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