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));
351 static void bfqg_stats_reset(struct bfqg_stats *stats)
353 #ifdef CONFIG_BFQ_CGROUP_DEBUG
354 /* queued stats shouldn't be cleared */
355 blkg_rwstat_reset(&stats->merged);
356 blkg_rwstat_reset(&stats->service_time);
357 blkg_rwstat_reset(&stats->wait_time);
358 bfq_stat_reset(&stats->time);
359 bfq_stat_reset(&stats->avg_queue_size_sum);
360 bfq_stat_reset(&stats->avg_queue_size_samples);
361 bfq_stat_reset(&stats->dequeue);
362 bfq_stat_reset(&stats->group_wait_time);
363 bfq_stat_reset(&stats->idle_time);
364 bfq_stat_reset(&stats->empty_time);
369 static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
374 #ifdef CONFIG_BFQ_CGROUP_DEBUG
375 /* queued stats shouldn't be cleared */
376 blkg_rwstat_add_aux(&to->merged, &from->merged);
377 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
378 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
379 bfq_stat_add_aux(&from->time, &from->time);
380 bfq_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
381 bfq_stat_add_aux(&to->avg_queue_size_samples,
382 &from->avg_queue_size_samples);
383 bfq_stat_add_aux(&to->dequeue, &from->dequeue);
384 bfq_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
385 bfq_stat_add_aux(&to->idle_time, &from->idle_time);
386 bfq_stat_add_aux(&to->empty_time, &from->empty_time);
391 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
392 * recursive stats can still account for the amount used by this bfqg after
395 static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
397 struct bfq_group *parent;
399 if (!bfqg) /* root_group */
402 parent = bfqg_parent(bfqg);
404 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
406 if (unlikely(!parent))
409 bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
410 bfqg_stats_reset(&bfqg->stats);
413 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
415 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
417 entity->weight = entity->new_weight;
418 entity->orig_weight = entity->new_weight;
420 bfqq->ioprio = bfqq->new_ioprio;
421 bfqq->ioprio_class = bfqq->new_ioprio_class;
423 * Make sure that bfqg and its associated blkg do not
424 * disappear before entity.
426 bfqg_and_blkg_get(bfqg);
428 entity->parent = bfqg->my_entity; /* NULL for root group */
429 entity->sched_data = &bfqg->sched_data;
432 static void bfqg_stats_exit(struct bfqg_stats *stats)
434 #ifdef CONFIG_BFQ_CGROUP_DEBUG
435 blkg_rwstat_exit(&stats->merged);
436 blkg_rwstat_exit(&stats->service_time);
437 blkg_rwstat_exit(&stats->wait_time);
438 blkg_rwstat_exit(&stats->queued);
439 bfq_stat_exit(&stats->time);
440 bfq_stat_exit(&stats->avg_queue_size_sum);
441 bfq_stat_exit(&stats->avg_queue_size_samples);
442 bfq_stat_exit(&stats->dequeue);
443 bfq_stat_exit(&stats->group_wait_time);
444 bfq_stat_exit(&stats->idle_time);
445 bfq_stat_exit(&stats->empty_time);
449 static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
451 #ifdef CONFIG_BFQ_CGROUP_DEBUG
452 if (blkg_rwstat_init(&stats->merged, gfp) ||
453 blkg_rwstat_init(&stats->service_time, gfp) ||
454 blkg_rwstat_init(&stats->wait_time, gfp) ||
455 blkg_rwstat_init(&stats->queued, gfp) ||
456 bfq_stat_init(&stats->time, gfp) ||
457 bfq_stat_init(&stats->avg_queue_size_sum, gfp) ||
458 bfq_stat_init(&stats->avg_queue_size_samples, gfp) ||
459 bfq_stat_init(&stats->dequeue, gfp) ||
460 bfq_stat_init(&stats->group_wait_time, gfp) ||
461 bfq_stat_init(&stats->idle_time, gfp) ||
462 bfq_stat_init(&stats->empty_time, gfp)) {
463 bfqg_stats_exit(stats);
471 static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
473 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
476 static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
478 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
481 static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
483 struct bfq_group_data *bgd;
485 bgd = kzalloc(sizeof(*bgd), gfp);
491 static void bfq_cpd_init(struct blkcg_policy_data *cpd)
493 struct bfq_group_data *d = cpd_to_bfqgd(cpd);
495 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
496 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
499 static void bfq_cpd_free(struct blkcg_policy_data *cpd)
501 kfree(cpd_to_bfqgd(cpd));
504 static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
507 struct bfq_group *bfqg;
509 bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
513 if (bfqg_stats_init(&bfqg->stats, gfp)) {
518 /* see comments in bfq_bic_update_cgroup for why refcounting */
523 static void bfq_pd_init(struct blkg_policy_data *pd)
525 struct blkcg_gq *blkg = pd_to_blkg(pd);
526 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
527 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
528 struct bfq_entity *entity = &bfqg->entity;
529 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
531 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
532 entity->my_sched_data = &bfqg->sched_data;
533 bfqg->my_entity = entity; /*
534 * the root_group's will be set to NULL
535 * in bfq_init_queue()
538 bfqg->active_entities = 0;
540 bfqg->rq_pos_tree = RB_ROOT;
543 static void bfq_pd_free(struct blkg_policy_data *pd)
545 struct bfq_group *bfqg = pd_to_bfqg(pd);
547 bfqg_stats_exit(&bfqg->stats);
551 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
553 struct bfq_group *bfqg = pd_to_bfqg(pd);
555 bfqg_stats_reset(&bfqg->stats);
558 static void bfq_group_set_parent(struct bfq_group *bfqg,
559 struct bfq_group *parent)
561 struct bfq_entity *entity;
563 entity = &bfqg->entity;
564 entity->parent = parent->my_entity;
565 entity->sched_data = &parent->sched_data;
568 static void bfq_link_bfqg(struct bfq_data *bfqd, struct bfq_group *bfqg)
570 struct bfq_group *parent;
571 struct bfq_entity *entity;
574 * Update chain of bfq_groups as we might be handling a leaf group
575 * which, along with some of its relatives, has not been hooked yet
576 * to the private hierarchy of BFQ.
578 entity = &bfqg->entity;
579 for_each_entity(entity) {
580 struct bfq_group *curr_bfqg = container_of(entity,
581 struct bfq_group, entity);
582 if (curr_bfqg != bfqd->root_group) {
583 parent = bfqg_parent(curr_bfqg);
585 parent = bfqd->root_group;
586 bfq_group_set_parent(curr_bfqg, parent);
591 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
593 struct blkcg_gq *blkg = bio->bi_blkg;
594 struct bfq_group *bfqg;
601 bfqg = blkg_to_bfqg(blkg);
603 bio_associate_blkg_from_css(bio, &blkg->blkcg->css);
608 bio_associate_blkg_from_css(bio,
609 &bfqg_to_blkg(bfqd->root_group)->blkcg->css);
610 return bfqd->root_group;
614 * bfq_bfqq_move - migrate @bfqq to @bfqg.
615 * @bfqd: queue descriptor.
616 * @bfqq: the queue to move.
617 * @bfqg: the group to move to.
619 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
620 * it on the new one. Avoid putting the entity on the old group idle tree.
622 * Must be called under the scheduler lock, to make sure that the blkg
623 * owning @bfqg does not disappear (see comments in
624 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
627 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
628 struct bfq_group *bfqg)
630 struct bfq_entity *entity = &bfqq->entity;
633 * oom_bfqq is not allowed to move, oom_bfqq will hold ref to root_group
634 * until elevator exit.
636 if (bfqq == &bfqd->oom_bfqq)
639 * Get extra reference to prevent bfqq from being freed in
640 * next possible expire or deactivate.
644 /* If bfqq is empty, then bfq_bfqq_expire also invokes
645 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
646 * from data structures related to current group. Otherwise we
647 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
650 if (bfqq == bfqd->in_service_queue)
651 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
652 false, BFQQE_PREEMPTED);
654 if (bfq_bfqq_busy(bfqq))
655 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
656 else if (entity->on_st)
657 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
658 bfqg_and_blkg_put(bfqq_group(bfqq));
660 entity->parent = bfqg->my_entity;
661 entity->sched_data = &bfqg->sched_data;
662 /* pin down bfqg and its associated blkg */
663 bfqg_and_blkg_get(bfqg);
665 if (bfq_bfqq_busy(bfqq)) {
666 if (unlikely(!bfqd->nonrot_with_queueing))
667 bfq_pos_tree_add_move(bfqd, bfqq);
668 bfq_activate_bfqq(bfqd, bfqq);
671 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
672 bfq_schedule_dispatch(bfqd);
673 /* release extra ref taken above, bfqq may happen to be freed now */
678 * __bfq_bic_change_cgroup - move @bic to @cgroup.
679 * @bfqd: the queue descriptor.
680 * @bic: the bic to move.
681 * @blkcg: the blk-cgroup to move to.
683 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
684 * sure that the reference to cgroup is valid across the call (see
685 * comments in bfq_bic_update_cgroup on this issue)
687 static void *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
688 struct bfq_io_cq *bic,
689 struct bfq_group *bfqg)
691 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
692 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
693 struct bfq_entity *entity;
696 entity = &async_bfqq->entity;
698 if (entity->sched_data != &bfqg->sched_data) {
699 bic_set_bfqq(bic, NULL, 0);
700 bfq_release_process_ref(bfqd, async_bfqq);
705 if (!sync_bfqq->new_bfqq && !bfq_bfqq_coop(sync_bfqq)) {
706 /* We are the only user of this bfqq, just move it */
707 if (sync_bfqq->entity.sched_data != &bfqg->sched_data)
708 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
710 struct bfq_queue *bfqq;
713 * The queue was merged to a different queue. Check
714 * that the merge chain still belongs to the same
717 for (bfqq = sync_bfqq; bfqq; bfqq = bfqq->new_bfqq)
718 if (bfqq->entity.sched_data !=
723 * Some queue changed cgroup so the merge is
724 * not valid anymore. We cannot easily just
725 * cancel the merge (by clearing new_bfqq) as
726 * there may be other processes using this
727 * queue and holding refs to all queues below
728 * sync_bfqq->new_bfqq. Similarly if the merge
729 * already happened, we need to detach from
730 * bfqq now so that we cannot merge bio to a
731 * request from the old cgroup.
733 bfq_put_cooperator(sync_bfqq);
734 bfq_release_process_ref(bfqd, sync_bfqq);
735 bic_set_bfqq(bic, NULL, 1);
743 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
745 struct bfq_data *bfqd = bic_to_bfqd(bic);
746 struct bfq_group *bfqg = bfq_bio_bfqg(bfqd, bio);
749 serial_nr = bfqg_to_blkg(bfqg)->blkcg->css.serial_nr;
752 * Check whether blkcg has changed. The condition may trigger
753 * spuriously on a newly created cic but there's no harm.
755 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
759 * New cgroup for this process. Make sure it is linked to bfq internal
762 bfq_link_bfqg(bfqd, bfqg);
763 __bfq_bic_change_cgroup(bfqd, bic, bfqg);
765 * Update blkg_path for bfq_log_* functions. We cache this
766 * path, and update it here, for the following
767 * reasons. Operations on blkg objects in blk-cgroup are
768 * protected with the request_queue lock, and not with the
769 * lock that protects the instances of this scheduler
770 * (bfqd->lock). This exposes BFQ to the following sort of
773 * The blkg_lookup performed in bfq_get_queue, protected
774 * through rcu, may happen to return the address of a copy of
775 * the original blkg. If this is the case, then the
776 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
777 * the blkg, is useless: it does not prevent blk-cgroup code
778 * from destroying both the original blkg and all objects
779 * directly or indirectly referred by the copy of the
782 * On the bright side, destroy operations on a blkg invoke, as
783 * a first step, hooks of the scheduler associated with the
784 * blkg. And these hooks are executed with bfqd->lock held for
785 * BFQ. As a consequence, for any blkg associated with the
786 * request queue this instance of the scheduler is attached
787 * to, we are guaranteed that such a blkg is not destroyed, and
788 * that all the pointers it contains are consistent, while we
789 * are holding bfqd->lock. A blkg_lookup performed with
790 * bfqd->lock held then returns a fully consistent blkg, which
791 * remains consistent until this lock is held.
793 * Thanks to the last fact, and to the fact that: (1) bfqg has
794 * been obtained through a blkg_lookup in the above
795 * assignment, and (2) bfqd->lock is being held, here we can
796 * safely use the policy data for the involved blkg (i.e., the
797 * field bfqg->pd) to get to the blkg associated with bfqg,
798 * and then we can safely use any field of blkg. After we
799 * release bfqd->lock, even just getting blkg through this
800 * bfqg may cause dangling references to be traversed, as
801 * bfqg->pd may not exist any more.
803 * In view of the above facts, here we cache, in the bfqg, any
804 * blkg data we may need for this bic, and for its associated
805 * bfq_queue. As of now, we need to cache only the path of the
806 * blkg, which is used in the bfq_log_* functions.
808 * Finally, note that bfqg itself needs to be protected from
809 * destruction on the blkg_free of the original blkg (which
810 * invokes bfq_pd_free). We use an additional private
811 * refcounter for bfqg, to let it disappear only after no
812 * bfq_queue refers to it any longer.
814 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
815 bic->blkcg_serial_nr = serial_nr;
819 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
820 * @st: the service tree being flushed.
822 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
824 struct bfq_entity *entity = st->first_idle;
826 for (; entity ; entity = st->first_idle)
827 __bfq_deactivate_entity(entity, false);
831 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
832 * @bfqd: the device data structure with the root group.
833 * @entity: the entity to move, if entity is a leaf; or the parent entity
834 * of an active leaf entity to move, if entity is not a leaf.
836 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
837 struct bfq_entity *entity,
840 struct bfq_queue *bfqq;
841 struct bfq_entity *child_entity = entity;
843 while (child_entity->my_sched_data) { /* leaf not reached yet */
844 struct bfq_sched_data *child_sd = child_entity->my_sched_data;
845 struct bfq_service_tree *child_st = child_sd->service_tree +
847 struct rb_root *child_active = &child_st->active;
849 child_entity = bfq_entity_of(rb_first(child_active));
852 child_entity = child_sd->in_service_entity;
855 bfqq = bfq_entity_to_bfqq(child_entity);
856 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
860 * bfq_reparent_active_queues - move to the root group all active queues.
861 * @bfqd: the device data structure with the root group.
862 * @bfqg: the group to move from.
863 * @st: the service tree to start the search from.
865 static void bfq_reparent_active_queues(struct bfq_data *bfqd,
866 struct bfq_group *bfqg,
867 struct bfq_service_tree *st,
870 struct rb_root *active = &st->active;
871 struct bfq_entity *entity;
873 while ((entity = bfq_entity_of(rb_first(active))))
874 bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
876 if (bfqg->sched_data.in_service_entity)
877 bfq_reparent_leaf_entity(bfqd,
878 bfqg->sched_data.in_service_entity,
883 * bfq_pd_offline - deactivate the entity associated with @pd,
884 * and reparent its children entities.
885 * @pd: descriptor of the policy going offline.
887 * blkio already grabs the queue_lock for us, so no need to use
890 static void bfq_pd_offline(struct blkg_policy_data *pd)
892 struct bfq_service_tree *st;
893 struct bfq_group *bfqg = pd_to_bfqg(pd);
894 struct bfq_data *bfqd = bfqg->bfqd;
895 struct bfq_entity *entity = bfqg->my_entity;
899 spin_lock_irqsave(&bfqd->lock, flags);
901 if (!entity) /* root group */
902 goto put_async_queues;
905 * Empty all service_trees belonging to this group before
906 * deactivating the group itself.
908 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
909 st = bfqg->sched_data.service_tree + i;
912 * It may happen that some queues are still active
913 * (busy) upon group destruction (if the corresponding
914 * processes have been forced to terminate). We move
915 * all the leaf entities corresponding to these queues
917 * Also, it may happen that the group has an entity
918 * in service, which is disconnected from the active
919 * tree: it must be moved, too.
920 * There is no need to put the sync queues, as the
921 * scheduler has taken no reference.
923 bfq_reparent_active_queues(bfqd, bfqg, st, i);
926 * The idle tree may still contain bfq_queues
927 * belonging to exited task because they never
928 * migrated to a different cgroup from the one being
929 * destroyed now. In addition, even
930 * bfq_reparent_active_queues() may happen to add some
931 * entities to the idle tree. It happens if, in some
932 * of the calls to bfq_bfqq_move() performed by
933 * bfq_reparent_active_queues(), the queue to move is
934 * empty and gets expired.
936 bfq_flush_idle_tree(st);
939 __bfq_deactivate_entity(entity, false);
942 bfq_put_async_queues(bfqd, bfqg);
943 bfqg->online = false;
945 spin_unlock_irqrestore(&bfqd->lock, flags);
947 * @blkg is going offline and will be ignored by
948 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
949 * that they don't get lost. If IOs complete after this point, the
950 * stats for them will be lost. Oh well...
952 bfqg_stats_xfer_dead(bfqg);
955 void bfq_end_wr_async(struct bfq_data *bfqd)
957 struct blkcg_gq *blkg;
959 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
960 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
962 bfq_end_wr_async_queues(bfqd, bfqg);
964 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
967 static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
969 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
970 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
971 unsigned int val = 0;
976 seq_printf(sf, "%u\n", val);
981 static u64 bfqg_prfill_weight_device(struct seq_file *sf,
982 struct blkg_policy_data *pd, int off)
984 struct bfq_group *bfqg = pd_to_bfqg(pd);
986 if (!bfqg->entity.dev_weight)
988 return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
991 static int bfq_io_show_weight(struct seq_file *sf, void *v)
993 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
994 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
996 seq_printf(sf, "default %u\n", bfqgd->weight);
997 blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
998 &blkcg_policy_bfq, 0, false);
1002 static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
1004 weight = dev_weight ?: weight;
1006 bfqg->entity.dev_weight = dev_weight;
1008 * Setting the prio_changed flag of the entity
1009 * to 1 with new_weight == weight would re-set
1010 * the value of the weight to its ioprio mapping.
1011 * Set the flag only if necessary.
1013 if ((unsigned short)weight != bfqg->entity.new_weight) {
1014 bfqg->entity.new_weight = (unsigned short)weight;
1016 * Make sure that the above new value has been
1017 * stored in bfqg->entity.new_weight before
1018 * setting the prio_changed flag. In fact,
1019 * this flag may be read asynchronously (in
1020 * critical sections protected by a different
1021 * lock than that held here), and finding this
1022 * flag set may cause the execution of the code
1023 * for updating parameters whose value may
1024 * depend also on bfqg->entity.new_weight (in
1025 * __bfq_entity_update_weight_prio).
1026 * This barrier makes sure that the new value
1027 * of bfqg->entity.new_weight is correctly
1028 * seen in that code.
1031 bfqg->entity.prio_changed = 1;
1035 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1036 struct cftype *cftype,
1039 struct blkcg *blkcg = css_to_blkcg(css);
1040 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1041 struct blkcg_gq *blkg;
1044 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1048 spin_lock_irq(&blkcg->lock);
1049 bfqgd->weight = (unsigned short)val;
1050 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1051 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1054 bfq_group_set_weight(bfqg, val, 0);
1056 spin_unlock_irq(&blkcg->lock);
1061 static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1062 char *buf, size_t nbytes,
1066 struct blkg_conf_ctx ctx;
1067 struct blkcg *blkcg = css_to_blkcg(of_css(of));
1068 struct bfq_group *bfqg;
1071 ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1075 if (sscanf(ctx.body, "%llu", &v) == 1) {
1076 /* require "default" on dfl */
1080 } else if (!strcmp(strim(ctx.body), "default")) {
1087 bfqg = blkg_to_bfqg(ctx.blkg);
1090 if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1091 bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1095 blkg_conf_finish(&ctx);
1096 return ret ?: nbytes;
1099 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1100 char *buf, size_t nbytes,
1109 /* "WEIGHT" or "default WEIGHT" sets the default weight */
1110 v = simple_strtoull(buf, &endp, 0);
1111 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1112 ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1113 return ret ?: nbytes;
1116 return bfq_io_set_device_weight(of, buf, nbytes, off);
1119 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1120 static int bfqg_print_stat(struct seq_file *sf, void *v)
1122 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1123 &blkcg_policy_bfq, seq_cft(sf)->private, false);
1127 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1129 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1130 &blkcg_policy_bfq, seq_cft(sf)->private, true);
1134 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1135 struct blkg_policy_data *pd, int off)
1137 struct blkcg_gq *blkg = pd_to_blkg(pd);
1138 struct blkcg_gq *pos_blkg;
1139 struct cgroup_subsys_state *pos_css;
1142 lockdep_assert_held(&blkg->q->queue_lock);
1145 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1146 struct bfq_stat *stat;
1148 if (!pos_blkg->online)
1151 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1152 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1156 return __blkg_prfill_u64(sf, pd, sum);
1159 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1160 struct blkg_policy_data *pd, int off)
1162 struct blkg_rwstat_sample sum;
1164 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1165 return __blkg_prfill_rwstat(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 int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1178 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1179 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1180 seq_cft(sf)->private, true);
1184 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1187 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1189 return __blkg_prfill_u64(sf, pd, sum >> 9);
1192 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1194 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1195 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1199 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1200 struct blkg_policy_data *pd, int off)
1202 struct blkg_rwstat_sample tmp;
1204 blkg_rwstat_recursive_sum(pd->blkg, NULL,
1205 offsetof(struct blkcg_gq, stat_bytes), &tmp);
1207 return __blkg_prfill_u64(sf, pd,
1208 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1211 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1213 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1214 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1219 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1220 struct blkg_policy_data *pd, int off)
1222 struct bfq_group *bfqg = pd_to_bfqg(pd);
1223 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1227 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1228 v = div64_u64(v, samples);
1230 __blkg_prfill_u64(sf, pd, v);
1234 /* print avg_queue_size */
1235 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1237 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1238 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1242 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1244 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1248 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1252 return blkg_to_bfqg(bfqd->queue->root_blkg);
1255 struct blkcg_policy blkcg_policy_bfq = {
1256 .dfl_cftypes = bfq_blkg_files,
1257 .legacy_cftypes = bfq_blkcg_legacy_files,
1259 .cpd_alloc_fn = bfq_cpd_alloc,
1260 .cpd_init_fn = bfq_cpd_init,
1261 .cpd_bind_fn = bfq_cpd_init,
1262 .cpd_free_fn = bfq_cpd_free,
1264 .pd_alloc_fn = bfq_pd_alloc,
1265 .pd_init_fn = bfq_pd_init,
1266 .pd_offline_fn = bfq_pd_offline,
1267 .pd_free_fn = bfq_pd_free,
1268 .pd_reset_stats_fn = bfq_pd_reset_stats,
1271 struct cftype bfq_blkcg_legacy_files[] = {
1273 .name = "bfq.weight",
1274 .flags = CFTYPE_NOT_ON_ROOT,
1275 .seq_show = bfq_io_show_weight_legacy,
1276 .write_u64 = bfq_io_set_weight_legacy,
1279 .name = "bfq.weight_device",
1280 .flags = CFTYPE_NOT_ON_ROOT,
1281 .seq_show = bfq_io_show_weight,
1282 .write = bfq_io_set_weight,
1285 /* statistics, covers only the tasks in the bfqg */
1287 .name = "bfq.io_service_bytes",
1288 .private = (unsigned long)&blkcg_policy_bfq,
1289 .seq_show = blkg_print_stat_bytes,
1292 .name = "bfq.io_serviced",
1293 .private = (unsigned long)&blkcg_policy_bfq,
1294 .seq_show = blkg_print_stat_ios,
1296 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1299 .private = offsetof(struct bfq_group, stats.time),
1300 .seq_show = bfqg_print_stat,
1303 .name = "bfq.sectors",
1304 .seq_show = bfqg_print_stat_sectors,
1307 .name = "bfq.io_service_time",
1308 .private = offsetof(struct bfq_group, stats.service_time),
1309 .seq_show = bfqg_print_rwstat,
1312 .name = "bfq.io_wait_time",
1313 .private = offsetof(struct bfq_group, stats.wait_time),
1314 .seq_show = bfqg_print_rwstat,
1317 .name = "bfq.io_merged",
1318 .private = offsetof(struct bfq_group, stats.merged),
1319 .seq_show = bfqg_print_rwstat,
1322 .name = "bfq.io_queued",
1323 .private = offsetof(struct bfq_group, stats.queued),
1324 .seq_show = bfqg_print_rwstat,
1326 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1328 /* the same statistics which cover the bfqg and its descendants */
1330 .name = "bfq.io_service_bytes_recursive",
1331 .private = (unsigned long)&blkcg_policy_bfq,
1332 .seq_show = blkg_print_stat_bytes_recursive,
1335 .name = "bfq.io_serviced_recursive",
1336 .private = (unsigned long)&blkcg_policy_bfq,
1337 .seq_show = blkg_print_stat_ios_recursive,
1339 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1341 .name = "bfq.time_recursive",
1342 .private = offsetof(struct bfq_group, stats.time),
1343 .seq_show = bfqg_print_stat_recursive,
1346 .name = "bfq.sectors_recursive",
1347 .seq_show = bfqg_print_stat_sectors_recursive,
1350 .name = "bfq.io_service_time_recursive",
1351 .private = offsetof(struct bfq_group, stats.service_time),
1352 .seq_show = bfqg_print_rwstat_recursive,
1355 .name = "bfq.io_wait_time_recursive",
1356 .private = offsetof(struct bfq_group, stats.wait_time),
1357 .seq_show = bfqg_print_rwstat_recursive,
1360 .name = "bfq.io_merged_recursive",
1361 .private = offsetof(struct bfq_group, stats.merged),
1362 .seq_show = bfqg_print_rwstat_recursive,
1365 .name = "bfq.io_queued_recursive",
1366 .private = offsetof(struct bfq_group, stats.queued),
1367 .seq_show = bfqg_print_rwstat_recursive,
1370 .name = "bfq.avg_queue_size",
1371 .seq_show = bfqg_print_avg_queue_size,
1374 .name = "bfq.group_wait_time",
1375 .private = offsetof(struct bfq_group, stats.group_wait_time),
1376 .seq_show = bfqg_print_stat,
1379 .name = "bfq.idle_time",
1380 .private = offsetof(struct bfq_group, stats.idle_time),
1381 .seq_show = bfqg_print_stat,
1384 .name = "bfq.empty_time",
1385 .private = offsetof(struct bfq_group, stats.empty_time),
1386 .seq_show = bfqg_print_stat,
1389 .name = "bfq.dequeue",
1390 .private = offsetof(struct bfq_group, stats.dequeue),
1391 .seq_show = bfqg_print_stat,
1393 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1397 struct cftype bfq_blkg_files[] = {
1399 .name = "bfq.weight",
1400 .flags = CFTYPE_NOT_ON_ROOT,
1401 .seq_show = bfq_io_show_weight,
1402 .write = bfq_io_set_weight,
1407 #else /* CONFIG_BFQ_GROUP_IOSCHED */
1409 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1410 struct bfq_group *bfqg) {}
1412 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1414 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1416 entity->weight = entity->new_weight;
1417 entity->orig_weight = entity->new_weight;
1419 bfqq->ioprio = bfqq->new_ioprio;
1420 bfqq->ioprio_class = bfqq->new_ioprio_class;
1422 entity->sched_data = &bfqg->sched_data;
1425 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1427 void bfq_end_wr_async(struct bfq_data *bfqd)
1429 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1432 struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio)
1434 return bfqd->root_group;
1437 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1439 return bfqq->bfqd->root_group;
1442 void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1444 void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1446 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1448 struct bfq_group *bfqg;
1451 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1455 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1456 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1460 #endif /* CONFIG_BFQ_GROUP_IOSCHED */