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);
85 static void __blkg_release(struct rcu_head *rcu)
87 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
89 percpu_ref_exit(&blkg->refcnt);
91 /* release the blkcg and parent blkg refs this blkg has been holding */
92 css_put(&blkg->blkcg->css);
94 blkg_put(blkg->parent);
96 wb_congested_put(blkg->wb_congested);
102 * A group is RCU protected, but having an rcu lock does not mean that one
103 * can access all the fields of blkg and assume these are valid. For
104 * example, don't try to follow throtl_data and request queue links.
106 * Having a reference to blkg under an rcu allows accesses to only values
107 * local to groups like group stats and group rate limits.
109 static void blkg_release(struct percpu_ref *ref)
111 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
113 call_rcu(&blkg->rcu_head, __blkg_release);
117 * blkg_alloc - allocate a blkg
118 * @blkcg: block cgroup the new blkg is associated with
119 * @q: request_queue the new blkg is associated with
120 * @gfp_mask: allocation mask to use
122 * Allocate a new blkg assocating @blkcg and @q.
124 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
127 struct blkcg_gq *blkg;
130 /* alloc and init base part */
131 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
135 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
136 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
140 INIT_LIST_HEAD(&blkg->q_node);
143 for (i = 0; i < BLKCG_MAX_POLS; i++) {
144 struct blkcg_policy *pol = blkcg_policy[i];
145 struct blkg_policy_data *pd;
147 if (!blkcg_policy_enabled(q, pol))
150 /* alloc per-policy data and attach it to blkg */
151 pd = pol->pd_alloc_fn(gfp_mask, q->node);
167 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
168 struct request_queue *q, bool update_hint)
170 struct blkcg_gq *blkg;
173 * Hint didn't match. Look up from the radix tree. Note that the
174 * hint can only be updated under queue_lock as otherwise @blkg
175 * could have already been removed from blkg_tree. The caller is
176 * responsible for grabbing queue_lock if @update_hint.
178 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
179 if (blkg && blkg->q == q) {
181 lockdep_assert_held(&q->queue_lock);
182 rcu_assign_pointer(blkcg->blkg_hint, blkg);
189 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
192 * If @new_blkg is %NULL, this function tries to allocate a new one as
193 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
195 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
196 struct request_queue *q,
197 struct blkcg_gq *new_blkg)
199 struct blkcg_gq *blkg;
200 struct bdi_writeback_congested *wb_congested;
203 WARN_ON_ONCE(!rcu_read_lock_held());
204 lockdep_assert_held(&q->queue_lock);
206 /* request_queue is dying, do not create/recreate a blkg */
207 if (blk_queue_dying(q)) {
212 /* blkg holds a reference to blkcg */
213 if (!css_tryget_online(&blkcg->css)) {
218 wb_congested = wb_congested_get_create(q->backing_dev_info,
220 GFP_NOWAIT | __GFP_NOWARN);
228 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
229 if (unlikely(!new_blkg)) {
231 goto err_put_congested;
235 blkg->wb_congested = wb_congested;
238 if (blkcg_parent(blkcg)) {
239 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
240 if (WARN_ON_ONCE(!blkg->parent)) {
242 goto err_put_congested;
244 blkg_get(blkg->parent);
247 ret = percpu_ref_init(&blkg->refcnt, blkg_release, 0,
248 GFP_NOWAIT | __GFP_NOWARN);
252 /* invoke per-policy init */
253 for (i = 0; i < BLKCG_MAX_POLS; i++) {
254 struct blkcg_policy *pol = blkcg_policy[i];
256 if (blkg->pd[i] && pol->pd_init_fn)
257 pol->pd_init_fn(blkg->pd[i]);
261 spin_lock(&blkcg->lock);
262 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
264 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
265 list_add(&blkg->q_node, &q->blkg_list);
267 for (i = 0; i < BLKCG_MAX_POLS; i++) {
268 struct blkcg_policy *pol = blkcg_policy[i];
270 if (blkg->pd[i] && pol->pd_online_fn)
271 pol->pd_online_fn(blkg->pd[i]);
275 spin_unlock(&blkcg->lock);
280 /* @blkg failed fully initialized, use the usual release path */
285 percpu_ref_exit(&blkg->refcnt);
287 wb_congested_put(wb_congested);
289 css_put(&blkcg->css);
296 * __blkg_lookup_create - lookup blkg, try to create one if not there
297 * @blkcg: blkcg of interest
298 * @q: request_queue of interest
300 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
301 * create one. blkg creation is performed recursively from blkcg_root such
302 * that all non-root blkg's have access to the parent blkg. This function
303 * should be called under RCU read lock and @q->queue_lock.
305 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
308 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
309 struct request_queue *q)
311 struct blkcg_gq *blkg;
313 WARN_ON_ONCE(!rcu_read_lock_held());
314 lockdep_assert_held(&q->queue_lock);
316 blkg = __blkg_lookup(blkcg, q, true);
321 * Create blkgs walking down from blkcg_root to @blkcg, so that all
322 * non-root blkgs have access to their parents. Returns the closest
323 * blkg to the intended blkg should blkg_create() fail.
326 struct blkcg *pos = blkcg;
327 struct blkcg *parent = blkcg_parent(blkcg);
328 struct blkcg_gq *ret_blkg = q->root_blkg;
331 blkg = __blkg_lookup(parent, q, false);
333 /* remember closest blkg */
338 parent = blkcg_parent(parent);
341 blkg = blkg_create(pos, q, NULL);
350 * blkg_lookup_create - find or create a blkg
351 * @blkcg: target block cgroup
352 * @q: target request_queue
354 * This looks up or creates the blkg representing the unique pair
355 * of the blkcg and the request_queue.
357 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
358 struct request_queue *q)
360 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
362 if (unlikely(!blkg)) {
365 spin_lock_irqsave(&q->queue_lock, flags);
366 blkg = __blkg_lookup_create(blkcg, q);
367 spin_unlock_irqrestore(&q->queue_lock, flags);
373 static void blkg_destroy(struct blkcg_gq *blkg)
375 struct blkcg *blkcg = blkg->blkcg;
376 struct blkcg_gq *parent = blkg->parent;
379 lockdep_assert_held(&blkg->q->queue_lock);
380 lockdep_assert_held(&blkcg->lock);
382 /* Something wrong if we are trying to remove same group twice */
383 WARN_ON_ONCE(list_empty(&blkg->q_node));
384 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
386 for (i = 0; i < BLKCG_MAX_POLS; i++) {
387 struct blkcg_policy *pol = blkcg_policy[i];
389 if (blkg->pd[i] && pol->pd_offline_fn)
390 pol->pd_offline_fn(blkg->pd[i]);
394 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
395 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
398 blkg->online = false;
400 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
401 list_del_init(&blkg->q_node);
402 hlist_del_init_rcu(&blkg->blkcg_node);
405 * Both setting lookup hint to and clearing it from @blkg are done
406 * under queue_lock. If it's not pointing to @blkg now, it never
407 * will. Hint assignment itself can race safely.
409 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
410 rcu_assign_pointer(blkcg->blkg_hint, NULL);
413 * Put the reference taken at the time of creation so that when all
414 * queues are gone, group can be destroyed.
416 percpu_ref_kill(&blkg->refcnt);
420 * blkg_destroy_all - destroy all blkgs associated with a request_queue
421 * @q: request_queue of interest
423 * Destroy all blkgs associated with @q.
425 static void blkg_destroy_all(struct request_queue *q)
427 struct blkcg_gq *blkg, *n;
429 spin_lock_irq(&q->queue_lock);
430 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
431 struct blkcg *blkcg = blkg->blkcg;
433 spin_lock(&blkcg->lock);
435 spin_unlock(&blkcg->lock);
439 spin_unlock_irq(&q->queue_lock);
442 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
443 struct cftype *cftype, u64 val)
445 struct blkcg *blkcg = css_to_blkcg(css);
446 struct blkcg_gq *blkg;
449 mutex_lock(&blkcg_pol_mutex);
450 spin_lock_irq(&blkcg->lock);
453 * Note that stat reset is racy - it doesn't synchronize against
454 * stat updates. This is a debug feature which shouldn't exist
455 * anyway. If you get hit by a race, retry.
457 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
458 blkg_rwstat_reset(&blkg->stat_bytes);
459 blkg_rwstat_reset(&blkg->stat_ios);
461 for (i = 0; i < BLKCG_MAX_POLS; i++) {
462 struct blkcg_policy *pol = blkcg_policy[i];
464 if (blkg->pd[i] && pol->pd_reset_stats_fn)
465 pol->pd_reset_stats_fn(blkg->pd[i]);
469 spin_unlock_irq(&blkcg->lock);
470 mutex_unlock(&blkcg_pol_mutex);
474 const char *blkg_dev_name(struct blkcg_gq *blkg)
476 /* some drivers (floppy) instantiate a queue w/o disk registered */
477 if (blkg->q->backing_dev_info->dev)
478 return dev_name(blkg->q->backing_dev_info->dev);
483 * blkcg_print_blkgs - helper for printing per-blkg data
484 * @sf: seq_file to print to
485 * @blkcg: blkcg of interest
486 * @prfill: fill function to print out a blkg
487 * @pol: policy in question
488 * @data: data to be passed to @prfill
489 * @show_total: to print out sum of prfill return values or not
491 * This function invokes @prfill on each blkg of @blkcg if pd for the
492 * policy specified by @pol exists. @prfill is invoked with @sf, the
493 * policy data and @data and the matching queue lock held. If @show_total
494 * is %true, the sum of the return values from @prfill is printed with
495 * "Total" label at the end.
497 * This is to be used to construct print functions for
498 * cftype->read_seq_string method.
500 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
501 u64 (*prfill)(struct seq_file *,
502 struct blkg_policy_data *, int),
503 const struct blkcg_policy *pol, int data,
506 struct blkcg_gq *blkg;
510 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
511 spin_lock_irq(&blkg->q->queue_lock);
512 if (blkcg_policy_enabled(blkg->q, pol))
513 total += prfill(sf, blkg->pd[pol->plid], data);
514 spin_unlock_irq(&blkg->q->queue_lock);
519 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
521 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
524 * __blkg_prfill_u64 - prfill helper for a single u64 value
525 * @sf: seq_file to print to
526 * @pd: policy private data of interest
529 * Print @v to @sf for the device assocaited with @pd.
531 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
533 const char *dname = blkg_dev_name(pd->blkg);
538 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
541 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
544 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
545 * @sf: seq_file to print to
546 * @pd: policy private data of interest
547 * @rwstat: rwstat to print
549 * Print @rwstat to @sf for the device assocaited with @pd.
551 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
552 const struct blkg_rwstat *rwstat)
554 static const char *rwstr[] = {
555 [BLKG_RWSTAT_READ] = "Read",
556 [BLKG_RWSTAT_WRITE] = "Write",
557 [BLKG_RWSTAT_SYNC] = "Sync",
558 [BLKG_RWSTAT_ASYNC] = "Async",
559 [BLKG_RWSTAT_DISCARD] = "Discard",
561 const char *dname = blkg_dev_name(pd->blkg);
568 for (i = 0; i < BLKG_RWSTAT_NR; i++)
569 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
570 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
572 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
573 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
574 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
575 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
578 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
581 * blkg_prfill_stat - prfill callback for blkg_stat
582 * @sf: seq_file to print to
583 * @pd: policy private data of interest
584 * @off: offset to the blkg_stat in @pd
586 * prfill callback for printing a blkg_stat.
588 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
590 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
592 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
595 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
596 * @sf: seq_file to print to
597 * @pd: policy private data of interest
598 * @off: offset to the blkg_rwstat in @pd
600 * prfill callback for printing a blkg_rwstat.
602 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
605 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
607 return __blkg_prfill_rwstat(sf, pd, &rwstat);
609 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
611 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
612 struct blkg_policy_data *pd, int off)
614 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
616 return __blkg_prfill_rwstat(sf, pd, &rwstat);
620 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
621 * @sf: seq_file to print to
624 * To be used as cftype->seq_show to print blkg->stat_bytes.
625 * cftype->private must be set to the blkcg_policy.
627 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
629 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
630 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
631 offsetof(struct blkcg_gq, stat_bytes), true);
634 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
637 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
638 * @sf: seq_file to print to
641 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
642 * must be set to the blkcg_policy.
644 int blkg_print_stat_ios(struct seq_file *sf, void *v)
646 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
647 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
648 offsetof(struct blkcg_gq, stat_ios), true);
651 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
653 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
654 struct blkg_policy_data *pd,
657 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
659 return __blkg_prfill_rwstat(sf, pd, &rwstat);
663 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
664 * @sf: seq_file to print to
667 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
669 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
670 blkg_prfill_rwstat_field_recursive,
671 (void *)seq_cft(sf)->private,
672 offsetof(struct blkcg_gq, stat_bytes), true);
675 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
678 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
679 * @sf: seq_file to print to
682 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
684 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
685 blkg_prfill_rwstat_field_recursive,
686 (void *)seq_cft(sf)->private,
687 offsetof(struct blkcg_gq, stat_ios), true);
690 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
693 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
694 * @blkg: blkg of interest
695 * @pol: blkcg_policy which contains the blkg_stat
696 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
698 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
699 * online descendants and their aux counts. The caller must be holding the
700 * queue lock for online tests.
702 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
703 * at @off bytes into @blkg's blkg_policy_data of the policy.
705 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
706 struct blkcg_policy *pol, int off)
708 struct blkcg_gq *pos_blkg;
709 struct cgroup_subsys_state *pos_css;
712 lockdep_assert_held(&blkg->q->queue_lock);
715 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
716 struct blkg_stat *stat;
718 if (!pos_blkg->online)
722 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
724 stat = (void *)blkg + off;
726 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
732 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
735 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
736 * @blkg: blkg of interest
737 * @pol: blkcg_policy which contains the blkg_rwstat
738 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
740 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
741 * online descendants and their aux counts. The caller must be holding the
742 * queue lock for online tests.
744 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
745 * is at @off bytes into @blkg's blkg_policy_data of the policy.
747 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
748 struct blkcg_policy *pol, int off)
750 struct blkcg_gq *pos_blkg;
751 struct cgroup_subsys_state *pos_css;
752 struct blkg_rwstat sum = { };
755 lockdep_assert_held(&blkg->q->queue_lock);
758 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
759 struct blkg_rwstat *rwstat;
761 if (!pos_blkg->online)
765 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
767 rwstat = (void *)pos_blkg + off;
769 for (i = 0; i < BLKG_RWSTAT_NR; i++)
770 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
771 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
778 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
780 /* Performs queue bypass and policy enabled checks then looks up blkg. */
781 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
782 const struct blkcg_policy *pol,
783 struct request_queue *q)
785 WARN_ON_ONCE(!rcu_read_lock_held());
786 lockdep_assert_held(&q->queue_lock);
788 if (!blkcg_policy_enabled(q, pol))
789 return ERR_PTR(-EOPNOTSUPP);
790 return __blkg_lookup(blkcg, q, true /* update_hint */);
794 * blkg_conf_prep - parse and prepare for per-blkg config update
795 * @blkcg: target block cgroup
796 * @pol: target policy
797 * @input: input string
798 * @ctx: blkg_conf_ctx to be filled
800 * Parse per-blkg config update from @input and initialize @ctx with the
801 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
802 * part of @input following MAJ:MIN. This function returns with RCU read
803 * lock and queue lock held and must be paired with blkg_conf_finish().
805 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
806 char *input, struct blkg_conf_ctx *ctx)
807 __acquires(rcu) __acquires(&disk->queue->queue_lock)
809 struct gendisk *disk;
810 struct request_queue *q;
811 struct blkcg_gq *blkg;
812 unsigned int major, minor;
813 int key_len, part, ret;
816 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
819 body = input + key_len;
822 body = skip_spaces(body);
824 disk = get_gendisk(MKDEV(major, minor), &part);
835 spin_lock_irq(&q->queue_lock);
837 blkg = blkg_lookup_check(blkcg, pol, q);
847 * Create blkgs walking down from blkcg_root to @blkcg, so that all
848 * non-root blkgs have access to their parents.
851 struct blkcg *pos = blkcg;
852 struct blkcg *parent;
853 struct blkcg_gq *new_blkg;
855 parent = blkcg_parent(blkcg);
856 while (parent && !__blkg_lookup(parent, q, false)) {
858 parent = blkcg_parent(parent);
861 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
862 spin_unlock_irq(&q->queue_lock);
865 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
866 if (unlikely(!new_blkg)) {
872 spin_lock_irq(&q->queue_lock);
874 blkg = blkg_lookup_check(pos, pol, q);
883 blkg = blkg_create(pos, q, new_blkg);
884 if (unlikely(IS_ERR(blkg))) {
900 spin_unlock_irq(&q->queue_lock);
903 put_disk_and_module(disk);
905 * If queue was bypassing, we should retry. Do so after a
906 * short msleep(). It isn't strictly necessary but queue
907 * can be bypassing for some time and it's always nice to
908 * avoid busy looping.
912 ret = restart_syscall();
918 * blkg_conf_finish - finish up per-blkg config update
919 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
921 * Finish up after per-blkg config update. This function must be paired
922 * with blkg_conf_prep().
924 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
925 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
927 spin_unlock_irq(&ctx->disk->queue->queue_lock);
929 put_disk_and_module(ctx->disk);
932 static int blkcg_print_stat(struct seq_file *sf, void *v)
934 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
935 struct blkcg_gq *blkg;
939 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
942 struct blkg_rwstat rwstat;
943 u64 rbytes, wbytes, rios, wios, dbytes, dios;
944 size_t size = seq_get_buf(sf, &buf), off = 0;
946 bool has_stats = false;
948 dname = blkg_dev_name(blkg);
953 * Hooray string manipulation, count is the size written NOT
954 * INCLUDING THE \0, so size is now count+1 less than what we
955 * had before, but we want to start writing the next bit from
956 * the \0 so we only add count to buf.
958 off += scnprintf(buf+off, size-off, "%s ", dname);
960 spin_lock_irq(&blkg->q->queue_lock);
962 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
963 offsetof(struct blkcg_gq, stat_bytes));
964 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
965 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
966 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
968 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
969 offsetof(struct blkcg_gq, stat_ios));
970 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
971 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
972 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
974 spin_unlock_irq(&blkg->q->queue_lock);
976 if (rbytes || wbytes || rios || wios) {
978 off += scnprintf(buf+off, size-off,
979 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
980 rbytes, wbytes, rios, wios,
984 if (!blkcg_debug_stats)
987 if (atomic_read(&blkg->use_delay)) {
989 off += scnprintf(buf+off, size-off,
990 " use_delay=%d delay_nsec=%llu",
991 atomic_read(&blkg->use_delay),
992 (unsigned long long)atomic64_read(&blkg->delay_nsec));
995 for (i = 0; i < BLKCG_MAX_POLS; i++) {
996 struct blkcg_policy *pol = blkcg_policy[i];
999 if (!blkg->pd[i] || !pol->pd_stat_fn)
1002 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1009 off += scnprintf(buf+off, size-off, "\n");
1010 seq_commit(sf, off);
1018 static struct cftype blkcg_files[] = {
1021 .flags = CFTYPE_NOT_ON_ROOT,
1022 .seq_show = blkcg_print_stat,
1027 static struct cftype blkcg_legacy_files[] = {
1029 .name = "reset_stats",
1030 .write_u64 = blkcg_reset_stats,
1036 * blkcg destruction is a three-stage process.
1038 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1039 * which offlines writeback. Here we tie the next stage of blkg destruction
1040 * to the completion of writeback associated with the blkcg. This lets us
1041 * avoid punting potentially large amounts of outstanding writeback to root
1042 * while maintaining any ongoing policies. The next stage is triggered when
1043 * the nr_cgwbs count goes to zero.
1045 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1046 * and handles the destruction of blkgs. Here the css reference held by
1047 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1048 * This work may occur in cgwb_release_workfn() on the cgwb_release
1049 * workqueue. Any submitted ios that fail to get the blkg ref will be
1050 * punted to the root_blkg.
1052 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1053 * This finally frees the blkcg.
1057 * blkcg_css_offline - cgroup css_offline callback
1058 * @css: css of interest
1060 * This function is called when @css is about to go away. Here the cgwbs are
1061 * offlined first and only once writeback associated with the blkcg has
1062 * finished do we start step 2 (see above).
1064 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1066 struct blkcg *blkcg = css_to_blkcg(css);
1068 /* this prevents anyone from attaching or migrating to this blkcg */
1069 wb_blkcg_offline(blkcg);
1071 /* put the base cgwb reference allowing step 2 to be triggered */
1072 blkcg_cgwb_put(blkcg);
1076 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1077 * @blkcg: blkcg of interest
1079 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1080 * is nested inside q lock, this function performs reverse double lock dancing.
1081 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1082 * blkcg_css_free to eventually be called.
1084 * This is the blkcg counterpart of ioc_release_fn().
1086 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1088 spin_lock_irq(&blkcg->lock);
1090 while (!hlist_empty(&blkcg->blkg_list)) {
1091 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1092 struct blkcg_gq, blkcg_node);
1093 struct request_queue *q = blkg->q;
1095 if (spin_trylock(&q->queue_lock)) {
1097 spin_unlock(&q->queue_lock);
1099 spin_unlock_irq(&blkcg->lock);
1101 spin_lock_irq(&blkcg->lock);
1105 spin_unlock_irq(&blkcg->lock);
1108 static void blkcg_css_free(struct cgroup_subsys_state *css)
1110 struct blkcg *blkcg = css_to_blkcg(css);
1113 mutex_lock(&blkcg_pol_mutex);
1115 list_del(&blkcg->all_blkcgs_node);
1117 for (i = 0; i < BLKCG_MAX_POLS; i++)
1119 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1121 mutex_unlock(&blkcg_pol_mutex);
1126 static struct cgroup_subsys_state *
1127 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1129 struct blkcg *blkcg;
1130 struct cgroup_subsys_state *ret;
1133 mutex_lock(&blkcg_pol_mutex);
1136 blkcg = &blkcg_root;
1138 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1140 ret = ERR_PTR(-ENOMEM);
1145 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1146 struct blkcg_policy *pol = blkcg_policy[i];
1147 struct blkcg_policy_data *cpd;
1150 * If the policy hasn't been attached yet, wait for it
1151 * to be attached before doing anything else. Otherwise,
1152 * check if the policy requires any specific per-cgroup
1153 * data: if it does, allocate and initialize it.
1155 if (!pol || !pol->cpd_alloc_fn)
1158 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1160 ret = ERR_PTR(-ENOMEM);
1163 blkcg->cpd[i] = cpd;
1166 if (pol->cpd_init_fn)
1167 pol->cpd_init_fn(cpd);
1170 spin_lock_init(&blkcg->lock);
1171 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1172 INIT_HLIST_HEAD(&blkcg->blkg_list);
1173 #ifdef CONFIG_CGROUP_WRITEBACK
1174 INIT_LIST_HEAD(&blkcg->cgwb_list);
1175 refcount_set(&blkcg->cgwb_refcnt, 1);
1177 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1179 mutex_unlock(&blkcg_pol_mutex);
1183 for (i--; i >= 0; i--)
1185 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1187 if (blkcg != &blkcg_root)
1190 mutex_unlock(&blkcg_pol_mutex);
1195 * blkcg_init_queue - initialize blkcg part of request queue
1196 * @q: request_queue to initialize
1198 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1199 * part of new request_queue @q.
1202 * 0 on success, -errno on failure.
1204 int blkcg_init_queue(struct request_queue *q)
1206 struct blkcg_gq *new_blkg, *blkg;
1210 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1214 preloaded = !radix_tree_preload(GFP_KERNEL);
1216 /* Make sure the root blkg exists. */
1218 spin_lock_irq(&q->queue_lock);
1219 blkg = blkg_create(&blkcg_root, q, new_blkg);
1222 q->root_blkg = blkg;
1223 spin_unlock_irq(&q->queue_lock);
1227 radix_tree_preload_end();
1229 ret = blk_iolatency_init(q);
1231 goto err_destroy_all;
1233 ret = blk_throtl_init(q);
1235 goto err_destroy_all;
1239 blkg_destroy_all(q);
1242 spin_unlock_irq(&q->queue_lock);
1245 radix_tree_preload_end();
1246 return PTR_ERR(blkg);
1250 * blkcg_drain_queue - drain blkcg part of request_queue
1251 * @q: request_queue to drain
1253 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1255 void blkcg_drain_queue(struct request_queue *q)
1257 lockdep_assert_held(&q->queue_lock);
1260 * @q could be exiting and already have destroyed all blkgs as
1261 * indicated by NULL root_blkg. If so, don't confuse policies.
1266 blk_throtl_drain(q);
1270 * blkcg_exit_queue - exit and release blkcg part of request_queue
1271 * @q: request_queue being released
1273 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1275 void blkcg_exit_queue(struct request_queue *q)
1277 blkg_destroy_all(q);
1282 * We cannot support shared io contexts, as we have no mean to support
1283 * two tasks with the same ioc in two different groups without major rework
1284 * of the main cic data structures. For now we allow a task to change
1285 * its cgroup only if it's the only owner of its ioc.
1287 static int blkcg_can_attach(struct cgroup_taskset *tset)
1289 struct task_struct *task;
1290 struct cgroup_subsys_state *dst_css;
1291 struct io_context *ioc;
1294 /* task_lock() is needed to avoid races with exit_io_context() */
1295 cgroup_taskset_for_each(task, dst_css, tset) {
1297 ioc = task->io_context;
1298 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1307 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1311 mutex_lock(&blkcg_pol_mutex);
1313 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1314 struct blkcg_policy *pol = blkcg_policy[i];
1315 struct blkcg *blkcg;
1317 if (!pol || !pol->cpd_bind_fn)
1320 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1321 if (blkcg->cpd[pol->plid])
1322 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1324 mutex_unlock(&blkcg_pol_mutex);
1327 static void blkcg_exit(struct task_struct *tsk)
1329 if (tsk->throttle_queue)
1330 blk_put_queue(tsk->throttle_queue);
1331 tsk->throttle_queue = NULL;
1334 struct cgroup_subsys io_cgrp_subsys = {
1335 .css_alloc = blkcg_css_alloc,
1336 .css_offline = blkcg_css_offline,
1337 .css_free = blkcg_css_free,
1338 .can_attach = blkcg_can_attach,
1340 .dfl_cftypes = blkcg_files,
1341 .legacy_cftypes = blkcg_legacy_files,
1342 .legacy_name = "blkio",
1346 * This ensures that, if available, memcg is automatically enabled
1347 * together on the default hierarchy so that the owner cgroup can
1348 * be retrieved from writeback pages.
1350 .depends_on = 1 << memory_cgrp_id,
1353 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1356 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1357 * @q: request_queue of interest
1358 * @pol: blkcg policy to activate
1360 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1361 * bypass mode to populate its blkgs with policy_data for @pol.
1363 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1364 * from IO path. Update of each blkg is protected by both queue and blkcg
1365 * locks so that holding either lock and testing blkcg_policy_enabled() is
1366 * always enough for dereferencing policy data.
1368 * The caller is responsible for synchronizing [de]activations and policy
1369 * [un]registerations. Returns 0 on success, -errno on failure.
1371 int blkcg_activate_policy(struct request_queue *q,
1372 const struct blkcg_policy *pol)
1374 struct blkg_policy_data *pd_prealloc = NULL;
1375 struct blkcg_gq *blkg;
1378 if (blkcg_policy_enabled(q, pol))
1382 blk_mq_freeze_queue(q);
1385 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1388 goto out_bypass_end;
1392 spin_lock_irq(&q->queue_lock);
1394 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1395 struct blkg_policy_data *pd;
1397 if (blkg->pd[pol->plid])
1400 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1402 swap(pd, pd_prealloc);
1404 spin_unlock_irq(&q->queue_lock);
1408 blkg->pd[pol->plid] = pd;
1410 pd->plid = pol->plid;
1411 if (pol->pd_init_fn)
1412 pol->pd_init_fn(pd);
1415 __set_bit(pol->plid, q->blkcg_pols);
1418 spin_unlock_irq(&q->queue_lock);
1421 blk_mq_unfreeze_queue(q);
1423 pol->pd_free_fn(pd_prealloc);
1426 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1429 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1430 * @q: request_queue of interest
1431 * @pol: blkcg policy to deactivate
1433 * Deactivate @pol on @q. Follows the same synchronization rules as
1434 * blkcg_activate_policy().
1436 void blkcg_deactivate_policy(struct request_queue *q,
1437 const struct blkcg_policy *pol)
1439 struct blkcg_gq *blkg;
1441 if (!blkcg_policy_enabled(q, pol))
1445 blk_mq_freeze_queue(q);
1447 spin_lock_irq(&q->queue_lock);
1449 __clear_bit(pol->plid, q->blkcg_pols);
1451 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1452 if (blkg->pd[pol->plid]) {
1453 if (pol->pd_offline_fn)
1454 pol->pd_offline_fn(blkg->pd[pol->plid]);
1455 pol->pd_free_fn(blkg->pd[pol->plid]);
1456 blkg->pd[pol->plid] = NULL;
1460 spin_unlock_irq(&q->queue_lock);
1463 blk_mq_unfreeze_queue(q);
1465 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1468 * blkcg_policy_register - register a blkcg policy
1469 * @pol: blkcg policy to register
1471 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1472 * successful registration. Returns 0 on success and -errno on failure.
1474 int blkcg_policy_register(struct blkcg_policy *pol)
1476 struct blkcg *blkcg;
1479 mutex_lock(&blkcg_pol_register_mutex);
1480 mutex_lock(&blkcg_pol_mutex);
1482 /* find an empty slot */
1484 for (i = 0; i < BLKCG_MAX_POLS; i++)
1485 if (!blkcg_policy[i])
1487 if (i >= BLKCG_MAX_POLS) {
1488 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1492 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1493 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1494 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1499 blkcg_policy[pol->plid] = pol;
1501 /* allocate and install cpd's */
1502 if (pol->cpd_alloc_fn) {
1503 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1504 struct blkcg_policy_data *cpd;
1506 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1510 blkcg->cpd[pol->plid] = cpd;
1512 cpd->plid = pol->plid;
1513 pol->cpd_init_fn(cpd);
1517 mutex_unlock(&blkcg_pol_mutex);
1519 /* everything is in place, add intf files for the new policy */
1520 if (pol->dfl_cftypes)
1521 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1523 if (pol->legacy_cftypes)
1524 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1525 pol->legacy_cftypes));
1526 mutex_unlock(&blkcg_pol_register_mutex);
1530 if (pol->cpd_free_fn) {
1531 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1532 if (blkcg->cpd[pol->plid]) {
1533 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1534 blkcg->cpd[pol->plid] = NULL;
1538 blkcg_policy[pol->plid] = NULL;
1540 mutex_unlock(&blkcg_pol_mutex);
1541 mutex_unlock(&blkcg_pol_register_mutex);
1544 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1547 * blkcg_policy_unregister - unregister a blkcg policy
1548 * @pol: blkcg policy to unregister
1550 * Undo blkcg_policy_register(@pol). Might sleep.
1552 void blkcg_policy_unregister(struct blkcg_policy *pol)
1554 struct blkcg *blkcg;
1556 mutex_lock(&blkcg_pol_register_mutex);
1558 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1561 /* kill the intf files first */
1562 if (pol->dfl_cftypes)
1563 cgroup_rm_cftypes(pol->dfl_cftypes);
1564 if (pol->legacy_cftypes)
1565 cgroup_rm_cftypes(pol->legacy_cftypes);
1567 /* remove cpds and unregister */
1568 mutex_lock(&blkcg_pol_mutex);
1570 if (pol->cpd_free_fn) {
1571 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1572 if (blkcg->cpd[pol->plid]) {
1573 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1574 blkcg->cpd[pol->plid] = NULL;
1578 blkcg_policy[pol->plid] = NULL;
1580 mutex_unlock(&blkcg_pol_mutex);
1582 mutex_unlock(&blkcg_pol_register_mutex);
1584 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1587 * Scale the accumulated delay based on how long it has been since we updated
1588 * the delay. We only call this when we are adding delay, in case it's been a
1589 * while since we added delay, and when we are checking to see if we need to
1590 * delay a task, to account for any delays that may have occurred.
1592 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1594 u64 old = atomic64_read(&blkg->delay_start);
1597 * We only want to scale down every second. The idea here is that we
1598 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1599 * time window. We only want to throttle tasks for recent delay that
1600 * has occurred, in 1 second time windows since that's the maximum
1601 * things can be throttled. We save the current delay window in
1602 * blkg->last_delay so we know what amount is still left to be charged
1603 * to the blkg from this point onward. blkg->last_use keeps track of
1604 * the use_delay counter. The idea is if we're unthrottling the blkg we
1605 * are ok with whatever is happening now, and we can take away more of
1606 * the accumulated delay as we've already throttled enough that
1607 * everybody is happy with their IO latencies.
1609 if (time_before64(old + NSEC_PER_SEC, now) &&
1610 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1611 u64 cur = atomic64_read(&blkg->delay_nsec);
1612 u64 sub = min_t(u64, blkg->last_delay, now - old);
1613 int cur_use = atomic_read(&blkg->use_delay);
1616 * We've been unthrottled, subtract a larger chunk of our
1617 * accumulated delay.
1619 if (cur_use < blkg->last_use)
1620 sub = max_t(u64, sub, blkg->last_delay >> 1);
1623 * This shouldn't happen, but handle it anyway. Our delay_nsec
1624 * should only ever be growing except here where we subtract out
1625 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1626 * rather not end up with negative numbers.
1628 if (unlikely(cur < sub)) {
1629 atomic64_set(&blkg->delay_nsec, 0);
1630 blkg->last_delay = 0;
1632 atomic64_sub(sub, &blkg->delay_nsec);
1633 blkg->last_delay = cur - sub;
1635 blkg->last_use = cur_use;
1640 * This is called when we want to actually walk up the hierarchy and check to
1641 * see if we need to throttle, and then actually throttle if there is some
1642 * accumulated delay. This should only be called upon return to user space so
1643 * we're not holding some lock that would induce a priority inversion.
1645 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1647 u64 now = ktime_to_ns(ktime_get());
1652 while (blkg->parent) {
1653 if (atomic_read(&blkg->use_delay)) {
1654 blkcg_scale_delay(blkg, now);
1655 delay_nsec = max_t(u64, delay_nsec,
1656 atomic64_read(&blkg->delay_nsec));
1658 blkg = blkg->parent;
1665 * Let's not sleep for all eternity if we've amassed a huge delay.
1666 * Swapping or metadata IO can accumulate 10's of seconds worth of
1667 * delay, and we want userspace to be able to do _something_ so cap the
1668 * delays at 1 second. If there's 10's of seconds worth of delay then
1669 * the tasks will be delayed for 1 second for every syscall.
1671 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1674 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1675 * that hasn't landed upstream yet. Once that stuff is in place we need
1676 * to do a psi_memstall_enter/leave if memdelay is set.
1679 exp = ktime_add_ns(now, delay_nsec);
1680 tok = io_schedule_prepare();
1682 __set_current_state(TASK_KILLABLE);
1683 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1685 } while (!fatal_signal_pending(current));
1686 io_schedule_finish(tok);
1690 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1692 * This is only called if we've been marked with set_notify_resume(). Obviously
1693 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1694 * check to see if current->throttle_queue is set and if not this doesn't do
1695 * anything. This should only ever be called by the resume code, it's not meant
1696 * to be called by people willy-nilly as it will actually do the work to
1697 * throttle the task if it is setup for throttling.
1699 void blkcg_maybe_throttle_current(void)
1701 struct request_queue *q = current->throttle_queue;
1702 struct cgroup_subsys_state *css;
1703 struct blkcg *blkcg;
1704 struct blkcg_gq *blkg;
1705 bool use_memdelay = current->use_memdelay;
1710 current->throttle_queue = NULL;
1711 current->use_memdelay = false;
1714 css = kthread_blkcg();
1716 blkcg = css_to_blkcg(css);
1718 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1722 blkg = blkg_lookup(blkcg, q);
1725 if (!blkg_tryget(blkg))
1729 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1739 * blkcg_schedule_throttle - this task needs to check for throttling
1740 * @q: the request queue IO was submitted on
1741 * @use_memdelay: do we charge this to memory delay for PSI
1743 * This is called by the IO controller when we know there's delay accumulated
1744 * for the blkg for this task. We do not pass the blkg because there are places
1745 * we call this that may not have that information, the swapping code for
1746 * instance will only have a request_queue at that point. This set's the
1747 * notify_resume for the task to check and see if it requires throttling before
1748 * returning to user space.
1750 * We will only schedule once per syscall. You can call this over and over
1751 * again and it will only do the check once upon return to user space, and only
1752 * throttle once. If the task needs to be throttled again it'll need to be
1753 * re-set at the next time we see the task.
1755 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1757 if (unlikely(current->flags & PF_KTHREAD))
1760 if (!blk_get_queue(q))
1763 if (current->throttle_queue)
1764 blk_put_queue(current->throttle_queue);
1765 current->throttle_queue = q;
1767 current->use_memdelay = use_memdelay;
1768 set_notify_resume(current);
1772 * blkcg_add_delay - add delay to this blkg
1773 * @blkg: blkg of interest
1774 * @now: the current time in nanoseconds
1775 * @delta: how many nanoseconds of delay to add
1777 * Charge @delta to the blkg's current delay accumulation. This is used to
1778 * throttle tasks if an IO controller thinks we need more throttling.
1780 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1782 blkcg_scale_delay(blkg, now);
1783 atomic64_add(delta, &blkg->delay_nsec);
1786 module_param(blkcg_debug_stats, bool, 0644);
1787 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");