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/ktime.h>
10 #include <linux/rbtree.h>
11 #include <linux/ioprio.h>
12 #include <linux/sbitmap.h>
13 #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, opf, 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, blk_opf_t opf)
233 blkg_rwstat_add(&bfqg->stats.queued, opf, -1);
236 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, blk_opf_t opf)
238 blkg_rwstat_add(&bfqg->stats.merged, opf, 1);
241 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
242 u64 io_start_time_ns, blk_opf_t opf)
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, opf,
249 now - io_start_time_ns);
250 if (io_start_time_ns > start_time_ns)
251 blkg_rwstat_add(&stats->wait_time, opf,
252 io_start_time_ns - start_time_ns);
255 #else /* CONFIG_BFQ_CGROUP_DEBUG */
257 void bfqg_stats_update_io_remove(struct bfq_group *bfqg, blk_opf_t opf) { }
258 void bfqg_stats_update_io_merged(struct bfq_group *bfqg, blk_opf_t opf) { }
259 void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
260 u64 io_start_time_ns, blk_opf_t opf) { }
261 void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
262 void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
264 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
266 #ifdef CONFIG_BFQ_GROUP_IOSCHED
269 * blk-cgroup policy-related handlers
270 * The following functions help in converting between blk-cgroup
271 * internal structures and BFQ-specific structures.
274 static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
276 return pd ? container_of(pd, struct bfq_group, pd) : NULL;
279 struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
281 return pd_to_blkg(&bfqg->pd);
284 static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
286 return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
291 * The following functions help in navigating the bfq_group hierarchy
292 * by allowing to find the parent of a bfq_group or the bfq_group
293 * associated to a bfq_queue.
296 static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
298 struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
300 return pblkg ? blkg_to_bfqg(pblkg) : NULL;
303 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
305 struct bfq_entity *group_entity = bfqq->entity.parent;
307 return group_entity ? container_of(group_entity, struct bfq_group,
309 bfqq->bfqd->root_group;
313 * The following two functions handle get and put of a bfq_group by
314 * wrapping the related blk-cgroup hooks.
317 static void bfqg_get(struct bfq_group *bfqg)
322 static void bfqg_put(struct bfq_group *bfqg)
330 static void bfqg_and_blkg_get(struct bfq_group *bfqg)
332 /* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
335 blkg_get(bfqg_to_blkg(bfqg));
338 void bfqg_and_blkg_put(struct bfq_group *bfqg)
340 blkg_put(bfqg_to_blkg(bfqg));
345 void bfqg_stats_update_legacy_io(struct request_queue *q, struct request *rq)
347 struct bfq_group *bfqg = blkg_to_bfqg(rq->bio->bi_blkg);
352 blkg_rwstat_add(&bfqg->stats.bytes, rq->cmd_flags, blk_rq_bytes(rq));
353 blkg_rwstat_add(&bfqg->stats.ios, rq->cmd_flags, 1);
357 static void bfqg_stats_reset(struct bfqg_stats *stats)
359 #ifdef CONFIG_BFQ_CGROUP_DEBUG
360 /* queued stats shouldn't be cleared */
361 blkg_rwstat_reset(&stats->merged);
362 blkg_rwstat_reset(&stats->service_time);
363 blkg_rwstat_reset(&stats->wait_time);
364 bfq_stat_reset(&stats->time);
365 bfq_stat_reset(&stats->avg_queue_size_sum);
366 bfq_stat_reset(&stats->avg_queue_size_samples);
367 bfq_stat_reset(&stats->dequeue);
368 bfq_stat_reset(&stats->group_wait_time);
369 bfq_stat_reset(&stats->idle_time);
370 bfq_stat_reset(&stats->empty_time);
375 static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
380 #ifdef CONFIG_BFQ_CGROUP_DEBUG
381 /* queued stats shouldn't be cleared */
382 blkg_rwstat_add_aux(&to->merged, &from->merged);
383 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
384 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
385 bfq_stat_add_aux(&from->time, &from->time);
386 bfq_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
387 bfq_stat_add_aux(&to->avg_queue_size_samples,
388 &from->avg_queue_size_samples);
389 bfq_stat_add_aux(&to->dequeue, &from->dequeue);
390 bfq_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
391 bfq_stat_add_aux(&to->idle_time, &from->idle_time);
392 bfq_stat_add_aux(&to->empty_time, &from->empty_time);
397 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
398 * recursive stats can still account for the amount used by this bfqg after
401 static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
403 struct bfq_group *parent;
405 if (!bfqg) /* root_group */
408 parent = bfqg_parent(bfqg);
410 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
412 if (unlikely(!parent))
415 bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
416 bfqg_stats_reset(&bfqg->stats);
419 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
421 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
423 entity->weight = entity->new_weight;
424 entity->orig_weight = entity->new_weight;
426 bfqq->ioprio = bfqq->new_ioprio;
427 bfqq->ioprio_class = bfqq->new_ioprio_class;
429 * Make sure that bfqg and its associated blkg do not
430 * disappear before entity.
432 bfqg_and_blkg_get(bfqg);
434 entity->parent = bfqg->my_entity; /* NULL for root group */
435 entity->sched_data = &bfqg->sched_data;
438 static void bfqg_stats_exit(struct bfqg_stats *stats)
440 blkg_rwstat_exit(&stats->bytes);
441 blkg_rwstat_exit(&stats->ios);
442 #ifdef CONFIG_BFQ_CGROUP_DEBUG
443 blkg_rwstat_exit(&stats->merged);
444 blkg_rwstat_exit(&stats->service_time);
445 blkg_rwstat_exit(&stats->wait_time);
446 blkg_rwstat_exit(&stats->queued);
447 bfq_stat_exit(&stats->time);
448 bfq_stat_exit(&stats->avg_queue_size_sum);
449 bfq_stat_exit(&stats->avg_queue_size_samples);
450 bfq_stat_exit(&stats->dequeue);
451 bfq_stat_exit(&stats->group_wait_time);
452 bfq_stat_exit(&stats->idle_time);
453 bfq_stat_exit(&stats->empty_time);
457 static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
459 if (blkg_rwstat_init(&stats->bytes, gfp) ||
460 blkg_rwstat_init(&stats->ios, gfp))
463 #ifdef CONFIG_BFQ_CGROUP_DEBUG
464 if (blkg_rwstat_init(&stats->merged, gfp) ||
465 blkg_rwstat_init(&stats->service_time, gfp) ||
466 blkg_rwstat_init(&stats->wait_time, gfp) ||
467 blkg_rwstat_init(&stats->queued, gfp) ||
468 bfq_stat_init(&stats->time, gfp) ||
469 bfq_stat_init(&stats->avg_queue_size_sum, gfp) ||
470 bfq_stat_init(&stats->avg_queue_size_samples, gfp) ||
471 bfq_stat_init(&stats->dequeue, gfp) ||
472 bfq_stat_init(&stats->group_wait_time, gfp) ||
473 bfq_stat_init(&stats->idle_time, gfp) ||
474 bfq_stat_init(&stats->empty_time, gfp))
481 bfqg_stats_exit(stats);
485 static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
487 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
490 static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
492 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
495 static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
497 struct bfq_group_data *bgd;
499 bgd = kzalloc(sizeof(*bgd), gfp);
505 static void bfq_cpd_init(struct blkcg_policy_data *cpd)
507 struct bfq_group_data *d = cpd_to_bfqgd(cpd);
509 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
510 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
513 static void bfq_cpd_free(struct blkcg_policy_data *cpd)
515 kfree(cpd_to_bfqgd(cpd));
518 static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
521 struct bfq_group *bfqg;
523 bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
527 if (bfqg_stats_init(&bfqg->stats, gfp)) {
532 /* see comments in bfq_bic_update_cgroup for why refcounting */
537 static void bfq_pd_init(struct blkg_policy_data *pd)
539 struct blkcg_gq *blkg = pd_to_blkg(pd);
540 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
541 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
542 struct bfq_entity *entity = &bfqg->entity;
543 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
545 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
546 entity->my_sched_data = &bfqg->sched_data;
547 entity->last_bfqq_created = NULL;
549 bfqg->my_entity = entity; /*
550 * the root_group's will be set to NULL
551 * in bfq_init_queue()
554 bfqg->active_entities = 0;
556 bfqg->rq_pos_tree = RB_ROOT;
559 static void bfq_pd_free(struct blkg_policy_data *pd)
561 struct bfq_group *bfqg = pd_to_bfqg(pd);
563 bfqg_stats_exit(&bfqg->stats);
567 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
569 struct bfq_group *bfqg = pd_to_bfqg(pd);
571 bfqg_stats_reset(&bfqg->stats);
574 static void bfq_group_set_parent(struct bfq_group *bfqg,
575 struct bfq_group *parent)
577 struct bfq_entity *entity;
579 entity = &bfqg->entity;
580 entity->parent = parent->my_entity;
581 entity->sched_data = &parent->sched_data;
584 static void bfq_link_bfqg(struct bfq_data *bfqd, struct bfq_group *bfqg)
586 struct bfq_group *parent;
587 struct bfq_entity *entity;
590 * Update chain of bfq_groups as we might be handling a leaf group
591 * which, along with some of its relatives, has not been hooked yet
592 * to the private hierarchy of BFQ.
594 entity = &bfqg->entity;
595 for_each_entity(entity) {
596 struct bfq_group *curr_bfqg = container_of(entity,
597 struct bfq_group, entity);
598 if (curr_bfqg != bfqd->root_group) {
599 parent = bfqg_parent(curr_bfqg);
601 parent = bfqd->root_group;
602 bfq_group_set_parent(curr_bfqg, parent);
607 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
609 struct blkcg_gq *blkg = bio->bi_blkg;
610 struct bfq_group *bfqg;
617 bfqg = blkg_to_bfqg(blkg);
619 bio_associate_blkg_from_css(bio, &blkg->blkcg->css);
624 bio_associate_blkg_from_css(bio,
625 &bfqg_to_blkg(bfqd->root_group)->blkcg->css);
626 return bfqd->root_group;
630 * bfq_bfqq_move - migrate @bfqq to @bfqg.
631 * @bfqd: queue descriptor.
632 * @bfqq: the queue to move.
633 * @bfqg: the group to move to.
635 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
636 * it on the new one. Avoid putting the entity on the old group idle tree.
638 * Must be called under the scheduler lock, to make sure that the blkg
639 * owning @bfqg does not disappear (see comments in
640 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
643 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
644 struct bfq_group *bfqg)
646 struct bfq_entity *entity = &bfqq->entity;
647 struct bfq_group *old_parent = bfqq_group(bfqq);
650 * No point to move bfqq to the same group, which can happen when
651 * root group is offlined
653 if (old_parent == bfqg)
657 * oom_bfqq is not allowed to move, oom_bfqq will hold ref to root_group
658 * until elevator exit.
660 if (bfqq == &bfqd->oom_bfqq)
663 * Get extra reference to prevent bfqq from being freed in
664 * next possible expire or deactivate.
668 /* If bfqq is empty, then bfq_bfqq_expire also invokes
669 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
670 * from data structures related to current group. Otherwise we
671 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
674 if (bfqq == bfqd->in_service_queue)
675 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
676 false, BFQQE_PREEMPTED);
678 if (bfq_bfqq_busy(bfqq))
679 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
680 else if (entity->on_st_or_in_serv)
681 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
682 bfqg_and_blkg_put(old_parent);
684 if (entity->parent &&
685 entity->parent->last_bfqq_created == bfqq)
686 entity->parent->last_bfqq_created = NULL;
687 else if (bfqd->last_bfqq_created == bfqq)
688 bfqd->last_bfqq_created = NULL;
690 entity->parent = bfqg->my_entity;
691 entity->sched_data = &bfqg->sched_data;
692 /* pin down bfqg and its associated blkg */
693 bfqg_and_blkg_get(bfqg);
695 if (bfq_bfqq_busy(bfqq)) {
696 if (unlikely(!bfqd->nonrot_with_queueing))
697 bfq_pos_tree_add_move(bfqd, bfqq);
698 bfq_activate_bfqq(bfqd, bfqq);
701 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
702 bfq_schedule_dispatch(bfqd);
703 /* release extra ref taken above, bfqq may happen to be freed now */
708 * __bfq_bic_change_cgroup - move @bic to @bfqg.
709 * @bfqd: the queue descriptor.
710 * @bic: the bic to move.
711 * @bfqg: the group to move to.
713 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
714 * sure that the reference to cgroup is valid across the call (see
715 * comments in bfq_bic_update_cgroup on this issue)
717 static void *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
718 struct bfq_io_cq *bic,
719 struct bfq_group *bfqg)
721 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, false);
722 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, true);
723 struct bfq_entity *entity;
726 entity = &async_bfqq->entity;
728 if (entity->sched_data != &bfqg->sched_data) {
729 bic_set_bfqq(bic, NULL, false);
730 bfq_release_process_ref(bfqd, async_bfqq);
735 if (!sync_bfqq->new_bfqq && !bfq_bfqq_coop(sync_bfqq)) {
736 /* We are the only user of this bfqq, just move it */
737 if (sync_bfqq->entity.sched_data != &bfqg->sched_data)
738 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
740 struct bfq_queue *bfqq;
743 * The queue was merged to a different queue. Check
744 * that the merge chain still belongs to the same
747 for (bfqq = sync_bfqq; bfqq; bfqq = bfqq->new_bfqq)
748 if (bfqq->entity.sched_data !=
753 * Some queue changed cgroup so the merge is
754 * not valid anymore. We cannot easily just
755 * cancel the merge (by clearing new_bfqq) as
756 * there may be other processes using this
757 * queue and holding refs to all queues below
758 * sync_bfqq->new_bfqq. Similarly if the merge
759 * already happened, we need to detach from
760 * bfqq now so that we cannot merge bio to a
761 * request from the old cgroup.
763 bfq_put_cooperator(sync_bfqq);
764 bfq_release_process_ref(bfqd, sync_bfqq);
765 bic_set_bfqq(bic, NULL, true);
773 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
775 struct bfq_data *bfqd = bic_to_bfqd(bic);
776 struct bfq_group *bfqg = bfq_bio_bfqg(bfqd, bio);
779 serial_nr = bfqg_to_blkg(bfqg)->blkcg->css.serial_nr;
782 * Check whether blkcg has changed. The condition may trigger
783 * spuriously on a newly created cic but there's no harm.
785 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
789 * New cgroup for this process. Make sure it is linked to bfq internal
792 bfq_link_bfqg(bfqd, bfqg);
793 __bfq_bic_change_cgroup(bfqd, bic, bfqg);
795 * Update blkg_path for bfq_log_* functions. We cache this
796 * path, and update it here, for the following
797 * reasons. Operations on blkg objects in blk-cgroup are
798 * protected with the request_queue lock, and not with the
799 * lock that protects the instances of this scheduler
800 * (bfqd->lock). This exposes BFQ to the following sort of
803 * The blkg_lookup performed in bfq_get_queue, protected
804 * through rcu, may happen to return the address of a copy of
805 * the original blkg. If this is the case, then the
806 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
807 * the blkg, is useless: it does not prevent blk-cgroup code
808 * from destroying both the original blkg and all objects
809 * directly or indirectly referred by the copy of the
812 * On the bright side, destroy operations on a blkg invoke, as
813 * a first step, hooks of the scheduler associated with the
814 * blkg. And these hooks are executed with bfqd->lock held for
815 * BFQ. As a consequence, for any blkg associated with the
816 * request queue this instance of the scheduler is attached
817 * to, we are guaranteed that such a blkg is not destroyed, and
818 * that all the pointers it contains are consistent, while we
819 * are holding bfqd->lock. A blkg_lookup performed with
820 * bfqd->lock held then returns a fully consistent blkg, which
821 * remains consistent until this lock is held.
823 * Thanks to the last fact, and to the fact that: (1) bfqg has
824 * been obtained through a blkg_lookup in the above
825 * assignment, and (2) bfqd->lock is being held, here we can
826 * safely use the policy data for the involved blkg (i.e., the
827 * field bfqg->pd) to get to the blkg associated with bfqg,
828 * and then we can safely use any field of blkg. After we
829 * release bfqd->lock, even just getting blkg through this
830 * bfqg may cause dangling references to be traversed, as
831 * bfqg->pd may not exist any more.
833 * In view of the above facts, here we cache, in the bfqg, any
834 * blkg data we may need for this bic, and for its associated
835 * bfq_queue. As of now, we need to cache only the path of the
836 * blkg, which is used in the bfq_log_* functions.
838 * Finally, note that bfqg itself needs to be protected from
839 * destruction on the blkg_free of the original blkg (which
840 * invokes bfq_pd_free). We use an additional private
841 * refcounter for bfqg, to let it disappear only after no
842 * bfq_queue refers to it any longer.
844 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
845 bic->blkcg_serial_nr = serial_nr;
849 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
850 * @st: the service tree being flushed.
852 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
854 struct bfq_entity *entity = st->first_idle;
856 for (; entity ; entity = st->first_idle)
857 __bfq_deactivate_entity(entity, false);
861 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
862 * @bfqd: the device data structure with the root group.
863 * @entity: the entity to move, if entity is a leaf; or the parent entity
864 * of an active leaf entity to move, if entity is not a leaf.
865 * @ioprio_class: I/O priority class to reparent.
867 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
868 struct bfq_entity *entity,
871 struct bfq_queue *bfqq;
872 struct bfq_entity *child_entity = entity;
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 +
878 struct rb_root *child_active = &child_st->active;
880 child_entity = bfq_entity_of(rb_first(child_active));
883 child_entity = child_sd->in_service_entity;
886 bfqq = bfq_entity_to_bfqq(child_entity);
887 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
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 * @ioprio_class: I/O priority class to reparent.
897 static void bfq_reparent_active_queues(struct bfq_data *bfqd,
898 struct bfq_group *bfqg,
899 struct bfq_service_tree *st,
902 struct rb_root *active = &st->active;
903 struct bfq_entity *entity;
905 while ((entity = bfq_entity_of(rb_first(active))))
906 bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
908 if (bfqg->sched_data.in_service_entity)
909 bfq_reparent_leaf_entity(bfqd,
910 bfqg->sched_data.in_service_entity,
915 * bfq_pd_offline - deactivate the entity associated with @pd,
916 * and reparent its children entities.
917 * @pd: descriptor of the policy going offline.
919 * blkio already grabs the queue_lock for us, so no need to use
922 static void bfq_pd_offline(struct blkg_policy_data *pd)
924 struct bfq_service_tree *st;
925 struct bfq_group *bfqg = pd_to_bfqg(pd);
926 struct bfq_data *bfqd = bfqg->bfqd;
927 struct bfq_entity *entity = bfqg->my_entity;
931 spin_lock_irqsave(&bfqd->lock, flags);
933 if (!entity) /* root group */
934 goto put_async_queues;
937 * Empty all service_trees belonging to this group before
938 * deactivating the group itself.
940 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
941 st = bfqg->sched_data.service_tree + i;
944 * It may happen that some queues are still active
945 * (busy) upon group destruction (if the corresponding
946 * processes have been forced to terminate). We move
947 * all the leaf entities corresponding to these queues
949 * Also, it may happen that the group has an entity
950 * in service, which is disconnected from the active
951 * tree: it must be moved, too.
952 * There is no need to put the sync queues, as the
953 * scheduler has taken no reference.
955 bfq_reparent_active_queues(bfqd, bfqg, st, i);
958 * The idle tree may still contain bfq_queues
959 * belonging to exited task because they never
960 * migrated to a different cgroup from the one being
961 * destroyed now. In addition, even
962 * bfq_reparent_active_queues() may happen to add some
963 * entities to the idle tree. It happens if, in some
964 * of the calls to bfq_bfqq_move() performed by
965 * bfq_reparent_active_queues(), the queue to move is
966 * empty and gets expired.
968 bfq_flush_idle_tree(st);
971 __bfq_deactivate_entity(entity, false);
974 bfq_put_async_queues(bfqd, bfqg);
975 bfqg->online = false;
977 spin_unlock_irqrestore(&bfqd->lock, flags);
979 * @blkg is going offline and will be ignored by
980 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
981 * that they don't get lost. If IOs complete after this point, the
982 * stats for them will be lost. Oh well...
984 bfqg_stats_xfer_dead(bfqg);
987 void bfq_end_wr_async(struct bfq_data *bfqd)
989 struct blkcg_gq *blkg;
991 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
992 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
994 bfq_end_wr_async_queues(bfqd, bfqg);
996 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
999 static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
1001 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1002 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1003 unsigned int val = 0;
1006 val = bfqgd->weight;
1008 seq_printf(sf, "%u\n", val);
1013 static u64 bfqg_prfill_weight_device(struct seq_file *sf,
1014 struct blkg_policy_data *pd, int off)
1016 struct bfq_group *bfqg = pd_to_bfqg(pd);
1018 if (!bfqg->entity.dev_weight)
1020 return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
1023 static int bfq_io_show_weight(struct seq_file *sf, void *v)
1025 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1026 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1028 seq_printf(sf, "default %u\n", bfqgd->weight);
1029 blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
1030 &blkcg_policy_bfq, 0, false);
1034 static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
1036 weight = dev_weight ?: weight;
1038 bfqg->entity.dev_weight = dev_weight;
1040 * Setting the prio_changed flag of the entity
1041 * to 1 with new_weight == weight would re-set
1042 * the value of the weight to its ioprio mapping.
1043 * Set the flag only if necessary.
1045 if ((unsigned short)weight != bfqg->entity.new_weight) {
1046 bfqg->entity.new_weight = (unsigned short)weight;
1048 * Make sure that the above new value has been
1049 * stored in bfqg->entity.new_weight before
1050 * setting the prio_changed flag. In fact,
1051 * this flag may be read asynchronously (in
1052 * critical sections protected by a different
1053 * lock than that held here), and finding this
1054 * flag set may cause the execution of the code
1055 * for updating parameters whose value may
1056 * depend also on bfqg->entity.new_weight (in
1057 * __bfq_entity_update_weight_prio).
1058 * This barrier makes sure that the new value
1059 * of bfqg->entity.new_weight is correctly
1060 * seen in that code.
1063 bfqg->entity.prio_changed = 1;
1067 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1068 struct cftype *cftype,
1071 struct blkcg *blkcg = css_to_blkcg(css);
1072 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1073 struct blkcg_gq *blkg;
1076 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1080 spin_lock_irq(&blkcg->lock);
1081 bfqgd->weight = (unsigned short)val;
1082 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1083 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1086 bfq_group_set_weight(bfqg, val, 0);
1088 spin_unlock_irq(&blkcg->lock);
1093 static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1094 char *buf, size_t nbytes,
1098 struct blkg_conf_ctx ctx;
1099 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1100 struct bfq_group *bfqg;
1103 ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1107 if (sscanf(ctx.body, "%llu", &v) == 1) {
1108 /* require "default" on dfl */
1112 } else if (!strcmp(strim(ctx.body), "default")) {
1119 bfqg = blkg_to_bfqg(ctx.blkg);
1122 if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1123 bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1127 blkg_conf_finish(&ctx);
1128 return ret ?: nbytes;
1131 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1132 char *buf, size_t nbytes,
1141 /* "WEIGHT" or "default WEIGHT" sets the default weight */
1142 v = simple_strtoull(buf, &endp, 0);
1143 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1144 ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1145 return ret ?: nbytes;
1148 return bfq_io_set_device_weight(of, buf, nbytes, off);
1151 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1153 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1154 &blkcg_policy_bfq, seq_cft(sf)->private, true);
1158 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1159 struct blkg_policy_data *pd, int off)
1161 struct blkg_rwstat_sample sum;
1163 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1164 return __blkg_prfill_rwstat(sf, pd, &sum);
1167 static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1169 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1170 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1171 seq_cft(sf)->private, true);
1175 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1176 static int bfqg_print_stat(struct seq_file *sf, void *v)
1178 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1179 &blkcg_policy_bfq, seq_cft(sf)->private, false);
1183 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1184 struct blkg_policy_data *pd, int off)
1186 struct blkcg_gq *blkg = pd_to_blkg(pd);
1187 struct blkcg_gq *pos_blkg;
1188 struct cgroup_subsys_state *pos_css;
1191 lockdep_assert_held(&blkg->q->queue_lock);
1194 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1195 struct bfq_stat *stat;
1197 if (!pos_blkg->online)
1200 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1201 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1205 return __blkg_prfill_u64(sf, pd, sum);
1208 static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1210 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1211 bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1212 seq_cft(sf)->private, false);
1216 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1219 struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1220 u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
1222 return __blkg_prfill_u64(sf, pd, sum >> 9);
1225 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1227 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1228 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1232 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1233 struct blkg_policy_data *pd, int off)
1235 struct blkg_rwstat_sample tmp;
1237 blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1238 offsetof(struct bfq_group, stats.bytes), &tmp);
1240 return __blkg_prfill_u64(sf, pd,
1241 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1244 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1246 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1247 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1252 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1253 struct blkg_policy_data *pd, int off)
1255 struct bfq_group *bfqg = pd_to_bfqg(pd);
1256 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1260 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1261 v = div64_u64(v, samples);
1263 __blkg_prfill_u64(sf, pd, v);
1267 /* print avg_queue_size */
1268 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1270 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1271 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1275 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1277 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1281 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1285 return blkg_to_bfqg(bfqd->queue->root_blkg);
1288 struct blkcg_policy blkcg_policy_bfq = {
1289 .dfl_cftypes = bfq_blkg_files,
1290 .legacy_cftypes = bfq_blkcg_legacy_files,
1292 .cpd_alloc_fn = bfq_cpd_alloc,
1293 .cpd_init_fn = bfq_cpd_init,
1294 .cpd_bind_fn = bfq_cpd_init,
1295 .cpd_free_fn = bfq_cpd_free,
1297 .pd_alloc_fn = bfq_pd_alloc,
1298 .pd_init_fn = bfq_pd_init,
1299 .pd_offline_fn = bfq_pd_offline,
1300 .pd_free_fn = bfq_pd_free,
1301 .pd_reset_stats_fn = bfq_pd_reset_stats,
1304 struct cftype bfq_blkcg_legacy_files[] = {
1306 .name = "bfq.weight",
1307 .flags = CFTYPE_NOT_ON_ROOT,
1308 .seq_show = bfq_io_show_weight_legacy,
1309 .write_u64 = bfq_io_set_weight_legacy,
1312 .name = "bfq.weight_device",
1313 .flags = CFTYPE_NOT_ON_ROOT,
1314 .seq_show = bfq_io_show_weight,
1315 .write = bfq_io_set_weight,
1318 /* statistics, covers only the tasks in the bfqg */
1320 .name = "bfq.io_service_bytes",
1321 .private = offsetof(struct bfq_group, stats.bytes),
1322 .seq_show = bfqg_print_rwstat,
1325 .name = "bfq.io_serviced",
1326 .private = offsetof(struct bfq_group, stats.ios),
1327 .seq_show = bfqg_print_rwstat,
1329 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1332 .private = offsetof(struct bfq_group, stats.time),
1333 .seq_show = bfqg_print_stat,
1336 .name = "bfq.sectors",
1337 .seq_show = bfqg_print_stat_sectors,
1340 .name = "bfq.io_service_time",
1341 .private = offsetof(struct bfq_group, stats.service_time),
1342 .seq_show = bfqg_print_rwstat,
1345 .name = "bfq.io_wait_time",
1346 .private = offsetof(struct bfq_group, stats.wait_time),
1347 .seq_show = bfqg_print_rwstat,
1350 .name = "bfq.io_merged",
1351 .private = offsetof(struct bfq_group, stats.merged),
1352 .seq_show = bfqg_print_rwstat,
1355 .name = "bfq.io_queued",
1356 .private = offsetof(struct bfq_group, stats.queued),
1357 .seq_show = bfqg_print_rwstat,
1359 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1361 /* the same statistics which cover the bfqg and its descendants */
1363 .name = "bfq.io_service_bytes_recursive",
1364 .private = offsetof(struct bfq_group, stats.bytes),
1365 .seq_show = bfqg_print_rwstat_recursive,
1368 .name = "bfq.io_serviced_recursive",
1369 .private = offsetof(struct bfq_group, stats.ios),
1370 .seq_show = bfqg_print_rwstat_recursive,
1372 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1374 .name = "bfq.time_recursive",
1375 .private = offsetof(struct bfq_group, stats.time),
1376 .seq_show = bfqg_print_stat_recursive,
1379 .name = "bfq.sectors_recursive",
1380 .seq_show = bfqg_print_stat_sectors_recursive,
1383 .name = "bfq.io_service_time_recursive",
1384 .private = offsetof(struct bfq_group, stats.service_time),
1385 .seq_show = bfqg_print_rwstat_recursive,
1388 .name = "bfq.io_wait_time_recursive",
1389 .private = offsetof(struct bfq_group, stats.wait_time),
1390 .seq_show = bfqg_print_rwstat_recursive,
1393 .name = "bfq.io_merged_recursive",
1394 .private = offsetof(struct bfq_group, stats.merged),
1395 .seq_show = bfqg_print_rwstat_recursive,
1398 .name = "bfq.io_queued_recursive",
1399 .private = offsetof(struct bfq_group, stats.queued),
1400 .seq_show = bfqg_print_rwstat_recursive,
1403 .name = "bfq.avg_queue_size",
1404 .seq_show = bfqg_print_avg_queue_size,
1407 .name = "bfq.group_wait_time",
1408 .private = offsetof(struct bfq_group, stats.group_wait_time),
1409 .seq_show = bfqg_print_stat,
1412 .name = "bfq.idle_time",
1413 .private = offsetof(struct bfq_group, stats.idle_time),
1414 .seq_show = bfqg_print_stat,
1417 .name = "bfq.empty_time",
1418 .private = offsetof(struct bfq_group, stats.empty_time),
1419 .seq_show = bfqg_print_stat,
1422 .name = "bfq.dequeue",
1423 .private = offsetof(struct bfq_group, stats.dequeue),
1424 .seq_show = bfqg_print_stat,
1426 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1430 struct cftype bfq_blkg_files[] = {
1432 .name = "bfq.weight",
1433 .flags = CFTYPE_NOT_ON_ROOT,
1434 .seq_show = bfq_io_show_weight,
1435 .write = bfq_io_set_weight,
1440 #else /* CONFIG_BFQ_GROUP_IOSCHED */
1442 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1443 struct bfq_group *bfqg) {}
1445 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1447 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1449 entity->weight = entity->new_weight;
1450 entity->orig_weight = entity->new_weight;
1452 bfqq->ioprio = bfqq->new_ioprio;
1453 bfqq->ioprio_class = bfqq->new_ioprio_class;
1455 entity->sched_data = &bfqg->sched_data;
1458 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1460 void bfq_end_wr_async(struct bfq_data *bfqd)
1462 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1465 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
1467 return bfqd->root_group;
1470 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1472 return bfqq->bfqd->root_group;
1475 void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1477 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1479 struct bfq_group *bfqg;
1482 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1486 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1487 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1491 #endif /* CONFIG_BFQ_GROUP_IOSCHED */