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 if (blkg->blkcg != &blkcg_root)
80 blk_exit_rl(blkg->q, &blkg->rl);
82 blkg_rwstat_exit(&blkg->stat_ios);
83 blkg_rwstat_exit(&blkg->stat_bytes);
88 * blkg_alloc - allocate a blkg
89 * @blkcg: block cgroup the new blkg is associated with
90 * @q: request_queue the new blkg is associated with
91 * @gfp_mask: allocation mask to use
93 * Allocate a new blkg assocating @blkcg and @q.
95 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
98 struct blkcg_gq *blkg;
101 /* alloc and init base part */
102 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
106 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
107 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
111 INIT_LIST_HEAD(&blkg->q_node);
113 atomic_set(&blkg->refcnt, 1);
115 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
116 if (blkcg != &blkcg_root) {
117 if (blk_init_rl(&blkg->rl, q, gfp_mask))
119 blkg->rl.blkg = blkg;
122 for (i = 0; i < BLKCG_MAX_POLS; i++) {
123 struct blkcg_policy *pol = blkcg_policy[i];
124 struct blkg_policy_data *pd;
126 if (!blkcg_policy_enabled(q, pol))
129 /* alloc per-policy data and attach it to blkg */
130 pd = pol->pd_alloc_fn(gfp_mask, q->node);
146 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
147 struct request_queue *q, bool update_hint)
149 struct blkcg_gq *blkg;
152 * Hint didn't match. Look up from the radix tree. Note that the
153 * hint can only be updated under queue_lock as otherwise @blkg
154 * could have already been removed from blkg_tree. The caller is
155 * responsible for grabbing queue_lock if @update_hint.
157 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
158 if (blkg && blkg->q == q) {
160 lockdep_assert_held(q->queue_lock);
161 rcu_assign_pointer(blkcg->blkg_hint, blkg);
168 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
171 * If @new_blkg is %NULL, this function tries to allocate a new one as
172 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
174 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
175 struct request_queue *q,
176 struct blkcg_gq *new_blkg)
178 struct blkcg_gq *blkg;
179 struct bdi_writeback_congested *wb_congested;
182 WARN_ON_ONCE(!rcu_read_lock_held());
183 lockdep_assert_held(q->queue_lock);
185 /* blkg holds a reference to blkcg */
186 if (!css_tryget_online(&blkcg->css)) {
191 wb_congested = wb_congested_get_create(q->backing_dev_info,
193 GFP_NOWAIT | __GFP_NOWARN);
201 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
202 if (unlikely(!new_blkg)) {
204 goto err_put_congested;
208 blkg->wb_congested = wb_congested;
211 if (blkcg_parent(blkcg)) {
212 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
213 if (WARN_ON_ONCE(!blkg->parent)) {
215 goto err_put_congested;
217 blkg_get(blkg->parent);
220 /* invoke per-policy init */
221 for (i = 0; i < BLKCG_MAX_POLS; i++) {
222 struct blkcg_policy *pol = blkcg_policy[i];
224 if (blkg->pd[i] && pol->pd_init_fn)
225 pol->pd_init_fn(blkg->pd[i]);
229 spin_lock(&blkcg->lock);
230 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
232 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
233 list_add(&blkg->q_node, &q->blkg_list);
235 for (i = 0; i < BLKCG_MAX_POLS; i++) {
236 struct blkcg_policy *pol = blkcg_policy[i];
238 if (blkg->pd[i] && pol->pd_online_fn)
239 pol->pd_online_fn(blkg->pd[i]);
243 spin_unlock(&blkcg->lock);
248 /* @blkg failed fully initialized, use the usual release path */
253 wb_congested_put(wb_congested);
255 css_put(&blkcg->css);
262 * __blkg_lookup_create - lookup blkg, try to create one if not there
263 * @blkcg: blkcg of interest
264 * @q: request_queue of interest
266 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
267 * create one. blkg creation is performed recursively from blkcg_root such
268 * that all non-root blkg's have access to the parent blkg. This function
269 * should be called under RCU read lock and @q->queue_lock.
271 * Returns the blkg or the closest blkg if blkg_create fails as it walks
274 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
275 struct request_queue *q)
277 struct blkcg_gq *blkg;
279 WARN_ON_ONCE(!rcu_read_lock_held());
280 lockdep_assert_held(q->queue_lock);
283 * This could be the first entry point of blkcg implementation and
284 * we shouldn't allow anything to go through for a bypassing queue.
286 if (unlikely(blk_queue_bypass(q)))
289 blkg = __blkg_lookup(blkcg, q, true);
294 * Create blkgs walking down from blkcg_root to @blkcg, so that all
295 * non-root blkgs have access to their parents. Returns the closest
296 * blkg to the intended blkg should blkg_create() fail.
299 struct blkcg *pos = blkcg;
300 struct blkcg *parent = blkcg_parent(blkcg);
301 struct blkcg_gq *ret_blkg = q->root_blkg;
304 blkg = __blkg_lookup(parent, q, false);
306 /* remember closest blkg */
311 parent = blkcg_parent(parent);
314 blkg = blkg_create(pos, q, NULL);
323 * blkg_lookup_create - find or create a blkg
324 * @blkcg: target block cgroup
325 * @q: target request_queue
327 * This looks up or creates the blkg representing the unique pair
328 * of the blkcg and the request_queue.
330 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
331 struct request_queue *q)
333 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
336 if (unlikely(!blkg)) {
337 spin_lock_irqsave(q->queue_lock, flags);
339 blkg = __blkg_lookup_create(blkcg, q);
341 spin_unlock_irqrestore(q->queue_lock, flags);
347 static void blkg_destroy(struct blkcg_gq *blkg)
349 struct blkcg *blkcg = blkg->blkcg;
350 struct blkcg_gq *parent = blkg->parent;
353 lockdep_assert_held(blkg->q->queue_lock);
354 lockdep_assert_held(&blkcg->lock);
356 /* Something wrong if we are trying to remove same group twice */
357 WARN_ON_ONCE(list_empty(&blkg->q_node));
358 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
360 for (i = 0; i < BLKCG_MAX_POLS; i++) {
361 struct blkcg_policy *pol = blkcg_policy[i];
363 if (blkg->pd[i] && pol->pd_offline_fn)
364 pol->pd_offline_fn(blkg->pd[i]);
368 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
369 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
372 blkg->online = false;
374 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
375 list_del_init(&blkg->q_node);
376 hlist_del_init_rcu(&blkg->blkcg_node);
379 * Both setting lookup hint to and clearing it from @blkg are done
380 * under queue_lock. If it's not pointing to @blkg now, it never
381 * will. Hint assignment itself can race safely.
383 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
384 rcu_assign_pointer(blkcg->blkg_hint, NULL);
387 * Put the reference taken at the time of creation so that when all
388 * queues are gone, group can be destroyed.
394 * blkg_destroy_all - destroy all blkgs associated with a request_queue
395 * @q: request_queue of interest
397 * Destroy all blkgs associated with @q.
399 static void blkg_destroy_all(struct request_queue *q)
401 struct blkcg_gq *blkg, *n;
403 lockdep_assert_held(q->queue_lock);
405 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
406 struct blkcg *blkcg = blkg->blkcg;
408 spin_lock(&blkcg->lock);
410 spin_unlock(&blkcg->lock);
414 q->root_rl.blkg = NULL;
418 * A group is RCU protected, but having an rcu lock does not mean that one
419 * can access all the fields of blkg and assume these are valid. For
420 * example, don't try to follow throtl_data and request queue links.
422 * Having a reference to blkg under an rcu allows accesses to only values
423 * local to groups like group stats and group rate limits.
425 void __blkg_release_rcu(struct rcu_head *rcu_head)
427 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
429 /* release the blkcg and parent blkg refs this blkg has been holding */
430 css_put(&blkg->blkcg->css);
432 blkg_put(blkg->parent);
434 wb_congested_put(blkg->wb_congested);
438 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
441 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
442 * because the root blkg uses @q->root_rl instead of its own rl.
444 struct request_list *__blk_queue_next_rl(struct request_list *rl,
445 struct request_queue *q)
447 struct list_head *ent;
448 struct blkcg_gq *blkg;
451 * Determine the current blkg list_head. The first entry is
452 * root_rl which is off @q->blkg_list and mapped to the head.
454 if (rl == &q->root_rl) {
456 /* There are no more block groups, hence no request lists */
460 blkg = container_of(rl, struct blkcg_gq, rl);
464 /* walk to the next list_head, skip root blkcg */
466 if (ent == &q->root_blkg->q_node)
468 if (ent == &q->blkg_list)
471 blkg = container_of(ent, struct blkcg_gq, q_node);
475 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
476 struct cftype *cftype, u64 val)
478 struct blkcg *blkcg = css_to_blkcg(css);
479 struct blkcg_gq *blkg;
482 mutex_lock(&blkcg_pol_mutex);
483 spin_lock_irq(&blkcg->lock);
486 * Note that stat reset is racy - it doesn't synchronize against
487 * stat updates. This is a debug feature which shouldn't exist
488 * anyway. If you get hit by a race, retry.
490 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
491 blkg_rwstat_reset(&blkg->stat_bytes);
492 blkg_rwstat_reset(&blkg->stat_ios);
494 for (i = 0; i < BLKCG_MAX_POLS; i++) {
495 struct blkcg_policy *pol = blkcg_policy[i];
497 if (blkg->pd[i] && pol->pd_reset_stats_fn)
498 pol->pd_reset_stats_fn(blkg->pd[i]);
502 spin_unlock_irq(&blkcg->lock);
503 mutex_unlock(&blkcg_pol_mutex);
507 const char *blkg_dev_name(struct blkcg_gq *blkg)
509 /* some drivers (floppy) instantiate a queue w/o disk registered */
510 if (blkg->q->backing_dev_info->dev)
511 return dev_name(blkg->q->backing_dev_info->dev);
514 EXPORT_SYMBOL_GPL(blkg_dev_name);
517 * blkcg_print_blkgs - helper for printing per-blkg data
518 * @sf: seq_file to print to
519 * @blkcg: blkcg of interest
520 * @prfill: fill function to print out a blkg
521 * @pol: policy in question
522 * @data: data to be passed to @prfill
523 * @show_total: to print out sum of prfill return values or not
525 * This function invokes @prfill on each blkg of @blkcg if pd for the
526 * policy specified by @pol exists. @prfill is invoked with @sf, the
527 * policy data and @data and the matching queue lock held. If @show_total
528 * is %true, the sum of the return values from @prfill is printed with
529 * "Total" label at the end.
531 * This is to be used to construct print functions for
532 * cftype->read_seq_string method.
534 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
535 u64 (*prfill)(struct seq_file *,
536 struct blkg_policy_data *, int),
537 const struct blkcg_policy *pol, int data,
540 struct blkcg_gq *blkg;
544 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
545 spin_lock_irq(blkg->q->queue_lock);
546 if (blkcg_policy_enabled(blkg->q, pol))
547 total += prfill(sf, blkg->pd[pol->plid], data);
548 spin_unlock_irq(blkg->q->queue_lock);
553 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
555 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
558 * __blkg_prfill_u64 - prfill helper for a single u64 value
559 * @sf: seq_file to print to
560 * @pd: policy private data of interest
563 * Print @v to @sf for the device assocaited with @pd.
565 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
567 const char *dname = blkg_dev_name(pd->blkg);
572 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
575 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
578 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
579 * @sf: seq_file to print to
580 * @pd: policy private data of interest
581 * @rwstat: rwstat to print
583 * Print @rwstat to @sf for the device assocaited with @pd.
585 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
586 const struct blkg_rwstat *rwstat)
588 static const char *rwstr[] = {
589 [BLKG_RWSTAT_READ] = "Read",
590 [BLKG_RWSTAT_WRITE] = "Write",
591 [BLKG_RWSTAT_SYNC] = "Sync",
592 [BLKG_RWSTAT_ASYNC] = "Async",
593 [BLKG_RWSTAT_DISCARD] = "Discard",
595 const char *dname = blkg_dev_name(pd->blkg);
602 for (i = 0; i < BLKG_RWSTAT_NR; i++)
603 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
604 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
606 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
607 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
608 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
609 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
612 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
615 * blkg_prfill_stat - prfill callback for blkg_stat
616 * @sf: seq_file to print to
617 * @pd: policy private data of interest
618 * @off: offset to the blkg_stat in @pd
620 * prfill callback for printing a blkg_stat.
622 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
624 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
626 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
629 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
630 * @sf: seq_file to print to
631 * @pd: policy private data of interest
632 * @off: offset to the blkg_rwstat in @pd
634 * prfill callback for printing a blkg_rwstat.
636 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
639 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
641 return __blkg_prfill_rwstat(sf, pd, &rwstat);
643 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
645 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
646 struct blkg_policy_data *pd, int off)
648 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
650 return __blkg_prfill_rwstat(sf, pd, &rwstat);
654 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
655 * @sf: seq_file to print to
658 * To be used as cftype->seq_show to print blkg->stat_bytes.
659 * cftype->private must be set to the blkcg_policy.
661 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
663 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
664 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
665 offsetof(struct blkcg_gq, stat_bytes), true);
668 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
671 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
672 * @sf: seq_file to print to
675 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
676 * must be set to the blkcg_policy.
678 int blkg_print_stat_ios(struct seq_file *sf, void *v)
680 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
681 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
682 offsetof(struct blkcg_gq, stat_ios), true);
685 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
687 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
688 struct blkg_policy_data *pd,
691 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
693 return __blkg_prfill_rwstat(sf, pd, &rwstat);
697 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
698 * @sf: seq_file to print to
701 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
703 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
704 blkg_prfill_rwstat_field_recursive,
705 (void *)seq_cft(sf)->private,
706 offsetof(struct blkcg_gq, stat_bytes), true);
709 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
712 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
713 * @sf: seq_file to print to
716 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
718 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
719 blkg_prfill_rwstat_field_recursive,
720 (void *)seq_cft(sf)->private,
721 offsetof(struct blkcg_gq, stat_ios), true);
724 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
727 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
728 * @blkg: blkg of interest
729 * @pol: blkcg_policy which contains the blkg_stat
730 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
732 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
733 * online descendants and their aux counts. The caller must be holding the
734 * queue lock for online tests.
736 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
737 * at @off bytes into @blkg's blkg_policy_data of the policy.
739 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
740 struct blkcg_policy *pol, int off)
742 struct blkcg_gq *pos_blkg;
743 struct cgroup_subsys_state *pos_css;
746 lockdep_assert_held(blkg->q->queue_lock);
749 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
750 struct blkg_stat *stat;
752 if (!pos_blkg->online)
756 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
758 stat = (void *)blkg + off;
760 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
766 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
769 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
770 * @blkg: blkg of interest
771 * @pol: blkcg_policy which contains the blkg_rwstat
772 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
774 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
775 * online descendants and their aux counts. The caller must be holding the
776 * queue lock for online tests.
778 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
779 * is at @off bytes into @blkg's blkg_policy_data of the policy.
781 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
782 struct blkcg_policy *pol, int off)
784 struct blkcg_gq *pos_blkg;
785 struct cgroup_subsys_state *pos_css;
786 struct blkg_rwstat sum = { };
789 lockdep_assert_held(blkg->q->queue_lock);
792 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
793 struct blkg_rwstat *rwstat;
795 if (!pos_blkg->online)
799 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
801 rwstat = (void *)pos_blkg + off;
803 for (i = 0; i < BLKG_RWSTAT_NR; i++)
804 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
805 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
812 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
814 /* Performs queue bypass and policy enabled checks then looks up blkg. */
815 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
816 const struct blkcg_policy *pol,
817 struct request_queue *q)
819 WARN_ON_ONCE(!rcu_read_lock_held());
820 lockdep_assert_held(q->queue_lock);
822 if (!blkcg_policy_enabled(q, pol))
823 return ERR_PTR(-EOPNOTSUPP);
826 * This could be the first entry point of blkcg implementation and
827 * we shouldn't allow anything to go through for a bypassing queue.
829 if (unlikely(blk_queue_bypass(q)))
830 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
832 return __blkg_lookup(blkcg, q, true /* update_hint */);
836 * blkg_conf_prep - parse and prepare for per-blkg config update
837 * @blkcg: target block cgroup
838 * @pol: target policy
839 * @input: input string
840 * @ctx: blkg_conf_ctx to be filled
842 * Parse per-blkg config update from @input and initialize @ctx with the
843 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
844 * part of @input following MAJ:MIN. This function returns with RCU read
845 * lock and queue lock held and must be paired with blkg_conf_finish().
847 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
848 char *input, struct blkg_conf_ctx *ctx)
849 __acquires(rcu) __acquires(disk->queue->queue_lock)
851 struct gendisk *disk;
852 struct request_queue *q;
853 struct blkcg_gq *blkg;
854 unsigned int major, minor;
855 int key_len, part, ret;
858 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
861 body = input + key_len;
864 body = skip_spaces(body);
866 disk = get_gendisk(MKDEV(major, minor), &part);
877 spin_lock_irq(q->queue_lock);
879 blkg = blkg_lookup_check(blkcg, pol, q);
889 * Create blkgs walking down from blkcg_root to @blkcg, so that all
890 * non-root blkgs have access to their parents.
893 struct blkcg *pos = blkcg;
894 struct blkcg *parent;
895 struct blkcg_gq *new_blkg;
897 parent = blkcg_parent(blkcg);
898 while (parent && !__blkg_lookup(parent, q, false)) {
900 parent = blkcg_parent(parent);
903 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
904 spin_unlock_irq(q->queue_lock);
907 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
908 if (unlikely(!new_blkg)) {
914 spin_lock_irq(q->queue_lock);
916 blkg = blkg_lookup_check(pos, pol, q);
925 blkg = blkg_create(pos, q, new_blkg);
926 if (unlikely(IS_ERR(blkg))) {
942 spin_unlock_irq(q->queue_lock);
945 put_disk_and_module(disk);
947 * If queue was bypassing, we should retry. Do so after a
948 * short msleep(). It isn't strictly necessary but queue
949 * can be bypassing for some time and it's always nice to
950 * avoid busy looping.
954 ret = restart_syscall();
958 EXPORT_SYMBOL_GPL(blkg_conf_prep);
961 * blkg_conf_finish - finish up per-blkg config update
962 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
964 * Finish up after per-blkg config update. This function must be paired
965 * with blkg_conf_prep().
967 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
968 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
970 spin_unlock_irq(ctx->disk->queue->queue_lock);
972 put_disk_and_module(ctx->disk);
974 EXPORT_SYMBOL_GPL(blkg_conf_finish);
976 static int blkcg_print_stat(struct seq_file *sf, void *v)
978 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
979 struct blkcg_gq *blkg;
983 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
986 struct blkg_rwstat rwstat;
987 u64 rbytes, wbytes, rios, wios, dbytes, dios;
988 size_t size = seq_get_buf(sf, &buf), off = 0;
990 bool has_stats = false;
992 dname = blkg_dev_name(blkg);
997 * Hooray string manipulation, count is the size written NOT
998 * INCLUDING THE \0, so size is now count+1 less than what we
999 * had before, but we want to start writing the next bit from
1000 * the \0 so we only add count to buf.
1002 off += scnprintf(buf+off, size-off, "%s ", dname);
1004 spin_lock_irq(blkg->q->queue_lock);
1006 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
1007 offsetof(struct blkcg_gq, stat_bytes));
1008 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
1009 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
1010 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
1012 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
1013 offsetof(struct blkcg_gq, stat_ios));
1014 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
1015 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
1016 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
1018 spin_unlock_irq(blkg->q->queue_lock);
1020 if (rbytes || wbytes || rios || wios) {
1022 off += scnprintf(buf+off, size-off,
1023 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1024 rbytes, wbytes, rios, wios,
1028 if (!blkcg_debug_stats)
1031 if (atomic_read(&blkg->use_delay)) {
1033 off += scnprintf(buf+off, size-off,
1034 " use_delay=%d delay_nsec=%llu",
1035 atomic_read(&blkg->use_delay),
1036 (unsigned long long)atomic64_read(&blkg->delay_nsec));
1039 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1040 struct blkcg_policy *pol = blkcg_policy[i];
1043 if (!blkg->pd[i] || !pol->pd_stat_fn)
1046 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1053 off += scnprintf(buf+off, size-off, "\n");
1054 seq_commit(sf, off);
1062 static struct cftype blkcg_files[] = {
1065 .flags = CFTYPE_NOT_ON_ROOT,
1066 .seq_show = blkcg_print_stat,
1071 static struct cftype blkcg_legacy_files[] = {
1073 .name = "reset_stats",
1074 .write_u64 = blkcg_reset_stats,
1080 * blkcg destruction is a three-stage process.
1082 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1083 * which offlines writeback. Here we tie the next stage of blkg destruction
1084 * to the completion of writeback associated with the blkcg. This lets us
1085 * avoid punting potentially large amounts of outstanding writeback to root
1086 * while maintaining any ongoing policies. The next stage is triggered when
1087 * the nr_cgwbs count goes to zero.
1089 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1090 * and handles the destruction of blkgs. Here the css reference held by
1091 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1092 * This work may occur in cgwb_release_workfn() on the cgwb_release
1093 * workqueue. Any submitted ios that fail to get the blkg ref will be
1094 * punted to the root_blkg.
1096 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1097 * This finally frees the blkcg.
1101 * blkcg_css_offline - cgroup css_offline callback
1102 * @css: css of interest
1104 * This function is called when @css is about to go away. Here the cgwbs are
1105 * offlined first and only once writeback associated with the blkcg has
1106 * finished do we start step 2 (see above).
1108 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1110 struct blkcg *blkcg = css_to_blkcg(css);
1112 /* this prevents anyone from attaching or migrating to this blkcg */
1113 wb_blkcg_offline(blkcg);
1115 /* put the base cgwb reference allowing step 2 to be triggered */
1116 blkcg_cgwb_put(blkcg);
1120 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1121 * @blkcg: blkcg of interest
1123 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1124 * is nested inside q lock, this function performs reverse double lock dancing.
1125 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1126 * blkcg_css_free to eventually be called.
1128 * This is the blkcg counterpart of ioc_release_fn().
1130 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1132 spin_lock_irq(&blkcg->lock);
1134 while (!hlist_empty(&blkcg->blkg_list)) {
1135 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1136 struct blkcg_gq, blkcg_node);
1137 struct request_queue *q = blkg->q;
1139 if (spin_trylock(q->queue_lock)) {
1141 spin_unlock(q->queue_lock);
1143 spin_unlock_irq(&blkcg->lock);
1145 spin_lock_irq(&blkcg->lock);
1149 spin_unlock_irq(&blkcg->lock);
1152 static void blkcg_css_free(struct cgroup_subsys_state *css)
1154 struct blkcg *blkcg = css_to_blkcg(css);
1157 mutex_lock(&blkcg_pol_mutex);
1159 list_del(&blkcg->all_blkcgs_node);
1161 for (i = 0; i < BLKCG_MAX_POLS; i++)
1163 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1165 mutex_unlock(&blkcg_pol_mutex);
1170 static struct cgroup_subsys_state *
1171 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1173 struct blkcg *blkcg;
1174 struct cgroup_subsys_state *ret;
1177 mutex_lock(&blkcg_pol_mutex);
1180 blkcg = &blkcg_root;
1182 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1184 ret = ERR_PTR(-ENOMEM);
1189 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1190 struct blkcg_policy *pol = blkcg_policy[i];
1191 struct blkcg_policy_data *cpd;
1194 * If the policy hasn't been attached yet, wait for it
1195 * to be attached before doing anything else. Otherwise,
1196 * check if the policy requires any specific per-cgroup
1197 * data: if it does, allocate and initialize it.
1199 if (!pol || !pol->cpd_alloc_fn)
1202 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1204 ret = ERR_PTR(-ENOMEM);
1207 blkcg->cpd[i] = cpd;
1210 if (pol->cpd_init_fn)
1211 pol->cpd_init_fn(cpd);
1214 spin_lock_init(&blkcg->lock);
1215 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1216 INIT_HLIST_HEAD(&blkcg->blkg_list);
1217 #ifdef CONFIG_CGROUP_WRITEBACK
1218 INIT_LIST_HEAD(&blkcg->cgwb_list);
1219 refcount_set(&blkcg->cgwb_refcnt, 1);
1221 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1223 mutex_unlock(&blkcg_pol_mutex);
1227 for (i--; i >= 0; i--)
1229 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1231 if (blkcg != &blkcg_root)
1234 mutex_unlock(&blkcg_pol_mutex);
1239 * blkcg_init_queue - initialize blkcg part of request queue
1240 * @q: request_queue to initialize
1242 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1243 * part of new request_queue @q.
1246 * 0 on success, -errno on failure.
1248 int blkcg_init_queue(struct request_queue *q)
1250 struct blkcg_gq *new_blkg, *blkg;
1254 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1258 preloaded = !radix_tree_preload(GFP_KERNEL);
1260 /* Make sure the root blkg exists. */
1262 spin_lock_irq(q->queue_lock);
1263 blkg = blkg_create(&blkcg_root, q, new_blkg);
1266 q->root_blkg = blkg;
1267 q->root_rl.blkg = blkg;
1268 spin_unlock_irq(q->queue_lock);
1272 radix_tree_preload_end();
1274 ret = blk_iolatency_init(q);
1276 spin_lock_irq(q->queue_lock);
1277 blkg_destroy_all(q);
1278 spin_unlock_irq(q->queue_lock);
1282 ret = blk_throtl_init(q);
1284 spin_lock_irq(q->queue_lock);
1285 blkg_destroy_all(q);
1286 spin_unlock_irq(q->queue_lock);
1291 spin_unlock_irq(q->queue_lock);
1294 radix_tree_preload_end();
1295 return PTR_ERR(blkg);
1299 * blkcg_drain_queue - drain blkcg part of request_queue
1300 * @q: request_queue to drain
1302 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1304 void blkcg_drain_queue(struct request_queue *q)
1306 lockdep_assert_held(q->queue_lock);
1309 * @q could be exiting and already have destroyed all blkgs as
1310 * indicated by NULL root_blkg. If so, don't confuse policies.
1315 blk_throtl_drain(q);
1319 * blkcg_exit_queue - exit and release blkcg part of request_queue
1320 * @q: request_queue being released
1322 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1324 void blkcg_exit_queue(struct request_queue *q)
1326 spin_lock_irq(q->queue_lock);
1327 blkg_destroy_all(q);
1328 spin_unlock_irq(q->queue_lock);
1334 * We cannot support shared io contexts, as we have no mean to support
1335 * two tasks with the same ioc in two different groups without major rework
1336 * of the main cic data structures. For now we allow a task to change
1337 * its cgroup only if it's the only owner of its ioc.
1339 static int blkcg_can_attach(struct cgroup_taskset *tset)
1341 struct task_struct *task;
1342 struct cgroup_subsys_state *dst_css;
1343 struct io_context *ioc;
1346 /* task_lock() is needed to avoid races with exit_io_context() */
1347 cgroup_taskset_for_each(task, dst_css, tset) {
1349 ioc = task->io_context;
1350 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1359 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1363 mutex_lock(&blkcg_pol_mutex);
1365 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1366 struct blkcg_policy *pol = blkcg_policy[i];
1367 struct blkcg *blkcg;
1369 if (!pol || !pol->cpd_bind_fn)
1372 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1373 if (blkcg->cpd[pol->plid])
1374 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1376 mutex_unlock(&blkcg_pol_mutex);
1379 static void blkcg_exit(struct task_struct *tsk)
1381 if (tsk->throttle_queue)
1382 blk_put_queue(tsk->throttle_queue);
1383 tsk->throttle_queue = NULL;
1386 struct cgroup_subsys io_cgrp_subsys = {
1387 .css_alloc = blkcg_css_alloc,
1388 .css_offline = blkcg_css_offline,
1389 .css_free = blkcg_css_free,
1390 .can_attach = blkcg_can_attach,
1392 .dfl_cftypes = blkcg_files,
1393 .legacy_cftypes = blkcg_legacy_files,
1394 .legacy_name = "blkio",
1398 * This ensures that, if available, memcg is automatically enabled
1399 * together on the default hierarchy so that the owner cgroup can
1400 * be retrieved from writeback pages.
1402 .depends_on = 1 << memory_cgrp_id,
1405 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1408 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1409 * @q: request_queue of interest
1410 * @pol: blkcg policy to activate
1412 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1413 * bypass mode to populate its blkgs with policy_data for @pol.
1415 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1416 * from IO path. Update of each blkg is protected by both queue and blkcg
1417 * locks so that holding either lock and testing blkcg_policy_enabled() is
1418 * always enough for dereferencing policy data.
1420 * The caller is responsible for synchronizing [de]activations and policy
1421 * [un]registerations. Returns 0 on success, -errno on failure.
1423 int blkcg_activate_policy(struct request_queue *q,
1424 const struct blkcg_policy *pol)
1426 struct blkg_policy_data *pd_prealloc = NULL;
1427 struct blkcg_gq *blkg;
1430 if (blkcg_policy_enabled(q, pol))
1434 blk_mq_freeze_queue(q);
1436 blk_queue_bypass_start(q);
1439 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1442 goto out_bypass_end;
1446 spin_lock_irq(q->queue_lock);
1448 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1449 struct blkg_policy_data *pd;
1451 if (blkg->pd[pol->plid])
1454 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1456 swap(pd, pd_prealloc);
1458 spin_unlock_irq(q->queue_lock);
1462 blkg->pd[pol->plid] = pd;
1464 pd->plid = pol->plid;
1465 if (pol->pd_init_fn)
1466 pol->pd_init_fn(pd);
1469 __set_bit(pol->plid, q->blkcg_pols);
1472 spin_unlock_irq(q->queue_lock);
1475 blk_mq_unfreeze_queue(q);
1477 blk_queue_bypass_end(q);
1479 pol->pd_free_fn(pd_prealloc);
1482 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1485 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1486 * @q: request_queue of interest
1487 * @pol: blkcg policy to deactivate
1489 * Deactivate @pol on @q. Follows the same synchronization rules as
1490 * blkcg_activate_policy().
1492 void blkcg_deactivate_policy(struct request_queue *q,
1493 const struct blkcg_policy *pol)
1495 struct blkcg_gq *blkg;
1497 if (!blkcg_policy_enabled(q, pol))
1501 blk_mq_freeze_queue(q);
1503 blk_queue_bypass_start(q);
1505 spin_lock_irq(q->queue_lock);
1507 __clear_bit(pol->plid, q->blkcg_pols);
1509 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1510 if (blkg->pd[pol->plid]) {
1511 if (pol->pd_offline_fn)
1512 pol->pd_offline_fn(blkg->pd[pol->plid]);
1513 pol->pd_free_fn(blkg->pd[pol->plid]);
1514 blkg->pd[pol->plid] = NULL;
1518 spin_unlock_irq(q->queue_lock);
1521 blk_mq_unfreeze_queue(q);
1523 blk_queue_bypass_end(q);
1525 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1528 * blkcg_policy_register - register a blkcg policy
1529 * @pol: blkcg policy to register
1531 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1532 * successful registration. Returns 0 on success and -errno on failure.
1534 int blkcg_policy_register(struct blkcg_policy *pol)
1536 struct blkcg *blkcg;
1539 mutex_lock(&blkcg_pol_register_mutex);
1540 mutex_lock(&blkcg_pol_mutex);
1542 /* find an empty slot */
1544 for (i = 0; i < BLKCG_MAX_POLS; i++)
1545 if (!blkcg_policy[i])
1547 if (i >= BLKCG_MAX_POLS)
1550 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1551 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1552 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1557 blkcg_policy[pol->plid] = pol;
1559 /* allocate and install cpd's */
1560 if (pol->cpd_alloc_fn) {
1561 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1562 struct blkcg_policy_data *cpd;
1564 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1568 blkcg->cpd[pol->plid] = cpd;
1570 cpd->plid = pol->plid;
1571 pol->cpd_init_fn(cpd);
1575 mutex_unlock(&blkcg_pol_mutex);
1577 /* everything is in place, add intf files for the new policy */
1578 if (pol->dfl_cftypes)
1579 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1581 if (pol->legacy_cftypes)
1582 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1583 pol->legacy_cftypes));
1584 mutex_unlock(&blkcg_pol_register_mutex);
1588 if (pol->cpd_free_fn) {
1589 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1590 if (blkcg->cpd[pol->plid]) {
1591 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1592 blkcg->cpd[pol->plid] = NULL;
1596 blkcg_policy[pol->plid] = NULL;
1598 mutex_unlock(&blkcg_pol_mutex);
1599 mutex_unlock(&blkcg_pol_register_mutex);
1602 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1605 * blkcg_policy_unregister - unregister a blkcg policy
1606 * @pol: blkcg policy to unregister
1608 * Undo blkcg_policy_register(@pol). Might sleep.
1610 void blkcg_policy_unregister(struct blkcg_policy *pol)
1612 struct blkcg *blkcg;
1614 mutex_lock(&blkcg_pol_register_mutex);
1616 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1619 /* kill the intf files first */
1620 if (pol->dfl_cftypes)
1621 cgroup_rm_cftypes(pol->dfl_cftypes);
1622 if (pol->legacy_cftypes)
1623 cgroup_rm_cftypes(pol->legacy_cftypes);
1625 /* remove cpds and unregister */
1626 mutex_lock(&blkcg_pol_mutex);
1628 if (pol->cpd_free_fn) {
1629 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1630 if (blkcg->cpd[pol->plid]) {
1631 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1632 blkcg->cpd[pol->plid] = NULL;
1636 blkcg_policy[pol->plid] = NULL;
1638 mutex_unlock(&blkcg_pol_mutex);
1640 mutex_unlock(&blkcg_pol_register_mutex);
1642 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1645 * Scale the accumulated delay based on how long it has been since we updated
1646 * the delay. We only call this when we are adding delay, in case it's been a
1647 * while since we added delay, and when we are checking to see if we need to
1648 * delay a task, to account for any delays that may have occurred.
1650 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1652 u64 old = atomic64_read(&blkg->delay_start);
1655 * We only want to scale down every second. The idea here is that we
1656 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1657 * time window. We only want to throttle tasks for recent delay that
1658 * has occurred, in 1 second time windows since that's the maximum
1659 * things can be throttled. We save the current delay window in
1660 * blkg->last_delay so we know what amount is still left to be charged
1661 * to the blkg from this point onward. blkg->last_use keeps track of
1662 * the use_delay counter. The idea is if we're unthrottling the blkg we
1663 * are ok with whatever is happening now, and we can take away more of
1664 * the accumulated delay as we've already throttled enough that
1665 * everybody is happy with their IO latencies.
1667 if (time_before64(old + NSEC_PER_SEC, now) &&
1668 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1669 u64 cur = atomic64_read(&blkg->delay_nsec);
1670 u64 sub = min_t(u64, blkg->last_delay, now - old);
1671 int cur_use = atomic_read(&blkg->use_delay);
1674 * We've been unthrottled, subtract a larger chunk of our
1675 * accumulated delay.
1677 if (cur_use < blkg->last_use)
1678 sub = max_t(u64, sub, blkg->last_delay >> 1);
1681 * This shouldn't happen, but handle it anyway. Our delay_nsec
1682 * should only ever be growing except here where we subtract out
1683 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1684 * rather not end up with negative numbers.
1686 if (unlikely(cur < sub)) {
1687 atomic64_set(&blkg->delay_nsec, 0);
1688 blkg->last_delay = 0;
1690 atomic64_sub(sub, &blkg->delay_nsec);
1691 blkg->last_delay = cur - sub;
1693 blkg->last_use = cur_use;
1698 * This is called when we want to actually walk up the hierarchy and check to
1699 * see if we need to throttle, and then actually throttle if there is some
1700 * accumulated delay. This should only be called upon return to user space so
1701 * we're not holding some lock that would induce a priority inversion.
1703 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1705 u64 now = ktime_to_ns(ktime_get());
1710 while (blkg->parent) {
1711 if (atomic_read(&blkg->use_delay)) {
1712 blkcg_scale_delay(blkg, now);
1713 delay_nsec = max_t(u64, delay_nsec,
1714 atomic64_read(&blkg->delay_nsec));
1716 blkg = blkg->parent;
1723 * Let's not sleep for all eternity if we've amassed a huge delay.
1724 * Swapping or metadata IO can accumulate 10's of seconds worth of
1725 * delay, and we want userspace to be able to do _something_ so cap the
1726 * delays at 1 second. If there's 10's of seconds worth of delay then
1727 * the tasks will be delayed for 1 second for every syscall.
1729 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1732 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1733 * that hasn't landed upstream yet. Once that stuff is in place we need
1734 * to do a psi_memstall_enter/leave if memdelay is set.
1737 exp = ktime_add_ns(now, delay_nsec);
1738 tok = io_schedule_prepare();
1740 __set_current_state(TASK_KILLABLE);
1741 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1743 } while (!fatal_signal_pending(current));
1744 io_schedule_finish(tok);
1748 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1750 * This is only called if we've been marked with set_notify_resume(). Obviously
1751 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1752 * check to see if current->throttle_queue is set and if not this doesn't do
1753 * anything. This should only ever be called by the resume code, it's not meant
1754 * to be called by people willy-nilly as it will actually do the work to
1755 * throttle the task if it is setup for throttling.
1757 void blkcg_maybe_throttle_current(void)
1759 struct request_queue *q = current->throttle_queue;
1760 struct cgroup_subsys_state *css;
1761 struct blkcg *blkcg;
1762 struct blkcg_gq *blkg;
1763 bool use_memdelay = current->use_memdelay;
1768 current->throttle_queue = NULL;
1769 current->use_memdelay = false;
1772 css = kthread_blkcg();
1774 blkcg = css_to_blkcg(css);
1776 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1780 blkg = blkg_lookup(blkcg, q);
1783 blkg = blkg_try_get(blkg);
1788 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1796 EXPORT_SYMBOL_GPL(blkcg_maybe_throttle_current);
1799 * blkcg_schedule_throttle - this task needs to check for throttling
1800 * @q - the request queue IO was submitted on
1801 * @use_memdelay - do we charge this to memory delay for PSI
1803 * This is called by the IO controller when we know there's delay accumulated
1804 * for the blkg for this task. We do not pass the blkg because there are places
1805 * we call this that may not have that information, the swapping code for
1806 * instance will only have a request_queue at that point. This set's the
1807 * notify_resume for the task to check and see if it requires throttling before
1808 * returning to user space.
1810 * We will only schedule once per syscall. You can call this over and over
1811 * again and it will only do the check once upon return to user space, and only
1812 * throttle once. If the task needs to be throttled again it'll need to be
1813 * re-set at the next time we see the task.
1815 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1817 if (unlikely(current->flags & PF_KTHREAD))
1820 if (!blk_get_queue(q))
1823 if (current->throttle_queue)
1824 blk_put_queue(current->throttle_queue);
1825 current->throttle_queue = q;
1827 current->use_memdelay = use_memdelay;
1828 set_notify_resume(current);
1830 EXPORT_SYMBOL_GPL(blkcg_schedule_throttle);
1833 * blkcg_add_delay - add delay to this blkg
1834 * @now - the current time in nanoseconds
1835 * @delta - how many nanoseconds of delay to add
1837 * Charge @delta to the blkg's current delay accumulation. This is used to
1838 * throttle tasks if an IO controller thinks we need more throttling.
1840 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1842 blkcg_scale_delay(blkg, now);
1843 atomic64_add(delta, &blkg->delay_nsec);
1845 EXPORT_SYMBOL_GPL(blkcg_add_delay);
1847 module_param(blkcg_debug_stats, bool, 0644);
1848 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");