2 * Common Block IO controller cgroup interface
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
13 * For policy-specific per-blkcg data:
14 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
15 * Arianna Avanzini <avanzini.arianna@gmail.com>
17 #include <linux/ioprio.h>
18 #include <linux/kdev_t.h>
19 #include <linux/module.h>
20 #include <linux/sched/signal.h>
21 #include <linux/err.h>
22 #include <linux/blkdev.h>
23 #include <linux/backing-dev.h>
24 #include <linux/slab.h>
25 #include <linux/genhd.h>
26 #include <linux/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/blk-cgroup.h>
30 #include <linux/tracehook.h>
33 #define MAX_KEY_LEN 100
36 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
37 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
38 * policy [un]register operations including cgroup file additions /
39 * removals. Putting cgroup file registration outside blkcg_pol_mutex
40 * allows grabbing it from cgroup callbacks.
42 static DEFINE_MUTEX(blkcg_pol_register_mutex);
43 static DEFINE_MUTEX(blkcg_pol_mutex);
45 struct blkcg blkcg_root;
46 EXPORT_SYMBOL_GPL(blkcg_root);
48 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
50 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
52 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
54 static bool blkcg_debug_stats = false;
56 static bool blkcg_policy_enabled(struct request_queue *q,
57 const struct blkcg_policy *pol)
59 return pol && test_bit(pol->plid, q->blkcg_pols);
63 * blkg_free - free a blkg
66 * Free @blkg which may be partially allocated.
68 static void blkg_free(struct blkcg_gq *blkg)
75 for (i = 0; i < BLKCG_MAX_POLS; i++)
77 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
79 blkg_rwstat_exit(&blkg->stat_ios);
80 blkg_rwstat_exit(&blkg->stat_bytes);
84 static void __blkg_release(struct rcu_head *rcu)
86 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
88 percpu_ref_exit(&blkg->refcnt);
90 /* release the blkcg and parent blkg refs this blkg has been holding */
91 css_put(&blkg->blkcg->css);
93 blkg_put(blkg->parent);
95 wb_congested_put(blkg->wb_congested);
101 * A group is RCU protected, but having an rcu lock does not mean that one
102 * can access all the fields of blkg and assume these are valid. For
103 * example, don't try to follow throtl_data and request queue links.
105 * Having a reference to blkg under an rcu allows accesses to only values
106 * local to groups like group stats and group rate limits.
108 static void blkg_release(struct percpu_ref *ref)
110 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
112 call_rcu(&blkg->rcu_head, __blkg_release);
116 * blkg_alloc - allocate a blkg
117 * @blkcg: block cgroup the new blkg is associated with
118 * @q: request_queue the new blkg is associated with
119 * @gfp_mask: allocation mask to use
121 * Allocate a new blkg assocating @blkcg and @q.
123 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
126 struct blkcg_gq *blkg;
129 /* alloc and init base part */
130 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
134 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
135 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
139 INIT_LIST_HEAD(&blkg->q_node);
142 for (i = 0; i < BLKCG_MAX_POLS; i++) {
143 struct blkcg_policy *pol = blkcg_policy[i];
144 struct blkg_policy_data *pd;
146 if (!blkcg_policy_enabled(q, pol))
149 /* alloc per-policy data and attach it to blkg */
150 pd = pol->pd_alloc_fn(gfp_mask, q->node);
166 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
167 struct request_queue *q, bool update_hint)
169 struct blkcg_gq *blkg;
172 * Hint didn't match. Look up from the radix tree. Note that the
173 * hint can only be updated under queue_lock as otherwise @blkg
174 * could have already been removed from blkg_tree. The caller is
175 * responsible for grabbing queue_lock if @update_hint.
177 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
178 if (blkg && blkg->q == q) {
180 lockdep_assert_held(&q->queue_lock);
181 rcu_assign_pointer(blkcg->blkg_hint, blkg);
188 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
191 * If @new_blkg is %NULL, this function tries to allocate a new one as
192 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
194 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
195 struct request_queue *q,
196 struct blkcg_gq *new_blkg)
198 struct blkcg_gq *blkg;
199 struct bdi_writeback_congested *wb_congested;
202 WARN_ON_ONCE(!rcu_read_lock_held());
203 lockdep_assert_held(&q->queue_lock);
205 /* request_queue is dying, do not create/recreate a blkg */
206 if (blk_queue_dying(q)) {
211 /* blkg holds a reference to blkcg */
212 if (!css_tryget_online(&blkcg->css)) {
217 wb_congested = wb_congested_get_create(q->backing_dev_info,
219 GFP_NOWAIT | __GFP_NOWARN);
227 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
228 if (unlikely(!new_blkg)) {
230 goto err_put_congested;
234 blkg->wb_congested = wb_congested;
237 if (blkcg_parent(blkcg)) {
238 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
239 if (WARN_ON_ONCE(!blkg->parent)) {
241 goto err_put_congested;
243 blkg_get(blkg->parent);
246 ret = percpu_ref_init(&blkg->refcnt, blkg_release, 0,
247 GFP_NOWAIT | __GFP_NOWARN);
251 /* invoke per-policy init */
252 for (i = 0; i < BLKCG_MAX_POLS; i++) {
253 struct blkcg_policy *pol = blkcg_policy[i];
255 if (blkg->pd[i] && pol->pd_init_fn)
256 pol->pd_init_fn(blkg->pd[i]);
260 spin_lock(&blkcg->lock);
261 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
263 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
264 list_add(&blkg->q_node, &q->blkg_list);
266 for (i = 0; i < BLKCG_MAX_POLS; i++) {
267 struct blkcg_policy *pol = blkcg_policy[i];
269 if (blkg->pd[i] && pol->pd_online_fn)
270 pol->pd_online_fn(blkg->pd[i]);
274 spin_unlock(&blkcg->lock);
279 /* @blkg failed fully initialized, use the usual release path */
284 percpu_ref_exit(&blkg->refcnt);
286 wb_congested_put(wb_congested);
288 css_put(&blkcg->css);
295 * __blkg_lookup_create - lookup blkg, try to create one if not there
296 * @blkcg: blkcg of interest
297 * @q: request_queue of interest
299 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
300 * create one. blkg creation is performed recursively from blkcg_root such
301 * that all non-root blkg's have access to the parent blkg. This function
302 * should be called under RCU read lock and @q->queue_lock.
304 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
307 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
308 struct request_queue *q)
310 struct blkcg_gq *blkg;
312 WARN_ON_ONCE(!rcu_read_lock_held());
313 lockdep_assert_held(&q->queue_lock);
315 blkg = __blkg_lookup(blkcg, q, true);
320 * Create blkgs walking down from blkcg_root to @blkcg, so that all
321 * non-root blkgs have access to their parents. Returns the closest
322 * blkg to the intended blkg should blkg_create() fail.
325 struct blkcg *pos = blkcg;
326 struct blkcg *parent = blkcg_parent(blkcg);
327 struct blkcg_gq *ret_blkg = q->root_blkg;
330 blkg = __blkg_lookup(parent, q, false);
332 /* remember closest blkg */
337 parent = blkcg_parent(parent);
340 blkg = blkg_create(pos, q, NULL);
349 * blkg_lookup_create - find or create a blkg
350 * @blkcg: target block cgroup
351 * @q: target request_queue
353 * This looks up or creates the blkg representing the unique pair
354 * of the blkcg and the request_queue.
356 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
357 struct request_queue *q)
359 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
361 if (unlikely(!blkg)) {
364 spin_lock_irqsave(&q->queue_lock, flags);
365 blkg = __blkg_lookup_create(blkcg, q);
366 spin_unlock_irqrestore(&q->queue_lock, flags);
372 static void blkg_destroy(struct blkcg_gq *blkg)
374 struct blkcg *blkcg = blkg->blkcg;
375 struct blkcg_gq *parent = blkg->parent;
378 lockdep_assert_held(&blkg->q->queue_lock);
379 lockdep_assert_held(&blkcg->lock);
381 /* Something wrong if we are trying to remove same group twice */
382 WARN_ON_ONCE(list_empty(&blkg->q_node));
383 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
385 for (i = 0; i < BLKCG_MAX_POLS; i++) {
386 struct blkcg_policy *pol = blkcg_policy[i];
388 if (blkg->pd[i] && pol->pd_offline_fn)
389 pol->pd_offline_fn(blkg->pd[i]);
393 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
394 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
397 blkg->online = false;
399 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
400 list_del_init(&blkg->q_node);
401 hlist_del_init_rcu(&blkg->blkcg_node);
404 * Both setting lookup hint to and clearing it from @blkg are done
405 * under queue_lock. If it's not pointing to @blkg now, it never
406 * will. Hint assignment itself can race safely.
408 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
409 rcu_assign_pointer(blkcg->blkg_hint, NULL);
412 * Put the reference taken at the time of creation so that when all
413 * queues are gone, group can be destroyed.
415 percpu_ref_kill(&blkg->refcnt);
419 * blkg_destroy_all - destroy all blkgs associated with a request_queue
420 * @q: request_queue of interest
422 * Destroy all blkgs associated with @q.
424 static void blkg_destroy_all(struct request_queue *q)
426 struct blkcg_gq *blkg, *n;
428 spin_lock_irq(&q->queue_lock);
429 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
430 struct blkcg *blkcg = blkg->blkcg;
432 spin_lock(&blkcg->lock);
434 spin_unlock(&blkcg->lock);
438 spin_unlock_irq(&q->queue_lock);
442 * A group is RCU protected, but having an rcu lock does not mean that one
443 * can access all the fields of blkg and assume these are valid. For
444 * example, don't try to follow throtl_data and request queue links.
446 * Having a reference to blkg under an rcu allows accesses to only values
447 * local to groups like group stats and group rate limits.
449 void __blkg_release_rcu(struct rcu_head *rcu_head)
451 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
453 /* release the blkcg and parent blkg refs this blkg has been holding */
454 css_put(&blkg->blkcg->css);
456 blkg_put(blkg->parent);
458 wb_congested_put(blkg->wb_congested);
462 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
464 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
465 struct cftype *cftype, u64 val)
467 struct blkcg *blkcg = css_to_blkcg(css);
468 struct blkcg_gq *blkg;
471 mutex_lock(&blkcg_pol_mutex);
472 spin_lock_irq(&blkcg->lock);
475 * Note that stat reset is racy - it doesn't synchronize against
476 * stat updates. This is a debug feature which shouldn't exist
477 * anyway. If you get hit by a race, retry.
479 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
480 blkg_rwstat_reset(&blkg->stat_bytes);
481 blkg_rwstat_reset(&blkg->stat_ios);
483 for (i = 0; i < BLKCG_MAX_POLS; i++) {
484 struct blkcg_policy *pol = blkcg_policy[i];
486 if (blkg->pd[i] && pol->pd_reset_stats_fn)
487 pol->pd_reset_stats_fn(blkg->pd[i]);
491 spin_unlock_irq(&blkcg->lock);
492 mutex_unlock(&blkcg_pol_mutex);
496 const char *blkg_dev_name(struct blkcg_gq *blkg)
498 /* some drivers (floppy) instantiate a queue w/o disk registered */
499 if (blkg->q->backing_dev_info->dev)
500 return dev_name(blkg->q->backing_dev_info->dev);
505 * blkcg_print_blkgs - helper for printing per-blkg data
506 * @sf: seq_file to print to
507 * @blkcg: blkcg of interest
508 * @prfill: fill function to print out a blkg
509 * @pol: policy in question
510 * @data: data to be passed to @prfill
511 * @show_total: to print out sum of prfill return values or not
513 * This function invokes @prfill on each blkg of @blkcg if pd for the
514 * policy specified by @pol exists. @prfill is invoked with @sf, the
515 * policy data and @data and the matching queue lock held. If @show_total
516 * is %true, the sum of the return values from @prfill is printed with
517 * "Total" label at the end.
519 * This is to be used to construct print functions for
520 * cftype->read_seq_string method.
522 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
523 u64 (*prfill)(struct seq_file *,
524 struct blkg_policy_data *, int),
525 const struct blkcg_policy *pol, int data,
528 struct blkcg_gq *blkg;
532 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
533 spin_lock_irq(&blkg->q->queue_lock);
534 if (blkcg_policy_enabled(blkg->q, pol))
535 total += prfill(sf, blkg->pd[pol->plid], data);
536 spin_unlock_irq(&blkg->q->queue_lock);
541 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
543 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
546 * __blkg_prfill_u64 - prfill helper for a single u64 value
547 * @sf: seq_file to print to
548 * @pd: policy private data of interest
551 * Print @v to @sf for the device assocaited with @pd.
553 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
555 const char *dname = blkg_dev_name(pd->blkg);
560 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
563 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
566 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
567 * @sf: seq_file to print to
568 * @pd: policy private data of interest
569 * @rwstat: rwstat to print
571 * Print @rwstat to @sf for the device assocaited with @pd.
573 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
574 const struct blkg_rwstat *rwstat)
576 static const char *rwstr[] = {
577 [BLKG_RWSTAT_READ] = "Read",
578 [BLKG_RWSTAT_WRITE] = "Write",
579 [BLKG_RWSTAT_SYNC] = "Sync",
580 [BLKG_RWSTAT_ASYNC] = "Async",
581 [BLKG_RWSTAT_DISCARD] = "Discard",
583 const char *dname = blkg_dev_name(pd->blkg);
590 for (i = 0; i < BLKG_RWSTAT_NR; i++)
591 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
592 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
594 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
595 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
596 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
597 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
600 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
603 * blkg_prfill_stat - prfill callback for blkg_stat
604 * @sf: seq_file to print to
605 * @pd: policy private data of interest
606 * @off: offset to the blkg_stat in @pd
608 * prfill callback for printing a blkg_stat.
610 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
612 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
614 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
617 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
618 * @sf: seq_file to print to
619 * @pd: policy private data of interest
620 * @off: offset to the blkg_rwstat in @pd
622 * prfill callback for printing a blkg_rwstat.
624 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
627 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
629 return __blkg_prfill_rwstat(sf, pd, &rwstat);
631 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
633 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
634 struct blkg_policy_data *pd, int off)
636 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
638 return __blkg_prfill_rwstat(sf, pd, &rwstat);
642 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
643 * @sf: seq_file to print to
646 * To be used as cftype->seq_show to print blkg->stat_bytes.
647 * cftype->private must be set to the blkcg_policy.
649 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
651 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
652 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
653 offsetof(struct blkcg_gq, stat_bytes), true);
656 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
659 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
660 * @sf: seq_file to print to
663 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
664 * must be set to the blkcg_policy.
666 int blkg_print_stat_ios(struct seq_file *sf, void *v)
668 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
669 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
670 offsetof(struct blkcg_gq, stat_ios), true);
673 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
675 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
676 struct blkg_policy_data *pd,
679 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
681 return __blkg_prfill_rwstat(sf, pd, &rwstat);
685 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
686 * @sf: seq_file to print to
689 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
691 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
692 blkg_prfill_rwstat_field_recursive,
693 (void *)seq_cft(sf)->private,
694 offsetof(struct blkcg_gq, stat_bytes), true);
697 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
700 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
701 * @sf: seq_file to print to
704 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
706 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
707 blkg_prfill_rwstat_field_recursive,
708 (void *)seq_cft(sf)->private,
709 offsetof(struct blkcg_gq, stat_ios), true);
712 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
715 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
716 * @blkg: blkg of interest
717 * @pol: blkcg_policy which contains the blkg_stat
718 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
720 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
721 * online descendants and their aux counts. The caller must be holding the
722 * queue lock for online tests.
724 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
725 * at @off bytes into @blkg's blkg_policy_data of the policy.
727 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
728 struct blkcg_policy *pol, int off)
730 struct blkcg_gq *pos_blkg;
731 struct cgroup_subsys_state *pos_css;
734 lockdep_assert_held(&blkg->q->queue_lock);
737 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
738 struct blkg_stat *stat;
740 if (!pos_blkg->online)
744 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
746 stat = (void *)blkg + off;
748 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
754 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
757 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
758 * @blkg: blkg of interest
759 * @pol: blkcg_policy which contains the blkg_rwstat
760 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
762 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
763 * online descendants and their aux counts. The caller must be holding the
764 * queue lock for online tests.
766 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
767 * is at @off bytes into @blkg's blkg_policy_data of the policy.
769 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
770 struct blkcg_policy *pol, int off)
772 struct blkcg_gq *pos_blkg;
773 struct cgroup_subsys_state *pos_css;
774 struct blkg_rwstat sum = { };
777 lockdep_assert_held(&blkg->q->queue_lock);
780 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
781 struct blkg_rwstat *rwstat;
783 if (!pos_blkg->online)
787 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
789 rwstat = (void *)pos_blkg + off;
791 for (i = 0; i < BLKG_RWSTAT_NR; i++)
792 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
793 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
800 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
802 /* Performs queue bypass and policy enabled checks then looks up blkg. */
803 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
804 const struct blkcg_policy *pol,
805 struct request_queue *q)
807 WARN_ON_ONCE(!rcu_read_lock_held());
808 lockdep_assert_held(&q->queue_lock);
810 if (!blkcg_policy_enabled(q, pol))
811 return ERR_PTR(-EOPNOTSUPP);
812 return __blkg_lookup(blkcg, q, true /* update_hint */);
816 * blkg_conf_prep - parse and prepare for per-blkg config update
817 * @blkcg: target block cgroup
818 * @pol: target policy
819 * @input: input string
820 * @ctx: blkg_conf_ctx to be filled
822 * Parse per-blkg config update from @input and initialize @ctx with the
823 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
824 * part of @input following MAJ:MIN. This function returns with RCU read
825 * lock and queue lock held and must be paired with blkg_conf_finish().
827 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
828 char *input, struct blkg_conf_ctx *ctx)
829 __acquires(rcu) __acquires(&disk->queue->queue_lock)
831 struct gendisk *disk;
832 struct request_queue *q;
833 struct blkcg_gq *blkg;
834 unsigned int major, minor;
835 int key_len, part, ret;
838 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
841 body = input + key_len;
844 body = skip_spaces(body);
846 disk = get_gendisk(MKDEV(major, minor), &part);
857 spin_lock_irq(&q->queue_lock);
859 blkg = blkg_lookup_check(blkcg, pol, q);
869 * Create blkgs walking down from blkcg_root to @blkcg, so that all
870 * non-root blkgs have access to their parents.
873 struct blkcg *pos = blkcg;
874 struct blkcg *parent;
875 struct blkcg_gq *new_blkg;
877 parent = blkcg_parent(blkcg);
878 while (parent && !__blkg_lookup(parent, q, false)) {
880 parent = blkcg_parent(parent);
883 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
884 spin_unlock_irq(&q->queue_lock);
887 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
888 if (unlikely(!new_blkg)) {
894 spin_lock_irq(&q->queue_lock);
896 blkg = blkg_lookup_check(pos, pol, q);
905 blkg = blkg_create(pos, q, new_blkg);
906 if (unlikely(IS_ERR(blkg))) {
922 spin_unlock_irq(&q->queue_lock);
925 put_disk_and_module(disk);
927 * If queue was bypassing, we should retry. Do so after a
928 * short msleep(). It isn't strictly necessary but queue
929 * can be bypassing for some time and it's always nice to
930 * avoid busy looping.
934 ret = restart_syscall();
940 * blkg_conf_finish - finish up per-blkg config update
941 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
943 * Finish up after per-blkg config update. This function must be paired
944 * with blkg_conf_prep().
946 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
947 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
949 spin_unlock_irq(&ctx->disk->queue->queue_lock);
951 put_disk_and_module(ctx->disk);
954 static int blkcg_print_stat(struct seq_file *sf, void *v)
956 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
957 struct blkcg_gq *blkg;
961 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
964 struct blkg_rwstat rwstat;
965 u64 rbytes, wbytes, rios, wios, dbytes, dios;
966 size_t size = seq_get_buf(sf, &buf), off = 0;
968 bool has_stats = false;
970 dname = blkg_dev_name(blkg);
975 * Hooray string manipulation, count is the size written NOT
976 * INCLUDING THE \0, so size is now count+1 less than what we
977 * had before, but we want to start writing the next bit from
978 * the \0 so we only add count to buf.
980 off += scnprintf(buf+off, size-off, "%s ", dname);
982 spin_lock_irq(&blkg->q->queue_lock);
984 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
985 offsetof(struct blkcg_gq, stat_bytes));
986 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
987 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
988 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
990 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
991 offsetof(struct blkcg_gq, stat_ios));
992 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
993 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
994 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
996 spin_unlock_irq(&blkg->q->queue_lock);
998 if (rbytes || wbytes || rios || wios) {
1000 off += scnprintf(buf+off, size-off,
1001 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1002 rbytes, wbytes, rios, wios,
1006 if (!blkcg_debug_stats)
1009 if (atomic_read(&blkg->use_delay)) {
1011 off += scnprintf(buf+off, size-off,
1012 " use_delay=%d delay_nsec=%llu",
1013 atomic_read(&blkg->use_delay),
1014 (unsigned long long)atomic64_read(&blkg->delay_nsec));
1017 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1018 struct blkcg_policy *pol = blkcg_policy[i];
1021 if (!blkg->pd[i] || !pol->pd_stat_fn)
1024 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1031 off += scnprintf(buf+off, size-off, "\n");
1032 seq_commit(sf, off);
1040 static struct cftype blkcg_files[] = {
1043 .flags = CFTYPE_NOT_ON_ROOT,
1044 .seq_show = blkcg_print_stat,
1049 static struct cftype blkcg_legacy_files[] = {
1051 .name = "reset_stats",
1052 .write_u64 = blkcg_reset_stats,
1058 * blkcg destruction is a three-stage process.
1060 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1061 * which offlines writeback. Here we tie the next stage of blkg destruction
1062 * to the completion of writeback associated with the blkcg. This lets us
1063 * avoid punting potentially large amounts of outstanding writeback to root
1064 * while maintaining any ongoing policies. The next stage is triggered when
1065 * the nr_cgwbs count goes to zero.
1067 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1068 * and handles the destruction of blkgs. Here the css reference held by
1069 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1070 * This work may occur in cgwb_release_workfn() on the cgwb_release
1071 * workqueue. Any submitted ios that fail to get the blkg ref will be
1072 * punted to the root_blkg.
1074 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1075 * This finally frees the blkcg.
1079 * blkcg_css_offline - cgroup css_offline callback
1080 * @css: css of interest
1082 * This function is called when @css is about to go away. Here the cgwbs are
1083 * offlined first and only once writeback associated with the blkcg has
1084 * finished do we start step 2 (see above).
1086 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1088 struct blkcg *blkcg = css_to_blkcg(css);
1090 /* this prevents anyone from attaching or migrating to this blkcg */
1091 wb_blkcg_offline(blkcg);
1093 /* put the base cgwb reference allowing step 2 to be triggered */
1094 blkcg_cgwb_put(blkcg);
1098 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1099 * @blkcg: blkcg of interest
1101 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1102 * is nested inside q lock, this function performs reverse double lock dancing.
1103 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1104 * blkcg_css_free to eventually be called.
1106 * This is the blkcg counterpart of ioc_release_fn().
1108 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1110 spin_lock_irq(&blkcg->lock);
1112 while (!hlist_empty(&blkcg->blkg_list)) {
1113 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1114 struct blkcg_gq, blkcg_node);
1115 struct request_queue *q = blkg->q;
1117 if (spin_trylock(&q->queue_lock)) {
1119 spin_unlock(&q->queue_lock);
1121 spin_unlock_irq(&blkcg->lock);
1123 spin_lock_irq(&blkcg->lock);
1127 spin_unlock_irq(&blkcg->lock);
1130 static void blkcg_css_free(struct cgroup_subsys_state *css)
1132 struct blkcg *blkcg = css_to_blkcg(css);
1135 mutex_lock(&blkcg_pol_mutex);
1137 list_del(&blkcg->all_blkcgs_node);
1139 for (i = 0; i < BLKCG_MAX_POLS; i++)
1141 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1143 mutex_unlock(&blkcg_pol_mutex);
1148 static struct cgroup_subsys_state *
1149 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1151 struct blkcg *blkcg;
1152 struct cgroup_subsys_state *ret;
1155 mutex_lock(&blkcg_pol_mutex);
1158 blkcg = &blkcg_root;
1160 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1162 ret = ERR_PTR(-ENOMEM);
1167 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1168 struct blkcg_policy *pol = blkcg_policy[i];
1169 struct blkcg_policy_data *cpd;
1172 * If the policy hasn't been attached yet, wait for it
1173 * to be attached before doing anything else. Otherwise,
1174 * check if the policy requires any specific per-cgroup
1175 * data: if it does, allocate and initialize it.
1177 if (!pol || !pol->cpd_alloc_fn)
1180 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1182 ret = ERR_PTR(-ENOMEM);
1185 blkcg->cpd[i] = cpd;
1188 if (pol->cpd_init_fn)
1189 pol->cpd_init_fn(cpd);
1192 spin_lock_init(&blkcg->lock);
1193 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1194 INIT_HLIST_HEAD(&blkcg->blkg_list);
1195 #ifdef CONFIG_CGROUP_WRITEBACK
1196 INIT_LIST_HEAD(&blkcg->cgwb_list);
1197 refcount_set(&blkcg->cgwb_refcnt, 1);
1199 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1201 mutex_unlock(&blkcg_pol_mutex);
1205 for (i--; i >= 0; i--)
1207 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1209 if (blkcg != &blkcg_root)
1212 mutex_unlock(&blkcg_pol_mutex);
1217 * blkcg_init_queue - initialize blkcg part of request queue
1218 * @q: request_queue to initialize
1220 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1221 * part of new request_queue @q.
1224 * 0 on success, -errno on failure.
1226 int blkcg_init_queue(struct request_queue *q)
1228 struct blkcg_gq *new_blkg, *blkg;
1232 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1236 preloaded = !radix_tree_preload(GFP_KERNEL);
1238 /* Make sure the root blkg exists. */
1240 spin_lock_irq(&q->queue_lock);
1241 blkg = blkg_create(&blkcg_root, q, new_blkg);
1244 q->root_blkg = blkg;
1245 spin_unlock_irq(&q->queue_lock);
1249 radix_tree_preload_end();
1251 ret = blk_iolatency_init(q);
1253 goto err_destroy_all;
1255 ret = blk_throtl_init(q);
1257 goto err_destroy_all;
1261 blkg_destroy_all(q);
1264 spin_unlock_irq(&q->queue_lock);
1267 radix_tree_preload_end();
1268 return PTR_ERR(blkg);
1272 * blkcg_drain_queue - drain blkcg part of request_queue
1273 * @q: request_queue to drain
1275 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1277 void blkcg_drain_queue(struct request_queue *q)
1279 lockdep_assert_held(&q->queue_lock);
1282 * @q could be exiting and already have destroyed all blkgs as
1283 * indicated by NULL root_blkg. If so, don't confuse policies.
1288 blk_throtl_drain(q);
1292 * blkcg_exit_queue - exit and release blkcg part of request_queue
1293 * @q: request_queue being released
1295 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1297 void blkcg_exit_queue(struct request_queue *q)
1299 blkg_destroy_all(q);
1304 * We cannot support shared io contexts, as we have no mean to support
1305 * two tasks with the same ioc in two different groups without major rework
1306 * of the main cic data structures. For now we allow a task to change
1307 * its cgroup only if it's the only owner of its ioc.
1309 static int blkcg_can_attach(struct cgroup_taskset *tset)
1311 struct task_struct *task;
1312 struct cgroup_subsys_state *dst_css;
1313 struct io_context *ioc;
1316 /* task_lock() is needed to avoid races with exit_io_context() */
1317 cgroup_taskset_for_each(task, dst_css, tset) {
1319 ioc = task->io_context;
1320 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1329 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1333 mutex_lock(&blkcg_pol_mutex);
1335 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1336 struct blkcg_policy *pol = blkcg_policy[i];
1337 struct blkcg *blkcg;
1339 if (!pol || !pol->cpd_bind_fn)
1342 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1343 if (blkcg->cpd[pol->plid])
1344 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1346 mutex_unlock(&blkcg_pol_mutex);
1349 static void blkcg_exit(struct task_struct *tsk)
1351 if (tsk->throttle_queue)
1352 blk_put_queue(tsk->throttle_queue);
1353 tsk->throttle_queue = NULL;
1356 struct cgroup_subsys io_cgrp_subsys = {
1357 .css_alloc = blkcg_css_alloc,
1358 .css_offline = blkcg_css_offline,
1359 .css_free = blkcg_css_free,
1360 .can_attach = blkcg_can_attach,
1362 .dfl_cftypes = blkcg_files,
1363 .legacy_cftypes = blkcg_legacy_files,
1364 .legacy_name = "blkio",
1368 * This ensures that, if available, memcg is automatically enabled
1369 * together on the default hierarchy so that the owner cgroup can
1370 * be retrieved from writeback pages.
1372 .depends_on = 1 << memory_cgrp_id,
1375 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1378 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1379 * @q: request_queue of interest
1380 * @pol: blkcg policy to activate
1382 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1383 * bypass mode to populate its blkgs with policy_data for @pol.
1385 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1386 * from IO path. Update of each blkg is protected by both queue and blkcg
1387 * locks so that holding either lock and testing blkcg_policy_enabled() is
1388 * always enough for dereferencing policy data.
1390 * The caller is responsible for synchronizing [de]activations and policy
1391 * [un]registerations. Returns 0 on success, -errno on failure.
1393 int blkcg_activate_policy(struct request_queue *q,
1394 const struct blkcg_policy *pol)
1396 struct blkg_policy_data *pd_prealloc = NULL;
1397 struct blkcg_gq *blkg;
1400 if (blkcg_policy_enabled(q, pol))
1404 blk_mq_freeze_queue(q);
1407 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1410 goto out_bypass_end;
1414 spin_lock_irq(&q->queue_lock);
1416 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1417 struct blkg_policy_data *pd;
1419 if (blkg->pd[pol->plid])
1422 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1424 swap(pd, pd_prealloc);
1426 spin_unlock_irq(&q->queue_lock);
1430 blkg->pd[pol->plid] = pd;
1432 pd->plid = pol->plid;
1433 if (pol->pd_init_fn)
1434 pol->pd_init_fn(pd);
1437 __set_bit(pol->plid, q->blkcg_pols);
1440 spin_unlock_irq(&q->queue_lock);
1443 blk_mq_unfreeze_queue(q);
1445 pol->pd_free_fn(pd_prealloc);
1448 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1451 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1452 * @q: request_queue of interest
1453 * @pol: blkcg policy to deactivate
1455 * Deactivate @pol on @q. Follows the same synchronization rules as
1456 * blkcg_activate_policy().
1458 void blkcg_deactivate_policy(struct request_queue *q,
1459 const struct blkcg_policy *pol)
1461 struct blkcg_gq *blkg;
1463 if (!blkcg_policy_enabled(q, pol))
1467 blk_mq_freeze_queue(q);
1469 spin_lock_irq(&q->queue_lock);
1471 __clear_bit(pol->plid, q->blkcg_pols);
1473 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1474 if (blkg->pd[pol->plid]) {
1475 if (pol->pd_offline_fn)
1476 pol->pd_offline_fn(blkg->pd[pol->plid]);
1477 pol->pd_free_fn(blkg->pd[pol->plid]);
1478 blkg->pd[pol->plid] = NULL;
1482 spin_unlock_irq(&q->queue_lock);
1485 blk_mq_unfreeze_queue(q);
1487 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1490 * blkcg_policy_register - register a blkcg policy
1491 * @pol: blkcg policy to register
1493 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1494 * successful registration. Returns 0 on success and -errno on failure.
1496 int blkcg_policy_register(struct blkcg_policy *pol)
1498 struct blkcg *blkcg;
1501 mutex_lock(&blkcg_pol_register_mutex);
1502 mutex_lock(&blkcg_pol_mutex);
1504 /* find an empty slot */
1506 for (i = 0; i < BLKCG_MAX_POLS; i++)
1507 if (!blkcg_policy[i])
1509 if (i >= BLKCG_MAX_POLS) {
1510 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1514 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1515 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1516 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1521 blkcg_policy[pol->plid] = pol;
1523 /* allocate and install cpd's */
1524 if (pol->cpd_alloc_fn) {
1525 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1526 struct blkcg_policy_data *cpd;
1528 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1532 blkcg->cpd[pol->plid] = cpd;
1534 cpd->plid = pol->plid;
1535 pol->cpd_init_fn(cpd);
1539 mutex_unlock(&blkcg_pol_mutex);
1541 /* everything is in place, add intf files for the new policy */
1542 if (pol->dfl_cftypes)
1543 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1545 if (pol->legacy_cftypes)
1546 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1547 pol->legacy_cftypes));
1548 mutex_unlock(&blkcg_pol_register_mutex);
1552 if (pol->cpd_free_fn) {
1553 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1554 if (blkcg->cpd[pol->plid]) {
1555 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1556 blkcg->cpd[pol->plid] = NULL;
1560 blkcg_policy[pol->plid] = NULL;
1562 mutex_unlock(&blkcg_pol_mutex);
1563 mutex_unlock(&blkcg_pol_register_mutex);
1566 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1569 * blkcg_policy_unregister - unregister a blkcg policy
1570 * @pol: blkcg policy to unregister
1572 * Undo blkcg_policy_register(@pol). Might sleep.
1574 void blkcg_policy_unregister(struct blkcg_policy *pol)
1576 struct blkcg *blkcg;
1578 mutex_lock(&blkcg_pol_register_mutex);
1580 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1583 /* kill the intf files first */
1584 if (pol->dfl_cftypes)
1585 cgroup_rm_cftypes(pol->dfl_cftypes);
1586 if (pol->legacy_cftypes)
1587 cgroup_rm_cftypes(pol->legacy_cftypes);
1589 /* remove cpds and unregister */
1590 mutex_lock(&blkcg_pol_mutex);
1592 if (pol->cpd_free_fn) {
1593 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1594 if (blkcg->cpd[pol->plid]) {
1595 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1596 blkcg->cpd[pol->plid] = NULL;
1600 blkcg_policy[pol->plid] = NULL;
1602 mutex_unlock(&blkcg_pol_mutex);
1604 mutex_unlock(&blkcg_pol_register_mutex);
1606 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1609 * Scale the accumulated delay based on how long it has been since we updated
1610 * the delay. We only call this when we are adding delay, in case it's been a
1611 * while since we added delay, and when we are checking to see if we need to
1612 * delay a task, to account for any delays that may have occurred.
1614 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1616 u64 old = atomic64_read(&blkg->delay_start);
1619 * We only want to scale down every second. The idea here is that we
1620 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1621 * time window. We only want to throttle tasks for recent delay that
1622 * has occurred, in 1 second time windows since that's the maximum
1623 * things can be throttled. We save the current delay window in
1624 * blkg->last_delay so we know what amount is still left to be charged
1625 * to the blkg from this point onward. blkg->last_use keeps track of
1626 * the use_delay counter. The idea is if we're unthrottling the blkg we
1627 * are ok with whatever is happening now, and we can take away more of
1628 * the accumulated delay as we've already throttled enough that
1629 * everybody is happy with their IO latencies.
1631 if (time_before64(old + NSEC_PER_SEC, now) &&
1632 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1633 u64 cur = atomic64_read(&blkg->delay_nsec);
1634 u64 sub = min_t(u64, blkg->last_delay, now - old);
1635 int cur_use = atomic_read(&blkg->use_delay);
1638 * We've been unthrottled, subtract a larger chunk of our
1639 * accumulated delay.
1641 if (cur_use < blkg->last_use)
1642 sub = max_t(u64, sub, blkg->last_delay >> 1);
1645 * This shouldn't happen, but handle it anyway. Our delay_nsec
1646 * should only ever be growing except here where we subtract out
1647 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1648 * rather not end up with negative numbers.
1650 if (unlikely(cur < sub)) {
1651 atomic64_set(&blkg->delay_nsec, 0);
1652 blkg->last_delay = 0;
1654 atomic64_sub(sub, &blkg->delay_nsec);
1655 blkg->last_delay = cur - sub;
1657 blkg->last_use = cur_use;
1662 * This is called when we want to actually walk up the hierarchy and check to
1663 * see if we need to throttle, and then actually throttle if there is some
1664 * accumulated delay. This should only be called upon return to user space so
1665 * we're not holding some lock that would induce a priority inversion.
1667 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1669 u64 now = ktime_to_ns(ktime_get());
1674 while (blkg->parent) {
1675 if (atomic_read(&blkg->use_delay)) {
1676 blkcg_scale_delay(blkg, now);
1677 delay_nsec = max_t(u64, delay_nsec,
1678 atomic64_read(&blkg->delay_nsec));
1680 blkg = blkg->parent;
1687 * Let's not sleep for all eternity if we've amassed a huge delay.
1688 * Swapping or metadata IO can accumulate 10's of seconds worth of
1689 * delay, and we want userspace to be able to do _something_ so cap the
1690 * delays at 1 second. If there's 10's of seconds worth of delay then
1691 * the tasks will be delayed for 1 second for every syscall.
1693 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1696 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1697 * that hasn't landed upstream yet. Once that stuff is in place we need
1698 * to do a psi_memstall_enter/leave if memdelay is set.
1701 exp = ktime_add_ns(now, delay_nsec);
1702 tok = io_schedule_prepare();
1704 __set_current_state(TASK_KILLABLE);
1705 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1707 } while (!fatal_signal_pending(current));
1708 io_schedule_finish(tok);
1712 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1714 * This is only called if we've been marked with set_notify_resume(). Obviously
1715 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1716 * check to see if current->throttle_queue is set and if not this doesn't do
1717 * anything. This should only ever be called by the resume code, it's not meant
1718 * to be called by people willy-nilly as it will actually do the work to
1719 * throttle the task if it is setup for throttling.
1721 void blkcg_maybe_throttle_current(void)
1723 struct request_queue *q = current->throttle_queue;
1724 struct cgroup_subsys_state *css;
1725 struct blkcg *blkcg;
1726 struct blkcg_gq *blkg;
1727 bool use_memdelay = current->use_memdelay;
1732 current->throttle_queue = NULL;
1733 current->use_memdelay = false;
1736 css = kthread_blkcg();
1738 blkcg = css_to_blkcg(css);
1740 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1744 blkg = blkg_lookup(blkcg, q);
1747 if (!blkg_tryget(blkg))
1751 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1761 * blkcg_schedule_throttle - this task needs to check for throttling
1762 * @q - the request queue IO was submitted on
1763 * @use_memdelay - do we charge this to memory delay for PSI
1765 * This is called by the IO controller when we know there's delay accumulated
1766 * for the blkg for this task. We do not pass the blkg because there are places
1767 * we call this that may not have that information, the swapping code for
1768 * instance will only have a request_queue at that point. This set's the
1769 * notify_resume for the task to check and see if it requires throttling before
1770 * returning to user space.
1772 * We will only schedule once per syscall. You can call this over and over
1773 * again and it will only do the check once upon return to user space, and only
1774 * throttle once. If the task needs to be throttled again it'll need to be
1775 * re-set at the next time we see the task.
1777 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1779 if (unlikely(current->flags & PF_KTHREAD))
1782 if (!blk_get_queue(q))
1785 if (current->throttle_queue)
1786 blk_put_queue(current->throttle_queue);
1787 current->throttle_queue = q;
1789 current->use_memdelay = use_memdelay;
1790 set_notify_resume(current);
1794 * blkcg_add_delay - add delay to this blkg
1795 * @now - the current time in nanoseconds
1796 * @delta - how many nanoseconds of delay to add
1798 * Charge @delta to the blkg's current delay accumulation. This is used to
1799 * throttle tasks if an IO controller thinks we need more throttling.
1801 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1803 blkcg_scale_delay(blkg, now);
1804 atomic64_add(delta, &blkg->delay_nsec);
1807 module_param(blkcg_debug_stats, bool, 0644);
1808 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");