xref: /openbmc/linux/block/bfq-cgroup.c (revision c059ee9d)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * cgroups support for the BFQ I/O scheduler.
4  */
5 #include <linux/module.h>
6 #include <linux/slab.h>
7 #include <linux/blkdev.h>
8 #include <linux/cgroup.h>
9 #include <linux/ktime.h>
10 #include <linux/rbtree.h>
11 #include <linux/ioprio.h>
12 #include <linux/sbitmap.h>
13 #include <linux/delay.h>
14 
15 #include "elevator.h"
16 #include "bfq-iosched.h"
17 
18 #ifdef CONFIG_BFQ_CGROUP_DEBUG
19 static int bfq_stat_init(struct bfq_stat *stat, gfp_t gfp)
20 {
21 	int ret;
22 
23 	ret = percpu_counter_init(&stat->cpu_cnt, 0, gfp);
24 	if (ret)
25 		return ret;
26 
27 	atomic64_set(&stat->aux_cnt, 0);
28 	return 0;
29 }
30 
31 static void bfq_stat_exit(struct bfq_stat *stat)
32 {
33 	percpu_counter_destroy(&stat->cpu_cnt);
34 }
35 
36 /**
37  * bfq_stat_add - add a value to a bfq_stat
38  * @stat: target bfq_stat
39  * @val: value to add
40  *
41  * Add @val to @stat.  The caller must ensure that IRQ on the same CPU
42  * don't re-enter this function for the same counter.
43  */
44 static inline void bfq_stat_add(struct bfq_stat *stat, uint64_t val)
45 {
46 	percpu_counter_add_batch(&stat->cpu_cnt, val, BLKG_STAT_CPU_BATCH);
47 }
48 
49 /**
50  * bfq_stat_read - read the current value of a bfq_stat
51  * @stat: bfq_stat to read
52  */
53 static inline uint64_t bfq_stat_read(struct bfq_stat *stat)
54 {
55 	return percpu_counter_sum_positive(&stat->cpu_cnt);
56 }
57 
58 /**
59  * bfq_stat_reset - reset a bfq_stat
60  * @stat: bfq_stat to reset
61  */
62 static inline void bfq_stat_reset(struct bfq_stat *stat)
63 {
64 	percpu_counter_set(&stat->cpu_cnt, 0);
65 	atomic64_set(&stat->aux_cnt, 0);
66 }
67 
68 /**
69  * bfq_stat_add_aux - add a bfq_stat into another's aux count
70  * @to: the destination bfq_stat
71  * @from: the source
72  *
73  * Add @from's count including the aux one to @to's aux count.
74  */
75 static inline void bfq_stat_add_aux(struct bfq_stat *to,
76 				     struct bfq_stat *from)
77 {
78 	atomic64_add(bfq_stat_read(from) + atomic64_read(&from->aux_cnt),
79 		     &to->aux_cnt);
80 }
81 
82 /**
83  * blkg_prfill_stat - prfill callback for bfq_stat
84  * @sf: seq_file to print to
85  * @pd: policy private data of interest
86  * @off: offset to the bfq_stat in @pd
87  *
88  * prfill callback for printing a bfq_stat.
89  */
90 static u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd,
91 		int off)
92 {
93 	return __blkg_prfill_u64(sf, pd, bfq_stat_read((void *)pd + off));
94 }
95 
96 /* bfqg stats flags */
97 enum bfqg_stats_flags {
98 	BFQG_stats_waiting = 0,
99 	BFQG_stats_idling,
100 	BFQG_stats_empty,
101 };
102 
103 #define BFQG_FLAG_FNS(name)						\
104 static void bfqg_stats_mark_##name(struct bfqg_stats *stats)	\
105 {									\
106 	stats->flags |= (1 << BFQG_stats_##name);			\
107 }									\
108 static void bfqg_stats_clear_##name(struct bfqg_stats *stats)	\
109 {									\
110 	stats->flags &= ~(1 << BFQG_stats_##name);			\
111 }									\
112 static int bfqg_stats_##name(struct bfqg_stats *stats)		\
113 {									\
114 	return (stats->flags & (1 << BFQG_stats_##name)) != 0;		\
115 }									\
116 
117 BFQG_FLAG_FNS(waiting)
118 BFQG_FLAG_FNS(idling)
119 BFQG_FLAG_FNS(empty)
120 #undef BFQG_FLAG_FNS
121 
122 /* This should be called with the scheduler lock held. */
123 static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
124 {
125 	u64 now;
126 
127 	if (!bfqg_stats_waiting(stats))
128 		return;
129 
130 	now = ktime_get_ns();
131 	if (now > stats->start_group_wait_time)
132 		bfq_stat_add(&stats->group_wait_time,
133 			      now - stats->start_group_wait_time);
134 	bfqg_stats_clear_waiting(stats);
135 }
136 
137 /* This should be called with the scheduler lock held. */
138 static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg,
139 						 struct bfq_group *curr_bfqg)
140 {
141 	struct bfqg_stats *stats = &bfqg->stats;
142 
143 	if (bfqg_stats_waiting(stats))
144 		return;
145 	if (bfqg == curr_bfqg)
146 		return;
147 	stats->start_group_wait_time = ktime_get_ns();
148 	bfqg_stats_mark_waiting(stats);
149 }
150 
151 /* This should be called with the scheduler lock held. */
152 static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
153 {
154 	u64 now;
155 
156 	if (!bfqg_stats_empty(stats))
157 		return;
158 
159 	now = ktime_get_ns();
160 	if (now > stats->start_empty_time)
161 		bfq_stat_add(&stats->empty_time,
162 			      now - stats->start_empty_time);
163 	bfqg_stats_clear_empty(stats);
164 }
165 
166 void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
167 {
168 	bfq_stat_add(&bfqg->stats.dequeue, 1);
169 }
170 
171 void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)
172 {
173 	struct bfqg_stats *stats = &bfqg->stats;
174 
175 	if (blkg_rwstat_total(&stats->queued))
176 		return;
177 
178 	/*
179 	 * group is already marked empty. This can happen if bfqq got new
180 	 * request in parent group and moved to this group while being added
181 	 * to service tree. Just ignore the event and move on.
182 	 */
183 	if (bfqg_stats_empty(stats))
184 		return;
185 
186 	stats->start_empty_time = ktime_get_ns();
187 	bfqg_stats_mark_empty(stats);
188 }
189 
190 void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
191 {
192 	struct bfqg_stats *stats = &bfqg->stats;
193 
194 	if (bfqg_stats_idling(stats)) {
195 		u64 now = ktime_get_ns();
196 
197 		if (now > stats->start_idle_time)
198 			bfq_stat_add(&stats->idle_time,
199 				      now - stats->start_idle_time);
200 		bfqg_stats_clear_idling(stats);
201 	}
202 }
203 
204 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
205 {
206 	struct bfqg_stats *stats = &bfqg->stats;
207 
208 	stats->start_idle_time = ktime_get_ns();
209 	bfqg_stats_mark_idling(stats);
210 }
211 
212 void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
213 {
214 	struct bfqg_stats *stats = &bfqg->stats;
215 
216 	bfq_stat_add(&stats->avg_queue_size_sum,
217 		      blkg_rwstat_total(&stats->queued));
218 	bfq_stat_add(&stats->avg_queue_size_samples, 1);
219 	bfqg_stats_update_group_wait_time(stats);
220 }
221 
222 void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
223 			      unsigned int op)
224 {
225 	blkg_rwstat_add(&bfqg->stats.queued, op, 1);
226 	bfqg_stats_end_empty_time(&bfqg->stats);
227 	if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue))
228 		bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq));
229 }
230 
231 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
232 {
233 	blkg_rwstat_add(&bfqg->stats.queued, op, -1);
234 }
235 
236 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
237 {
238 	blkg_rwstat_add(&bfqg->stats.merged, op, 1);
239 }
240 
241 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
242 				  u64 io_start_time_ns, unsigned int op)
243 {
244 	struct bfqg_stats *stats = &bfqg->stats;
245 	u64 now = ktime_get_ns();
246 
247 	if (now > io_start_time_ns)
248 		blkg_rwstat_add(&stats->service_time, op,
249 				now - io_start_time_ns);
250 	if (io_start_time_ns > start_time_ns)
251 		blkg_rwstat_add(&stats->wait_time, op,
252 				io_start_time_ns - start_time_ns);
253 }
254 
255 #else /* CONFIG_BFQ_CGROUP_DEBUG */
256 
257 void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
258 			      unsigned int op) { }
259 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { }
260 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { }
261 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
262 				  u64 io_start_time_ns, unsigned int op) { }
263 void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
264 void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { }
265 void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { }
266 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
267 void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { }
268 
269 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
270 
271 #ifdef CONFIG_BFQ_GROUP_IOSCHED
272 
273 /*
274  * blk-cgroup policy-related handlers
275  * The following functions help in converting between blk-cgroup
276  * internal structures and BFQ-specific structures.
277  */
278 
279 static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
280 {
281 	return pd ? container_of(pd, struct bfq_group, pd) : NULL;
282 }
283 
284 struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
285 {
286 	return pd_to_blkg(&bfqg->pd);
287 }
288 
289 static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
290 {
291 	return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
292 }
293 
294 /*
295  * bfq_group handlers
296  * The following functions help in navigating the bfq_group hierarchy
297  * by allowing to find the parent of a bfq_group or the bfq_group
298  * associated to a bfq_queue.
299  */
300 
301 static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
302 {
303 	struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
304 
305 	return pblkg ? blkg_to_bfqg(pblkg) : NULL;
306 }
307 
308 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
309 {
310 	struct bfq_entity *group_entity = bfqq->entity.parent;
311 
312 	return group_entity ? container_of(group_entity, struct bfq_group,
313 					   entity) :
314 			      bfqq->bfqd->root_group;
315 }
316 
317 /*
318  * The following two functions handle get and put of a bfq_group by
319  * wrapping the related blk-cgroup hooks.
320  */
321 
322 static void bfqg_get(struct bfq_group *bfqg)
323 {
324 	bfqg->ref++;
325 }
326 
327 static void bfqg_put(struct bfq_group *bfqg)
328 {
329 	bfqg->ref--;
330 
331 	if (bfqg->ref == 0)
332 		kfree(bfqg);
333 }
334 
335 static void bfqg_and_blkg_get(struct bfq_group *bfqg)
336 {
337 	/* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
338 	bfqg_get(bfqg);
339 
340 	blkg_get(bfqg_to_blkg(bfqg));
341 }
342 
343 void bfqg_and_blkg_put(struct bfq_group *bfqg)
344 {
345 	blkg_put(bfqg_to_blkg(bfqg));
346 
347 	bfqg_put(bfqg);
348 }
349 
350 void bfqg_stats_update_legacy_io(struct request_queue *q, struct request *rq)
351 {
352 	struct bfq_group *bfqg = blkg_to_bfqg(rq->bio->bi_blkg);
353 
354 	if (!bfqg)
355 		return;
356 
357 	blkg_rwstat_add(&bfqg->stats.bytes, rq->cmd_flags, blk_rq_bytes(rq));
358 	blkg_rwstat_add(&bfqg->stats.ios, rq->cmd_flags, 1);
359 }
360 
361 /* @stats = 0 */
362 static void bfqg_stats_reset(struct bfqg_stats *stats)
363 {
364 #ifdef CONFIG_BFQ_CGROUP_DEBUG
365 	/* queued stats shouldn't be cleared */
366 	blkg_rwstat_reset(&stats->merged);
367 	blkg_rwstat_reset(&stats->service_time);
368 	blkg_rwstat_reset(&stats->wait_time);
369 	bfq_stat_reset(&stats->time);
370 	bfq_stat_reset(&stats->avg_queue_size_sum);
371 	bfq_stat_reset(&stats->avg_queue_size_samples);
372 	bfq_stat_reset(&stats->dequeue);
373 	bfq_stat_reset(&stats->group_wait_time);
374 	bfq_stat_reset(&stats->idle_time);
375 	bfq_stat_reset(&stats->empty_time);
376 #endif
377 }
378 
379 /* @to += @from */
380 static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
381 {
382 	if (!to || !from)
383 		return;
384 
385 #ifdef CONFIG_BFQ_CGROUP_DEBUG
386 	/* queued stats shouldn't be cleared */
387 	blkg_rwstat_add_aux(&to->merged, &from->merged);
388 	blkg_rwstat_add_aux(&to->service_time, &from->service_time);
389 	blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
390 	bfq_stat_add_aux(&from->time, &from->time);
391 	bfq_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
392 	bfq_stat_add_aux(&to->avg_queue_size_samples,
393 			  &from->avg_queue_size_samples);
394 	bfq_stat_add_aux(&to->dequeue, &from->dequeue);
395 	bfq_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
396 	bfq_stat_add_aux(&to->idle_time, &from->idle_time);
397 	bfq_stat_add_aux(&to->empty_time, &from->empty_time);
398 #endif
399 }
400 
401 /*
402  * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
403  * recursive stats can still account for the amount used by this bfqg after
404  * it's gone.
405  */
406 static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
407 {
408 	struct bfq_group *parent;
409 
410 	if (!bfqg) /* root_group */
411 		return;
412 
413 	parent = bfqg_parent(bfqg);
414 
415 	lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
416 
417 	if (unlikely(!parent))
418 		return;
419 
420 	bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
421 	bfqg_stats_reset(&bfqg->stats);
422 }
423 
424 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
425 {
426 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
427 
428 	entity->weight = entity->new_weight;
429 	entity->orig_weight = entity->new_weight;
430 	if (bfqq) {
431 		bfqq->ioprio = bfqq->new_ioprio;
432 		bfqq->ioprio_class = bfqq->new_ioprio_class;
433 		/*
434 		 * Make sure that bfqg and its associated blkg do not
435 		 * disappear before entity.
436 		 */
437 		bfqg_and_blkg_get(bfqg);
438 	}
439 	entity->parent = bfqg->my_entity; /* NULL for root group */
440 	entity->sched_data = &bfqg->sched_data;
441 }
442 
443 static void bfqg_stats_exit(struct bfqg_stats *stats)
444 {
445 	blkg_rwstat_exit(&stats->bytes);
446 	blkg_rwstat_exit(&stats->ios);
447 #ifdef CONFIG_BFQ_CGROUP_DEBUG
448 	blkg_rwstat_exit(&stats->merged);
449 	blkg_rwstat_exit(&stats->service_time);
450 	blkg_rwstat_exit(&stats->wait_time);
451 	blkg_rwstat_exit(&stats->queued);
452 	bfq_stat_exit(&stats->time);
453 	bfq_stat_exit(&stats->avg_queue_size_sum);
454 	bfq_stat_exit(&stats->avg_queue_size_samples);
455 	bfq_stat_exit(&stats->dequeue);
456 	bfq_stat_exit(&stats->group_wait_time);
457 	bfq_stat_exit(&stats->idle_time);
458 	bfq_stat_exit(&stats->empty_time);
459 #endif
460 }
461 
462 static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
463 {
464 	if (blkg_rwstat_init(&stats->bytes, gfp) ||
465 	    blkg_rwstat_init(&stats->ios, gfp))
466 		goto error;
467 
468 #ifdef CONFIG_BFQ_CGROUP_DEBUG
469 	if (blkg_rwstat_init(&stats->merged, gfp) ||
470 	    blkg_rwstat_init(&stats->service_time, gfp) ||
471 	    blkg_rwstat_init(&stats->wait_time, gfp) ||
472 	    blkg_rwstat_init(&stats->queued, gfp) ||
473 	    bfq_stat_init(&stats->time, gfp) ||
474 	    bfq_stat_init(&stats->avg_queue_size_sum, gfp) ||
475 	    bfq_stat_init(&stats->avg_queue_size_samples, gfp) ||
476 	    bfq_stat_init(&stats->dequeue, gfp) ||
477 	    bfq_stat_init(&stats->group_wait_time, gfp) ||
478 	    bfq_stat_init(&stats->idle_time, gfp) ||
479 	    bfq_stat_init(&stats->empty_time, gfp))
480 		goto error;
481 #endif
482 
483 	return 0;
484 
485 error:
486 	bfqg_stats_exit(stats);
487 	return -ENOMEM;
488 }
489 
490 static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
491 {
492 	return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
493 }
494 
495 static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
496 {
497 	return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
498 }
499 
500 static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
501 {
502 	struct bfq_group_data *bgd;
503 
504 	bgd = kzalloc(sizeof(*bgd), gfp);
505 	if (!bgd)
506 		return NULL;
507 	return &bgd->pd;
508 }
509 
510 static void bfq_cpd_init(struct blkcg_policy_data *cpd)
511 {
512 	struct bfq_group_data *d = cpd_to_bfqgd(cpd);
513 
514 	d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
515 		CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
516 }
517 
518 static void bfq_cpd_free(struct blkcg_policy_data *cpd)
519 {
520 	kfree(cpd_to_bfqgd(cpd));
521 }
522 
523 static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
524 					     struct blkcg *blkcg)
525 {
526 	struct bfq_group *bfqg;
527 
528 	bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
529 	if (!bfqg)
530 		return NULL;
531 
532 	if (bfqg_stats_init(&bfqg->stats, gfp)) {
533 		kfree(bfqg);
534 		return NULL;
535 	}
536 
537 	/* see comments in bfq_bic_update_cgroup for why refcounting */
538 	bfqg_get(bfqg);
539 	return &bfqg->pd;
540 }
541 
542 static void bfq_pd_init(struct blkg_policy_data *pd)
543 {
544 	struct blkcg_gq *blkg = pd_to_blkg(pd);
545 	struct bfq_group *bfqg = blkg_to_bfqg(blkg);
546 	struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
547 	struct bfq_entity *entity = &bfqg->entity;
548 	struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
549 
550 	entity->orig_weight = entity->weight = entity->new_weight = d->weight;
551 	entity->my_sched_data = &bfqg->sched_data;
552 	entity->last_bfqq_created = NULL;
553 
554 	bfqg->my_entity = entity; /*
555 				   * the root_group's will be set to NULL
556 				   * in bfq_init_queue()
557 				   */
558 	bfqg->bfqd = bfqd;
559 	bfqg->active_entities = 0;
560 	bfqg->online = true;
561 	bfqg->rq_pos_tree = RB_ROOT;
562 }
563 
564 static void bfq_pd_free(struct blkg_policy_data *pd)
565 {
566 	struct bfq_group *bfqg = pd_to_bfqg(pd);
567 
568 	bfqg_stats_exit(&bfqg->stats);
569 	bfqg_put(bfqg);
570 }
571 
572 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
573 {
574 	struct bfq_group *bfqg = pd_to_bfqg(pd);
575 
576 	bfqg_stats_reset(&bfqg->stats);
577 }
578 
579 static void bfq_group_set_parent(struct bfq_group *bfqg,
580 					struct bfq_group *parent)
581 {
582 	struct bfq_entity *entity;
583 
584 	entity = &bfqg->entity;
585 	entity->parent = parent->my_entity;
586 	entity->sched_data = &parent->sched_data;
587 }
588 
589 static void bfq_link_bfqg(struct bfq_data *bfqd, struct bfq_group *bfqg)
590 {
591 	struct bfq_group *parent;
592 	struct bfq_entity *entity;
593 
594 	/*
595 	 * Update chain of bfq_groups as we might be handling a leaf group
596 	 * which, along with some of its relatives, has not been hooked yet
597 	 * to the private hierarchy of BFQ.
598 	 */
599 	entity = &bfqg->entity;
600 	for_each_entity(entity) {
601 		struct bfq_group *curr_bfqg = container_of(entity,
602 						struct bfq_group, entity);
603 		if (curr_bfqg != bfqd->root_group) {
604 			parent = bfqg_parent(curr_bfqg);
605 			if (!parent)
606 				parent = bfqd->root_group;
607 			bfq_group_set_parent(curr_bfqg, parent);
608 		}
609 	}
610 }
611 
612 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
613 {
614 	struct blkcg_gq *blkg = bio->bi_blkg;
615 	struct bfq_group *bfqg;
616 
617 	while (blkg) {
618 		bfqg = blkg_to_bfqg(blkg);
619 		if (bfqg->online) {
620 			bio_associate_blkg_from_css(bio, &blkg->blkcg->css);
621 			return bfqg;
622 		}
623 		blkg = blkg->parent;
624 	}
625 	bio_associate_blkg_from_css(bio,
626 				&bfqg_to_blkg(bfqd->root_group)->blkcg->css);
627 	return bfqd->root_group;
628 }
629 
630 /**
631  * bfq_bfqq_move - migrate @bfqq to @bfqg.
632  * @bfqd: queue descriptor.
633  * @bfqq: the queue to move.
634  * @bfqg: the group to move to.
635  *
636  * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
637  * it on the new one.  Avoid putting the entity on the old group idle tree.
638  *
639  * Must be called under the scheduler lock, to make sure that the blkg
640  * owning @bfqg does not disappear (see comments in
641  * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
642  * objects).
643  */
644 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
645 		   struct bfq_group *bfqg)
646 {
647 	struct bfq_entity *entity = &bfqq->entity;
648 	struct bfq_group *old_parent = bfqq_group(bfqq);
649 
650 	/*
651 	 * No point to move bfqq to the same group, which can happen when
652 	 * root group is offlined
653 	 */
654 	if (old_parent == bfqg)
655 		return;
656 
657 	/*
658 	 * oom_bfqq is not allowed to move, oom_bfqq will hold ref to root_group
659 	 * until elevator exit.
660 	 */
661 	if (bfqq == &bfqd->oom_bfqq)
662 		return;
663 	/*
664 	 * Get extra reference to prevent bfqq from being freed in
665 	 * next possible expire or deactivate.
666 	 */
667 	bfqq->ref++;
668 
669 	/* If bfqq is empty, then bfq_bfqq_expire also invokes
670 	 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
671 	 * from data structures related to current group. Otherwise we
672 	 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
673 	 * we do below.
674 	 */
675 	if (bfqq == bfqd->in_service_queue)
676 		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
677 				false, BFQQE_PREEMPTED);
678 
679 	if (bfq_bfqq_busy(bfqq))
680 		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
681 	else if (entity->on_st_or_in_serv)
682 		bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
683 	bfqg_and_blkg_put(old_parent);
684 
685 	if (entity->parent &&
686 	    entity->parent->last_bfqq_created == bfqq)
687 		entity->parent->last_bfqq_created = NULL;
688 	else if (bfqd->last_bfqq_created == bfqq)
689 		bfqd->last_bfqq_created = NULL;
690 
691 	entity->parent = bfqg->my_entity;
692 	entity->sched_data = &bfqg->sched_data;
693 	/* pin down bfqg and its associated blkg  */
694 	bfqg_and_blkg_get(bfqg);
695 
696 	if (bfq_bfqq_busy(bfqq)) {
697 		if (unlikely(!bfqd->nonrot_with_queueing))
698 			bfq_pos_tree_add_move(bfqd, bfqq);
699 		bfq_activate_bfqq(bfqd, bfqq);
700 	}
701 
702 	if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
703 		bfq_schedule_dispatch(bfqd);
704 	/* release extra ref taken above, bfqq may happen to be freed now */
705 	bfq_put_queue(bfqq);
706 }
707 
708 /**
709  * __bfq_bic_change_cgroup - move @bic to @cgroup.
710  * @bfqd: the queue descriptor.
711  * @bic: the bic to move.
712  * @blkcg: the blk-cgroup to move to.
713  *
714  * Move bic to blkcg, assuming that bfqd->lock is held; which makes
715  * sure that the reference to cgroup is valid across the call (see
716  * comments in bfq_bic_update_cgroup on this issue)
717  */
718 static void *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
719 				     struct bfq_io_cq *bic,
720 				     struct bfq_group *bfqg)
721 {
722 	struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
723 	struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
724 	struct bfq_entity *entity;
725 
726 	if (async_bfqq) {
727 		entity = &async_bfqq->entity;
728 
729 		if (entity->sched_data != &bfqg->sched_data) {
730 			bic_set_bfqq(bic, NULL, 0);
731 			bfq_release_process_ref(bfqd, async_bfqq);
732 		}
733 	}
734 
735 	if (sync_bfqq) {
736 		if (!sync_bfqq->new_bfqq && !bfq_bfqq_coop(sync_bfqq)) {
737 			/* We are the only user of this bfqq, just move it */
738 			if (sync_bfqq->entity.sched_data != &bfqg->sched_data)
739 				bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
740 		} else {
741 			struct bfq_queue *bfqq;
742 
743 			/*
744 			 * The queue was merged to a different queue. Check
745 			 * that the merge chain still belongs to the same
746 			 * cgroup.
747 			 */
748 			for (bfqq = sync_bfqq; bfqq; bfqq = bfqq->new_bfqq)
749 				if (bfqq->entity.sched_data !=
750 				    &bfqg->sched_data)
751 					break;
752 			if (bfqq) {
753 				/*
754 				 * Some queue changed cgroup so the merge is
755 				 * not valid anymore. We cannot easily just
756 				 * cancel the merge (by clearing new_bfqq) as
757 				 * there may be other processes using this
758 				 * queue and holding refs to all queues below
759 				 * sync_bfqq->new_bfqq. Similarly if the merge
760 				 * already happened, we need to detach from
761 				 * bfqq now so that we cannot merge bio to a
762 				 * request from the old cgroup.
763 				 */
764 				bfq_put_cooperator(sync_bfqq);
765 				bfq_release_process_ref(bfqd, sync_bfqq);
766 				bic_set_bfqq(bic, NULL, 1);
767 			}
768 		}
769 	}
770 
771 	return bfqg;
772 }
773 
774 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
775 {
776 	struct bfq_data *bfqd = bic_to_bfqd(bic);
777 	struct bfq_group *bfqg = bfq_bio_bfqg(bfqd, bio);
778 	uint64_t serial_nr;
779 
780 	serial_nr = bfqg_to_blkg(bfqg)->blkcg->css.serial_nr;
781 
782 	/*
783 	 * Check whether blkcg has changed.  The condition may trigger
784 	 * spuriously on a newly created cic but there's no harm.
785 	 */
786 	if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
787 		return;
788 
789 	/*
790 	 * New cgroup for this process. Make sure it is linked to bfq internal
791 	 * cgroup hierarchy.
792 	 */
793 	bfq_link_bfqg(bfqd, bfqg);
794 	__bfq_bic_change_cgroup(bfqd, bic, bfqg);
795 	/*
796 	 * Update blkg_path for bfq_log_* functions. We cache this
797 	 * path, and update it here, for the following
798 	 * reasons. Operations on blkg objects in blk-cgroup are
799 	 * protected with the request_queue lock, and not with the
800 	 * lock that protects the instances of this scheduler
801 	 * (bfqd->lock). This exposes BFQ to the following sort of
802 	 * race.
803 	 *
804 	 * The blkg_lookup performed in bfq_get_queue, protected
805 	 * through rcu, may happen to return the address of a copy of
806 	 * the original blkg. If this is the case, then the
807 	 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
808 	 * the blkg, is useless: it does not prevent blk-cgroup code
809 	 * from destroying both the original blkg and all objects
810 	 * directly or indirectly referred by the copy of the
811 	 * blkg.
812 	 *
813 	 * On the bright side, destroy operations on a blkg invoke, as
814 	 * a first step, hooks of the scheduler associated with the
815 	 * blkg. And these hooks are executed with bfqd->lock held for
816 	 * BFQ. As a consequence, for any blkg associated with the
817 	 * request queue this instance of the scheduler is attached
818 	 * to, we are guaranteed that such a blkg is not destroyed, and
819 	 * that all the pointers it contains are consistent, while we
820 	 * are holding bfqd->lock. A blkg_lookup performed with
821 	 * bfqd->lock held then returns a fully consistent blkg, which
822 	 * remains consistent until this lock is held.
823 	 *
824 	 * Thanks to the last fact, and to the fact that: (1) bfqg has
825 	 * been obtained through a blkg_lookup in the above
826 	 * assignment, and (2) bfqd->lock is being held, here we can
827 	 * safely use the policy data for the involved blkg (i.e., the
828 	 * field bfqg->pd) to get to the blkg associated with bfqg,
829 	 * and then we can safely use any field of blkg. After we
830 	 * release bfqd->lock, even just getting blkg through this
831 	 * bfqg may cause dangling references to be traversed, as
832 	 * bfqg->pd may not exist any more.
833 	 *
834 	 * In view of the above facts, here we cache, in the bfqg, any
835 	 * blkg data we may need for this bic, and for its associated
836 	 * bfq_queue. As of now, we need to cache only the path of the
837 	 * blkg, which is used in the bfq_log_* functions.
838 	 *
839 	 * Finally, note that bfqg itself needs to be protected from
840 	 * destruction on the blkg_free of the original blkg (which
841 	 * invokes bfq_pd_free). We use an additional private
842 	 * refcounter for bfqg, to let it disappear only after no
843 	 * bfq_queue refers to it any longer.
844 	 */
845 	blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
846 	bic->blkcg_serial_nr = serial_nr;
847 }
848 
849 /**
850  * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
851  * @st: the service tree being flushed.
852  */
853 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
854 {
855 	struct bfq_entity *entity = st->first_idle;
856 
857 	for (; entity ; entity = st->first_idle)
858 		__bfq_deactivate_entity(entity, false);
859 }
860 
861 /**
862  * bfq_reparent_leaf_entity - move leaf entity to the root_group.
863  * @bfqd: the device data structure with the root group.
864  * @entity: the entity to move, if entity is a leaf; or the parent entity
865  *	    of an active leaf entity to move, if entity is not a leaf.
866  */
867 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
868 				     struct bfq_entity *entity,
869 				     int ioprio_class)
870 {
871 	struct bfq_queue *bfqq;
872 	struct bfq_entity *child_entity = entity;
873 
874 	while (child_entity->my_sched_data) { /* leaf not reached yet */
875 		struct bfq_sched_data *child_sd = child_entity->my_sched_data;
876 		struct bfq_service_tree *child_st = child_sd->service_tree +
877 			ioprio_class;
878 		struct rb_root *child_active = &child_st->active;
879 
880 		child_entity = bfq_entity_of(rb_first(child_active));
881 
882 		if (!child_entity)
883 			child_entity = child_sd->in_service_entity;
884 	}
885 
886 	bfqq = bfq_entity_to_bfqq(child_entity);
887 	bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
888 }
889 
890 /**
891  * bfq_reparent_active_queues - move to the root group all active queues.
892  * @bfqd: the device data structure with the root group.
893  * @bfqg: the group to move from.
894  * @st: the service tree to start the search from.
895  */
896 static void bfq_reparent_active_queues(struct bfq_data *bfqd,
897 				       struct bfq_group *bfqg,
898 				       struct bfq_service_tree *st,
899 				       int ioprio_class)
900 {
901 	struct rb_root *active = &st->active;
902 	struct bfq_entity *entity;
903 
904 	while ((entity = bfq_entity_of(rb_first(active))))
905 		bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
906 
907 	if (bfqg->sched_data.in_service_entity)
908 		bfq_reparent_leaf_entity(bfqd,
909 					 bfqg->sched_data.in_service_entity,
910 					 ioprio_class);
911 }
912 
913 /**
914  * bfq_pd_offline - deactivate the entity associated with @pd,
915  *		    and reparent its children entities.
916  * @pd: descriptor of the policy going offline.
917  *
918  * blkio already grabs the queue_lock for us, so no need to use
919  * RCU-based magic
920  */
921 static void bfq_pd_offline(struct blkg_policy_data *pd)
922 {
923 	struct bfq_service_tree *st;
924 	struct bfq_group *bfqg = pd_to_bfqg(pd);
925 	struct bfq_data *bfqd = bfqg->bfqd;
926 	struct bfq_entity *entity = bfqg->my_entity;
927 	unsigned long flags;
928 	int i;
929 
930 	spin_lock_irqsave(&bfqd->lock, flags);
931 
932 	if (!entity) /* root group */
933 		goto put_async_queues;
934 
935 	/*
936 	 * Empty all service_trees belonging to this group before
937 	 * deactivating the group itself.
938 	 */
939 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
940 		st = bfqg->sched_data.service_tree + i;
941 
942 		/*
943 		 * It may happen that some queues are still active
944 		 * (busy) upon group destruction (if the corresponding
945 		 * processes have been forced to terminate). We move
946 		 * all the leaf entities corresponding to these queues
947 		 * to the root_group.
948 		 * Also, it may happen that the group has an entity
949 		 * in service, which is disconnected from the active
950 		 * tree: it must be moved, too.
951 		 * There is no need to put the sync queues, as the
952 		 * scheduler has taken no reference.
953 		 */
954 		bfq_reparent_active_queues(bfqd, bfqg, st, i);
955 
956 		/*
957 		 * The idle tree may still contain bfq_queues
958 		 * belonging to exited task because they never
959 		 * migrated to a different cgroup from the one being
960 		 * destroyed now. In addition, even
961 		 * bfq_reparent_active_queues() may happen to add some
962 		 * entities to the idle tree. It happens if, in some
963 		 * of the calls to bfq_bfqq_move() performed by
964 		 * bfq_reparent_active_queues(), the queue to move is
965 		 * empty and gets expired.
966 		 */
967 		bfq_flush_idle_tree(st);
968 	}
969 
970 	__bfq_deactivate_entity(entity, false);
971 
972 put_async_queues:
973 	bfq_put_async_queues(bfqd, bfqg);
974 	bfqg->online = false;
975 
976 	spin_unlock_irqrestore(&bfqd->lock, flags);
977 	/*
978 	 * @blkg is going offline and will be ignored by
979 	 * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
980 	 * that they don't get lost.  If IOs complete after this point, the
981 	 * stats for them will be lost.  Oh well...
982 	 */
983 	bfqg_stats_xfer_dead(bfqg);
984 }
985 
986 void bfq_end_wr_async(struct bfq_data *bfqd)
987 {
988 	struct blkcg_gq *blkg;
989 
990 	list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
991 		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
992 
993 		bfq_end_wr_async_queues(bfqd, bfqg);
994 	}
995 	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
996 }
997 
998 static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
999 {
1000 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1001 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1002 	unsigned int val = 0;
1003 
1004 	if (bfqgd)
1005 		val = bfqgd->weight;
1006 
1007 	seq_printf(sf, "%u\n", val);
1008 
1009 	return 0;
1010 }
1011 
1012 static u64 bfqg_prfill_weight_device(struct seq_file *sf,
1013 				     struct blkg_policy_data *pd, int off)
1014 {
1015 	struct bfq_group *bfqg = pd_to_bfqg(pd);
1016 
1017 	if (!bfqg->entity.dev_weight)
1018 		return 0;
1019 	return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
1020 }
1021 
1022 static int bfq_io_show_weight(struct seq_file *sf, void *v)
1023 {
1024 	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1025 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1026 
1027 	seq_printf(sf, "default %u\n", bfqgd->weight);
1028 	blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
1029 			  &blkcg_policy_bfq, 0, false);
1030 	return 0;
1031 }
1032 
1033 static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
1034 {
1035 	weight = dev_weight ?: weight;
1036 
1037 	bfqg->entity.dev_weight = dev_weight;
1038 	/*
1039 	 * Setting the prio_changed flag of the entity
1040 	 * to 1 with new_weight == weight would re-set
1041 	 * the value of the weight to its ioprio mapping.
1042 	 * Set the flag only if necessary.
1043 	 */
1044 	if ((unsigned short)weight != bfqg->entity.new_weight) {
1045 		bfqg->entity.new_weight = (unsigned short)weight;
1046 		/*
1047 		 * Make sure that the above new value has been
1048 		 * stored in bfqg->entity.new_weight before
1049 		 * setting the prio_changed flag. In fact,
1050 		 * this flag may be read asynchronously (in
1051 		 * critical sections protected by a different
1052 		 * lock than that held here), and finding this
1053 		 * flag set may cause the execution of the code
1054 		 * for updating parameters whose value may
1055 		 * depend also on bfqg->entity.new_weight (in
1056 		 * __bfq_entity_update_weight_prio).
1057 		 * This barrier makes sure that the new value
1058 		 * of bfqg->entity.new_weight is correctly
1059 		 * seen in that code.
1060 		 */
1061 		smp_wmb();
1062 		bfqg->entity.prio_changed = 1;
1063 	}
1064 }
1065 
1066 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1067 				    struct cftype *cftype,
1068 				    u64 val)
1069 {
1070 	struct blkcg *blkcg = css_to_blkcg(css);
1071 	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1072 	struct blkcg_gq *blkg;
1073 	int ret = -ERANGE;
1074 
1075 	if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1076 		return ret;
1077 
1078 	ret = 0;
1079 	spin_lock_irq(&blkcg->lock);
1080 	bfqgd->weight = (unsigned short)val;
1081 	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1082 		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1083 
1084 		if (bfqg)
1085 			bfq_group_set_weight(bfqg, val, 0);
1086 	}
1087 	spin_unlock_irq(&blkcg->lock);
1088 
1089 	return ret;
1090 }
1091 
1092 static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1093 					char *buf, size_t nbytes,
1094 					loff_t off)
1095 {
1096 	int ret;
1097 	struct blkg_conf_ctx ctx;
1098 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1099 	struct bfq_group *bfqg;
1100 	u64 v;
1101 
1102 	ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1103 	if (ret)
1104 		return ret;
1105 
1106 	if (sscanf(ctx.body, "%llu", &v) == 1) {
1107 		/* require "default" on dfl */
1108 		ret = -ERANGE;
1109 		if (!v)
1110 			goto out;
1111 	} else if (!strcmp(strim(ctx.body), "default")) {
1112 		v = 0;
1113 	} else {
1114 		ret = -EINVAL;
1115 		goto out;
1116 	}
1117 
1118 	bfqg = blkg_to_bfqg(ctx.blkg);
1119 
1120 	ret = -ERANGE;
1121 	if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1122 		bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1123 		ret = 0;
1124 	}
1125 out:
1126 	blkg_conf_finish(&ctx);
1127 	return ret ?: nbytes;
1128 }
1129 
1130 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1131 				 char *buf, size_t nbytes,
1132 				 loff_t off)
1133 {
1134 	char *endp;
1135 	int ret;
1136 	u64 v;
1137 
1138 	buf = strim(buf);
1139 
1140 	/* "WEIGHT" or "default WEIGHT" sets the default weight */
1141 	v = simple_strtoull(buf, &endp, 0);
1142 	if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1143 		ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1144 		return ret ?: nbytes;
1145 	}
1146 
1147 	return bfq_io_set_device_weight(of, buf, nbytes, off);
1148 }
1149 
1150 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1151 {
1152 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1153 			  &blkcg_policy_bfq, seq_cft(sf)->private, true);
1154 	return 0;
1155 }
1156 
1157 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1158 					struct blkg_policy_data *pd, int off)
1159 {
1160 	struct blkg_rwstat_sample sum;
1161 
1162 	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1163 	return __blkg_prfill_rwstat(sf, pd, &sum);
1164 }
1165 
1166 static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1167 {
1168 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1169 			  bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1170 			  seq_cft(sf)->private, true);
1171 	return 0;
1172 }
1173 
1174 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1175 static int bfqg_print_stat(struct seq_file *sf, void *v)
1176 {
1177 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1178 			  &blkcg_policy_bfq, seq_cft(sf)->private, false);
1179 	return 0;
1180 }
1181 
1182 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1183 				      struct blkg_policy_data *pd, int off)
1184 {
1185 	struct blkcg_gq *blkg = pd_to_blkg(pd);
1186 	struct blkcg_gq *pos_blkg;
1187 	struct cgroup_subsys_state *pos_css;
1188 	u64 sum = 0;
1189 
1190 	lockdep_assert_held(&blkg->q->queue_lock);
1191 
1192 	rcu_read_lock();
1193 	blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1194 		struct bfq_stat *stat;
1195 
1196 		if (!pos_blkg->online)
1197 			continue;
1198 
1199 		stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1200 		sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1201 	}
1202 	rcu_read_unlock();
1203 
1204 	return __blkg_prfill_u64(sf, pd, sum);
1205 }
1206 
1207 static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1208 {
1209 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1210 			  bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1211 			  seq_cft(sf)->private, false);
1212 	return 0;
1213 }
1214 
1215 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1216 			       int off)
1217 {
1218 	struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1219 	u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
1220 
1221 	return __blkg_prfill_u64(sf, pd, sum >> 9);
1222 }
1223 
1224 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1225 {
1226 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1227 			  bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1228 	return 0;
1229 }
1230 
1231 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1232 					 struct blkg_policy_data *pd, int off)
1233 {
1234 	struct blkg_rwstat_sample tmp;
1235 
1236 	blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1237 			offsetof(struct bfq_group, stats.bytes), &tmp);
1238 
1239 	return __blkg_prfill_u64(sf, pd,
1240 		(tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1241 }
1242 
1243 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1244 {
1245 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1246 			  bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1247 			  false);
1248 	return 0;
1249 }
1250 
1251 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1252 				      struct blkg_policy_data *pd, int off)
1253 {
1254 	struct bfq_group *bfqg = pd_to_bfqg(pd);
1255 	u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1256 	u64 v = 0;
1257 
1258 	if (samples) {
1259 		v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1260 		v = div64_u64(v, samples);
1261 	}
1262 	__blkg_prfill_u64(sf, pd, v);
1263 	return 0;
1264 }
1265 
1266 /* print avg_queue_size */
1267 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1268 {
1269 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1270 			  bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1271 			  0, false);
1272 	return 0;
1273 }
1274 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1275 
1276 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1277 {
1278 	int ret;
1279 
1280 	ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1281 	if (ret)
1282 		return NULL;
1283 
1284 	return blkg_to_bfqg(bfqd->queue->root_blkg);
1285 }
1286 
1287 struct blkcg_policy blkcg_policy_bfq = {
1288 	.dfl_cftypes		= bfq_blkg_files,
1289 	.legacy_cftypes		= bfq_blkcg_legacy_files,
1290 
1291 	.cpd_alloc_fn		= bfq_cpd_alloc,
1292 	.cpd_init_fn		= bfq_cpd_init,
1293 	.cpd_bind_fn	        = bfq_cpd_init,
1294 	.cpd_free_fn		= bfq_cpd_free,
1295 
1296 	.pd_alloc_fn		= bfq_pd_alloc,
1297 	.pd_init_fn		= bfq_pd_init,
1298 	.pd_offline_fn		= bfq_pd_offline,
1299 	.pd_free_fn		= bfq_pd_free,
1300 	.pd_reset_stats_fn	= bfq_pd_reset_stats,
1301 };
1302 
1303 struct cftype bfq_blkcg_legacy_files[] = {
1304 	{
1305 		.name = "bfq.weight",
1306 		.flags = CFTYPE_NOT_ON_ROOT,
1307 		.seq_show = bfq_io_show_weight_legacy,
1308 		.write_u64 = bfq_io_set_weight_legacy,
1309 	},
1310 	{
1311 		.name = "bfq.weight_device",
1312 		.flags = CFTYPE_NOT_ON_ROOT,
1313 		.seq_show = bfq_io_show_weight,
1314 		.write = bfq_io_set_weight,
1315 	},
1316 
1317 	/* statistics, covers only the tasks in the bfqg */
1318 	{
1319 		.name = "bfq.io_service_bytes",
1320 		.private = offsetof(struct bfq_group, stats.bytes),
1321 		.seq_show = bfqg_print_rwstat,
1322 	},
1323 	{
1324 		.name = "bfq.io_serviced",
1325 		.private = offsetof(struct bfq_group, stats.ios),
1326 		.seq_show = bfqg_print_rwstat,
1327 	},
1328 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1329 	{
1330 		.name = "bfq.time",
1331 		.private = offsetof(struct bfq_group, stats.time),
1332 		.seq_show = bfqg_print_stat,
1333 	},
1334 	{
1335 		.name = "bfq.sectors",
1336 		.seq_show = bfqg_print_stat_sectors,
1337 	},
1338 	{
1339 		.name = "bfq.io_service_time",
1340 		.private = offsetof(struct bfq_group, stats.service_time),
1341 		.seq_show = bfqg_print_rwstat,
1342 	},
1343 	{
1344 		.name = "bfq.io_wait_time",
1345 		.private = offsetof(struct bfq_group, stats.wait_time),
1346 		.seq_show = bfqg_print_rwstat,
1347 	},
1348 	{
1349 		.name = "bfq.io_merged",
1350 		.private = offsetof(struct bfq_group, stats.merged),
1351 		.seq_show = bfqg_print_rwstat,
1352 	},
1353 	{
1354 		.name = "bfq.io_queued",
1355 		.private = offsetof(struct bfq_group, stats.queued),
1356 		.seq_show = bfqg_print_rwstat,
1357 	},
1358 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1359 
1360 	/* the same statistics which cover the bfqg and its descendants */
1361 	{
1362 		.name = "bfq.io_service_bytes_recursive",
1363 		.private = offsetof(struct bfq_group, stats.bytes),
1364 		.seq_show = bfqg_print_rwstat_recursive,
1365 	},
1366 	{
1367 		.name = "bfq.io_serviced_recursive",
1368 		.private = offsetof(struct bfq_group, stats.ios),
1369 		.seq_show = bfqg_print_rwstat_recursive,
1370 	},
1371 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1372 	{
1373 		.name = "bfq.time_recursive",
1374 		.private = offsetof(struct bfq_group, stats.time),
1375 		.seq_show = bfqg_print_stat_recursive,
1376 	},
1377 	{
1378 		.name = "bfq.sectors_recursive",
1379 		.seq_show = bfqg_print_stat_sectors_recursive,
1380 	},
1381 	{
1382 		.name = "bfq.io_service_time_recursive",
1383 		.private = offsetof(struct bfq_group, stats.service_time),
1384 		.seq_show = bfqg_print_rwstat_recursive,
1385 	},
1386 	{
1387 		.name = "bfq.io_wait_time_recursive",
1388 		.private = offsetof(struct bfq_group, stats.wait_time),
1389 		.seq_show = bfqg_print_rwstat_recursive,
1390 	},
1391 	{
1392 		.name = "bfq.io_merged_recursive",
1393 		.private = offsetof(struct bfq_group, stats.merged),
1394 		.seq_show = bfqg_print_rwstat_recursive,
1395 	},
1396 	{
1397 		.name = "bfq.io_queued_recursive",
1398 		.private = offsetof(struct bfq_group, stats.queued),
1399 		.seq_show = bfqg_print_rwstat_recursive,
1400 	},
1401 	{
1402 		.name = "bfq.avg_queue_size",
1403 		.seq_show = bfqg_print_avg_queue_size,
1404 	},
1405 	{
1406 		.name = "bfq.group_wait_time",
1407 		.private = offsetof(struct bfq_group, stats.group_wait_time),
1408 		.seq_show = bfqg_print_stat,
1409 	},
1410 	{
1411 		.name = "bfq.idle_time",
1412 		.private = offsetof(struct bfq_group, stats.idle_time),
1413 		.seq_show = bfqg_print_stat,
1414 	},
1415 	{
1416 		.name = "bfq.empty_time",
1417 		.private = offsetof(struct bfq_group, stats.empty_time),
1418 		.seq_show = bfqg_print_stat,
1419 	},
1420 	{
1421 		.name = "bfq.dequeue",
1422 		.private = offsetof(struct bfq_group, stats.dequeue),
1423 		.seq_show = bfqg_print_stat,
1424 	},
1425 #endif	/* CONFIG_BFQ_CGROUP_DEBUG */
1426 	{ }	/* terminate */
1427 };
1428 
1429 struct cftype bfq_blkg_files[] = {
1430 	{
1431 		.name = "bfq.weight",
1432 		.flags = CFTYPE_NOT_ON_ROOT,
1433 		.seq_show = bfq_io_show_weight,
1434 		.write = bfq_io_set_weight,
1435 	},
1436 	{} /* terminate */
1437 };
1438 
1439 #else	/* CONFIG_BFQ_GROUP_IOSCHED */
1440 
1441 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1442 		   struct bfq_group *bfqg) {}
1443 
1444 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1445 {
1446 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1447 
1448 	entity->weight = entity->new_weight;
1449 	entity->orig_weight = entity->new_weight;
1450 	if (bfqq) {
1451 		bfqq->ioprio = bfqq->new_ioprio;
1452 		bfqq->ioprio_class = bfqq->new_ioprio_class;
1453 	}
1454 	entity->sched_data = &bfqg->sched_data;
1455 }
1456 
1457 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1458 
1459 void bfq_end_wr_async(struct bfq_data *bfqd)
1460 {
1461 	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1462 }
1463 
1464 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
1465 {
1466 	return bfqd->root_group;
1467 }
1468 
1469 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1470 {
1471 	return bfqq->bfqd->root_group;
1472 }
1473 
1474 void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1475 
1476 void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1477 
1478 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1479 {
1480 	struct bfq_group *bfqg;
1481 	int i;
1482 
1483 	bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1484 	if (!bfqg)
1485 		return NULL;
1486 
1487 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1488 		bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1489 
1490 	return bfqg;
1491 }
1492 #endif	/* CONFIG_BFQ_GROUP_IOSCHED */
1493