1 // SPDX-License-Identifier: GPL-2.0
3 * Common Block IO controller cgroup interface
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
18 #include <linux/ioprio.h>
19 #include <linux/kdev_t.h>
20 #include <linux/module.h>
21 #include <linux/sched/signal.h>
22 #include <linux/err.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/slab.h>
26 #include <linux/genhd.h>
27 #include <linux/delay.h>
28 #include <linux/atomic.h>
29 #include <linux/ctype.h>
30 #include <linux/blk-cgroup.h>
31 #include <linux/tracehook.h>
34 #define MAX_KEY_LEN 100
37 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
38 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
39 * policy [un]register operations including cgroup file additions /
40 * removals. Putting cgroup file registration outside blkcg_pol_mutex
41 * allows grabbing it from cgroup callbacks.
43 static DEFINE_MUTEX(blkcg_pol_register_mutex);
44 static DEFINE_MUTEX(blkcg_pol_mutex);
46 struct blkcg blkcg_root;
47 EXPORT_SYMBOL_GPL(blkcg_root);
49 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
53 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
55 static bool blkcg_debug_stats = false;
57 static bool blkcg_policy_enabled(struct request_queue *q,
58 const struct blkcg_policy *pol)
60 return pol && test_bit(pol->plid, q->blkcg_pols);
64 * blkg_free - free a blkg
67 * Free @blkg which may be partially allocated.
69 static void blkg_free(struct blkcg_gq *blkg)
76 for (i = 0; i < BLKCG_MAX_POLS; i++)
78 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
80 blkg_rwstat_exit(&blkg->stat_ios);
81 blkg_rwstat_exit(&blkg->stat_bytes);
82 percpu_ref_exit(&blkg->refcnt);
86 static void __blkg_release(struct rcu_head *rcu)
88 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
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 (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
137 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
138 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
142 INIT_LIST_HEAD(&blkg->q_node);
145 for (i = 0; i < BLKCG_MAX_POLS; i++) {
146 struct blkcg_policy *pol = blkcg_policy[i];
147 struct blkg_policy_data *pd;
149 if (!blkcg_policy_enabled(q, pol))
152 /* alloc per-policy data and attach it to blkg */
153 pd = pol->pd_alloc_fn(gfp_mask, q->node);
169 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
170 struct request_queue *q, bool update_hint)
172 struct blkcg_gq *blkg;
175 * Hint didn't match. Look up from the radix tree. Note that the
176 * hint can only be updated under queue_lock as otherwise @blkg
177 * could have already been removed from blkg_tree. The caller is
178 * responsible for grabbing queue_lock if @update_hint.
180 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
181 if (blkg && blkg->q == q) {
183 lockdep_assert_held(&q->queue_lock);
184 rcu_assign_pointer(blkcg->blkg_hint, blkg);
191 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
194 * If @new_blkg is %NULL, this function tries to allocate a new one as
195 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
197 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
198 struct request_queue *q,
199 struct blkcg_gq *new_blkg)
201 struct blkcg_gq *blkg;
202 struct bdi_writeback_congested *wb_congested;
205 WARN_ON_ONCE(!rcu_read_lock_held());
206 lockdep_assert_held(&q->queue_lock);
208 /* request_queue is dying, do not create/recreate a blkg */
209 if (blk_queue_dying(q)) {
214 /* blkg holds a reference to blkcg */
215 if (!css_tryget_online(&blkcg->css)) {
220 wb_congested = wb_congested_get_create(q->backing_dev_info,
222 GFP_NOWAIT | __GFP_NOWARN);
230 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
231 if (unlikely(!new_blkg)) {
233 goto err_put_congested;
237 blkg->wb_congested = wb_congested;
240 if (blkcg_parent(blkcg)) {
241 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
242 if (WARN_ON_ONCE(!blkg->parent)) {
244 goto err_put_congested;
246 blkg_get(blkg->parent);
249 /* invoke per-policy init */
250 for (i = 0; i < BLKCG_MAX_POLS; i++) {
251 struct blkcg_policy *pol = blkcg_policy[i];
253 if (blkg->pd[i] && pol->pd_init_fn)
254 pol->pd_init_fn(blkg->pd[i]);
258 spin_lock(&blkcg->lock);
259 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
261 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
262 list_add(&blkg->q_node, &q->blkg_list);
264 for (i = 0; i < BLKCG_MAX_POLS; i++) {
265 struct blkcg_policy *pol = blkcg_policy[i];
267 if (blkg->pd[i] && pol->pd_online_fn)
268 pol->pd_online_fn(blkg->pd[i]);
272 spin_unlock(&blkcg->lock);
277 /* @blkg failed fully initialized, use the usual release path */
282 wb_congested_put(wb_congested);
284 css_put(&blkcg->css);
291 * __blkg_lookup_create - lookup blkg, try to create one if not there
292 * @blkcg: blkcg of interest
293 * @q: request_queue of interest
295 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
296 * create one. blkg creation is performed recursively from blkcg_root such
297 * that all non-root blkg's have access to the parent blkg. This function
298 * should be called under RCU read lock and @q->queue_lock.
300 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
303 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
304 struct request_queue *q)
306 struct blkcg_gq *blkg;
308 WARN_ON_ONCE(!rcu_read_lock_held());
309 lockdep_assert_held(&q->queue_lock);
311 blkg = __blkg_lookup(blkcg, q, true);
316 * Create blkgs walking down from blkcg_root to @blkcg, so that all
317 * non-root blkgs have access to their parents. Returns the closest
318 * blkg to the intended blkg should blkg_create() fail.
321 struct blkcg *pos = blkcg;
322 struct blkcg *parent = blkcg_parent(blkcg);
323 struct blkcg_gq *ret_blkg = q->root_blkg;
326 blkg = __blkg_lookup(parent, q, false);
328 /* remember closest blkg */
333 parent = blkcg_parent(parent);
336 blkg = blkg_create(pos, q, NULL);
345 * blkg_lookup_create - find or create a blkg
346 * @blkcg: target block cgroup
347 * @q: target request_queue
349 * This looks up or creates the blkg representing the unique pair
350 * of the blkcg and the request_queue.
352 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
353 struct request_queue *q)
355 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
357 if (unlikely(!blkg)) {
360 spin_lock_irqsave(&q->queue_lock, flags);
361 blkg = __blkg_lookup_create(blkcg, q);
362 spin_unlock_irqrestore(&q->queue_lock, flags);
368 static void blkg_destroy(struct blkcg_gq *blkg)
370 struct blkcg *blkcg = blkg->blkcg;
371 struct blkcg_gq *parent = blkg->parent;
374 lockdep_assert_held(&blkg->q->queue_lock);
375 lockdep_assert_held(&blkcg->lock);
377 /* Something wrong if we are trying to remove same group twice */
378 WARN_ON_ONCE(list_empty(&blkg->q_node));
379 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
381 for (i = 0; i < BLKCG_MAX_POLS; i++) {
382 struct blkcg_policy *pol = blkcg_policy[i];
384 if (blkg->pd[i] && pol->pd_offline_fn)
385 pol->pd_offline_fn(blkg->pd[i]);
389 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
390 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
393 blkg->online = false;
395 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
396 list_del_init(&blkg->q_node);
397 hlist_del_init_rcu(&blkg->blkcg_node);
400 * Both setting lookup hint to and clearing it from @blkg are done
401 * under queue_lock. If it's not pointing to @blkg now, it never
402 * will. Hint assignment itself can race safely.
404 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
405 rcu_assign_pointer(blkcg->blkg_hint, NULL);
408 * Put the reference taken at the time of creation so that when all
409 * queues are gone, group can be destroyed.
411 percpu_ref_kill(&blkg->refcnt);
415 * blkg_destroy_all - destroy all blkgs associated with a request_queue
416 * @q: request_queue of interest
418 * Destroy all blkgs associated with @q.
420 static void blkg_destroy_all(struct request_queue *q)
422 struct blkcg_gq *blkg, *n;
424 spin_lock_irq(&q->queue_lock);
425 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
426 struct blkcg *blkcg = blkg->blkcg;
428 spin_lock(&blkcg->lock);
430 spin_unlock(&blkcg->lock);
434 spin_unlock_irq(&q->queue_lock);
437 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
438 struct cftype *cftype, u64 val)
440 struct blkcg *blkcg = css_to_blkcg(css);
441 struct blkcg_gq *blkg;
444 mutex_lock(&blkcg_pol_mutex);
445 spin_lock_irq(&blkcg->lock);
448 * Note that stat reset is racy - it doesn't synchronize against
449 * stat updates. This is a debug feature which shouldn't exist
450 * anyway. If you get hit by a race, retry.
452 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
453 blkg_rwstat_reset(&blkg->stat_bytes);
454 blkg_rwstat_reset(&blkg->stat_ios);
456 for (i = 0; i < BLKCG_MAX_POLS; i++) {
457 struct blkcg_policy *pol = blkcg_policy[i];
459 if (blkg->pd[i] && pol->pd_reset_stats_fn)
460 pol->pd_reset_stats_fn(blkg->pd[i]);
464 spin_unlock_irq(&blkcg->lock);
465 mutex_unlock(&blkcg_pol_mutex);
469 const char *blkg_dev_name(struct blkcg_gq *blkg)
471 /* some drivers (floppy) instantiate a queue w/o disk registered */
472 if (blkg->q->backing_dev_info->dev)
473 return dev_name(blkg->q->backing_dev_info->dev);
478 * blkcg_print_blkgs - helper for printing per-blkg data
479 * @sf: seq_file to print to
480 * @blkcg: blkcg of interest
481 * @prfill: fill function to print out a blkg
482 * @pol: policy in question
483 * @data: data to be passed to @prfill
484 * @show_total: to print out sum of prfill return values or not
486 * This function invokes @prfill on each blkg of @blkcg if pd for the
487 * policy specified by @pol exists. @prfill is invoked with @sf, the
488 * policy data and @data and the matching queue lock held. If @show_total
489 * is %true, the sum of the return values from @prfill is printed with
490 * "Total" label at the end.
492 * This is to be used to construct print functions for
493 * cftype->read_seq_string method.
495 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
496 u64 (*prfill)(struct seq_file *,
497 struct blkg_policy_data *, int),
498 const struct blkcg_policy *pol, int data,
501 struct blkcg_gq *blkg;
505 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
506 spin_lock_irq(&blkg->q->queue_lock);
507 if (blkcg_policy_enabled(blkg->q, pol))
508 total += prfill(sf, blkg->pd[pol->plid], data);
509 spin_unlock_irq(&blkg->q->queue_lock);
514 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
516 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
519 * __blkg_prfill_u64 - prfill helper for a single u64 value
520 * @sf: seq_file to print to
521 * @pd: policy private data of interest
524 * Print @v to @sf for the device assocaited with @pd.
526 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
528 const char *dname = blkg_dev_name(pd->blkg);
533 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
536 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
539 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
540 * @sf: seq_file to print to
541 * @pd: policy private data of interest
542 * @rwstat: rwstat to print
544 * Print @rwstat to @sf for the device assocaited with @pd.
546 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
547 const struct blkg_rwstat *rwstat)
549 static const char *rwstr[] = {
550 [BLKG_RWSTAT_READ] = "Read",
551 [BLKG_RWSTAT_WRITE] = "Write",
552 [BLKG_RWSTAT_SYNC] = "Sync",
553 [BLKG_RWSTAT_ASYNC] = "Async",
554 [BLKG_RWSTAT_DISCARD] = "Discard",
556 const char *dname = blkg_dev_name(pd->blkg);
563 for (i = 0; i < BLKG_RWSTAT_NR; i++)
564 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
565 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
567 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
568 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
569 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
570 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
573 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
576 * blkg_prfill_stat - prfill callback for blkg_stat
577 * @sf: seq_file to print to
578 * @pd: policy private data of interest
579 * @off: offset to the blkg_stat in @pd
581 * prfill callback for printing a blkg_stat.
583 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
585 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
587 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
590 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
591 * @sf: seq_file to print to
592 * @pd: policy private data of interest
593 * @off: offset to the blkg_rwstat in @pd
595 * prfill callback for printing a blkg_rwstat.
597 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
600 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
602 return __blkg_prfill_rwstat(sf, pd, &rwstat);
604 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
606 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
607 struct blkg_policy_data *pd, int off)
609 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
611 return __blkg_prfill_rwstat(sf, pd, &rwstat);
615 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
616 * @sf: seq_file to print to
619 * To be used as cftype->seq_show to print blkg->stat_bytes.
620 * cftype->private must be set to the blkcg_policy.
622 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
624 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
625 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
626 offsetof(struct blkcg_gq, stat_bytes), true);
629 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
632 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
633 * @sf: seq_file to print to
636 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
637 * must be set to the blkcg_policy.
639 int blkg_print_stat_ios(struct seq_file *sf, void *v)
641 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
642 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
643 offsetof(struct blkcg_gq, stat_ios), true);
646 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
648 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
649 struct blkg_policy_data *pd,
652 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
654 return __blkg_prfill_rwstat(sf, pd, &rwstat);
658 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
659 * @sf: seq_file to print to
662 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
664 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
665 blkg_prfill_rwstat_field_recursive,
666 (void *)seq_cft(sf)->private,
667 offsetof(struct blkcg_gq, stat_bytes), true);
670 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
673 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
674 * @sf: seq_file to print to
677 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
679 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
680 blkg_prfill_rwstat_field_recursive,
681 (void *)seq_cft(sf)->private,
682 offsetof(struct blkcg_gq, stat_ios), true);
685 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
688 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
689 * @blkg: blkg of interest
690 * @pol: blkcg_policy which contains the blkg_stat
691 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
693 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
694 * online descendants and their aux counts. The caller must be holding the
695 * queue lock for online tests.
697 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
698 * at @off bytes into @blkg's blkg_policy_data of the policy.
700 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
701 struct blkcg_policy *pol, int off)
703 struct blkcg_gq *pos_blkg;
704 struct cgroup_subsys_state *pos_css;
707 lockdep_assert_held(&blkg->q->queue_lock);
710 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
711 struct blkg_stat *stat;
713 if (!pos_blkg->online)
717 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
719 stat = (void *)blkg + off;
721 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
727 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
730 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
731 * @blkg: blkg of interest
732 * @pol: blkcg_policy which contains the blkg_rwstat
733 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
735 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
736 * online descendants and their aux counts. The caller must be holding the
737 * queue lock for online tests.
739 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
740 * is at @off bytes into @blkg's blkg_policy_data of the policy.
742 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
743 struct blkcg_policy *pol, int off)
745 struct blkcg_gq *pos_blkg;
746 struct cgroup_subsys_state *pos_css;
747 struct blkg_rwstat sum = { };
750 lockdep_assert_held(&blkg->q->queue_lock);
753 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
754 struct blkg_rwstat *rwstat;
756 if (!pos_blkg->online)
760 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
762 rwstat = (void *)pos_blkg + off;
764 for (i = 0; i < BLKG_RWSTAT_NR; i++)
765 atomic64_add(blkg_rwstat_read_counter(rwstat, i),
772 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
774 /* Performs queue bypass and policy enabled checks then looks up blkg. */
775 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
776 const struct blkcg_policy *pol,
777 struct request_queue *q)
779 WARN_ON_ONCE(!rcu_read_lock_held());
780 lockdep_assert_held(&q->queue_lock);
782 if (!blkcg_policy_enabled(q, pol))
783 return ERR_PTR(-EOPNOTSUPP);
784 return __blkg_lookup(blkcg, q, true /* update_hint */);
788 * blkg_conf_prep - parse and prepare for per-blkg config update
789 * @blkcg: target block cgroup
790 * @pol: target policy
791 * @input: input string
792 * @ctx: blkg_conf_ctx to be filled
794 * Parse per-blkg config update from @input and initialize @ctx with the
795 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
796 * part of @input following MAJ:MIN. This function returns with RCU read
797 * lock and queue lock held and must be paired with blkg_conf_finish().
799 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
800 char *input, struct blkg_conf_ctx *ctx)
801 __acquires(rcu) __acquires(&disk->queue->queue_lock)
803 struct gendisk *disk;
804 struct request_queue *q;
805 struct blkcg_gq *blkg;
806 unsigned int major, minor;
807 int key_len, part, ret;
810 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
813 body = input + key_len;
816 body = skip_spaces(body);
818 disk = get_gendisk(MKDEV(major, minor), &part);
829 spin_lock_irq(&q->queue_lock);
831 blkg = blkg_lookup_check(blkcg, pol, q);
841 * Create blkgs walking down from blkcg_root to @blkcg, so that all
842 * non-root blkgs have access to their parents.
845 struct blkcg *pos = blkcg;
846 struct blkcg *parent;
847 struct blkcg_gq *new_blkg;
849 parent = blkcg_parent(blkcg);
850 while (parent && !__blkg_lookup(parent, q, false)) {
852 parent = blkcg_parent(parent);
855 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
856 spin_unlock_irq(&q->queue_lock);
859 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
860 if (unlikely(!new_blkg)) {
866 spin_lock_irq(&q->queue_lock);
868 blkg = blkg_lookup_check(pos, pol, q);
877 blkg = blkg_create(pos, q, new_blkg);
894 spin_unlock_irq(&q->queue_lock);
897 put_disk_and_module(disk);
899 * If queue was bypassing, we should retry. Do so after a
900 * short msleep(). It isn't strictly necessary but queue
901 * can be bypassing for some time and it's always nice to
902 * avoid busy looping.
906 ret = restart_syscall();
912 * blkg_conf_finish - finish up per-blkg config update
913 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
915 * Finish up after per-blkg config update. This function must be paired
916 * with blkg_conf_prep().
918 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
919 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
921 spin_unlock_irq(&ctx->disk->queue->queue_lock);
923 put_disk_and_module(ctx->disk);
926 static int blkcg_print_stat(struct seq_file *sf, void *v)
928 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
929 struct blkcg_gq *blkg;
933 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
936 struct blkg_rwstat rwstat;
937 u64 rbytes, wbytes, rios, wios, dbytes, dios;
938 size_t size = seq_get_buf(sf, &buf), off = 0;
940 bool has_stats = false;
942 dname = blkg_dev_name(blkg);
947 * Hooray string manipulation, count is the size written NOT
948 * INCLUDING THE \0, so size is now count+1 less than what we
949 * had before, but we want to start writing the next bit from
950 * the \0 so we only add count to buf.
952 off += scnprintf(buf+off, size-off, "%s ", dname);
954 spin_lock_irq(&blkg->q->queue_lock);
956 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
957 offsetof(struct blkcg_gq, stat_bytes));
958 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
959 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
960 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
962 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
963 offsetof(struct blkcg_gq, stat_ios));
964 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
965 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
966 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
968 spin_unlock_irq(&blkg->q->queue_lock);
970 if (rbytes || wbytes || rios || wios) {
972 off += scnprintf(buf+off, size-off,
973 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
974 rbytes, wbytes, rios, wios,
978 if (!blkcg_debug_stats)
981 if (atomic_read(&blkg->use_delay)) {
983 off += scnprintf(buf+off, size-off,
984 " use_delay=%d delay_nsec=%llu",
985 atomic_read(&blkg->use_delay),
986 (unsigned long long)atomic64_read(&blkg->delay_nsec));
989 for (i = 0; i < BLKCG_MAX_POLS; i++) {
990 struct blkcg_policy *pol = blkcg_policy[i];
993 if (!blkg->pd[i] || !pol->pd_stat_fn)
996 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1003 if (off < size - 1) {
1004 off += scnprintf(buf+off, size-off, "\n");
1005 seq_commit(sf, off);
1016 static struct cftype blkcg_files[] = {
1019 .flags = CFTYPE_NOT_ON_ROOT,
1020 .seq_show = blkcg_print_stat,
1025 static struct cftype blkcg_legacy_files[] = {
1027 .name = "reset_stats",
1028 .write_u64 = blkcg_reset_stats,
1034 * blkcg destruction is a three-stage process.
1036 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1037 * which offlines writeback. Here we tie the next stage of blkg destruction
1038 * to the completion of writeback associated with the blkcg. This lets us
1039 * avoid punting potentially large amounts of outstanding writeback to root
1040 * while maintaining any ongoing policies. The next stage is triggered when
1041 * the nr_cgwbs count goes to zero.
1043 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1044 * and handles the destruction of blkgs. Here the css reference held by
1045 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1046 * This work may occur in cgwb_release_workfn() on the cgwb_release
1047 * workqueue. Any submitted ios that fail to get the blkg ref will be
1048 * punted to the root_blkg.
1050 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1051 * This finally frees the blkcg.
1055 * blkcg_css_offline - cgroup css_offline callback
1056 * @css: css of interest
1058 * This function is called when @css is about to go away. Here the cgwbs are
1059 * offlined first and only once writeback associated with the blkcg has
1060 * finished do we start step 2 (see above).
1062 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1064 struct blkcg *blkcg = css_to_blkcg(css);
1066 /* this prevents anyone from attaching or migrating to this blkcg */
1067 wb_blkcg_offline(blkcg);
1069 /* put the base cgwb reference allowing step 2 to be triggered */
1070 blkcg_cgwb_put(blkcg);
1074 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1075 * @blkcg: blkcg of interest
1077 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1078 * is nested inside q lock, this function performs reverse double lock dancing.
1079 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1080 * blkcg_css_free to eventually be called.
1082 * This is the blkcg counterpart of ioc_release_fn().
1084 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1086 spin_lock_irq(&blkcg->lock);
1088 while (!hlist_empty(&blkcg->blkg_list)) {
1089 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1090 struct blkcg_gq, blkcg_node);
1091 struct request_queue *q = blkg->q;
1093 if (spin_trylock(&q->queue_lock)) {
1095 spin_unlock(&q->queue_lock);
1097 spin_unlock_irq(&blkcg->lock);
1099 spin_lock_irq(&blkcg->lock);
1103 spin_unlock_irq(&blkcg->lock);
1106 static void blkcg_css_free(struct cgroup_subsys_state *css)
1108 struct blkcg *blkcg = css_to_blkcg(css);
1111 mutex_lock(&blkcg_pol_mutex);
1113 list_del(&blkcg->all_blkcgs_node);
1115 for (i = 0; i < BLKCG_MAX_POLS; i++)
1117 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1119 mutex_unlock(&blkcg_pol_mutex);
1124 static struct cgroup_subsys_state *
1125 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1127 struct blkcg *blkcg;
1128 struct cgroup_subsys_state *ret;
1131 mutex_lock(&blkcg_pol_mutex);
1134 blkcg = &blkcg_root;
1136 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1138 ret = ERR_PTR(-ENOMEM);
1143 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1144 struct blkcg_policy *pol = blkcg_policy[i];
1145 struct blkcg_policy_data *cpd;
1148 * If the policy hasn't been attached yet, wait for it
1149 * to be attached before doing anything else. Otherwise,
1150 * check if the policy requires any specific per-cgroup
1151 * data: if it does, allocate and initialize it.
1153 if (!pol || !pol->cpd_alloc_fn)
1156 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1158 ret = ERR_PTR(-ENOMEM);
1161 blkcg->cpd[i] = cpd;
1164 if (pol->cpd_init_fn)
1165 pol->cpd_init_fn(cpd);
1168 spin_lock_init(&blkcg->lock);
1169 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1170 INIT_HLIST_HEAD(&blkcg->blkg_list);
1171 #ifdef CONFIG_CGROUP_WRITEBACK
1172 INIT_LIST_HEAD(&blkcg->cgwb_list);
1173 refcount_set(&blkcg->cgwb_refcnt, 1);
1175 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1177 mutex_unlock(&blkcg_pol_mutex);
1181 for (i--; i >= 0; i--)
1183 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1185 if (blkcg != &blkcg_root)
1188 mutex_unlock(&blkcg_pol_mutex);
1193 * blkcg_init_queue - initialize blkcg part of request queue
1194 * @q: request_queue to initialize
1196 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1197 * part of new request_queue @q.
1200 * 0 on success, -errno on failure.
1202 int blkcg_init_queue(struct request_queue *q)
1204 struct blkcg_gq *new_blkg, *blkg;
1208 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1212 preloaded = !radix_tree_preload(GFP_KERNEL);
1214 /* Make sure the root blkg exists. */
1216 spin_lock_irq(&q->queue_lock);
1217 blkg = blkg_create(&blkcg_root, q, new_blkg);
1220 q->root_blkg = blkg;
1221 spin_unlock_irq(&q->queue_lock);
1225 radix_tree_preload_end();
1227 ret = blk_iolatency_init(q);
1229 goto err_destroy_all;
1231 ret = blk_throtl_init(q);
1233 goto err_destroy_all;
1237 blkg_destroy_all(q);
1240 spin_unlock_irq(&q->queue_lock);
1243 radix_tree_preload_end();
1244 return PTR_ERR(blkg);
1248 * blkcg_drain_queue - drain blkcg part of request_queue
1249 * @q: request_queue to drain
1251 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1253 void blkcg_drain_queue(struct request_queue *q)
1255 lockdep_assert_held(&q->queue_lock);
1258 * @q could be exiting and already have destroyed all blkgs as
1259 * indicated by NULL root_blkg. If so, don't confuse policies.
1264 blk_throtl_drain(q);
1268 * blkcg_exit_queue - exit and release blkcg part of request_queue
1269 * @q: request_queue being released
1271 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1273 void blkcg_exit_queue(struct request_queue *q)
1275 blkg_destroy_all(q);
1280 * We cannot support shared io contexts, as we have no mean to support
1281 * two tasks with the same ioc in two different groups without major rework
1282 * of the main cic data structures. For now we allow a task to change
1283 * its cgroup only if it's the only owner of its ioc.
1285 static int blkcg_can_attach(struct cgroup_taskset *tset)
1287 struct task_struct *task;
1288 struct cgroup_subsys_state *dst_css;
1289 struct io_context *ioc;
1292 /* task_lock() is needed to avoid races with exit_io_context() */
1293 cgroup_taskset_for_each(task, dst_css, tset) {
1295 ioc = task->io_context;
1296 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1305 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1309 mutex_lock(&blkcg_pol_mutex);
1311 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1312 struct blkcg_policy *pol = blkcg_policy[i];
1313 struct blkcg *blkcg;
1315 if (!pol || !pol->cpd_bind_fn)
1318 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1319 if (blkcg->cpd[pol->plid])
1320 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1322 mutex_unlock(&blkcg_pol_mutex);
1325 static void blkcg_exit(struct task_struct *tsk)
1327 if (tsk->throttle_queue)
1328 blk_put_queue(tsk->throttle_queue);
1329 tsk->throttle_queue = NULL;
1332 struct cgroup_subsys io_cgrp_subsys = {
1333 .css_alloc = blkcg_css_alloc,
1334 .css_offline = blkcg_css_offline,
1335 .css_free = blkcg_css_free,
1336 .can_attach = blkcg_can_attach,
1338 .dfl_cftypes = blkcg_files,
1339 .legacy_cftypes = blkcg_legacy_files,
1340 .legacy_name = "blkio",
1344 * This ensures that, if available, memcg is automatically enabled
1345 * together on the default hierarchy so that the owner cgroup can
1346 * be retrieved from writeback pages.
1348 .depends_on = 1 << memory_cgrp_id,
1351 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1354 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1355 * @q: request_queue of interest
1356 * @pol: blkcg policy to activate
1358 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1359 * bypass mode to populate its blkgs with policy_data for @pol.
1361 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1362 * from IO path. Update of each blkg is protected by both queue and blkcg
1363 * locks so that holding either lock and testing blkcg_policy_enabled() is
1364 * always enough for dereferencing policy data.
1366 * The caller is responsible for synchronizing [de]activations and policy
1367 * [un]registerations. Returns 0 on success, -errno on failure.
1369 int blkcg_activate_policy(struct request_queue *q,
1370 const struct blkcg_policy *pol)
1372 struct blkg_policy_data *pd_prealloc = NULL;
1373 struct blkcg_gq *blkg;
1376 if (blkcg_policy_enabled(q, pol))
1380 blk_mq_freeze_queue(q);
1383 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1386 goto out_bypass_end;
1390 spin_lock_irq(&q->queue_lock);
1392 /* blkg_list is pushed at the head, reverse walk to init parents first */
1393 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1394 struct blkg_policy_data *pd;
1396 if (blkg->pd[pol->plid])
1399 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1401 swap(pd, pd_prealloc);
1403 spin_unlock_irq(&q->queue_lock);
1407 blkg->pd[pol->plid] = pd;
1409 pd->plid = pol->plid;
1410 if (pol->pd_init_fn)
1411 pol->pd_init_fn(pd);
1414 __set_bit(pol->plid, q->blkcg_pols);
1417 spin_unlock_irq(&q->queue_lock);
1420 blk_mq_unfreeze_queue(q);
1422 pol->pd_free_fn(pd_prealloc);
1425 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1428 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1429 * @q: request_queue of interest
1430 * @pol: blkcg policy to deactivate
1432 * Deactivate @pol on @q. Follows the same synchronization rules as
1433 * blkcg_activate_policy().
1435 void blkcg_deactivate_policy(struct request_queue *q,
1436 const struct blkcg_policy *pol)
1438 struct blkcg_gq *blkg;
1440 if (!blkcg_policy_enabled(q, pol))
1444 blk_mq_freeze_queue(q);
1446 spin_lock_irq(&q->queue_lock);
1448 __clear_bit(pol->plid, q->blkcg_pols);
1450 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1451 if (blkg->pd[pol->plid]) {
1452 if (pol->pd_offline_fn)
1453 pol->pd_offline_fn(blkg->pd[pol->plid]);
1454 pol->pd_free_fn(blkg->pd[pol->plid]);
1455 blkg->pd[pol->plid] = NULL;
1459 spin_unlock_irq(&q->queue_lock);
1462 blk_mq_unfreeze_queue(q);
1464 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1467 * blkcg_policy_register - register a blkcg policy
1468 * @pol: blkcg policy to register
1470 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1471 * successful registration. Returns 0 on success and -errno on failure.
1473 int blkcg_policy_register(struct blkcg_policy *pol)
1475 struct blkcg *blkcg;
1478 mutex_lock(&blkcg_pol_register_mutex);
1479 mutex_lock(&blkcg_pol_mutex);
1481 /* find an empty slot */
1483 for (i = 0; i < BLKCG_MAX_POLS; i++)
1484 if (!blkcg_policy[i])
1486 if (i >= BLKCG_MAX_POLS) {
1487 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1491 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1492 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1493 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1498 blkcg_policy[pol->plid] = pol;
1500 /* allocate and install cpd's */
1501 if (pol->cpd_alloc_fn) {
1502 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1503 struct blkcg_policy_data *cpd;
1505 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1509 blkcg->cpd[pol->plid] = cpd;
1511 cpd->plid = pol->plid;
1512 pol->cpd_init_fn(cpd);
1516 mutex_unlock(&blkcg_pol_mutex);
1518 /* everything is in place, add intf files for the new policy */
1519 if (pol->dfl_cftypes)
1520 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1522 if (pol->legacy_cftypes)
1523 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1524 pol->legacy_cftypes));
1525 mutex_unlock(&blkcg_pol_register_mutex);
1529 if (pol->cpd_free_fn) {
1530 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1531 if (blkcg->cpd[pol->plid]) {
1532 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1533 blkcg->cpd[pol->plid] = NULL;
1537 blkcg_policy[pol->plid] = NULL;
1539 mutex_unlock(&blkcg_pol_mutex);
1540 mutex_unlock(&blkcg_pol_register_mutex);
1543 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1546 * blkcg_policy_unregister - unregister a blkcg policy
1547 * @pol: blkcg policy to unregister
1549 * Undo blkcg_policy_register(@pol). Might sleep.
1551 void blkcg_policy_unregister(struct blkcg_policy *pol)
1553 struct blkcg *blkcg;
1555 mutex_lock(&blkcg_pol_register_mutex);
1557 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1560 /* kill the intf files first */
1561 if (pol->dfl_cftypes)
1562 cgroup_rm_cftypes(pol->dfl_cftypes);
1563 if (pol->legacy_cftypes)
1564 cgroup_rm_cftypes(pol->legacy_cftypes);
1566 /* remove cpds and unregister */
1567 mutex_lock(&blkcg_pol_mutex);
1569 if (pol->cpd_free_fn) {
1570 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1571 if (blkcg->cpd[pol->plid]) {
1572 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1573 blkcg->cpd[pol->plid] = NULL;
1577 blkcg_policy[pol->plid] = NULL;
1579 mutex_unlock(&blkcg_pol_mutex);
1581 mutex_unlock(&blkcg_pol_register_mutex);
1583 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1586 * Scale the accumulated delay based on how long it has been since we updated
1587 * the delay. We only call this when we are adding delay, in case it's been a
1588 * while since we added delay, and when we are checking to see if we need to
1589 * delay a task, to account for any delays that may have occurred.
1591 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1593 u64 old = atomic64_read(&blkg->delay_start);
1596 * We only want to scale down every second. The idea here is that we
1597 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1598 * time window. We only want to throttle tasks for recent delay that
1599 * has occurred, in 1 second time windows since that's the maximum
1600 * things can be throttled. We save the current delay window in
1601 * blkg->last_delay so we know what amount is still left to be charged
1602 * to the blkg from this point onward. blkg->last_use keeps track of
1603 * the use_delay counter. The idea is if we're unthrottling the blkg we
1604 * are ok with whatever is happening now, and we can take away more of
1605 * the accumulated delay as we've already throttled enough that
1606 * everybody is happy with their IO latencies.
1608 if (time_before64(old + NSEC_PER_SEC, now) &&
1609 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1610 u64 cur = atomic64_read(&blkg->delay_nsec);
1611 u64 sub = min_t(u64, blkg->last_delay, now - old);
1612 int cur_use = atomic_read(&blkg->use_delay);
1615 * We've been unthrottled, subtract a larger chunk of our
1616 * accumulated delay.
1618 if (cur_use < blkg->last_use)
1619 sub = max_t(u64, sub, blkg->last_delay >> 1);
1622 * This shouldn't happen, but handle it anyway. Our delay_nsec
1623 * should only ever be growing except here where we subtract out
1624 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1625 * rather not end up with negative numbers.
1627 if (unlikely(cur < sub)) {
1628 atomic64_set(&blkg->delay_nsec, 0);
1629 blkg->last_delay = 0;
1631 atomic64_sub(sub, &blkg->delay_nsec);
1632 blkg->last_delay = cur - sub;
1634 blkg->last_use = cur_use;
1639 * This is called when we want to actually walk up the hierarchy and check to
1640 * see if we need to throttle, and then actually throttle if there is some
1641 * accumulated delay. This should only be called upon return to user space so
1642 * we're not holding some lock that would induce a priority inversion.
1644 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1646 u64 now = ktime_to_ns(ktime_get());
1651 while (blkg->parent) {
1652 if (atomic_read(&blkg->use_delay)) {
1653 blkcg_scale_delay(blkg, now);
1654 delay_nsec = max_t(u64, delay_nsec,
1655 atomic64_read(&blkg->delay_nsec));
1657 blkg = blkg->parent;
1664 * Let's not sleep for all eternity if we've amassed a huge delay.
1665 * Swapping or metadata IO can accumulate 10's of seconds worth of
1666 * delay, and we want userspace to be able to do _something_ so cap the
1667 * delays at 1 second. If there's 10's of seconds worth of delay then
1668 * the tasks will be delayed for 1 second for every syscall.
1670 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1673 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1674 * that hasn't landed upstream yet. Once that stuff is in place we need
1675 * to do a psi_memstall_enter/leave if memdelay is set.
1678 exp = ktime_add_ns(now, delay_nsec);
1679 tok = io_schedule_prepare();
1681 __set_current_state(TASK_KILLABLE);
1682 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1684 } while (!fatal_signal_pending(current));
1685 io_schedule_finish(tok);
1689 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1691 * This is only called if we've been marked with set_notify_resume(). Obviously
1692 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1693 * check to see if current->throttle_queue is set and if not this doesn't do
1694 * anything. This should only ever be called by the resume code, it's not meant
1695 * to be called by people willy-nilly as it will actually do the work to
1696 * throttle the task if it is setup for throttling.
1698 void blkcg_maybe_throttle_current(void)
1700 struct request_queue *q = current->throttle_queue;
1701 struct cgroup_subsys_state *css;
1702 struct blkcg *blkcg;
1703 struct blkcg_gq *blkg;
1704 bool use_memdelay = current->use_memdelay;
1709 current->throttle_queue = NULL;
1710 current->use_memdelay = false;
1713 css = kthread_blkcg();
1715 blkcg = css_to_blkcg(css);
1717 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1721 blkg = blkg_lookup(blkcg, q);
1724 if (!blkg_tryget(blkg))
1728 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1738 * blkcg_schedule_throttle - this task needs to check for throttling
1739 * @q: the request queue IO was submitted on
1740 * @use_memdelay: do we charge this to memory delay for PSI
1742 * This is called by the IO controller when we know there's delay accumulated
1743 * for the blkg for this task. We do not pass the blkg because there are places
1744 * we call this that may not have that information, the swapping code for
1745 * instance will only have a request_queue at that point. This set's the
1746 * notify_resume for the task to check and see if it requires throttling before
1747 * returning to user space.
1749 * We will only schedule once per syscall. You can call this over and over
1750 * again and it will only do the check once upon return to user space, and only
1751 * throttle once. If the task needs to be throttled again it'll need to be
1752 * re-set at the next time we see the task.
1754 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1756 if (unlikely(current->flags & PF_KTHREAD))
1759 if (!blk_get_queue(q))
1762 if (current->throttle_queue)
1763 blk_put_queue(current->throttle_queue);
1764 current->throttle_queue = q;
1766 current->use_memdelay = use_memdelay;
1767 set_notify_resume(current);
1771 * blkcg_add_delay - add delay to this blkg
1772 * @blkg: blkg of interest
1773 * @now: the current time in nanoseconds
1774 * @delta: how many nanoseconds of delay to add
1776 * Charge @delta to the blkg's current delay accumulation. This is used to
1777 * throttle tasks if an IO controller thinks we need more throttling.
1779 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1781 blkcg_scale_delay(blkg, now);
1782 atomic64_add(delta, &blkg->delay_nsec);
1785 module_param(blkcg_debug_stats, bool, 0644);
1786 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");