1	// SPDX-License-Identifier: GPL-2.0
2	/*
3	* Common Block IO controller cgroup interface
4	*
5	* Based on ideas and code from CFQ, CFS and BFQ:
6	* Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7	*
8	* Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9	* Paolo Valente <paolo.valente@unimore.it>
10	*
11	* Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12	* Nauman Rafique <nauman@google.com>
13	*
14	* For policy-specific per-blkcg data:
15	* Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16	* Arianna Avanzini <avanzini.arianna@gmail.com>
17	*/
18	#include <linux/ioprio.h>
19	#include <linux/kdev_t.h>
20	#include <linux/module.h>
21	#include <linux/sched/signal.h>
22	#include <linux/err.h>
23	#include <linux/blkdev.h>
24	#include <linux/backing-dev.h>
25	#include <linux/slab.h>
26	#include <linux/delay.h>
27	#include <linux/atomic.h>
28	#include <linux/ctype.h>
29	#include <linux/resume_user_mode.h>
30	#include <linux/psi.h>
31	#include <linux/part_stat.h>
32	#include "blk.h"
33	#include "blk-cgroup.h"
34	#include "blk-ioprio.h"
35	#include "blk-throttle.h"
36
37	static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu);
38
39	/*
40	* blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
41	* blkcg_pol_register_mutex nests outside of it and synchronizes entire
42	* policy [un]register operations including cgroup file additions /
43	* removals. Putting cgroup file registration outside blkcg_pol_mutex
44	* allows grabbing it from cgroup callbacks.
45	*/
46	static DEFINE_MUTEX(blkcg_pol_register_mutex);
47	static DEFINE_MUTEX(blkcg_pol_mutex);
48
49	struct blkcg blkcg_root;
50	EXPORT_SYMBOL_GPL(blkcg_root);
51
52	struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
53	EXPORT_SYMBOL_GPL(blkcg_root_css);
54
55	static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
56
57	static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
58
59	bool blkcg_debug_stats = false;
60
61	static DEFINE_RAW_SPINLOCK(blkg_stat_lock);
62
63	#define BLKG_DESTROY_BATCH_SIZE 64
64
65	/*
66	* Lockless lists for tracking IO stats update
67	*
68	* New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
69	* There are multiple blkg's (one for each block device) attached to each
70	* blkcg. The rstat code keeps track of which cpu has IO stats updated,
71	* but it doesn't know which blkg has the updated stats. If there are many
72	* block devices in a system, the cost of iterating all the blkg's to flush
73	* out the IO stats can be high. To reduce such overhead, a set of percpu
74	* lockless lists (lhead) per blkcg are used to track the set of recently
75	* updated iostat_cpu's since the last flush. An iostat_cpu will be put
76	* onto the lockless list on the update side [blk_cgroup_bio_start()] if
77	* not there yet and then removed when being flushed [blkcg_rstat_flush()].
78	* References to blkg are gotten and then put back in the process to
79	* protect against blkg removal.
80	*
81	* Return: 0 if successful or -ENOMEM if allocation fails.
82	*/
83	static int init_blkcg_llists(struct blkcg *blkcg)
84	{
85	int cpu;
86
87	blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL);
88	if (!blkcg->lhead)
89	return -ENOMEM;
90
91	for_each_possible_cpu(cpu)
92	init_llist_head(per_cpu_ptr(blkcg->lhead, cpu));
93	return 0;
94	}
95
96	/**
97	* blkcg_css - find the current css
98	*
99	* Find the css associated with either the kthread or the current task.
100	* This may return a dying css, so it is up to the caller to use tryget logic
101	* to confirm it is alive and well.
102	*/
103	static struct cgroup_subsys_state *blkcg_css(void)
104	{
105	struct cgroup_subsys_state *css;
106
107	css = kthread_blkcg();
108	if (css)
109	return css;
110	return task_css(current, subsys_id: io_cgrp_id);
111	}
112
113	static bool blkcg_policy_enabled(struct request_queue *q,
114	const struct blkcg_policy *pol)
115	{
116	return pol && test_bit(pol->plid, q->blkcg_pols);
117	}
118
119	static void blkg_free_workfn(struct work_struct *work)
120	{
121	struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
122	free_work);
123	struct request_queue *q = blkg->q;
124	int i;
125
126	/*
127	* pd_free_fn() can also be called from blkcg_deactivate_policy(),
128	* in order to make sure pd_free_fn() is called in order, the deletion
129	* of the list blkg->q_node is delayed to here from blkg_destroy(), and
130	* blkcg_mutex is used to synchronize blkg_free_workfn() and
131	* blkcg_deactivate_policy().
132	*/
133	mutex_lock(&q->blkcg_mutex);
134	for (i = 0; i < BLKCG_MAX_POLS; i++)
135	if (blkg->pd[i])
136	blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
137	if (blkg->parent)
138	blkg_put(blkg: blkg->parent);
139	spin_lock_irq(lock: &q->queue_lock);
140	list_del_init(entry: &blkg->q_node);
141	spin_unlock_irq(lock: &q->queue_lock);
142	mutex_unlock(lock: &q->blkcg_mutex);
143
144	blk_put_queue(q);
145	free_percpu(pdata: blkg->iostat_cpu);
146	percpu_ref_exit(ref: &blkg->refcnt);
147	kfree(objp: blkg);
148	}
149
150	/**
151	* blkg_free - free a blkg
152	* @blkg: blkg to free
153	*
154	* Free @blkg which may be partially allocated.
155	*/
156	static void blkg_free(struct blkcg_gq *blkg)
157	{
158	if (!blkg)
159	return;
160
161	/*
162	* Both ->pd_free_fn() and request queue's release handler may
163	* sleep, so free us by scheduling one work func
164	*/
165	INIT_WORK(&blkg->free_work, blkg_free_workfn);
166	schedule_work(work: &blkg->free_work);
167	}
168
169	static void __blkg_release(struct rcu_head *rcu)
170	{
171	struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
172	struct blkcg *blkcg = blkg->blkcg;
173	int cpu;
174
175	#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
176	WARN_ON(!bio_list_empty(&blkg->async_bios));
177	#endif
178	/*
179	* Flush all the non-empty percpu lockless lists before releasing
180	* us, given these stat belongs to us.
181	*
182	* blkg_stat_lock is for serializing blkg stat update
183	*/
184	for_each_possible_cpu(cpu)
185	__blkcg_rstat_flush(blkcg, cpu);
186
187	/* release the blkcg and parent blkg refs this blkg has been holding */
188	css_put(css: &blkg->blkcg->css);
189	blkg_free(blkg);
190	}
191
192	/*
193	* A group is RCU protected, but having an rcu lock does not mean that one
194	* can access all the fields of blkg and assume these are valid. For
195	* example, don't try to follow throtl_data and request queue links.
196	*
197	* Having a reference to blkg under an rcu allows accesses to only values
198	* local to groups like group stats and group rate limits.
199	*/
200	static void blkg_release(struct percpu_ref *ref)
201	{
202	struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
203
204	call_rcu(head: &blkg->rcu_head, func: __blkg_release);
205	}
206
207	#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
208	static struct workqueue_struct *blkcg_punt_bio_wq;
209
210	static void blkg_async_bio_workfn(struct work_struct *work)
211	{
212	struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
213	async_bio_work);
214	struct bio_list bios = BIO_EMPTY_LIST;
215	struct bio *bio;
216	struct blk_plug plug;
217	bool need_plug = false;
218
219	/* as long as there are pending bios, @blkg can't go away */
220	spin_lock(lock: &blkg->async_bio_lock);
221	bio_list_merge(bl: &bios, bl2: &blkg->async_bios);
222	bio_list_init(bl: &blkg->async_bios);
223	spin_unlock(lock: &blkg->async_bio_lock);
224
225	/* start plug only when bio_list contains at least 2 bios */
226	if (bios.head && bios.head->bi_next) {
227	need_plug = true;
228	blk_start_plug(&plug);
229	}
230	while ((bio = bio_list_pop(bl: &bios)))
231	submit_bio(bio);
232	if (need_plug)
233	blk_finish_plug(&plug);
234	}
235
236	/*
237	* When a shared kthread issues a bio for a cgroup, doing so synchronously can
238	* lead to priority inversions as the kthread can be trapped waiting for that
239	* cgroup. Use this helper instead of submit_bio to punt the actual issuing to
240	* a dedicated per-blkcg work item to avoid such priority inversions.
241	*/
242	void blkcg_punt_bio_submit(struct bio *bio)
243	{
244	struct blkcg_gq *blkg = bio->bi_blkg;
245
246	if (blkg->parent) {
247	spin_lock(lock: &blkg->async_bio_lock);
248	bio_list_add(bl: &blkg->async_bios, bio);
249	spin_unlock(lock: &blkg->async_bio_lock);
250	queue_work(wq: blkcg_punt_bio_wq, work: &blkg->async_bio_work);
251	} else {
252	/* never bounce for the root cgroup */
253	submit_bio(bio);
254	}
255	}
256	EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit);
257
258	static int __init blkcg_punt_bio_init(void)
259	{
260	blkcg_punt_bio_wq = alloc_workqueue(fmt: "blkcg_punt_bio",
261	flags: WQ_MEM_RECLAIM | WQ_FREEZABLE |
262	WQ_UNBOUND | WQ_SYSFS, max_active: 0);
263	if (!blkcg_punt_bio_wq)
264	return -ENOMEM;
265	return 0;
266	}
267	subsys_initcall(blkcg_punt_bio_init);
268	#endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
269
270	/**
271	* bio_blkcg_css - return the blkcg CSS associated with a bio
272	* @bio: target bio
273	*
274	* This returns the CSS for the blkcg associated with a bio, or %NULL if not
275	* associated. Callers are expected to either handle %NULL or know association
276	* has been done prior to calling this.
277	*/
278	struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
279	{
280	if (!bio || !bio->bi_blkg)
281	return NULL;
282	return &bio->bi_blkg->blkcg->css;
283	}
284	EXPORT_SYMBOL_GPL(bio_blkcg_css);
285
286	/**
287	* blkcg_parent - get the parent of a blkcg
288	* @blkcg: blkcg of interest
289	*
290	* Return the parent blkcg of @blkcg. Can be called anytime.
291	*/
292	static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
293	{
294	return css_to_blkcg(css: blkcg->css.parent);
295	}
296
297	/**
298	* blkg_alloc - allocate a blkg
299	* @blkcg: block cgroup the new blkg is associated with
300	* @disk: gendisk the new blkg is associated with
301	* @gfp_mask: allocation mask to use
302	*
303	* Allocate a new blkg associating @blkcg and @disk.
304	*/
305	static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
306	gfp_t gfp_mask)
307	{
308	struct blkcg_gq *blkg;
309	int i, cpu;
310
311	/* alloc and init base part */
312	blkg = kzalloc_node(size: sizeof(*blkg), flags: gfp_mask, node: disk->queue->node);
313	if (!blkg)
314	return NULL;
315	if (percpu_ref_init(ref: &blkg->refcnt, release: blkg_release, flags: 0, gfp: gfp_mask))
316	goto out_free_blkg;
317	blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
318	if (!blkg->iostat_cpu)
319	goto out_exit_refcnt;
320	if (!blk_get_queue(disk->queue))
321	goto out_free_iostat;
322
323	blkg->q = disk->queue;
324	INIT_LIST_HEAD(list: &blkg->q_node);
325	blkg->blkcg = blkcg;
326	#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
327	spin_lock_init(&blkg->async_bio_lock);
328	bio_list_init(bl: &blkg->async_bios);
329	INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
330	#endif
331
332	u64_stats_init(syncp: &blkg->iostat.sync);
333	for_each_possible_cpu(cpu) {
334	u64_stats_init(syncp: &per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
335	per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
336	}
337
338	for (i = 0; i < BLKCG_MAX_POLS; i++) {
339	struct blkcg_policy *pol = blkcg_policy[i];
340	struct blkg_policy_data *pd;
341
342	if (!blkcg_policy_enabled(q: disk->queue, pol))
343	continue;
344
345	/* alloc per-policy data and attach it to blkg */
346	pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
347	if (!pd)
348	goto out_free_pds;
349	blkg->pd[i] = pd;
350	pd->blkg = blkg;
351	pd->plid = i;
352	pd->online = false;
353	}
354
355	return blkg;
356
357	out_free_pds:
358	while (--i >= 0)
359	if (blkg->pd[i])
360	blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
361	blk_put_queue(disk->queue);
362	out_free_iostat:
363	free_percpu(pdata: blkg->iostat_cpu);
364	out_exit_refcnt:
365	percpu_ref_exit(ref: &blkg->refcnt);
366	out_free_blkg:
367	kfree(objp: blkg);
368	return NULL;
369	}
370
371	/*
372	* If @new_blkg is %NULL, this function tries to allocate a new one as
373	* necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
374	*/
375	static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
376	struct blkcg_gq *new_blkg)
377	{
378	struct blkcg_gq *blkg;
379	int i, ret;
380
381	lockdep_assert_held(&disk->queue->queue_lock);
382
383	/* request_queue is dying, do not create/recreate a blkg */
384	if (blk_queue_dying(disk->queue)) {
385	ret = -ENODEV;
386	goto err_free_blkg;
387	}
388
389	/* blkg holds a reference to blkcg */
390	if (!css_tryget_online(css: &blkcg->css)) {
391	ret = -ENODEV;
392	goto err_free_blkg;
393	}
394
395	/* allocate */
396	if (!new_blkg) {
397	new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
398	if (unlikely(!new_blkg)) {
399	ret = -ENOMEM;
400	goto err_put_css;
401	}
402	}
403	blkg = new_blkg;
404
405	/* link parent */
406	if (blkcg_parent(blkcg)) {
407	blkg->parent = blkg_lookup(blkcg: blkcg_parent(blkcg), q: disk->queue);
408	if (WARN_ON_ONCE(!blkg->parent)) {
409	ret = -ENODEV;
410	goto err_put_css;
411	}
412	blkg_get(blkg: blkg->parent);
413	}
414
415	/* invoke per-policy init */
416	for (i = 0; i < BLKCG_MAX_POLS; i++) {
417	struct blkcg_policy *pol = blkcg_policy[i];
418
419	if (blkg->pd[i] && pol->pd_init_fn)
420	pol->pd_init_fn(blkg->pd[i]);
421	}
422
423	/* insert */
424	spin_lock(lock: &blkcg->lock);
425	ret = radix_tree_insert(&blkcg->blkg_tree, index: disk->queue->id, blkg);
426	if (likely(!ret)) {
427	hlist_add_head_rcu(n: &blkg->blkcg_node, h: &blkcg->blkg_list);
428	list_add(new: &blkg->q_node, head: &disk->queue->blkg_list);
429
430	for (i = 0; i < BLKCG_MAX_POLS; i++) {
431	struct blkcg_policy *pol = blkcg_policy[i];
432
433	if (blkg->pd[i]) {
434	if (pol->pd_online_fn)
435	pol->pd_online_fn(blkg->pd[i]);
436	blkg->pd[i]->online = true;
437	}
438	}
439	}
440	blkg->online = true;
441	spin_unlock(lock: &blkcg->lock);
442
443	if (!ret)
444	return blkg;
445
446	/* @blkg failed fully initialized, use the usual release path */
447	blkg_put(blkg);
448	return ERR_PTR(error: ret);
449
450	err_put_css:
451	css_put(css: &blkcg->css);
452	err_free_blkg:
453	if (new_blkg)
454	blkg_free(blkg: new_blkg);
455	return ERR_PTR(error: ret);
456	}
457
458	/**
459	* blkg_lookup_create - lookup blkg, try to create one if not there
460	* @blkcg: blkcg of interest
461	* @disk: gendisk of interest
462	*
463	* Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
464	* create one. blkg creation is performed recursively from blkcg_root such
465	* that all non-root blkg's have access to the parent blkg. This function
466	* should be called under RCU read lock and takes @disk->queue->queue_lock.
467	*
468	* Returns the blkg or the closest blkg if blkg_create() fails as it walks
469	* down from root.
470	*/
471	static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
472	struct gendisk *disk)
473	{
474	struct request_queue *q = disk->queue;
475	struct blkcg_gq *blkg;
476	unsigned long flags;
477
478	WARN_ON_ONCE(!rcu_read_lock_held());
479
480	blkg = blkg_lookup(blkcg, q);
481	if (blkg)
482	return blkg;
483
484	spin_lock_irqsave(&q->queue_lock, flags);
485	blkg = blkg_lookup(blkcg, q);
486	if (blkg) {
487	if (blkcg != &blkcg_root &&
488	blkg != rcu_dereference(blkcg->blkg_hint))
489	rcu_assign_pointer(blkcg->blkg_hint, blkg);
490	goto found;
491	}
492
493	/*
494	* Create blkgs walking down from blkcg_root to @blkcg, so that all
495	* non-root blkgs have access to their parents. Returns the closest
496	* blkg to the intended blkg should blkg_create() fail.
497	*/
498	while (true) {
499	struct blkcg *pos = blkcg;
500	struct blkcg *parent = blkcg_parent(blkcg);
501	struct blkcg_gq *ret_blkg = q->root_blkg;
502
503	while (parent) {
504	blkg = blkg_lookup(blkcg: parent, q);
505	if (blkg) {
506	/* remember closest blkg */
507	ret_blkg = blkg;
508	break;
509	}
510	pos = parent;
511	parent = blkcg_parent(blkcg: parent);
512	}
513
514	blkg = blkg_create(blkcg: pos, disk, NULL);
515	if (IS_ERR(ptr: blkg)) {
516	blkg = ret_blkg;
517	break;
518	}
519	if (pos == blkcg)
520	break;
521	}
522
523	found:
524	spin_unlock_irqrestore(lock: &q->queue_lock, flags);
525	return blkg;
526	}
527
528	static void blkg_destroy(struct blkcg_gq *blkg)
529	{
530	struct blkcg *blkcg = blkg->blkcg;
531	int i;
532
533	lockdep_assert_held(&blkg->q->queue_lock);
534	lockdep_assert_held(&blkcg->lock);
535
536	/*
537	* blkg stays on the queue list until blkg_free_workfn(), see details in
538	* blkg_free_workfn(), hence this function can be called from
539	* blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
540	* blkg_free_workfn().
541	*/
542	if (hlist_unhashed(h: &blkg->blkcg_node))
543	return;
544
545	for (i = 0; i < BLKCG_MAX_POLS; i++) {
546	struct blkcg_policy *pol = blkcg_policy[i];
547
548	if (blkg->pd[i] && blkg->pd[i]->online) {
549	blkg->pd[i]->online = false;
550	if (pol->pd_offline_fn)
551	pol->pd_offline_fn(blkg->pd[i]);
552	}
553	}
554
555	blkg->online = false;
556
557	radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
558	hlist_del_init_rcu(n: &blkg->blkcg_node);
559
560	/*
561	* Both setting lookup hint to and clearing it from @blkg are done
562	* under queue_lock. If it's not pointing to @blkg now, it never
563	* will. Hint assignment itself can race safely.
564	*/
565	if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
566	rcu_assign_pointer(blkcg->blkg_hint, NULL);
567
568	/*
569	* Put the reference taken at the time of creation so that when all
570	* queues are gone, group can be destroyed.
571	*/
572	percpu_ref_kill(ref: &blkg->refcnt);
573	}
574
575	static void blkg_destroy_all(struct gendisk *disk)
576	{
577	struct request_queue *q = disk->queue;
578	struct blkcg_gq *blkg;
579	int count = BLKG_DESTROY_BATCH_SIZE;
580	int i;
581
582	restart:
583	spin_lock_irq(lock: &q->queue_lock);
584	list_for_each_entry(blkg, &q->blkg_list, q_node) {
585	struct blkcg *blkcg = blkg->blkcg;
586
587	if (hlist_unhashed(h: &blkg->blkcg_node))
588	continue;
589
590	spin_lock(lock: &blkcg->lock);
591	blkg_destroy(blkg);
592	spin_unlock(lock: &blkcg->lock);
593
594	/*
595	* in order to avoid holding the spin lock for too long, release
596	* it when a batch of blkgs are destroyed.
597	*/
598	if (!(--count)) {
599	count = BLKG_DESTROY_BATCH_SIZE;
600	spin_unlock_irq(lock: &q->queue_lock);
601	cond_resched();
602	goto restart;
603	}
604	}
605
606	/*
607	* Mark policy deactivated since policy offline has been done, and
608	* the free is scheduled, so future blkcg_deactivate_policy() can
609	* be bypassed
610	*/
611	for (i = 0; i < BLKCG_MAX_POLS; i++) {
612	struct blkcg_policy *pol = blkcg_policy[i];
613
614	if (pol)
615	__clear_bit(pol->plid, q->blkcg_pols);
616	}
617
618	q->root_blkg = NULL;
619	spin_unlock_irq(lock: &q->queue_lock);
620	}
621
622	static int blkcg_reset_stats(struct cgroup_subsys_state *css,
623	struct cftype *cftype, u64 val)
624	{
625	struct blkcg *blkcg = css_to_blkcg(css);
626	struct blkcg_gq *blkg;
627	int i, cpu;
628
629	mutex_lock(&blkcg_pol_mutex);
630	spin_lock_irq(lock: &blkcg->lock);
631
632	/*
633	* Note that stat reset is racy - it doesn't synchronize against
634	* stat updates. This is a debug feature which shouldn't exist
635	* anyway. If you get hit by a race, retry.
636	*/
637	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
638	for_each_possible_cpu(cpu) {
639	struct blkg_iostat_set *bis =
640	per_cpu_ptr(blkg->iostat_cpu, cpu);
641	memset(bis, 0, sizeof(*bis));
642
643	/* Re-initialize the cleared blkg_iostat_set */
644	u64_stats_init(syncp: &bis->sync);
645	bis->blkg = blkg;
646	}
647	memset(&blkg->iostat, 0, sizeof(blkg->iostat));
648	u64_stats_init(syncp: &blkg->iostat.sync);
649
650	for (i = 0; i < BLKCG_MAX_POLS; i++) {
651	struct blkcg_policy *pol = blkcg_policy[i];
652
653	if (blkg->pd[i] && pol->pd_reset_stats_fn)
654	pol->pd_reset_stats_fn(blkg->pd[i]);
655	}
656	}
657
658	spin_unlock_irq(lock: &blkcg->lock);
659	mutex_unlock(lock: &blkcg_pol_mutex);
660	return 0;
661	}
662
663	const char *blkg_dev_name(struct blkcg_gq *blkg)
664	{
665	if (!blkg->q->disk)
666	return NULL;
667	return bdi_dev_name(bdi: blkg->q->disk->bdi);
668	}
669
670	/**
671	* blkcg_print_blkgs - helper for printing per-blkg data
672	* @sf: seq_file to print to
673	* @blkcg: blkcg of interest
674	* @prfill: fill function to print out a blkg
675	* @pol: policy in question
676	* @data: data to be passed to @prfill
677	* @show_total: to print out sum of prfill return values or not
678	*
679	* This function invokes @prfill on each blkg of @blkcg if pd for the
680	* policy specified by @pol exists. @prfill is invoked with @sf, the
681	* policy data and @data and the matching queue lock held. If @show_total
682	* is %true, the sum of the return values from @prfill is printed with
683	* "Total" label at the end.
684	*
685	* This is to be used to construct print functions for
686	* cftype->read_seq_string method.
687	*/
688	void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
689	u64 (*prfill)(struct seq_file *,
690	struct blkg_policy_data *, int),
691	const struct blkcg_policy *pol, int data,
692	bool show_total)
693	{
694	struct blkcg_gq *blkg;
695	u64 total = 0;
696
697	rcu_read_lock();
698	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
699	spin_lock_irq(lock: &blkg->q->queue_lock);
700	if (blkcg_policy_enabled(q: blkg->q, pol))
701	total += prfill(sf, blkg->pd[pol->plid], data);
702	spin_unlock_irq(lock: &blkg->q->queue_lock);
703	}
704	rcu_read_unlock();
705
706	if (show_total)
707	seq_printf(m: sf, fmt: "Total %llu\n", (unsigned long long)total);
708	}
709	EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
710
711	/**
712	* __blkg_prfill_u64 - prfill helper for a single u64 value
713	* @sf: seq_file to print to
714	* @pd: policy private data of interest
715	* @v: value to print
716	*
717	* Print @v to @sf for the device associated with @pd.
718	*/
719	u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
720	{
721	const char *dname = blkg_dev_name(blkg: pd->blkg);
722
723	if (!dname)
724	return 0;
725
726	seq_printf(m: sf, fmt: "%s %llu\n", dname, (unsigned long long)v);
727	return v;
728	}
729	EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
730
731	/**
732	* blkg_conf_init - initialize a blkg_conf_ctx
733	* @ctx: blkg_conf_ctx to initialize
734	* @input: input string
735	*
736	* Initialize @ctx which can be used to parse blkg config input string @input.
737	* Once initialized, @ctx can be used with blkg_conf_open_bdev() and
738	* blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
739	*/
740	void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
741	{
742	*ctx = (struct blkg_conf_ctx){ .input = input };
743	}
744	EXPORT_SYMBOL_GPL(blkg_conf_init);
745
746	/**
747	* blkg_conf_open_bdev - parse and open bdev for per-blkg config update
748	* @ctx: blkg_conf_ctx initialized with blkg_conf_init()
749	*
750	* Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
751	* @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
752	* set to point past the device node prefix.
753	*
754	* This function may be called multiple times on @ctx and the extra calls become
755	* NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
756	* explicitly if bdev access is needed without resolving the blkcg / policy part
757	* of @ctx->input. Returns -errno on error.
758	*/
759	int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
760	{
761	char *input = ctx->input;
762	unsigned int major, minor;
763	struct block_device *bdev;
764	int key_len;
765
766	if (ctx->bdev)
767	return 0;
768
769	if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
770	return -EINVAL;
771
772	input += key_len;
773	if (!isspace(*input))
774	return -EINVAL;
775	input = skip_spaces(input);
776
777	bdev = blkdev_get_no_open(MKDEV(major, minor));
778	if (!bdev)
779	return -ENODEV;
780	if (bdev_is_partition(bdev)) {
781	blkdev_put_no_open(bdev);
782	return -ENODEV;
783	}
784
785	mutex_lock(&bdev->bd_queue->rq_qos_mutex);
786	if (!disk_live(disk: bdev->bd_disk)) {
787	blkdev_put_no_open(bdev);
788	mutex_unlock(lock: &bdev->bd_queue->rq_qos_mutex);
789	return -ENODEV;
790	}
791
792	ctx->body = input;
793	ctx->bdev = bdev;
794	return 0;
795	}
796
797	/**
798	* blkg_conf_prep - parse and prepare for per-blkg config update
799	* @blkcg: target block cgroup
800	* @pol: target policy
801	* @ctx: blkg_conf_ctx initialized with blkg_conf_init()
802	*
803	* Parse per-blkg config update from @ctx->input and initialize @ctx
804	* accordingly. On success, @ctx->body points to the part of @ctx->input
805	* following MAJ:MIN, @ctx->bdev points to the target block device and
806	* @ctx->blkg to the blkg being configured.
807	*
808	* blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
809	* function returns with queue lock held and must be followed by
810	* blkg_conf_exit().
811	*/
812	int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
813	struct blkg_conf_ctx *ctx)
814	__acquires(&bdev->bd_queue->queue_lock)
815	{
816	struct gendisk *disk;
817	struct request_queue *q;
818	struct blkcg_gq *blkg;
819	int ret;
820
821	ret = blkg_conf_open_bdev(ctx);
822	if (ret)
823	return ret;
824
825	disk = ctx->bdev->bd_disk;
826	q = disk->queue;
827
828	/*
829	* blkcg_deactivate_policy() requires queue to be frozen, we can grab
830	* q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
831	*/
832	ret = blk_queue_enter(q, flags: 0);
833	if (ret)
834	goto fail;
835
836	spin_lock_irq(lock: &q->queue_lock);
837
838	if (!blkcg_policy_enabled(q, pol)) {
839	ret = -EOPNOTSUPP;
840	goto fail_unlock;
841	}
842
843	blkg = blkg_lookup(blkcg, q);
844	if (blkg)
845	goto success;
846
847	/*
848	* Create blkgs walking down from blkcg_root to @blkcg, so that all
849	* non-root blkgs have access to their parents.
850	*/
851	while (true) {
852	struct blkcg *pos = blkcg;
853	struct blkcg *parent;
854	struct blkcg_gq *new_blkg;
855
856	parent = blkcg_parent(blkcg);
857	while (parent && !blkg_lookup(blkcg: parent, q)) {
858	pos = parent;
859	parent = blkcg_parent(blkcg: parent);
860	}
861
862	/* Drop locks to do new blkg allocation with GFP_KERNEL. */
863	spin_unlock_irq(lock: &q->queue_lock);
864
865	new_blkg = blkg_alloc(blkcg: pos, disk, GFP_KERNEL);
866	if (unlikely(!new_blkg)) {
867	ret = -ENOMEM;
868	goto fail_exit_queue;
869	}
870
871	if (radix_tree_preload(GFP_KERNEL)) {
872	blkg_free(blkg: new_blkg);
873	ret = -ENOMEM;
874	goto fail_exit_queue;
875	}
876
877	spin_lock_irq(lock: &q->queue_lock);
878
879	if (!blkcg_policy_enabled(q, pol)) {
880	blkg_free(blkg: new_blkg);
881	ret = -EOPNOTSUPP;
882	goto fail_preloaded;
883	}
884
885	blkg = blkg_lookup(blkcg: pos, q);
886	if (blkg) {
887	blkg_free(blkg: new_blkg);
888	} else {
889	blkg = blkg_create(blkcg: pos, disk, new_blkg);
890	if (IS_ERR(ptr: blkg)) {
891	ret = PTR_ERR(ptr: blkg);
892	goto fail_preloaded;
893	}
894	}
895
896	radix_tree_preload_end();
897
898	if (pos == blkcg)
899	goto success;
900	}
901	success:
902	blk_queue_exit(q);
903	ctx->blkg = blkg;
904	return 0;
905
906	fail_preloaded:
907	radix_tree_preload_end();
908	fail_unlock:
909	spin_unlock_irq(lock: &q->queue_lock);
910	fail_exit_queue:
911	blk_queue_exit(q);
912	fail:
913	/*
914	* If queue was bypassing, we should retry. Do so after a
915	* short msleep(). It isn't strictly necessary but queue
916	* can be bypassing for some time and it's always nice to
917	* avoid busy looping.
918	*/
919	if (ret == -EBUSY) {
920	msleep(msecs: 10);
921	ret = restart_syscall();
922	}
923	return ret;
924	}
925	EXPORT_SYMBOL_GPL(blkg_conf_prep);
926
927	/**
928	* blkg_conf_exit - clean up per-blkg config update
929	* @ctx: blkg_conf_ctx initialized with blkg_conf_init()
930	*
931	* Clean up after per-blkg config update. This function must be called on all
932	* blkg_conf_ctx's initialized with blkg_conf_init().
933	*/
934	void blkg_conf_exit(struct blkg_conf_ctx *ctx)
935	__releases(&ctx->bdev->bd_queue->queue_lock)
936	__releases(&ctx->bdev->bd_queue->rq_qos_mutex)
937	{
938	if (ctx->blkg) {
939	spin_unlock_irq(lock: &bdev_get_queue(bdev: ctx->bdev)->queue_lock);
940	ctx->blkg = NULL;
941	}
942
943	if (ctx->bdev) {
944	mutex_unlock(lock: &ctx->bdev->bd_queue->rq_qos_mutex);
945	blkdev_put_no_open(bdev: ctx->bdev);
946	ctx->body = NULL;
947	ctx->bdev = NULL;
948	}
949	}
950	EXPORT_SYMBOL_GPL(blkg_conf_exit);
951
952	static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
953	{
954	int i;
955
956	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
957	dst->bytes[i] = src->bytes[i];
958	dst->ios[i] = src->ios[i];
959	}
960	}
961
962	static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
963	{
964	int i;
965
966	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
967	dst->bytes[i] += src->bytes[i];
968	dst->ios[i] += src->ios[i];
969	}
970	}
971
972	static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
973	{
974	int i;
975
976	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
977	dst->bytes[i] -= src->bytes[i];
978	dst->ios[i] -= src->ios[i];
979	}
980	}
981
982	static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
983	struct blkg_iostat *last)
984	{
985	struct blkg_iostat delta;
986	unsigned long flags;
987
988	/* propagate percpu delta to global */
989	flags = u64_stats_update_begin_irqsave(syncp: &blkg->iostat.sync);
990	blkg_iostat_set(dst: &delta, src: cur);
991	blkg_iostat_sub(dst: &delta, src: last);
992	blkg_iostat_add(dst: &blkg->iostat.cur, src: &delta);
993	blkg_iostat_add(dst: last, src: &delta);
994	u64_stats_update_end_irqrestore(syncp: &blkg->iostat.sync, flags);
995	}
996
997	static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
998	{
999	struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
1000	struct llist_node *lnode;
1001	struct blkg_iostat_set *bisc, *next_bisc;
1002	unsigned long flags;
1003
1004	rcu_read_lock();
1005
1006	lnode = llist_del_all(head: lhead);
1007	if (!lnode)
1008	goto out;
1009
1010	/*
1011	* For covering concurrent parent blkg update from blkg_release().
1012	*
1013	* When flushing from cgroup, cgroup_rstat_lock is always held, so
1014	* this lock won't cause contention most of time.
1015	*/
1016	raw_spin_lock_irqsave(&blkg_stat_lock, flags);
1017
1018	/*
1019	* Iterate only the iostat_cpu's queued in the lockless list.
1020	*/
1021	llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
1022	struct blkcg_gq *blkg = bisc->blkg;
1023	struct blkcg_gq *parent = blkg->parent;
1024	struct blkg_iostat cur;
1025	unsigned int seq;
1026
1027	WRITE_ONCE(bisc->lqueued, false);
1028
1029	/* fetch the current per-cpu values */
1030	do {
1031	seq = u64_stats_fetch_begin(syncp: &bisc->sync);
1032	blkg_iostat_set(dst: &cur, src: &bisc->cur);
1033	} while (u64_stats_fetch_retry(syncp: &bisc->sync, start: seq));
1034
1035	blkcg_iostat_update(blkg, cur: &cur, last: &bisc->last);
1036
1037	/* propagate global delta to parent (unless that's root) */
1038	if (parent && parent->parent)
1039	blkcg_iostat_update(blkg: parent, cur: &blkg->iostat.cur,
1040	last: &blkg->iostat.last);
1041	}
1042	raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
1043	out:
1044	rcu_read_unlock();
1045	}
1046
1047	static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1048	{
1049	/* Root-level stats are sourced from system-wide IO stats */
1050	if (cgroup_parent(cgrp: css->cgroup))
1051	__blkcg_rstat_flush(blkcg: css_to_blkcg(css), cpu);
1052	}
1053
1054	/*
1055	* We source root cgroup stats from the system-wide stats to avoid
1056	* tracking the same information twice and incurring overhead when no
1057	* cgroups are defined. For that reason, cgroup_rstat_flush in
1058	* blkcg_print_stat does not actually fill out the iostat in the root
1059	* cgroup's blkcg_gq.
1060	*
1061	* However, we would like to re-use the printing code between the root and
1062	* non-root cgroups to the extent possible. For that reason, we simulate
1063	* flushing the root cgroup's stats by explicitly filling in the iostat
1064	* with disk level statistics.
1065	*/
1066	static void blkcg_fill_root_iostats(void)
1067	{
1068	struct class_dev_iter iter;
1069	struct device *dev;
1070
1071	class_dev_iter_init(iter: &iter, class: &block_class, NULL, type: &disk_type);
1072	while ((dev = class_dev_iter_next(iter: &iter))) {
1073	struct block_device *bdev = dev_to_bdev(dev);
1074	struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
1075	struct blkg_iostat tmp;
1076	int cpu;
1077	unsigned long flags;
1078
1079	memset(&tmp, 0, sizeof(tmp));
1080	for_each_possible_cpu(cpu) {
1081	struct disk_stats *cpu_dkstats;
1082
1083	cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1084	tmp.ios[BLKG_IOSTAT_READ] +=
1085	cpu_dkstats->ios[STAT_READ];
1086	tmp.ios[BLKG_IOSTAT_WRITE] +=
1087	cpu_dkstats->ios[STAT_WRITE];
1088	tmp.ios[BLKG_IOSTAT_DISCARD] +=
1089	cpu_dkstats->ios[STAT_DISCARD];
1090	// convert sectors to bytes
1091	tmp.bytes[BLKG_IOSTAT_READ] +=
1092	cpu_dkstats->sectors[STAT_READ] << 9;
1093	tmp.bytes[BLKG_IOSTAT_WRITE] +=
1094	cpu_dkstats->sectors[STAT_WRITE] << 9;
1095	tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1096	cpu_dkstats->sectors[STAT_DISCARD] << 9;
1097	}
1098
1099	flags = u64_stats_update_begin_irqsave(syncp: &blkg->iostat.sync);
1100	blkg_iostat_set(dst: &blkg->iostat.cur, src: &tmp);
1101	u64_stats_update_end_irqrestore(syncp: &blkg->iostat.sync, flags);
1102	}
1103	}
1104
1105	static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1106	{
1107	struct blkg_iostat_set *bis = &blkg->iostat;
1108	u64 rbytes, wbytes, rios, wios, dbytes, dios;
1109	const char *dname;
1110	unsigned seq;
1111	int i;
1112
1113	if (!blkg->online)
1114	return;
1115
1116	dname = blkg_dev_name(blkg);
1117	if (!dname)
1118	return;
1119
1120	seq_printf(m: s, fmt: "%s ", dname);
1121
1122	do {
1123	seq = u64_stats_fetch_begin(syncp: &bis->sync);
1124
1125	rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1126	wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1127	dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1128	rios = bis->cur.ios[BLKG_IOSTAT_READ];
1129	wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1130	dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1131	} while (u64_stats_fetch_retry(syncp: &bis->sync, start: seq));
1132
1133	if (rbytes || wbytes || rios || wios) {
1134	seq_printf(m: s, fmt: "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1135	rbytes, wbytes, rios, wios,
1136	dbytes, dios);
1137	}
1138
1139	if (blkcg_debug_stats && atomic_read(v: &blkg->use_delay)) {
1140	seq_printf(m: s, fmt: " use_delay=%d delay_nsec=%llu",
1141	atomic_read(v: &blkg->use_delay),
1142	atomic64_read(v: &blkg->delay_nsec));
1143	}
1144
1145	for (i = 0; i < BLKCG_MAX_POLS; i++) {
1146	struct blkcg_policy *pol = blkcg_policy[i];
1147
1148	if (!blkg->pd[i] || !pol->pd_stat_fn)
1149	continue;
1150
1151	pol->pd_stat_fn(blkg->pd[i], s);
1152	}
1153
1154	seq_puts(m: s, s: "\n");
1155	}
1156
1157	static int blkcg_print_stat(struct seq_file *sf, void *v)
1158	{
1159	struct blkcg *blkcg = css_to_blkcg(css: seq_css(seq: sf));
1160	struct blkcg_gq *blkg;
1161
1162	if (!seq_css(seq: sf)->parent)
1163	blkcg_fill_root_iostats();
1164	else
1165	cgroup_rstat_flush(cgrp: blkcg->css.cgroup);
1166
1167	rcu_read_lock();
1168	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1169	spin_lock_irq(lock: &blkg->q->queue_lock);
1170	blkcg_print_one_stat(blkg, s: sf);
1171	spin_unlock_irq(lock: &blkg->q->queue_lock);
1172	}
1173	rcu_read_unlock();
1174	return 0;
1175	}
1176
1177	static struct cftype blkcg_files[] = {
1178	{
1179	.name = "stat",
1180	.seq_show = blkcg_print_stat,
1181	},
1182	{ } /* terminate */
1183	};
1184
1185	static struct cftype blkcg_legacy_files[] = {
1186	{
1187	.name = "reset_stats",
1188	.write_u64 = blkcg_reset_stats,
1189	},
1190	{ } /* terminate */
1191	};
1192
1193	#ifdef CONFIG_CGROUP_WRITEBACK
1194	struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1195	{
1196	return &css_to_blkcg(css)->cgwb_list;
1197	}
1198	#endif
1199
1200	/*
1201	* blkcg destruction is a three-stage process.
1202	*
1203	* 1. Destruction starts. The blkcg_css_offline() callback is invoked
1204	* which offlines writeback. Here we tie the next stage of blkg destruction
1205	* to the completion of writeback associated with the blkcg. This lets us
1206	* avoid punting potentially large amounts of outstanding writeback to root
1207	* while maintaining any ongoing policies. The next stage is triggered when
1208	* the nr_cgwbs count goes to zero.
1209	*
1210	* 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1211	* and handles the destruction of blkgs. Here the css reference held by
1212	* the blkg is put back eventually allowing blkcg_css_free() to be called.
1213	* This work may occur in cgwb_release_workfn() on the cgwb_release
1214	* workqueue. Any submitted ios that fail to get the blkg ref will be
1215	* punted to the root_blkg.
1216	*
1217	* 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1218	* This finally frees the blkcg.
1219	*/
1220
1221	/**
1222	* blkcg_destroy_blkgs - responsible for shooting down blkgs
1223	* @blkcg: blkcg of interest
1224	*
1225	* blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1226	* is nested inside q lock, this function performs reverse double lock dancing.
1227	* Destroying the blkgs releases the reference held on the blkcg's css allowing
1228	* blkcg_css_free to eventually be called.
1229	*
1230	* This is the blkcg counterpart of ioc_release_fn().
1231	*/
1232	static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1233	{
1234	might_sleep();
1235
1236	spin_lock_irq(lock: &blkcg->lock);
1237
1238	while (!hlist_empty(h: &blkcg->blkg_list)) {
1239	struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1240	struct blkcg_gq, blkcg_node);
1241	struct request_queue *q = blkg->q;
1242
1243	if (need_resched() || !spin_trylock(lock: &q->queue_lock)) {
1244	/*
1245	* Given that the system can accumulate a huge number
1246	* of blkgs in pathological cases, check to see if we
1247	* need to rescheduling to avoid softlockup.
1248	*/
1249	spin_unlock_irq(lock: &blkcg->lock);
1250	cond_resched();
1251	spin_lock_irq(lock: &blkcg->lock);
1252	continue;
1253	}
1254
1255	blkg_destroy(blkg);
1256	spin_unlock(lock: &q->queue_lock);
1257	}
1258
1259	spin_unlock_irq(lock: &blkcg->lock);
1260	}
1261
1262	/**
1263	* blkcg_pin_online - pin online state
1264	* @blkcg_css: blkcg of interest
1265	*
1266	* While pinned, a blkcg is kept online. This is primarily used to
1267	* impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1268	* while an associated cgwb is still active.
1269	*/
1270	void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1271	{
1272	refcount_inc(r: &css_to_blkcg(css: blkcg_css)->online_pin);
1273	}
1274
1275	/**
1276	* blkcg_unpin_online - unpin online state
1277	* @blkcg_css: blkcg of interest
1278	*
1279	* This is primarily used to impedance-match blkg and cgwb lifetimes so
1280	* that blkg doesn't go offline while an associated cgwb is still active.
1281	* When this count goes to zero, all active cgwbs have finished so the
1282	* blkcg can continue destruction by calling blkcg_destroy_blkgs().
1283	*/
1284	void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1285	{
1286	struct blkcg *blkcg = css_to_blkcg(css: blkcg_css);
1287
1288	do {
1289	if (!refcount_dec_and_test(r: &blkcg->online_pin))
1290	break;
1291	blkcg_destroy_blkgs(blkcg);
1292	blkcg = blkcg_parent(blkcg);
1293	} while (blkcg);
1294	}
1295
1296	/**
1297	* blkcg_css_offline - cgroup css_offline callback
1298	* @css: css of interest
1299	*
1300	* This function is called when @css is about to go away. Here the cgwbs are
1301	* offlined first and only once writeback associated with the blkcg has
1302	* finished do we start step 2 (see above).
1303	*/
1304	static void blkcg_css_offline(struct cgroup_subsys_state *css)
1305	{
1306	/* this prevents anyone from attaching or migrating to this blkcg */
1307	wb_blkcg_offline(css);
1308
1309	/* put the base online pin allowing step 2 to be triggered */
1310	blkcg_unpin_online(blkcg_css: css);
1311	}
1312
1313	static void blkcg_css_free(struct cgroup_subsys_state *css)
1314	{
1315	struct blkcg *blkcg = css_to_blkcg(css);
1316	int i;
1317
1318	mutex_lock(&blkcg_pol_mutex);
1319
1320	list_del(entry: &blkcg->all_blkcgs_node);
1321
1322	for (i = 0; i < BLKCG_MAX_POLS; i++)
1323	if (blkcg->cpd[i])
1324	blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1325
1326	mutex_unlock(lock: &blkcg_pol_mutex);
1327
1328	free_percpu(pdata: blkcg->lhead);
1329	kfree(objp: blkcg);
1330	}
1331
1332	static struct cgroup_subsys_state *
1333	blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1334	{
1335	struct blkcg *blkcg;
1336	int i;
1337
1338	mutex_lock(&blkcg_pol_mutex);
1339
1340	if (!parent_css) {
1341	blkcg = &blkcg_root;
1342	} else {
1343	blkcg = kzalloc(size: sizeof(*blkcg), GFP_KERNEL);
1344	if (!blkcg)
1345	goto unlock;
1346	}
1347
1348	if (init_blkcg_llists(blkcg))
1349	goto free_blkcg;
1350
1351	for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1352	struct blkcg_policy *pol = blkcg_policy[i];
1353	struct blkcg_policy_data *cpd;
1354
1355	/*
1356	* If the policy hasn't been attached yet, wait for it
1357	* to be attached before doing anything else. Otherwise,
1358	* check if the policy requires any specific per-cgroup
1359	* data: if it does, allocate and initialize it.
1360	*/
1361	if (!pol || !pol->cpd_alloc_fn)
1362	continue;
1363
1364	cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1365	if (!cpd)
1366	goto free_pd_blkcg;
1367
1368	blkcg->cpd[i] = cpd;
1369	cpd->blkcg = blkcg;
1370	cpd->plid = i;
1371	}
1372
1373	spin_lock_init(&blkcg->lock);
1374	refcount_set(r: &blkcg->online_pin, n: 1);
1375	INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1376	INIT_HLIST_HEAD(&blkcg->blkg_list);
1377	#ifdef CONFIG_CGROUP_WRITEBACK
1378	INIT_LIST_HEAD(list: &blkcg->cgwb_list);
1379	#endif
1380	list_add_tail(new: &blkcg->all_blkcgs_node, head: &all_blkcgs);
1381
1382	mutex_unlock(lock: &blkcg_pol_mutex);
1383	return &blkcg->css;
1384
1385	free_pd_blkcg:
1386	for (i--; i >= 0; i--)
1387	if (blkcg->cpd[i])
1388	blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1389	free_percpu(pdata: blkcg->lhead);
1390	free_blkcg:
1391	if (blkcg != &blkcg_root)
1392	kfree(objp: blkcg);
1393	unlock:
1394	mutex_unlock(lock: &blkcg_pol_mutex);
1395	return ERR_PTR(error: -ENOMEM);
1396	}
1397
1398	static int blkcg_css_online(struct cgroup_subsys_state *css)
1399	{
1400	struct blkcg *parent = blkcg_parent(blkcg: css_to_blkcg(css));
1401
1402	/*
1403	* blkcg_pin_online() is used to delay blkcg offline so that blkgs
1404	* don't go offline while cgwbs are still active on them. Pin the
1405	* parent so that offline always happens towards the root.
1406	*/
1407	if (parent)
1408	blkcg_pin_online(blkcg_css: &parent->css);
1409	return 0;
1410	}
1411
1412	void blkg_init_queue(struct request_queue *q)
1413	{
1414	INIT_LIST_HEAD(list: &q->blkg_list);
1415	mutex_init(&q->blkcg_mutex);
1416	}
1417
1418	int blkcg_init_disk(struct gendisk *disk)
1419	{
1420	struct request_queue *q = disk->queue;
1421	struct blkcg_gq *new_blkg, *blkg;
1422	bool preloaded;
1423	int ret;
1424
1425	new_blkg = blkg_alloc(blkcg: &blkcg_root, disk, GFP_KERNEL);
1426	if (!new_blkg)
1427	return -ENOMEM;
1428
1429	preloaded = !radix_tree_preload(GFP_KERNEL);
1430
1431	/* Make sure the root blkg exists. */
1432	/* spin_lock_irq can serve as RCU read-side critical section. */
1433	spin_lock_irq(lock: &q->queue_lock);
1434	blkg = blkg_create(blkcg: &blkcg_root, disk, new_blkg);
1435	if (IS_ERR(ptr: blkg))
1436	goto err_unlock;
1437	q->root_blkg = blkg;
1438	spin_unlock_irq(lock: &q->queue_lock);
1439
1440	if (preloaded)
1441	radix_tree_preload_end();
1442
1443	ret = blk_ioprio_init(disk);
1444	if (ret)
1445	goto err_destroy_all;
1446
1447	ret = blk_throtl_init(disk);
1448	if (ret)
1449	goto err_ioprio_exit;
1450
1451	return 0;
1452
1453	err_ioprio_exit:
1454	blk_ioprio_exit(disk);
1455	err_destroy_all:
1456	blkg_destroy_all(disk);
1457	return ret;
1458	err_unlock:
1459	spin_unlock_irq(lock: &q->queue_lock);
1460	if (preloaded)
1461	radix_tree_preload_end();
1462	return PTR_ERR(ptr: blkg);
1463	}
1464
1465	void blkcg_exit_disk(struct gendisk *disk)
1466	{
1467	blkg_destroy_all(disk);
1468	blk_throtl_exit(disk);
1469	}
1470
1471	static void blkcg_exit(struct task_struct *tsk)
1472	{
1473	if (tsk->throttle_disk)
1474	put_disk(disk: tsk->throttle_disk);
1475	tsk->throttle_disk = NULL;
1476	}
1477
1478	struct cgroup_subsys io_cgrp_subsys = {
1479	.css_alloc = blkcg_css_alloc,
1480	.css_online = blkcg_css_online,
1481	.css_offline = blkcg_css_offline,
1482	.css_free = blkcg_css_free,
1483	.css_rstat_flush = blkcg_rstat_flush,
1484	.dfl_cftypes = blkcg_files,
1485	.legacy_cftypes = blkcg_legacy_files,
1486	.legacy_name = "blkio",
1487	.exit = blkcg_exit,
1488	#ifdef CONFIG_MEMCG
1489	/*
1490	* This ensures that, if available, memcg is automatically enabled
1491	* together on the default hierarchy so that the owner cgroup can
1492	* be retrieved from writeback pages.
1493	*/
1494	.depends_on = 1 << memory_cgrp_id,
1495	#endif
1496	};
1497	EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1498
1499	/**
1500	* blkcg_activate_policy - activate a blkcg policy on a gendisk
1501	* @disk: gendisk of interest
1502	* @pol: blkcg policy to activate
1503	*
1504	* Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1505	* bypass mode to populate its blkgs with policy_data for @pol.
1506	*
1507	* Activation happens with @disk bypassed, so nobody would be accessing blkgs
1508	* from IO path. Update of each blkg is protected by both queue and blkcg
1509	* locks so that holding either lock and testing blkcg_policy_enabled() is
1510	* always enough for dereferencing policy data.
1511	*
1512	* The caller is responsible for synchronizing [de]activations and policy
1513	* [un]registerations. Returns 0 on success, -errno on failure.
1514	*/
1515	int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1516	{
1517	struct request_queue *q = disk->queue;
1518	struct blkg_policy_data *pd_prealloc = NULL;
1519	struct blkcg_gq *blkg, *pinned_blkg = NULL;
1520	int ret;
1521
1522	if (blkcg_policy_enabled(q, pol))
1523	return 0;
1524
1525	if (queue_is_mq(q))
1526	blk_mq_freeze_queue(q);
1527	retry:
1528	spin_lock_irq(lock: &q->queue_lock);
1529
1530	/* blkg_list is pushed at the head, reverse walk to initialize parents first */
1531	list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1532	struct blkg_policy_data *pd;
1533
1534	if (blkg->pd[pol->plid])
1535	continue;
1536
1537	/* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1538	if (blkg == pinned_blkg) {
1539	pd = pd_prealloc;
1540	pd_prealloc = NULL;
1541	} else {
1542	pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1543	GFP_NOWAIT | __GFP_NOWARN);
1544	}
1545
1546	if (!pd) {
1547	/*
1548	* GFP_NOWAIT failed. Free the existing one and
1549	* prealloc for @blkg w/ GFP_KERNEL.
1550	*/
1551	if (pinned_blkg)
1552	blkg_put(blkg: pinned_blkg);
1553	blkg_get(blkg);
1554	pinned_blkg = blkg;
1555
1556	spin_unlock_irq(lock: &q->queue_lock);
1557
1558	if (pd_prealloc)
1559	pol->pd_free_fn(pd_prealloc);
1560	pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1561	GFP_KERNEL);
1562	if (pd_prealloc)
1563	goto retry;
1564	else
1565	goto enomem;
1566	}
1567
1568	spin_lock(lock: &blkg->blkcg->lock);
1569
1570	pd->blkg = blkg;
1571	pd->plid = pol->plid;
1572	blkg->pd[pol->plid] = pd;
1573
1574	if (pol->pd_init_fn)
1575	pol->pd_init_fn(pd);
1576
1577	if (pol->pd_online_fn)
1578	pol->pd_online_fn(pd);
1579	pd->online = true;
1580
1581	spin_unlock(lock: &blkg->blkcg->lock);
1582	}
1583
1584	__set_bit(pol->plid, q->blkcg_pols);
1585	ret = 0;
1586
1587	spin_unlock_irq(lock: &q->queue_lock);
1588	out:
1589	if (queue_is_mq(q))
1590	blk_mq_unfreeze_queue(q);
1591	if (pinned_blkg)
1592	blkg_put(blkg: pinned_blkg);
1593	if (pd_prealloc)
1594	pol->pd_free_fn(pd_prealloc);
1595	return ret;
1596
1597	enomem:
1598	/* alloc failed, take down everything */
1599	spin_lock_irq(lock: &q->queue_lock);
1600	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1601	struct blkcg *blkcg = blkg->blkcg;
1602	struct blkg_policy_data *pd;
1603
1604	spin_lock(lock: &blkcg->lock);
1605	pd = blkg->pd[pol->plid];
1606	if (pd) {
1607	if (pd->online && pol->pd_offline_fn)
1608	pol->pd_offline_fn(pd);
1609	pd->online = false;
1610	pol->pd_free_fn(pd);
1611	blkg->pd[pol->plid] = NULL;
1612	}
1613	spin_unlock(lock: &blkcg->lock);
1614	}
1615	spin_unlock_irq(lock: &q->queue_lock);
1616	ret = -ENOMEM;
1617	goto out;
1618	}
1619	EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1620
1621	/**
1622	* blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1623	* @disk: gendisk of interest
1624	* @pol: blkcg policy to deactivate
1625	*
1626	* Deactivate @pol on @disk. Follows the same synchronization rules as
1627	* blkcg_activate_policy().
1628	*/
1629	void blkcg_deactivate_policy(struct gendisk *disk,
1630	const struct blkcg_policy *pol)
1631	{
1632	struct request_queue *q = disk->queue;
1633	struct blkcg_gq *blkg;
1634
1635	if (!blkcg_policy_enabled(q, pol))
1636	return;
1637
1638	if (queue_is_mq(q))
1639	blk_mq_freeze_queue(q);
1640
1641	mutex_lock(&q->blkcg_mutex);
1642	spin_lock_irq(lock: &q->queue_lock);
1643
1644	__clear_bit(pol->plid, q->blkcg_pols);
1645
1646	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1647	struct blkcg *blkcg = blkg->blkcg;
1648
1649	spin_lock(lock: &blkcg->lock);
1650	if (blkg->pd[pol->plid]) {
1651	if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1652	pol->pd_offline_fn(blkg->pd[pol->plid]);
1653	pol->pd_free_fn(blkg->pd[pol->plid]);
1654	blkg->pd[pol->plid] = NULL;
1655	}
1656	spin_unlock(lock: &blkcg->lock);
1657	}
1658
1659	spin_unlock_irq(lock: &q->queue_lock);
1660	mutex_unlock(lock: &q->blkcg_mutex);
1661
1662	if (queue_is_mq(q))
1663	blk_mq_unfreeze_queue(q);
1664	}
1665	EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1666
1667	static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1668	{
1669	struct blkcg *blkcg;
1670
1671	list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1672	if (blkcg->cpd[pol->plid]) {
1673	pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1674	blkcg->cpd[pol->plid] = NULL;
1675	}
1676	}
1677	}
1678
1679	/**
1680	* blkcg_policy_register - register a blkcg policy
1681	* @pol: blkcg policy to register
1682	*
1683	* Register @pol with blkcg core. Might sleep and @pol may be modified on
1684	* successful registration. Returns 0 on success and -errno on failure.
1685	*/
1686	int blkcg_policy_register(struct blkcg_policy *pol)
1687	{
1688	struct blkcg *blkcg;
1689	int i, ret;
1690
1691	mutex_lock(&blkcg_pol_register_mutex);
1692	mutex_lock(&blkcg_pol_mutex);
1693
1694	/* find an empty slot */
1695	ret = -ENOSPC;
1696	for (i = 0; i < BLKCG_MAX_POLS; i++)
1697	if (!blkcg_policy[i])
1698	break;
1699	if (i >= BLKCG_MAX_POLS) {
1700	pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1701	goto err_unlock;
1702	}
1703
1704	/* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1705	if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1706	(!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1707	goto err_unlock;
1708
1709	/* register @pol */
1710	pol->plid = i;
1711	blkcg_policy[pol->plid] = pol;
1712
1713	/* allocate and install cpd's */
1714	if (pol->cpd_alloc_fn) {
1715	list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1716	struct blkcg_policy_data *cpd;
1717
1718	cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1719	if (!cpd)
1720	goto err_free_cpds;
1721
1722	blkcg->cpd[pol->plid] = cpd;
1723	cpd->blkcg = blkcg;
1724	cpd->plid = pol->plid;
1725	}
1726	}
1727
1728	mutex_unlock(lock: &blkcg_pol_mutex);
1729
1730	/* everything is in place, add intf files for the new policy */
1731	if (pol->dfl_cftypes)
1732	WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1733	pol->dfl_cftypes));
1734	if (pol->legacy_cftypes)
1735	WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1736	pol->legacy_cftypes));
1737	mutex_unlock(lock: &blkcg_pol_register_mutex);
1738	return 0;
1739
1740	err_free_cpds:
1741	if (pol->cpd_free_fn)
1742	blkcg_free_all_cpd(pol);
1743
1744	blkcg_policy[pol->plid] = NULL;
1745	err_unlock:
1746	mutex_unlock(lock: &blkcg_pol_mutex);
1747	mutex_unlock(lock: &blkcg_pol_register_mutex);
1748	return ret;
1749	}
1750	EXPORT_SYMBOL_GPL(blkcg_policy_register);
1751
1752	/**
1753	* blkcg_policy_unregister - unregister a blkcg policy
1754	* @pol: blkcg policy to unregister
1755	*
1756	* Undo blkcg_policy_register(@pol). Might sleep.
1757	*/
1758	void blkcg_policy_unregister(struct blkcg_policy *pol)
1759	{
1760	mutex_lock(&blkcg_pol_register_mutex);
1761
1762	if (WARN_ON(blkcg_policy[pol->plid] != pol))
1763	goto out_unlock;
1764
1765	/* kill the intf files first */
1766	if (pol->dfl_cftypes)
1767	cgroup_rm_cftypes(cfts: pol->dfl_cftypes);
1768	if (pol->legacy_cftypes)
1769	cgroup_rm_cftypes(cfts: pol->legacy_cftypes);
1770
1771	/* remove cpds and unregister */
1772	mutex_lock(&blkcg_pol_mutex);
1773
1774	if (pol->cpd_free_fn)
1775	blkcg_free_all_cpd(pol);
1776
1777	blkcg_policy[pol->plid] = NULL;
1778
1779	mutex_unlock(lock: &blkcg_pol_mutex);
1780	out_unlock:
1781	mutex_unlock(lock: &blkcg_pol_register_mutex);
1782	}
1783	EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1784
1785	/*
1786	* Scale the accumulated delay based on how long it has been since we updated
1787	* the delay. We only call this when we are adding delay, in case it's been a
1788	* while since we added delay, and when we are checking to see if we need to
1789	* delay a task, to account for any delays that may have occurred.
1790	*/
1791	static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1792	{
1793	u64 old = atomic64_read(v: &blkg->delay_start);
1794
1795	/* negative use_delay means no scaling, see blkcg_set_delay() */
1796	if (atomic_read(v: &blkg->use_delay) < 0)
1797	return;
1798
1799	/*
1800	* We only want to scale down every second. The idea here is that we
1801	* want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1802	* time window. We only want to throttle tasks for recent delay that
1803	* has occurred, in 1 second time windows since that's the maximum
1804	* things can be throttled. We save the current delay window in
1805	* blkg->last_delay so we know what amount is still left to be charged
1806	* to the blkg from this point onward. blkg->last_use keeps track of
1807	* the use_delay counter. The idea is if we're unthrottling the blkg we
1808	* are ok with whatever is happening now, and we can take away more of
1809	* the accumulated delay as we've already throttled enough that
1810	* everybody is happy with their IO latencies.
1811	*/
1812	if (time_before64(old + NSEC_PER_SEC, now) &&
1813	atomic64_try_cmpxchg(v: &blkg->delay_start, old: &old, new: now)) {
1814	u64 cur = atomic64_read(v: &blkg->delay_nsec);
1815	u64 sub = min_t(u64, blkg->last_delay, now - old);
1816	int cur_use = atomic_read(v: &blkg->use_delay);
1817
1818	/*
1819	* We've been unthrottled, subtract a larger chunk of our
1820	* accumulated delay.
1821	*/
1822	if (cur_use < blkg->last_use)
1823	sub = max_t(u64, sub, blkg->last_delay >> 1);
1824
1825	/*
1826	* This shouldn't happen, but handle it anyway. Our delay_nsec
1827	* should only ever be growing except here where we subtract out
1828	* min(last_delay, 1 second), but lord knows bugs happen and I'd
1829	* rather not end up with negative numbers.
1830	*/
1831	if (unlikely(cur < sub)) {
1832	atomic64_set(v: &blkg->delay_nsec, i: 0);
1833	blkg->last_delay = 0;
1834	} else {
1835	atomic64_sub(i: sub, v: &blkg->delay_nsec);
1836	blkg->last_delay = cur - sub;
1837	}
1838	blkg->last_use = cur_use;
1839	}
1840	}
1841
1842	/*
1843	* This is called when we want to actually walk up the hierarchy and check to
1844	* see if we need to throttle, and then actually throttle if there is some
1845	* accumulated delay. This should only be called upon return to user space so
1846	* we're not holding some lock that would induce a priority inversion.
1847	*/
1848	static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1849	{
1850	unsigned long pflags;
1851	bool clamp;
1852	u64 now = blk_time_get_ns();
1853	u64 exp;
1854	u64 delay_nsec = 0;
1855	int tok;
1856
1857	while (blkg->parent) {
1858	int use_delay = atomic_read(v: &blkg->use_delay);
1859
1860	if (use_delay) {
1861	u64 this_delay;
1862
1863	blkcg_scale_delay(blkg, now);
1864	this_delay = atomic64_read(v: &blkg->delay_nsec);
1865	if (this_delay > delay_nsec) {
1866	delay_nsec = this_delay;
1867	clamp = use_delay > 0;
1868	}
1869	}
1870	blkg = blkg->parent;
1871	}
1872
1873	if (!delay_nsec)
1874	return;
1875
1876	/*
1877	* Let's not sleep for all eternity if we've amassed a huge delay.
1878	* Swapping or metadata IO can accumulate 10's of seconds worth of
1879	* delay, and we want userspace to be able to do _something_ so cap the
1880	* delays at 0.25s. If there's 10's of seconds worth of delay then the
1881	* tasks will be delayed for 0.25 second for every syscall. If
1882	* blkcg_set_delay() was used as indicated by negative use_delay, the
1883	* caller is responsible for regulating the range.
1884	*/
1885	if (clamp)
1886	delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1887
1888	if (use_memdelay)
1889	psi_memstall_enter(flags: &pflags);
1890
1891	exp = ktime_add_ns(now, delay_nsec);
1892	tok = io_schedule_prepare();
1893	do {
1894	__set_current_state(TASK_KILLABLE);
1895	if (!schedule_hrtimeout(expires: &exp, mode: HRTIMER_MODE_ABS))
1896	break;
1897	} while (!fatal_signal_pending(current));
1898	io_schedule_finish(token: tok);
1899
1900	if (use_memdelay)
1901	psi_memstall_leave(flags: &pflags);
1902	}
1903
1904	/**
1905	* blkcg_maybe_throttle_current - throttle the current task if it has been marked
1906	*
1907	* This is only called if we've been marked with set_notify_resume(). Obviously
1908	* we can be set_notify_resume() for reasons other than blkcg throttling, so we
1909	* check to see if current->throttle_disk is set and if not this doesn't do
1910	* anything. This should only ever be called by the resume code, it's not meant
1911	* to be called by people willy-nilly as it will actually do the work to
1912	* throttle the task if it is setup for throttling.
1913	*/
1914	void blkcg_maybe_throttle_current(void)
1915	{
1916	struct gendisk *disk = current->throttle_disk;
1917	struct blkcg *blkcg;
1918	struct blkcg_gq *blkg;
1919	bool use_memdelay = current->use_memdelay;
1920
1921	if (!disk)
1922	return;
1923
1924	current->throttle_disk = NULL;
1925	current->use_memdelay = false;
1926
1927	rcu_read_lock();
1928	blkcg = css_to_blkcg(css: blkcg_css());
1929	if (!blkcg)
1930	goto out;
1931	blkg = blkg_lookup(blkcg, q: disk->queue);
1932	if (!blkg)
1933	goto out;
1934	if (!blkg_tryget(blkg))
1935	goto out;
1936	rcu_read_unlock();
1937
1938	blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1939	blkg_put(blkg);
1940	put_disk(disk);
1941	return;
1942	out:
1943	rcu_read_unlock();
1944	}
1945
1946	/**
1947	* blkcg_schedule_throttle - this task needs to check for throttling
1948	* @disk: disk to throttle
1949	* @use_memdelay: do we charge this to memory delay for PSI
1950	*
1951	* This is called by the IO controller when we know there's delay accumulated
1952	* for the blkg for this task. We do not pass the blkg because there are places
1953	* we call this that may not have that information, the swapping code for
1954	* instance will only have a block_device at that point. This set's the
1955	* notify_resume for the task to check and see if it requires throttling before
1956	* returning to user space.
1957	*
1958	* We will only schedule once per syscall. You can call this over and over
1959	* again and it will only do the check once upon return to user space, and only
1960	* throttle once. If the task needs to be throttled again it'll need to be
1961	* re-set at the next time we see the task.
1962	*/
1963	void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1964	{
1965	if (unlikely(current->flags & PF_KTHREAD))
1966	return;
1967
1968	if (current->throttle_disk != disk) {
1969	if (test_bit(GD_DEAD, &disk->state))
1970	return;
1971	get_device(disk_to_dev(disk));
1972
1973	if (current->throttle_disk)
1974	put_disk(current->throttle_disk);
1975	current->throttle_disk = disk;
1976	}
1977
1978	if (use_memdelay)
1979	current->use_memdelay = use_memdelay;
1980	set_notify_resume(current);
1981	}
1982
1983	/**
1984	* blkcg_add_delay - add delay to this blkg
1985	* @blkg: blkg of interest
1986	* @now: the current time in nanoseconds
1987	* @delta: how many nanoseconds of delay to add
1988	*
1989	* Charge @delta to the blkg's current delay accumulation. This is used to
1990	* throttle tasks if an IO controller thinks we need more throttling.
1991	*/
1992	void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1993	{
1994	if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1995	return;
1996	blkcg_scale_delay(blkg, now);
1997	atomic64_add(i: delta, v: &blkg->delay_nsec);
1998	}
1999
2000	/**
2001	* blkg_tryget_closest - try and get a blkg ref on the closet blkg
2002	* @bio: target bio
2003	* @css: target css
2004	*
2005	* As the failure mode here is to walk up the blkg tree, this ensure that the
2006	* blkg->parent pointers are always valid. This returns the blkg that it ended
2007	* up taking a reference on or %NULL if no reference was taken.
2008	*/
2009	static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
2010	struct cgroup_subsys_state *css)
2011	{
2012	struct blkcg_gq *blkg, *ret_blkg = NULL;
2013
2014	rcu_read_lock();
2015	blkg = blkg_lookup_create(blkcg: css_to_blkcg(css), disk: bio->bi_bdev->bd_disk);
2016	while (blkg) {
2017	if (blkg_tryget(blkg)) {
2018	ret_blkg = blkg;
2019	break;
2020	}
2021	blkg = blkg->parent;
2022	}
2023	rcu_read_unlock();
2024
2025	return ret_blkg;
2026	}
2027
2028	/**
2029	* bio_associate_blkg_from_css - associate a bio with a specified css
2030	* @bio: target bio
2031	* @css: target css
2032	*
2033	* Associate @bio with the blkg found by combining the css's blkg and the
2034	* request_queue of the @bio. An association failure is handled by walking up
2035	* the blkg tree. Therefore, the blkg associated can be anything between @blkg
2036	* and q->root_blkg. This situation only happens when a cgroup is dying and
2037	* then the remaining bios will spill to the closest alive blkg.
2038	*
2039	* A reference will be taken on the blkg and will be released when @bio is
2040	* freed.
2041	*/
2042	void bio_associate_blkg_from_css(struct bio *bio,
2043	struct cgroup_subsys_state *css)
2044	{
2045	if (bio->bi_blkg)
2046	blkg_put(blkg: bio->bi_blkg);
2047
2048	if (css && css->parent) {
2049	bio->bi_blkg = blkg_tryget_closest(bio, css);
2050	} else {
2051	blkg_get(blkg: bdev_get_queue(bdev: bio->bi_bdev)->root_blkg);
2052	bio->bi_blkg = bdev_get_queue(bdev: bio->bi_bdev)->root_blkg;
2053	}
2054	}
2055	EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2056
2057	/**
2058	* bio_associate_blkg - associate a bio with a blkg
2059	* @bio: target bio
2060	*
2061	* Associate @bio with the blkg found from the bio's css and request_queue.
2062	* If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2063	* already associated, the css is reused and association redone as the
2064	* request_queue may have changed.
2065	*/
2066	void bio_associate_blkg(struct bio *bio)
2067	{
2068	struct cgroup_subsys_state *css;
2069
2070	if (blk_op_is_passthrough(op: bio->bi_opf))
2071	return;
2072
2073	rcu_read_lock();
2074
2075	if (bio->bi_blkg)
2076	css = bio_blkcg_css(bio);
2077	else
2078	css = blkcg_css();
2079
2080	bio_associate_blkg_from_css(bio, css);
2081
2082	rcu_read_unlock();
2083	}
2084	EXPORT_SYMBOL_GPL(bio_associate_blkg);
2085
2086	/**
2087	* bio_clone_blkg_association - clone blkg association from src to dst bio
2088	* @dst: destination bio
2089	* @src: source bio
2090	*/
2091	void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2092	{
2093	if (src->bi_blkg)
2094	bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2095	}
2096	EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2097
2098	static int blk_cgroup_io_type(struct bio *bio)
2099	{
2100	if (op_is_discard(op: bio->bi_opf))
2101	return BLKG_IOSTAT_DISCARD;
2102	if (op_is_write(op: bio->bi_opf))
2103	return BLKG_IOSTAT_WRITE;
2104	return BLKG_IOSTAT_READ;
2105	}
2106
2107	void blk_cgroup_bio_start(struct bio *bio)
2108	{
2109	struct blkcg *blkcg = bio->bi_blkg->blkcg;
2110	int rwd = blk_cgroup_io_type(bio), cpu;
2111	struct blkg_iostat_set *bis;
2112	unsigned long flags;
2113
2114	if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
2115	return;
2116
2117	/* Root-level stats are sourced from system-wide IO stats */
2118	if (!cgroup_parent(cgrp: blkcg->css.cgroup))
2119	return;
2120
2121	cpu = get_cpu();
2122	bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2123	flags = u64_stats_update_begin_irqsave(syncp: &bis->sync);
2124
2125	/*
2126	* If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2127	* bio and we would have already accounted for the size of the bio.
2128	*/
2129	if (!bio_flagged(bio, bit: BIO_CGROUP_ACCT)) {
2130	bio_set_flag(bio, bit: BIO_CGROUP_ACCT);
2131	bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2132	}
2133	bis->cur.ios[rwd]++;
2134
2135	/*
2136	* If the iostat_cpu isn't in a lockless list, put it into the
2137	* list to indicate that a stat update is pending.
2138	*/
2139	if (!READ_ONCE(bis->lqueued)) {
2140	struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2141
2142	llist_add(new: &bis->lnode, head: lhead);
2143	WRITE_ONCE(bis->lqueued, true);
2144	}
2145
2146	u64_stats_update_end_irqrestore(syncp: &bis->sync, flags);
2147	cgroup_rstat_updated(cgrp: blkcg->css.cgroup, cpu);
2148	put_cpu();
2149	}
2150
2151	bool blk_cgroup_congested(void)
2152	{
2153	struct cgroup_subsys_state *css;
2154	bool ret = false;
2155
2156	rcu_read_lock();
2157	for (css = blkcg_css(); css; css = css->parent) {
2158	if (atomic_read(v: &css->cgroup->congestion_count)) {
2159	ret = true;
2160	break;
2161	}
2162	}
2163	rcu_read_unlock();
2164	return ret;
2165	}
2166
2167	module_param(blkcg_debug_stats, bool, 0644);
2168	MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");
2169