xref: /openbmc/linux/block/blk-cgroup.c (revision bfa87ac8)
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 /*
38  * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39  * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40  * policy [un]register operations including cgroup file additions /
41  * removals.  Putting cgroup file registration outside blkcg_pol_mutex
42  * allows grabbing it from cgroup callbacks.
43  */
44 static DEFINE_MUTEX(blkcg_pol_register_mutex);
45 static DEFINE_MUTEX(blkcg_pol_mutex);
46 
47 struct blkcg blkcg_root;
48 EXPORT_SYMBOL_GPL(blkcg_root);
49 
50 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51 EXPORT_SYMBOL_GPL(blkcg_root_css);
52 
53 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
54 
55 static LIST_HEAD(all_blkcgs);		/* protected by blkcg_pol_mutex */
56 
57 bool blkcg_debug_stats = false;
58 static struct workqueue_struct *blkcg_punt_bio_wq;
59 
60 #define BLKG_DESTROY_BATCH_SIZE  64
61 
62 /**
63  * blkcg_css - find the current css
64  *
65  * Find the css associated with either the kthread or the current task.
66  * This may return a dying css, so it is up to the caller to use tryget logic
67  * to confirm it is alive and well.
68  */
69 static struct cgroup_subsys_state *blkcg_css(void)
70 {
71 	struct cgroup_subsys_state *css;
72 
73 	css = kthread_blkcg();
74 	if (css)
75 		return css;
76 	return task_css(current, io_cgrp_id);
77 }
78 
79 static bool blkcg_policy_enabled(struct request_queue *q,
80 				 const struct blkcg_policy *pol)
81 {
82 	return pol && test_bit(pol->plid, q->blkcg_pols);
83 }
84 
85 static void blkg_free_workfn(struct work_struct *work)
86 {
87 	struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
88 					     free_work);
89 	int i;
90 
91 	for (i = 0; i < BLKCG_MAX_POLS; i++)
92 		if (blkg->pd[i])
93 			blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
94 
95 	if (blkg->q)
96 		blk_put_queue(blkg->q);
97 	free_percpu(blkg->iostat_cpu);
98 	percpu_ref_exit(&blkg->refcnt);
99 	kfree(blkg);
100 }
101 
102 /**
103  * blkg_free - free a blkg
104  * @blkg: blkg to free
105  *
106  * Free @blkg which may be partially allocated.
107  */
108 static void blkg_free(struct blkcg_gq *blkg)
109 {
110 	if (!blkg)
111 		return;
112 
113 	/*
114 	 * Both ->pd_free_fn() and request queue's release handler may
115 	 * sleep, so free us by scheduling one work func
116 	 */
117 	INIT_WORK(&blkg->free_work, blkg_free_workfn);
118 	schedule_work(&blkg->free_work);
119 }
120 
121 static void __blkg_release(struct rcu_head *rcu)
122 {
123 	struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
124 
125 	WARN_ON(!bio_list_empty(&blkg->async_bios));
126 
127 	/* release the blkcg and parent blkg refs this blkg has been holding */
128 	css_put(&blkg->blkcg->css);
129 	if (blkg->parent)
130 		blkg_put(blkg->parent);
131 	blkg_free(blkg);
132 }
133 
134 /*
135  * A group is RCU protected, but having an rcu lock does not mean that one
136  * can access all the fields of blkg and assume these are valid.  For
137  * example, don't try to follow throtl_data and request queue links.
138  *
139  * Having a reference to blkg under an rcu allows accesses to only values
140  * local to groups like group stats and group rate limits.
141  */
142 static void blkg_release(struct percpu_ref *ref)
143 {
144 	struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
145 
146 	call_rcu(&blkg->rcu_head, __blkg_release);
147 }
148 
149 static void blkg_async_bio_workfn(struct work_struct *work)
150 {
151 	struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
152 					     async_bio_work);
153 	struct bio_list bios = BIO_EMPTY_LIST;
154 	struct bio *bio;
155 	struct blk_plug plug;
156 	bool need_plug = false;
157 
158 	/* as long as there are pending bios, @blkg can't go away */
159 	spin_lock_bh(&blkg->async_bio_lock);
160 	bio_list_merge(&bios, &blkg->async_bios);
161 	bio_list_init(&blkg->async_bios);
162 	spin_unlock_bh(&blkg->async_bio_lock);
163 
164 	/* start plug only when bio_list contains at least 2 bios */
165 	if (bios.head && bios.head->bi_next) {
166 		need_plug = true;
167 		blk_start_plug(&plug);
168 	}
169 	while ((bio = bio_list_pop(&bios)))
170 		submit_bio(bio);
171 	if (need_plug)
172 		blk_finish_plug(&plug);
173 }
174 
175 /**
176  * bio_blkcg_css - return the blkcg CSS associated with a bio
177  * @bio: target bio
178  *
179  * This returns the CSS for the blkcg associated with a bio, or %NULL if not
180  * associated. Callers are expected to either handle %NULL or know association
181  * has been done prior to calling this.
182  */
183 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
184 {
185 	if (!bio || !bio->bi_blkg)
186 		return NULL;
187 	return &bio->bi_blkg->blkcg->css;
188 }
189 EXPORT_SYMBOL_GPL(bio_blkcg_css);
190 
191 /**
192  * blkcg_parent - get the parent of a blkcg
193  * @blkcg: blkcg of interest
194  *
195  * Return the parent blkcg of @blkcg.  Can be called anytime.
196  */
197 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
198 {
199 	return css_to_blkcg(blkcg->css.parent);
200 }
201 
202 /**
203  * blkg_alloc - allocate a blkg
204  * @blkcg: block cgroup the new blkg is associated with
205  * @disk: gendisk the new blkg is associated with
206  * @gfp_mask: allocation mask to use
207  *
208  * Allocate a new blkg assocating @blkcg and @q.
209  */
210 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
211 				   gfp_t gfp_mask)
212 {
213 	struct blkcg_gq *blkg;
214 	int i, cpu;
215 
216 	/* alloc and init base part */
217 	blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
218 	if (!blkg)
219 		return NULL;
220 
221 	if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
222 		goto err_free;
223 
224 	blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
225 	if (!blkg->iostat_cpu)
226 		goto err_free;
227 
228 	if (!blk_get_queue(disk->queue))
229 		goto err_free;
230 
231 	blkg->q = disk->queue;
232 	INIT_LIST_HEAD(&blkg->q_node);
233 	spin_lock_init(&blkg->async_bio_lock);
234 	bio_list_init(&blkg->async_bios);
235 	INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
236 	blkg->blkcg = blkcg;
237 
238 	u64_stats_init(&blkg->iostat.sync);
239 	for_each_possible_cpu(cpu)
240 		u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
241 
242 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
243 		struct blkcg_policy *pol = blkcg_policy[i];
244 		struct blkg_policy_data *pd;
245 
246 		if (!blkcg_policy_enabled(disk->queue, pol))
247 			continue;
248 
249 		/* alloc per-policy data and attach it to blkg */
250 		pd = pol->pd_alloc_fn(gfp_mask, disk->queue, blkcg);
251 		if (!pd)
252 			goto err_free;
253 
254 		blkg->pd[i] = pd;
255 		pd->blkg = blkg;
256 		pd->plid = i;
257 	}
258 
259 	return blkg;
260 
261 err_free:
262 	blkg_free(blkg);
263 	return NULL;
264 }
265 
266 /*
267  * If @new_blkg is %NULL, this function tries to allocate a new one as
268  * necessary using %GFP_NOWAIT.  @new_blkg is always consumed on return.
269  */
270 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
271 				    struct blkcg_gq *new_blkg)
272 {
273 	struct blkcg_gq *blkg;
274 	int i, ret;
275 
276 	lockdep_assert_held(&disk->queue->queue_lock);
277 
278 	/* request_queue is dying, do not create/recreate a blkg */
279 	if (blk_queue_dying(disk->queue)) {
280 		ret = -ENODEV;
281 		goto err_free_blkg;
282 	}
283 
284 	/* blkg holds a reference to blkcg */
285 	if (!css_tryget_online(&blkcg->css)) {
286 		ret = -ENODEV;
287 		goto err_free_blkg;
288 	}
289 
290 	/* allocate */
291 	if (!new_blkg) {
292 		new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
293 		if (unlikely(!new_blkg)) {
294 			ret = -ENOMEM;
295 			goto err_put_css;
296 		}
297 	}
298 	blkg = new_blkg;
299 
300 	/* link parent */
301 	if (blkcg_parent(blkcg)) {
302 		blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
303 		if (WARN_ON_ONCE(!blkg->parent)) {
304 			ret = -ENODEV;
305 			goto err_put_css;
306 		}
307 		blkg_get(blkg->parent);
308 	}
309 
310 	/* invoke per-policy init */
311 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
312 		struct blkcg_policy *pol = blkcg_policy[i];
313 
314 		if (blkg->pd[i] && pol->pd_init_fn)
315 			pol->pd_init_fn(blkg->pd[i]);
316 	}
317 
318 	/* insert */
319 	spin_lock(&blkcg->lock);
320 	ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
321 	if (likely(!ret)) {
322 		hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
323 		list_add(&blkg->q_node, &disk->queue->blkg_list);
324 
325 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
326 			struct blkcg_policy *pol = blkcg_policy[i];
327 
328 			if (blkg->pd[i] && pol->pd_online_fn)
329 				pol->pd_online_fn(blkg->pd[i]);
330 		}
331 	}
332 	blkg->online = true;
333 	spin_unlock(&blkcg->lock);
334 
335 	if (!ret)
336 		return blkg;
337 
338 	/* @blkg failed fully initialized, use the usual release path */
339 	blkg_put(blkg);
340 	return ERR_PTR(ret);
341 
342 err_put_css:
343 	css_put(&blkcg->css);
344 err_free_blkg:
345 	blkg_free(new_blkg);
346 	return ERR_PTR(ret);
347 }
348 
349 /**
350  * blkg_lookup_create - lookup blkg, try to create one if not there
351  * @blkcg: blkcg of interest
352  * @disk: gendisk of interest
353  *
354  * Lookup blkg for the @blkcg - @disk pair.  If it doesn't exist, try to
355  * create one.  blkg creation is performed recursively from blkcg_root such
356  * that all non-root blkg's have access to the parent blkg.  This function
357  * should be called under RCU read lock and takes @disk->queue->queue_lock.
358  *
359  * Returns the blkg or the closest blkg if blkg_create() fails as it walks
360  * down from root.
361  */
362 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
363 		struct gendisk *disk)
364 {
365 	struct request_queue *q = disk->queue;
366 	struct blkcg_gq *blkg;
367 	unsigned long flags;
368 
369 	WARN_ON_ONCE(!rcu_read_lock_held());
370 
371 	blkg = blkg_lookup(blkcg, q);
372 	if (blkg)
373 		return blkg;
374 
375 	spin_lock_irqsave(&q->queue_lock, flags);
376 	blkg = blkg_lookup(blkcg, q);
377 	if (blkg) {
378 		if (blkcg != &blkcg_root &&
379 		    blkg != rcu_dereference(blkcg->blkg_hint))
380 			rcu_assign_pointer(blkcg->blkg_hint, blkg);
381 		goto found;
382 	}
383 
384 	/*
385 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
386 	 * non-root blkgs have access to their parents.  Returns the closest
387 	 * blkg to the intended blkg should blkg_create() fail.
388 	 */
389 	while (true) {
390 		struct blkcg *pos = blkcg;
391 		struct blkcg *parent = blkcg_parent(blkcg);
392 		struct blkcg_gq *ret_blkg = q->root_blkg;
393 
394 		while (parent) {
395 			blkg = blkg_lookup(parent, q);
396 			if (blkg) {
397 				/* remember closest blkg */
398 				ret_blkg = blkg;
399 				break;
400 			}
401 			pos = parent;
402 			parent = blkcg_parent(parent);
403 		}
404 
405 		blkg = blkg_create(pos, disk, NULL);
406 		if (IS_ERR(blkg)) {
407 			blkg = ret_blkg;
408 			break;
409 		}
410 		if (pos == blkcg)
411 			break;
412 	}
413 
414 found:
415 	spin_unlock_irqrestore(&q->queue_lock, flags);
416 	return blkg;
417 }
418 
419 static void blkg_destroy(struct blkcg_gq *blkg)
420 {
421 	struct blkcg *blkcg = blkg->blkcg;
422 	int i;
423 
424 	lockdep_assert_held(&blkg->q->queue_lock);
425 	lockdep_assert_held(&blkcg->lock);
426 
427 	/* Something wrong if we are trying to remove same group twice */
428 	WARN_ON_ONCE(list_empty(&blkg->q_node));
429 	WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
430 
431 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
432 		struct blkcg_policy *pol = blkcg_policy[i];
433 
434 		if (blkg->pd[i] && pol->pd_offline_fn)
435 			pol->pd_offline_fn(blkg->pd[i]);
436 	}
437 
438 	blkg->online = false;
439 
440 	radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
441 	list_del_init(&blkg->q_node);
442 	hlist_del_init_rcu(&blkg->blkcg_node);
443 
444 	/*
445 	 * Both setting lookup hint to and clearing it from @blkg are done
446 	 * under queue_lock.  If it's not pointing to @blkg now, it never
447 	 * will.  Hint assignment itself can race safely.
448 	 */
449 	if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
450 		rcu_assign_pointer(blkcg->blkg_hint, NULL);
451 
452 	/*
453 	 * Put the reference taken at the time of creation so that when all
454 	 * queues are gone, group can be destroyed.
455 	 */
456 	percpu_ref_kill(&blkg->refcnt);
457 }
458 
459 static void blkg_destroy_all(struct gendisk *disk)
460 {
461 	struct request_queue *q = disk->queue;
462 	struct blkcg_gq *blkg, *n;
463 	int count = BLKG_DESTROY_BATCH_SIZE;
464 
465 restart:
466 	spin_lock_irq(&q->queue_lock);
467 	list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
468 		struct blkcg *blkcg = blkg->blkcg;
469 
470 		spin_lock(&blkcg->lock);
471 		blkg_destroy(blkg);
472 		spin_unlock(&blkcg->lock);
473 
474 		/*
475 		 * in order to avoid holding the spin lock for too long, release
476 		 * it when a batch of blkgs are destroyed.
477 		 */
478 		if (!(--count)) {
479 			count = BLKG_DESTROY_BATCH_SIZE;
480 			spin_unlock_irq(&q->queue_lock);
481 			cond_resched();
482 			goto restart;
483 		}
484 	}
485 
486 	q->root_blkg = NULL;
487 	spin_unlock_irq(&q->queue_lock);
488 }
489 
490 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
491 			     struct cftype *cftype, u64 val)
492 {
493 	struct blkcg *blkcg = css_to_blkcg(css);
494 	struct blkcg_gq *blkg;
495 	int i, cpu;
496 
497 	mutex_lock(&blkcg_pol_mutex);
498 	spin_lock_irq(&blkcg->lock);
499 
500 	/*
501 	 * Note that stat reset is racy - it doesn't synchronize against
502 	 * stat updates.  This is a debug feature which shouldn't exist
503 	 * anyway.  If you get hit by a race, retry.
504 	 */
505 	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
506 		for_each_possible_cpu(cpu) {
507 			struct blkg_iostat_set *bis =
508 				per_cpu_ptr(blkg->iostat_cpu, cpu);
509 			memset(bis, 0, sizeof(*bis));
510 		}
511 		memset(&blkg->iostat, 0, sizeof(blkg->iostat));
512 
513 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
514 			struct blkcg_policy *pol = blkcg_policy[i];
515 
516 			if (blkg->pd[i] && pol->pd_reset_stats_fn)
517 				pol->pd_reset_stats_fn(blkg->pd[i]);
518 		}
519 	}
520 
521 	spin_unlock_irq(&blkcg->lock);
522 	mutex_unlock(&blkcg_pol_mutex);
523 	return 0;
524 }
525 
526 const char *blkg_dev_name(struct blkcg_gq *blkg)
527 {
528 	if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
529 		return NULL;
530 	return bdi_dev_name(blkg->q->disk->bdi);
531 }
532 
533 /**
534  * blkcg_print_blkgs - helper for printing per-blkg data
535  * @sf: seq_file to print to
536  * @blkcg: blkcg of interest
537  * @prfill: fill function to print out a blkg
538  * @pol: policy in question
539  * @data: data to be passed to @prfill
540  * @show_total: to print out sum of prfill return values or not
541  *
542  * This function invokes @prfill on each blkg of @blkcg if pd for the
543  * policy specified by @pol exists.  @prfill is invoked with @sf, the
544  * policy data and @data and the matching queue lock held.  If @show_total
545  * is %true, the sum of the return values from @prfill is printed with
546  * "Total" label at the end.
547  *
548  * This is to be used to construct print functions for
549  * cftype->read_seq_string method.
550  */
551 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
552 		       u64 (*prfill)(struct seq_file *,
553 				     struct blkg_policy_data *, int),
554 		       const struct blkcg_policy *pol, int data,
555 		       bool show_total)
556 {
557 	struct blkcg_gq *blkg;
558 	u64 total = 0;
559 
560 	rcu_read_lock();
561 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
562 		spin_lock_irq(&blkg->q->queue_lock);
563 		if (blkcg_policy_enabled(blkg->q, pol))
564 			total += prfill(sf, blkg->pd[pol->plid], data);
565 		spin_unlock_irq(&blkg->q->queue_lock);
566 	}
567 	rcu_read_unlock();
568 
569 	if (show_total)
570 		seq_printf(sf, "Total %llu\n", (unsigned long long)total);
571 }
572 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
573 
574 /**
575  * __blkg_prfill_u64 - prfill helper for a single u64 value
576  * @sf: seq_file to print to
577  * @pd: policy private data of interest
578  * @v: value to print
579  *
580  * Print @v to @sf for the device assocaited with @pd.
581  */
582 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
583 {
584 	const char *dname = blkg_dev_name(pd->blkg);
585 
586 	if (!dname)
587 		return 0;
588 
589 	seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
590 	return v;
591 }
592 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
593 
594 /**
595  * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
596  * @inputp: input string pointer
597  *
598  * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
599  * from @input and get and return the matching bdev.  *@inputp is
600  * updated to point past the device node prefix.  Returns an ERR_PTR()
601  * value on error.
602  *
603  * Use this function iff blkg_conf_prep() can't be used for some reason.
604  */
605 struct block_device *blkcg_conf_open_bdev(char **inputp)
606 {
607 	char *input = *inputp;
608 	unsigned int major, minor;
609 	struct block_device *bdev;
610 	int key_len;
611 
612 	if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
613 		return ERR_PTR(-EINVAL);
614 
615 	input += key_len;
616 	if (!isspace(*input))
617 		return ERR_PTR(-EINVAL);
618 	input = skip_spaces(input);
619 
620 	bdev = blkdev_get_no_open(MKDEV(major, minor));
621 	if (!bdev)
622 		return ERR_PTR(-ENODEV);
623 	if (bdev_is_partition(bdev)) {
624 		blkdev_put_no_open(bdev);
625 		return ERR_PTR(-ENODEV);
626 	}
627 
628 	*inputp = input;
629 	return bdev;
630 }
631 
632 /**
633  * blkg_conf_prep - parse and prepare for per-blkg config update
634  * @blkcg: target block cgroup
635  * @pol: target policy
636  * @input: input string
637  * @ctx: blkg_conf_ctx to be filled
638  *
639  * Parse per-blkg config update from @input and initialize @ctx with the
640  * result.  @ctx->blkg points to the blkg to be updated and @ctx->body the
641  * part of @input following MAJ:MIN.  This function returns with RCU read
642  * lock and queue lock held and must be paired with blkg_conf_finish().
643  */
644 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
645 		   char *input, struct blkg_conf_ctx *ctx)
646 	__acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
647 {
648 	struct block_device *bdev;
649 	struct gendisk *disk;
650 	struct request_queue *q;
651 	struct blkcg_gq *blkg;
652 	int ret;
653 
654 	bdev = blkcg_conf_open_bdev(&input);
655 	if (IS_ERR(bdev))
656 		return PTR_ERR(bdev);
657 	disk = bdev->bd_disk;
658 	q = disk->queue;
659 
660 	/*
661 	 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
662 	 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
663 	 */
664 	ret = blk_queue_enter(q, 0);
665 	if (ret)
666 		goto fail;
667 
668 	rcu_read_lock();
669 	spin_lock_irq(&q->queue_lock);
670 
671 	if (!blkcg_policy_enabled(q, pol)) {
672 		ret = -EOPNOTSUPP;
673 		goto fail_unlock;
674 	}
675 
676 	blkg = blkg_lookup(blkcg, q);
677 	if (blkg)
678 		goto success;
679 
680 	/*
681 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
682 	 * non-root blkgs have access to their parents.
683 	 */
684 	while (true) {
685 		struct blkcg *pos = blkcg;
686 		struct blkcg *parent;
687 		struct blkcg_gq *new_blkg;
688 
689 		parent = blkcg_parent(blkcg);
690 		while (parent && !blkg_lookup(parent, q)) {
691 			pos = parent;
692 			parent = blkcg_parent(parent);
693 		}
694 
695 		/* Drop locks to do new blkg allocation with GFP_KERNEL. */
696 		spin_unlock_irq(&q->queue_lock);
697 		rcu_read_unlock();
698 
699 		new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
700 		if (unlikely(!new_blkg)) {
701 			ret = -ENOMEM;
702 			goto fail_exit_queue;
703 		}
704 
705 		if (radix_tree_preload(GFP_KERNEL)) {
706 			blkg_free(new_blkg);
707 			ret = -ENOMEM;
708 			goto fail_exit_queue;
709 		}
710 
711 		rcu_read_lock();
712 		spin_lock_irq(&q->queue_lock);
713 
714 		if (!blkcg_policy_enabled(q, pol)) {
715 			blkg_free(new_blkg);
716 			ret = -EOPNOTSUPP;
717 			goto fail_preloaded;
718 		}
719 
720 		blkg = blkg_lookup(pos, q);
721 		if (blkg) {
722 			blkg_free(new_blkg);
723 		} else {
724 			blkg = blkg_create(pos, disk, new_blkg);
725 			if (IS_ERR(blkg)) {
726 				ret = PTR_ERR(blkg);
727 				goto fail_preloaded;
728 			}
729 		}
730 
731 		radix_tree_preload_end();
732 
733 		if (pos == blkcg)
734 			goto success;
735 	}
736 success:
737 	blk_queue_exit(q);
738 	ctx->bdev = bdev;
739 	ctx->blkg = blkg;
740 	ctx->body = input;
741 	return 0;
742 
743 fail_preloaded:
744 	radix_tree_preload_end();
745 fail_unlock:
746 	spin_unlock_irq(&q->queue_lock);
747 	rcu_read_unlock();
748 fail_exit_queue:
749 	blk_queue_exit(q);
750 fail:
751 	blkdev_put_no_open(bdev);
752 	/*
753 	 * If queue was bypassing, we should retry.  Do so after a
754 	 * short msleep().  It isn't strictly necessary but queue
755 	 * can be bypassing for some time and it's always nice to
756 	 * avoid busy looping.
757 	 */
758 	if (ret == -EBUSY) {
759 		msleep(10);
760 		ret = restart_syscall();
761 	}
762 	return ret;
763 }
764 EXPORT_SYMBOL_GPL(blkg_conf_prep);
765 
766 /**
767  * blkg_conf_finish - finish up per-blkg config update
768  * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
769  *
770  * Finish up after per-blkg config update.  This function must be paired
771  * with blkg_conf_prep().
772  */
773 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
774 	__releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
775 {
776 	spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
777 	rcu_read_unlock();
778 	blkdev_put_no_open(ctx->bdev);
779 }
780 EXPORT_SYMBOL_GPL(blkg_conf_finish);
781 
782 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
783 {
784 	int i;
785 
786 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
787 		dst->bytes[i] = src->bytes[i];
788 		dst->ios[i] = src->ios[i];
789 	}
790 }
791 
792 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
793 {
794 	int i;
795 
796 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
797 		dst->bytes[i] += src->bytes[i];
798 		dst->ios[i] += src->ios[i];
799 	}
800 }
801 
802 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
803 {
804 	int i;
805 
806 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
807 		dst->bytes[i] -= src->bytes[i];
808 		dst->ios[i] -= src->ios[i];
809 	}
810 }
811 
812 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
813 				struct blkg_iostat *last)
814 {
815 	struct blkg_iostat delta;
816 	unsigned long flags;
817 
818 	/* propagate percpu delta to global */
819 	flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
820 	blkg_iostat_set(&delta, cur);
821 	blkg_iostat_sub(&delta, last);
822 	blkg_iostat_add(&blkg->iostat.cur, &delta);
823 	blkg_iostat_add(last, &delta);
824 	u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
825 }
826 
827 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
828 {
829 	struct blkcg *blkcg = css_to_blkcg(css);
830 	struct blkcg_gq *blkg;
831 
832 	/* Root-level stats are sourced from system-wide IO stats */
833 	if (!cgroup_parent(css->cgroup))
834 		return;
835 
836 	rcu_read_lock();
837 
838 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
839 		struct blkcg_gq *parent = blkg->parent;
840 		struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
841 		struct blkg_iostat cur;
842 		unsigned int seq;
843 
844 		/* fetch the current per-cpu values */
845 		do {
846 			seq = u64_stats_fetch_begin(&bisc->sync);
847 			blkg_iostat_set(&cur, &bisc->cur);
848 		} while (u64_stats_fetch_retry(&bisc->sync, seq));
849 
850 		blkcg_iostat_update(blkg, &cur, &bisc->last);
851 
852 		/* propagate global delta to parent (unless that's root) */
853 		if (parent && parent->parent)
854 			blkcg_iostat_update(parent, &blkg->iostat.cur,
855 					    &blkg->iostat.last);
856 	}
857 
858 	rcu_read_unlock();
859 }
860 
861 /*
862  * We source root cgroup stats from the system-wide stats to avoid
863  * tracking the same information twice and incurring overhead when no
864  * cgroups are defined. For that reason, cgroup_rstat_flush in
865  * blkcg_print_stat does not actually fill out the iostat in the root
866  * cgroup's blkcg_gq.
867  *
868  * However, we would like to re-use the printing code between the root and
869  * non-root cgroups to the extent possible. For that reason, we simulate
870  * flushing the root cgroup's stats by explicitly filling in the iostat
871  * with disk level statistics.
872  */
873 static void blkcg_fill_root_iostats(void)
874 {
875 	struct class_dev_iter iter;
876 	struct device *dev;
877 
878 	class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
879 	while ((dev = class_dev_iter_next(&iter))) {
880 		struct block_device *bdev = dev_to_bdev(dev);
881 		struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
882 		struct blkg_iostat tmp;
883 		int cpu;
884 		unsigned long flags;
885 
886 		memset(&tmp, 0, sizeof(tmp));
887 		for_each_possible_cpu(cpu) {
888 			struct disk_stats *cpu_dkstats;
889 
890 			cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
891 			tmp.ios[BLKG_IOSTAT_READ] +=
892 				cpu_dkstats->ios[STAT_READ];
893 			tmp.ios[BLKG_IOSTAT_WRITE] +=
894 				cpu_dkstats->ios[STAT_WRITE];
895 			tmp.ios[BLKG_IOSTAT_DISCARD] +=
896 				cpu_dkstats->ios[STAT_DISCARD];
897 			// convert sectors to bytes
898 			tmp.bytes[BLKG_IOSTAT_READ] +=
899 				cpu_dkstats->sectors[STAT_READ] << 9;
900 			tmp.bytes[BLKG_IOSTAT_WRITE] +=
901 				cpu_dkstats->sectors[STAT_WRITE] << 9;
902 			tmp.bytes[BLKG_IOSTAT_DISCARD] +=
903 				cpu_dkstats->sectors[STAT_DISCARD] << 9;
904 		}
905 
906 		flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
907 		blkg_iostat_set(&blkg->iostat.cur, &tmp);
908 		u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
909 	}
910 }
911 
912 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
913 {
914 	struct blkg_iostat_set *bis = &blkg->iostat;
915 	u64 rbytes, wbytes, rios, wios, dbytes, dios;
916 	const char *dname;
917 	unsigned seq;
918 	int i;
919 
920 	if (!blkg->online)
921 		return;
922 
923 	dname = blkg_dev_name(blkg);
924 	if (!dname)
925 		return;
926 
927 	seq_printf(s, "%s ", dname);
928 
929 	do {
930 		seq = u64_stats_fetch_begin(&bis->sync);
931 
932 		rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
933 		wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
934 		dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
935 		rios = bis->cur.ios[BLKG_IOSTAT_READ];
936 		wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
937 		dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
938 	} while (u64_stats_fetch_retry(&bis->sync, seq));
939 
940 	if (rbytes || wbytes || rios || wios) {
941 		seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
942 			rbytes, wbytes, rios, wios,
943 			dbytes, dios);
944 	}
945 
946 	if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
947 		seq_printf(s, " use_delay=%d delay_nsec=%llu",
948 			atomic_read(&blkg->use_delay),
949 			atomic64_read(&blkg->delay_nsec));
950 	}
951 
952 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
953 		struct blkcg_policy *pol = blkcg_policy[i];
954 
955 		if (!blkg->pd[i] || !pol->pd_stat_fn)
956 			continue;
957 
958 		pol->pd_stat_fn(blkg->pd[i], s);
959 	}
960 
961 	seq_puts(s, "\n");
962 }
963 
964 static int blkcg_print_stat(struct seq_file *sf, void *v)
965 {
966 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
967 	struct blkcg_gq *blkg;
968 
969 	if (!seq_css(sf)->parent)
970 		blkcg_fill_root_iostats();
971 	else
972 		cgroup_rstat_flush(blkcg->css.cgroup);
973 
974 	rcu_read_lock();
975 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
976 		spin_lock_irq(&blkg->q->queue_lock);
977 		blkcg_print_one_stat(blkg, sf);
978 		spin_unlock_irq(&blkg->q->queue_lock);
979 	}
980 	rcu_read_unlock();
981 	return 0;
982 }
983 
984 static struct cftype blkcg_files[] = {
985 	{
986 		.name = "stat",
987 		.seq_show = blkcg_print_stat,
988 	},
989 	{ }	/* terminate */
990 };
991 
992 static struct cftype blkcg_legacy_files[] = {
993 	{
994 		.name = "reset_stats",
995 		.write_u64 = blkcg_reset_stats,
996 	},
997 	{ }	/* terminate */
998 };
999 
1000 #ifdef CONFIG_CGROUP_WRITEBACK
1001 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1002 {
1003 	return &css_to_blkcg(css)->cgwb_list;
1004 }
1005 #endif
1006 
1007 /*
1008  * blkcg destruction is a three-stage process.
1009  *
1010  * 1. Destruction starts.  The blkcg_css_offline() callback is invoked
1011  *    which offlines writeback.  Here we tie the next stage of blkg destruction
1012  *    to the completion of writeback associated with the blkcg.  This lets us
1013  *    avoid punting potentially large amounts of outstanding writeback to root
1014  *    while maintaining any ongoing policies.  The next stage is triggered when
1015  *    the nr_cgwbs count goes to zero.
1016  *
1017  * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1018  *    and handles the destruction of blkgs.  Here the css reference held by
1019  *    the blkg is put back eventually allowing blkcg_css_free() to be called.
1020  *    This work may occur in cgwb_release_workfn() on the cgwb_release
1021  *    workqueue.  Any submitted ios that fail to get the blkg ref will be
1022  *    punted to the root_blkg.
1023  *
1024  * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1025  *    This finally frees the blkcg.
1026  */
1027 
1028 /**
1029  * blkcg_destroy_blkgs - responsible for shooting down blkgs
1030  * @blkcg: blkcg of interest
1031  *
1032  * blkgs should be removed while holding both q and blkcg locks.  As blkcg lock
1033  * is nested inside q lock, this function performs reverse double lock dancing.
1034  * Destroying the blkgs releases the reference held on the blkcg's css allowing
1035  * blkcg_css_free to eventually be called.
1036  *
1037  * This is the blkcg counterpart of ioc_release_fn().
1038  */
1039 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1040 {
1041 	might_sleep();
1042 
1043 	spin_lock_irq(&blkcg->lock);
1044 
1045 	while (!hlist_empty(&blkcg->blkg_list)) {
1046 		struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1047 						struct blkcg_gq, blkcg_node);
1048 		struct request_queue *q = blkg->q;
1049 
1050 		if (need_resched() || !spin_trylock(&q->queue_lock)) {
1051 			/*
1052 			 * Given that the system can accumulate a huge number
1053 			 * of blkgs in pathological cases, check to see if we
1054 			 * need to rescheduling to avoid softlockup.
1055 			 */
1056 			spin_unlock_irq(&blkcg->lock);
1057 			cond_resched();
1058 			spin_lock_irq(&blkcg->lock);
1059 			continue;
1060 		}
1061 
1062 		blkg_destroy(blkg);
1063 		spin_unlock(&q->queue_lock);
1064 	}
1065 
1066 	spin_unlock_irq(&blkcg->lock);
1067 }
1068 
1069 /**
1070  * blkcg_pin_online - pin online state
1071  * @blkcg_css: blkcg of interest
1072  *
1073  * While pinned, a blkcg is kept online.  This is primarily used to
1074  * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1075  * while an associated cgwb is still active.
1076  */
1077 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1078 {
1079 	refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1080 }
1081 
1082 /**
1083  * blkcg_unpin_online - unpin online state
1084  * @blkcg_css: blkcg of interest
1085  *
1086  * This is primarily used to impedance-match blkg and cgwb lifetimes so
1087  * that blkg doesn't go offline while an associated cgwb is still active.
1088  * When this count goes to zero, all active cgwbs have finished so the
1089  * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1090  */
1091 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1092 {
1093 	struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1094 
1095 	do {
1096 		if (!refcount_dec_and_test(&blkcg->online_pin))
1097 			break;
1098 		blkcg_destroy_blkgs(blkcg);
1099 		blkcg = blkcg_parent(blkcg);
1100 	} while (blkcg);
1101 }
1102 
1103 /**
1104  * blkcg_css_offline - cgroup css_offline callback
1105  * @css: css of interest
1106  *
1107  * This function is called when @css is about to go away.  Here the cgwbs are
1108  * offlined first and only once writeback associated with the blkcg has
1109  * finished do we start step 2 (see above).
1110  */
1111 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1112 {
1113 	/* this prevents anyone from attaching or migrating to this blkcg */
1114 	wb_blkcg_offline(css);
1115 
1116 	/* put the base online pin allowing step 2 to be triggered */
1117 	blkcg_unpin_online(css);
1118 }
1119 
1120 static void blkcg_css_free(struct cgroup_subsys_state *css)
1121 {
1122 	struct blkcg *blkcg = css_to_blkcg(css);
1123 	int i;
1124 
1125 	mutex_lock(&blkcg_pol_mutex);
1126 
1127 	list_del(&blkcg->all_blkcgs_node);
1128 
1129 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1130 		if (blkcg->cpd[i])
1131 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1132 
1133 	mutex_unlock(&blkcg_pol_mutex);
1134 
1135 	kfree(blkcg);
1136 }
1137 
1138 static struct cgroup_subsys_state *
1139 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1140 {
1141 	struct blkcg *blkcg;
1142 	struct cgroup_subsys_state *ret;
1143 	int i;
1144 
1145 	mutex_lock(&blkcg_pol_mutex);
1146 
1147 	if (!parent_css) {
1148 		blkcg = &blkcg_root;
1149 	} else {
1150 		blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1151 		if (!blkcg) {
1152 			ret = ERR_PTR(-ENOMEM);
1153 			goto unlock;
1154 		}
1155 	}
1156 
1157 	for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1158 		struct blkcg_policy *pol = blkcg_policy[i];
1159 		struct blkcg_policy_data *cpd;
1160 
1161 		/*
1162 		 * If the policy hasn't been attached yet, wait for it
1163 		 * to be attached before doing anything else. Otherwise,
1164 		 * check if the policy requires any specific per-cgroup
1165 		 * data: if it does, allocate and initialize it.
1166 		 */
1167 		if (!pol || !pol->cpd_alloc_fn)
1168 			continue;
1169 
1170 		cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1171 		if (!cpd) {
1172 			ret = ERR_PTR(-ENOMEM);
1173 			goto free_pd_blkcg;
1174 		}
1175 		blkcg->cpd[i] = cpd;
1176 		cpd->blkcg = blkcg;
1177 		cpd->plid = i;
1178 		if (pol->cpd_init_fn)
1179 			pol->cpd_init_fn(cpd);
1180 	}
1181 
1182 	spin_lock_init(&blkcg->lock);
1183 	refcount_set(&blkcg->online_pin, 1);
1184 	INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1185 	INIT_HLIST_HEAD(&blkcg->blkg_list);
1186 #ifdef CONFIG_CGROUP_WRITEBACK
1187 	INIT_LIST_HEAD(&blkcg->cgwb_list);
1188 #endif
1189 	list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1190 
1191 	mutex_unlock(&blkcg_pol_mutex);
1192 	return &blkcg->css;
1193 
1194 free_pd_blkcg:
1195 	for (i--; i >= 0; i--)
1196 		if (blkcg->cpd[i])
1197 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1198 
1199 	if (blkcg != &blkcg_root)
1200 		kfree(blkcg);
1201 unlock:
1202 	mutex_unlock(&blkcg_pol_mutex);
1203 	return ret;
1204 }
1205 
1206 static int blkcg_css_online(struct cgroup_subsys_state *css)
1207 {
1208 	struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1209 
1210 	/*
1211 	 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1212 	 * don't go offline while cgwbs are still active on them.  Pin the
1213 	 * parent so that offline always happens towards the root.
1214 	 */
1215 	if (parent)
1216 		blkcg_pin_online(&parent->css);
1217 	return 0;
1218 }
1219 
1220 int blkcg_init_disk(struct gendisk *disk)
1221 {
1222 	struct request_queue *q = disk->queue;
1223 	struct blkcg_gq *new_blkg, *blkg;
1224 	bool preloaded;
1225 	int ret;
1226 
1227 	INIT_LIST_HEAD(&q->blkg_list);
1228 
1229 	new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1230 	if (!new_blkg)
1231 		return -ENOMEM;
1232 
1233 	preloaded = !radix_tree_preload(GFP_KERNEL);
1234 
1235 	/* Make sure the root blkg exists. */
1236 	/* spin_lock_irq can serve as RCU read-side critical section. */
1237 	spin_lock_irq(&q->queue_lock);
1238 	blkg = blkg_create(&blkcg_root, disk, new_blkg);
1239 	if (IS_ERR(blkg))
1240 		goto err_unlock;
1241 	q->root_blkg = blkg;
1242 	spin_unlock_irq(&q->queue_lock);
1243 
1244 	if (preloaded)
1245 		radix_tree_preload_end();
1246 
1247 	ret = blk_ioprio_init(disk);
1248 	if (ret)
1249 		goto err_destroy_all;
1250 
1251 	ret = blk_throtl_init(disk);
1252 	if (ret)
1253 		goto err_ioprio_exit;
1254 
1255 	ret = blk_iolatency_init(disk);
1256 	if (ret)
1257 		goto err_throtl_exit;
1258 
1259 	return 0;
1260 
1261 err_throtl_exit:
1262 	blk_throtl_exit(disk);
1263 err_ioprio_exit:
1264 	blk_ioprio_exit(disk);
1265 err_destroy_all:
1266 	blkg_destroy_all(disk);
1267 	return ret;
1268 err_unlock:
1269 	spin_unlock_irq(&q->queue_lock);
1270 	if (preloaded)
1271 		radix_tree_preload_end();
1272 	return PTR_ERR(blkg);
1273 }
1274 
1275 void blkcg_exit_disk(struct gendisk *disk)
1276 {
1277 	blkg_destroy_all(disk);
1278 	blk_throtl_exit(disk);
1279 }
1280 
1281 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1282 {
1283 	int i;
1284 
1285 	mutex_lock(&blkcg_pol_mutex);
1286 
1287 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
1288 		struct blkcg_policy *pol = blkcg_policy[i];
1289 		struct blkcg *blkcg;
1290 
1291 		if (!pol || !pol->cpd_bind_fn)
1292 			continue;
1293 
1294 		list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1295 			if (blkcg->cpd[pol->plid])
1296 				pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1297 	}
1298 	mutex_unlock(&blkcg_pol_mutex);
1299 }
1300 
1301 static void blkcg_exit(struct task_struct *tsk)
1302 {
1303 	if (tsk->throttle_queue)
1304 		blk_put_queue(tsk->throttle_queue);
1305 	tsk->throttle_queue = NULL;
1306 }
1307 
1308 struct cgroup_subsys io_cgrp_subsys = {
1309 	.css_alloc = blkcg_css_alloc,
1310 	.css_online = blkcg_css_online,
1311 	.css_offline = blkcg_css_offline,
1312 	.css_free = blkcg_css_free,
1313 	.css_rstat_flush = blkcg_rstat_flush,
1314 	.bind = blkcg_bind,
1315 	.dfl_cftypes = blkcg_files,
1316 	.legacy_cftypes = blkcg_legacy_files,
1317 	.legacy_name = "blkio",
1318 	.exit = blkcg_exit,
1319 #ifdef CONFIG_MEMCG
1320 	/*
1321 	 * This ensures that, if available, memcg is automatically enabled
1322 	 * together on the default hierarchy so that the owner cgroup can
1323 	 * be retrieved from writeback pages.
1324 	 */
1325 	.depends_on = 1 << memory_cgrp_id,
1326 #endif
1327 };
1328 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1329 
1330 /**
1331  * blkcg_activate_policy - activate a blkcg policy on a request_queue
1332  * @q: request_queue of interest
1333  * @pol: blkcg policy to activate
1334  *
1335  * Activate @pol on @q.  Requires %GFP_KERNEL context.  @q goes through
1336  * bypass mode to populate its blkgs with policy_data for @pol.
1337  *
1338  * Activation happens with @q bypassed, so nobody would be accessing blkgs
1339  * from IO path.  Update of each blkg is protected by both queue and blkcg
1340  * locks so that holding either lock and testing blkcg_policy_enabled() is
1341  * always enough for dereferencing policy data.
1342  *
1343  * The caller is responsible for synchronizing [de]activations and policy
1344  * [un]registerations.  Returns 0 on success, -errno on failure.
1345  */
1346 int blkcg_activate_policy(struct request_queue *q,
1347 			  const struct blkcg_policy *pol)
1348 {
1349 	struct blkg_policy_data *pd_prealloc = NULL;
1350 	struct blkcg_gq *blkg, *pinned_blkg = NULL;
1351 	int ret;
1352 
1353 	if (blkcg_policy_enabled(q, pol))
1354 		return 0;
1355 
1356 	if (queue_is_mq(q))
1357 		blk_mq_freeze_queue(q);
1358 retry:
1359 	spin_lock_irq(&q->queue_lock);
1360 
1361 	/* blkg_list is pushed at the head, reverse walk to allocate parents first */
1362 	list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1363 		struct blkg_policy_data *pd;
1364 
1365 		if (blkg->pd[pol->plid])
1366 			continue;
1367 
1368 		/* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1369 		if (blkg == pinned_blkg) {
1370 			pd = pd_prealloc;
1371 			pd_prealloc = NULL;
1372 		} else {
1373 			pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1374 					      blkg->blkcg);
1375 		}
1376 
1377 		if (!pd) {
1378 			/*
1379 			 * GFP_NOWAIT failed.  Free the existing one and
1380 			 * prealloc for @blkg w/ GFP_KERNEL.
1381 			 */
1382 			if (pinned_blkg)
1383 				blkg_put(pinned_blkg);
1384 			blkg_get(blkg);
1385 			pinned_blkg = blkg;
1386 
1387 			spin_unlock_irq(&q->queue_lock);
1388 
1389 			if (pd_prealloc)
1390 				pol->pd_free_fn(pd_prealloc);
1391 			pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1392 						       blkg->blkcg);
1393 			if (pd_prealloc)
1394 				goto retry;
1395 			else
1396 				goto enomem;
1397 		}
1398 
1399 		blkg->pd[pol->plid] = pd;
1400 		pd->blkg = blkg;
1401 		pd->plid = pol->plid;
1402 	}
1403 
1404 	/* all allocated, init in the same order */
1405 	if (pol->pd_init_fn)
1406 		list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1407 			pol->pd_init_fn(blkg->pd[pol->plid]);
1408 
1409 	__set_bit(pol->plid, q->blkcg_pols);
1410 	ret = 0;
1411 
1412 	spin_unlock_irq(&q->queue_lock);
1413 out:
1414 	if (queue_is_mq(q))
1415 		blk_mq_unfreeze_queue(q);
1416 	if (pinned_blkg)
1417 		blkg_put(pinned_blkg);
1418 	if (pd_prealloc)
1419 		pol->pd_free_fn(pd_prealloc);
1420 	return ret;
1421 
1422 enomem:
1423 	/* alloc failed, nothing's initialized yet, free everything */
1424 	spin_lock_irq(&q->queue_lock);
1425 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1426 		struct blkcg *blkcg = blkg->blkcg;
1427 
1428 		spin_lock(&blkcg->lock);
1429 		if (blkg->pd[pol->plid]) {
1430 			pol->pd_free_fn(blkg->pd[pol->plid]);
1431 			blkg->pd[pol->plid] = NULL;
1432 		}
1433 		spin_unlock(&blkcg->lock);
1434 	}
1435 	spin_unlock_irq(&q->queue_lock);
1436 	ret = -ENOMEM;
1437 	goto out;
1438 }
1439 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1440 
1441 /**
1442  * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1443  * @q: request_queue of interest
1444  * @pol: blkcg policy to deactivate
1445  *
1446  * Deactivate @pol on @q.  Follows the same synchronization rules as
1447  * blkcg_activate_policy().
1448  */
1449 void blkcg_deactivate_policy(struct request_queue *q,
1450 			     const struct blkcg_policy *pol)
1451 {
1452 	struct blkcg_gq *blkg;
1453 
1454 	if (!blkcg_policy_enabled(q, pol))
1455 		return;
1456 
1457 	if (queue_is_mq(q))
1458 		blk_mq_freeze_queue(q);
1459 
1460 	spin_lock_irq(&q->queue_lock);
1461 
1462 	__clear_bit(pol->plid, q->blkcg_pols);
1463 
1464 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1465 		struct blkcg *blkcg = blkg->blkcg;
1466 
1467 		spin_lock(&blkcg->lock);
1468 		if (blkg->pd[pol->plid]) {
1469 			if (pol->pd_offline_fn)
1470 				pol->pd_offline_fn(blkg->pd[pol->plid]);
1471 			pol->pd_free_fn(blkg->pd[pol->plid]);
1472 			blkg->pd[pol->plid] = NULL;
1473 		}
1474 		spin_unlock(&blkcg->lock);
1475 	}
1476 
1477 	spin_unlock_irq(&q->queue_lock);
1478 
1479 	if (queue_is_mq(q))
1480 		blk_mq_unfreeze_queue(q);
1481 }
1482 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1483 
1484 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1485 {
1486 	struct blkcg *blkcg;
1487 
1488 	list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1489 		if (blkcg->cpd[pol->plid]) {
1490 			pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1491 			blkcg->cpd[pol->plid] = NULL;
1492 		}
1493 	}
1494 }
1495 
1496 /**
1497  * blkcg_policy_register - register a blkcg policy
1498  * @pol: blkcg policy to register
1499  *
1500  * Register @pol with blkcg core.  Might sleep and @pol may be modified on
1501  * successful registration.  Returns 0 on success and -errno on failure.
1502  */
1503 int blkcg_policy_register(struct blkcg_policy *pol)
1504 {
1505 	struct blkcg *blkcg;
1506 	int i, ret;
1507 
1508 	mutex_lock(&blkcg_pol_register_mutex);
1509 	mutex_lock(&blkcg_pol_mutex);
1510 
1511 	/* find an empty slot */
1512 	ret = -ENOSPC;
1513 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1514 		if (!blkcg_policy[i])
1515 			break;
1516 	if (i >= BLKCG_MAX_POLS) {
1517 		pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1518 		goto err_unlock;
1519 	}
1520 
1521 	/* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1522 	if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1523 		(!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1524 		goto err_unlock;
1525 
1526 	/* register @pol */
1527 	pol->plid = i;
1528 	blkcg_policy[pol->plid] = pol;
1529 
1530 	/* allocate and install cpd's */
1531 	if (pol->cpd_alloc_fn) {
1532 		list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1533 			struct blkcg_policy_data *cpd;
1534 
1535 			cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1536 			if (!cpd)
1537 				goto err_free_cpds;
1538 
1539 			blkcg->cpd[pol->plid] = cpd;
1540 			cpd->blkcg = blkcg;
1541 			cpd->plid = pol->plid;
1542 			if (pol->cpd_init_fn)
1543 				pol->cpd_init_fn(cpd);
1544 		}
1545 	}
1546 
1547 	mutex_unlock(&blkcg_pol_mutex);
1548 
1549 	/* everything is in place, add intf files for the new policy */
1550 	if (pol->dfl_cftypes)
1551 		WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1552 					       pol->dfl_cftypes));
1553 	if (pol->legacy_cftypes)
1554 		WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1555 						  pol->legacy_cftypes));
1556 	mutex_unlock(&blkcg_pol_register_mutex);
1557 	return 0;
1558 
1559 err_free_cpds:
1560 	if (pol->cpd_free_fn)
1561 		blkcg_free_all_cpd(pol);
1562 
1563 	blkcg_policy[pol->plid] = NULL;
1564 err_unlock:
1565 	mutex_unlock(&blkcg_pol_mutex);
1566 	mutex_unlock(&blkcg_pol_register_mutex);
1567 	return ret;
1568 }
1569 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1570 
1571 /**
1572  * blkcg_policy_unregister - unregister a blkcg policy
1573  * @pol: blkcg policy to unregister
1574  *
1575  * Undo blkcg_policy_register(@pol).  Might sleep.
1576  */
1577 void blkcg_policy_unregister(struct blkcg_policy *pol)
1578 {
1579 	mutex_lock(&blkcg_pol_register_mutex);
1580 
1581 	if (WARN_ON(blkcg_policy[pol->plid] != pol))
1582 		goto out_unlock;
1583 
1584 	/* kill the intf files first */
1585 	if (pol->dfl_cftypes)
1586 		cgroup_rm_cftypes(pol->dfl_cftypes);
1587 	if (pol->legacy_cftypes)
1588 		cgroup_rm_cftypes(pol->legacy_cftypes);
1589 
1590 	/* remove cpds and unregister */
1591 	mutex_lock(&blkcg_pol_mutex);
1592 
1593 	if (pol->cpd_free_fn)
1594 		blkcg_free_all_cpd(pol);
1595 
1596 	blkcg_policy[pol->plid] = NULL;
1597 
1598 	mutex_unlock(&blkcg_pol_mutex);
1599 out_unlock:
1600 	mutex_unlock(&blkcg_pol_register_mutex);
1601 }
1602 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1603 
1604 bool __blkcg_punt_bio_submit(struct bio *bio)
1605 {
1606 	struct blkcg_gq *blkg = bio->bi_blkg;
1607 
1608 	/* consume the flag first */
1609 	bio->bi_opf &= ~REQ_CGROUP_PUNT;
1610 
1611 	/* never bounce for the root cgroup */
1612 	if (!blkg->parent)
1613 		return false;
1614 
1615 	spin_lock_bh(&blkg->async_bio_lock);
1616 	bio_list_add(&blkg->async_bios, bio);
1617 	spin_unlock_bh(&blkg->async_bio_lock);
1618 
1619 	queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1620 	return true;
1621 }
1622 
1623 /*
1624  * Scale the accumulated delay based on how long it has been since we updated
1625  * the delay.  We only call this when we are adding delay, in case it's been a
1626  * while since we added delay, and when we are checking to see if we need to
1627  * delay a task, to account for any delays that may have occurred.
1628  */
1629 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1630 {
1631 	u64 old = atomic64_read(&blkg->delay_start);
1632 
1633 	/* negative use_delay means no scaling, see blkcg_set_delay() */
1634 	if (atomic_read(&blkg->use_delay) < 0)
1635 		return;
1636 
1637 	/*
1638 	 * We only want to scale down every second.  The idea here is that we
1639 	 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1640 	 * time window.  We only want to throttle tasks for recent delay that
1641 	 * has occurred, in 1 second time windows since that's the maximum
1642 	 * things can be throttled.  We save the current delay window in
1643 	 * blkg->last_delay so we know what amount is still left to be charged
1644 	 * to the blkg from this point onward.  blkg->last_use keeps track of
1645 	 * the use_delay counter.  The idea is if we're unthrottling the blkg we
1646 	 * are ok with whatever is happening now, and we can take away more of
1647 	 * the accumulated delay as we've already throttled enough that
1648 	 * everybody is happy with their IO latencies.
1649 	 */
1650 	if (time_before64(old + NSEC_PER_SEC, now) &&
1651 	    atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1652 		u64 cur = atomic64_read(&blkg->delay_nsec);
1653 		u64 sub = min_t(u64, blkg->last_delay, now - old);
1654 		int cur_use = atomic_read(&blkg->use_delay);
1655 
1656 		/*
1657 		 * We've been unthrottled, subtract a larger chunk of our
1658 		 * accumulated delay.
1659 		 */
1660 		if (cur_use < blkg->last_use)
1661 			sub = max_t(u64, sub, blkg->last_delay >> 1);
1662 
1663 		/*
1664 		 * This shouldn't happen, but handle it anyway.  Our delay_nsec
1665 		 * should only ever be growing except here where we subtract out
1666 		 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1667 		 * rather not end up with negative numbers.
1668 		 */
1669 		if (unlikely(cur < sub)) {
1670 			atomic64_set(&blkg->delay_nsec, 0);
1671 			blkg->last_delay = 0;
1672 		} else {
1673 			atomic64_sub(sub, &blkg->delay_nsec);
1674 			blkg->last_delay = cur - sub;
1675 		}
1676 		blkg->last_use = cur_use;
1677 	}
1678 }
1679 
1680 /*
1681  * This is called when we want to actually walk up the hierarchy and check to
1682  * see if we need to throttle, and then actually throttle if there is some
1683  * accumulated delay.  This should only be called upon return to user space so
1684  * we're not holding some lock that would induce a priority inversion.
1685  */
1686 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1687 {
1688 	unsigned long pflags;
1689 	bool clamp;
1690 	u64 now = ktime_to_ns(ktime_get());
1691 	u64 exp;
1692 	u64 delay_nsec = 0;
1693 	int tok;
1694 
1695 	while (blkg->parent) {
1696 		int use_delay = atomic_read(&blkg->use_delay);
1697 
1698 		if (use_delay) {
1699 			u64 this_delay;
1700 
1701 			blkcg_scale_delay(blkg, now);
1702 			this_delay = atomic64_read(&blkg->delay_nsec);
1703 			if (this_delay > delay_nsec) {
1704 				delay_nsec = this_delay;
1705 				clamp = use_delay > 0;
1706 			}
1707 		}
1708 		blkg = blkg->parent;
1709 	}
1710 
1711 	if (!delay_nsec)
1712 		return;
1713 
1714 	/*
1715 	 * Let's not sleep for all eternity if we've amassed a huge delay.
1716 	 * Swapping or metadata IO can accumulate 10's of seconds worth of
1717 	 * delay, and we want userspace to be able to do _something_ so cap the
1718 	 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1719 	 * tasks will be delayed for 0.25 second for every syscall. If
1720 	 * blkcg_set_delay() was used as indicated by negative use_delay, the
1721 	 * caller is responsible for regulating the range.
1722 	 */
1723 	if (clamp)
1724 		delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1725 
1726 	if (use_memdelay)
1727 		psi_memstall_enter(&pflags);
1728 
1729 	exp = ktime_add_ns(now, delay_nsec);
1730 	tok = io_schedule_prepare();
1731 	do {
1732 		__set_current_state(TASK_KILLABLE);
1733 		if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1734 			break;
1735 	} while (!fatal_signal_pending(current));
1736 	io_schedule_finish(tok);
1737 
1738 	if (use_memdelay)
1739 		psi_memstall_leave(&pflags);
1740 }
1741 
1742 /**
1743  * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1744  *
1745  * This is only called if we've been marked with set_notify_resume().  Obviously
1746  * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1747  * check to see if current->throttle_queue is set and if not this doesn't do
1748  * anything.  This should only ever be called by the resume code, it's not meant
1749  * to be called by people willy-nilly as it will actually do the work to
1750  * throttle the task if it is setup for throttling.
1751  */
1752 void blkcg_maybe_throttle_current(void)
1753 {
1754 	struct request_queue *q = current->throttle_queue;
1755 	struct blkcg *blkcg;
1756 	struct blkcg_gq *blkg;
1757 	bool use_memdelay = current->use_memdelay;
1758 
1759 	if (!q)
1760 		return;
1761 
1762 	current->throttle_queue = NULL;
1763 	current->use_memdelay = false;
1764 
1765 	rcu_read_lock();
1766 	blkcg = css_to_blkcg(blkcg_css());
1767 	if (!blkcg)
1768 		goto out;
1769 	blkg = blkg_lookup(blkcg, q);
1770 	if (!blkg)
1771 		goto out;
1772 	if (!blkg_tryget(blkg))
1773 		goto out;
1774 	rcu_read_unlock();
1775 
1776 	blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1777 	blkg_put(blkg);
1778 	blk_put_queue(q);
1779 	return;
1780 out:
1781 	rcu_read_unlock();
1782 	blk_put_queue(q);
1783 }
1784 
1785 /**
1786  * blkcg_schedule_throttle - this task needs to check for throttling
1787  * @gendisk: disk to throttle
1788  * @use_memdelay: do we charge this to memory delay for PSI
1789  *
1790  * This is called by the IO controller when we know there's delay accumulated
1791  * for the blkg for this task.  We do not pass the blkg because there are places
1792  * we call this that may not have that information, the swapping code for
1793  * instance will only have a block_device at that point.  This set's the
1794  * notify_resume for the task to check and see if it requires throttling before
1795  * returning to user space.
1796  *
1797  * We will only schedule once per syscall.  You can call this over and over
1798  * again and it will only do the check once upon return to user space, and only
1799  * throttle once.  If the task needs to be throttled again it'll need to be
1800  * re-set at the next time we see the task.
1801  */
1802 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1803 {
1804 	struct request_queue *q = disk->queue;
1805 
1806 	if (unlikely(current->flags & PF_KTHREAD))
1807 		return;
1808 
1809 	if (current->throttle_queue != q) {
1810 		if (!blk_get_queue(q))
1811 			return;
1812 
1813 		if (current->throttle_queue)
1814 			blk_put_queue(current->throttle_queue);
1815 		current->throttle_queue = q;
1816 	}
1817 
1818 	if (use_memdelay)
1819 		current->use_memdelay = use_memdelay;
1820 	set_notify_resume(current);
1821 }
1822 
1823 /**
1824  * blkcg_add_delay - add delay to this blkg
1825  * @blkg: blkg of interest
1826  * @now: the current time in nanoseconds
1827  * @delta: how many nanoseconds of delay to add
1828  *
1829  * Charge @delta to the blkg's current delay accumulation.  This is used to
1830  * throttle tasks if an IO controller thinks we need more throttling.
1831  */
1832 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1833 {
1834 	if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1835 		return;
1836 	blkcg_scale_delay(blkg, now);
1837 	atomic64_add(delta, &blkg->delay_nsec);
1838 }
1839 
1840 /**
1841  * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1842  * @bio: target bio
1843  * @css: target css
1844  *
1845  * As the failure mode here is to walk up the blkg tree, this ensure that the
1846  * blkg->parent pointers are always valid.  This returns the blkg that it ended
1847  * up taking a reference on or %NULL if no reference was taken.
1848  */
1849 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1850 		struct cgroup_subsys_state *css)
1851 {
1852 	struct blkcg_gq *blkg, *ret_blkg = NULL;
1853 
1854 	rcu_read_lock();
1855 	blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1856 	while (blkg) {
1857 		if (blkg_tryget(blkg)) {
1858 			ret_blkg = blkg;
1859 			break;
1860 		}
1861 		blkg = blkg->parent;
1862 	}
1863 	rcu_read_unlock();
1864 
1865 	return ret_blkg;
1866 }
1867 
1868 /**
1869  * bio_associate_blkg_from_css - associate a bio with a specified css
1870  * @bio: target bio
1871  * @css: target css
1872  *
1873  * Associate @bio with the blkg found by combining the css's blkg and the
1874  * request_queue of the @bio.  An association failure is handled by walking up
1875  * the blkg tree.  Therefore, the blkg associated can be anything between @blkg
1876  * and q->root_blkg.  This situation only happens when a cgroup is dying and
1877  * then the remaining bios will spill to the closest alive blkg.
1878  *
1879  * A reference will be taken on the blkg and will be released when @bio is
1880  * freed.
1881  */
1882 void bio_associate_blkg_from_css(struct bio *bio,
1883 				 struct cgroup_subsys_state *css)
1884 {
1885 	if (bio->bi_blkg)
1886 		blkg_put(bio->bi_blkg);
1887 
1888 	if (css && css->parent) {
1889 		bio->bi_blkg = blkg_tryget_closest(bio, css);
1890 	} else {
1891 		blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1892 		bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1893 	}
1894 }
1895 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1896 
1897 /**
1898  * bio_associate_blkg - associate a bio with a blkg
1899  * @bio: target bio
1900  *
1901  * Associate @bio with the blkg found from the bio's css and request_queue.
1902  * If one is not found, bio_lookup_blkg() creates the blkg.  If a blkg is
1903  * already associated, the css is reused and association redone as the
1904  * request_queue may have changed.
1905  */
1906 void bio_associate_blkg(struct bio *bio)
1907 {
1908 	struct cgroup_subsys_state *css;
1909 
1910 	rcu_read_lock();
1911 
1912 	if (bio->bi_blkg)
1913 		css = bio_blkcg_css(bio);
1914 	else
1915 		css = blkcg_css();
1916 
1917 	bio_associate_blkg_from_css(bio, css);
1918 
1919 	rcu_read_unlock();
1920 }
1921 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1922 
1923 /**
1924  * bio_clone_blkg_association - clone blkg association from src to dst bio
1925  * @dst: destination bio
1926  * @src: source bio
1927  */
1928 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1929 {
1930 	if (src->bi_blkg)
1931 		bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
1932 }
1933 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1934 
1935 static int blk_cgroup_io_type(struct bio *bio)
1936 {
1937 	if (op_is_discard(bio->bi_opf))
1938 		return BLKG_IOSTAT_DISCARD;
1939 	if (op_is_write(bio->bi_opf))
1940 		return BLKG_IOSTAT_WRITE;
1941 	return BLKG_IOSTAT_READ;
1942 }
1943 
1944 void blk_cgroup_bio_start(struct bio *bio)
1945 {
1946 	int rwd = blk_cgroup_io_type(bio), cpu;
1947 	struct blkg_iostat_set *bis;
1948 	unsigned long flags;
1949 
1950 	cpu = get_cpu();
1951 	bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1952 	flags = u64_stats_update_begin_irqsave(&bis->sync);
1953 
1954 	/*
1955 	 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1956 	 * bio and we would have already accounted for the size of the bio.
1957 	 */
1958 	if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1959 		bio_set_flag(bio, BIO_CGROUP_ACCT);
1960 		bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1961 	}
1962 	bis->cur.ios[rwd]++;
1963 
1964 	u64_stats_update_end_irqrestore(&bis->sync, flags);
1965 	if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1966 		cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1967 	put_cpu();
1968 }
1969 
1970 bool blk_cgroup_congested(void)
1971 {
1972 	struct cgroup_subsys_state *css;
1973 	bool ret = false;
1974 
1975 	rcu_read_lock();
1976 	for (css = blkcg_css(); css; css = css->parent) {
1977 		if (atomic_read(&css->cgroup->congestion_count)) {
1978 			ret = true;
1979 			break;
1980 		}
1981 	}
1982 	rcu_read_unlock();
1983 	return ret;
1984 }
1985 
1986 static int __init blkcg_init(void)
1987 {
1988 	blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1989 					    WQ_MEM_RECLAIM | WQ_FREEZABLE |
1990 					    WQ_UNBOUND | WQ_SYSFS, 0);
1991 	if (!blkcg_punt_bio_wq)
1992 		return -ENOMEM;
1993 	return 0;
1994 }
1995 subsys_initcall(blkcg_init);
1996 
1997 module_param(blkcg_debug_stats, bool, 0644);
1998 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");
1999