xref: /openbmc/linux/block/blk-cgroup.c (revision 16f6ccde74a6f8538c62f127f17207c75f4dba7a)
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  */
init_blkcg_llists(struct blkcg * blkcg)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  */
blkcg_css(void)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, io_cgrp_id);
111 }
112 
blkcg_policy_enabled(struct request_queue * q,const struct blkcg_policy * pol)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 
blkg_free_workfn(struct work_struct * work)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->parent);
139 	spin_lock_irq(&q->queue_lock);
140 	list_del_init(&blkg->q_node);
141 	spin_unlock_irq(&q->queue_lock);
142 	mutex_unlock(&q->blkcg_mutex);
143 
144 	blk_put_queue(q);
145 	free_percpu(blkg->iostat_cpu);
146 	percpu_ref_exit(&blkg->refcnt);
147 	kfree(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  */
blkg_free(struct blkcg_gq * blkg)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(&blkg->free_work);
167 }
168 
__blkg_release(struct rcu_head * rcu)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(&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  */
blkg_release(struct percpu_ref * ref)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(&blkg->rcu_head, __blkg_release);
205 }
206 
207 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
208 static struct workqueue_struct *blkcg_punt_bio_wq;
209 
blkg_async_bio_workfn(struct work_struct * work)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(&blkg->async_bio_lock);
221 	bio_list_merge(&bios, &blkg->async_bios);
222 	bio_list_init(&blkg->async_bios);
223 	spin_unlock(&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(&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  */
blkcg_punt_bio_submit(struct bio * bio)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(&blkg->async_bio_lock);
248 		bio_list_add(&blkg->async_bios, bio);
249 		spin_unlock(&blkg->async_bio_lock);
250 		queue_work(blkcg_punt_bio_wq, &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 
blkcg_punt_bio_init(void)258 static int __init blkcg_punt_bio_init(void)
259 {
260 	blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
261 					    WQ_MEM_RECLAIM | WQ_FREEZABLE |
262 					    WQ_UNBOUND | WQ_SYSFS, 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  */
bio_blkcg_css(struct bio * bio)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  */
blkcg_parent(struct blkcg * blkcg)292 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
293 {
294 	return css_to_blkcg(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 assocating @blkcg and @q.
304  */
blkg_alloc(struct blkcg * blkcg,struct gendisk * disk,gfp_t gfp_mask)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(sizeof(*blkg), gfp_mask, disk->queue->node);
313 	if (!blkg)
314 		return NULL;
315 	if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, 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(&blkg->q_node);
325 	blkg->blkcg = blkcg;
326 	blkg->iostat.blkg = blkg;
327 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
328 	spin_lock_init(&blkg->async_bio_lock);
329 	bio_list_init(&blkg->async_bios);
330 	INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
331 #endif
332 
333 	u64_stats_init(&blkg->iostat.sync);
334 	for_each_possible_cpu(cpu) {
335 		u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
336 		per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
337 	}
338 
339 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
340 		struct blkcg_policy *pol = blkcg_policy[i];
341 		struct blkg_policy_data *pd;
342 
343 		if (!blkcg_policy_enabled(disk->queue, pol))
344 			continue;
345 
346 		/* alloc per-policy data and attach it to blkg */
347 		pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
348 		if (!pd)
349 			goto out_free_pds;
350 		blkg->pd[i] = pd;
351 		pd->blkg = blkg;
352 		pd->plid = i;
353 		pd->online = false;
354 	}
355 
356 	return blkg;
357 
358 out_free_pds:
359 	while (--i >= 0)
360 		if (blkg->pd[i])
361 			blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
362 	blk_put_queue(disk->queue);
363 out_free_iostat:
364 	free_percpu(blkg->iostat_cpu);
365 out_exit_refcnt:
366 	percpu_ref_exit(&blkg->refcnt);
367 out_free_blkg:
368 	kfree(blkg);
369 	return NULL;
370 }
371 
372 /*
373  * If @new_blkg is %NULL, this function tries to allocate a new one as
374  * necessary using %GFP_NOWAIT.  @new_blkg is always consumed on return.
375  */
blkg_create(struct blkcg * blkcg,struct gendisk * disk,struct blkcg_gq * new_blkg)376 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
377 				    struct blkcg_gq *new_blkg)
378 {
379 	struct blkcg_gq *blkg;
380 	int i, ret;
381 
382 	lockdep_assert_held(&disk->queue->queue_lock);
383 
384 	/* request_queue is dying, do not create/recreate a blkg */
385 	if (blk_queue_dying(disk->queue)) {
386 		ret = -ENODEV;
387 		goto err_free_blkg;
388 	}
389 
390 	/* blkg holds a reference to blkcg */
391 	if (!css_tryget_online(&blkcg->css)) {
392 		ret = -ENODEV;
393 		goto err_free_blkg;
394 	}
395 
396 	/* allocate */
397 	if (!new_blkg) {
398 		new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
399 		if (unlikely(!new_blkg)) {
400 			ret = -ENOMEM;
401 			goto err_put_css;
402 		}
403 	}
404 	blkg = new_blkg;
405 
406 	/* link parent */
407 	if (blkcg_parent(blkcg)) {
408 		blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
409 		if (WARN_ON_ONCE(!blkg->parent)) {
410 			ret = -ENODEV;
411 			goto err_put_css;
412 		}
413 		blkg_get(blkg->parent);
414 	}
415 
416 	/* invoke per-policy init */
417 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
418 		struct blkcg_policy *pol = blkcg_policy[i];
419 
420 		if (blkg->pd[i] && pol->pd_init_fn)
421 			pol->pd_init_fn(blkg->pd[i]);
422 	}
423 
424 	/* insert */
425 	spin_lock(&blkcg->lock);
426 	ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
427 	if (likely(!ret)) {
428 		hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
429 		list_add(&blkg->q_node, &disk->queue->blkg_list);
430 
431 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
432 			struct blkcg_policy *pol = blkcg_policy[i];
433 
434 			if (blkg->pd[i]) {
435 				if (pol->pd_online_fn)
436 					pol->pd_online_fn(blkg->pd[i]);
437 				blkg->pd[i]->online = true;
438 			}
439 		}
440 	}
441 	blkg->online = true;
442 	spin_unlock(&blkcg->lock);
443 
444 	if (!ret)
445 		return blkg;
446 
447 	/* @blkg failed fully initialized, use the usual release path */
448 	blkg_put(blkg);
449 	return ERR_PTR(ret);
450 
451 err_put_css:
452 	css_put(&blkcg->css);
453 err_free_blkg:
454 	if (new_blkg)
455 		blkg_free(new_blkg);
456 	return ERR_PTR(ret);
457 }
458 
459 /**
460  * blkg_lookup_create - lookup blkg, try to create one if not there
461  * @blkcg: blkcg of interest
462  * @disk: gendisk of interest
463  *
464  * Lookup blkg for the @blkcg - @disk pair.  If it doesn't exist, try to
465  * create one.  blkg creation is performed recursively from blkcg_root such
466  * that all non-root blkg's have access to the parent blkg.  This function
467  * should be called under RCU read lock and takes @disk->queue->queue_lock.
468  *
469  * Returns the blkg or the closest blkg if blkg_create() fails as it walks
470  * down from root.
471  */
blkg_lookup_create(struct blkcg * blkcg,struct gendisk * disk)472 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
473 		struct gendisk *disk)
474 {
475 	struct request_queue *q = disk->queue;
476 	struct blkcg_gq *blkg;
477 	unsigned long flags;
478 
479 	WARN_ON_ONCE(!rcu_read_lock_held());
480 
481 	blkg = blkg_lookup(blkcg, q);
482 	if (blkg)
483 		return blkg;
484 
485 	spin_lock_irqsave(&q->queue_lock, flags);
486 	blkg = blkg_lookup(blkcg, q);
487 	if (blkg) {
488 		if (blkcg != &blkcg_root &&
489 		    blkg != rcu_dereference(blkcg->blkg_hint))
490 			rcu_assign_pointer(blkcg->blkg_hint, blkg);
491 		goto found;
492 	}
493 
494 	/*
495 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
496 	 * non-root blkgs have access to their parents.  Returns the closest
497 	 * blkg to the intended blkg should blkg_create() fail.
498 	 */
499 	while (true) {
500 		struct blkcg *pos = blkcg;
501 		struct blkcg *parent = blkcg_parent(blkcg);
502 		struct blkcg_gq *ret_blkg = q->root_blkg;
503 
504 		while (parent) {
505 			blkg = blkg_lookup(parent, q);
506 			if (blkg) {
507 				/* remember closest blkg */
508 				ret_blkg = blkg;
509 				break;
510 			}
511 			pos = parent;
512 			parent = blkcg_parent(parent);
513 		}
514 
515 		blkg = blkg_create(pos, disk, NULL);
516 		if (IS_ERR(blkg)) {
517 			blkg = ret_blkg;
518 			break;
519 		}
520 		if (pos == blkcg)
521 			break;
522 	}
523 
524 found:
525 	spin_unlock_irqrestore(&q->queue_lock, flags);
526 	return blkg;
527 }
528 
blkg_destroy(struct blkcg_gq * blkg)529 static void blkg_destroy(struct blkcg_gq *blkg)
530 {
531 	struct blkcg *blkcg = blkg->blkcg;
532 	int i;
533 
534 	lockdep_assert_held(&blkg->q->queue_lock);
535 	lockdep_assert_held(&blkcg->lock);
536 
537 	/*
538 	 * blkg stays on the queue list until blkg_free_workfn(), see details in
539 	 * blkg_free_workfn(), hence this function can be called from
540 	 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
541 	 * blkg_free_workfn().
542 	 */
543 	if (hlist_unhashed(&blkg->blkcg_node))
544 		return;
545 
546 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
547 		struct blkcg_policy *pol = blkcg_policy[i];
548 
549 		if (blkg->pd[i] && blkg->pd[i]->online) {
550 			blkg->pd[i]->online = false;
551 			if (pol->pd_offline_fn)
552 				pol->pd_offline_fn(blkg->pd[i]);
553 		}
554 	}
555 
556 	blkg->online = false;
557 
558 	radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
559 	hlist_del_init_rcu(&blkg->blkcg_node);
560 
561 	/*
562 	 * Both setting lookup hint to and clearing it from @blkg are done
563 	 * under queue_lock.  If it's not pointing to @blkg now, it never
564 	 * will.  Hint assignment itself can race safely.
565 	 */
566 	if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
567 		rcu_assign_pointer(blkcg->blkg_hint, NULL);
568 
569 	/*
570 	 * Put the reference taken at the time of creation so that when all
571 	 * queues are gone, group can be destroyed.
572 	 */
573 	percpu_ref_kill(&blkg->refcnt);
574 }
575 
blkg_destroy_all(struct gendisk * disk)576 static void blkg_destroy_all(struct gendisk *disk)
577 {
578 	struct request_queue *q = disk->queue;
579 	struct blkcg_gq *blkg, *n;
580 	int count = BLKG_DESTROY_BATCH_SIZE;
581 	int i;
582 
583 restart:
584 	spin_lock_irq(&q->queue_lock);
585 	list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
586 		struct blkcg *blkcg = blkg->blkcg;
587 
588 		if (hlist_unhashed(&blkg->blkcg_node))
589 			continue;
590 
591 		spin_lock(&blkcg->lock);
592 		blkg_destroy(blkg);
593 		spin_unlock(&blkcg->lock);
594 
595 		/*
596 		 * in order to avoid holding the spin lock for too long, release
597 		 * it when a batch of blkgs are destroyed.
598 		 */
599 		if (!(--count)) {
600 			count = BLKG_DESTROY_BATCH_SIZE;
601 			spin_unlock_irq(&q->queue_lock);
602 			cond_resched();
603 			goto restart;
604 		}
605 	}
606 
607 	/*
608 	 * Mark policy deactivated since policy offline has been done, and
609 	 * the free is scheduled, so future blkcg_deactivate_policy() can
610 	 * be bypassed
611 	 */
612 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
613 		struct blkcg_policy *pol = blkcg_policy[i];
614 
615 		if (pol)
616 			__clear_bit(pol->plid, q->blkcg_pols);
617 	}
618 
619 	q->root_blkg = NULL;
620 	spin_unlock_irq(&q->queue_lock);
621 }
622 
blkg_iostat_set(struct blkg_iostat * dst,struct blkg_iostat * src)623 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
624 {
625 	int i;
626 
627 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
628 		dst->bytes[i] = src->bytes[i];
629 		dst->ios[i] = src->ios[i];
630 	}
631 }
632 
__blkg_clear_stat(struct blkg_iostat_set * bis)633 static void __blkg_clear_stat(struct blkg_iostat_set *bis)
634 {
635 	struct blkg_iostat cur = {0};
636 	unsigned long flags;
637 
638 	flags = u64_stats_update_begin_irqsave(&bis->sync);
639 	blkg_iostat_set(&bis->cur, &cur);
640 	blkg_iostat_set(&bis->last, &cur);
641 	u64_stats_update_end_irqrestore(&bis->sync, flags);
642 }
643 
blkg_clear_stat(struct blkcg_gq * blkg)644 static void blkg_clear_stat(struct blkcg_gq *blkg)
645 {
646 	int cpu;
647 
648 	for_each_possible_cpu(cpu) {
649 		struct blkg_iostat_set *s = per_cpu_ptr(blkg->iostat_cpu, cpu);
650 
651 		__blkg_clear_stat(s);
652 	}
653 	__blkg_clear_stat(&blkg->iostat);
654 }
655 
blkcg_reset_stats(struct cgroup_subsys_state * css,struct cftype * cftype,u64 val)656 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
657 			     struct cftype *cftype, u64 val)
658 {
659 	struct blkcg *blkcg = css_to_blkcg(css);
660 	struct blkcg_gq *blkg;
661 	int i;
662 
663 	mutex_lock(&blkcg_pol_mutex);
664 	spin_lock_irq(&blkcg->lock);
665 
666 	/*
667 	 * Note that stat reset is racy - it doesn't synchronize against
668 	 * stat updates.  This is a debug feature which shouldn't exist
669 	 * anyway.  If you get hit by a race, retry.
670 	 */
671 	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
672 		blkg_clear_stat(blkg);
673 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
674 			struct blkcg_policy *pol = blkcg_policy[i];
675 
676 			if (blkg->pd[i] && pol->pd_reset_stats_fn)
677 				pol->pd_reset_stats_fn(blkg->pd[i]);
678 		}
679 	}
680 
681 	spin_unlock_irq(&blkcg->lock);
682 	mutex_unlock(&blkcg_pol_mutex);
683 	return 0;
684 }
685 
blkg_dev_name(struct blkcg_gq * blkg)686 const char *blkg_dev_name(struct blkcg_gq *blkg)
687 {
688 	if (!blkg->q->disk)
689 		return NULL;
690 	return bdi_dev_name(blkg->q->disk->bdi);
691 }
692 
693 /**
694  * blkcg_print_blkgs - helper for printing per-blkg data
695  * @sf: seq_file to print to
696  * @blkcg: blkcg of interest
697  * @prfill: fill function to print out a blkg
698  * @pol: policy in question
699  * @data: data to be passed to @prfill
700  * @show_total: to print out sum of prfill return values or not
701  *
702  * This function invokes @prfill on each blkg of @blkcg if pd for the
703  * policy specified by @pol exists.  @prfill is invoked with @sf, the
704  * policy data and @data and the matching queue lock held.  If @show_total
705  * is %true, the sum of the return values from @prfill is printed with
706  * "Total" label at the end.
707  *
708  * This is to be used to construct print functions for
709  * cftype->read_seq_string method.
710  */
blkcg_print_blkgs(struct seq_file * sf,struct blkcg * blkcg,u64 (* prfill)(struct seq_file *,struct blkg_policy_data *,int),const struct blkcg_policy * pol,int data,bool show_total)711 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
712 		       u64 (*prfill)(struct seq_file *,
713 				     struct blkg_policy_data *, int),
714 		       const struct blkcg_policy *pol, int data,
715 		       bool show_total)
716 {
717 	struct blkcg_gq *blkg;
718 	u64 total = 0;
719 
720 	rcu_read_lock();
721 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
722 		spin_lock_irq(&blkg->q->queue_lock);
723 		if (blkcg_policy_enabled(blkg->q, pol))
724 			total += prfill(sf, blkg->pd[pol->plid], data);
725 		spin_unlock_irq(&blkg->q->queue_lock);
726 	}
727 	rcu_read_unlock();
728 
729 	if (show_total)
730 		seq_printf(sf, "Total %llu\n", (unsigned long long)total);
731 }
732 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
733 
734 /**
735  * __blkg_prfill_u64 - prfill helper for a single u64 value
736  * @sf: seq_file to print to
737  * @pd: policy private data of interest
738  * @v: value to print
739  *
740  * Print @v to @sf for the device associated with @pd.
741  */
__blkg_prfill_u64(struct seq_file * sf,struct blkg_policy_data * pd,u64 v)742 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
743 {
744 	const char *dname = blkg_dev_name(pd->blkg);
745 
746 	if (!dname)
747 		return 0;
748 
749 	seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
750 	return v;
751 }
752 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
753 
754 /**
755  * blkg_conf_init - initialize a blkg_conf_ctx
756  * @ctx: blkg_conf_ctx to initialize
757  * @input: input string
758  *
759  * Initialize @ctx which can be used to parse blkg config input string @input.
760  * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
761  * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
762  */
blkg_conf_init(struct blkg_conf_ctx * ctx,char * input)763 void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
764 {
765 	*ctx = (struct blkg_conf_ctx){ .input = input };
766 }
767 EXPORT_SYMBOL_GPL(blkg_conf_init);
768 
769 /**
770  * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
771  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
772  *
773  * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
774  * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
775  * set to point past the device node prefix.
776  *
777  * This function may be called multiple times on @ctx and the extra calls become
778  * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
779  * explicitly if bdev access is needed without resolving the blkcg / policy part
780  * of @ctx->input. Returns -errno on error.
781  */
blkg_conf_open_bdev(struct blkg_conf_ctx * ctx)782 int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
783 {
784 	char *input = ctx->input;
785 	unsigned int major, minor;
786 	struct block_device *bdev;
787 	int key_len;
788 
789 	if (ctx->bdev)
790 		return 0;
791 
792 	if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
793 		return -EINVAL;
794 
795 	input += key_len;
796 	if (!isspace(*input))
797 		return -EINVAL;
798 	input = skip_spaces(input);
799 
800 	bdev = blkdev_get_no_open(MKDEV(major, minor));
801 	if (!bdev)
802 		return -ENODEV;
803 	if (bdev_is_partition(bdev)) {
804 		blkdev_put_no_open(bdev);
805 		return -ENODEV;
806 	}
807 
808 	mutex_lock(&bdev->bd_queue->rq_qos_mutex);
809 	if (!disk_live(bdev->bd_disk)) {
810 		blkdev_put_no_open(bdev);
811 		mutex_unlock(&bdev->bd_queue->rq_qos_mutex);
812 		return -ENODEV;
813 	}
814 
815 	ctx->body = input;
816 	ctx->bdev = bdev;
817 	return 0;
818 }
819 
820 /**
821  * blkg_conf_prep - parse and prepare for per-blkg config update
822  * @blkcg: target block cgroup
823  * @pol: target policy
824  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
825  *
826  * Parse per-blkg config update from @ctx->input and initialize @ctx
827  * accordingly. On success, @ctx->body points to the part of @ctx->input
828  * following MAJ:MIN, @ctx->bdev points to the target block device and
829  * @ctx->blkg to the blkg being configured.
830  *
831  * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
832  * function returns with queue lock held and must be followed by
833  * blkg_conf_exit().
834  */
blkg_conf_prep(struct blkcg * blkcg,const struct blkcg_policy * pol,struct blkg_conf_ctx * ctx)835 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
836 		   struct blkg_conf_ctx *ctx)
837 	__acquires(&bdev->bd_queue->queue_lock)
838 {
839 	struct gendisk *disk;
840 	struct request_queue *q;
841 	struct blkcg_gq *blkg;
842 	int ret;
843 
844 	ret = blkg_conf_open_bdev(ctx);
845 	if (ret)
846 		return ret;
847 
848 	disk = ctx->bdev->bd_disk;
849 	q = disk->queue;
850 
851 	/*
852 	 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
853 	 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
854 	 */
855 	ret = blk_queue_enter(q, 0);
856 	if (ret)
857 		goto fail;
858 
859 	spin_lock_irq(&q->queue_lock);
860 
861 	if (!blkcg_policy_enabled(q, pol)) {
862 		ret = -EOPNOTSUPP;
863 		goto fail_unlock;
864 	}
865 
866 	blkg = blkg_lookup(blkcg, q);
867 	if (blkg)
868 		goto success;
869 
870 	/*
871 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
872 	 * non-root blkgs have access to their parents.
873 	 */
874 	while (true) {
875 		struct blkcg *pos = blkcg;
876 		struct blkcg *parent;
877 		struct blkcg_gq *new_blkg;
878 
879 		parent = blkcg_parent(blkcg);
880 		while (parent && !blkg_lookup(parent, q)) {
881 			pos = parent;
882 			parent = blkcg_parent(parent);
883 		}
884 
885 		/* Drop locks to do new blkg allocation with GFP_KERNEL. */
886 		spin_unlock_irq(&q->queue_lock);
887 
888 		new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
889 		if (unlikely(!new_blkg)) {
890 			ret = -ENOMEM;
891 			goto fail_exit_queue;
892 		}
893 
894 		if (radix_tree_preload(GFP_KERNEL)) {
895 			blkg_free(new_blkg);
896 			ret = -ENOMEM;
897 			goto fail_exit_queue;
898 		}
899 
900 		spin_lock_irq(&q->queue_lock);
901 
902 		if (!blkcg_policy_enabled(q, pol)) {
903 			blkg_free(new_blkg);
904 			ret = -EOPNOTSUPP;
905 			goto fail_preloaded;
906 		}
907 
908 		blkg = blkg_lookup(pos, q);
909 		if (blkg) {
910 			blkg_free(new_blkg);
911 		} else {
912 			blkg = blkg_create(pos, disk, new_blkg);
913 			if (IS_ERR(blkg)) {
914 				ret = PTR_ERR(blkg);
915 				goto fail_preloaded;
916 			}
917 		}
918 
919 		radix_tree_preload_end();
920 
921 		if (pos == blkcg)
922 			goto success;
923 	}
924 success:
925 	blk_queue_exit(q);
926 	ctx->blkg = blkg;
927 	return 0;
928 
929 fail_preloaded:
930 	radix_tree_preload_end();
931 fail_unlock:
932 	spin_unlock_irq(&q->queue_lock);
933 fail_exit_queue:
934 	blk_queue_exit(q);
935 fail:
936 	/*
937 	 * If queue was bypassing, we should retry.  Do so after a
938 	 * short msleep().  It isn't strictly necessary but queue
939 	 * can be bypassing for some time and it's always nice to
940 	 * avoid busy looping.
941 	 */
942 	if (ret == -EBUSY) {
943 		msleep(10);
944 		ret = restart_syscall();
945 	}
946 	return ret;
947 }
948 EXPORT_SYMBOL_GPL(blkg_conf_prep);
949 
950 /**
951  * blkg_conf_exit - clean up per-blkg config update
952  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
953  *
954  * Clean up after per-blkg config update. This function must be called on all
955  * blkg_conf_ctx's initialized with blkg_conf_init().
956  */
blkg_conf_exit(struct blkg_conf_ctx * ctx)957 void blkg_conf_exit(struct blkg_conf_ctx *ctx)
958 	__releases(&ctx->bdev->bd_queue->queue_lock)
959 	__releases(&ctx->bdev->bd_queue->rq_qos_mutex)
960 {
961 	if (ctx->blkg) {
962 		spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
963 		ctx->blkg = NULL;
964 	}
965 
966 	if (ctx->bdev) {
967 		mutex_unlock(&ctx->bdev->bd_queue->rq_qos_mutex);
968 		blkdev_put_no_open(ctx->bdev);
969 		ctx->body = NULL;
970 		ctx->bdev = NULL;
971 	}
972 }
973 EXPORT_SYMBOL_GPL(blkg_conf_exit);
974 
blkg_iostat_add(struct blkg_iostat * dst,struct blkg_iostat * src)975 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
976 {
977 	int i;
978 
979 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
980 		dst->bytes[i] += src->bytes[i];
981 		dst->ios[i] += src->ios[i];
982 	}
983 }
984 
blkg_iostat_sub(struct blkg_iostat * dst,struct blkg_iostat * src)985 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
986 {
987 	int i;
988 
989 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
990 		dst->bytes[i] -= src->bytes[i];
991 		dst->ios[i] -= src->ios[i];
992 	}
993 }
994 
blkcg_iostat_update(struct blkcg_gq * blkg,struct blkg_iostat * cur,struct blkg_iostat * last)995 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
996 				struct blkg_iostat *last)
997 {
998 	struct blkg_iostat delta;
999 	unsigned long flags;
1000 
1001 	/* propagate percpu delta to global */
1002 	flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1003 	blkg_iostat_set(&delta, cur);
1004 	blkg_iostat_sub(&delta, last);
1005 	blkg_iostat_add(&blkg->iostat.cur, &delta);
1006 	blkg_iostat_add(last, &delta);
1007 	u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1008 }
1009 
__blkcg_rstat_flush(struct blkcg * blkcg,int cpu)1010 static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
1011 {
1012 	struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
1013 	struct llist_node *lnode;
1014 	struct blkg_iostat_set *bisc, *next_bisc;
1015 	unsigned long flags;
1016 
1017 	rcu_read_lock();
1018 
1019 	lnode = llist_del_all(lhead);
1020 	if (!lnode)
1021 		goto out;
1022 
1023 	/*
1024 	 * For covering concurrent parent blkg update from blkg_release().
1025 	 *
1026 	 * When flushing from cgroup, cgroup_rstat_lock is always held, so
1027 	 * this lock won't cause contention most of time.
1028 	 */
1029 	raw_spin_lock_irqsave(&blkg_stat_lock, flags);
1030 
1031 	/*
1032 	 * Iterate only the iostat_cpu's queued in the lockless list.
1033 	 */
1034 	llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
1035 		struct blkcg_gq *blkg = bisc->blkg;
1036 		struct blkcg_gq *parent = blkg->parent;
1037 		struct blkg_iostat cur;
1038 		unsigned int seq;
1039 
1040 		/*
1041 		 * Order assignment of `next_bisc` from `bisc->lnode.next` in
1042 		 * llist_for_each_entry_safe and clearing `bisc->lqueued` for
1043 		 * avoiding to assign `next_bisc` with new next pointer added
1044 		 * in blk_cgroup_bio_start() in case of re-ordering.
1045 		 *
1046 		 * The pair barrier is implied in llist_add() in blk_cgroup_bio_start().
1047 		 */
1048 		smp_mb();
1049 
1050 		WRITE_ONCE(bisc->lqueued, false);
1051 		if (bisc == &blkg->iostat)
1052 			goto propagate_up; /* propagate up to parent only */
1053 
1054 		/* fetch the current per-cpu values */
1055 		do {
1056 			seq = u64_stats_fetch_begin(&bisc->sync);
1057 			blkg_iostat_set(&cur, &bisc->cur);
1058 		} while (u64_stats_fetch_retry(&bisc->sync, seq));
1059 
1060 		blkcg_iostat_update(blkg, &cur, &bisc->last);
1061 
1062 propagate_up:
1063 		/* propagate global delta to parent (unless that's root) */
1064 		if (parent && parent->parent) {
1065 			blkcg_iostat_update(parent, &blkg->iostat.cur,
1066 					    &blkg->iostat.last);
1067 			/*
1068 			 * Queue parent->iostat to its blkcg's lockless
1069 			 * list to propagate up to the grandparent if the
1070 			 * iostat hasn't been queued yet.
1071 			 */
1072 			if (!parent->iostat.lqueued) {
1073 				struct llist_head *plhead;
1074 
1075 				plhead = per_cpu_ptr(parent->blkcg->lhead, cpu);
1076 				llist_add(&parent->iostat.lnode, plhead);
1077 				parent->iostat.lqueued = true;
1078 			}
1079 		}
1080 	}
1081 	raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
1082 out:
1083 	rcu_read_unlock();
1084 }
1085 
blkcg_rstat_flush(struct cgroup_subsys_state * css,int cpu)1086 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1087 {
1088 	/* Root-level stats are sourced from system-wide IO stats */
1089 	if (cgroup_parent(css->cgroup))
1090 		__blkcg_rstat_flush(css_to_blkcg(css), cpu);
1091 }
1092 
1093 /*
1094  * We source root cgroup stats from the system-wide stats to avoid
1095  * tracking the same information twice and incurring overhead when no
1096  * cgroups are defined. For that reason, cgroup_rstat_flush in
1097  * blkcg_print_stat does not actually fill out the iostat in the root
1098  * cgroup's blkcg_gq.
1099  *
1100  * However, we would like to re-use the printing code between the root and
1101  * non-root cgroups to the extent possible. For that reason, we simulate
1102  * flushing the root cgroup's stats by explicitly filling in the iostat
1103  * with disk level statistics.
1104  */
blkcg_fill_root_iostats(void)1105 static void blkcg_fill_root_iostats(void)
1106 {
1107 	struct class_dev_iter iter;
1108 	struct device *dev;
1109 
1110 	class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
1111 	while ((dev = class_dev_iter_next(&iter))) {
1112 		struct block_device *bdev = dev_to_bdev(dev);
1113 		struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
1114 		struct blkg_iostat tmp;
1115 		int cpu;
1116 		unsigned long flags;
1117 
1118 		memset(&tmp, 0, sizeof(tmp));
1119 		for_each_possible_cpu(cpu) {
1120 			struct disk_stats *cpu_dkstats;
1121 
1122 			cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1123 			tmp.ios[BLKG_IOSTAT_READ] +=
1124 				cpu_dkstats->ios[STAT_READ];
1125 			tmp.ios[BLKG_IOSTAT_WRITE] +=
1126 				cpu_dkstats->ios[STAT_WRITE];
1127 			tmp.ios[BLKG_IOSTAT_DISCARD] +=
1128 				cpu_dkstats->ios[STAT_DISCARD];
1129 			// convert sectors to bytes
1130 			tmp.bytes[BLKG_IOSTAT_READ] +=
1131 				cpu_dkstats->sectors[STAT_READ] << 9;
1132 			tmp.bytes[BLKG_IOSTAT_WRITE] +=
1133 				cpu_dkstats->sectors[STAT_WRITE] << 9;
1134 			tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1135 				cpu_dkstats->sectors[STAT_DISCARD] << 9;
1136 		}
1137 
1138 		flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1139 		blkg_iostat_set(&blkg->iostat.cur, &tmp);
1140 		u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1141 	}
1142 }
1143 
blkcg_print_one_stat(struct blkcg_gq * blkg,struct seq_file * s)1144 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1145 {
1146 	struct blkg_iostat_set *bis = &blkg->iostat;
1147 	u64 rbytes, wbytes, rios, wios, dbytes, dios;
1148 	const char *dname;
1149 	unsigned seq;
1150 	int i;
1151 
1152 	if (!blkg->online)
1153 		return;
1154 
1155 	dname = blkg_dev_name(blkg);
1156 	if (!dname)
1157 		return;
1158 
1159 	seq_printf(s, "%s ", dname);
1160 
1161 	do {
1162 		seq = u64_stats_fetch_begin(&bis->sync);
1163 
1164 		rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1165 		wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1166 		dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1167 		rios = bis->cur.ios[BLKG_IOSTAT_READ];
1168 		wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1169 		dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1170 	} while (u64_stats_fetch_retry(&bis->sync, seq));
1171 
1172 	if (rbytes || wbytes || rios || wios) {
1173 		seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1174 			rbytes, wbytes, rios, wios,
1175 			dbytes, dios);
1176 	}
1177 
1178 	if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1179 		seq_printf(s, " use_delay=%d delay_nsec=%llu",
1180 			atomic_read(&blkg->use_delay),
1181 			atomic64_read(&blkg->delay_nsec));
1182 	}
1183 
1184 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
1185 		struct blkcg_policy *pol = blkcg_policy[i];
1186 
1187 		if (!blkg->pd[i] || !pol->pd_stat_fn)
1188 			continue;
1189 
1190 		pol->pd_stat_fn(blkg->pd[i], s);
1191 	}
1192 
1193 	seq_puts(s, "\n");
1194 }
1195 
blkcg_print_stat(struct seq_file * sf,void * v)1196 static int blkcg_print_stat(struct seq_file *sf, void *v)
1197 {
1198 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1199 	struct blkcg_gq *blkg;
1200 
1201 	if (!seq_css(sf)->parent)
1202 		blkcg_fill_root_iostats();
1203 	else
1204 		cgroup_rstat_flush(blkcg->css.cgroup);
1205 
1206 	rcu_read_lock();
1207 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1208 		spin_lock_irq(&blkg->q->queue_lock);
1209 		blkcg_print_one_stat(blkg, sf);
1210 		spin_unlock_irq(&blkg->q->queue_lock);
1211 	}
1212 	rcu_read_unlock();
1213 	return 0;
1214 }
1215 
1216 static struct cftype blkcg_files[] = {
1217 	{
1218 		.name = "stat",
1219 		.seq_show = blkcg_print_stat,
1220 	},
1221 	{ }	/* terminate */
1222 };
1223 
1224 static struct cftype blkcg_legacy_files[] = {
1225 	{
1226 		.name = "reset_stats",
1227 		.write_u64 = blkcg_reset_stats,
1228 	},
1229 	{ }	/* terminate */
1230 };
1231 
1232 #ifdef CONFIG_CGROUP_WRITEBACK
blkcg_get_cgwb_list(struct cgroup_subsys_state * css)1233 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1234 {
1235 	return &css_to_blkcg(css)->cgwb_list;
1236 }
1237 #endif
1238 
1239 /*
1240  * blkcg destruction is a three-stage process.
1241  *
1242  * 1. Destruction starts.  The blkcg_css_offline() callback is invoked
1243  *    which offlines writeback.  Here we tie the next stage of blkg destruction
1244  *    to the completion of writeback associated with the blkcg.  This lets us
1245  *    avoid punting potentially large amounts of outstanding writeback to root
1246  *    while maintaining any ongoing policies.  The next stage is triggered when
1247  *    the nr_cgwbs count goes to zero.
1248  *
1249  * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1250  *    and handles the destruction of blkgs.  Here the css reference held by
1251  *    the blkg is put back eventually allowing blkcg_css_free() to be called.
1252  *    This work may occur in cgwb_release_workfn() on the cgwb_release
1253  *    workqueue.  Any submitted ios that fail to get the blkg ref will be
1254  *    punted to the root_blkg.
1255  *
1256  * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1257  *    This finally frees the blkcg.
1258  */
1259 
1260 /**
1261  * blkcg_destroy_blkgs - responsible for shooting down blkgs
1262  * @blkcg: blkcg of interest
1263  *
1264  * blkgs should be removed while holding both q and blkcg locks.  As blkcg lock
1265  * is nested inside q lock, this function performs reverse double lock dancing.
1266  * Destroying the blkgs releases the reference held on the blkcg's css allowing
1267  * blkcg_css_free to eventually be called.
1268  *
1269  * This is the blkcg counterpart of ioc_release_fn().
1270  */
blkcg_destroy_blkgs(struct blkcg * blkcg)1271 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1272 {
1273 	might_sleep();
1274 
1275 	spin_lock_irq(&blkcg->lock);
1276 
1277 	while (!hlist_empty(&blkcg->blkg_list)) {
1278 		struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1279 						struct blkcg_gq, blkcg_node);
1280 		struct request_queue *q = blkg->q;
1281 
1282 		if (need_resched() || !spin_trylock(&q->queue_lock)) {
1283 			/*
1284 			 * Given that the system can accumulate a huge number
1285 			 * of blkgs in pathological cases, check to see if we
1286 			 * need to rescheduling to avoid softlockup.
1287 			 */
1288 			spin_unlock_irq(&blkcg->lock);
1289 			cond_resched();
1290 			spin_lock_irq(&blkcg->lock);
1291 			continue;
1292 		}
1293 
1294 		blkg_destroy(blkg);
1295 		spin_unlock(&q->queue_lock);
1296 	}
1297 
1298 	spin_unlock_irq(&blkcg->lock);
1299 }
1300 
1301 /**
1302  * blkcg_pin_online - pin online state
1303  * @blkcg_css: blkcg of interest
1304  *
1305  * While pinned, a blkcg is kept online.  This is primarily used to
1306  * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1307  * while an associated cgwb is still active.
1308  */
blkcg_pin_online(struct cgroup_subsys_state * blkcg_css)1309 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1310 {
1311 	refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1312 }
1313 
1314 /**
1315  * blkcg_unpin_online - unpin online state
1316  * @blkcg_css: blkcg of interest
1317  *
1318  * This is primarily used to impedance-match blkg and cgwb lifetimes so
1319  * that blkg doesn't go offline while an associated cgwb is still active.
1320  * When this count goes to zero, all active cgwbs have finished so the
1321  * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1322  */
blkcg_unpin_online(struct cgroup_subsys_state * blkcg_css)1323 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1324 {
1325 	struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1326 
1327 	do {
1328 		struct blkcg *parent;
1329 
1330 		if (!refcount_dec_and_test(&blkcg->online_pin))
1331 			break;
1332 
1333 		parent = blkcg_parent(blkcg);
1334 		blkcg_destroy_blkgs(blkcg);
1335 		blkcg = parent;
1336 	} while (blkcg);
1337 }
1338 
1339 /**
1340  * blkcg_css_offline - cgroup css_offline callback
1341  * @css: css of interest
1342  *
1343  * This function is called when @css is about to go away.  Here the cgwbs are
1344  * offlined first and only once writeback associated with the blkcg has
1345  * finished do we start step 2 (see above).
1346  */
blkcg_css_offline(struct cgroup_subsys_state * css)1347 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1348 {
1349 	/* this prevents anyone from attaching or migrating to this blkcg */
1350 	wb_blkcg_offline(css);
1351 
1352 	/* put the base online pin allowing step 2 to be triggered */
1353 	blkcg_unpin_online(css);
1354 }
1355 
blkcg_css_free(struct cgroup_subsys_state * css)1356 static void blkcg_css_free(struct cgroup_subsys_state *css)
1357 {
1358 	struct blkcg *blkcg = css_to_blkcg(css);
1359 	int i;
1360 
1361 	mutex_lock(&blkcg_pol_mutex);
1362 
1363 	list_del(&blkcg->all_blkcgs_node);
1364 
1365 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1366 		if (blkcg->cpd[i])
1367 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1368 
1369 	mutex_unlock(&blkcg_pol_mutex);
1370 
1371 	free_percpu(blkcg->lhead);
1372 	kfree(blkcg);
1373 }
1374 
1375 static struct cgroup_subsys_state *
blkcg_css_alloc(struct cgroup_subsys_state * parent_css)1376 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1377 {
1378 	struct blkcg *blkcg;
1379 	int i;
1380 
1381 	mutex_lock(&blkcg_pol_mutex);
1382 
1383 	if (!parent_css) {
1384 		blkcg = &blkcg_root;
1385 	} else {
1386 		blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1387 		if (!blkcg)
1388 			goto unlock;
1389 	}
1390 
1391 	if (init_blkcg_llists(blkcg))
1392 		goto free_blkcg;
1393 
1394 	for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1395 		struct blkcg_policy *pol = blkcg_policy[i];
1396 		struct blkcg_policy_data *cpd;
1397 
1398 		/*
1399 		 * If the policy hasn't been attached yet, wait for it
1400 		 * to be attached before doing anything else. Otherwise,
1401 		 * check if the policy requires any specific per-cgroup
1402 		 * data: if it does, allocate and initialize it.
1403 		 */
1404 		if (!pol || !pol->cpd_alloc_fn)
1405 			continue;
1406 
1407 		cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1408 		if (!cpd)
1409 			goto free_pd_blkcg;
1410 
1411 		blkcg->cpd[i] = cpd;
1412 		cpd->blkcg = blkcg;
1413 		cpd->plid = i;
1414 	}
1415 
1416 	spin_lock_init(&blkcg->lock);
1417 	refcount_set(&blkcg->online_pin, 1);
1418 	INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1419 	INIT_HLIST_HEAD(&blkcg->blkg_list);
1420 #ifdef CONFIG_CGROUP_WRITEBACK
1421 	INIT_LIST_HEAD(&blkcg->cgwb_list);
1422 #endif
1423 	list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1424 
1425 	mutex_unlock(&blkcg_pol_mutex);
1426 	return &blkcg->css;
1427 
1428 free_pd_blkcg:
1429 	for (i--; i >= 0; i--)
1430 		if (blkcg->cpd[i])
1431 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1432 	free_percpu(blkcg->lhead);
1433 free_blkcg:
1434 	if (blkcg != &blkcg_root)
1435 		kfree(blkcg);
1436 unlock:
1437 	mutex_unlock(&blkcg_pol_mutex);
1438 	return ERR_PTR(-ENOMEM);
1439 }
1440 
blkcg_css_online(struct cgroup_subsys_state * css)1441 static int blkcg_css_online(struct cgroup_subsys_state *css)
1442 {
1443 	struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1444 
1445 	/*
1446 	 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1447 	 * don't go offline while cgwbs are still active on them.  Pin the
1448 	 * parent so that offline always happens towards the root.
1449 	 */
1450 	if (parent)
1451 		blkcg_pin_online(&parent->css);
1452 	return 0;
1453 }
1454 
blkg_init_queue(struct request_queue * q)1455 void blkg_init_queue(struct request_queue *q)
1456 {
1457 	INIT_LIST_HEAD(&q->blkg_list);
1458 	mutex_init(&q->blkcg_mutex);
1459 }
1460 
blkcg_init_disk(struct gendisk * disk)1461 int blkcg_init_disk(struct gendisk *disk)
1462 {
1463 	struct request_queue *q = disk->queue;
1464 	struct blkcg_gq *new_blkg, *blkg;
1465 	bool preloaded;
1466 	int ret;
1467 
1468 	new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1469 	if (!new_blkg)
1470 		return -ENOMEM;
1471 
1472 	preloaded = !radix_tree_preload(GFP_KERNEL);
1473 
1474 	/* Make sure the root blkg exists. */
1475 	/* spin_lock_irq can serve as RCU read-side critical section. */
1476 	spin_lock_irq(&q->queue_lock);
1477 	blkg = blkg_create(&blkcg_root, disk, new_blkg);
1478 	if (IS_ERR(blkg))
1479 		goto err_unlock;
1480 	q->root_blkg = blkg;
1481 	spin_unlock_irq(&q->queue_lock);
1482 
1483 	if (preloaded)
1484 		radix_tree_preload_end();
1485 
1486 	ret = blk_ioprio_init(disk);
1487 	if (ret)
1488 		goto err_destroy_all;
1489 
1490 	ret = blk_throtl_init(disk);
1491 	if (ret)
1492 		goto err_ioprio_exit;
1493 
1494 	return 0;
1495 
1496 err_ioprio_exit:
1497 	blk_ioprio_exit(disk);
1498 err_destroy_all:
1499 	blkg_destroy_all(disk);
1500 	return ret;
1501 err_unlock:
1502 	spin_unlock_irq(&q->queue_lock);
1503 	if (preloaded)
1504 		radix_tree_preload_end();
1505 	return PTR_ERR(blkg);
1506 }
1507 
blkcg_exit_disk(struct gendisk * disk)1508 void blkcg_exit_disk(struct gendisk *disk)
1509 {
1510 	blkg_destroy_all(disk);
1511 	blk_throtl_exit(disk);
1512 }
1513 
blkcg_exit(struct task_struct * tsk)1514 static void blkcg_exit(struct task_struct *tsk)
1515 {
1516 	if (tsk->throttle_disk)
1517 		put_disk(tsk->throttle_disk);
1518 	tsk->throttle_disk = NULL;
1519 }
1520 
1521 struct cgroup_subsys io_cgrp_subsys = {
1522 	.css_alloc = blkcg_css_alloc,
1523 	.css_online = blkcg_css_online,
1524 	.css_offline = blkcg_css_offline,
1525 	.css_free = blkcg_css_free,
1526 	.css_rstat_flush = blkcg_rstat_flush,
1527 	.dfl_cftypes = blkcg_files,
1528 	.legacy_cftypes = blkcg_legacy_files,
1529 	.legacy_name = "blkio",
1530 	.exit = blkcg_exit,
1531 #ifdef CONFIG_MEMCG
1532 	/*
1533 	 * This ensures that, if available, memcg is automatically enabled
1534 	 * together on the default hierarchy so that the owner cgroup can
1535 	 * be retrieved from writeback pages.
1536 	 */
1537 	.depends_on = 1 << memory_cgrp_id,
1538 #endif
1539 };
1540 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1541 
1542 /**
1543  * blkcg_activate_policy - activate a blkcg policy on a gendisk
1544  * @disk: gendisk of interest
1545  * @pol: blkcg policy to activate
1546  *
1547  * Activate @pol on @disk.  Requires %GFP_KERNEL context.  @disk goes through
1548  * bypass mode to populate its blkgs with policy_data for @pol.
1549  *
1550  * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1551  * from IO path.  Update of each blkg is protected by both queue and blkcg
1552  * locks so that holding either lock and testing blkcg_policy_enabled() is
1553  * always enough for dereferencing policy data.
1554  *
1555  * The caller is responsible for synchronizing [de]activations and policy
1556  * [un]registerations.  Returns 0 on success, -errno on failure.
1557  */
blkcg_activate_policy(struct gendisk * disk,const struct blkcg_policy * pol)1558 int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1559 {
1560 	struct request_queue *q = disk->queue;
1561 	struct blkg_policy_data *pd_prealloc = NULL;
1562 	struct blkcg_gq *blkg, *pinned_blkg = NULL;
1563 	int ret;
1564 
1565 	if (blkcg_policy_enabled(q, pol))
1566 		return 0;
1567 
1568 	if (queue_is_mq(q))
1569 		blk_mq_freeze_queue(q);
1570 retry:
1571 	spin_lock_irq(&q->queue_lock);
1572 
1573 	/* blkg_list is pushed at the head, reverse walk to initialize parents first */
1574 	list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1575 		struct blkg_policy_data *pd;
1576 
1577 		if (blkg->pd[pol->plid])
1578 			continue;
1579 
1580 		/* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1581 		if (blkg == pinned_blkg) {
1582 			pd = pd_prealloc;
1583 			pd_prealloc = NULL;
1584 		} else {
1585 			pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1586 					      GFP_NOWAIT | __GFP_NOWARN);
1587 		}
1588 
1589 		if (!pd) {
1590 			/*
1591 			 * GFP_NOWAIT failed.  Free the existing one and
1592 			 * prealloc for @blkg w/ GFP_KERNEL.
1593 			 */
1594 			if (pinned_blkg)
1595 				blkg_put(pinned_blkg);
1596 			blkg_get(blkg);
1597 			pinned_blkg = blkg;
1598 
1599 			spin_unlock_irq(&q->queue_lock);
1600 
1601 			if (pd_prealloc)
1602 				pol->pd_free_fn(pd_prealloc);
1603 			pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1604 						       GFP_KERNEL);
1605 			if (pd_prealloc)
1606 				goto retry;
1607 			else
1608 				goto enomem;
1609 		}
1610 
1611 		spin_lock(&blkg->blkcg->lock);
1612 
1613 		pd->blkg = blkg;
1614 		pd->plid = pol->plid;
1615 		blkg->pd[pol->plid] = pd;
1616 
1617 		if (pol->pd_init_fn)
1618 			pol->pd_init_fn(pd);
1619 
1620 		if (pol->pd_online_fn)
1621 			pol->pd_online_fn(pd);
1622 		pd->online = true;
1623 
1624 		spin_unlock(&blkg->blkcg->lock);
1625 	}
1626 
1627 	__set_bit(pol->plid, q->blkcg_pols);
1628 	ret = 0;
1629 
1630 	spin_unlock_irq(&q->queue_lock);
1631 out:
1632 	if (queue_is_mq(q))
1633 		blk_mq_unfreeze_queue(q);
1634 	if (pinned_blkg)
1635 		blkg_put(pinned_blkg);
1636 	if (pd_prealloc)
1637 		pol->pd_free_fn(pd_prealloc);
1638 	return ret;
1639 
1640 enomem:
1641 	/* alloc failed, take down everything */
1642 	spin_lock_irq(&q->queue_lock);
1643 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1644 		struct blkcg *blkcg = blkg->blkcg;
1645 		struct blkg_policy_data *pd;
1646 
1647 		spin_lock(&blkcg->lock);
1648 		pd = blkg->pd[pol->plid];
1649 		if (pd) {
1650 			if (pd->online && pol->pd_offline_fn)
1651 				pol->pd_offline_fn(pd);
1652 			pd->online = false;
1653 			pol->pd_free_fn(pd);
1654 			blkg->pd[pol->plid] = NULL;
1655 		}
1656 		spin_unlock(&blkcg->lock);
1657 	}
1658 	spin_unlock_irq(&q->queue_lock);
1659 	ret = -ENOMEM;
1660 	goto out;
1661 }
1662 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1663 
1664 /**
1665  * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1666  * @disk: gendisk of interest
1667  * @pol: blkcg policy to deactivate
1668  *
1669  * Deactivate @pol on @disk.  Follows the same synchronization rules as
1670  * blkcg_activate_policy().
1671  */
blkcg_deactivate_policy(struct gendisk * disk,const struct blkcg_policy * pol)1672 void blkcg_deactivate_policy(struct gendisk *disk,
1673 			     const struct blkcg_policy *pol)
1674 {
1675 	struct request_queue *q = disk->queue;
1676 	struct blkcg_gq *blkg;
1677 
1678 	if (!blkcg_policy_enabled(q, pol))
1679 		return;
1680 
1681 	if (queue_is_mq(q))
1682 		blk_mq_freeze_queue(q);
1683 
1684 	mutex_lock(&q->blkcg_mutex);
1685 	spin_lock_irq(&q->queue_lock);
1686 
1687 	__clear_bit(pol->plid, q->blkcg_pols);
1688 
1689 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1690 		struct blkcg *blkcg = blkg->blkcg;
1691 
1692 		spin_lock(&blkcg->lock);
1693 		if (blkg->pd[pol->plid]) {
1694 			if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1695 				pol->pd_offline_fn(blkg->pd[pol->plid]);
1696 			pol->pd_free_fn(blkg->pd[pol->plid]);
1697 			blkg->pd[pol->plid] = NULL;
1698 		}
1699 		spin_unlock(&blkcg->lock);
1700 	}
1701 
1702 	spin_unlock_irq(&q->queue_lock);
1703 	mutex_unlock(&q->blkcg_mutex);
1704 
1705 	if (queue_is_mq(q))
1706 		blk_mq_unfreeze_queue(q);
1707 }
1708 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1709 
blkcg_free_all_cpd(struct blkcg_policy * pol)1710 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1711 {
1712 	struct blkcg *blkcg;
1713 
1714 	list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1715 		if (blkcg->cpd[pol->plid]) {
1716 			pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1717 			blkcg->cpd[pol->plid] = NULL;
1718 		}
1719 	}
1720 }
1721 
1722 /**
1723  * blkcg_policy_register - register a blkcg policy
1724  * @pol: blkcg policy to register
1725  *
1726  * Register @pol with blkcg core.  Might sleep and @pol may be modified on
1727  * successful registration.  Returns 0 on success and -errno on failure.
1728  */
blkcg_policy_register(struct blkcg_policy * pol)1729 int blkcg_policy_register(struct blkcg_policy *pol)
1730 {
1731 	struct blkcg *blkcg;
1732 	int i, ret;
1733 
1734 	mutex_lock(&blkcg_pol_register_mutex);
1735 	mutex_lock(&blkcg_pol_mutex);
1736 
1737 	/* find an empty slot */
1738 	ret = -ENOSPC;
1739 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1740 		if (!blkcg_policy[i])
1741 			break;
1742 	if (i >= BLKCG_MAX_POLS) {
1743 		pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1744 		goto err_unlock;
1745 	}
1746 
1747 	/* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1748 	if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1749 		(!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1750 		goto err_unlock;
1751 
1752 	/* register @pol */
1753 	pol->plid = i;
1754 	blkcg_policy[pol->plid] = pol;
1755 
1756 	/* allocate and install cpd's */
1757 	if (pol->cpd_alloc_fn) {
1758 		list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1759 			struct blkcg_policy_data *cpd;
1760 
1761 			cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1762 			if (!cpd)
1763 				goto err_free_cpds;
1764 
1765 			blkcg->cpd[pol->plid] = cpd;
1766 			cpd->blkcg = blkcg;
1767 			cpd->plid = pol->plid;
1768 		}
1769 	}
1770 
1771 	mutex_unlock(&blkcg_pol_mutex);
1772 
1773 	/* everything is in place, add intf files for the new policy */
1774 	if (pol->dfl_cftypes)
1775 		WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1776 					       pol->dfl_cftypes));
1777 	if (pol->legacy_cftypes)
1778 		WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1779 						  pol->legacy_cftypes));
1780 	mutex_unlock(&blkcg_pol_register_mutex);
1781 	return 0;
1782 
1783 err_free_cpds:
1784 	if (pol->cpd_free_fn)
1785 		blkcg_free_all_cpd(pol);
1786 
1787 	blkcg_policy[pol->plid] = NULL;
1788 err_unlock:
1789 	mutex_unlock(&blkcg_pol_mutex);
1790 	mutex_unlock(&blkcg_pol_register_mutex);
1791 	return ret;
1792 }
1793 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1794 
1795 /**
1796  * blkcg_policy_unregister - unregister a blkcg policy
1797  * @pol: blkcg policy to unregister
1798  *
1799  * Undo blkcg_policy_register(@pol).  Might sleep.
1800  */
blkcg_policy_unregister(struct blkcg_policy * pol)1801 void blkcg_policy_unregister(struct blkcg_policy *pol)
1802 {
1803 	mutex_lock(&blkcg_pol_register_mutex);
1804 
1805 	if (WARN_ON(blkcg_policy[pol->plid] != pol))
1806 		goto out_unlock;
1807 
1808 	/* kill the intf files first */
1809 	if (pol->dfl_cftypes)
1810 		cgroup_rm_cftypes(pol->dfl_cftypes);
1811 	if (pol->legacy_cftypes)
1812 		cgroup_rm_cftypes(pol->legacy_cftypes);
1813 
1814 	/* remove cpds and unregister */
1815 	mutex_lock(&blkcg_pol_mutex);
1816 
1817 	if (pol->cpd_free_fn)
1818 		blkcg_free_all_cpd(pol);
1819 
1820 	blkcg_policy[pol->plid] = NULL;
1821 
1822 	mutex_unlock(&blkcg_pol_mutex);
1823 out_unlock:
1824 	mutex_unlock(&blkcg_pol_register_mutex);
1825 }
1826 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1827 
1828 /*
1829  * Scale the accumulated delay based on how long it has been since we updated
1830  * the delay.  We only call this when we are adding delay, in case it's been a
1831  * while since we added delay, and when we are checking to see if we need to
1832  * delay a task, to account for any delays that may have occurred.
1833  */
blkcg_scale_delay(struct blkcg_gq * blkg,u64 now)1834 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1835 {
1836 	u64 old = atomic64_read(&blkg->delay_start);
1837 
1838 	/* negative use_delay means no scaling, see blkcg_set_delay() */
1839 	if (atomic_read(&blkg->use_delay) < 0)
1840 		return;
1841 
1842 	/*
1843 	 * We only want to scale down every second.  The idea here is that we
1844 	 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1845 	 * time window.  We only want to throttle tasks for recent delay that
1846 	 * has occurred, in 1 second time windows since that's the maximum
1847 	 * things can be throttled.  We save the current delay window in
1848 	 * blkg->last_delay so we know what amount is still left to be charged
1849 	 * to the blkg from this point onward.  blkg->last_use keeps track of
1850 	 * the use_delay counter.  The idea is if we're unthrottling the blkg we
1851 	 * are ok with whatever is happening now, and we can take away more of
1852 	 * the accumulated delay as we've already throttled enough that
1853 	 * everybody is happy with their IO latencies.
1854 	 */
1855 	if (time_before64(old + NSEC_PER_SEC, now) &&
1856 	    atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1857 		u64 cur = atomic64_read(&blkg->delay_nsec);
1858 		u64 sub = min_t(u64, blkg->last_delay, now - old);
1859 		int cur_use = atomic_read(&blkg->use_delay);
1860 
1861 		/*
1862 		 * We've been unthrottled, subtract a larger chunk of our
1863 		 * accumulated delay.
1864 		 */
1865 		if (cur_use < blkg->last_use)
1866 			sub = max_t(u64, sub, blkg->last_delay >> 1);
1867 
1868 		/*
1869 		 * This shouldn't happen, but handle it anyway.  Our delay_nsec
1870 		 * should only ever be growing except here where we subtract out
1871 		 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1872 		 * rather not end up with negative numbers.
1873 		 */
1874 		if (unlikely(cur < sub)) {
1875 			atomic64_set(&blkg->delay_nsec, 0);
1876 			blkg->last_delay = 0;
1877 		} else {
1878 			atomic64_sub(sub, &blkg->delay_nsec);
1879 			blkg->last_delay = cur - sub;
1880 		}
1881 		blkg->last_use = cur_use;
1882 	}
1883 }
1884 
1885 /*
1886  * This is called when we want to actually walk up the hierarchy and check to
1887  * see if we need to throttle, and then actually throttle if there is some
1888  * accumulated delay.  This should only be called upon return to user space so
1889  * we're not holding some lock that would induce a priority inversion.
1890  */
blkcg_maybe_throttle_blkg(struct blkcg_gq * blkg,bool use_memdelay)1891 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1892 {
1893 	unsigned long pflags;
1894 	bool clamp;
1895 	u64 now = ktime_to_ns(ktime_get());
1896 	u64 exp;
1897 	u64 delay_nsec = 0;
1898 	int tok;
1899 
1900 	while (blkg->parent) {
1901 		int use_delay = atomic_read(&blkg->use_delay);
1902 
1903 		if (use_delay) {
1904 			u64 this_delay;
1905 
1906 			blkcg_scale_delay(blkg, now);
1907 			this_delay = atomic64_read(&blkg->delay_nsec);
1908 			if (this_delay > delay_nsec) {
1909 				delay_nsec = this_delay;
1910 				clamp = use_delay > 0;
1911 			}
1912 		}
1913 		blkg = blkg->parent;
1914 	}
1915 
1916 	if (!delay_nsec)
1917 		return;
1918 
1919 	/*
1920 	 * Let's not sleep for all eternity if we've amassed a huge delay.
1921 	 * Swapping or metadata IO can accumulate 10's of seconds worth of
1922 	 * delay, and we want userspace to be able to do _something_ so cap the
1923 	 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1924 	 * tasks will be delayed for 0.25 second for every syscall. If
1925 	 * blkcg_set_delay() was used as indicated by negative use_delay, the
1926 	 * caller is responsible for regulating the range.
1927 	 */
1928 	if (clamp)
1929 		delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1930 
1931 	if (use_memdelay)
1932 		psi_memstall_enter(&pflags);
1933 
1934 	exp = ktime_add_ns(now, delay_nsec);
1935 	tok = io_schedule_prepare();
1936 	do {
1937 		__set_current_state(TASK_KILLABLE);
1938 		if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1939 			break;
1940 	} while (!fatal_signal_pending(current));
1941 	io_schedule_finish(tok);
1942 
1943 	if (use_memdelay)
1944 		psi_memstall_leave(&pflags);
1945 }
1946 
1947 /**
1948  * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1949  *
1950  * This is only called if we've been marked with set_notify_resume().  Obviously
1951  * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1952  * check to see if current->throttle_disk is set and if not this doesn't do
1953  * anything.  This should only ever be called by the resume code, it's not meant
1954  * to be called by people willy-nilly as it will actually do the work to
1955  * throttle the task if it is setup for throttling.
1956  */
blkcg_maybe_throttle_current(void)1957 void blkcg_maybe_throttle_current(void)
1958 {
1959 	struct gendisk *disk = current->throttle_disk;
1960 	struct blkcg *blkcg;
1961 	struct blkcg_gq *blkg;
1962 	bool use_memdelay = current->use_memdelay;
1963 
1964 	if (!disk)
1965 		return;
1966 
1967 	current->throttle_disk = NULL;
1968 	current->use_memdelay = false;
1969 
1970 	rcu_read_lock();
1971 	blkcg = css_to_blkcg(blkcg_css());
1972 	if (!blkcg)
1973 		goto out;
1974 	blkg = blkg_lookup(blkcg, disk->queue);
1975 	if (!blkg)
1976 		goto out;
1977 	if (!blkg_tryget(blkg))
1978 		goto out;
1979 	rcu_read_unlock();
1980 
1981 	blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1982 	blkg_put(blkg);
1983 	put_disk(disk);
1984 	return;
1985 out:
1986 	rcu_read_unlock();
1987 }
1988 
1989 /**
1990  * blkcg_schedule_throttle - this task needs to check for throttling
1991  * @disk: disk to throttle
1992  * @use_memdelay: do we charge this to memory delay for PSI
1993  *
1994  * This is called by the IO controller when we know there's delay accumulated
1995  * for the blkg for this task.  We do not pass the blkg because there are places
1996  * we call this that may not have that information, the swapping code for
1997  * instance will only have a block_device at that point.  This set's the
1998  * notify_resume for the task to check and see if it requires throttling before
1999  * returning to user space.
2000  *
2001  * We will only schedule once per syscall.  You can call this over and over
2002  * again and it will only do the check once upon return to user space, and only
2003  * throttle once.  If the task needs to be throttled again it'll need to be
2004  * re-set at the next time we see the task.
2005  */
blkcg_schedule_throttle(struct gendisk * disk,bool use_memdelay)2006 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
2007 {
2008 	if (unlikely(current->flags & PF_KTHREAD))
2009 		return;
2010 
2011 	if (current->throttle_disk != disk) {
2012 		if (test_bit(GD_DEAD, &disk->state))
2013 			return;
2014 		get_device(disk_to_dev(disk));
2015 
2016 		if (current->throttle_disk)
2017 			put_disk(current->throttle_disk);
2018 		current->throttle_disk = disk;
2019 	}
2020 
2021 	if (use_memdelay)
2022 		current->use_memdelay = use_memdelay;
2023 	set_notify_resume(current);
2024 }
2025 
2026 /**
2027  * blkcg_add_delay - add delay to this blkg
2028  * @blkg: blkg of interest
2029  * @now: the current time in nanoseconds
2030  * @delta: how many nanoseconds of delay to add
2031  *
2032  * Charge @delta to the blkg's current delay accumulation.  This is used to
2033  * throttle tasks if an IO controller thinks we need more throttling.
2034  */
blkcg_add_delay(struct blkcg_gq * blkg,u64 now,u64 delta)2035 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
2036 {
2037 	if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
2038 		return;
2039 	blkcg_scale_delay(blkg, now);
2040 	atomic64_add(delta, &blkg->delay_nsec);
2041 }
2042 
2043 /**
2044  * blkg_tryget_closest - try and get a blkg ref on the closet blkg
2045  * @bio: target bio
2046  * @css: target css
2047  *
2048  * As the failure mode here is to walk up the blkg tree, this ensure that the
2049  * blkg->parent pointers are always valid.  This returns the blkg that it ended
2050  * up taking a reference on or %NULL if no reference was taken.
2051  */
blkg_tryget_closest(struct bio * bio,struct cgroup_subsys_state * css)2052 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
2053 		struct cgroup_subsys_state *css)
2054 {
2055 	struct blkcg_gq *blkg, *ret_blkg = NULL;
2056 
2057 	rcu_read_lock();
2058 	blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
2059 	while (blkg) {
2060 		if (blkg_tryget(blkg)) {
2061 			ret_blkg = blkg;
2062 			break;
2063 		}
2064 		blkg = blkg->parent;
2065 	}
2066 	rcu_read_unlock();
2067 
2068 	return ret_blkg;
2069 }
2070 
2071 /**
2072  * bio_associate_blkg_from_css - associate a bio with a specified css
2073  * @bio: target bio
2074  * @css: target css
2075  *
2076  * Associate @bio with the blkg found by combining the css's blkg and the
2077  * request_queue of the @bio.  An association failure is handled by walking up
2078  * the blkg tree.  Therefore, the blkg associated can be anything between @blkg
2079  * and q->root_blkg.  This situation only happens when a cgroup is dying and
2080  * then the remaining bios will spill to the closest alive blkg.
2081  *
2082  * A reference will be taken on the blkg and will be released when @bio is
2083  * freed.
2084  */
bio_associate_blkg_from_css(struct bio * bio,struct cgroup_subsys_state * css)2085 void bio_associate_blkg_from_css(struct bio *bio,
2086 				 struct cgroup_subsys_state *css)
2087 {
2088 	if (bio->bi_blkg)
2089 		blkg_put(bio->bi_blkg);
2090 
2091 	if (css && css->parent) {
2092 		bio->bi_blkg = blkg_tryget_closest(bio, css);
2093 	} else {
2094 		blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
2095 		bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
2096 	}
2097 }
2098 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2099 
2100 /**
2101  * bio_associate_blkg - associate a bio with a blkg
2102  * @bio: target bio
2103  *
2104  * Associate @bio with the blkg found from the bio's css and request_queue.
2105  * If one is not found, bio_lookup_blkg() creates the blkg.  If a blkg is
2106  * already associated, the css is reused and association redone as the
2107  * request_queue may have changed.
2108  */
bio_associate_blkg(struct bio * bio)2109 void bio_associate_blkg(struct bio *bio)
2110 {
2111 	struct cgroup_subsys_state *css;
2112 
2113 	rcu_read_lock();
2114 
2115 	if (bio->bi_blkg)
2116 		css = bio_blkcg_css(bio);
2117 	else
2118 		css = blkcg_css();
2119 
2120 	bio_associate_blkg_from_css(bio, css);
2121 
2122 	rcu_read_unlock();
2123 }
2124 EXPORT_SYMBOL_GPL(bio_associate_blkg);
2125 
2126 /**
2127  * bio_clone_blkg_association - clone blkg association from src to dst bio
2128  * @dst: destination bio
2129  * @src: source bio
2130  */
bio_clone_blkg_association(struct bio * dst,struct bio * src)2131 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2132 {
2133 	if (src->bi_blkg)
2134 		bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2135 }
2136 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2137 
blk_cgroup_io_type(struct bio * bio)2138 static int blk_cgroup_io_type(struct bio *bio)
2139 {
2140 	if (op_is_discard(bio->bi_opf))
2141 		return BLKG_IOSTAT_DISCARD;
2142 	if (op_is_write(bio->bi_opf))
2143 		return BLKG_IOSTAT_WRITE;
2144 	return BLKG_IOSTAT_READ;
2145 }
2146 
blk_cgroup_bio_start(struct bio * bio)2147 void blk_cgroup_bio_start(struct bio *bio)
2148 {
2149 	struct blkcg *blkcg = bio->bi_blkg->blkcg;
2150 	int rwd = blk_cgroup_io_type(bio), cpu;
2151 	struct blkg_iostat_set *bis;
2152 	unsigned long flags;
2153 
2154 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
2155 		return;
2156 
2157 	/* Root-level stats are sourced from system-wide IO stats */
2158 	if (!cgroup_parent(blkcg->css.cgroup))
2159 		return;
2160 
2161 	cpu = get_cpu();
2162 	bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2163 	flags = u64_stats_update_begin_irqsave(&bis->sync);
2164 
2165 	/*
2166 	 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2167 	 * bio and we would have already accounted for the size of the bio.
2168 	 */
2169 	if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2170 		bio_set_flag(bio, BIO_CGROUP_ACCT);
2171 		bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2172 	}
2173 	bis->cur.ios[rwd]++;
2174 
2175 	/*
2176 	 * If the iostat_cpu isn't in a lockless list, put it into the
2177 	 * list to indicate that a stat update is pending.
2178 	 */
2179 	if (!READ_ONCE(bis->lqueued)) {
2180 		struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2181 
2182 		llist_add(&bis->lnode, lhead);
2183 		WRITE_ONCE(bis->lqueued, true);
2184 	}
2185 
2186 	u64_stats_update_end_irqrestore(&bis->sync, flags);
2187 	cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2188 	put_cpu();
2189 }
2190 
blk_cgroup_congested(void)2191 bool blk_cgroup_congested(void)
2192 {
2193 	struct cgroup_subsys_state *css;
2194 	bool ret = false;
2195 
2196 	rcu_read_lock();
2197 	for (css = blkcg_css(); css; css = css->parent) {
2198 		if (atomic_read(&css->cgroup->congestion_count)) {
2199 			ret = true;
2200 			break;
2201 		}
2202 	}
2203 	rcu_read_unlock();
2204 	return ret;
2205 }
2206 
2207 module_param(blkcg_debug_stats, bool, 0644);
2208 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");
2209