1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * cgroups support for the BFQ I/O scheduler.
5 #include <linux/module.h>
6 #include <linux/slab.h>
7 #include <linux/blkdev.h>
8 #include <linux/cgroup.h>
9 #include <linux/elevator.h>
10 #include <linux/ktime.h>
11 #include <linux/rbtree.h>
12 #include <linux/ioprio.h>
13 #include <linux/sbitmap.h>
14 #include <linux/delay.h>
16 #include "bfq-iosched.h"
18 #ifdef CONFIG_BFQ_CGROUP_DEBUG
19 static int bfq_stat_init(struct bfq_stat *stat, gfp_t gfp)
23 ret = percpu_counter_init(&stat->cpu_cnt, 0, gfp);
27 atomic64_set(&stat->aux_cnt, 0);
31 static void bfq_stat_exit(struct bfq_stat *stat)
33 percpu_counter_destroy(&stat->cpu_cnt);
37 * bfq_stat_add - add a value to a bfq_stat
38 * @stat: target bfq_stat
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.
44 static inline void bfq_stat_add(struct bfq_stat *stat, uint64_t val)
46 percpu_counter_add_batch(&stat->cpu_cnt, val, BLKG_STAT_CPU_BATCH);
50 * bfq_stat_read - read the current value of a bfq_stat
51 * @stat: bfq_stat to read
53 static inline uint64_t bfq_stat_read(struct bfq_stat *stat)
55 return percpu_counter_sum_positive(&stat->cpu_cnt);
59 * bfq_stat_reset - reset a bfq_stat
60 * @stat: bfq_stat to reset
62 static inline void bfq_stat_reset(struct bfq_stat *stat)
64 percpu_counter_set(&stat->cpu_cnt, 0);
65 atomic64_set(&stat->aux_cnt, 0);
69 * bfq_stat_add_aux - add a bfq_stat into another's aux count
70 * @to: the destination bfq_stat
73 * Add @from's count including the aux one to @to's aux count.
75 static inline void bfq_stat_add_aux(struct bfq_stat *to,
76 struct bfq_stat *from)
78 atomic64_add(bfq_stat_read(from) + atomic64_read(&from->aux_cnt),
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
88 * prfill callback for printing a bfq_stat.
90 static u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd,
93 return __blkg_prfill_u64(sf, pd, bfq_stat_read((void *)pd + off));
96 /* bfqg stats flags */
97 enum bfqg_stats_flags {
98 BFQG_stats_waiting = 0,
103 #define BFQG_FLAG_FNS(name) \
104 static void bfqg_stats_mark_##name(struct bfqg_stats *stats) \
106 stats->flags |= (1 << BFQG_stats_##name); \
108 static void bfqg_stats_clear_##name(struct bfqg_stats *stats) \
110 stats->flags &= ~(1 << BFQG_stats_##name); \
112 static int bfqg_stats_##name(struct bfqg_stats *stats) \
114 return (stats->flags & (1 << BFQG_stats_##name)) != 0; \
117 BFQG_FLAG_FNS(waiting)
118 BFQG_FLAG_FNS(idling)
122 /* This should be called with the scheduler lock held. */
123 static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
127 if (!bfqg_stats_waiting(stats))
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);
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)
141 struct bfqg_stats *stats = &bfqg->stats;
143 if (bfqg_stats_waiting(stats))
145 if (bfqg == curr_bfqg)
147 stats->start_group_wait_time = ktime_get_ns();
148 bfqg_stats_mark_waiting(stats);
151 /* This should be called with the scheduler lock held. */
152 static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
156 if (!bfqg_stats_empty(stats))
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);
166 void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
168 bfq_stat_add(&bfqg->stats.dequeue, 1);
171 void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)
173 struct bfqg_stats *stats = &bfqg->stats;
175 if (blkg_rwstat_total(&stats->queued))
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.
183 if (bfqg_stats_empty(stats))
186 stats->start_empty_time = ktime_get_ns();
187 bfqg_stats_mark_empty(stats);
190 void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
192 struct bfqg_stats *stats = &bfqg->stats;
194 if (bfqg_stats_idling(stats)) {
195 u64 now = ktime_get_ns();
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);
204 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
206 struct bfqg_stats *stats = &bfqg->stats;
208 stats->start_idle_time = ktime_get_ns();
209 bfqg_stats_mark_idling(stats);
212 void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
214 struct bfqg_stats *stats = &bfqg->stats;
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);
222 void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
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));
231 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
233 blkg_rwstat_add(&bfqg->stats.queued, op, -1);
236 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
238 blkg_rwstat_add(&bfqg->stats.merged, op, 1);
241 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
242 u64 io_start_time_ns, unsigned int op)
244 struct bfqg_stats *stats = &bfqg->stats;
245 u64 now = ktime_get_ns();
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);
255 #else /* CONFIG_BFQ_CGROUP_DEBUG */
257 void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
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) { }
269 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
271 #ifdef CONFIG_BFQ_GROUP_IOSCHED
274 * blk-cgroup policy-related handlers
275 * The following functions help in converting between blk-cgroup
276 * internal structures and BFQ-specific structures.
279 static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
281 return pd ? container_of(pd, struct bfq_group, pd) : NULL;
284 struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
286 return pd_to_blkg(&bfqg->pd);
289 static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
291 return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
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.
301 static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
303 struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
305 return pblkg ? blkg_to_bfqg(pblkg) : NULL;
308 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
310 struct bfq_entity *group_entity = bfqq->entity.parent;
312 return group_entity ? container_of(group_entity, struct bfq_group,
314 bfqq->bfqd->root_group;
318 * The following two functions handle get and put of a bfq_group by
319 * wrapping the related blk-cgroup hooks.
322 static void bfqg_get(struct bfq_group *bfqg)
327 static void bfqg_put(struct bfq_group *bfqg)
335 static void bfqg_and_blkg_get(struct bfq_group *bfqg)
337 /* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
340 blkg_get(bfqg_to_blkg(bfqg));
343 void bfqg_and_blkg_put(struct bfq_group *bfqg)
345 blkg_put(bfqg_to_blkg(bfqg));
350 void bfqg_stats_update_legacy_io(struct request_queue *q, struct request *rq)
352 struct bfq_group *bfqg = blkg_to_bfqg(rq->bio->bi_blkg);
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);
362 static void bfqg_stats_reset(struct bfqg_stats *stats)
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);
380 static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
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);
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
406 static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
408 struct bfq_group *parent;
410 if (!bfqg) /* root_group */
413 parent = bfqg_parent(bfqg);
415 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
417 if (unlikely(!parent))
420 bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
421 bfqg_stats_reset(&bfqg->stats);
424 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
426 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
428 entity->weight = entity->new_weight;
429 entity->orig_weight = entity->new_weight;
431 bfqq->ioprio = bfqq->new_ioprio;
432 bfqq->ioprio_class = bfqq->new_ioprio_class;
434 * Make sure that bfqg and its associated blkg do not
435 * disappear before entity.
437 bfqg_and_blkg_get(bfqg);
439 entity->parent = bfqg->my_entity; /* NULL for root group */
440 entity->sched_data = &bfqg->sched_data;
443 static void bfqg_stats_exit(struct bfqg_stats *stats)
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);
462 static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
464 if (blkg_rwstat_init(&stats->bytes, gfp) ||
465 blkg_rwstat_init(&stats->ios, gfp))
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 bfqg_stats_exit(stats);
488 static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
490 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
493 static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
495 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
498 static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
500 struct bfq_group_data *bgd;
502 bgd = kzalloc(sizeof(*bgd), gfp);
508 static void bfq_cpd_init(struct blkcg_policy_data *cpd)
510 struct bfq_group_data *d = cpd_to_bfqgd(cpd);
512 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
513 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
516 static void bfq_cpd_free(struct blkcg_policy_data *cpd)
518 kfree(cpd_to_bfqgd(cpd));
521 static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
524 struct bfq_group *bfqg;
526 bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
530 if (bfqg_stats_init(&bfqg->stats, gfp)) {
535 /* see comments in bfq_bic_update_cgroup for why refcounting */
540 static void bfq_pd_init(struct blkg_policy_data *pd)
542 struct blkcg_gq *blkg = pd_to_blkg(pd);
543 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
544 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
545 struct bfq_entity *entity = &bfqg->entity;
546 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
548 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
549 entity->my_sched_data = &bfqg->sched_data;
550 entity->last_bfqq_created = NULL;
552 bfqg->my_entity = entity; /*
553 * the root_group's will be set to NULL
554 * in bfq_init_queue()
557 bfqg->active_entities = 0;
558 bfqg->rq_pos_tree = RB_ROOT;
561 static void bfq_pd_free(struct blkg_policy_data *pd)
563 struct bfq_group *bfqg = pd_to_bfqg(pd);
565 bfqg_stats_exit(&bfqg->stats);
569 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
571 struct bfq_group *bfqg = pd_to_bfqg(pd);
573 bfqg_stats_reset(&bfqg->stats);
576 static void bfq_group_set_parent(struct bfq_group *bfqg,
577 struct bfq_group *parent)
579 struct bfq_entity *entity;
581 entity = &bfqg->entity;
582 entity->parent = parent->my_entity;
583 entity->sched_data = &parent->sched_data;
586 static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
589 struct blkcg_gq *blkg;
591 blkg = blkg_lookup(blkcg, bfqd->queue);
593 return blkg_to_bfqg(blkg);
597 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
600 struct bfq_group *bfqg, *parent;
601 struct bfq_entity *entity;
603 bfqg = bfq_lookup_bfqg(bfqd, blkcg);
609 * Update chain of bfq_groups as we might be handling a leaf group
610 * which, along with some of its relatives, has not been hooked yet
611 * to the private hierarchy of BFQ.
613 entity = &bfqg->entity;
614 for_each_entity(entity) {
615 struct bfq_group *curr_bfqg = container_of(entity,
616 struct bfq_group, entity);
617 if (curr_bfqg != bfqd->root_group) {
618 parent = bfqg_parent(curr_bfqg);
620 parent = bfqd->root_group;
621 bfq_group_set_parent(curr_bfqg, parent);
629 * bfq_bfqq_move - migrate @bfqq to @bfqg.
630 * @bfqd: queue descriptor.
631 * @bfqq: the queue to move.
632 * @bfqg: the group to move to.
634 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
635 * it on the new one. Avoid putting the entity on the old group idle tree.
637 * Must be called under the scheduler lock, to make sure that the blkg
638 * owning @bfqg does not disappear (see comments in
639 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
642 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
643 struct bfq_group *bfqg)
645 struct bfq_entity *entity = &bfqq->entity;
648 * Get extra reference to prevent bfqq from being freed in
649 * next possible expire or deactivate.
653 /* If bfqq is empty, then bfq_bfqq_expire also invokes
654 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
655 * from data structures related to current group. Otherwise we
656 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
659 if (bfqq == bfqd->in_service_queue)
660 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
661 false, BFQQE_PREEMPTED);
663 if (bfq_bfqq_busy(bfqq))
664 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
665 else if (entity->on_st_or_in_serv)
666 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
667 bfqg_and_blkg_put(bfqq_group(bfqq));
669 if (entity->parent &&
670 entity->parent->last_bfqq_created == bfqq)
671 entity->parent->last_bfqq_created = NULL;
672 else if (bfqd->last_bfqq_created == bfqq)
673 bfqd->last_bfqq_created = NULL;
675 entity->parent = bfqg->my_entity;
676 entity->sched_data = &bfqg->sched_data;
677 /* pin down bfqg and its associated blkg */
678 bfqg_and_blkg_get(bfqg);
680 if (bfq_bfqq_busy(bfqq)) {
681 if (unlikely(!bfqd->nonrot_with_queueing))
682 bfq_pos_tree_add_move(bfqd, bfqq);
683 bfq_activate_bfqq(bfqd, bfqq);
686 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
687 bfq_schedule_dispatch(bfqd);
688 /* release extra ref taken above, bfqq may happen to be freed now */
693 * __bfq_bic_change_cgroup - move @bic to @cgroup.
694 * @bfqd: the queue descriptor.
695 * @bic: the bic to move.
696 * @blkcg: the blk-cgroup to move to.
698 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
699 * sure that the reference to cgroup is valid across the call (see
700 * comments in bfq_bic_update_cgroup on this issue)
702 * NOTE: an alternative approach might have been to store the current
703 * cgroup in bfqq and getting a reference to it, reducing the lookup
704 * time here, at the price of slightly more complex code.
706 static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
707 struct bfq_io_cq *bic,
710 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
711 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
712 struct bfq_group *bfqg;
713 struct bfq_entity *entity;
715 bfqg = bfq_find_set_group(bfqd, blkcg);
718 bfqg = bfqd->root_group;
721 entity = &async_bfqq->entity;
723 if (entity->sched_data != &bfqg->sched_data) {
724 bic_set_bfqq(bic, NULL, 0);
725 bfq_release_process_ref(bfqd, async_bfqq);
730 entity = &sync_bfqq->entity;
731 if (entity->sched_data != &bfqg->sched_data)
732 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
738 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
740 struct bfq_data *bfqd = bic_to_bfqd(bic);
741 struct bfq_group *bfqg = NULL;
745 serial_nr = __bio_blkcg(bio)->css.serial_nr;
748 * Check whether blkcg has changed. The condition may trigger
749 * spuriously on a newly created cic but there's no harm.
751 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
754 bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
756 * Update blkg_path for bfq_log_* functions. We cache this
757 * path, and update it here, for the following
758 * reasons. Operations on blkg objects in blk-cgroup are
759 * protected with the request_queue lock, and not with the
760 * lock that protects the instances of this scheduler
761 * (bfqd->lock). This exposes BFQ to the following sort of
764 * The blkg_lookup performed in bfq_get_queue, protected
765 * through rcu, may happen to return the address of a copy of
766 * the original blkg. If this is the case, then the
767 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
768 * the blkg, is useless: it does not prevent blk-cgroup code
769 * from destroying both the original blkg and all objects
770 * directly or indirectly referred by the copy of the
773 * On the bright side, destroy operations on a blkg invoke, as
774 * a first step, hooks of the scheduler associated with the
775 * blkg. And these hooks are executed with bfqd->lock held for
776 * BFQ. As a consequence, for any blkg associated with the
777 * request queue this instance of the scheduler is attached
778 * to, we are guaranteed that such a blkg is not destroyed, and
779 * that all the pointers it contains are consistent, while we
780 * are holding bfqd->lock. A blkg_lookup performed with
781 * bfqd->lock held then returns a fully consistent blkg, which
782 * remains consistent until this lock is held.
784 * Thanks to the last fact, and to the fact that: (1) bfqg has
785 * been obtained through a blkg_lookup in the above
786 * assignment, and (2) bfqd->lock is being held, here we can
787 * safely use the policy data for the involved blkg (i.e., the
788 * field bfqg->pd) to get to the blkg associated with bfqg,
789 * and then we can safely use any field of blkg. After we
790 * release bfqd->lock, even just getting blkg through this
791 * bfqg may cause dangling references to be traversed, as
792 * bfqg->pd may not exist any more.
794 * In view of the above facts, here we cache, in the bfqg, any
795 * blkg data we may need for this bic, and for its associated
796 * bfq_queue. As of now, we need to cache only the path of the
797 * blkg, which is used in the bfq_log_* functions.
799 * Finally, note that bfqg itself needs to be protected from
800 * destruction on the blkg_free of the original blkg (which
801 * invokes bfq_pd_free). We use an additional private
802 * refcounter for bfqg, to let it disappear only after no
803 * bfq_queue refers to it any longer.
805 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
806 bic->blkcg_serial_nr = serial_nr;
812 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
813 * @st: the service tree being flushed.
815 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
817 struct bfq_entity *entity = st->first_idle;
819 for (; entity ; entity = st->first_idle)
820 __bfq_deactivate_entity(entity, false);
824 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
825 * @bfqd: the device data structure with the root group.
826 * @entity: the entity to move, if entity is a leaf; or the parent entity
827 * of an active leaf entity to move, if entity is not a leaf.
829 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
830 struct bfq_entity *entity,
833 struct bfq_queue *bfqq;
834 struct bfq_entity *child_entity = entity;
836 while (child_entity->my_sched_data) { /* leaf not reached yet */
837 struct bfq_sched_data *child_sd = child_entity->my_sched_data;
838 struct bfq_service_tree *child_st = child_sd->service_tree +
840 struct rb_root *child_active = &child_st->active;
842 child_entity = bfq_entity_of(rb_first(child_active));
845 child_entity = child_sd->in_service_entity;
848 bfqq = bfq_entity_to_bfqq(child_entity);
849 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
853 * bfq_reparent_active_queues - move to the root group all active queues.
854 * @bfqd: the device data structure with the root group.
855 * @bfqg: the group to move from.
856 * @st: the service tree to start the search from.
858 static void bfq_reparent_active_queues(struct bfq_data *bfqd,
859 struct bfq_group *bfqg,
860 struct bfq_service_tree *st,
863 struct rb_root *active = &st->active;
864 struct bfq_entity *entity;
866 while ((entity = bfq_entity_of(rb_first(active))))
867 bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
869 if (bfqg->sched_data.in_service_entity)
870 bfq_reparent_leaf_entity(bfqd,
871 bfqg->sched_data.in_service_entity,
876 * bfq_pd_offline - deactivate the entity associated with @pd,
877 * and reparent its children entities.
878 * @pd: descriptor of the policy going offline.
880 * blkio already grabs the queue_lock for us, so no need to use
883 static void bfq_pd_offline(struct blkg_policy_data *pd)
885 struct bfq_service_tree *st;
886 struct bfq_group *bfqg = pd_to_bfqg(pd);
887 struct bfq_data *bfqd = bfqg->bfqd;
888 struct bfq_entity *entity = bfqg->my_entity;
892 spin_lock_irqsave(&bfqd->lock, flags);
894 if (!entity) /* root group */
895 goto put_async_queues;
898 * Empty all service_trees belonging to this group before
899 * deactivating the group itself.
901 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
902 st = bfqg->sched_data.service_tree + i;
905 * It may happen that some queues are still active
906 * (busy) upon group destruction (if the corresponding
907 * processes have been forced to terminate). We move
908 * all the leaf entities corresponding to these queues
910 * Also, it may happen that the group has an entity
911 * in service, which is disconnected from the active
912 * tree: it must be moved, too.
913 * There is no need to put the sync queues, as the
914 * scheduler has taken no reference.
916 bfq_reparent_active_queues(bfqd, bfqg, st, i);
919 * The idle tree may still contain bfq_queues
920 * belonging to exited task because they never
921 * migrated to a different cgroup from the one being
922 * destroyed now. In addition, even
923 * bfq_reparent_active_queues() may happen to add some
924 * entities to the idle tree. It happens if, in some
925 * of the calls to bfq_bfqq_move() performed by
926 * bfq_reparent_active_queues(), the queue to move is
927 * empty and gets expired.
929 bfq_flush_idle_tree(st);
932 __bfq_deactivate_entity(entity, false);
935 bfq_put_async_queues(bfqd, bfqg);
937 spin_unlock_irqrestore(&bfqd->lock, flags);
939 * @blkg is going offline and will be ignored by
940 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
941 * that they don't get lost. If IOs complete after this point, the
942 * stats for them will be lost. Oh well...
944 bfqg_stats_xfer_dead(bfqg);
947 void bfq_end_wr_async(struct bfq_data *bfqd)
949 struct blkcg_gq *blkg;
951 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
952 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
954 bfq_end_wr_async_queues(bfqd, bfqg);
956 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
959 static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
961 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
962 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
963 unsigned int val = 0;
968 seq_printf(sf, "%u\n", val);
973 static u64 bfqg_prfill_weight_device(struct seq_file *sf,
974 struct blkg_policy_data *pd, int off)
976 struct bfq_group *bfqg = pd_to_bfqg(pd);
978 if (!bfqg->entity.dev_weight)
980 return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
983 static int bfq_io_show_weight(struct seq_file *sf, void *v)
985 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
986 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
988 seq_printf(sf, "default %u\n", bfqgd->weight);
989 blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
990 &blkcg_policy_bfq, 0, false);
994 static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
996 weight = dev_weight ?: weight;
998 bfqg->entity.dev_weight = dev_weight;
1000 * Setting the prio_changed flag of the entity
1001 * to 1 with new_weight == weight would re-set
1002 * the value of the weight to its ioprio mapping.
1003 * Set the flag only if necessary.
1005 if ((unsigned short)weight != bfqg->entity.new_weight) {
1006 bfqg->entity.new_weight = (unsigned short)weight;
1008 * Make sure that the above new value has been
1009 * stored in bfqg->entity.new_weight before
1010 * setting the prio_changed flag. In fact,
1011 * this flag may be read asynchronously (in
1012 * critical sections protected by a different
1013 * lock than that held here), and finding this
1014 * flag set may cause the execution of the code
1015 * for updating parameters whose value may
1016 * depend also on bfqg->entity.new_weight (in
1017 * __bfq_entity_update_weight_prio).
1018 * This barrier makes sure that the new value
1019 * of bfqg->entity.new_weight is correctly
1020 * seen in that code.
1023 bfqg->entity.prio_changed = 1;
1027 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1028 struct cftype *cftype,
1031 struct blkcg *blkcg = css_to_blkcg(css);
1032 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1033 struct blkcg_gq *blkg;
1036 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1040 spin_lock_irq(&blkcg->lock);
1041 bfqgd->weight = (unsigned short)val;
1042 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1043 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1046 bfq_group_set_weight(bfqg, val, 0);
1048 spin_unlock_irq(&blkcg->lock);
1053 static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1054 char *buf, size_t nbytes,
1058 struct blkg_conf_ctx ctx;
1059 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1060 struct bfq_group *bfqg;
1063 ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1067 if (sscanf(ctx.body, "%llu", &v) == 1) {
1068 /* require "default" on dfl */
1072 } else if (!strcmp(strim(ctx.body), "default")) {
1079 bfqg = blkg_to_bfqg(ctx.blkg);
1082 if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1083 bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1087 blkg_conf_finish(&ctx);
1088 return ret ?: nbytes;
1091 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1092 char *buf, size_t nbytes,
1101 /* "WEIGHT" or "default WEIGHT" sets the default weight */
1102 v = simple_strtoull(buf, &endp, 0);
1103 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1104 ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1105 return ret ?: nbytes;
1108 return bfq_io_set_device_weight(of, buf, nbytes, off);
1111 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1113 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1114 &blkcg_policy_bfq, seq_cft(sf)->private, true);
1118 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1119 struct blkg_policy_data *pd, int off)
1121 struct blkg_rwstat_sample sum;
1123 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1124 return __blkg_prfill_rwstat(sf, pd, &sum);
1127 static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1129 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1130 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1131 seq_cft(sf)->private, true);
1135 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1136 static int bfqg_print_stat(struct seq_file *sf, void *v)
1138 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1139 &blkcg_policy_bfq, seq_cft(sf)->private, false);
1143 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1144 struct blkg_policy_data *pd, int off)
1146 struct blkcg_gq *blkg = pd_to_blkg(pd);
1147 struct blkcg_gq *pos_blkg;
1148 struct cgroup_subsys_state *pos_css;
1151 lockdep_assert_held(&blkg->q->queue_lock);
1154 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1155 struct bfq_stat *stat;
1157 if (!pos_blkg->online)
1160 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1161 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1165 return __blkg_prfill_u64(sf, pd, sum);
1168 static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1170 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1171 bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1172 seq_cft(sf)->private, false);
1176 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1179 struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1180 u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
1182 return __blkg_prfill_u64(sf, pd, sum >> 9);
1185 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1187 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1188 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1192 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1193 struct blkg_policy_data *pd, int off)
1195 struct blkg_rwstat_sample tmp;
1197 blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1198 offsetof(struct bfq_group, stats.bytes), &tmp);
1200 return __blkg_prfill_u64(sf, pd,
1201 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1204 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1206 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1207 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1212 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1213 struct blkg_policy_data *pd, int off)
1215 struct bfq_group *bfqg = pd_to_bfqg(pd);
1216 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1220 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1221 v = div64_u64(v, samples);
1223 __blkg_prfill_u64(sf, pd, v);
1227 /* print avg_queue_size */
1228 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1230 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1231 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1235 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1237 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1241 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1245 return blkg_to_bfqg(bfqd->queue->root_blkg);
1248 struct blkcg_policy blkcg_policy_bfq = {
1249 .dfl_cftypes = bfq_blkg_files,
1250 .legacy_cftypes = bfq_blkcg_legacy_files,
1252 .cpd_alloc_fn = bfq_cpd_alloc,
1253 .cpd_init_fn = bfq_cpd_init,
1254 .cpd_bind_fn = bfq_cpd_init,
1255 .cpd_free_fn = bfq_cpd_free,
1257 .pd_alloc_fn = bfq_pd_alloc,
1258 .pd_init_fn = bfq_pd_init,
1259 .pd_offline_fn = bfq_pd_offline,
1260 .pd_free_fn = bfq_pd_free,
1261 .pd_reset_stats_fn = bfq_pd_reset_stats,
1264 struct cftype bfq_blkcg_legacy_files[] = {
1266 .name = "bfq.weight",
1267 .flags = CFTYPE_NOT_ON_ROOT,
1268 .seq_show = bfq_io_show_weight_legacy,
1269 .write_u64 = bfq_io_set_weight_legacy,
1272 .name = "bfq.weight_device",
1273 .flags = CFTYPE_NOT_ON_ROOT,
1274 .seq_show = bfq_io_show_weight,
1275 .write = bfq_io_set_weight,
1278 /* statistics, covers only the tasks in the bfqg */
1280 .name = "bfq.io_service_bytes",
1281 .private = offsetof(struct bfq_group, stats.bytes),
1282 .seq_show = bfqg_print_rwstat,
1285 .name = "bfq.io_serviced",
1286 .private = offsetof(struct bfq_group, stats.ios),
1287 .seq_show = bfqg_print_rwstat,
1289 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1292 .private = offsetof(struct bfq_group, stats.time),
1293 .seq_show = bfqg_print_stat,
1296 .name = "bfq.sectors",
1297 .seq_show = bfqg_print_stat_sectors,
1300 .name = "bfq.io_service_time",
1301 .private = offsetof(struct bfq_group, stats.service_time),
1302 .seq_show = bfqg_print_rwstat,
1305 .name = "bfq.io_wait_time",
1306 .private = offsetof(struct bfq_group, stats.wait_time),
1307 .seq_show = bfqg_print_rwstat,
1310 .name = "bfq.io_merged",
1311 .private = offsetof(struct bfq_group, stats.merged),
1312 .seq_show = bfqg_print_rwstat,
1315 .name = "bfq.io_queued",
1316 .private = offsetof(struct bfq_group, stats.queued),
1317 .seq_show = bfqg_print_rwstat,
1319 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1321 /* the same statistics which cover the bfqg and its descendants */
1323 .name = "bfq.io_service_bytes_recursive",
1324 .private = offsetof(struct bfq_group, stats.bytes),
1325 .seq_show = bfqg_print_rwstat_recursive,
1328 .name = "bfq.io_serviced_recursive",
1329 .private = offsetof(struct bfq_group, stats.ios),
1330 .seq_show = bfqg_print_rwstat_recursive,
1332 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1334 .name = "bfq.time_recursive",
1335 .private = offsetof(struct bfq_group, stats.time),
1336 .seq_show = bfqg_print_stat_recursive,
1339 .name = "bfq.sectors_recursive",
1340 .seq_show = bfqg_print_stat_sectors_recursive,
1343 .name = "bfq.io_service_time_recursive",
1344 .private = offsetof(struct bfq_group, stats.service_time),
1345 .seq_show = bfqg_print_rwstat_recursive,
1348 .name = "bfq.io_wait_time_recursive",
1349 .private = offsetof(struct bfq_group, stats.wait_time),
1350 .seq_show = bfqg_print_rwstat_recursive,
1353 .name = "bfq.io_merged_recursive",
1354 .private = offsetof(struct bfq_group, stats.merged),
1355 .seq_show = bfqg_print_rwstat_recursive,
1358 .name = "bfq.io_queued_recursive",
1359 .private = offsetof(struct bfq_group, stats.queued),
1360 .seq_show = bfqg_print_rwstat_recursive,
1363 .name = "bfq.avg_queue_size",
1364 .seq_show = bfqg_print_avg_queue_size,
1367 .name = "bfq.group_wait_time",
1368 .private = offsetof(struct bfq_group, stats.group_wait_time),
1369 .seq_show = bfqg_print_stat,
1372 .name = "bfq.idle_time",
1373 .private = offsetof(struct bfq_group, stats.idle_time),
1374 .seq_show = bfqg_print_stat,
1377 .name = "bfq.empty_time",
1378 .private = offsetof(struct bfq_group, stats.empty_time),
1379 .seq_show = bfqg_print_stat,
1382 .name = "bfq.dequeue",
1383 .private = offsetof(struct bfq_group, stats.dequeue),
1384 .seq_show = bfqg_print_stat,
1386 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1390 struct cftype bfq_blkg_files[] = {
1392 .name = "bfq.weight",
1393 .flags = CFTYPE_NOT_ON_ROOT,
1394 .seq_show = bfq_io_show_weight,
1395 .write = bfq_io_set_weight,
1400 #else /* CONFIG_BFQ_GROUP_IOSCHED */
1402 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1403 struct bfq_group *bfqg) {}
1405 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1407 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1409 entity->weight = entity->new_weight;
1410 entity->orig_weight = entity->new_weight;
1412 bfqq->ioprio = bfqq->new_ioprio;
1413 bfqq->ioprio_class = bfqq->new_ioprio_class;
1415 entity->sched_data = &bfqg->sched_data;
1418 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1420 void bfq_end_wr_async(struct bfq_data *bfqd)
1422 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1425 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1427 return bfqd->root_group;
1430 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1432 return bfqq->bfqd->root_group;
1435 void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1437 void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1439 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1441 struct bfq_group *bfqg;
1444 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1448 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1449 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1453 #endif /* CONFIG_BFQ_GROUP_IOSCHED */