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>
32 #include <linux/psi.h>
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;
49 EXPORT_SYMBOL_GPL(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 bool blkcg_debug_stats = false;
56 static struct workqueue_struct *blkcg_punt_bio_wq;
58 static bool blkcg_policy_enabled(struct request_queue *q,
59 const struct blkcg_policy *pol)
61 return pol && test_bit(pol->plid, q->blkcg_pols);
65 * blkg_free - free a blkg
68 * Free @blkg which may be partially allocated.
70 static void blkg_free(struct blkcg_gq *blkg)
77 for (i = 0; i < BLKCG_MAX_POLS; i++)
79 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
81 free_percpu(blkg->iostat_cpu);
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 WARN_ON(!bio_list_empty(&blkg->async_bios));
92 /* release the blkcg and parent blkg refs this blkg has been holding */
93 css_put(&blkg->blkcg->css);
95 blkg_put(blkg->parent);
100 * A group is RCU protected, but having an rcu lock does not mean that one
101 * can access all the fields of blkg and assume these are valid. For
102 * example, don't try to follow throtl_data and request queue links.
104 * Having a reference to blkg under an rcu allows accesses to only values
105 * local to groups like group stats and group rate limits.
107 static void blkg_release(struct percpu_ref *ref)
109 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
111 call_rcu(&blkg->rcu_head, __blkg_release);
114 static void blkg_async_bio_workfn(struct work_struct *work)
116 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
118 struct bio_list bios = BIO_EMPTY_LIST;
120 struct blk_plug plug;
121 bool need_plug = false;
123 /* as long as there are pending bios, @blkg can't go away */
124 spin_lock_bh(&blkg->async_bio_lock);
125 bio_list_merge(&bios, &blkg->async_bios);
126 bio_list_init(&blkg->async_bios);
127 spin_unlock_bh(&blkg->async_bio_lock);
129 /* start plug only when bio_list contains at least 2 bios */
130 if (bios.head && bios.head->bi_next) {
132 blk_start_plug(&plug);
134 while ((bio = bio_list_pop(&bios)))
137 blk_finish_plug(&plug);
141 * blkg_alloc - allocate a blkg
142 * @blkcg: block cgroup the new blkg is associated with
143 * @q: request_queue the new blkg is associated with
144 * @gfp_mask: allocation mask to use
146 * Allocate a new blkg assocating @blkcg and @q.
148 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
151 struct blkcg_gq *blkg;
154 /* alloc and init base part */
155 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
159 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
162 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
163 if (!blkg->iostat_cpu)
167 INIT_LIST_HEAD(&blkg->q_node);
168 spin_lock_init(&blkg->async_bio_lock);
169 bio_list_init(&blkg->async_bios);
170 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
173 u64_stats_init(&blkg->iostat.sync);
174 for_each_possible_cpu(cpu)
175 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
177 for (i = 0; i < BLKCG_MAX_POLS; i++) {
178 struct blkcg_policy *pol = blkcg_policy[i];
179 struct blkg_policy_data *pd;
181 if (!blkcg_policy_enabled(q, pol))
184 /* alloc per-policy data and attach it to blkg */
185 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
201 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
202 struct request_queue *q, bool update_hint)
204 struct blkcg_gq *blkg;
207 * Hint didn't match. Look up from the radix tree. Note that the
208 * hint can only be updated under queue_lock as otherwise @blkg
209 * could have already been removed from blkg_tree. The caller is
210 * responsible for grabbing queue_lock if @update_hint.
212 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
213 if (blkg && blkg->q == q) {
215 lockdep_assert_held(&q->queue_lock);
216 rcu_assign_pointer(blkcg->blkg_hint, blkg);
223 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
226 * If @new_blkg is %NULL, this function tries to allocate a new one as
227 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
229 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
230 struct request_queue *q,
231 struct blkcg_gq *new_blkg)
233 struct blkcg_gq *blkg;
236 WARN_ON_ONCE(!rcu_read_lock_held());
237 lockdep_assert_held(&q->queue_lock);
239 /* request_queue is dying, do not create/recreate a blkg */
240 if (blk_queue_dying(q)) {
245 /* blkg holds a reference to blkcg */
246 if (!css_tryget_online(&blkcg->css)) {
253 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
254 if (unlikely(!new_blkg)) {
262 if (blkcg_parent(blkcg)) {
263 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
264 if (WARN_ON_ONCE(!blkg->parent)) {
268 blkg_get(blkg->parent);
271 /* invoke per-policy init */
272 for (i = 0; i < BLKCG_MAX_POLS; i++) {
273 struct blkcg_policy *pol = blkcg_policy[i];
275 if (blkg->pd[i] && pol->pd_init_fn)
276 pol->pd_init_fn(blkg->pd[i]);
280 spin_lock(&blkcg->lock);
281 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
283 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
284 list_add(&blkg->q_node, &q->blkg_list);
286 for (i = 0; i < BLKCG_MAX_POLS; i++) {
287 struct blkcg_policy *pol = blkcg_policy[i];
289 if (blkg->pd[i] && pol->pd_online_fn)
290 pol->pd_online_fn(blkg->pd[i]);
294 spin_unlock(&blkcg->lock);
299 /* @blkg failed fully initialized, use the usual release path */
304 css_put(&blkcg->css);
311 * blkg_lookup_create - lookup blkg, try to create one if not there
312 * @blkcg: blkcg of interest
313 * @q: request_queue of interest
315 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
316 * create one. blkg creation is performed recursively from blkcg_root such
317 * that all non-root blkg's have access to the parent blkg. This function
318 * should be called under RCU read lock and takes @q->queue_lock.
320 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
323 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
324 struct request_queue *q)
326 struct blkcg_gq *blkg;
329 WARN_ON_ONCE(!rcu_read_lock_held());
331 blkg = blkg_lookup(blkcg, q);
335 spin_lock_irqsave(&q->queue_lock, flags);
336 blkg = __blkg_lookup(blkcg, q, true);
341 * Create blkgs walking down from blkcg_root to @blkcg, so that all
342 * non-root blkgs have access to their parents. Returns the closest
343 * blkg to the intended blkg should blkg_create() fail.
346 struct blkcg *pos = blkcg;
347 struct blkcg *parent = blkcg_parent(blkcg);
348 struct blkcg_gq *ret_blkg = q->root_blkg;
351 blkg = __blkg_lookup(parent, q, false);
353 /* remember closest blkg */
358 parent = blkcg_parent(parent);
361 blkg = blkg_create(pos, q, NULL);
371 spin_unlock_irqrestore(&q->queue_lock, flags);
375 static void blkg_destroy(struct blkcg_gq *blkg)
377 struct blkcg *blkcg = blkg->blkcg;
380 lockdep_assert_held(&blkg->q->queue_lock);
381 lockdep_assert_held(&blkcg->lock);
383 /* Something wrong if we are trying to remove same group twice */
384 WARN_ON_ONCE(list_empty(&blkg->q_node));
385 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
387 for (i = 0; i < BLKCG_MAX_POLS; i++) {
388 struct blkcg_policy *pol = blkcg_policy[i];
390 if (blkg->pd[i] && pol->pd_offline_fn)
391 pol->pd_offline_fn(blkg->pd[i]);
394 blkg->online = false;
396 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
397 list_del_init(&blkg->q_node);
398 hlist_del_init_rcu(&blkg->blkcg_node);
401 * Both setting lookup hint to and clearing it from @blkg are done
402 * under queue_lock. If it's not pointing to @blkg now, it never
403 * will. Hint assignment itself can race safely.
405 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
406 rcu_assign_pointer(blkcg->blkg_hint, NULL);
409 * Put the reference taken at the time of creation so that when all
410 * queues are gone, group can be destroyed.
412 percpu_ref_kill(&blkg->refcnt);
416 * blkg_destroy_all - destroy all blkgs associated with a request_queue
417 * @q: request_queue of interest
419 * Destroy all blkgs associated with @q.
421 static void blkg_destroy_all(struct request_queue *q)
423 struct blkcg_gq *blkg, *n;
425 spin_lock_irq(&q->queue_lock);
426 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
427 struct blkcg *blkcg = blkg->blkcg;
429 spin_lock(&blkcg->lock);
431 spin_unlock(&blkcg->lock);
435 spin_unlock_irq(&q->queue_lock);
438 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
439 struct cftype *cftype, u64 val)
441 struct blkcg *blkcg = css_to_blkcg(css);
442 struct blkcg_gq *blkg;
445 mutex_lock(&blkcg_pol_mutex);
446 spin_lock_irq(&blkcg->lock);
449 * Note that stat reset is racy - it doesn't synchronize against
450 * stat updates. This is a debug feature which shouldn't exist
451 * anyway. If you get hit by a race, retry.
453 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
454 for_each_possible_cpu(cpu) {
455 struct blkg_iostat_set *bis =
456 per_cpu_ptr(blkg->iostat_cpu, cpu);
457 memset(bis, 0, sizeof(*bis));
459 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
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 bdi_dev_name(blkg->q->backing_dev_info);
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);
543 /* Performs queue bypass and policy enabled checks then looks up blkg. */
544 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
545 const struct blkcg_policy *pol,
546 struct request_queue *q)
548 WARN_ON_ONCE(!rcu_read_lock_held());
549 lockdep_assert_held(&q->queue_lock);
551 if (!blkcg_policy_enabled(q, pol))
552 return ERR_PTR(-EOPNOTSUPP);
553 return __blkg_lookup(blkcg, q, true /* update_hint */);
557 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
558 * @inputp: input string pointer
560 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
561 * from @input and get and return the matching bdev. *@inputp is
562 * updated to point past the device node prefix. Returns an ERR_PTR()
565 * Use this function iff blkg_conf_prep() can't be used for some reason.
567 struct block_device *blkcg_conf_open_bdev(char **inputp)
569 char *input = *inputp;
570 unsigned int major, minor;
571 struct block_device *bdev;
574 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
575 return ERR_PTR(-EINVAL);
578 if (!isspace(*input))
579 return ERR_PTR(-EINVAL);
580 input = skip_spaces(input);
582 bdev = blkdev_get_no_open(MKDEV(major, minor));
584 return ERR_PTR(-ENODEV);
585 if (bdev_is_partition(bdev)) {
586 blkdev_put_no_open(bdev);
587 return ERR_PTR(-ENODEV);
595 * blkg_conf_prep - parse and prepare for per-blkg config update
596 * @blkcg: target block cgroup
597 * @pol: target policy
598 * @input: input string
599 * @ctx: blkg_conf_ctx to be filled
601 * Parse per-blkg config update from @input and initialize @ctx with the
602 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
603 * part of @input following MAJ:MIN. This function returns with RCU read
604 * lock and queue lock held and must be paired with blkg_conf_finish().
606 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
607 char *input, struct blkg_conf_ctx *ctx)
608 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
610 struct block_device *bdev;
611 struct request_queue *q;
612 struct blkcg_gq *blkg;
615 bdev = blkcg_conf_open_bdev(&input);
617 return PTR_ERR(bdev);
619 q = bdev->bd_disk->queue;
622 spin_lock_irq(&q->queue_lock);
624 blkg = blkg_lookup_check(blkcg, pol, q);
634 * Create blkgs walking down from blkcg_root to @blkcg, so that all
635 * non-root blkgs have access to their parents.
638 struct blkcg *pos = blkcg;
639 struct blkcg *parent;
640 struct blkcg_gq *new_blkg;
642 parent = blkcg_parent(blkcg);
643 while (parent && !__blkg_lookup(parent, q, false)) {
645 parent = blkcg_parent(parent);
648 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
649 spin_unlock_irq(&q->queue_lock);
652 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
653 if (unlikely(!new_blkg)) {
658 if (radix_tree_preload(GFP_KERNEL)) {
665 spin_lock_irq(&q->queue_lock);
667 blkg = blkg_lookup_check(pos, pol, q);
677 blkg = blkg_create(pos, q, new_blkg);
684 radix_tree_preload_end();
696 radix_tree_preload_end();
698 spin_unlock_irq(&q->queue_lock);
701 blkdev_put_no_open(bdev);
703 * If queue was bypassing, we should retry. Do so after a
704 * short msleep(). It isn't strictly necessary but queue
705 * can be bypassing for some time and it's always nice to
706 * avoid busy looping.
710 ret = restart_syscall();
714 EXPORT_SYMBOL_GPL(blkg_conf_prep);
717 * blkg_conf_finish - finish up per-blkg config update
718 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
720 * Finish up after per-blkg config update. This function must be paired
721 * with blkg_conf_prep().
723 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
724 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
726 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
728 blkdev_put_no_open(ctx->bdev);
730 EXPORT_SYMBOL_GPL(blkg_conf_finish);
732 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
736 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
737 dst->bytes[i] = src->bytes[i];
738 dst->ios[i] = src->ios[i];
742 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
746 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
747 dst->bytes[i] += src->bytes[i];
748 dst->ios[i] += src->ios[i];
752 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
756 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
757 dst->bytes[i] -= src->bytes[i];
758 dst->ios[i] -= src->ios[i];
762 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
764 struct blkcg *blkcg = css_to_blkcg(css);
765 struct blkcg_gq *blkg;
767 /* Root-level stats are sourced from system-wide IO stats */
768 if (!cgroup_parent(css->cgroup))
773 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
774 struct blkcg_gq *parent = blkg->parent;
775 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
776 struct blkg_iostat cur, delta;
779 /* fetch the current per-cpu values */
781 seq = u64_stats_fetch_begin(&bisc->sync);
782 blkg_iostat_set(&cur, &bisc->cur);
783 } while (u64_stats_fetch_retry(&bisc->sync, seq));
785 /* propagate percpu delta to global */
786 u64_stats_update_begin(&blkg->iostat.sync);
787 blkg_iostat_set(&delta, &cur);
788 blkg_iostat_sub(&delta, &bisc->last);
789 blkg_iostat_add(&blkg->iostat.cur, &delta);
790 blkg_iostat_add(&bisc->last, &delta);
791 u64_stats_update_end(&blkg->iostat.sync);
793 /* propagate global delta to parent (unless that's root) */
794 if (parent && parent->parent) {
795 u64_stats_update_begin(&parent->iostat.sync);
796 blkg_iostat_set(&delta, &blkg->iostat.cur);
797 blkg_iostat_sub(&delta, &blkg->iostat.last);
798 blkg_iostat_add(&parent->iostat.cur, &delta);
799 blkg_iostat_add(&blkg->iostat.last, &delta);
800 u64_stats_update_end(&parent->iostat.sync);
808 * We source root cgroup stats from the system-wide stats to avoid
809 * tracking the same information twice and incurring overhead when no
810 * cgroups are defined. For that reason, cgroup_rstat_flush in
811 * blkcg_print_stat does not actually fill out the iostat in the root
814 * However, we would like to re-use the printing code between the root and
815 * non-root cgroups to the extent possible. For that reason, we simulate
816 * flushing the root cgroup's stats by explicitly filling in the iostat
817 * with disk level statistics.
819 static void blkcg_fill_root_iostats(void)
821 struct class_dev_iter iter;
824 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
825 while ((dev = class_dev_iter_next(&iter))) {
826 struct block_device *bdev = dev_to_bdev(dev);
827 struct blkcg_gq *blkg =
828 blk_queue_root_blkg(bdev->bd_disk->queue);
829 struct blkg_iostat tmp;
832 memset(&tmp, 0, sizeof(tmp));
833 for_each_possible_cpu(cpu) {
834 struct disk_stats *cpu_dkstats;
836 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
837 tmp.ios[BLKG_IOSTAT_READ] +=
838 cpu_dkstats->ios[STAT_READ];
839 tmp.ios[BLKG_IOSTAT_WRITE] +=
840 cpu_dkstats->ios[STAT_WRITE];
841 tmp.ios[BLKG_IOSTAT_DISCARD] +=
842 cpu_dkstats->ios[STAT_DISCARD];
843 // convert sectors to bytes
844 tmp.bytes[BLKG_IOSTAT_READ] +=
845 cpu_dkstats->sectors[STAT_READ] << 9;
846 tmp.bytes[BLKG_IOSTAT_WRITE] +=
847 cpu_dkstats->sectors[STAT_WRITE] << 9;
848 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
849 cpu_dkstats->sectors[STAT_DISCARD] << 9;
851 u64_stats_update_begin(&blkg->iostat.sync);
852 blkg_iostat_set(&blkg->iostat.cur, &tmp);
853 u64_stats_update_end(&blkg->iostat.sync);
858 static int blkcg_print_stat(struct seq_file *sf, void *v)
860 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
861 struct blkcg_gq *blkg;
863 if (!seq_css(sf)->parent)
864 blkcg_fill_root_iostats();
866 cgroup_rstat_flush(blkcg->css.cgroup);
870 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
871 struct blkg_iostat_set *bis = &blkg->iostat;
874 u64 rbytes, wbytes, rios, wios, dbytes, dios;
875 size_t size = seq_get_buf(sf, &buf), off = 0;
877 bool has_stats = false;
880 spin_lock_irq(&blkg->q->queue_lock);
885 dname = blkg_dev_name(blkg);
890 * Hooray string manipulation, count is the size written NOT
891 * INCLUDING THE \0, so size is now count+1 less than what we
892 * had before, but we want to start writing the next bit from
893 * the \0 so we only add count to buf.
895 off += scnprintf(buf+off, size-off, "%s ", dname);
898 seq = u64_stats_fetch_begin(&bis->sync);
900 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
901 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
902 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
903 rios = bis->cur.ios[BLKG_IOSTAT_READ];
904 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
905 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
906 } while (u64_stats_fetch_retry(&bis->sync, seq));
908 if (rbytes || wbytes || rios || wios) {
910 off += scnprintf(buf+off, size-off,
911 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
912 rbytes, wbytes, rios, wios,
916 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
918 off += scnprintf(buf+off, size-off,
919 " use_delay=%d delay_nsec=%llu",
920 atomic_read(&blkg->use_delay),
921 (unsigned long long)atomic64_read(&blkg->delay_nsec));
924 for (i = 0; i < BLKCG_MAX_POLS; i++) {
925 struct blkcg_policy *pol = blkcg_policy[i];
928 if (!blkg->pd[i] || !pol->pd_stat_fn)
931 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
938 if (off < size - 1) {
939 off += scnprintf(buf+off, size-off, "\n");
946 spin_unlock_irq(&blkg->q->queue_lock);
953 static struct cftype blkcg_files[] = {
956 .seq_show = blkcg_print_stat,
961 static struct cftype blkcg_legacy_files[] = {
963 .name = "reset_stats",
964 .write_u64 = blkcg_reset_stats,
970 * blkcg destruction is a three-stage process.
972 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
973 * which offlines writeback. Here we tie the next stage of blkg destruction
974 * to the completion of writeback associated with the blkcg. This lets us
975 * avoid punting potentially large amounts of outstanding writeback to root
976 * while maintaining any ongoing policies. The next stage is triggered when
977 * the nr_cgwbs count goes to zero.
979 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
980 * and handles the destruction of blkgs. Here the css reference held by
981 * the blkg is put back eventually allowing blkcg_css_free() to be called.
982 * This work may occur in cgwb_release_workfn() on the cgwb_release
983 * workqueue. Any submitted ios that fail to get the blkg ref will be
984 * punted to the root_blkg.
986 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
987 * This finally frees the blkcg.
991 * blkcg_css_offline - cgroup css_offline callback
992 * @css: css of interest
994 * This function is called when @css is about to go away. Here the cgwbs are
995 * offlined first and only once writeback associated with the blkcg has
996 * finished do we start step 2 (see above).
998 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1000 struct blkcg *blkcg = css_to_blkcg(css);
1002 /* this prevents anyone from attaching or migrating to this blkcg */
1003 wb_blkcg_offline(blkcg);
1005 /* put the base online pin allowing step 2 to be triggered */
1006 blkcg_unpin_online(blkcg);
1010 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1011 * @blkcg: blkcg of interest
1013 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1014 * is nested inside q lock, this function performs reverse double lock dancing.
1015 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1016 * blkcg_css_free to eventually be called.
1018 * This is the blkcg counterpart of ioc_release_fn().
1020 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1024 spin_lock_irq(&blkcg->lock);
1026 while (!hlist_empty(&blkcg->blkg_list)) {
1027 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1028 struct blkcg_gq, blkcg_node);
1029 struct request_queue *q = blkg->q;
1031 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1033 * Given that the system can accumulate a huge number
1034 * of blkgs in pathological cases, check to see if we
1035 * need to rescheduling to avoid softlockup.
1037 spin_unlock_irq(&blkcg->lock);
1039 spin_lock_irq(&blkcg->lock);
1044 spin_unlock(&q->queue_lock);
1047 spin_unlock_irq(&blkcg->lock);
1050 static void blkcg_css_free(struct cgroup_subsys_state *css)
1052 struct blkcg *blkcg = css_to_blkcg(css);
1055 mutex_lock(&blkcg_pol_mutex);
1057 list_del(&blkcg->all_blkcgs_node);
1059 for (i = 0; i < BLKCG_MAX_POLS; i++)
1061 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1063 mutex_unlock(&blkcg_pol_mutex);
1068 static struct cgroup_subsys_state *
1069 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1071 struct blkcg *blkcg;
1072 struct cgroup_subsys_state *ret;
1075 mutex_lock(&blkcg_pol_mutex);
1078 blkcg = &blkcg_root;
1080 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1082 ret = ERR_PTR(-ENOMEM);
1087 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1088 struct blkcg_policy *pol = blkcg_policy[i];
1089 struct blkcg_policy_data *cpd;
1092 * If the policy hasn't been attached yet, wait for it
1093 * to be attached before doing anything else. Otherwise,
1094 * check if the policy requires any specific per-cgroup
1095 * data: if it does, allocate and initialize it.
1097 if (!pol || !pol->cpd_alloc_fn)
1100 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1102 ret = ERR_PTR(-ENOMEM);
1105 blkcg->cpd[i] = cpd;
1108 if (pol->cpd_init_fn)
1109 pol->cpd_init_fn(cpd);
1112 spin_lock_init(&blkcg->lock);
1113 refcount_set(&blkcg->online_pin, 1);
1114 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1115 INIT_HLIST_HEAD(&blkcg->blkg_list);
1116 #ifdef CONFIG_CGROUP_WRITEBACK
1117 INIT_LIST_HEAD(&blkcg->cgwb_list);
1119 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1121 mutex_unlock(&blkcg_pol_mutex);
1125 for (i--; i >= 0; i--)
1127 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1129 if (blkcg != &blkcg_root)
1132 mutex_unlock(&blkcg_pol_mutex);
1136 static int blkcg_css_online(struct cgroup_subsys_state *css)
1138 struct blkcg *blkcg = css_to_blkcg(css);
1139 struct blkcg *parent = blkcg_parent(blkcg);
1142 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1143 * don't go offline while cgwbs are still active on them. Pin the
1144 * parent so that offline always happens towards the root.
1147 blkcg_pin_online(parent);
1152 * blkcg_init_queue - initialize blkcg part of request queue
1153 * @q: request_queue to initialize
1155 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1156 * part of new request_queue @q.
1159 * 0 on success, -errno on failure.
1161 int blkcg_init_queue(struct request_queue *q)
1163 struct blkcg_gq *new_blkg, *blkg;
1167 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1171 preloaded = !radix_tree_preload(GFP_KERNEL);
1173 /* Make sure the root blkg exists. */
1175 spin_lock_irq(&q->queue_lock);
1176 blkg = blkg_create(&blkcg_root, q, new_blkg);
1179 q->root_blkg = blkg;
1180 spin_unlock_irq(&q->queue_lock);
1184 radix_tree_preload_end();
1186 ret = blk_throtl_init(q);
1188 goto err_destroy_all;
1190 ret = blk_iolatency_init(q);
1193 goto err_destroy_all;
1198 blkg_destroy_all(q);
1201 spin_unlock_irq(&q->queue_lock);
1204 radix_tree_preload_end();
1205 return PTR_ERR(blkg);
1209 * blkcg_exit_queue - exit and release blkcg part of request_queue
1210 * @q: request_queue being released
1212 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1214 void blkcg_exit_queue(struct request_queue *q)
1216 blkg_destroy_all(q);
1221 * We cannot support shared io contexts, as we have no mean to support
1222 * two tasks with the same ioc in two different groups without major rework
1223 * of the main cic data structures. For now we allow a task to change
1224 * its cgroup only if it's the only owner of its ioc.
1226 static int blkcg_can_attach(struct cgroup_taskset *tset)
1228 struct task_struct *task;
1229 struct cgroup_subsys_state *dst_css;
1230 struct io_context *ioc;
1233 /* task_lock() is needed to avoid races with exit_io_context() */
1234 cgroup_taskset_for_each(task, dst_css, tset) {
1236 ioc = task->io_context;
1237 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1246 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1250 mutex_lock(&blkcg_pol_mutex);
1252 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1253 struct blkcg_policy *pol = blkcg_policy[i];
1254 struct blkcg *blkcg;
1256 if (!pol || !pol->cpd_bind_fn)
1259 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1260 if (blkcg->cpd[pol->plid])
1261 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1263 mutex_unlock(&blkcg_pol_mutex);
1266 static void blkcg_exit(struct task_struct *tsk)
1268 if (tsk->throttle_queue)
1269 blk_put_queue(tsk->throttle_queue);
1270 tsk->throttle_queue = NULL;
1273 struct cgroup_subsys io_cgrp_subsys = {
1274 .css_alloc = blkcg_css_alloc,
1275 .css_online = blkcg_css_online,
1276 .css_offline = blkcg_css_offline,
1277 .css_free = blkcg_css_free,
1278 .can_attach = blkcg_can_attach,
1279 .css_rstat_flush = blkcg_rstat_flush,
1281 .dfl_cftypes = blkcg_files,
1282 .legacy_cftypes = blkcg_legacy_files,
1283 .legacy_name = "blkio",
1287 * This ensures that, if available, memcg is automatically enabled
1288 * together on the default hierarchy so that the owner cgroup can
1289 * be retrieved from writeback pages.
1291 .depends_on = 1 << memory_cgrp_id,
1294 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1297 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1298 * @q: request_queue of interest
1299 * @pol: blkcg policy to activate
1301 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1302 * bypass mode to populate its blkgs with policy_data for @pol.
1304 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1305 * from IO path. Update of each blkg is protected by both queue and blkcg
1306 * locks so that holding either lock and testing blkcg_policy_enabled() is
1307 * always enough for dereferencing policy data.
1309 * The caller is responsible for synchronizing [de]activations and policy
1310 * [un]registerations. Returns 0 on success, -errno on failure.
1312 int blkcg_activate_policy(struct request_queue *q,
1313 const struct blkcg_policy *pol)
1315 struct blkg_policy_data *pd_prealloc = NULL;
1316 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1319 if (blkcg_policy_enabled(q, pol))
1323 blk_mq_freeze_queue(q);
1325 spin_lock_irq(&q->queue_lock);
1327 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1328 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1329 struct blkg_policy_data *pd;
1331 if (blkg->pd[pol->plid])
1334 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1335 if (blkg == pinned_blkg) {
1339 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1345 * GFP_NOWAIT failed. Free the existing one and
1346 * prealloc for @blkg w/ GFP_KERNEL.
1349 blkg_put(pinned_blkg);
1353 spin_unlock_irq(&q->queue_lock);
1356 pol->pd_free_fn(pd_prealloc);
1357 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1365 blkg->pd[pol->plid] = pd;
1367 pd->plid = pol->plid;
1370 /* all allocated, init in the same order */
1371 if (pol->pd_init_fn)
1372 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1373 pol->pd_init_fn(blkg->pd[pol->plid]);
1375 __set_bit(pol->plid, q->blkcg_pols);
1378 spin_unlock_irq(&q->queue_lock);
1381 blk_mq_unfreeze_queue(q);
1383 blkg_put(pinned_blkg);
1385 pol->pd_free_fn(pd_prealloc);
1389 /* alloc failed, nothing's initialized yet, free everything */
1390 spin_lock_irq(&q->queue_lock);
1391 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1392 if (blkg->pd[pol->plid]) {
1393 pol->pd_free_fn(blkg->pd[pol->plid]);
1394 blkg->pd[pol->plid] = NULL;
1397 spin_unlock_irq(&q->queue_lock);
1401 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1404 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1405 * @q: request_queue of interest
1406 * @pol: blkcg policy to deactivate
1408 * Deactivate @pol on @q. Follows the same synchronization rules as
1409 * blkcg_activate_policy().
1411 void blkcg_deactivate_policy(struct request_queue *q,
1412 const struct blkcg_policy *pol)
1414 struct blkcg_gq *blkg;
1416 if (!blkcg_policy_enabled(q, pol))
1420 blk_mq_freeze_queue(q);
1422 spin_lock_irq(&q->queue_lock);
1424 __clear_bit(pol->plid, q->blkcg_pols);
1426 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1427 if (blkg->pd[pol->plid]) {
1428 if (pol->pd_offline_fn)
1429 pol->pd_offline_fn(blkg->pd[pol->plid]);
1430 pol->pd_free_fn(blkg->pd[pol->plid]);
1431 blkg->pd[pol->plid] = NULL;
1435 spin_unlock_irq(&q->queue_lock);
1438 blk_mq_unfreeze_queue(q);
1440 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1443 * blkcg_policy_register - register a blkcg policy
1444 * @pol: blkcg policy to register
1446 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1447 * successful registration. Returns 0 on success and -errno on failure.
1449 int blkcg_policy_register(struct blkcg_policy *pol)
1451 struct blkcg *blkcg;
1454 mutex_lock(&blkcg_pol_register_mutex);
1455 mutex_lock(&blkcg_pol_mutex);
1457 /* find an empty slot */
1459 for (i = 0; i < BLKCG_MAX_POLS; i++)
1460 if (!blkcg_policy[i])
1462 if (i >= BLKCG_MAX_POLS) {
1463 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1467 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1468 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1469 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1474 blkcg_policy[pol->plid] = pol;
1476 /* allocate and install cpd's */
1477 if (pol->cpd_alloc_fn) {
1478 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1479 struct blkcg_policy_data *cpd;
1481 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1485 blkcg->cpd[pol->plid] = cpd;
1487 cpd->plid = pol->plid;
1488 if (pol->cpd_init_fn)
1489 pol->cpd_init_fn(cpd);
1493 mutex_unlock(&blkcg_pol_mutex);
1495 /* everything is in place, add intf files for the new policy */
1496 if (pol->dfl_cftypes)
1497 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1499 if (pol->legacy_cftypes)
1500 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1501 pol->legacy_cftypes));
1502 mutex_unlock(&blkcg_pol_register_mutex);
1506 if (pol->cpd_free_fn) {
1507 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1508 if (blkcg->cpd[pol->plid]) {
1509 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1510 blkcg->cpd[pol->plid] = NULL;
1514 blkcg_policy[pol->plid] = NULL;
1516 mutex_unlock(&blkcg_pol_mutex);
1517 mutex_unlock(&blkcg_pol_register_mutex);
1520 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1523 * blkcg_policy_unregister - unregister a blkcg policy
1524 * @pol: blkcg policy to unregister
1526 * Undo blkcg_policy_register(@pol). Might sleep.
1528 void blkcg_policy_unregister(struct blkcg_policy *pol)
1530 struct blkcg *blkcg;
1532 mutex_lock(&blkcg_pol_register_mutex);
1534 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1537 /* kill the intf files first */
1538 if (pol->dfl_cftypes)
1539 cgroup_rm_cftypes(pol->dfl_cftypes);
1540 if (pol->legacy_cftypes)
1541 cgroup_rm_cftypes(pol->legacy_cftypes);
1543 /* remove cpds and unregister */
1544 mutex_lock(&blkcg_pol_mutex);
1546 if (pol->cpd_free_fn) {
1547 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1548 if (blkcg->cpd[pol->plid]) {
1549 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1550 blkcg->cpd[pol->plid] = NULL;
1554 blkcg_policy[pol->plid] = NULL;
1556 mutex_unlock(&blkcg_pol_mutex);
1558 mutex_unlock(&blkcg_pol_register_mutex);
1560 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1562 bool __blkcg_punt_bio_submit(struct bio *bio)
1564 struct blkcg_gq *blkg = bio->bi_blkg;
1566 /* consume the flag first */
1567 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1569 /* never bounce for the root cgroup */
1573 spin_lock_bh(&blkg->async_bio_lock);
1574 bio_list_add(&blkg->async_bios, bio);
1575 spin_unlock_bh(&blkg->async_bio_lock);
1577 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1582 * Scale the accumulated delay based on how long it has been since we updated
1583 * the delay. We only call this when we are adding delay, in case it's been a
1584 * while since we added delay, and when we are checking to see if we need to
1585 * delay a task, to account for any delays that may have occurred.
1587 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1589 u64 old = atomic64_read(&blkg->delay_start);
1591 /* negative use_delay means no scaling, see blkcg_set_delay() */
1592 if (atomic_read(&blkg->use_delay) < 0)
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 unsigned long pflags;
1648 u64 now = ktime_to_ns(ktime_get());
1653 while (blkg->parent) {
1654 int use_delay = atomic_read(&blkg->use_delay);
1659 blkcg_scale_delay(blkg, now);
1660 this_delay = atomic64_read(&blkg->delay_nsec);
1661 if (this_delay > delay_nsec) {
1662 delay_nsec = this_delay;
1663 clamp = use_delay > 0;
1666 blkg = blkg->parent;
1673 * Let's not sleep for all eternity if we've amassed a huge delay.
1674 * Swapping or metadata IO can accumulate 10's of seconds worth of
1675 * delay, and we want userspace to be able to do _something_ so cap the
1676 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1677 * tasks will be delayed for 0.25 second for every syscall. If
1678 * blkcg_set_delay() was used as indicated by negative use_delay, the
1679 * caller is responsible for regulating the range.
1682 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1685 psi_memstall_enter(&pflags);
1687 exp = ktime_add_ns(now, delay_nsec);
1688 tok = io_schedule_prepare();
1690 __set_current_state(TASK_KILLABLE);
1691 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1693 } while (!fatal_signal_pending(current));
1694 io_schedule_finish(tok);
1697 psi_memstall_leave(&pflags);
1701 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1703 * This is only called if we've been marked with set_notify_resume(). Obviously
1704 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1705 * check to see if current->throttle_queue is set and if not this doesn't do
1706 * anything. This should only ever be called by the resume code, it's not meant
1707 * to be called by people willy-nilly as it will actually do the work to
1708 * throttle the task if it is setup for throttling.
1710 void blkcg_maybe_throttle_current(void)
1712 struct request_queue *q = current->throttle_queue;
1713 struct cgroup_subsys_state *css;
1714 struct blkcg *blkcg;
1715 struct blkcg_gq *blkg;
1716 bool use_memdelay = current->use_memdelay;
1721 current->throttle_queue = NULL;
1722 current->use_memdelay = false;
1725 css = kthread_blkcg();
1727 blkcg = css_to_blkcg(css);
1729 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1733 blkg = blkg_lookup(blkcg, q);
1736 if (!blkg_tryget(blkg))
1740 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1750 * blkcg_schedule_throttle - this task needs to check for throttling
1751 * @q: the request queue IO was submitted on
1752 * @use_memdelay: do we charge this to memory delay for PSI
1754 * This is called by the IO controller when we know there's delay accumulated
1755 * for the blkg for this task. We do not pass the blkg because there are places
1756 * we call this that may not have that information, the swapping code for
1757 * instance will only have a request_queue at that point. This set's the
1758 * notify_resume for the task to check and see if it requires throttling before
1759 * returning to user space.
1761 * We will only schedule once per syscall. You can call this over and over
1762 * again and it will only do the check once upon return to user space, and only
1763 * throttle once. If the task needs to be throttled again it'll need to be
1764 * re-set at the next time we see the task.
1766 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1768 if (unlikely(current->flags & PF_KTHREAD))
1771 if (current->throttle_queue != q) {
1772 if (!blk_get_queue(q))
1775 if (current->throttle_queue)
1776 blk_put_queue(current->throttle_queue);
1777 current->throttle_queue = q;
1781 current->use_memdelay = use_memdelay;
1782 set_notify_resume(current);
1786 * blkcg_add_delay - add delay to this blkg
1787 * @blkg: blkg of interest
1788 * @now: the current time in nanoseconds
1789 * @delta: how many nanoseconds of delay to add
1791 * Charge @delta to the blkg's current delay accumulation. This is used to
1792 * throttle tasks if an IO controller thinks we need more throttling.
1794 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1796 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1798 blkcg_scale_delay(blkg, now);
1799 atomic64_add(delta, &blkg->delay_nsec);
1803 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1807 * As the failure mode here is to walk up the blkg tree, this ensure that the
1808 * blkg->parent pointers are always valid. This returns the blkg that it ended
1809 * up taking a reference on or %NULL if no reference was taken.
1811 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1812 struct cgroup_subsys_state *css)
1814 struct blkcg_gq *blkg, *ret_blkg = NULL;
1817 blkg = blkg_lookup_create(css_to_blkcg(css),
1818 bio->bi_bdev->bd_disk->queue);
1820 if (blkg_tryget(blkg)) {
1824 blkg = blkg->parent;
1832 * bio_associate_blkg_from_css - associate a bio with a specified css
1836 * Associate @bio with the blkg found by combining the css's blkg and the
1837 * request_queue of the @bio. An association failure is handled by walking up
1838 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1839 * and q->root_blkg. This situation only happens when a cgroup is dying and
1840 * then the remaining bios will spill to the closest alive blkg.
1842 * A reference will be taken on the blkg and will be released when @bio is
1845 void bio_associate_blkg_from_css(struct bio *bio,
1846 struct cgroup_subsys_state *css)
1849 blkg_put(bio->bi_blkg);
1851 if (css && css->parent) {
1852 bio->bi_blkg = blkg_tryget_closest(bio, css);
1854 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg);
1855 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg;
1858 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1861 * bio_associate_blkg - associate a bio with a blkg
1864 * Associate @bio with the blkg found from the bio's css and request_queue.
1865 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1866 * already associated, the css is reused and association redone as the
1867 * request_queue may have changed.
1869 void bio_associate_blkg(struct bio *bio)
1871 struct cgroup_subsys_state *css;
1876 css = &bio_blkcg(bio)->css;
1880 bio_associate_blkg_from_css(bio, css);
1884 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1887 * bio_clone_blkg_association - clone blkg association from src to dst bio
1888 * @dst: destination bio
1891 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1895 blkg_put(dst->bi_blkg);
1896 blkg_get(src->bi_blkg);
1897 dst->bi_blkg = src->bi_blkg;
1900 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1902 static int blk_cgroup_io_type(struct bio *bio)
1904 if (op_is_discard(bio->bi_opf))
1905 return BLKG_IOSTAT_DISCARD;
1906 if (op_is_write(bio->bi_opf))
1907 return BLKG_IOSTAT_WRITE;
1908 return BLKG_IOSTAT_READ;
1911 void blk_cgroup_bio_start(struct bio *bio)
1913 int rwd = blk_cgroup_io_type(bio), cpu;
1914 struct blkg_iostat_set *bis;
1917 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1918 u64_stats_update_begin(&bis->sync);
1921 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1922 * bio and we would have already accounted for the size of the bio.
1924 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1925 bio_set_flag(bio, BIO_CGROUP_ACCT);
1926 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1928 bis->cur.ios[rwd]++;
1930 u64_stats_update_end(&bis->sync);
1931 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1932 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1936 static int __init blkcg_init(void)
1938 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1939 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1940 WQ_UNBOUND | WQ_SYSFS, 0);
1941 if (!blkcg_punt_bio_wq)
1945 subsys_initcall(blkcg_init);
1947 module_param(blkcg_debug_stats, bool, 0644);
1948 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");