xref: /linux/block/blk-cgroup.c (revision 26bb0d3f38a764b743a3ad5c8b6e5b5044d7ceb4)
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_init(&bios, &blkg->async_bios);
222 	spin_unlock(&blkg->async_bio_lock);
223 
224 	/* start plug only when bio_list contains at least 2 bios */
225 	if (bios.head && bios.head->bi_next) {
226 		need_plug = true;
227 		blk_start_plug(&plug);
228 	}
229 	while ((bio = bio_list_pop(&bios)))
230 		submit_bio(bio);
231 	if (need_plug)
232 		blk_finish_plug(&plug);
233 }
234 
235 /*
236  * When a shared kthread issues a bio for a cgroup, doing so synchronously can
237  * lead to priority inversions as the kthread can be trapped waiting for that
238  * cgroup.  Use this helper instead of submit_bio to punt the actual issuing to
239  * a dedicated per-blkcg work item to avoid such priority inversions.
240  */
blkcg_punt_bio_submit(struct bio * bio)241 void blkcg_punt_bio_submit(struct bio *bio)
242 {
243 	struct blkcg_gq *blkg = bio->bi_blkg;
244 
245 	if (blkg->parent) {
246 		spin_lock(&blkg->async_bio_lock);
247 		bio_list_add(&blkg->async_bios, bio);
248 		spin_unlock(&blkg->async_bio_lock);
249 		queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
250 	} else {
251 		/* never bounce for the root cgroup */
252 		submit_bio(bio);
253 	}
254 }
255 EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit);
256 
blkcg_punt_bio_init(void)257 static int __init blkcg_punt_bio_init(void)
258 {
259 	blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
260 					    WQ_MEM_RECLAIM | WQ_FREEZABLE |
261 					    WQ_UNBOUND | WQ_SYSFS, 0);
262 	if (!blkcg_punt_bio_wq)
263 		return -ENOMEM;
264 	return 0;
265 }
266 subsys_initcall(blkcg_punt_bio_init);
267 #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
268 
269 /**
270  * bio_blkcg_css - return the blkcg CSS associated with a bio
271  * @bio: target bio
272  *
273  * This returns the CSS for the blkcg associated with a bio, or %NULL if not
274  * associated. Callers are expected to either handle %NULL or know association
275  * has been done prior to calling this.
276  */
bio_blkcg_css(struct bio * bio)277 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
278 {
279 	if (!bio || !bio->bi_blkg)
280 		return NULL;
281 	return &bio->bi_blkg->blkcg->css;
282 }
283 EXPORT_SYMBOL_GPL(bio_blkcg_css);
284 
285 /**
286  * blkcg_parent - get the parent of a blkcg
287  * @blkcg: blkcg of interest
288  *
289  * Return the parent blkcg of @blkcg.  Can be called anytime.
290  */
blkcg_parent(struct blkcg * blkcg)291 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
292 {
293 	return css_to_blkcg(blkcg->css.parent);
294 }
295 
296 /**
297  * blkg_alloc - allocate a blkg
298  * @blkcg: block cgroup the new blkg is associated with
299  * @disk: gendisk the new blkg is associated with
300  * @gfp_mask: allocation mask to use
301  *
302  * Allocate a new blkg associating @blkcg and @disk.
303  */
blkg_alloc(struct blkcg * blkcg,struct gendisk * disk,gfp_t gfp_mask)304 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
305 				   gfp_t gfp_mask)
306 {
307 	struct blkcg_gq *blkg;
308 	int i, cpu;
309 
310 	/* alloc and init base part */
311 	blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
312 	if (!blkg)
313 		return NULL;
314 	if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
315 		goto out_free_blkg;
316 	blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
317 	if (!blkg->iostat_cpu)
318 		goto out_exit_refcnt;
319 	if (!blk_get_queue(disk->queue))
320 		goto out_free_iostat;
321 
322 	blkg->q = disk->queue;
323 	INIT_LIST_HEAD(&blkg->q_node);
324 	blkg->blkcg = blkcg;
325 	blkg->iostat.blkg = blkg;
326 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
327 	spin_lock_init(&blkg->async_bio_lock);
328 	bio_list_init(&blkg->async_bios);
329 	INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
330 #endif
331 
332 	u64_stats_init(&blkg->iostat.sync);
333 	for_each_possible_cpu(cpu) {
334 		u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
335 		per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
336 	}
337 
338 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
339 		struct blkcg_policy *pol = blkcg_policy[i];
340 		struct blkg_policy_data *pd;
341 
342 		if (!blkcg_policy_enabled(disk->queue, pol))
343 			continue;
344 
345 		/* alloc per-policy data and attach it to blkg */
346 		pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
347 		if (!pd)
348 			goto out_free_pds;
349 		blkg->pd[i] = pd;
350 		pd->blkg = blkg;
351 		pd->plid = i;
352 		pd->online = false;
353 	}
354 
355 	return blkg;
356 
357 out_free_pds:
358 	while (--i >= 0)
359 		if (blkg->pd[i])
360 			blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
361 	blk_put_queue(disk->queue);
362 out_free_iostat:
363 	free_percpu(blkg->iostat_cpu);
364 out_exit_refcnt:
365 	percpu_ref_exit(&blkg->refcnt);
366 out_free_blkg:
367 	kfree(blkg);
368 	return NULL;
369 }
370 
371 /*
372  * If @new_blkg is %NULL, this function tries to allocate a new one as
373  * necessary using %GFP_NOWAIT.  @new_blkg is always consumed on return.
374  */
blkg_create(struct blkcg * blkcg,struct gendisk * disk,struct blkcg_gq * new_blkg)375 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
376 				    struct blkcg_gq *new_blkg)
377 {
378 	struct blkcg_gq *blkg;
379 	int i, ret;
380 
381 	lockdep_assert_held(&disk->queue->queue_lock);
382 
383 	/* request_queue is dying, do not create/recreate a blkg */
384 	if (blk_queue_dying(disk->queue)) {
385 		ret = -ENODEV;
386 		goto err_free_blkg;
387 	}
388 
389 	/* blkg holds a reference to blkcg */
390 	if (!css_tryget_online(&blkcg->css)) {
391 		ret = -ENODEV;
392 		goto err_free_blkg;
393 	}
394 
395 	/* allocate */
396 	if (!new_blkg) {
397 		new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
398 		if (unlikely(!new_blkg)) {
399 			ret = -ENOMEM;
400 			goto err_put_css;
401 		}
402 	}
403 	blkg = new_blkg;
404 
405 	/* link parent */
406 	if (blkcg_parent(blkcg)) {
407 		blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
408 		if (WARN_ON_ONCE(!blkg->parent)) {
409 			ret = -ENODEV;
410 			goto err_put_css;
411 		}
412 		blkg_get(blkg->parent);
413 	}
414 
415 	/* invoke per-policy init */
416 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
417 		struct blkcg_policy *pol = blkcg_policy[i];
418 
419 		if (blkg->pd[i] && pol->pd_init_fn)
420 			pol->pd_init_fn(blkg->pd[i]);
421 	}
422 
423 	/* insert */
424 	spin_lock(&blkcg->lock);
425 	ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
426 	if (likely(!ret)) {
427 		hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
428 		list_add(&blkg->q_node, &disk->queue->blkg_list);
429 
430 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
431 			struct blkcg_policy *pol = blkcg_policy[i];
432 
433 			if (blkg->pd[i]) {
434 				if (pol->pd_online_fn)
435 					pol->pd_online_fn(blkg->pd[i]);
436 				blkg->pd[i]->online = true;
437 			}
438 		}
439 	}
440 	blkg->online = true;
441 	spin_unlock(&blkcg->lock);
442 
443 	if (!ret)
444 		return blkg;
445 
446 	/* @blkg failed fully initialized, use the usual release path */
447 	blkg_put(blkg);
448 	return ERR_PTR(ret);
449 
450 err_put_css:
451 	css_put(&blkcg->css);
452 err_free_blkg:
453 	if (new_blkg)
454 		blkg_free(new_blkg);
455 	return ERR_PTR(ret);
456 }
457 
458 /**
459  * blkg_lookup_create - lookup blkg, try to create one if not there
460  * @blkcg: blkcg of interest
461  * @disk: gendisk of interest
462  *
463  * Lookup blkg for the @blkcg - @disk pair.  If it doesn't exist, try to
464  * create one.  blkg creation is performed recursively from blkcg_root such
465  * that all non-root blkg's have access to the parent blkg.  This function
466  * should be called under RCU read lock and takes @disk->queue->queue_lock.
467  *
468  * Returns the blkg or the closest blkg if blkg_create() fails as it walks
469  * down from root.
470  */
blkg_lookup_create(struct blkcg * blkcg,struct gendisk * disk)471 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
472 		struct gendisk *disk)
473 {
474 	struct request_queue *q = disk->queue;
475 	struct blkcg_gq *blkg;
476 	unsigned long flags;
477 
478 	WARN_ON_ONCE(!rcu_read_lock_held());
479 
480 	blkg = blkg_lookup(blkcg, q);
481 	if (blkg)
482 		return blkg;
483 
484 	spin_lock_irqsave(&q->queue_lock, flags);
485 	blkg = blkg_lookup(blkcg, q);
486 	if (blkg) {
487 		if (blkcg != &blkcg_root &&
488 		    blkg != rcu_dereference(blkcg->blkg_hint))
489 			rcu_assign_pointer(blkcg->blkg_hint, blkg);
490 		goto found;
491 	}
492 
493 	/*
494 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
495 	 * non-root blkgs have access to their parents.  Returns the closest
496 	 * blkg to the intended blkg should blkg_create() fail.
497 	 */
498 	while (true) {
499 		struct blkcg *pos = blkcg;
500 		struct blkcg *parent = blkcg_parent(blkcg);
501 		struct blkcg_gq *ret_blkg = q->root_blkg;
502 
503 		while (parent) {
504 			blkg = blkg_lookup(parent, q);
505 			if (blkg) {
506 				/* remember closest blkg */
507 				ret_blkg = blkg;
508 				break;
509 			}
510 			pos = parent;
511 			parent = blkcg_parent(parent);
512 		}
513 
514 		blkg = blkg_create(pos, disk, NULL);
515 		if (IS_ERR(blkg)) {
516 			blkg = ret_blkg;
517 			break;
518 		}
519 		if (pos == blkcg)
520 			break;
521 	}
522 
523 found:
524 	spin_unlock_irqrestore(&q->queue_lock, flags);
525 	return blkg;
526 }
527 
blkg_destroy(struct blkcg_gq * blkg)528 static void blkg_destroy(struct blkcg_gq *blkg)
529 {
530 	struct blkcg *blkcg = blkg->blkcg;
531 	int i;
532 
533 	lockdep_assert_held(&blkg->q->queue_lock);
534 	lockdep_assert_held(&blkcg->lock);
535 
536 	/*
537 	 * blkg stays on the queue list until blkg_free_workfn(), see details in
538 	 * blkg_free_workfn(), hence this function can be called from
539 	 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
540 	 * blkg_free_workfn().
541 	 */
542 	if (hlist_unhashed(&blkg->blkcg_node))
543 		return;
544 
545 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
546 		struct blkcg_policy *pol = blkcg_policy[i];
547 
548 		if (blkg->pd[i] && blkg->pd[i]->online) {
549 			blkg->pd[i]->online = false;
550 			if (pol->pd_offline_fn)
551 				pol->pd_offline_fn(blkg->pd[i]);
552 		}
553 	}
554 
555 	blkg->online = false;
556 
557 	radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
558 	hlist_del_init_rcu(&blkg->blkcg_node);
559 
560 	/*
561 	 * Both setting lookup hint to and clearing it from @blkg are done
562 	 * under queue_lock.  If it's not pointing to @blkg now, it never
563 	 * will.  Hint assignment itself can race safely.
564 	 */
565 	if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
566 		rcu_assign_pointer(blkcg->blkg_hint, NULL);
567 
568 	/*
569 	 * Put the reference taken at the time of creation so that when all
570 	 * queues are gone, group can be destroyed.
571 	 */
572 	percpu_ref_kill(&blkg->refcnt);
573 }
574 
blkg_destroy_all(struct gendisk * disk)575 static void blkg_destroy_all(struct gendisk *disk)
576 {
577 	struct request_queue *q = disk->queue;
578 	struct blkcg_gq *blkg;
579 	int count = BLKG_DESTROY_BATCH_SIZE;
580 	int i;
581 
582 restart:
583 	spin_lock_irq(&q->queue_lock);
584 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
585 		struct blkcg *blkcg = blkg->blkcg;
586 
587 		if (hlist_unhashed(&blkg->blkcg_node))
588 			continue;
589 
590 		spin_lock(&blkcg->lock);
591 		blkg_destroy(blkg);
592 		spin_unlock(&blkcg->lock);
593 
594 		/*
595 		 * in order to avoid holding the spin lock for too long, release
596 		 * it when a batch of blkgs are destroyed.
597 		 */
598 		if (!(--count)) {
599 			count = BLKG_DESTROY_BATCH_SIZE;
600 			spin_unlock_irq(&q->queue_lock);
601 			cond_resched();
602 			goto restart;
603 		}
604 	}
605 
606 	/*
607 	 * Mark policy deactivated since policy offline has been done, and
608 	 * the free is scheduled, so future blkcg_deactivate_policy() can
609 	 * be bypassed
610 	 */
611 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
612 		struct blkcg_policy *pol = blkcg_policy[i];
613 
614 		if (pol)
615 			__clear_bit(pol->plid, q->blkcg_pols);
616 	}
617 
618 	q->root_blkg = NULL;
619 	spin_unlock_irq(&q->queue_lock);
620 }
621 
blkg_iostat_set(struct blkg_iostat * dst,struct blkg_iostat * src)622 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
623 {
624 	int i;
625 
626 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
627 		dst->bytes[i] = src->bytes[i];
628 		dst->ios[i] = src->ios[i];
629 	}
630 }
631 
__blkg_clear_stat(struct blkg_iostat_set * bis)632 static void __blkg_clear_stat(struct blkg_iostat_set *bis)
633 {
634 	struct blkg_iostat cur = {0};
635 	unsigned long flags;
636 
637 	flags = u64_stats_update_begin_irqsave(&bis->sync);
638 	blkg_iostat_set(&bis->cur, &cur);
639 	blkg_iostat_set(&bis->last, &cur);
640 	u64_stats_update_end_irqrestore(&bis->sync, flags);
641 }
642 
blkg_clear_stat(struct blkcg_gq * blkg)643 static void blkg_clear_stat(struct blkcg_gq *blkg)
644 {
645 	int cpu;
646 
647 	for_each_possible_cpu(cpu) {
648 		struct blkg_iostat_set *s = per_cpu_ptr(blkg->iostat_cpu, cpu);
649 
650 		__blkg_clear_stat(s);
651 	}
652 	__blkg_clear_stat(&blkg->iostat);
653 }
654 
blkcg_reset_stats(struct cgroup_subsys_state * css,struct cftype * cftype,u64 val)655 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
656 			     struct cftype *cftype, u64 val)
657 {
658 	struct blkcg *blkcg = css_to_blkcg(css);
659 	struct blkcg_gq *blkg;
660 	int i;
661 
662 	mutex_lock(&blkcg_pol_mutex);
663 	spin_lock_irq(&blkcg->lock);
664 
665 	/*
666 	 * Note that stat reset is racy - it doesn't synchronize against
667 	 * stat updates.  This is a debug feature which shouldn't exist
668 	 * anyway.  If you get hit by a race, retry.
669 	 */
670 	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
671 		blkg_clear_stat(blkg);
672 		for (i = 0; i < BLKCG_MAX_POLS; i++) {
673 			struct blkcg_policy *pol = blkcg_policy[i];
674 
675 			if (blkg->pd[i] && pol->pd_reset_stats_fn)
676 				pol->pd_reset_stats_fn(blkg->pd[i]);
677 		}
678 	}
679 
680 	spin_unlock_irq(&blkcg->lock);
681 	mutex_unlock(&blkcg_pol_mutex);
682 	return 0;
683 }
684 
blkg_dev_name(struct blkcg_gq * blkg)685 const char *blkg_dev_name(struct blkcg_gq *blkg)
686 {
687 	if (!blkg->q->disk)
688 		return NULL;
689 	return bdi_dev_name(blkg->q->disk->bdi);
690 }
691 
692 /**
693  * blkcg_print_blkgs - helper for printing per-blkg data
694  * @sf: seq_file to print to
695  * @blkcg: blkcg of interest
696  * @prfill: fill function to print out a blkg
697  * @pol: policy in question
698  * @data: data to be passed to @prfill
699  * @show_total: to print out sum of prfill return values or not
700  *
701  * This function invokes @prfill on each blkg of @blkcg if pd for the
702  * policy specified by @pol exists.  @prfill is invoked with @sf, the
703  * policy data and @data and the matching queue lock held.  If @show_total
704  * is %true, the sum of the return values from @prfill is printed with
705  * "Total" label at the end.
706  *
707  * This is to be used to construct print functions for
708  * cftype->read_seq_string method.
709  */
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)710 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
711 		       u64 (*prfill)(struct seq_file *,
712 				     struct blkg_policy_data *, int),
713 		       const struct blkcg_policy *pol, int data,
714 		       bool show_total)
715 {
716 	struct blkcg_gq *blkg;
717 	u64 total = 0;
718 
719 	rcu_read_lock();
720 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
721 		spin_lock_irq(&blkg->q->queue_lock);
722 		if (blkcg_policy_enabled(blkg->q, pol))
723 			total += prfill(sf, blkg->pd[pol->plid], data);
724 		spin_unlock_irq(&blkg->q->queue_lock);
725 	}
726 	rcu_read_unlock();
727 
728 	if (show_total)
729 		seq_printf(sf, "Total %llu\n", (unsigned long long)total);
730 }
731 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
732 
733 /**
734  * __blkg_prfill_u64 - prfill helper for a single u64 value
735  * @sf: seq_file to print to
736  * @pd: policy private data of interest
737  * @v: value to print
738  *
739  * Print @v to @sf for the device associated with @pd.
740  */
__blkg_prfill_u64(struct seq_file * sf,struct blkg_policy_data * pd,u64 v)741 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
742 {
743 	const char *dname = blkg_dev_name(pd->blkg);
744 
745 	if (!dname)
746 		return 0;
747 
748 	seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
749 	return v;
750 }
751 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
752 
753 /**
754  * blkg_conf_init - initialize a blkg_conf_ctx
755  * @ctx: blkg_conf_ctx to initialize
756  * @input: input string
757  *
758  * Initialize @ctx which can be used to parse blkg config input string @input.
759  * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
760  * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
761  */
blkg_conf_init(struct blkg_conf_ctx * ctx,char * input)762 void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
763 {
764 	*ctx = (struct blkg_conf_ctx){ .input = input };
765 }
766 EXPORT_SYMBOL_GPL(blkg_conf_init);
767 
768 /**
769  * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
770  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
771  *
772  * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
773  * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
774  * set to point past the device node prefix.
775  *
776  * This function may be called multiple times on @ctx and the extra calls become
777  * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
778  * explicitly if bdev access is needed without resolving the blkcg / policy part
779  * of @ctx->input. Returns -errno on error.
780  */
blkg_conf_open_bdev(struct blkg_conf_ctx * ctx)781 int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
782 {
783 	char *input = ctx->input;
784 	unsigned int major, minor;
785 	struct block_device *bdev;
786 	int key_len;
787 
788 	if (ctx->bdev)
789 		return 0;
790 
791 	if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
792 		return -EINVAL;
793 
794 	input += key_len;
795 	if (!isspace(*input))
796 		return -EINVAL;
797 	input = skip_spaces(input);
798 
799 	bdev = blkdev_get_no_open(MKDEV(major, minor));
800 	if (!bdev)
801 		return -ENODEV;
802 	if (bdev_is_partition(bdev)) {
803 		blkdev_put_no_open(bdev);
804 		return -ENODEV;
805 	}
806 
807 	mutex_lock(&bdev->bd_queue->rq_qos_mutex);
808 	if (!disk_live(bdev->bd_disk)) {
809 		blkdev_put_no_open(bdev);
810 		mutex_unlock(&bdev->bd_queue->rq_qos_mutex);
811 		return -ENODEV;
812 	}
813 
814 	ctx->body = input;
815 	ctx->bdev = bdev;
816 	return 0;
817 }
818 
819 /**
820  * blkg_conf_prep - parse and prepare for per-blkg config update
821  * @blkcg: target block cgroup
822  * @pol: target policy
823  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
824  *
825  * Parse per-blkg config update from @ctx->input and initialize @ctx
826  * accordingly. On success, @ctx->body points to the part of @ctx->input
827  * following MAJ:MIN, @ctx->bdev points to the target block device and
828  * @ctx->blkg to the blkg being configured.
829  *
830  * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
831  * function returns with queue lock held and must be followed by
832  * blkg_conf_exit().
833  */
blkg_conf_prep(struct blkcg * blkcg,const struct blkcg_policy * pol,struct blkg_conf_ctx * ctx)834 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
835 		   struct blkg_conf_ctx *ctx)
836 	__acquires(&bdev->bd_queue->queue_lock)
837 {
838 	struct gendisk *disk;
839 	struct request_queue *q;
840 	struct blkcg_gq *blkg;
841 	int ret;
842 
843 	ret = blkg_conf_open_bdev(ctx);
844 	if (ret)
845 		return ret;
846 
847 	disk = ctx->bdev->bd_disk;
848 	q = disk->queue;
849 
850 	/*
851 	 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
852 	 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
853 	 */
854 	ret = blk_queue_enter(q, 0);
855 	if (ret)
856 		goto fail;
857 
858 	spin_lock_irq(&q->queue_lock);
859 
860 	if (!blkcg_policy_enabled(q, pol)) {
861 		ret = -EOPNOTSUPP;
862 		goto fail_unlock;
863 	}
864 
865 	blkg = blkg_lookup(blkcg, q);
866 	if (blkg)
867 		goto success;
868 
869 	/*
870 	 * Create blkgs walking down from blkcg_root to @blkcg, so that all
871 	 * non-root blkgs have access to their parents.
872 	 */
873 	while (true) {
874 		struct blkcg *pos = blkcg;
875 		struct blkcg *parent;
876 		struct blkcg_gq *new_blkg;
877 
878 		parent = blkcg_parent(blkcg);
879 		while (parent && !blkg_lookup(parent, q)) {
880 			pos = parent;
881 			parent = blkcg_parent(parent);
882 		}
883 
884 		/* Drop locks to do new blkg allocation with GFP_KERNEL. */
885 		spin_unlock_irq(&q->queue_lock);
886 
887 		new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
888 		if (unlikely(!new_blkg)) {
889 			ret = -ENOMEM;
890 			goto fail_exit_queue;
891 		}
892 
893 		if (radix_tree_preload(GFP_KERNEL)) {
894 			blkg_free(new_blkg);
895 			ret = -ENOMEM;
896 			goto fail_exit_queue;
897 		}
898 
899 		spin_lock_irq(&q->queue_lock);
900 
901 		if (!blkcg_policy_enabled(q, pol)) {
902 			blkg_free(new_blkg);
903 			ret = -EOPNOTSUPP;
904 			goto fail_preloaded;
905 		}
906 
907 		blkg = blkg_lookup(pos, q);
908 		if (blkg) {
909 			blkg_free(new_blkg);
910 		} else {
911 			blkg = blkg_create(pos, disk, new_blkg);
912 			if (IS_ERR(blkg)) {
913 				ret = PTR_ERR(blkg);
914 				goto fail_preloaded;
915 			}
916 		}
917 
918 		radix_tree_preload_end();
919 
920 		if (pos == blkcg)
921 			goto success;
922 	}
923 success:
924 	blk_queue_exit(q);
925 	ctx->blkg = blkg;
926 	return 0;
927 
928 fail_preloaded:
929 	radix_tree_preload_end();
930 fail_unlock:
931 	spin_unlock_irq(&q->queue_lock);
932 fail_exit_queue:
933 	blk_queue_exit(q);
934 fail:
935 	/*
936 	 * If queue was bypassing, we should retry.  Do so after a
937 	 * short msleep().  It isn't strictly necessary but queue
938 	 * can be bypassing for some time and it's always nice to
939 	 * avoid busy looping.
940 	 */
941 	if (ret == -EBUSY) {
942 		msleep(10);
943 		ret = restart_syscall();
944 	}
945 	return ret;
946 }
947 EXPORT_SYMBOL_GPL(blkg_conf_prep);
948 
949 /**
950  * blkg_conf_exit - clean up per-blkg config update
951  * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
952  *
953  * Clean up after per-blkg config update. This function must be called on all
954  * blkg_conf_ctx's initialized with blkg_conf_init().
955  */
blkg_conf_exit(struct blkg_conf_ctx * ctx)956 void blkg_conf_exit(struct blkg_conf_ctx *ctx)
957 	__releases(&ctx->bdev->bd_queue->queue_lock)
958 	__releases(&ctx->bdev->bd_queue->rq_qos_mutex)
959 {
960 	if (ctx->blkg) {
961 		spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
962 		ctx->blkg = NULL;
963 	}
964 
965 	if (ctx->bdev) {
966 		mutex_unlock(&ctx->bdev->bd_queue->rq_qos_mutex);
967 		blkdev_put_no_open(ctx->bdev);
968 		ctx->body = NULL;
969 		ctx->bdev = NULL;
970 	}
971 }
972 EXPORT_SYMBOL_GPL(blkg_conf_exit);
973 
blkg_iostat_add(struct blkg_iostat * dst,struct blkg_iostat * src)974 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
975 {
976 	int i;
977 
978 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
979 		dst->bytes[i] += src->bytes[i];
980 		dst->ios[i] += src->ios[i];
981 	}
982 }
983 
blkg_iostat_sub(struct blkg_iostat * dst,struct blkg_iostat * src)984 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
985 {
986 	int i;
987 
988 	for (i = 0; i < BLKG_IOSTAT_NR; i++) {
989 		dst->bytes[i] -= src->bytes[i];
990 		dst->ios[i] -= src->ios[i];
991 	}
992 }
993 
blkcg_iostat_update(struct blkcg_gq * blkg,struct blkg_iostat * cur,struct blkg_iostat * last)994 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
995 				struct blkg_iostat *last)
996 {
997 	struct blkg_iostat delta;
998 	unsigned long flags;
999 
1000 	/* propagate percpu delta to global */
1001 	flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1002 	blkg_iostat_set(&delta, cur);
1003 	blkg_iostat_sub(&delta, last);
1004 	blkg_iostat_add(&blkg->iostat.cur, &delta);
1005 	blkg_iostat_add(last, &delta);
1006 	u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1007 }
1008 
__blkcg_rstat_flush(struct blkcg * blkcg,int cpu)1009 static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
1010 {
1011 	struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
1012 	struct llist_node *lnode;
1013 	struct blkg_iostat_set *bisc, *next_bisc;
1014 	unsigned long flags;
1015 
1016 	rcu_read_lock();
1017 
1018 	lnode = llist_del_all(lhead);
1019 	if (!lnode)
1020 		goto out;
1021 
1022 	/*
1023 	 * For covering concurrent parent blkg update from blkg_release().
1024 	 *
1025 	 * When flushing from cgroup, cgroup_rstat_lock is always held, so
1026 	 * this lock won't cause contention most of time.
1027 	 */
1028 	raw_spin_lock_irqsave(&blkg_stat_lock, flags);
1029 
1030 	/*
1031 	 * Iterate only the iostat_cpu's queued in the lockless list.
1032 	 */
1033 	llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
1034 		struct blkcg_gq *blkg = bisc->blkg;
1035 		struct blkcg_gq *parent = blkg->parent;
1036 		struct blkg_iostat cur;
1037 		unsigned int seq;
1038 
1039 		/*
1040 		 * Order assignment of `next_bisc` from `bisc->lnode.next` in
1041 		 * llist_for_each_entry_safe and clearing `bisc->lqueued` for
1042 		 * avoiding to assign `next_bisc` with new next pointer added
1043 		 * in blk_cgroup_bio_start() in case of re-ordering.
1044 		 *
1045 		 * The pair barrier is implied in llist_add() in blk_cgroup_bio_start().
1046 		 */
1047 		smp_mb();
1048 
1049 		WRITE_ONCE(bisc->lqueued, false);
1050 		if (bisc == &blkg->iostat)
1051 			goto propagate_up; /* propagate up to parent only */
1052 
1053 		/* fetch the current per-cpu values */
1054 		do {
1055 			seq = u64_stats_fetch_begin(&bisc->sync);
1056 			blkg_iostat_set(&cur, &bisc->cur);
1057 		} while (u64_stats_fetch_retry(&bisc->sync, seq));
1058 
1059 		blkcg_iostat_update(blkg, &cur, &bisc->last);
1060 
1061 propagate_up:
1062 		/* propagate global delta to parent (unless that's root) */
1063 		if (parent && parent->parent) {
1064 			blkcg_iostat_update(parent, &blkg->iostat.cur,
1065 					    &blkg->iostat.last);
1066 			/*
1067 			 * Queue parent->iostat to its blkcg's lockless
1068 			 * list to propagate up to the grandparent if the
1069 			 * iostat hasn't been queued yet.
1070 			 */
1071 			if (!parent->iostat.lqueued) {
1072 				struct llist_head *plhead;
1073 
1074 				plhead = per_cpu_ptr(parent->blkcg->lhead, cpu);
1075 				llist_add(&parent->iostat.lnode, plhead);
1076 				parent->iostat.lqueued = true;
1077 			}
1078 		}
1079 	}
1080 	raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
1081 out:
1082 	rcu_read_unlock();
1083 }
1084 
blkcg_rstat_flush(struct cgroup_subsys_state * css,int cpu)1085 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1086 {
1087 	/* Root-level stats are sourced from system-wide IO stats */
1088 	if (cgroup_parent(css->cgroup))
1089 		__blkcg_rstat_flush(css_to_blkcg(css), cpu);
1090 }
1091 
1092 /*
1093  * We source root cgroup stats from the system-wide stats to avoid
1094  * tracking the same information twice and incurring overhead when no
1095  * cgroups are defined. For that reason, cgroup_rstat_flush in
1096  * blkcg_print_stat does not actually fill out the iostat in the root
1097  * cgroup's blkcg_gq.
1098  *
1099  * However, we would like to re-use the printing code between the root and
1100  * non-root cgroups to the extent possible. For that reason, we simulate
1101  * flushing the root cgroup's stats by explicitly filling in the iostat
1102  * with disk level statistics.
1103  */
blkcg_fill_root_iostats(void)1104 static void blkcg_fill_root_iostats(void)
1105 {
1106 	struct class_dev_iter iter;
1107 	struct device *dev;
1108 
1109 	class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
1110 	while ((dev = class_dev_iter_next(&iter))) {
1111 		struct block_device *bdev = dev_to_bdev(dev);
1112 		struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
1113 		struct blkg_iostat tmp;
1114 		int cpu;
1115 		unsigned long flags;
1116 
1117 		memset(&tmp, 0, sizeof(tmp));
1118 		for_each_possible_cpu(cpu) {
1119 			struct disk_stats *cpu_dkstats;
1120 
1121 			cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1122 			tmp.ios[BLKG_IOSTAT_READ] +=
1123 				cpu_dkstats->ios[STAT_READ];
1124 			tmp.ios[BLKG_IOSTAT_WRITE] +=
1125 				cpu_dkstats->ios[STAT_WRITE];
1126 			tmp.ios[BLKG_IOSTAT_DISCARD] +=
1127 				cpu_dkstats->ios[STAT_DISCARD];
1128 			// convert sectors to bytes
1129 			tmp.bytes[BLKG_IOSTAT_READ] +=
1130 				cpu_dkstats->sectors[STAT_READ] << 9;
1131 			tmp.bytes[BLKG_IOSTAT_WRITE] +=
1132 				cpu_dkstats->sectors[STAT_WRITE] << 9;
1133 			tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1134 				cpu_dkstats->sectors[STAT_DISCARD] << 9;
1135 		}
1136 
1137 		flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1138 		blkg_iostat_set(&blkg->iostat.cur, &tmp);
1139 		u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1140 	}
1141 }
1142 
blkcg_print_one_stat(struct blkcg_gq * blkg,struct seq_file * s)1143 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1144 {
1145 	struct blkg_iostat_set *bis = &blkg->iostat;
1146 	u64 rbytes, wbytes, rios, wios, dbytes, dios;
1147 	const char *dname;
1148 	unsigned seq;
1149 	int i;
1150 
1151 	if (!blkg->online)
1152 		return;
1153 
1154 	dname = blkg_dev_name(blkg);
1155 	if (!dname)
1156 		return;
1157 
1158 	seq_printf(s, "%s ", dname);
1159 
1160 	do {
1161 		seq = u64_stats_fetch_begin(&bis->sync);
1162 
1163 		rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1164 		wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1165 		dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1166 		rios = bis->cur.ios[BLKG_IOSTAT_READ];
1167 		wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1168 		dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1169 	} while (u64_stats_fetch_retry(&bis->sync, seq));
1170 
1171 	if (rbytes || wbytes || rios || wios) {
1172 		seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1173 			rbytes, wbytes, rios, wios,
1174 			dbytes, dios);
1175 	}
1176 
1177 	if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1178 		seq_printf(s, " use_delay=%d delay_nsec=%llu",
1179 			atomic_read(&blkg->use_delay),
1180 			atomic64_read(&blkg->delay_nsec));
1181 	}
1182 
1183 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
1184 		struct blkcg_policy *pol = blkcg_policy[i];
1185 
1186 		if (!blkg->pd[i] || !pol->pd_stat_fn)
1187 			continue;
1188 
1189 		pol->pd_stat_fn(blkg->pd[i], s);
1190 	}
1191 
1192 	seq_puts(s, "\n");
1193 }
1194 
blkcg_print_stat(struct seq_file * sf,void * v)1195 static int blkcg_print_stat(struct seq_file *sf, void *v)
1196 {
1197 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1198 	struct blkcg_gq *blkg;
1199 
1200 	if (!seq_css(sf)->parent)
1201 		blkcg_fill_root_iostats();
1202 	else
1203 		cgroup_rstat_flush(blkcg->css.cgroup);
1204 
1205 	rcu_read_lock();
1206 	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1207 		spin_lock_irq(&blkg->q->queue_lock);
1208 		blkcg_print_one_stat(blkg, sf);
1209 		spin_unlock_irq(&blkg->q->queue_lock);
1210 	}
1211 	rcu_read_unlock();
1212 	return 0;
1213 }
1214 
1215 static struct cftype blkcg_files[] = {
1216 	{
1217 		.name = "stat",
1218 		.seq_show = blkcg_print_stat,
1219 	},
1220 	{ }	/* terminate */
1221 };
1222 
1223 static struct cftype blkcg_legacy_files[] = {
1224 	{
1225 		.name = "reset_stats",
1226 		.write_u64 = blkcg_reset_stats,
1227 	},
1228 	{ }	/* terminate */
1229 };
1230 
1231 #ifdef CONFIG_CGROUP_WRITEBACK
blkcg_get_cgwb_list(struct cgroup_subsys_state * css)1232 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1233 {
1234 	return &css_to_blkcg(css)->cgwb_list;
1235 }
1236 #endif
1237 
1238 /*
1239  * blkcg destruction is a three-stage process.
1240  *
1241  * 1. Destruction starts.  The blkcg_css_offline() callback is invoked
1242  *    which offlines writeback.  Here we tie the next stage of blkg destruction
1243  *    to the completion of writeback associated with the blkcg.  This lets us
1244  *    avoid punting potentially large amounts of outstanding writeback to root
1245  *    while maintaining any ongoing policies.  The next stage is triggered when
1246  *    the nr_cgwbs count goes to zero.
1247  *
1248  * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1249  *    and handles the destruction of blkgs.  Here the css reference held by
1250  *    the blkg is put back eventually allowing blkcg_css_free() to be called.
1251  *    This work may occur in cgwb_release_workfn() on the cgwb_release
1252  *    workqueue.  Any submitted ios that fail to get the blkg ref will be
1253  *    punted to the root_blkg.
1254  *
1255  * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1256  *    This finally frees the blkcg.
1257  */
1258 
1259 /**
1260  * blkcg_destroy_blkgs - responsible for shooting down blkgs
1261  * @blkcg: blkcg of interest
1262  *
1263  * blkgs should be removed while holding both q and blkcg locks.  As blkcg lock
1264  * is nested inside q lock, this function performs reverse double lock dancing.
1265  * Destroying the blkgs releases the reference held on the blkcg's css allowing
1266  * blkcg_css_free to eventually be called.
1267  *
1268  * This is the blkcg counterpart of ioc_release_fn().
1269  */
blkcg_destroy_blkgs(struct blkcg * blkcg)1270 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1271 {
1272 	might_sleep();
1273 
1274 	spin_lock_irq(&blkcg->lock);
1275 
1276 	while (!hlist_empty(&blkcg->blkg_list)) {
1277 		struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1278 						struct blkcg_gq, blkcg_node);
1279 		struct request_queue *q = blkg->q;
1280 
1281 		if (need_resched() || !spin_trylock(&q->queue_lock)) {
1282 			/*
1283 			 * Given that the system can accumulate a huge number
1284 			 * of blkgs in pathological cases, check to see if we
1285 			 * need to rescheduling to avoid softlockup.
1286 			 */
1287 			spin_unlock_irq(&blkcg->lock);
1288 			cond_resched();
1289 			spin_lock_irq(&blkcg->lock);
1290 			continue;
1291 		}
1292 
1293 		blkg_destroy(blkg);
1294 		spin_unlock(&q->queue_lock);
1295 	}
1296 
1297 	spin_unlock_irq(&blkcg->lock);
1298 }
1299 
1300 /**
1301  * blkcg_pin_online - pin online state
1302  * @blkcg_css: blkcg of interest
1303  *
1304  * While pinned, a blkcg is kept online.  This is primarily used to
1305  * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1306  * while an associated cgwb is still active.
1307  */
blkcg_pin_online(struct cgroup_subsys_state * blkcg_css)1308 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1309 {
1310 	refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1311 }
1312 
1313 /**
1314  * blkcg_unpin_online - unpin online state
1315  * @blkcg_css: blkcg of interest
1316  *
1317  * This is primarily used to impedance-match blkg and cgwb lifetimes so
1318  * that blkg doesn't go offline while an associated cgwb is still active.
1319  * When this count goes to zero, all active cgwbs have finished so the
1320  * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1321  */
blkcg_unpin_online(struct cgroup_subsys_state * blkcg_css)1322 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1323 {
1324 	struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1325 
1326 	do {
1327 		if (!refcount_dec_and_test(&blkcg->online_pin))
1328 			break;
1329 		blkcg_destroy_blkgs(blkcg);
1330 		blkcg = blkcg_parent(blkcg);
1331 	} while (blkcg);
1332 }
1333 
1334 /**
1335  * blkcg_css_offline - cgroup css_offline callback
1336  * @css: css of interest
1337  *
1338  * This function is called when @css is about to go away.  Here the cgwbs are
1339  * offlined first and only once writeback associated with the blkcg has
1340  * finished do we start step 2 (see above).
1341  */
blkcg_css_offline(struct cgroup_subsys_state * css)1342 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1343 {
1344 	/* this prevents anyone from attaching or migrating to this blkcg */
1345 	wb_blkcg_offline(css);
1346 
1347 	/* put the base online pin allowing step 2 to be triggered */
1348 	blkcg_unpin_online(css);
1349 }
1350 
blkcg_css_free(struct cgroup_subsys_state * css)1351 static void blkcg_css_free(struct cgroup_subsys_state *css)
1352 {
1353 	struct blkcg *blkcg = css_to_blkcg(css);
1354 	int i;
1355 
1356 	mutex_lock(&blkcg_pol_mutex);
1357 
1358 	list_del(&blkcg->all_blkcgs_node);
1359 
1360 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1361 		if (blkcg->cpd[i])
1362 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1363 
1364 	mutex_unlock(&blkcg_pol_mutex);
1365 
1366 	free_percpu(blkcg->lhead);
1367 	kfree(blkcg);
1368 }
1369 
1370 static struct cgroup_subsys_state *
blkcg_css_alloc(struct cgroup_subsys_state * parent_css)1371 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1372 {
1373 	struct blkcg *blkcg;
1374 	int i;
1375 
1376 	mutex_lock(&blkcg_pol_mutex);
1377 
1378 	if (!parent_css) {
1379 		blkcg = &blkcg_root;
1380 	} else {
1381 		blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1382 		if (!blkcg)
1383 			goto unlock;
1384 	}
1385 
1386 	if (init_blkcg_llists(blkcg))
1387 		goto free_blkcg;
1388 
1389 	for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1390 		struct blkcg_policy *pol = blkcg_policy[i];
1391 		struct blkcg_policy_data *cpd;
1392 
1393 		/*
1394 		 * If the policy hasn't been attached yet, wait for it
1395 		 * to be attached before doing anything else. Otherwise,
1396 		 * check if the policy requires any specific per-cgroup
1397 		 * data: if it does, allocate and initialize it.
1398 		 */
1399 		if (!pol || !pol->cpd_alloc_fn)
1400 			continue;
1401 
1402 		cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1403 		if (!cpd)
1404 			goto free_pd_blkcg;
1405 
1406 		blkcg->cpd[i] = cpd;
1407 		cpd->blkcg = blkcg;
1408 		cpd->plid = i;
1409 	}
1410 
1411 	spin_lock_init(&blkcg->lock);
1412 	refcount_set(&blkcg->online_pin, 1);
1413 	INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1414 	INIT_HLIST_HEAD(&blkcg->blkg_list);
1415 #ifdef CONFIG_CGROUP_WRITEBACK
1416 	INIT_LIST_HEAD(&blkcg->cgwb_list);
1417 #endif
1418 	list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1419 
1420 	mutex_unlock(&blkcg_pol_mutex);
1421 	return &blkcg->css;
1422 
1423 free_pd_blkcg:
1424 	for (i--; i >= 0; i--)
1425 		if (blkcg->cpd[i])
1426 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1427 	free_percpu(blkcg->lhead);
1428 free_blkcg:
1429 	if (blkcg != &blkcg_root)
1430 		kfree(blkcg);
1431 unlock:
1432 	mutex_unlock(&blkcg_pol_mutex);
1433 	return ERR_PTR(-ENOMEM);
1434 }
1435 
blkcg_css_online(struct cgroup_subsys_state * css)1436 static int blkcg_css_online(struct cgroup_subsys_state *css)
1437 {
1438 	struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1439 
1440 	/*
1441 	 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1442 	 * don't go offline while cgwbs are still active on them.  Pin the
1443 	 * parent so that offline always happens towards the root.
1444 	 */
1445 	if (parent)
1446 		blkcg_pin_online(&parent->css);
1447 	return 0;
1448 }
1449 
blkg_init_queue(struct request_queue * q)1450 void blkg_init_queue(struct request_queue *q)
1451 {
1452 	INIT_LIST_HEAD(&q->blkg_list);
1453 	mutex_init(&q->blkcg_mutex);
1454 }
1455 
blkcg_init_disk(struct gendisk * disk)1456 int blkcg_init_disk(struct gendisk *disk)
1457 {
1458 	struct request_queue *q = disk->queue;
1459 	struct blkcg_gq *new_blkg, *blkg;
1460 	bool preloaded;
1461 
1462 	new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1463 	if (!new_blkg)
1464 		return -ENOMEM;
1465 
1466 	preloaded = !radix_tree_preload(GFP_KERNEL);
1467 
1468 	/* Make sure the root blkg exists. */
1469 	/* spin_lock_irq can serve as RCU read-side critical section. */
1470 	spin_lock_irq(&q->queue_lock);
1471 	blkg = blkg_create(&blkcg_root, disk, new_blkg);
1472 	if (IS_ERR(blkg))
1473 		goto err_unlock;
1474 	q->root_blkg = blkg;
1475 	spin_unlock_irq(&q->queue_lock);
1476 
1477 	if (preloaded)
1478 		radix_tree_preload_end();
1479 
1480 	return 0;
1481 
1482 err_unlock:
1483 	spin_unlock_irq(&q->queue_lock);
1484 	if (preloaded)
1485 		radix_tree_preload_end();
1486 	return PTR_ERR(blkg);
1487 }
1488 
blkcg_exit_disk(struct gendisk * disk)1489 void blkcg_exit_disk(struct gendisk *disk)
1490 {
1491 	blkg_destroy_all(disk);
1492 	blk_throtl_exit(disk);
1493 }
1494 
blkcg_exit(struct task_struct * tsk)1495 static void blkcg_exit(struct task_struct *tsk)
1496 {
1497 	if (tsk->throttle_disk)
1498 		put_disk(tsk->throttle_disk);
1499 	tsk->throttle_disk = NULL;
1500 }
1501 
1502 struct cgroup_subsys io_cgrp_subsys = {
1503 	.css_alloc = blkcg_css_alloc,
1504 	.css_online = blkcg_css_online,
1505 	.css_offline = blkcg_css_offline,
1506 	.css_free = blkcg_css_free,
1507 	.css_rstat_flush = blkcg_rstat_flush,
1508 	.dfl_cftypes = blkcg_files,
1509 	.legacy_cftypes = blkcg_legacy_files,
1510 	.legacy_name = "blkio",
1511 	.exit = blkcg_exit,
1512 #ifdef CONFIG_MEMCG
1513 	/*
1514 	 * This ensures that, if available, memcg is automatically enabled
1515 	 * together on the default hierarchy so that the owner cgroup can
1516 	 * be retrieved from writeback pages.
1517 	 */
1518 	.depends_on = 1 << memory_cgrp_id,
1519 #endif
1520 };
1521 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1522 
1523 /**
1524  * blkcg_activate_policy - activate a blkcg policy on a gendisk
1525  * @disk: gendisk of interest
1526  * @pol: blkcg policy to activate
1527  *
1528  * Activate @pol on @disk.  Requires %GFP_KERNEL context.  @disk goes through
1529  * bypass mode to populate its blkgs with policy_data for @pol.
1530  *
1531  * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1532  * from IO path.  Update of each blkg is protected by both queue and blkcg
1533  * locks so that holding either lock and testing blkcg_policy_enabled() is
1534  * always enough for dereferencing policy data.
1535  *
1536  * The caller is responsible for synchronizing [de]activations and policy
1537  * [un]registerations.  Returns 0 on success, -errno on failure.
1538  */
blkcg_activate_policy(struct gendisk * disk,const struct blkcg_policy * pol)1539 int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1540 {
1541 	struct request_queue *q = disk->queue;
1542 	struct blkg_policy_data *pd_prealloc = NULL;
1543 	struct blkcg_gq *blkg, *pinned_blkg = NULL;
1544 	int ret;
1545 
1546 	if (blkcg_policy_enabled(q, pol))
1547 		return 0;
1548 
1549 	/*
1550 	 * Policy is allowed to be registered without pd_alloc_fn/pd_free_fn,
1551 	 * for example, ioprio. Such policy will work on blkcg level, not disk
1552 	 * level, and don't need to be activated.
1553 	 */
1554 	if (WARN_ON_ONCE(!pol->pd_alloc_fn || !pol->pd_free_fn))
1555 		return -EINVAL;
1556 
1557 	if (queue_is_mq(q))
1558 		blk_mq_freeze_queue(q);
1559 retry:
1560 	spin_lock_irq(&q->queue_lock);
1561 
1562 	/* blkg_list is pushed at the head, reverse walk to initialize parents first */
1563 	list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1564 		struct blkg_policy_data *pd;
1565 
1566 		if (blkg->pd[pol->plid])
1567 			continue;
1568 
1569 		/* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1570 		if (blkg == pinned_blkg) {
1571 			pd = pd_prealloc;
1572 			pd_prealloc = NULL;
1573 		} else {
1574 			pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1575 					      GFP_NOWAIT | __GFP_NOWARN);
1576 		}
1577 
1578 		if (!pd) {
1579 			/*
1580 			 * GFP_NOWAIT failed.  Free the existing one and
1581 			 * prealloc for @blkg w/ GFP_KERNEL.
1582 			 */
1583 			if (pinned_blkg)
1584 				blkg_put(pinned_blkg);
1585 			blkg_get(blkg);
1586 			pinned_blkg = blkg;
1587 
1588 			spin_unlock_irq(&q->queue_lock);
1589 
1590 			if (pd_prealloc)
1591 				pol->pd_free_fn(pd_prealloc);
1592 			pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1593 						       GFP_KERNEL);
1594 			if (pd_prealloc)
1595 				goto retry;
1596 			else
1597 				goto enomem;
1598 		}
1599 
1600 		spin_lock(&blkg->blkcg->lock);
1601 
1602 		pd->blkg = blkg;
1603 		pd->plid = pol->plid;
1604 		blkg->pd[pol->plid] = pd;
1605 
1606 		if (pol->pd_init_fn)
1607 			pol->pd_init_fn(pd);
1608 
1609 		if (pol->pd_online_fn)
1610 			pol->pd_online_fn(pd);
1611 		pd->online = true;
1612 
1613 		spin_unlock(&blkg->blkcg->lock);
1614 	}
1615 
1616 	__set_bit(pol->plid, q->blkcg_pols);
1617 	ret = 0;
1618 
1619 	spin_unlock_irq(&q->queue_lock);
1620 out:
1621 	if (queue_is_mq(q))
1622 		blk_mq_unfreeze_queue(q);
1623 	if (pinned_blkg)
1624 		blkg_put(pinned_blkg);
1625 	if (pd_prealloc)
1626 		pol->pd_free_fn(pd_prealloc);
1627 	return ret;
1628 
1629 enomem:
1630 	/* alloc failed, take down everything */
1631 	spin_lock_irq(&q->queue_lock);
1632 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1633 		struct blkcg *blkcg = blkg->blkcg;
1634 		struct blkg_policy_data *pd;
1635 
1636 		spin_lock(&blkcg->lock);
1637 		pd = blkg->pd[pol->plid];
1638 		if (pd) {
1639 			if (pd->online && pol->pd_offline_fn)
1640 				pol->pd_offline_fn(pd);
1641 			pd->online = false;
1642 			pol->pd_free_fn(pd);
1643 			blkg->pd[pol->plid] = NULL;
1644 		}
1645 		spin_unlock(&blkcg->lock);
1646 	}
1647 	spin_unlock_irq(&q->queue_lock);
1648 	ret = -ENOMEM;
1649 	goto out;
1650 }
1651 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1652 
1653 /**
1654  * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1655  * @disk: gendisk of interest
1656  * @pol: blkcg policy to deactivate
1657  *
1658  * Deactivate @pol on @disk.  Follows the same synchronization rules as
1659  * blkcg_activate_policy().
1660  */
blkcg_deactivate_policy(struct gendisk * disk,const struct blkcg_policy * pol)1661 void blkcg_deactivate_policy(struct gendisk *disk,
1662 			     const struct blkcg_policy *pol)
1663 {
1664 	struct request_queue *q = disk->queue;
1665 	struct blkcg_gq *blkg;
1666 
1667 	if (!blkcg_policy_enabled(q, pol))
1668 		return;
1669 
1670 	if (queue_is_mq(q))
1671 		blk_mq_freeze_queue(q);
1672 
1673 	mutex_lock(&q->blkcg_mutex);
1674 	spin_lock_irq(&q->queue_lock);
1675 
1676 	__clear_bit(pol->plid, q->blkcg_pols);
1677 
1678 	list_for_each_entry(blkg, &q->blkg_list, q_node) {
1679 		struct blkcg *blkcg = blkg->blkcg;
1680 
1681 		spin_lock(&blkcg->lock);
1682 		if (blkg->pd[pol->plid]) {
1683 			if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1684 				pol->pd_offline_fn(blkg->pd[pol->plid]);
1685 			pol->pd_free_fn(blkg->pd[pol->plid]);
1686 			blkg->pd[pol->plid] = NULL;
1687 		}
1688 		spin_unlock(&blkcg->lock);
1689 	}
1690 
1691 	spin_unlock_irq(&q->queue_lock);
1692 	mutex_unlock(&q->blkcg_mutex);
1693 
1694 	if (queue_is_mq(q))
1695 		blk_mq_unfreeze_queue(q);
1696 }
1697 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1698 
blkcg_free_all_cpd(struct blkcg_policy * pol)1699 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1700 {
1701 	struct blkcg *blkcg;
1702 
1703 	list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1704 		if (blkcg->cpd[pol->plid]) {
1705 			pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1706 			blkcg->cpd[pol->plid] = NULL;
1707 		}
1708 	}
1709 }
1710 
1711 /**
1712  * blkcg_policy_register - register a blkcg policy
1713  * @pol: blkcg policy to register
1714  *
1715  * Register @pol with blkcg core.  Might sleep and @pol may be modified on
1716  * successful registration.  Returns 0 on success and -errno on failure.
1717  */
blkcg_policy_register(struct blkcg_policy * pol)1718 int blkcg_policy_register(struct blkcg_policy *pol)
1719 {
1720 	struct blkcg *blkcg;
1721 	int i, ret;
1722 
1723 	mutex_lock(&blkcg_pol_register_mutex);
1724 	mutex_lock(&blkcg_pol_mutex);
1725 
1726 	/* find an empty slot */
1727 	ret = -ENOSPC;
1728 	for (i = 0; i < BLKCG_MAX_POLS; i++)
1729 		if (!blkcg_policy[i])
1730 			break;
1731 	if (i >= BLKCG_MAX_POLS) {
1732 		pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1733 		goto err_unlock;
1734 	}
1735 
1736 	/*
1737 	 * Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs, and policy
1738 	 * without pd_alloc_fn/pd_free_fn can't be activated.
1739 	 */
1740 	if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1741 	    (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1742 		goto err_unlock;
1743 
1744 	/* register @pol */
1745 	pol->plid = i;
1746 	blkcg_policy[pol->plid] = pol;
1747 
1748 	/* allocate and install cpd's */
1749 	if (pol->cpd_alloc_fn) {
1750 		list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1751 			struct blkcg_policy_data *cpd;
1752 
1753 			cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1754 			if (!cpd)
1755 				goto err_free_cpds;
1756 
1757 			blkcg->cpd[pol->plid] = cpd;
1758 			cpd->blkcg = blkcg;
1759 			cpd->plid = pol->plid;
1760 		}
1761 	}
1762 
1763 	mutex_unlock(&blkcg_pol_mutex);
1764 
1765 	/* everything is in place, add intf files for the new policy */
1766 	if (pol->dfl_cftypes)
1767 		WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1768 					       pol->dfl_cftypes));
1769 	if (pol->legacy_cftypes)
1770 		WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1771 						  pol->legacy_cftypes));
1772 	mutex_unlock(&blkcg_pol_register_mutex);
1773 	return 0;
1774 
1775 err_free_cpds:
1776 	if (pol->cpd_free_fn)
1777 		blkcg_free_all_cpd(pol);
1778 
1779 	blkcg_policy[pol->plid] = NULL;
1780 err_unlock:
1781 	mutex_unlock(&blkcg_pol_mutex);
1782 	mutex_unlock(&blkcg_pol_register_mutex);
1783 	return ret;
1784 }
1785 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1786 
1787 /**
1788  * blkcg_policy_unregister - unregister a blkcg policy
1789  * @pol: blkcg policy to unregister
1790  *
1791  * Undo blkcg_policy_register(@pol).  Might sleep.
1792  */
blkcg_policy_unregister(struct blkcg_policy * pol)1793 void blkcg_policy_unregister(struct blkcg_policy *pol)
1794 {
1795 	mutex_lock(&blkcg_pol_register_mutex);
1796 
1797 	if (WARN_ON(blkcg_policy[pol->plid] != pol))
1798 		goto out_unlock;
1799 
1800 	/* kill the intf files first */
1801 	if (pol->dfl_cftypes)
1802 		cgroup_rm_cftypes(pol->dfl_cftypes);
1803 	if (pol->legacy_cftypes)
1804 		cgroup_rm_cftypes(pol->legacy_cftypes);
1805 
1806 	/* remove cpds and unregister */
1807 	mutex_lock(&blkcg_pol_mutex);
1808 
1809 	if (pol->cpd_free_fn)
1810 		blkcg_free_all_cpd(pol);
1811 
1812 	blkcg_policy[pol->plid] = NULL;
1813 
1814 	mutex_unlock(&blkcg_pol_mutex);
1815 out_unlock:
1816 	mutex_unlock(&blkcg_pol_register_mutex);
1817 }
1818 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1819 
1820 /*
1821  * Scale the accumulated delay based on how long it has been since we updated
1822  * the delay.  We only call this when we are adding delay, in case it's been a
1823  * while since we added delay, and when we are checking to see if we need to
1824  * delay a task, to account for any delays that may have occurred.
1825  */
blkcg_scale_delay(struct blkcg_gq * blkg,u64 now)1826 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1827 {
1828 	u64 old = atomic64_read(&blkg->delay_start);
1829 
1830 	/* negative use_delay means no scaling, see blkcg_set_delay() */
1831 	if (atomic_read(&blkg->use_delay) < 0)
1832 		return;
1833 
1834 	/*
1835 	 * We only want to scale down every second.  The idea here is that we
1836 	 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1837 	 * time window.  We only want to throttle tasks for recent delay that
1838 	 * has occurred, in 1 second time windows since that's the maximum
1839 	 * things can be throttled.  We save the current delay window in
1840 	 * blkg->last_delay so we know what amount is still left to be charged
1841 	 * to the blkg from this point onward.  blkg->last_use keeps track of
1842 	 * the use_delay counter.  The idea is if we're unthrottling the blkg we
1843 	 * are ok with whatever is happening now, and we can take away more of
1844 	 * the accumulated delay as we've already throttled enough that
1845 	 * everybody is happy with their IO latencies.
1846 	 */
1847 	if (time_before64(old + NSEC_PER_SEC, now) &&
1848 	    atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1849 		u64 cur = atomic64_read(&blkg->delay_nsec);
1850 		u64 sub = min_t(u64, blkg->last_delay, now - old);
1851 		int cur_use = atomic_read(&blkg->use_delay);
1852 
1853 		/*
1854 		 * We've been unthrottled, subtract a larger chunk of our
1855 		 * accumulated delay.
1856 		 */
1857 		if (cur_use < blkg->last_use)
1858 			sub = max_t(u64, sub, blkg->last_delay >> 1);
1859 
1860 		/*
1861 		 * This shouldn't happen, but handle it anyway.  Our delay_nsec
1862 		 * should only ever be growing except here where we subtract out
1863 		 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1864 		 * rather not end up with negative numbers.
1865 		 */
1866 		if (unlikely(cur < sub)) {
1867 			atomic64_set(&blkg->delay_nsec, 0);
1868 			blkg->last_delay = 0;
1869 		} else {
1870 			atomic64_sub(sub, &blkg->delay_nsec);
1871 			blkg->last_delay = cur - sub;
1872 		}
1873 		blkg->last_use = cur_use;
1874 	}
1875 }
1876 
1877 /*
1878  * This is called when we want to actually walk up the hierarchy and check to
1879  * see if we need to throttle, and then actually throttle if there is some
1880  * accumulated delay.  This should only be called upon return to user space so
1881  * we're not holding some lock that would induce a priority inversion.
1882  */
blkcg_maybe_throttle_blkg(struct blkcg_gq * blkg,bool use_memdelay)1883 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1884 {
1885 	unsigned long pflags;
1886 	bool clamp;
1887 	u64 now = blk_time_get_ns();
1888 	u64 exp;
1889 	u64 delay_nsec = 0;
1890 	int tok;
1891 
1892 	while (blkg->parent) {
1893 		int use_delay = atomic_read(&blkg->use_delay);
1894 
1895 		if (use_delay) {
1896 			u64 this_delay;
1897 
1898 			blkcg_scale_delay(blkg, now);
1899 			this_delay = atomic64_read(&blkg->delay_nsec);
1900 			if (this_delay > delay_nsec) {
1901 				delay_nsec = this_delay;
1902 				clamp = use_delay > 0;
1903 			}
1904 		}
1905 		blkg = blkg->parent;
1906 	}
1907 
1908 	if (!delay_nsec)
1909 		return;
1910 
1911 	/*
1912 	 * Let's not sleep for all eternity if we've amassed a huge delay.
1913 	 * Swapping or metadata IO can accumulate 10's of seconds worth of
1914 	 * delay, and we want userspace to be able to do _something_ so cap the
1915 	 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1916 	 * tasks will be delayed for 0.25 second for every syscall. If
1917 	 * blkcg_set_delay() was used as indicated by negative use_delay, the
1918 	 * caller is responsible for regulating the range.
1919 	 */
1920 	if (clamp)
1921 		delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1922 
1923 	if (use_memdelay)
1924 		psi_memstall_enter(&pflags);
1925 
1926 	exp = ktime_add_ns(now, delay_nsec);
1927 	tok = io_schedule_prepare();
1928 	do {
1929 		__set_current_state(TASK_KILLABLE);
1930 		if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1931 			break;
1932 	} while (!fatal_signal_pending(current));
1933 	io_schedule_finish(tok);
1934 
1935 	if (use_memdelay)
1936 		psi_memstall_leave(&pflags);
1937 }
1938 
1939 /**
1940  * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1941  *
1942  * This is only called if we've been marked with set_notify_resume().  Obviously
1943  * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1944  * check to see if current->throttle_disk is set and if not this doesn't do
1945  * anything.  This should only ever be called by the resume code, it's not meant
1946  * to be called by people willy-nilly as it will actually do the work to
1947  * throttle the task if it is setup for throttling.
1948  */
blkcg_maybe_throttle_current(void)1949 void blkcg_maybe_throttle_current(void)
1950 {
1951 	struct gendisk *disk = current->throttle_disk;
1952 	struct blkcg *blkcg;
1953 	struct blkcg_gq *blkg;
1954 	bool use_memdelay = current->use_memdelay;
1955 
1956 	if (!disk)
1957 		return;
1958 
1959 	current->throttle_disk = NULL;
1960 	current->use_memdelay = false;
1961 
1962 	rcu_read_lock();
1963 	blkcg = css_to_blkcg(blkcg_css());
1964 	if (!blkcg)
1965 		goto out;
1966 	blkg = blkg_lookup(blkcg, disk->queue);
1967 	if (!blkg)
1968 		goto out;
1969 	if (!blkg_tryget(blkg))
1970 		goto out;
1971 	rcu_read_unlock();
1972 
1973 	blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1974 	blkg_put(blkg);
1975 	put_disk(disk);
1976 	return;
1977 out:
1978 	rcu_read_unlock();
1979 }
1980 
1981 /**
1982  * blkcg_schedule_throttle - this task needs to check for throttling
1983  * @disk: disk to throttle
1984  * @use_memdelay: do we charge this to memory delay for PSI
1985  *
1986  * This is called by the IO controller when we know there's delay accumulated
1987  * for the blkg for this task.  We do not pass the blkg because there are places
1988  * we call this that may not have that information, the swapping code for
1989  * instance will only have a block_device at that point.  This set's the
1990  * notify_resume for the task to check and see if it requires throttling before
1991  * returning to user space.
1992  *
1993  * We will only schedule once per syscall.  You can call this over and over
1994  * again and it will only do the check once upon return to user space, and only
1995  * throttle once.  If the task needs to be throttled again it'll need to be
1996  * re-set at the next time we see the task.
1997  */
blkcg_schedule_throttle(struct gendisk * disk,bool use_memdelay)1998 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1999 {
2000 	if (unlikely(current->flags & PF_KTHREAD))
2001 		return;
2002 
2003 	if (current->throttle_disk != disk) {
2004 		if (test_bit(GD_DEAD, &disk->state))
2005 			return;
2006 		get_device(disk_to_dev(disk));
2007 
2008 		if (current->throttle_disk)
2009 			put_disk(current->throttle_disk);
2010 		current->throttle_disk = disk;
2011 	}
2012 
2013 	if (use_memdelay)
2014 		current->use_memdelay = use_memdelay;
2015 	set_notify_resume(current);
2016 }
2017 
2018 /**
2019  * blkcg_add_delay - add delay to this blkg
2020  * @blkg: blkg of interest
2021  * @now: the current time in nanoseconds
2022  * @delta: how many nanoseconds of delay to add
2023  *
2024  * Charge @delta to the blkg's current delay accumulation.  This is used to
2025  * throttle tasks if an IO controller thinks we need more throttling.
2026  */
blkcg_add_delay(struct blkcg_gq * blkg,u64 now,u64 delta)2027 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
2028 {
2029 	if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
2030 		return;
2031 	blkcg_scale_delay(blkg, now);
2032 	atomic64_add(delta, &blkg->delay_nsec);
2033 }
2034 
2035 /**
2036  * blkg_tryget_closest - try and get a blkg ref on the closet blkg
2037  * @bio: target bio
2038  * @css: target css
2039  *
2040  * As the failure mode here is to walk up the blkg tree, this ensure that the
2041  * blkg->parent pointers are always valid.  This returns the blkg that it ended
2042  * up taking a reference on or %NULL if no reference was taken.
2043  */
blkg_tryget_closest(struct bio * bio,struct cgroup_subsys_state * css)2044 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
2045 		struct cgroup_subsys_state *css)
2046 {
2047 	struct blkcg_gq *blkg, *ret_blkg = NULL;
2048 
2049 	rcu_read_lock();
2050 	blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
2051 	while (blkg) {
2052 		if (blkg_tryget(blkg)) {
2053 			ret_blkg = blkg;
2054 			break;
2055 		}
2056 		blkg = blkg->parent;
2057 	}
2058 	rcu_read_unlock();
2059 
2060 	return ret_blkg;
2061 }
2062 
2063 /**
2064  * bio_associate_blkg_from_css - associate a bio with a specified css
2065  * @bio: target bio
2066  * @css: target css
2067  *
2068  * Associate @bio with the blkg found by combining the css's blkg and the
2069  * request_queue of the @bio.  An association failure is handled by walking up
2070  * the blkg tree.  Therefore, the blkg associated can be anything between @blkg
2071  * and q->root_blkg.  This situation only happens when a cgroup is dying and
2072  * then the remaining bios will spill to the closest alive blkg.
2073  *
2074  * A reference will be taken on the blkg and will be released when @bio is
2075  * freed.
2076  */
bio_associate_blkg_from_css(struct bio * bio,struct cgroup_subsys_state * css)2077 void bio_associate_blkg_from_css(struct bio *bio,
2078 				 struct cgroup_subsys_state *css)
2079 {
2080 	if (bio->bi_blkg)
2081 		blkg_put(bio->bi_blkg);
2082 
2083 	if (css && css->parent) {
2084 		bio->bi_blkg = blkg_tryget_closest(bio, css);
2085 	} else {
2086 		blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
2087 		bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
2088 	}
2089 }
2090 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2091 
2092 /**
2093  * bio_associate_blkg - associate a bio with a blkg
2094  * @bio: target bio
2095  *
2096  * Associate @bio with the blkg found from the bio's css and request_queue.
2097  * If one is not found, bio_lookup_blkg() creates the blkg.  If a blkg is
2098  * already associated, the css is reused and association redone as the
2099  * request_queue may have changed.
2100  */
bio_associate_blkg(struct bio * bio)2101 void bio_associate_blkg(struct bio *bio)
2102 {
2103 	struct cgroup_subsys_state *css;
2104 
2105 	if (blk_op_is_passthrough(bio->bi_opf))
2106 		return;
2107 
2108 	rcu_read_lock();
2109 
2110 	if (bio->bi_blkg)
2111 		css = bio_blkcg_css(bio);
2112 	else
2113 		css = blkcg_css();
2114 
2115 	bio_associate_blkg_from_css(bio, css);
2116 
2117 	rcu_read_unlock();
2118 }
2119 EXPORT_SYMBOL_GPL(bio_associate_blkg);
2120 
2121 /**
2122  * bio_clone_blkg_association - clone blkg association from src to dst bio
2123  * @dst: destination bio
2124  * @src: source bio
2125  */
bio_clone_blkg_association(struct bio * dst,struct bio * src)2126 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2127 {
2128 	if (src->bi_blkg)
2129 		bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2130 }
2131 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2132 
blk_cgroup_io_type(struct bio * bio)2133 static int blk_cgroup_io_type(struct bio *bio)
2134 {
2135 	if (op_is_discard(bio->bi_opf))
2136 		return BLKG_IOSTAT_DISCARD;
2137 	if (op_is_write(bio->bi_opf))
2138 		return BLKG_IOSTAT_WRITE;
2139 	return BLKG_IOSTAT_READ;
2140 }
2141 
blk_cgroup_bio_start(struct bio * bio)2142 void blk_cgroup_bio_start(struct bio *bio)
2143 {
2144 	struct blkcg *blkcg = bio->bi_blkg->blkcg;
2145 	int rwd = blk_cgroup_io_type(bio), cpu;
2146 	struct blkg_iostat_set *bis;
2147 	unsigned long flags;
2148 
2149 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
2150 		return;
2151 
2152 	/* Root-level stats are sourced from system-wide IO stats */
2153 	if (!cgroup_parent(blkcg->css.cgroup))
2154 		return;
2155 
2156 	cpu = get_cpu();
2157 	bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2158 	flags = u64_stats_update_begin_irqsave(&bis->sync);
2159 
2160 	/*
2161 	 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2162 	 * bio and we would have already accounted for the size of the bio.
2163 	 */
2164 	if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2165 		bio_set_flag(bio, BIO_CGROUP_ACCT);
2166 		bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2167 	}
2168 	bis->cur.ios[rwd]++;
2169 
2170 	/*
2171 	 * If the iostat_cpu isn't in a lockless list, put it into the
2172 	 * list to indicate that a stat update is pending.
2173 	 */
2174 	if (!READ_ONCE(bis->lqueued)) {
2175 		struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2176 
2177 		llist_add(&bis->lnode, lhead);
2178 		WRITE_ONCE(bis->lqueued, true);
2179 	}
2180 
2181 	u64_stats_update_end_irqrestore(&bis->sync, flags);
2182 	cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2183 	put_cpu();
2184 }
2185 
blk_cgroup_congested(void)2186 bool blk_cgroup_congested(void)
2187 {
2188 	struct blkcg *blkcg;
2189 	bool ret = false;
2190 
2191 	rcu_read_lock();
2192 	for (blkcg = css_to_blkcg(blkcg_css()); blkcg;
2193 	     blkcg = blkcg_parent(blkcg)) {
2194 		if (atomic_read(&blkcg->congestion_count)) {
2195 			ret = true;
2196 			break;
2197 		}
2198 	}
2199 	rcu_read_unlock();
2200 	return ret;
2201 }
2202 
2203 module_param(blkcg_debug_stats, bool, 0644);
2204 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");
2205