bps_fn)
114			blkiop->ops.blkio_update_group_write_bps_fn(blkg->q,
115								blkg, bps);
116	}
117}
118
119static inline void blkio_update_group_iops(struct blkio_group *blkg,
120			unsigned int iops, int fileid)
121{
122	struct blkio_policy_type *blkiop;
123
124	list_for_each_entry(blkiop, &blkio_list, list) {
125
126		/* If this policy does not own the blkg, do not send updates */
127		if (blkiop->plid != blkg->plid)
128			continue;
129
130		if (fileid == BLKIO_THROTL_read_iops_device
131		    && blkiop->ops.blkio_update_group_read_iops_fn)
132			blkiop->ops.blkio_update_group_read_iops_fn(blkg->q,
133								blkg, iops);
134
135		if (fileid == BLKIO_THROTL_write_iops_device
136		    && blkiop->ops.blkio_update_group_write_iops_fn)
137			blkiop->ops.blkio_update_group_write_iops_fn(blkg->q,
138								blkg,iops);
139	}
140}
141
142/*
143 * Add to the appropriate stat variable depending on the request type.
144 * This should be called with the blkg->stats_lock held.
145 */
146static void blkio_add_stat(uint64_t *stat, uint64_t add, bool direction,
147				bool sync)
148{
149	if (direction)
150		stat[BLKIO_STAT_WRITE] += add;
151	else
152		stat[BLKIO_STAT_READ] += add;
153	if (sync)
154		stat[BLKIO_STAT_SYNC] += add;
155	else
156		stat[BLKIO_STAT_ASYNC] += add;
157}
158
159/*
160 * Decrements the appropriate stat variable if non-zero depending on the
161 * request type. Panics on value being zero.
162 * This should be called with the blkg->stats_lock held.
163 */
164static void blkio_check_and_dec_stat(uint64_t *stat, bool direction, bool sync)
165{
166	if (direction) {
167		BUG_ON(stat[BLKIO_STAT_WRITE] == 0);
168		stat[BLKIO_STAT_WRITE]--;
169	} else {
170		BUG_ON(stat[BLKIO_STAT_READ] == 0);
171		stat[BLKIO_STAT_READ]--;
172	}
173	if (sync) {
174		BUG_ON(stat[BLKIO_STAT_SYNC] == 0);
175		stat[BLKIO_STAT_SYNC]--;
176	} else {
177		BUG_ON(stat[BLKIO_STAT_ASYNC] == 0);
178		stat[BLKIO_STAT_ASYNC]--;
179	}
180}
181
182#ifdef CONFIG_DEBUG_BLK_CGROUP
183/* This should be called with the blkg->stats_lock held. */
184static void blkio_set_start_group_wait_time(struct blkio_group *blkg,
185						struct blkio_group *curr_blkg)
186{
187	if (blkio_blkg_waiting(&blkg->stats))
188		return;
189	if (blkg == curr_blkg)
190		return;
191	blkg->stats.start_group_wait_time = sched_clock();
192	blkio_mark_blkg_waiting(&blkg->stats);
193}
194
195/* This should be called with the blkg->stats_lock held. */
196static void blkio_update_group_wait_time(struct blkio_group_stats *stats)
197{
198	unsigned long long now;
199
200	if (!blkio_blkg_waiting(stats))
201		return;
202
203	now = sched_clock();
204	if (time_after64(now, stats->start_group_wait_time))
205		stats->group_wait_time += now - stats->start_group_wait_time;
206	blkio_clear_blkg_waiting(stats);
207}
208
209/* This should be called with the blkg->stats_lock held. */
210static void blkio_end_empty_time(struct blkio_group_stats *stats)
211{
212	unsigned long long now;
213
214	if (!blkio_blkg_empty(stats))
215		return;
216
217	now = sched_clock();
218	if (time_after64(now, stats->start_empty_time))
219		stats->empty_time += now - stats->start_empty_time;
220	blkio_clear_blkg_empty(stats);
221}
222
223void blkiocg_update_set_idle_time_stats(struct blkio_group *blkg)
224{
225	unsigned long flags;
226
227	spin_lock_irqsave(&blkg->stats_lock, flags);
228	BUG_ON(blkio_blkg_idling(&blkg->stats));
229	blkg->stats.start_idle_time = sched_clock();
230	blkio_mark_blkg_idling(&blkg->stats);
231	spin_unlock_irqrestore(&blkg->stats_lock, flags);
232}
233EXPORT_SYMBOL_GPL(blkiocg_update_set_idle_time_stats);
234
235void blkiocg_update_idle_time_stats(struct blkio_group *blkg)
236{
237	unsigned long flags;
238	unsigned long long now;
239	struct blkio_group_stats *stats;
240
241	spin_lock_irqsave(&blkg->stats_lock, flags);
242	stats = &blkg->stats;
243	if (blkio_blkg_idling(stats)) {
244		now = sched_clock();
245		if (time_after64(now, stats->start_idle_time))
246			stats->idle_time += now - stats->start_idle_time;
247		blkio_clear_blkg_idling(stats);
248	}
249	spin_unlock_irqrestore(&blkg->stats_lock, flags);
250}
251EXPORT_SYMBOL_GPL(blkiocg_update_idle_time_stats);
252
253void blkiocg_update_avg_queue_size_stats(struct blkio_group *blkg)
254{
255	unsigned long flags;
256	struct blkio_group_stats *stats;
257
258	spin_lock_irqsave(&blkg->stats_lock, flags);
259	stats = &blkg->stats;
260	stats->avg_queue_size_sum +=
261			stats->stat_arr[BLKIO_STAT_QUEUED][BLKIO_STAT_READ] +
262			stats->stat_arr[BLKIO_STAT_QUEUED][BLKIO_STAT_WRITE];
263	stats->avg_queue_size_samples++;
264	blkio_update_group_wait_time(stats);
265	spin_unlock_irqrestore(&blkg->stats_lock, flags);
266}
267EXPORT_SYMBOL_GPL(blkiocg_update_avg_queue_size_stats);
268
269void blkiocg_set_start_empty_time(struct blkio_group *blkg)
270{
271	unsigned long flags;
272	struct blkio_group_stats *stats;
273
274	spin_lock_irqsave(&blkg->stats_lock, flags);
275	stats = &blkg->stats;
276
277	if (stats->stat_arr[BLKIO_STAT_QUEUED][BLKIO_STAT_READ] ||
278			stats->stat_arr[BLKIO_STAT_QUEUED][BLKIO_STAT_WRITE]) {
279		spin_unlock_irqrestore(&blkg->stats_lock, flags);
280		return;
281	}
282
283	/*
284	 * group is already marked empty. This can happen if cfqq got new
285	 * request in parent group and moved to this group while being added
286	 * to service tree. Just ignore the event and move on.
287	 */
288	if(blkio_blkg_empty(stats)) {
289		spin_unlock_irqrestore(&blkg->stats_lock, flags);
290		return;
291	}
292
293	stats->start_empty_time = sched_clock();
294	blkio_mark_blkg_empty(stats);
295	spin_unlock_irqrestore(&blkg->stats_lock, flags);
296}
297EXPORT_SYMBOL_GPL(blkiocg_set_start_empty_time);
298
299void blkiocg_update_dequeue_stats(struct blkio_group *blkg,
300			unsigned long dequeue)
301{
302	blkg->stats.dequeue += dequeue;
303}
304EXPORT_SYMBOL_GPL(blkiocg_update_dequeue_stats);
305#else
306static inline void blkio_set_start_group_wait_time(struct blkio_group *blkg,
307					struct blkio_group *curr_blkg) {}
308static inline void blkio_end_empty_time(struct blkio_group_stats *stats) {}
309#endif
310
311void blkiocg_update_io_add_stats(struct blkio_group *blkg,
312			struct blkio_group *curr_blkg, bool direction,
313			bool sync)
314{
315	unsigned long flags;
316
317	spin_lock_irqsave(&blkg->stats_lock, flags);
318	blkio_add_stat(blkg->stats.stat_arr[BLKIO_STAT_QUEUED], 1, direction,
319			sync);
320	blkio_end_empty_time(&blkg->stats);
321	blkio_set_start_group_wait_time(blkg, curr_blkg);
322	spin_unlock_irqrestore(&blkg->stats_lock, flags);
323}
324EXPORT_SYMBOL_GPL(blkiocg_update_io_add_stats);
325
326void blkiocg_update_io_remove_stats(struct blkio_group *blkg,
327						bool direction, bool sync)
328{
329	unsigned long flags;
330
331	spin_lock_irqsave(&blkg->stats_lock, flags);
332	blkio_check_and_dec_stat(blkg->stats.stat_arr[BLKIO_STAT_QUEUED],
333					direction, sync);
334	spin_unlock_irqrestore(&blkg->stats_lock, flags);
335}
336EXPORT_SYMBOL_GPL(blkiocg_update_io_remove_stats);
337
338void blkiocg_update_timeslice_used(struct blkio_group *blkg, unsigned long time,
339				unsigned long unaccounted_time)
340{
341	unsigned long flags;
342
343	spin_lock_irqsave(&blkg->stats_lock, flags);
344	blkg->stats.time += time;
345#ifdef CONFIG_DEBUG_BLK_CGROUP
346	blkg->stats.unaccounted_time += unaccounted_time;
347#endif
348	spin_unlock_irqrestore(&blkg->stats_lock, flags);
349}
350EXPORT_SYMBOL_GPL(blkiocg_update_timeslice_used);
351
352/*
353 * should be called under rcu read lock or queue lock to make sure blkg pointer
354 * is valid.
355 */
356void blkiocg_update_dispatch_stats(struct blkio_group *blkg,
357				uint64_t bytes, bool direction, bool sync)
358{
359	struct blkio_group_stats_cpu *stats_cpu;
360	unsigned long flags;
361
362	/*
363	 * Disabling interrupts to provide mutual exclusion between two
364	 * writes on same cpu. It probably is not needed for 64bit. Not
365	 * optimizing that case yet.
366	 */
367	local_irq_save(flags);
368
369	stats_cpu = this_cpu_ptr(blkg->stats_cpu);
370
371	u64_stats_update_begin(&stats_cpu->syncp);
372	stats_cpu->sectors += bytes >> 9;
373	blkio_add_stat(stats_cpu->stat_arr_cpu[BLKIO_STAT_CPU_SERVICED],
374			1, direction, sync);
375	blkio_add_stat(stats_cpu->stat_arr_cpu[BLKIO_STAT_CPU_SERVICE_BYTES],
376			bytes, direction, sync);
377	u64_stats_update_end(&stats_cpu->syncp);
378	local_irq_restore(flags);
379}
380EXPORT_SYMBOL_GPL(blkiocg_update_dispatch_stats);
381
382void blkiocg_update_completion_stats(struct blkio_group *blkg,
383	uint64_t start_time, uint64_t io_start_time, bool direction, bool sync)
384{
385	struct blkio_group_stats *stats;
386	unsigned long flags;
387	unsigned long long now = sched_clock();
388
389	spin_lock_irqsave(&blkg->stats_lock, flags);
390	stats = &blkg->stats;
391	if (time_after64(now, io_start_time))
392		blkio_add_stat(stats->stat_arr[BLKIO_STAT_SERVICE_TIME],
393				now - io_start_time, direction, sync);
394	if (time_after64(io_start_time, start_time))
395		blkio_add_stat(stats->stat_arr[BLKIO_STAT_WAIT_TIME],
396				io_start_time - start_time, direction, sync);
397	spin_unlock_irqrestore(&blkg->stats_lock, flags);
398}
399EXPORT_SYMBOL_GPL(blkiocg_update_completion_stats);
400
401/*  Merged stats are per cpu.  */
402void blkiocg_update_io_merged_stats(struct blkio_group *blkg, bool direction,
403					bool sync)
404{
405	struct blkio_group_stats_cpu *stats_cpu;
406	unsigned long flags;
407
408	/*
409	 * Disabling interrupts to provide mutual exclusion between two
410	 * writes on same cpu. It probably is not needed for 64bit. Not
411	 * optimizing that case yet.
412	 */
413	local_irq_save(flags);
414
415	stats_cpu = this_cpu_ptr(blkg->stats_cpu);
416
417	u64_stats_update_begin(&stats_cpu->syncp);
418	blkio_add_stat(stats_cpu->stat_arr_cpu[BLKIO_STAT_CPU_MERGED], 1,
419				direction, sync);
420	u64_stats_update_end(&stats_cpu->syncp);
421	local_irq_restore(flags);
422}
423EXPORT_SYMBOL_GPL(blkiocg_update_io_merged_stats);
424
425/**
426 * blkg_free - free a blkg
427 * @blkg: blkg to free
428 *
429 * Free @blkg which may be partially allocated.
430 */
431static void blkg_free(struct blkio_group *blkg)
432{
433	if (blkg) {
434		free_percpu(blkg->stats_cpu);
435		kfree(blkg->pd);
436		kfree(blkg);
437	}
438}
439
440/**
441 * blkg_alloc - allocate a blkg
442 * @blkcg: block cgroup the new blkg is associated with
443 * @q: request_queue the new blkg is associated with
444 * @pol: policy the new blkg is associated with
445 *
446 * Allocate a new blkg assocating @blkcg and @q for @pol.
447 *
448 * FIXME: Should be called with queue locked but currently isn't due to
449 *        percpu stat breakage.
450 */
451static struct blkio_group *blkg_alloc(struct blkio_cgroup *blkcg,
452				      struct request_queue *q,
453				      struct blkio_policy_type *pol)
454{
455	struct blkio_group *blkg;
456
457	/* alloc and init base part */
458	blkg = kzalloc_node(sizeof(*blkg), GFP_ATOMIC, q->node);
459	if (!blkg)
460		return NULL;
461
462	spin_lock_init(&blkg->stats_lock);
463	rcu_assign_pointer(blkg->q, q);
464	blkg->blkcg = blkcg;
465	blkg->plid = pol->plid;
466	blkg->refcnt = 1;
467	cgroup_path(blkcg->css.cgroup, blkg->path, sizeof(blkg->path));
468
469	/* alloc per-policy data */
470	blkg->pd = kzalloc_node(sizeof(*blkg->pd) + pol->pdata_size, GFP_ATOMIC,
471				q->node);
472	if (!blkg->pd) {
473		blkg_free(blkg);
474		return NULL;
475	}
476
477	/* broken, read comment in the callsite */
478	blkg->stats_cpu = alloc_percpu(struct blkio_group_stats_cpu);
479	if (!blkg->stats_cpu) {
480		blkg_free(blkg);
481		return NULL;
482	}
483
484	/* attach pd to blkg and invoke per-policy init */
485	blkg->pd->blkg = blkg;
486	pol->ops.blkio_init_group_fn(blkg);
487	return blkg;
488}
489
490struct blkio_group *blkg_lookup_create(struct blkio_cgroup *blkcg,
491				       struct request_queue *q,
492				       enum blkio_policy_id plid,
493				       bool for_root)
494	__releases(q->queue_lock) __acquires(q->queue_lock)
495{
496	struct blkio_policy_type *pol = blkio_policy[plid];
497	struct blkio_group *blkg, *new_blkg;
498
499	WARN_ON_ONCE(!rcu_read_lock_held());
500	lockdep_assert_held(q->queue_lock);
501
502	/*
503	 * This could be the first entry point of blkcg implementation and
504	 * we shouldn't allow anything to go through for a bypassing queue.
505	 * The following can be removed if blkg lookup is guaranteed to
506	 * fail on a bypassing queue.
507	 */
508	if (unlikely(blk_queue_bypass(q)) && !for_root)
509		return ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
510
511	blkg = blkg_lookup(blkcg, q, plid);
512	if (blkg)
513		return blkg;
514
515	/* blkg holds a reference to blkcg */
516	if (!css_tryget(&blkcg->css))
517		return ERR_PTR(-EINVAL);
518
519	/*
520	 * Allocate and initialize.
521	 *
522	 * FIXME: The following is broken.  Percpu memory allocation
523	 * requires %GFP_KERNEL context and can't be performed from IO
524	 * path.  Allocation here should inherently be atomic and the
525	 * following lock dancing can be removed once the broken percpu
526	 * allocation is fixed.
527	 */
528	spin_unlock_irq(q->queue_lock);
529	rcu_read_unlock();
530
531	new_blkg = blkg_alloc(blkcg, q, pol);
532
533	rcu_read_lock();
534	spin_lock_irq(q->queue_lock);
535
536	/* did bypass get turned on inbetween? */
537	if (unlikely(blk_queue_bypass(q)) && !for_root) {
538		blkg = ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
539		goto out;
540	}
541
542	/* did someone beat us to it? */
543	blkg = blkg_lookup(blkcg, q, plid);
544	if (unlikely(blkg))
545		goto out;
546
547	/* did alloc fail? */
548	if (unlikely(!new_blkg)) {
549		blkg = ERR_PTR(-ENOMEM);
550		goto out;
551	}
552
553	/* insert */
554	spin_lock(&blkcg->lock);
555	swap(blkg, new_blkg);
556	hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
557	pol->ops.blkio_link_group_fn(q, blkg);
558	spin_unlock(&blkcg->lock);
559out:
560	blkg_free(new_blkg);
561	return blkg;
562}
563EXPORT_SYMBOL_GPL(blkg_lookup_create);
564
565static void __blkiocg_del_blkio_group(struct blkio_group *blkg)
566{
567	hlist_del_init_rcu(&blkg->blkcg_node);
568}
569
570/*
571 * returns 0 if blkio_group was still on cgroup list. Otherwise returns 1
572 * indicating that blk_group was unhashed by the time we got to it.
573 */
574int blkiocg_del_blkio_group(struct blkio_group *blkg)
575{
576	struct blkio_cgroup *blkcg = blkg->blkcg;
577	unsigned long flags;
578	int ret = 1;
579
580	spin_lock_irqsave(&blkcg->lock, flags);
581	if (!hlist_unhashed(&blkg->blkcg_node)) {
582		__blkiocg_del_blkio_group(blkg);
583		ret = 0;
584	}
585	spin_unlock_irqrestore(&blkcg->lock, flags);
586
587	return ret;
588}
589EXPORT_SYMBOL_GPL(blkiocg_del_blkio_group);
590
591/* called under rcu_read_lock(). */
592struct blkio_group *blkg_lookup(struct blkio_cgroup *blkcg,
593				struct request_queue *q,
594				enum blkio_policy_id plid)
595{
596	struct blkio_group *blkg;
597	struct hlist_node *n;
598
599	hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node)
600		if (blkg->q == q && blkg->plid == plid)
601			return blkg;
602	return NULL;
603}
604EXPORT_SYMBOL_GPL(blkg_lookup);
605
606void blkg_destroy_all(struct request_queue *q)
607{
608	struct blkio_policy_type *pol;
609
610	while (true) {
611		bool done = true;
612
613		spin_lock(&blkio_list_lock);
614		spin_lock_irq(q->queue_lock);
615
616		/*
617		 * clear_queue_fn() might return with non-empty group list
618		 * if it raced cgroup removal and lost.  cgroup removal is
619		 * guaranteed to make forward progress and retrying after a
620		 * while is enough.  This ugliness is scheduled to be
621		 * removed after locking update.
622		 */
623		list_for_each_entry(pol, &blkio_list, list)
624			if (!pol->ops.blkio_clear_queue_fn(q))
625				done = false;
626
627		spin_unlock_irq(q->queue_lock);
628		spin_unlock(&blkio_list_lock);
629
630		if (done)
631			break;
632
633		msleep(10);	/* just some random duration I like */
634	}
635}
636
637static void blkg_rcu_free(struct rcu_head *rcu_head)
638{
639	blkg_free(container_of(rcu_head, struct blkio_group, rcu_head));
640}
641
642void __blkg_release(struct blkio_group *blkg)
643{
644	/* release the extra blkcg reference this blkg has been holding */
645	css_put(&blkg->blkcg->css);
646
647	/*
648	 * A group is freed in rcu manner. But having an rcu lock does not
649	 * mean that one can access all the fields of blkg and assume these
650	 * are valid. For example, don't try to follow throtl_data and
651	 * request queue links.
652	 *
653	 * Having a reference to blkg under an rcu allows acess to only
654	 * values local to groups like group stats and group rate limits
655	 */
656	call_rcu(&blkg->rcu_head, blkg_rcu_free);
657}
658EXPORT_SYMBOL_GPL(__blkg_release);
659
660static void blkio_reset_stats_cpu(struct blkio_group *blkg)
661{
662	struct blkio_group_stats_cpu *stats_cpu;
663	int i, j, k;
664	/*
665	 * Note: On 64 bit arch this should not be an issue. This has the
666	 * possibility of returning some inconsistent value on 32bit arch
667	 * as 64bit update on 32bit is non atomic. Taking care of this
668	 * corner case makes code very complicated, like sending IPIs to
669	 * cpus, taking care of stats of offline cpus etc.
670	 *
671	 * reset stats is anyway more of a debug feature and this sounds a
672	 * corner case. So I am not complicating the code yet until and
673	 * unless this becomes a real issue.
674	 */
675	for_each_possible_cpu(i) {
676		stats_cpu = per_cpu_ptr(blkg->stats_cpu, i);
677		stats_cpu->sectors = 0;
678		for(j = 0; j < BLKIO_STAT_CPU_NR; j++)
679			for (k = 0; k < BLKIO_STAT_TOTAL; k++)
680				stats_cpu->stat_arr_cpu[j][k] = 0;
681	}
682}
683
684static int
685blkiocg_reset_stats(struct cgroup *cgroup, struct cftype *cftype, u64 val)
686{
687	struct blkio_cgroup *blkcg;
688	struct blkio_group *blkg;
689	struct blkio_group_stats *stats;
690	struct hlist_node *n;
691	uint64_t queued[BLKIO_STAT_TOTAL];
692	int i;
693#ifdef CONFIG_DEBUG_BLK_CGROUP
694	bool idling, waiting, empty;
695	unsigned long long now = sched_clock();
696#endif
697
698	blkcg = cgroup_to_blkio_cgroup(cgroup);
699	spin_lock_irq(&blkcg->lock);
700	hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
701		spin_lock(&blkg->stats_lock);
702		stats = &blkg->stats;
703#ifdef CONFIG_DEBUG_BLK_CGROUP
704		idling = blkio_blkg_idling(stats);
705		waiting = blkio_blkg_waiting(stats);
706		empty = blkio_blkg_empty(stats);
707#endif
708		for (i = 0; i < BLKIO_STAT_TOTAL; i++)
709			queued[i] = stats->stat_arr[BLKIO_STAT_QUEUED][i];
710		memset(stats, 0, sizeof(struct blkio_group_stats));
711		for (i = 0; i < BLKIO_STAT_TOTAL; i++)
712			stats->stat_arr[BLKIO_STAT_QUEUED][i] = queued[i];
713#ifdef CONFIG_DEBUG_BLK_CGROUP
714		if (idling) {
715			blkio_mark_blkg_idling(stats);
716			stats->start_idle_time = now;
717		}
718		if (waiting) {
719			blkio_mark_blkg_waiting(stats);
720			stats->start_group_wait_time = now;
721		}
722		if (empty) {
723			blkio_mark_blkg_empty(stats);
724			stats->start_empty_time = now;
725		}
726#endif
727		spin_unlock(&blkg->stats_lock);
728
729		/* Reset Per cpu stats which don't take blkg->stats_lock */
730		blkio_reset_stats_cpu(blkg);
731	}
732
733	spin_unlock_irq(&blkcg->lock);
734	return 0;
735}
736
737static void blkio_get_key_name(enum stat_sub_type type, const char *dname,
738			       char *str, int chars_left, bool diskname_only)
739{
740	snprintf(str, chars_left, "%s", dname);
741	chars_left -= strlen(str);
742	if (chars_left <= 0) {
743		printk(KERN_WARNING
744			"Possibly incorrect cgroup stat display format");
745		return;
746	}
747	if (diskname_only)
748		return;
749	switch (type) {
750	case BLKIO_STAT_READ:
751		strlcat(str, " Read", chars_left);
752		break;
753	case BLKIO_STAT_WRITE:
754		strlcat(str, " Write", chars_left);
755		break;
756	case BLKIO_STAT_SYNC:
757		strlcat(str, " Sync", chars_left);
758		break;
759	case BLKIO_STAT_ASYNC:
760		strlcat(str, " Async", chars_left);
761		break;
762	case BLKIO_STAT_TOTAL:
763		strlcat(str, " Total", chars_left);
764		break;
765	default:
766		strlcat(str, " Invalid", chars_left);
767	}
768}
769
770static uint64_t blkio_fill_stat(char *str, int chars_left, uint64_t val,
771				struct cgroup_map_cb *cb, const char *dname)
772{
773	blkio_get_key_name(0, dname, str, chars_left, true);
774	cb->fill(cb, str, val);
775	return val;
776}
777
778
779static uint64_t blkio_read_stat_cpu(struct blkio_group *blkg,
780			enum stat_type_cpu type, enum stat_sub_type sub_type)
781{
782	int cpu;
783	struct blkio_group_stats_cpu *stats_cpu;
784	u64 val = 0, tval;
785
786	for_each_possible_cpu(cpu) {
787		unsigned int start;
788		stats_cpu  = per_cpu_ptr(blkg->stats_cpu, cpu);
789
790		do {
791			start = u64_stats_fetch_begin(&stats_cpu->syncp);
792			if (type == BLKIO_STAT_CPU_SECTORS)
793				tval = stats_cpu->sectors;
794			else
795				tval = stats_cpu->stat_arr_cpu[type][sub_type];
796		} while(u64_stats_fetch_retry(&stats_cpu->syncp, start));
797
798		val += tval;
799	}
800
801	return val;
802}
803
804static uint64_t blkio_get_stat_cpu(struct blkio_group *blkg,
805				   struct cgroup_map_cb *cb, const char *dname,
806				   enum stat_type_cpu type)
807{
808	uint64_t disk_total, val;
809	char key_str[MAX_KEY_LEN];
810	enum stat_sub_type sub_type;
811
812	if (type == BLKIO_STAT_CPU_SECTORS) {
813		val = blkio_read_stat_cpu(blkg, type, 0);
814		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1, val, cb,
815				       dname);
816	}
817
818	for (sub_type = BLKIO_STAT_READ; sub_type < BLKIO_STAT_TOTAL;
819			sub_type++) {
820		blkio_get_key_name(sub_type, dname, key_str, MAX_KEY_LEN,
821				   false);
822		val = blkio_read_stat_cpu(blkg, type, sub_type);
823		cb->fill(cb, key_str, val);
824	}
825
826	disk_total = blkio_read_stat_cpu(blkg, type, BLKIO_STAT_READ) +
827			blkio_read_stat_cpu(blkg, type, BLKIO_STAT_WRITE);
828
829	blkio_get_key_name(BLKIO_STAT_TOTAL, dname, key_str, MAX_KEY_LEN,
830			   false);
831	cb->fill(cb, key_str, disk_total);
832	return disk_total;
833}
834
835/* This should be called with blkg->stats_lock held */
836static uint64_t blkio_get_stat(struct blkio_group *blkg,
837			       struct cgroup_map_cb *cb, const char *dname,
838			       enum stat_type type)
839{
840	uint64_t disk_total;
841	char key_str[MAX_KEY_LEN];
842	enum stat_sub_type sub_type;
843
844	if (type == BLKIO_STAT_TIME)
845		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
846					blkg->stats.time, cb, dname);
847#ifdef CONFIG_DEBUG_BLK_CGROUP
848	if (type == BLKIO_STAT_UNACCOUNTED_TIME)
849		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
850				       blkg->stats.unaccounted_time, cb, dname);
851	if (type == BLKIO_STAT_AVG_QUEUE_SIZE) {
852		uint64_t sum = blkg->stats.avg_queue_size_sum;
853		uint64_t samples = blkg->stats.avg_queue_size_samples;
854		if (samples)
855			do_div(sum, samples);
856		else
857			sum = 0;
858		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
859				       sum, cb, dname);
860	}
861	if (type == BLKIO_STAT_GROUP_WAIT_TIME)
862		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
863				       blkg->stats.group_wait_time, cb, dname);
864	if (type == BLKIO_STAT_IDLE_TIME)
865		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
866				       blkg->stats.idle_time, cb, dname);
867	if (type == BLKIO_STAT_EMPTY_TIME)
868		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
869				       blkg->stats.empty_time, cb, dname);
870	if (type == BLKIO_STAT_DEQUEUE)
871		return blkio_fill_stat(key_str, MAX_KEY_LEN - 1,
872				       blkg->stats.dequeue, cb, dname);
873#endif
874
875	for (sub_type = BLKIO_STAT_READ; sub_type < BLKIO_STAT_TOTAL;
876			sub_type++) {
877		blkio_get_key_name(sub_type, dname, key_str, MAX_KEY_LEN,
878				   false);
879		cb->fill(cb, key_str, blkg->stats.stat_arr[type][sub_type]);
880	}
881	disk_total = blkg->stats.stat_arr[type][BLKIO_STAT_READ] +
882			blkg->stats.stat_arr[type][BLKIO_STAT_WRITE];
883	blkio_get_key_name(BLKIO_STAT_TOTAL, dname, key_str, MAX_KEY_LEN,
884			   false);
885	cb->fill(cb, key_str, disk_total);
886	return disk_total;
887}
888
889static int blkio_policy_parse_and_set(char *buf, enum blkio_policy_id plid,
890				      int fileid, struct blkio_cgroup *blkcg)
891{
892	struct gendisk *disk = NULL;
893	struct blkio_group *blkg = NULL;
894	char *s[4], *p, *major_s = NULL, *minor_s = NULL;
895	unsigned long major, minor;
896	int i = 0, ret = -EINVAL;
897	int part;
898	dev_t dev;
899	u64 temp;
900
901	memset(s, 0, sizeof(s));
902
903	while ((p = strsep(&buf, " ")) != NULL) {
904		if (!*p)
905			continue;
906
907		s[i++] = p;
908
909		/* Prevent from inputing too many things */
910		if (i == 3)
911			break;
912	}
913
914	if (i != 2)
915		goto out;
916
917	p = strsep(&s[0], ":");
918	if (p != NULL)
919		major_s = p;
920	else
921		goto out;
922
923	minor_s = s[0];
924	if (!minor_s)
925		goto out;
926
927	if (strict_strtoul(major_s, 10, &major))
928		goto out;
929
930	if (strict_strtoul(minor_s, 10, &minor))
931		goto out;
932
933	dev = MKDEV(major, minor);
934
935	if (strict_strtoull(s[1], 10, &temp))
936		goto out;
937
938	disk = get_gendisk(dev, &part);
939	if (!disk || part)
940		goto out;
941
942	rcu_read_lock();
943
944	spin_lock_irq(disk->queue->queue_lock);
945	blkg = blkg_lookup_create(blkcg, disk->queue, plid, false);
946	spin_unlock_irq(disk->queue->queue_lock);
947
948	if (IS_ERR(blkg)) {
949		ret = PTR_ERR(blkg);
950		goto out_unlock;
951	}
952
953	switch (plid) {
954	case BLKIO_POLICY_PROP:
955		if ((temp < BLKIO_WEIGHT_MIN && temp > 0) ||
956		     temp > BLKIO_WEIGHT_MAX)
957			goto out_unlock;
958
959		blkg->conf.weight = temp;
960		blkio_update_group_weight(blkg, temp ?: blkcg->weight);
961		break;
962	case BLKIO_POLICY_THROTL:
963		switch(fileid) {
964		case BLKIO_THROTL_read_bps_device:
965			blkg->conf.bps[READ] = temp;
966			blkio_update_group_bps(blkg, temp ?: -1, fileid);
967			break;
968		case BLKIO_THROTL_write_bps_device:
969			blkg->conf.bps[WRITE] = temp;
970			blkio_update_group_bps(blkg, temp ?: -1, fileid);
971			break;
972		case BLKIO_THROTL_read_iops_device:
973			if (temp > THROTL_IOPS_MAX)
974				goto out_unlock;
975			blkg->conf.iops[READ] = temp;
976			blkio_update_group_iops(blkg, temp ?: -1, fileid);
977			break;
978		case BLKIO_THROTL_write_iops_device:
979			if (temp > THROTL_IOPS_MAX)
980				goto out_unlock;
981			blkg->conf.iops[WRITE] = temp;
982			blkio_update_group_iops(blkg, temp ?: -1, fileid);
983			break;
984		}
985		break;
986	default:
987		BUG();
988	}
989	ret = 0;
990out_unlock:
991	rcu_read_unlock();
992out:
993	put_disk(disk);
994
995	/*
996	 * If queue was bypassing, we should retry.  Do so after a short
997	 * msleep().  It isn't strictly necessary but queue can be
998	 * bypassing for some time and it's always nice to avoid busy
999	 * looping.
1000	 */
1001	if (ret == -EBUSY) {
1002		msleep(10);
1003		return restart_syscall();
1004	}
1005	return ret;
1006}
1007
1008static int blkiocg_file_write(struct cgroup *cgrp, struct cftype *cft,
1009 				       const char *buffer)
1010{
1011	int ret = 0;
1012	char *buf;
1013	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp);
1014	enum blkio_policy_id plid = BLKIOFILE_POLICY(cft->private);
1015	int fileid = BLKIOFILE_ATTR(cft->private);
1016
1017	buf = kstrdup(buffer, GFP_KERNEL);
1018	if (!buf)
1019		return -ENOMEM;
1020
1021	ret = blkio_policy_parse_and_set(buf, plid, fileid, blkcg);
1022	kfree(buf);
1023	return ret;
1024}
1025
1026static const char *blkg_dev_name(struct blkio_group *blkg)
1027{
1028	/* some drivers (floppy) instantiate a queue w/o disk registered */
1029	if (blkg->q->backing_dev_info.dev)
1030		return dev_name(blkg->q->backing_dev_info.dev);
1031	return NULL;
1032}
1033
1034static void blkio_print_group_conf(struct cftype *cft, struct blkio_group *blkg,
1035				   struct seq_file *m)
1036{
1037	const char *dname = blkg_dev_name(blkg);
1038	int fileid = BLKIOFILE_ATTR(cft->private);
1039	int rw = WRITE;
1040
1041	if (!dname)
1042		return;
1043
1044	switch (blkg->plid) {
1045		case BLKIO_POLICY_PROP:
1046			if (blkg->conf.weight)
1047				seq_printf(m, "%s\t%u\n",
1048					   dname, blkg->conf.weight);
1049			break;
1050		case BLKIO_POLICY_THROTL:
1051			switch (fileid) {
1052			case BLKIO_THROTL_read_bps_device:
1053				rw = READ;
1054			case BLKIO_THROTL_write_bps_device:
1055				if (blkg->conf.bps[rw])
1056					seq_printf(m, "%s\t%llu\n",
1057						   dname, blkg->conf.bps[rw]);
1058				break;
1059			case BLKIO_THROTL_read_iops_device:
1060				rw = READ;
1061			case BLKIO_THROTL_write_iops_device:
1062				if (blkg->conf.iops[rw])
1063					seq_printf(m, "%s\t%u\n",
1064						   dname, blkg->conf.iops[rw]);
1065				break;
1066			}
1067			break;
1068		default:
1069			BUG();
1070	}
1071}
1072
1073/* cgroup files which read their data from policy nodes end up here */
1074static void blkio_read_conf(struct cftype *cft, struct blkio_cgroup *blkcg,
1075			    struct seq_file *m)
1076{
1077	struct blkio_group *blkg;
1078	struct hlist_node *n;
1079
1080	spin_lock_irq(&blkcg->lock);
1081	hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node)
1082		if (BLKIOFILE_POLICY(cft->private) == blkg->plid)
1083			blkio_print_group_conf(cft, blkg, m);
1084	spin_unlock_irq(&blkcg->lock);
1085}
1086
1087static int blkiocg_file_read(struct cgroup *cgrp, struct cftype *cft,
1088				struct seq_file *m)
1089{
1090	struct blkio_cgroup *blkcg;
1091	enum blkio_policy_id plid = BLKIOFILE_POLICY(cft->private);
1092	int name = BLKIOFILE_ATTR(cft->private);
1093
1094	blkcg = cgroup_to_blkio_cgroup(cgrp);
1095
1096	switch(plid) {
1097	case BLKIO_POLICY_PROP:
1098		switch(name) {
1099		case BLKIO_PROP_weight_device:
1100			blkio_read_conf(cft, blkcg, m);
1101			return 0;
1102		default:
1103			BUG();
1104		}
1105		break;
1106	case BLKIO_POLICY_THROTL:
1107		switch(name){
1108		case BLKIO_THROTL_read_bps_device:
1109		case BLKIO_THROTL_write_bps_device:
1110		case BLKIO_THROTL_read_iops_device:
1111		case BLKIO_THROTL_write_iops_device:
1112			blkio_read_conf(cft, blkcg, m);
1113			return 0;
1114		default:
1115			BUG();
1116		}
1117		break;
1118	default:
1119		BUG();
1120	}
1121
1122	return 0;
1123}
1124
1125static int blkio_read_blkg_stats(struct blkio_cgroup *blkcg,
1126		struct cftype *cft, struct cgroup_map_cb *cb,
1127		enum stat_type type, bool show_total, bool pcpu)
1128{
1129	struct blkio_group *blkg;
1130	struct hlist_node *n;
1131	uint64_t cgroup_total = 0;
1132
1133	rcu_read_lock();
1134	hlist_for_each_entry_rcu(blkg, n, &blkcg->blkg_list, blkcg_node) {
1135		const char *dname = blkg_dev_name(blkg);
1136
1137		if (!dname || BLKIOFILE_POLICY(cft->private) != blkg->plid)
1138			continue;
1139		if (pcpu)
1140			cgroup_total += blkio_get_stat_cpu(blkg, cb, dname,
1141							   type);
1142		else {
1143			spin_lock_irq(&blkg->stats_lock);
1144			cgroup_total += blkio_get_stat(blkg, cb, dname, type);
1145			spin_unlock_irq(&blkg->stats_lock);
1146		}
1147	}
1148	if (show_total)
1149		cb->fill(cb, "Total", cgroup_total);
1150	rcu_read_unlock();
1151	return 0;
1152}
1153
1154/* All map kind of cgroup file get serviced by this function */
1155static int blkiocg_file_read_map(struct cgroup *cgrp, struct cftype *cft,
1156				struct cgroup_map_cb *cb)
1157{
1158	struct blkio_cgroup *blkcg;
1159	enum blkio_policy_id plid = BLKIOFILE_POLICY(cft->private);
1160	int name = BLKIOFILE_ATTR(cft->private);
1161
1162	blkcg = cgroup_to_blkio_cgroup(cgrp);
1163
1164	switch(plid) {
1165	case BLKIO_POLICY_PROP:
1166		switch(name) {
1167		case BLKIO_PROP_time:
1168			return blkio_read_blkg_stats(blkcg, cft, cb,
1169						BLKIO_STAT_TIME, 0, 0);
1170		case BLKIO_PROP_sectors:
1171			return blkio_read_blkg_stats(blkcg, cft, cb,
1172						BLKIO_STAT_CPU_SECTORS, 0, 1);
1173		case BLKIO_PROP_io_service_bytes:
1174			return blkio_read_blkg_stats(blkcg, cft, cb,
1175					BLKIO_STAT_CPU_SERVICE_BYTES, 1, 1);
1176		case BLKIO_PROP_io_serviced:
1177			return blkio_read_blkg_stats(blkcg, cft, cb,
1178						BLKIO_STAT_CPU_SERVICED, 1, 1);
1179		case BLKIO_PROP_io_service_time:
1180			return blkio_read_blkg_stats(blkcg, cft, cb,
1181						BLKIO_STAT_SERVICE_TIME, 1, 0);
1182		case BLKIO_PROP_io_wait_time:
1183			return blkio_read_blkg_stats(blkcg, cft, cb,
1184						BLKIO_STAT_WAIT_TIME, 1, 0);
1185		case BLKIO_PROP_io_merged:
1186			return blkio_read_blkg_stats(blkcg, cft, cb,
1187						BLKIO_STAT_CPU_MERGED, 1, 1);
1188		case BLKIO_PROP_io_queued:
1189			return blkio_read_blkg_stats(blkcg, cft, cb,
1190						BLKIO_STAT_QUEUED, 1, 0);
1191#ifdef CONFIG_DEBUG_BLK_CGROUP
1192		case BLKIO_PROP_unaccounted_time:
1193			return blkio_read_blkg_stats(blkcg, cft, cb,
1194					BLKIO_STAT_UNACCOUNTED_TIME, 0, 0);
1195		case BLKIO_PROP_dequeue:
1196			return blkio_read_blkg_stats(blkcg, cft, cb,
1197						BLKIO_STAT_DEQUEUE, 0, 0);
1198		case BLKIO_PROP_avg_queue_size:
1199			return blkio_read_blkg_stats(blkcg, cft, cb,
1200					BLKIO_STAT_AVG_QUEUE_SIZE, 0, 0);
1201		case BLKIO_PROP_group_wait_time:
1202			return blkio_read_blkg_stats(blkcg, cft, cb,
1203					BLKIO_STAT_GROUP_WAIT_TIME, 0, 0);
1204		case BLKIO_PROP_idle_time:
1205			return blkio_read_blkg_stats(blkcg, cft, cb,
1206						BLKIO_STAT_IDLE_TIME, 0, 0);
1207		case BLKIO_PROP_empty_time:
1208			return blkio_read_blkg_stats(blkcg, cft, cb,
1209						BLKIO_STAT_EMPTY_TIME, 0, 0);
1210#endif
1211		default:
1212			BUG();
1213		}
1214		break;
1215	case BLKIO_POLICY_THROTL:
1216		switch(name){
1217		case BLKIO_THROTL_io_service_bytes:
1218			return blkio_read_blkg_stats(blkcg, cft, cb,
1219						BLKIO_STAT_CPU_SERVICE_BYTES, 1, 1);
1220		case BLKIO_THROTL_io_serviced:
1221			return blkio_read_blkg_stats(blkcg, cft, cb,
1222						BLKIO_STAT_CPU_SERVICED, 1, 1);
1223		default:
1224			BUG();
1225		}
1226		break;
1227	default:
1228		BUG();
1229	}
1230
1231	return 0;
1232}
1233
1234static int blkio_weight_write(struct blkio_cgroup *blkcg, int plid, u64 val)
1235{
1236	struct blkio_group *blkg;
1237	struct hlist_node *n;
1238
1239	if (val < BLKIO_WEIGHT_MIN || val > BLKIO_WEIGHT_MAX)
1240		return -EINVAL;
1241
1242	spin_lock(&blkio_list_lock);
1243	spin_lock_irq(&blkcg->lock);
1244	blkcg->weight = (unsigned int)val;
1245
1246	hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node)
1247		if (blkg->plid == plid && !blkg->conf.weight)
1248			blkio_update_group_weight(blkg, blkcg->weight);
1249
1250	spin_unlock_irq(&blkcg->lock);
1251	spin_unlock(&blkio_list_lock);
1252	return 0;
1253}
1254
1255static u64 blkiocg_file_read_u64 (struct cgroup *cgrp, struct cftype *cft) {
1256	struct blkio_cgroup *blkcg;
1257	enum blkio_policy_id plid = BLKIOFILE_POLICY(cft->private);
1258	int name = BLKIOFILE_ATTR(cft->private);
1259
1260	blkcg = cgroup_to_blkio_cgroup(cgrp);
1261
1262	switch(plid) {
1263	case BLKIO_POLICY_PROP:
1264		switch(name) {
1265		case BLKIO_PROP_weight:
1266			return (u64)blkcg->weight;
1267		}
1268		break;
1269	default:
1270		BUG();
1271	}
1272	return 0;
1273}
1274
1275static int
1276blkiocg_file_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1277{
1278	struct blkio_cgroup *blkcg;
1279	enum blkio_policy_id plid = BLKIOFILE_POLICY(cft->private);
1280	int name = BLKIOFILE_ATTR(cft->private);
1281
1282	blkcg = cgroup_to_blkio_cgroup(cgrp);
1283
1284	switch(plid) {
1285	case BLKIO_POLICY_PROP:
1286		switch(name) {
1287		case BLKIO_PROP_weight:
1288			return blkio_weight_write(blkcg, plid, val);
1289		}
1290		break;
1291	default:
1292		BUG();
1293	}
1294
1295	return 0;
1296}
1297
1298struct cftype blkio_files[] = {
1299	{
1300		.name = "weight_device",
1301		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1302				BLKIO_PROP_weight_device),
1303		.read_seq_string = blkiocg_file_read,
1304		.write_string = blkiocg_file_write,
1305		.max_write_len = 256,
1306	},
1307	{
1308		.name = "weight",
1309		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1310				BLKIO_PROP_weight),
1311		.read_u64 = blkiocg_file_read_u64,
1312		.write_u64 = blkiocg_file_write_u64,
1313	},
1314	{
1315		.name = "time",
1316		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1317				BLKIO_PROP_time),
1318		.read_map = blkiocg_file_read_map,
1319	},
1320	{
1321		.name = "sectors",
1322		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1323				BLKIO_PROP_sectors),
1324		.read_map = blkiocg_file_read_map,
1325	},
1326	{
1327		.name = "io_service_bytes",
1328		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1329				BLKIO_PROP_io_service_bytes),
1330		.read_map = blkiocg_file_read_map,
1331	},
1332	{
1333		.name = "io_serviced",
1334		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1335				BLKIO_PROP_io_serviced),
1336		.read_map = blkiocg_file_read_map,
1337	},
1338	{
1339		.name = "io_service_time",
1340		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1341				BLKIO_PROP_io_service_time),
1342		.read_map = blkiocg_file_read_map,
1343	},
1344	{
1345		.name = "io_wait_time",
1346		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1347				BLKIO_PROP_io_wait_time),
1348		.read_map = blkiocg_file_read_map,
1349	},
1350	{
1351		.name = "io_merged",
1352		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1353				BLKIO_PROP_io_merged),
1354		.read_map = blkiocg_file_read_map,
1355	},
1356	{
1357		.name = "io_queued",
1358		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1359				BLKIO_PROP_io_queued),
1360		.read_map = blkiocg_file_read_map,
1361	},
1362	{
1363		.name = "reset_stats",
1364		.write_u64 = blkiocg_reset_stats,
1365	},
1366#ifdef CONFIG_BLK_DEV_THROTTLING
1367	{
1368		.name = "throttle.read_bps_device",
1369		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1370				BLKIO_THROTL_read_bps_device),
1371		.read_seq_string = blkiocg_file_read,
1372		.write_string = blkiocg_file_write,
1373		.max_write_len = 256,
1374	},
1375
1376	{
1377		.name = "throttle.write_bps_device",
1378		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1379				BLKIO_THROTL_write_bps_device),
1380		.read_seq_string = blkiocg_file_read,
1381		.write_string = blkiocg_file_write,
1382		.max_write_len = 256,
1383	},
1384
1385	{
1386		.name = "throttle.read_iops_device",
1387		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1388				BLKIO_THROTL_read_iops_device),
1389		.read_seq_string = blkiocg_file_read,
1390		.write_string = blkiocg_file_write,
1391		.max_write_len = 256,
1392	},
1393
1394	{
1395		.name = "throttle.write_iops_device",
1396		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1397				BLKIO_THROTL_write_iops_device),
1398		.read_seq_string = blkiocg_file_read,
1399		.write_string = blkiocg_file_write,
1400		.max_write_len = 256,
1401	},
1402	{
1403		.name = "throttle.io_service_bytes",
1404		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1405				BLKIO_THROTL_io_service_bytes),
1406		.read_map = blkiocg_file_read_map,
1407	},
1408	{
1409		.name = "throttle.io_serviced",
1410		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_THROTL,
1411				BLKIO_THROTL_io_serviced),
1412		.read_map = blkiocg_file_read_map,
1413	},
1414#endif /* CONFIG_BLK_DEV_THROTTLING */
1415
1416#ifdef CONFIG_DEBUG_BLK_CGROUP
1417	{
1418		.name = "avg_queue_size",
1419		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1420				BLKIO_PROP_avg_queue_size),
1421		.read_map = blkiocg_file_read_map,
1422	},
1423	{
1424		.name = "group_wait_time",
1425		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1426				BLKIO_PROP_group_wait_time),
1427		.read_map = blkiocg_file_read_map,
1428	},
1429	{
1430		.name = "idle_time",
1431		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1432				BLKIO_PROP_idle_time),
1433		.read_map = blkiocg_file_read_map,
1434	},
1435	{
1436		.name = "empty_time",
1437		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1438				BLKIO_PROP_empty_time),
1439		.read_map = blkiocg_file_read_map,
1440	},
1441	{
1442		.name = "dequeue",
1443		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1444				BLKIO_PROP_dequeue),
1445		.read_map = blkiocg_file_read_map,
1446	},
1447	{
1448		.name = "unaccounted_time",
1449		.private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
1450				BLKIO_PROP_unaccounted_time),
1451		.read_map = blkiocg_file_read_map,
1452	},
1453#endif
1454};
1455
1456static int blkiocg_populate(struct cgroup_subsys *subsys, struct cgroup *cgroup)
1457{
1458	return cgroup_add_files(cgroup, subsys, blkio_files,
1459				ARRAY_SIZE(blkio_files));
1460}
1461
1462static int blkiocg_pre_destroy(struct cgroup_subsys *subsys,
1463			       struct cgroup *cgroup)
1464{
1465	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
1466	unsigned long flags;
1467	struct blkio_group *blkg;
1468	struct request_queue *q;
1469	struct blkio_policy_type *blkiop;
1470
1471	rcu_read_lock();
1472
1473	do {
1474		spin_lock_irqsave(&blkcg->lock, flags);
1475
1476		if (hlist_empty(&blkcg->blkg_list)) {
1477			spin_unlock_irqrestore(&blkcg->lock, flags);
1478			break;
1479		}
1480
1481		blkg = hlist_entry(blkcg->blkg_list.first, struct blkio_group,
1482					blkcg_node);
1483		q = rcu_dereference(blkg->q);
1484		__blkiocg_del_blkio_group(blkg);
1485
1486		spin_unlock_irqrestore(&blkcg->lock, flags);
1487
1488		/*
1489		 * This blkio_group is being unlinked as associated cgroup is
1490		 * going away. Let all the IO controlling policies know about
1491		 * this event.
1492		 */
1493		spin_lock(&blkio_list_lock);
1494		list_for_each_entry(blkiop, &blkio_list, list) {
1495			if (blkiop->plid != blkg->plid)
1496				continue;
1497			blkiop->ops.blkio_unlink_group_fn(q, blkg);
1498		}
1499		spin_unlock(&blkio_list_lock);
1500	} while (1);
1501
1502	rcu_read_unlock();
1503
1504	return 0;
1505}
1506
1507static void blkiocg_destroy(struct cgroup_subsys *subsys, struct cgroup *cgroup)
1508{
1509	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
1510
1511	if (blkcg != &blkio_root_cgroup)
1512		kfree(blkcg);
1513}
1514
1515static struct cgroup_subsys_state *
1516blkiocg_create(struct cgroup_subsys *subsys, struct cgroup *cgroup)
1517{
1518	struct blkio_cgroup *blkcg;
1519	struct cgroup *parent = cgroup->parent;
1520
1521	if (!parent) {
1522		blkcg = &blkio_root_cgroup;
1523		goto done;
1524	}
1525
1526	blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1527	if (!blkcg)
1528		return ERR_PTR(-ENOMEM);
1529
1530	blkcg->weight = BLKIO_WEIGHT_DEFAULT;
1531done:
1532	spin_lock_init(&blkcg->lock);
1533	INIT_HLIST_HEAD(&blkcg->blkg_list);
1534
1535	return &blkcg->css;
1536}
1537
1538/**
1539 * blkcg_init_queue - initialize blkcg part of request queue
1540 * @q: request_queue to initialize
1541 *
1542 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1543 * part of new request_queue @q.
1544 *
1545 * RETURNS:
1546 * 0 on success, -errno on failure.
1547 */
1548int blkcg_init_queue(struct request_queue *q)
1549{
1550	int ret;
1551
1552	might_sleep();
1553
1554	ret = blk_throtl_init(q);
1555	if (ret)
1556		return ret;
1557
1558	mutex_lock(&all_q_mutex);
1559	INIT_LIST_HEAD(&q->all_q_node);
1560	list_add_tail(&q->all_q_node, &all_q_list);
1561	mutex_unlock(&all_q_mutex);
1562
1563	return 0;
1564}
1565
1566/**
1567 * blkcg_drain_queue - drain blkcg part of request_queue
1568 * @q: request_queue to drain
1569 *
1570 * Called from blk_drain_queue().  Responsible for draining blkcg part.
1571 */
1572void blkcg_drain_queue(struct request_queue *q)
1573{
1574	lockdep_assert_held(q->queue_lock);
1575
1576	blk_throtl_drain(q);
1577}
1578
1579/**
1580 * blkcg_exit_queue - exit and release blkcg part of request_queue
1581 * @q: request_queue being released
1582 *
1583 * Called from blk_release_queue().  Responsible for exiting blkcg part.
1584 */
1585void blkcg_exit_queue(struct request_queue *q)
1586{
1587	mutex_lock(&all_q_mutex);
1588	list_del_init(&q->all_q_node);
1589	mutex_unlock(&all_q_mutex);
1590
1591	blk_throtl_exit(q);
1592}
1593
1594/*
1595 * We cannot support shared io contexts, as we have no mean to support
1596 * two tasks with the same ioc in two different groups without major rework
1597 * of the main cic data structures.  For now we allow a task to change
1598 * its cgroup only if it's the only owner of its ioc.
1599 */
1600static int blkiocg_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1601			      struct cgroup_taskset *tset)
1602{
1603	struct task_struct *task;
1604	struct io_context *ioc;
1605	int ret = 0;
1606
1607	/* task_lock() is needed to avoid races with exit_io_context() */
1608	cgroup_taskset_for_each(task, cgrp, tset) {
1609		task_lock(task);
1610		ioc = task->io_context;
1611		if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1612			ret = -EINVAL;
1613		task_unlock(task);
1614		if (ret)
1615			break;
1616	}
1617	return ret;
1618}
1619
1620static void blkiocg_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
1621			   struct cgroup_taskset *tset)
1622{
1623	struct task_struct *task;
1624	struct io_context *ioc;
1625
1626	cgroup_taskset_for_each(task, cgrp, tset) {
1627		/* we don't lose anything even if ioc allocation fails */
1628		ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE);
1629		if (ioc) {
1630			ioc_cgroup_changed(ioc);
1631			put_io_context(ioc);
1632		}
1633	}
1634}
1635
1636static void blkcg_bypass_start(void)
1637	__acquires(&all_q_mutex)
1638{
1639	struct request_queue *q;
1640
1641	mutex_lock(&all_q_mutex);
1642
1643	list_for_each_entry(q, &all_q_list, all_q_node) {
1644		blk_queue_bypass_start(q);
1645		blkg_destroy_all(q);
1646	}
1647}
1648
1649static void blkcg_bypass_end(void)
1650	__releases(&all_q_mutex)
1651{
1652	struct request_queue *q;
1653
1654	list_for_each_entry(q, &all_q_list, all_q_node)
1655		blk_queue_bypass_end(q);
1656
1657	mutex_unlock(&all_q_mutex);
1658}
1659
1660void blkio_policy_register(struct blkio_policy_type *blkiop)
1661{
1662	blkcg_bypass_start();
1663	spin_lock(&blkio_list_lock);
1664
1665	BUG_ON(blkio_policy[blkiop->plid]);
1666	blkio_policy[blkiop->plid] = blkiop;
1667	list_add_tail(&blkiop->list, &blkio_list);
1668
1669	spin_unlock(&blkio_list_lock);
1670	blkcg_bypass_end();
1671}
1672EXPORT_SYMBOL_GPL(blkio_policy_register);
1673
1674void blkio_policy_unregister(struct blkio_policy_type *blkiop)
1675{
1676	blkcg_bypass_start();
1677	spin_lock(&blkio_list_lock);
1678
1679	BUG_ON(blkio_policy[blkiop->plid] != blkiop);
1680	blkio_policy[blkiop->plid] = NULL;
1681	list_del_init(&blkiop->list);
1682
1683	spin_unlock(&blkio_list_lock);
1684	blkcg_bypass_end();
1685}
1686EXPORT_SYMBOL_GPL(blkio_policy_unregister);
1687