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, int node)
506 struct bfq_group *bfqg;
508 bfqg = kzalloc_node(sizeof(*bfqg), gfp, node);
512 if (bfqg_stats_init(&bfqg->stats, gfp)) {
517 /* see comments in bfq_bic_update_cgroup for why refcounting */
522 static void bfq_pd_init(struct blkg_policy_data *pd)
524 struct blkcg_gq *blkg = pd_to_blkg(pd);
525 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
526 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
527 struct bfq_entity *entity = &bfqg->entity;
528 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
530 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
531 entity->my_sched_data = &bfqg->sched_data;
532 bfqg->my_entity = entity; /*
533 * the root_group's will be set to NULL
534 * in bfq_init_queue()
537 bfqg->active_entities = 0;
538 bfqg->rq_pos_tree = RB_ROOT;
541 static void bfq_pd_free(struct blkg_policy_data *pd)
543 struct bfq_group *bfqg = pd_to_bfqg(pd);
545 bfqg_stats_exit(&bfqg->stats);
549 static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
551 struct bfq_group *bfqg = pd_to_bfqg(pd);
553 bfqg_stats_reset(&bfqg->stats);
556 static void bfq_group_set_parent(struct bfq_group *bfqg,
557 struct bfq_group *parent)
559 struct bfq_entity *entity;
561 entity = &bfqg->entity;
562 entity->parent = parent->my_entity;
563 entity->sched_data = &parent->sched_data;
566 static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
569 struct blkcg_gq *blkg;
571 blkg = blkg_lookup(blkcg, bfqd->queue);
573 return blkg_to_bfqg(blkg);
577 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
580 struct bfq_group *bfqg, *parent;
581 struct bfq_entity *entity;
583 bfqg = bfq_lookup_bfqg(bfqd, blkcg);
589 * Update chain of bfq_groups as we might be handling a leaf group
590 * which, along with some of its relatives, has not been hooked yet
591 * to the private hierarchy of BFQ.
593 entity = &bfqg->entity;
594 for_each_entity(entity) {
595 bfqg = container_of(entity, struct bfq_group, entity);
596 if (bfqg != bfqd->root_group) {
597 parent = bfqg_parent(bfqg);
599 parent = bfqd->root_group;
600 bfq_group_set_parent(bfqg, parent);
608 * bfq_bfqq_move - migrate @bfqq to @bfqg.
609 * @bfqd: queue descriptor.
610 * @bfqq: the queue to move.
611 * @bfqg: the group to move to.
613 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
614 * it on the new one. Avoid putting the entity on the old group idle tree.
616 * Must be called under the scheduler lock, to make sure that the blkg
617 * owning @bfqg does not disappear (see comments in
618 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
621 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
622 struct bfq_group *bfqg)
624 struct bfq_entity *entity = &bfqq->entity;
626 /* If bfqq is empty, then bfq_bfqq_expire also invokes
627 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
628 * from data structures related to current group. Otherwise we
629 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
632 if (bfqq == bfqd->in_service_queue)
633 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
634 false, BFQQE_PREEMPTED);
636 if (bfq_bfqq_busy(bfqq))
637 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
638 else if (entity->on_st)
639 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
640 bfqg_and_blkg_put(bfqq_group(bfqq));
642 entity->parent = bfqg->my_entity;
643 entity->sched_data = &bfqg->sched_data;
644 /* pin down bfqg and its associated blkg */
645 bfqg_and_blkg_get(bfqg);
647 if (bfq_bfqq_busy(bfqq)) {
648 if (unlikely(!bfqd->nonrot_with_queueing))
649 bfq_pos_tree_add_move(bfqd, bfqq);
650 bfq_activate_bfqq(bfqd, bfqq);
653 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
654 bfq_schedule_dispatch(bfqd);
658 * __bfq_bic_change_cgroup - move @bic to @cgroup.
659 * @bfqd: the queue descriptor.
660 * @bic: the bic to move.
661 * @blkcg: the blk-cgroup to move to.
663 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
664 * sure that the reference to cgroup is valid across the call (see
665 * comments in bfq_bic_update_cgroup on this issue)
667 * NOTE: an alternative approach might have been to store the current
668 * cgroup in bfqq and getting a reference to it, reducing the lookup
669 * time here, at the price of slightly more complex code.
671 static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
672 struct bfq_io_cq *bic,
675 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
676 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
677 struct bfq_group *bfqg;
678 struct bfq_entity *entity;
680 bfqg = bfq_find_set_group(bfqd, blkcg);
683 bfqg = bfqd->root_group;
686 entity = &async_bfqq->entity;
688 if (entity->sched_data != &bfqg->sched_data) {
689 bic_set_bfqq(bic, NULL, 0);
690 bfq_log_bfqq(bfqd, async_bfqq,
691 "bic_change_group: %p %d",
692 async_bfqq, async_bfqq->ref);
693 bfq_put_queue(async_bfqq);
698 entity = &sync_bfqq->entity;
699 if (entity->sched_data != &bfqg->sched_data)
700 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
706 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
708 struct bfq_data *bfqd = bic_to_bfqd(bic);
709 struct bfq_group *bfqg = NULL;
713 serial_nr = __bio_blkcg(bio)->css.serial_nr;
716 * Check whether blkcg has changed. The condition may trigger
717 * spuriously on a newly created cic but there's no harm.
719 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
722 bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
724 * Update blkg_path for bfq_log_* functions. We cache this
725 * path, and update it here, for the following
726 * reasons. Operations on blkg objects in blk-cgroup are
727 * protected with the request_queue lock, and not with the
728 * lock that protects the instances of this scheduler
729 * (bfqd->lock). This exposes BFQ to the following sort of
732 * The blkg_lookup performed in bfq_get_queue, protected
733 * through rcu, may happen to return the address of a copy of
734 * the original blkg. If this is the case, then the
735 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
736 * the blkg, is useless: it does not prevent blk-cgroup code
737 * from destroying both the original blkg and all objects
738 * directly or indirectly referred by the copy of the
741 * On the bright side, destroy operations on a blkg invoke, as
742 * a first step, hooks of the scheduler associated with the
743 * blkg. And these hooks are executed with bfqd->lock held for
744 * BFQ. As a consequence, for any blkg associated with the
745 * request queue this instance of the scheduler is attached
746 * to, we are guaranteed that such a blkg is not destroyed, and
747 * that all the pointers it contains are consistent, while we
748 * are holding bfqd->lock. A blkg_lookup performed with
749 * bfqd->lock held then returns a fully consistent blkg, which
750 * remains consistent until this lock is held.
752 * Thanks to the last fact, and to the fact that: (1) bfqg has
753 * been obtained through a blkg_lookup in the above
754 * assignment, and (2) bfqd->lock is being held, here we can
755 * safely use the policy data for the involved blkg (i.e., the
756 * field bfqg->pd) to get to the blkg associated with bfqg,
757 * and then we can safely use any field of blkg. After we
758 * release bfqd->lock, even just getting blkg through this
759 * bfqg may cause dangling references to be traversed, as
760 * bfqg->pd may not exist any more.
762 * In view of the above facts, here we cache, in the bfqg, any
763 * blkg data we may need for this bic, and for its associated
764 * bfq_queue. As of now, we need to cache only the path of the
765 * blkg, which is used in the bfq_log_* functions.
767 * Finally, note that bfqg itself needs to be protected from
768 * destruction on the blkg_free of the original blkg (which
769 * invokes bfq_pd_free). We use an additional private
770 * refcounter for bfqg, to let it disappear only after no
771 * bfq_queue refers to it any longer.
773 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
774 bic->blkcg_serial_nr = serial_nr;
780 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
781 * @st: the service tree being flushed.
783 static void bfq_flush_idle_tree(struct bfq_service_tree *st)
785 struct bfq_entity *entity = st->first_idle;
787 for (; entity ; entity = st->first_idle)
788 __bfq_deactivate_entity(entity, false);
792 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
793 * @bfqd: the device data structure with the root group.
794 * @entity: the entity to move.
796 static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
797 struct bfq_entity *entity)
799 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
801 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
805 * bfq_reparent_active_entities - move to the root group all active
807 * @bfqd: the device data structure with the root group.
808 * @bfqg: the group to move from.
809 * @st: the service tree with the entities.
811 static void bfq_reparent_active_entities(struct bfq_data *bfqd,
812 struct bfq_group *bfqg,
813 struct bfq_service_tree *st)
815 struct rb_root *active = &st->active;
816 struct bfq_entity *entity = NULL;
818 if (!RB_EMPTY_ROOT(&st->active))
819 entity = bfq_entity_of(rb_first(active));
821 for (; entity ; entity = bfq_entity_of(rb_first(active)))
822 bfq_reparent_leaf_entity(bfqd, entity);
824 if (bfqg->sched_data.in_service_entity)
825 bfq_reparent_leaf_entity(bfqd,
826 bfqg->sched_data.in_service_entity);
830 * bfq_pd_offline - deactivate the entity associated with @pd,
831 * and reparent its children entities.
832 * @pd: descriptor of the policy going offline.
834 * blkio already grabs the queue_lock for us, so no need to use
837 static void bfq_pd_offline(struct blkg_policy_data *pd)
839 struct bfq_service_tree *st;
840 struct bfq_group *bfqg = pd_to_bfqg(pd);
841 struct bfq_data *bfqd = bfqg->bfqd;
842 struct bfq_entity *entity = bfqg->my_entity;
846 spin_lock_irqsave(&bfqd->lock, flags);
848 if (!entity) /* root group */
849 goto put_async_queues;
852 * Empty all service_trees belonging to this group before
853 * deactivating the group itself.
855 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
856 st = bfqg->sched_data.service_tree + i;
859 * The idle tree may still contain bfq_queues belonging
860 * to exited task because they never migrated to a different
861 * cgroup from the one being destroyed now.
863 bfq_flush_idle_tree(st);
866 * It may happen that some queues are still active
867 * (busy) upon group destruction (if the corresponding
868 * processes have been forced to terminate). We move
869 * all the leaf entities corresponding to these queues
871 * Also, it may happen that the group has an entity
872 * in service, which is disconnected from the active
873 * tree: it must be moved, too.
874 * There is no need to put the sync queues, as the
875 * scheduler has taken no reference.
877 bfq_reparent_active_entities(bfqd, bfqg, st);
880 __bfq_deactivate_entity(entity, false);
883 bfq_put_async_queues(bfqd, bfqg);
885 spin_unlock_irqrestore(&bfqd->lock, flags);
887 * @blkg is going offline and will be ignored by
888 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
889 * that they don't get lost. If IOs complete after this point, the
890 * stats for them will be lost. Oh well...
892 bfqg_stats_xfer_dead(bfqg);
895 void bfq_end_wr_async(struct bfq_data *bfqd)
897 struct blkcg_gq *blkg;
899 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
900 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
902 bfq_end_wr_async_queues(bfqd, bfqg);
904 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
907 static int bfq_io_show_weight(struct seq_file *sf, void *v)
909 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
910 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
911 unsigned int val = 0;
916 seq_printf(sf, "%u\n", val);
921 static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
922 struct cftype *cftype,
925 struct blkcg *blkcg = css_to_blkcg(css);
926 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
927 struct blkcg_gq *blkg;
930 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
934 spin_lock_irq(&blkcg->lock);
935 bfqgd->weight = (unsigned short)val;
936 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
937 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
942 * Setting the prio_changed flag of the entity
943 * to 1 with new_weight == weight would re-set
944 * the value of the weight to its ioprio mapping.
945 * Set the flag only if necessary.
947 if ((unsigned short)val != bfqg->entity.new_weight) {
948 bfqg->entity.new_weight = (unsigned short)val;
950 * Make sure that the above new value has been
951 * stored in bfqg->entity.new_weight before
952 * setting the prio_changed flag. In fact,
953 * this flag may be read asynchronously (in
954 * critical sections protected by a different
955 * lock than that held here), and finding this
956 * flag set may cause the execution of the code
957 * for updating parameters whose value may
958 * depend also on bfqg->entity.new_weight (in
959 * __bfq_entity_update_weight_prio).
960 * This barrier makes sure that the new value
961 * of bfqg->entity.new_weight is correctly
965 bfqg->entity.prio_changed = 1;
968 spin_unlock_irq(&blkcg->lock);
973 static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
974 char *buf, size_t nbytes,
978 /* First unsigned long found in the file is used */
979 int ret = kstrtoull(strim(buf), 0, &weight);
984 ret = bfq_io_set_weight_legacy(of_css(of), NULL, weight);
985 return ret ?: nbytes;
988 #ifdef CONFIG_BFQ_CGROUP_DEBUG
989 static int bfqg_print_stat(struct seq_file *sf, void *v)
991 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
992 &blkcg_policy_bfq, seq_cft(sf)->private, false);
996 static int bfqg_print_rwstat(struct seq_file *sf, void *v)
998 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
999 &blkcg_policy_bfq, seq_cft(sf)->private, true);
1003 static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1004 struct blkg_policy_data *pd, int off)
1006 struct blkcg_gq *blkg = pd_to_blkg(pd);
1007 struct blkcg_gq *pos_blkg;
1008 struct cgroup_subsys_state *pos_css;
1011 lockdep_assert_held(&blkg->q->queue_lock);
1014 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1015 struct bfq_stat *stat;
1017 if (!pos_blkg->online)
1020 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1021 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1025 return __blkg_prfill_u64(sf, pd, sum);
1028 static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1029 struct blkg_policy_data *pd, int off)
1031 struct blkg_rwstat_sample sum;
1033 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1034 return __blkg_prfill_rwstat(sf, pd, &sum);
1037 static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1039 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1040 bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1041 seq_cft(sf)->private, false);
1045 static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1047 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1048 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1049 seq_cft(sf)->private, true);
1053 static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1056 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1058 return __blkg_prfill_u64(sf, pd, sum >> 9);
1061 static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1063 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1064 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1068 static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1069 struct blkg_policy_data *pd, int off)
1071 struct blkg_rwstat_sample tmp;
1073 blkg_rwstat_recursive_sum(pd->blkg, NULL,
1074 offsetof(struct blkcg_gq, stat_bytes), &tmp);
1076 return __blkg_prfill_u64(sf, pd,
1077 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1080 static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1082 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1083 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1088 static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1089 struct blkg_policy_data *pd, int off)
1091 struct bfq_group *bfqg = pd_to_bfqg(pd);
1092 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1096 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1097 v = div64_u64(v, samples);
1099 __blkg_prfill_u64(sf, pd, v);
1103 /* print avg_queue_size */
1104 static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1106 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1107 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1111 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1113 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1117 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1121 return blkg_to_bfqg(bfqd->queue->root_blkg);
1124 struct blkcg_policy blkcg_policy_bfq = {
1125 .dfl_cftypes = bfq_blkg_files,
1126 .legacy_cftypes = bfq_blkcg_legacy_files,
1128 .cpd_alloc_fn = bfq_cpd_alloc,
1129 .cpd_init_fn = bfq_cpd_init,
1130 .cpd_bind_fn = bfq_cpd_init,
1131 .cpd_free_fn = bfq_cpd_free,
1133 .pd_alloc_fn = bfq_pd_alloc,
1134 .pd_init_fn = bfq_pd_init,
1135 .pd_offline_fn = bfq_pd_offline,
1136 .pd_free_fn = bfq_pd_free,
1137 .pd_reset_stats_fn = bfq_pd_reset_stats,
1140 struct cftype bfq_blkcg_legacy_files[] = {
1142 .name = "bfq.weight",
1143 .flags = CFTYPE_NOT_ON_ROOT,
1144 .seq_show = bfq_io_show_weight,
1145 .write_u64 = bfq_io_set_weight_legacy,
1148 /* statistics, covers only the tasks in the bfqg */
1150 .name = "bfq.io_service_bytes",
1151 .private = (unsigned long)&blkcg_policy_bfq,
1152 .seq_show = blkg_print_stat_bytes,
1155 .name = "bfq.io_serviced",
1156 .private = (unsigned long)&blkcg_policy_bfq,
1157 .seq_show = blkg_print_stat_ios,
1159 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1162 .private = offsetof(struct bfq_group, stats.time),
1163 .seq_show = bfqg_print_stat,
1166 .name = "bfq.sectors",
1167 .seq_show = bfqg_print_stat_sectors,
1170 .name = "bfq.io_service_time",
1171 .private = offsetof(struct bfq_group, stats.service_time),
1172 .seq_show = bfqg_print_rwstat,
1175 .name = "bfq.io_wait_time",
1176 .private = offsetof(struct bfq_group, stats.wait_time),
1177 .seq_show = bfqg_print_rwstat,
1180 .name = "bfq.io_merged",
1181 .private = offsetof(struct bfq_group, stats.merged),
1182 .seq_show = bfqg_print_rwstat,
1185 .name = "bfq.io_queued",
1186 .private = offsetof(struct bfq_group, stats.queued),
1187 .seq_show = bfqg_print_rwstat,
1189 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1191 /* the same statistics which cover the bfqg and its descendants */
1193 .name = "bfq.io_service_bytes_recursive",
1194 .private = (unsigned long)&blkcg_policy_bfq,
1195 .seq_show = blkg_print_stat_bytes_recursive,
1198 .name = "bfq.io_serviced_recursive",
1199 .private = (unsigned long)&blkcg_policy_bfq,
1200 .seq_show = blkg_print_stat_ios_recursive,
1202 #ifdef CONFIG_BFQ_CGROUP_DEBUG
1204 .name = "bfq.time_recursive",
1205 .private = offsetof(struct bfq_group, stats.time),
1206 .seq_show = bfqg_print_stat_recursive,
1209 .name = "bfq.sectors_recursive",
1210 .seq_show = bfqg_print_stat_sectors_recursive,
1213 .name = "bfq.io_service_time_recursive",
1214 .private = offsetof(struct bfq_group, stats.service_time),
1215 .seq_show = bfqg_print_rwstat_recursive,
1218 .name = "bfq.io_wait_time_recursive",
1219 .private = offsetof(struct bfq_group, stats.wait_time),
1220 .seq_show = bfqg_print_rwstat_recursive,
1223 .name = "bfq.io_merged_recursive",
1224 .private = offsetof(struct bfq_group, stats.merged),
1225 .seq_show = bfqg_print_rwstat_recursive,
1228 .name = "bfq.io_queued_recursive",
1229 .private = offsetof(struct bfq_group, stats.queued),
1230 .seq_show = bfqg_print_rwstat_recursive,
1233 .name = "bfq.avg_queue_size",
1234 .seq_show = bfqg_print_avg_queue_size,
1237 .name = "bfq.group_wait_time",
1238 .private = offsetof(struct bfq_group, stats.group_wait_time),
1239 .seq_show = bfqg_print_stat,
1242 .name = "bfq.idle_time",
1243 .private = offsetof(struct bfq_group, stats.idle_time),
1244 .seq_show = bfqg_print_stat,
1247 .name = "bfq.empty_time",
1248 .private = offsetof(struct bfq_group, stats.empty_time),
1249 .seq_show = bfqg_print_stat,
1252 .name = "bfq.dequeue",
1253 .private = offsetof(struct bfq_group, stats.dequeue),
1254 .seq_show = bfqg_print_stat,
1256 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
1260 struct cftype bfq_blkg_files[] = {
1262 .name = "bfq.weight",
1263 .flags = CFTYPE_NOT_ON_ROOT,
1264 .seq_show = bfq_io_show_weight,
1265 .write = bfq_io_set_weight,
1270 #else /* CONFIG_BFQ_GROUP_IOSCHED */
1272 void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1273 struct bfq_group *bfqg) {}
1275 void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1277 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1279 entity->weight = entity->new_weight;
1280 entity->orig_weight = entity->new_weight;
1282 bfqq->ioprio = bfqq->new_ioprio;
1283 bfqq->ioprio_class = bfqq->new_ioprio_class;
1285 entity->sched_data = &bfqg->sched_data;
1288 void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1290 void bfq_end_wr_async(struct bfq_data *bfqd)
1292 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1295 struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1297 return bfqd->root_group;
1300 struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1302 return bfqq->bfqd->root_group;
1305 struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1307 struct bfq_group *bfqg;
1310 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1314 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1315 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1319 #endif /* CONFIG_BFQ_GROUP_IOSCHED */