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/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/tracehook.h>
30 #include <linux/psi.h>
31 #include <linux/part_stat.h>
33 #include "blk-cgroup.h"
34 #include "blk-ioprio.h"
35 #include "blk-throttle.h"
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
44 static DEFINE_MUTEX(blkcg_pol_register_mutex);
45 static DEFINE_MUTEX(blkcg_pol_mutex);
47 struct blkcg blkcg_root;
48 EXPORT_SYMBOL_GPL(blkcg_root);
50 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51 EXPORT_SYMBOL_GPL(blkcg_root_css);
53 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
55 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
57 bool blkcg_debug_stats = false;
58 static struct workqueue_struct *blkcg_punt_bio_wq;
60 #define BLKG_DESTROY_BATCH_SIZE 64
62 static bool blkcg_policy_enabled(struct request_queue *q,
63 const struct blkcg_policy *pol)
65 return pol && test_bit(pol->plid, q->blkcg_pols);
69 * blkg_free - free a blkg
72 * Free @blkg which may be partially allocated.
74 static void blkg_free(struct blkcg_gq *blkg)
81 for (i = 0; i < BLKCG_MAX_POLS; i++)
83 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
86 blk_put_queue(blkg->q);
87 free_percpu(blkg->iostat_cpu);
88 percpu_ref_exit(&blkg->refcnt);
92 static void __blkg_release(struct rcu_head *rcu)
94 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
96 WARN_ON(!bio_list_empty(&blkg->async_bios));
98 /* release the blkcg and parent blkg refs this blkg has been holding */
99 css_put(&blkg->blkcg->css);
101 blkg_put(blkg->parent);
106 * A group is RCU protected, but having an rcu lock does not mean that one
107 * can access all the fields of blkg and assume these are valid. For
108 * example, don't try to follow throtl_data and request queue links.
110 * Having a reference to blkg under an rcu allows accesses to only values
111 * local to groups like group stats and group rate limits.
113 static void blkg_release(struct percpu_ref *ref)
115 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
117 call_rcu(&blkg->rcu_head, __blkg_release);
120 static void blkg_async_bio_workfn(struct work_struct *work)
122 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
124 struct bio_list bios = BIO_EMPTY_LIST;
126 struct blk_plug plug;
127 bool need_plug = false;
129 /* as long as there are pending bios, @blkg can't go away */
130 spin_lock_bh(&blkg->async_bio_lock);
131 bio_list_merge(&bios, &blkg->async_bios);
132 bio_list_init(&blkg->async_bios);
133 spin_unlock_bh(&blkg->async_bio_lock);
135 /* start plug only when bio_list contains at least 2 bios */
136 if (bios.head && bios.head->bi_next) {
138 blk_start_plug(&plug);
140 while ((bio = bio_list_pop(&bios)))
143 blk_finish_plug(&plug);
147 * blkg_alloc - allocate a blkg
148 * @blkcg: block cgroup the new blkg is associated with
149 * @q: request_queue the new blkg is associated with
150 * @gfp_mask: allocation mask to use
152 * Allocate a new blkg assocating @blkcg and @q.
154 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
157 struct blkcg_gq *blkg;
160 /* alloc and init base part */
161 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
165 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
168 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
169 if (!blkg->iostat_cpu)
172 if (!blk_get_queue(q))
176 INIT_LIST_HEAD(&blkg->q_node);
177 spin_lock_init(&blkg->async_bio_lock);
178 bio_list_init(&blkg->async_bios);
179 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
182 u64_stats_init(&blkg->iostat.sync);
183 for_each_possible_cpu(cpu)
184 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
186 for (i = 0; i < BLKCG_MAX_POLS; i++) {
187 struct blkcg_policy *pol = blkcg_policy[i];
188 struct blkg_policy_data *pd;
190 if (!blkcg_policy_enabled(q, pol))
193 /* alloc per-policy data and attach it to blkg */
194 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
210 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
211 struct request_queue *q, bool update_hint)
213 struct blkcg_gq *blkg;
216 * Hint didn't match. Look up from the radix tree. Note that the
217 * hint can only be updated under queue_lock as otherwise @blkg
218 * could have already been removed from blkg_tree. The caller is
219 * responsible for grabbing queue_lock if @update_hint.
221 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
222 if (blkg && blkg->q == q) {
224 lockdep_assert_held(&q->queue_lock);
225 rcu_assign_pointer(blkcg->blkg_hint, blkg);
232 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
235 * If @new_blkg is %NULL, this function tries to allocate a new one as
236 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
238 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
239 struct request_queue *q,
240 struct blkcg_gq *new_blkg)
242 struct blkcg_gq *blkg;
245 WARN_ON_ONCE(!rcu_read_lock_held());
246 lockdep_assert_held(&q->queue_lock);
248 /* request_queue is dying, do not create/recreate a blkg */
249 if (blk_queue_dying(q)) {
254 /* blkg holds a reference to blkcg */
255 if (!css_tryget_online(&blkcg->css)) {
262 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
263 if (unlikely(!new_blkg)) {
271 if (blkcg_parent(blkcg)) {
272 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
273 if (WARN_ON_ONCE(!blkg->parent)) {
277 blkg_get(blkg->parent);
280 /* invoke per-policy init */
281 for (i = 0; i < BLKCG_MAX_POLS; i++) {
282 struct blkcg_policy *pol = blkcg_policy[i];
284 if (blkg->pd[i] && pol->pd_init_fn)
285 pol->pd_init_fn(blkg->pd[i]);
289 spin_lock(&blkcg->lock);
290 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
292 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
293 list_add(&blkg->q_node, &q->blkg_list);
295 for (i = 0; i < BLKCG_MAX_POLS; i++) {
296 struct blkcg_policy *pol = blkcg_policy[i];
298 if (blkg->pd[i] && pol->pd_online_fn)
299 pol->pd_online_fn(blkg->pd[i]);
303 spin_unlock(&blkcg->lock);
308 /* @blkg failed fully initialized, use the usual release path */
313 css_put(&blkcg->css);
320 * blkg_lookup_create - lookup blkg, try to create one if not there
321 * @blkcg: blkcg of interest
322 * @q: request_queue of interest
324 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
325 * create one. blkg creation is performed recursively from blkcg_root such
326 * that all non-root blkg's have access to the parent blkg. This function
327 * should be called under RCU read lock and takes @q->queue_lock.
329 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
332 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
333 struct request_queue *q)
335 struct blkcg_gq *blkg;
338 WARN_ON_ONCE(!rcu_read_lock_held());
340 blkg = blkg_lookup(blkcg, q);
344 spin_lock_irqsave(&q->queue_lock, flags);
345 blkg = __blkg_lookup(blkcg, q, true);
350 * Create blkgs walking down from blkcg_root to @blkcg, so that all
351 * non-root blkgs have access to their parents. Returns the closest
352 * blkg to the intended blkg should blkg_create() fail.
355 struct blkcg *pos = blkcg;
356 struct blkcg *parent = blkcg_parent(blkcg);
357 struct blkcg_gq *ret_blkg = q->root_blkg;
360 blkg = __blkg_lookup(parent, q, false);
362 /* remember closest blkg */
367 parent = blkcg_parent(parent);
370 blkg = blkg_create(pos, q, NULL);
380 spin_unlock_irqrestore(&q->queue_lock, flags);
384 static void blkg_destroy(struct blkcg_gq *blkg)
386 struct blkcg *blkcg = blkg->blkcg;
389 lockdep_assert_held(&blkg->q->queue_lock);
390 lockdep_assert_held(&blkcg->lock);
392 /* Something wrong if we are trying to remove same group twice */
393 WARN_ON_ONCE(list_empty(&blkg->q_node));
394 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
396 for (i = 0; i < BLKCG_MAX_POLS; i++) {
397 struct blkcg_policy *pol = blkcg_policy[i];
399 if (blkg->pd[i] && pol->pd_offline_fn)
400 pol->pd_offline_fn(blkg->pd[i]);
403 blkg->online = false;
405 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
406 list_del_init(&blkg->q_node);
407 hlist_del_init_rcu(&blkg->blkcg_node);
410 * Both setting lookup hint to and clearing it from @blkg are done
411 * under queue_lock. If it's not pointing to @blkg now, it never
412 * will. Hint assignment itself can race safely.
414 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
415 rcu_assign_pointer(blkcg->blkg_hint, NULL);
418 * Put the reference taken at the time of creation so that when all
419 * queues are gone, group can be destroyed.
421 percpu_ref_kill(&blkg->refcnt);
425 * blkg_destroy_all - destroy all blkgs associated with a request_queue
426 * @q: request_queue of interest
428 * Destroy all blkgs associated with @q.
430 static void blkg_destroy_all(struct request_queue *q)
432 struct blkcg_gq *blkg, *n;
433 int count = BLKG_DESTROY_BATCH_SIZE;
436 spin_lock_irq(&q->queue_lock);
437 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
438 struct blkcg *blkcg = blkg->blkcg;
440 spin_lock(&blkcg->lock);
442 spin_unlock(&blkcg->lock);
445 * in order to avoid holding the spin lock for too long, release
446 * it when a batch of blkgs are destroyed.
449 count = BLKG_DESTROY_BATCH_SIZE;
450 spin_unlock_irq(&q->queue_lock);
457 spin_unlock_irq(&q->queue_lock);
460 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
461 struct cftype *cftype, u64 val)
463 struct blkcg *blkcg = css_to_blkcg(css);
464 struct blkcg_gq *blkg;
467 mutex_lock(&blkcg_pol_mutex);
468 spin_lock_irq(&blkcg->lock);
471 * Note that stat reset is racy - it doesn't synchronize against
472 * stat updates. This is a debug feature which shouldn't exist
473 * anyway. If you get hit by a race, retry.
475 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
476 for_each_possible_cpu(cpu) {
477 struct blkg_iostat_set *bis =
478 per_cpu_ptr(blkg->iostat_cpu, cpu);
479 memset(bis, 0, sizeof(*bis));
481 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
483 for (i = 0; i < BLKCG_MAX_POLS; i++) {
484 struct blkcg_policy *pol = blkcg_policy[i];
486 if (blkg->pd[i] && pol->pd_reset_stats_fn)
487 pol->pd_reset_stats_fn(blkg->pd[i]);
491 spin_unlock_irq(&blkcg->lock);
492 mutex_unlock(&blkcg_pol_mutex);
496 const char *blkg_dev_name(struct blkcg_gq *blkg)
498 if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
500 return bdi_dev_name(blkg->q->disk->bdi);
504 * blkcg_print_blkgs - helper for printing per-blkg data
505 * @sf: seq_file to print to
506 * @blkcg: blkcg of interest
507 * @prfill: fill function to print out a blkg
508 * @pol: policy in question
509 * @data: data to be passed to @prfill
510 * @show_total: to print out sum of prfill return values or not
512 * This function invokes @prfill on each blkg of @blkcg if pd for the
513 * policy specified by @pol exists. @prfill is invoked with @sf, the
514 * policy data and @data and the matching queue lock held. If @show_total
515 * is %true, the sum of the return values from @prfill is printed with
516 * "Total" label at the end.
518 * This is to be used to construct print functions for
519 * cftype->read_seq_string method.
521 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
522 u64 (*prfill)(struct seq_file *,
523 struct blkg_policy_data *, int),
524 const struct blkcg_policy *pol, int data,
527 struct blkcg_gq *blkg;
531 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
532 spin_lock_irq(&blkg->q->queue_lock);
533 if (blkcg_policy_enabled(blkg->q, pol))
534 total += prfill(sf, blkg->pd[pol->plid], data);
535 spin_unlock_irq(&blkg->q->queue_lock);
540 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
542 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
545 * __blkg_prfill_u64 - prfill helper for a single u64 value
546 * @sf: seq_file to print to
547 * @pd: policy private data of interest
550 * Print @v to @sf for the device assocaited with @pd.
552 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
554 const char *dname = blkg_dev_name(pd->blkg);
559 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
562 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
564 /* Performs queue bypass and policy enabled checks then looks up blkg. */
565 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
566 const struct blkcg_policy *pol,
567 struct request_queue *q)
569 WARN_ON_ONCE(!rcu_read_lock_held());
570 lockdep_assert_held(&q->queue_lock);
572 if (!blkcg_policy_enabled(q, pol))
573 return ERR_PTR(-EOPNOTSUPP);
574 return __blkg_lookup(blkcg, q, true /* update_hint */);
578 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
579 * @inputp: input string pointer
581 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
582 * from @input and get and return the matching bdev. *@inputp is
583 * updated to point past the device node prefix. Returns an ERR_PTR()
586 * Use this function iff blkg_conf_prep() can't be used for some reason.
588 struct block_device *blkcg_conf_open_bdev(char **inputp)
590 char *input = *inputp;
591 unsigned int major, minor;
592 struct block_device *bdev;
595 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
596 return ERR_PTR(-EINVAL);
599 if (!isspace(*input))
600 return ERR_PTR(-EINVAL);
601 input = skip_spaces(input);
603 bdev = blkdev_get_no_open(MKDEV(major, minor));
605 return ERR_PTR(-ENODEV);
606 if (bdev_is_partition(bdev)) {
607 blkdev_put_no_open(bdev);
608 return ERR_PTR(-ENODEV);
616 * blkg_conf_prep - parse and prepare for per-blkg config update
617 * @blkcg: target block cgroup
618 * @pol: target policy
619 * @input: input string
620 * @ctx: blkg_conf_ctx to be filled
622 * Parse per-blkg config update from @input and initialize @ctx with the
623 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
624 * part of @input following MAJ:MIN. This function returns with RCU read
625 * lock and queue lock held and must be paired with blkg_conf_finish().
627 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
628 char *input, struct blkg_conf_ctx *ctx)
629 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
631 struct block_device *bdev;
632 struct request_queue *q;
633 struct blkcg_gq *blkg;
636 bdev = blkcg_conf_open_bdev(&input);
638 return PTR_ERR(bdev);
640 q = bdev_get_queue(bdev);
643 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
644 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
646 ret = blk_queue_enter(q, 0);
651 spin_lock_irq(&q->queue_lock);
653 blkg = blkg_lookup_check(blkcg, pol, q);
663 * Create blkgs walking down from blkcg_root to @blkcg, so that all
664 * non-root blkgs have access to their parents.
667 struct blkcg *pos = blkcg;
668 struct blkcg *parent;
669 struct blkcg_gq *new_blkg;
671 parent = blkcg_parent(blkcg);
672 while (parent && !__blkg_lookup(parent, q, false)) {
674 parent = blkcg_parent(parent);
677 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
678 spin_unlock_irq(&q->queue_lock);
681 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
682 if (unlikely(!new_blkg)) {
684 goto fail_exit_queue;
687 if (radix_tree_preload(GFP_KERNEL)) {
690 goto fail_exit_queue;
694 spin_lock_irq(&q->queue_lock);
696 blkg = blkg_lookup_check(pos, pol, q);
706 blkg = blkg_create(pos, q, new_blkg);
713 radix_tree_preload_end();
726 radix_tree_preload_end();
728 spin_unlock_irq(&q->queue_lock);
733 blkdev_put_no_open(bdev);
735 * If queue was bypassing, we should retry. Do so after a
736 * short msleep(). It isn't strictly necessary but queue
737 * can be bypassing for some time and it's always nice to
738 * avoid busy looping.
742 ret = restart_syscall();
746 EXPORT_SYMBOL_GPL(blkg_conf_prep);
749 * blkg_conf_finish - finish up per-blkg config update
750 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
752 * Finish up after per-blkg config update. This function must be paired
753 * with blkg_conf_prep().
755 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
756 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
758 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
760 blkdev_put_no_open(ctx->bdev);
762 EXPORT_SYMBOL_GPL(blkg_conf_finish);
764 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
768 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
769 dst->bytes[i] = src->bytes[i];
770 dst->ios[i] = src->ios[i];
774 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
778 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
779 dst->bytes[i] += src->bytes[i];
780 dst->ios[i] += src->ios[i];
784 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
788 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
789 dst->bytes[i] -= src->bytes[i];
790 dst->ios[i] -= src->ios[i];
794 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
796 struct blkcg *blkcg = css_to_blkcg(css);
797 struct blkcg_gq *blkg;
799 /* Root-level stats are sourced from system-wide IO stats */
800 if (!cgroup_parent(css->cgroup))
805 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
806 struct blkcg_gq *parent = blkg->parent;
807 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
808 struct blkg_iostat cur, delta;
812 /* fetch the current per-cpu values */
814 seq = u64_stats_fetch_begin(&bisc->sync);
815 blkg_iostat_set(&cur, &bisc->cur);
816 } while (u64_stats_fetch_retry(&bisc->sync, seq));
818 /* propagate percpu delta to global */
819 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
820 blkg_iostat_set(&delta, &cur);
821 blkg_iostat_sub(&delta, &bisc->last);
822 blkg_iostat_add(&blkg->iostat.cur, &delta);
823 blkg_iostat_add(&bisc->last, &delta);
824 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
826 /* propagate global delta to parent (unless that's root) */
827 if (parent && parent->parent) {
828 flags = u64_stats_update_begin_irqsave(&parent->iostat.sync);
829 blkg_iostat_set(&delta, &blkg->iostat.cur);
830 blkg_iostat_sub(&delta, &blkg->iostat.last);
831 blkg_iostat_add(&parent->iostat.cur, &delta);
832 blkg_iostat_add(&blkg->iostat.last, &delta);
833 u64_stats_update_end_irqrestore(&parent->iostat.sync, flags);
841 * We source root cgroup stats from the system-wide stats to avoid
842 * tracking the same information twice and incurring overhead when no
843 * cgroups are defined. For that reason, cgroup_rstat_flush in
844 * blkcg_print_stat does not actually fill out the iostat in the root
847 * However, we would like to re-use the printing code between the root and
848 * non-root cgroups to the extent possible. For that reason, we simulate
849 * flushing the root cgroup's stats by explicitly filling in the iostat
850 * with disk level statistics.
852 static void blkcg_fill_root_iostats(void)
854 struct class_dev_iter iter;
857 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
858 while ((dev = class_dev_iter_next(&iter))) {
859 struct block_device *bdev = dev_to_bdev(dev);
860 struct blkcg_gq *blkg =
861 blk_queue_root_blkg(bdev_get_queue(bdev));
862 struct blkg_iostat tmp;
866 memset(&tmp, 0, sizeof(tmp));
867 for_each_possible_cpu(cpu) {
868 struct disk_stats *cpu_dkstats;
870 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
871 tmp.ios[BLKG_IOSTAT_READ] +=
872 cpu_dkstats->ios[STAT_READ];
873 tmp.ios[BLKG_IOSTAT_WRITE] +=
874 cpu_dkstats->ios[STAT_WRITE];
875 tmp.ios[BLKG_IOSTAT_DISCARD] +=
876 cpu_dkstats->ios[STAT_DISCARD];
877 // convert sectors to bytes
878 tmp.bytes[BLKG_IOSTAT_READ] +=
879 cpu_dkstats->sectors[STAT_READ] << 9;
880 tmp.bytes[BLKG_IOSTAT_WRITE] +=
881 cpu_dkstats->sectors[STAT_WRITE] << 9;
882 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
883 cpu_dkstats->sectors[STAT_DISCARD] << 9;
886 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
887 blkg_iostat_set(&blkg->iostat.cur, &tmp);
888 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
892 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
894 struct blkg_iostat_set *bis = &blkg->iostat;
895 u64 rbytes, wbytes, rios, wios, dbytes, dios;
896 bool has_stats = false;
904 dname = blkg_dev_name(blkg);
908 seq_printf(s, "%s ", dname);
911 seq = u64_stats_fetch_begin(&bis->sync);
913 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
914 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
915 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
916 rios = bis->cur.ios[BLKG_IOSTAT_READ];
917 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
918 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
919 } while (u64_stats_fetch_retry(&bis->sync, seq));
921 if (rbytes || wbytes || rios || wios) {
923 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
924 rbytes, wbytes, rios, wios,
928 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
930 seq_printf(s, " use_delay=%d delay_nsec=%llu",
931 atomic_read(&blkg->use_delay),
932 atomic64_read(&blkg->delay_nsec));
935 for (i = 0; i < BLKCG_MAX_POLS; i++) {
936 struct blkcg_policy *pol = blkcg_policy[i];
938 if (!blkg->pd[i] || !pol->pd_stat_fn)
941 if (pol->pd_stat_fn(blkg->pd[i], s))
949 static int blkcg_print_stat(struct seq_file *sf, void *v)
951 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
952 struct blkcg_gq *blkg;
954 if (!seq_css(sf)->parent)
955 blkcg_fill_root_iostats();
957 cgroup_rstat_flush(blkcg->css.cgroup);
960 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
961 spin_lock_irq(&blkg->q->queue_lock);
962 blkcg_print_one_stat(blkg, sf);
963 spin_unlock_irq(&blkg->q->queue_lock);
969 static struct cftype blkcg_files[] = {
972 .seq_show = blkcg_print_stat,
977 static struct cftype blkcg_legacy_files[] = {
979 .name = "reset_stats",
980 .write_u64 = blkcg_reset_stats,
986 * blkcg destruction is a three-stage process.
988 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
989 * which offlines writeback. Here we tie the next stage of blkg destruction
990 * to the completion of writeback associated with the blkcg. This lets us
991 * avoid punting potentially large amounts of outstanding writeback to root
992 * while maintaining any ongoing policies. The next stage is triggered when
993 * the nr_cgwbs count goes to zero.
995 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
996 * and handles the destruction of blkgs. Here the css reference held by
997 * the blkg is put back eventually allowing blkcg_css_free() to be called.
998 * This work may occur in cgwb_release_workfn() on the cgwb_release
999 * workqueue. Any submitted ios that fail to get the blkg ref will be
1000 * punted to the root_blkg.
1002 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1003 * This finally frees the blkcg.
1007 * blkcg_css_offline - cgroup css_offline callback
1008 * @css: css of interest
1010 * This function is called when @css is about to go away. Here the cgwbs are
1011 * offlined first and only once writeback associated with the blkcg has
1012 * finished do we start step 2 (see above).
1014 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1016 struct blkcg *blkcg = css_to_blkcg(css);
1018 /* this prevents anyone from attaching or migrating to this blkcg */
1019 wb_blkcg_offline(blkcg);
1021 /* put the base online pin allowing step 2 to be triggered */
1022 blkcg_unpin_online(blkcg);
1026 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1027 * @blkcg: blkcg of interest
1029 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1030 * is nested inside q lock, this function performs reverse double lock dancing.
1031 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1032 * blkcg_css_free to eventually be called.
1034 * This is the blkcg counterpart of ioc_release_fn().
1036 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1040 spin_lock_irq(&blkcg->lock);
1042 while (!hlist_empty(&blkcg->blkg_list)) {
1043 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1044 struct blkcg_gq, blkcg_node);
1045 struct request_queue *q = blkg->q;
1047 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1049 * Given that the system can accumulate a huge number
1050 * of blkgs in pathological cases, check to see if we
1051 * need to rescheduling to avoid softlockup.
1053 spin_unlock_irq(&blkcg->lock);
1055 spin_lock_irq(&blkcg->lock);
1060 spin_unlock(&q->queue_lock);
1063 spin_unlock_irq(&blkcg->lock);
1066 static void blkcg_css_free(struct cgroup_subsys_state *css)
1068 struct blkcg *blkcg = css_to_blkcg(css);
1071 mutex_lock(&blkcg_pol_mutex);
1073 list_del(&blkcg->all_blkcgs_node);
1075 for (i = 0; i < BLKCG_MAX_POLS; i++)
1077 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1079 mutex_unlock(&blkcg_pol_mutex);
1084 static struct cgroup_subsys_state *
1085 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1087 struct blkcg *blkcg;
1088 struct cgroup_subsys_state *ret;
1091 mutex_lock(&blkcg_pol_mutex);
1094 blkcg = &blkcg_root;
1096 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1098 ret = ERR_PTR(-ENOMEM);
1103 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1104 struct blkcg_policy *pol = blkcg_policy[i];
1105 struct blkcg_policy_data *cpd;
1108 * If the policy hasn't been attached yet, wait for it
1109 * to be attached before doing anything else. Otherwise,
1110 * check if the policy requires any specific per-cgroup
1111 * data: if it does, allocate and initialize it.
1113 if (!pol || !pol->cpd_alloc_fn)
1116 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1118 ret = ERR_PTR(-ENOMEM);
1121 blkcg->cpd[i] = cpd;
1124 if (pol->cpd_init_fn)
1125 pol->cpd_init_fn(cpd);
1128 spin_lock_init(&blkcg->lock);
1129 refcount_set(&blkcg->online_pin, 1);
1130 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1131 INIT_HLIST_HEAD(&blkcg->blkg_list);
1132 #ifdef CONFIG_CGROUP_WRITEBACK
1133 INIT_LIST_HEAD(&blkcg->cgwb_list);
1135 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1137 mutex_unlock(&blkcg_pol_mutex);
1141 for (i--; i >= 0; i--)
1143 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1145 if (blkcg != &blkcg_root)
1148 mutex_unlock(&blkcg_pol_mutex);
1152 static int blkcg_css_online(struct cgroup_subsys_state *css)
1154 struct blkcg *blkcg = css_to_blkcg(css);
1155 struct blkcg *parent = blkcg_parent(blkcg);
1158 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1159 * don't go offline while cgwbs are still active on them. Pin the
1160 * parent so that offline always happens towards the root.
1163 blkcg_pin_online(parent);
1168 * blkcg_init_queue - initialize blkcg part of request queue
1169 * @q: request_queue to initialize
1171 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1172 * part of new request_queue @q.
1175 * 0 on success, -errno on failure.
1177 int blkcg_init_queue(struct request_queue *q)
1179 struct blkcg_gq *new_blkg, *blkg;
1183 INIT_LIST_HEAD(&q->blkg_list);
1185 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1189 preloaded = !radix_tree_preload(GFP_KERNEL);
1191 /* Make sure the root blkg exists. */
1193 spin_lock_irq(&q->queue_lock);
1194 blkg = blkg_create(&blkcg_root, q, new_blkg);
1197 q->root_blkg = blkg;
1198 spin_unlock_irq(&q->queue_lock);
1202 radix_tree_preload_end();
1204 ret = blk_ioprio_init(q);
1206 goto err_destroy_all;
1208 ret = blk_throtl_init(q);
1210 goto err_destroy_all;
1212 ret = blk_iolatency_init(q);
1215 goto err_destroy_all;
1221 blkg_destroy_all(q);
1224 spin_unlock_irq(&q->queue_lock);
1227 radix_tree_preload_end();
1228 return PTR_ERR(blkg);
1232 * blkcg_exit_queue - exit and release blkcg part of request_queue
1233 * @q: request_queue being released
1235 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1237 void blkcg_exit_queue(struct request_queue *q)
1239 blkg_destroy_all(q);
1243 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1247 mutex_lock(&blkcg_pol_mutex);
1249 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1250 struct blkcg_policy *pol = blkcg_policy[i];
1251 struct blkcg *blkcg;
1253 if (!pol || !pol->cpd_bind_fn)
1256 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1257 if (blkcg->cpd[pol->plid])
1258 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1260 mutex_unlock(&blkcg_pol_mutex);
1263 static void blkcg_exit(struct task_struct *tsk)
1265 if (tsk->throttle_queue)
1266 blk_put_queue(tsk->throttle_queue);
1267 tsk->throttle_queue = NULL;
1270 struct cgroup_subsys io_cgrp_subsys = {
1271 .css_alloc = blkcg_css_alloc,
1272 .css_online = blkcg_css_online,
1273 .css_offline = blkcg_css_offline,
1274 .css_free = blkcg_css_free,
1275 .css_rstat_flush = blkcg_rstat_flush,
1277 .dfl_cftypes = blkcg_files,
1278 .legacy_cftypes = blkcg_legacy_files,
1279 .legacy_name = "blkio",
1283 * This ensures that, if available, memcg is automatically enabled
1284 * together on the default hierarchy so that the owner cgroup can
1285 * be retrieved from writeback pages.
1287 .depends_on = 1 << memory_cgrp_id,
1290 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1293 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1294 * @q: request_queue of interest
1295 * @pol: blkcg policy to activate
1297 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1298 * bypass mode to populate its blkgs with policy_data for @pol.
1300 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1301 * from IO path. Update of each blkg is protected by both queue and blkcg
1302 * locks so that holding either lock and testing blkcg_policy_enabled() is
1303 * always enough for dereferencing policy data.
1305 * The caller is responsible for synchronizing [de]activations and policy
1306 * [un]registerations. Returns 0 on success, -errno on failure.
1308 int blkcg_activate_policy(struct request_queue *q,
1309 const struct blkcg_policy *pol)
1311 struct blkg_policy_data *pd_prealloc = NULL;
1312 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1315 if (blkcg_policy_enabled(q, pol))
1319 blk_mq_freeze_queue(q);
1321 spin_lock_irq(&q->queue_lock);
1323 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1324 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1325 struct blkg_policy_data *pd;
1327 if (blkg->pd[pol->plid])
1330 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1331 if (blkg == pinned_blkg) {
1335 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1341 * GFP_NOWAIT failed. Free the existing one and
1342 * prealloc for @blkg w/ GFP_KERNEL.
1345 blkg_put(pinned_blkg);
1349 spin_unlock_irq(&q->queue_lock);
1352 pol->pd_free_fn(pd_prealloc);
1353 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1361 blkg->pd[pol->plid] = pd;
1363 pd->plid = pol->plid;
1366 /* all allocated, init in the same order */
1367 if (pol->pd_init_fn)
1368 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1369 pol->pd_init_fn(blkg->pd[pol->plid]);
1371 __set_bit(pol->plid, q->blkcg_pols);
1374 spin_unlock_irq(&q->queue_lock);
1377 blk_mq_unfreeze_queue(q);
1379 blkg_put(pinned_blkg);
1381 pol->pd_free_fn(pd_prealloc);
1385 /* alloc failed, nothing's initialized yet, free everything */
1386 spin_lock_irq(&q->queue_lock);
1387 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1388 struct blkcg *blkcg = blkg->blkcg;
1390 spin_lock(&blkcg->lock);
1391 if (blkg->pd[pol->plid]) {
1392 pol->pd_free_fn(blkg->pd[pol->plid]);
1393 blkg->pd[pol->plid] = NULL;
1395 spin_unlock(&blkcg->lock);
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 struct blkcg *blkcg = blkg->blkcg;
1429 spin_lock(&blkcg->lock);
1430 if (blkg->pd[pol->plid]) {
1431 if (pol->pd_offline_fn)
1432 pol->pd_offline_fn(blkg->pd[pol->plid]);
1433 pol->pd_free_fn(blkg->pd[pol->plid]);
1434 blkg->pd[pol->plid] = NULL;
1436 spin_unlock(&blkcg->lock);
1439 spin_unlock_irq(&q->queue_lock);
1442 blk_mq_unfreeze_queue(q);
1444 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1447 * blkcg_policy_register - register a blkcg policy
1448 * @pol: blkcg policy to register
1450 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1451 * successful registration. Returns 0 on success and -errno on failure.
1453 int blkcg_policy_register(struct blkcg_policy *pol)
1455 struct blkcg *blkcg;
1458 mutex_lock(&blkcg_pol_register_mutex);
1459 mutex_lock(&blkcg_pol_mutex);
1461 /* find an empty slot */
1463 for (i = 0; i < BLKCG_MAX_POLS; i++)
1464 if (!blkcg_policy[i])
1466 if (i >= BLKCG_MAX_POLS) {
1467 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1471 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1472 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1473 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1478 blkcg_policy[pol->plid] = pol;
1480 /* allocate and install cpd's */
1481 if (pol->cpd_alloc_fn) {
1482 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1483 struct blkcg_policy_data *cpd;
1485 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1489 blkcg->cpd[pol->plid] = cpd;
1491 cpd->plid = pol->plid;
1492 if (pol->cpd_init_fn)
1493 pol->cpd_init_fn(cpd);
1497 mutex_unlock(&blkcg_pol_mutex);
1499 /* everything is in place, add intf files for the new policy */
1500 if (pol->dfl_cftypes)
1501 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1503 if (pol->legacy_cftypes)
1504 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1505 pol->legacy_cftypes));
1506 mutex_unlock(&blkcg_pol_register_mutex);
1510 if (pol->cpd_free_fn) {
1511 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1512 if (blkcg->cpd[pol->plid]) {
1513 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1514 blkcg->cpd[pol->plid] = NULL;
1518 blkcg_policy[pol->plid] = NULL;
1520 mutex_unlock(&blkcg_pol_mutex);
1521 mutex_unlock(&blkcg_pol_register_mutex);
1524 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1527 * blkcg_policy_unregister - unregister a blkcg policy
1528 * @pol: blkcg policy to unregister
1530 * Undo blkcg_policy_register(@pol). Might sleep.
1532 void blkcg_policy_unregister(struct blkcg_policy *pol)
1534 struct blkcg *blkcg;
1536 mutex_lock(&blkcg_pol_register_mutex);
1538 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1541 /* kill the intf files first */
1542 if (pol->dfl_cftypes)
1543 cgroup_rm_cftypes(pol->dfl_cftypes);
1544 if (pol->legacy_cftypes)
1545 cgroup_rm_cftypes(pol->legacy_cftypes);
1547 /* remove cpds and unregister */
1548 mutex_lock(&blkcg_pol_mutex);
1550 if (pol->cpd_free_fn) {
1551 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1552 if (blkcg->cpd[pol->plid]) {
1553 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1554 blkcg->cpd[pol->plid] = NULL;
1558 blkcg_policy[pol->plid] = NULL;
1560 mutex_unlock(&blkcg_pol_mutex);
1562 mutex_unlock(&blkcg_pol_register_mutex);
1564 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1566 bool __blkcg_punt_bio_submit(struct bio *bio)
1568 struct blkcg_gq *blkg = bio->bi_blkg;
1570 /* consume the flag first */
1571 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1573 /* never bounce for the root cgroup */
1577 spin_lock_bh(&blkg->async_bio_lock);
1578 bio_list_add(&blkg->async_bios, bio);
1579 spin_unlock_bh(&blkg->async_bio_lock);
1581 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1586 * Scale the accumulated delay based on how long it has been since we updated
1587 * the delay. We only call this when we are adding delay, in case it's been a
1588 * while since we added delay, and when we are checking to see if we need to
1589 * delay a task, to account for any delays that may have occurred.
1591 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1593 u64 old = atomic64_read(&blkg->delay_start);
1595 /* negative use_delay means no scaling, see blkcg_set_delay() */
1596 if (atomic_read(&blkg->use_delay) < 0)
1600 * We only want to scale down every second. The idea here is that we
1601 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1602 * time window. We only want to throttle tasks for recent delay that
1603 * has occurred, in 1 second time windows since that's the maximum
1604 * things can be throttled. We save the current delay window in
1605 * blkg->last_delay so we know what amount is still left to be charged
1606 * to the blkg from this point onward. blkg->last_use keeps track of
1607 * the use_delay counter. The idea is if we're unthrottling the blkg we
1608 * are ok with whatever is happening now, and we can take away more of
1609 * the accumulated delay as we've already throttled enough that
1610 * everybody is happy with their IO latencies.
1612 if (time_before64(old + NSEC_PER_SEC, now) &&
1613 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1614 u64 cur = atomic64_read(&blkg->delay_nsec);
1615 u64 sub = min_t(u64, blkg->last_delay, now - old);
1616 int cur_use = atomic_read(&blkg->use_delay);
1619 * We've been unthrottled, subtract a larger chunk of our
1620 * accumulated delay.
1622 if (cur_use < blkg->last_use)
1623 sub = max_t(u64, sub, blkg->last_delay >> 1);
1626 * This shouldn't happen, but handle it anyway. Our delay_nsec
1627 * should only ever be growing except here where we subtract out
1628 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1629 * rather not end up with negative numbers.
1631 if (unlikely(cur < sub)) {
1632 atomic64_set(&blkg->delay_nsec, 0);
1633 blkg->last_delay = 0;
1635 atomic64_sub(sub, &blkg->delay_nsec);
1636 blkg->last_delay = cur - sub;
1638 blkg->last_use = cur_use;
1643 * This is called when we want to actually walk up the hierarchy and check to
1644 * see if we need to throttle, and then actually throttle if there is some
1645 * accumulated delay. This should only be called upon return to user space so
1646 * we're not holding some lock that would induce a priority inversion.
1648 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1650 unsigned long pflags;
1652 u64 now = ktime_to_ns(ktime_get());
1657 while (blkg->parent) {
1658 int use_delay = atomic_read(&blkg->use_delay);
1663 blkcg_scale_delay(blkg, now);
1664 this_delay = atomic64_read(&blkg->delay_nsec);
1665 if (this_delay > delay_nsec) {
1666 delay_nsec = this_delay;
1667 clamp = use_delay > 0;
1670 blkg = blkg->parent;
1677 * Let's not sleep for all eternity if we've amassed a huge delay.
1678 * Swapping or metadata IO can accumulate 10's of seconds worth of
1679 * delay, and we want userspace to be able to do _something_ so cap the
1680 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1681 * tasks will be delayed for 0.25 second for every syscall. If
1682 * blkcg_set_delay() was used as indicated by negative use_delay, the
1683 * caller is responsible for regulating the range.
1686 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1689 psi_memstall_enter(&pflags);
1691 exp = ktime_add_ns(now, delay_nsec);
1692 tok = io_schedule_prepare();
1694 __set_current_state(TASK_KILLABLE);
1695 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1697 } while (!fatal_signal_pending(current));
1698 io_schedule_finish(tok);
1701 psi_memstall_leave(&pflags);
1705 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1707 * This is only called if we've been marked with set_notify_resume(). Obviously
1708 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1709 * check to see if current->throttle_queue is set and if not this doesn't do
1710 * anything. This should only ever be called by the resume code, it's not meant
1711 * to be called by people willy-nilly as it will actually do the work to
1712 * throttle the task if it is setup for throttling.
1714 void blkcg_maybe_throttle_current(void)
1716 struct request_queue *q = current->throttle_queue;
1717 struct cgroup_subsys_state *css;
1718 struct blkcg *blkcg;
1719 struct blkcg_gq *blkg;
1720 bool use_memdelay = current->use_memdelay;
1725 current->throttle_queue = NULL;
1726 current->use_memdelay = false;
1729 css = kthread_blkcg();
1731 blkcg = css_to_blkcg(css);
1733 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1737 blkg = blkg_lookup(blkcg, q);
1740 if (!blkg_tryget(blkg))
1744 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1754 * blkcg_schedule_throttle - this task needs to check for throttling
1755 * @q: the request queue IO was submitted on
1756 * @use_memdelay: do we charge this to memory delay for PSI
1758 * This is called by the IO controller when we know there's delay accumulated
1759 * for the blkg for this task. We do not pass the blkg because there are places
1760 * we call this that may not have that information, the swapping code for
1761 * instance will only have a request_queue at that point. This set's the
1762 * notify_resume for the task to check and see if it requires throttling before
1763 * returning to user space.
1765 * We will only schedule once per syscall. You can call this over and over
1766 * again and it will only do the check once upon return to user space, and only
1767 * throttle once. If the task needs to be throttled again it'll need to be
1768 * re-set at the next time we see the task.
1770 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1772 if (unlikely(current->flags & PF_KTHREAD))
1775 if (current->throttle_queue != q) {
1776 if (!blk_get_queue(q))
1779 if (current->throttle_queue)
1780 blk_put_queue(current->throttle_queue);
1781 current->throttle_queue = q;
1785 current->use_memdelay = use_memdelay;
1786 set_notify_resume(current);
1790 * blkcg_add_delay - add delay to this blkg
1791 * @blkg: blkg of interest
1792 * @now: the current time in nanoseconds
1793 * @delta: how many nanoseconds of delay to add
1795 * Charge @delta to the blkg's current delay accumulation. This is used to
1796 * throttle tasks if an IO controller thinks we need more throttling.
1798 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1800 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1802 blkcg_scale_delay(blkg, now);
1803 atomic64_add(delta, &blkg->delay_nsec);
1807 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1811 * As the failure mode here is to walk up the blkg tree, this ensure that the
1812 * blkg->parent pointers are always valid. This returns the blkg that it ended
1813 * up taking a reference on or %NULL if no reference was taken.
1815 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1816 struct cgroup_subsys_state *css)
1818 struct blkcg_gq *blkg, *ret_blkg = NULL;
1821 blkg = blkg_lookup_create(css_to_blkcg(css),
1822 bdev_get_queue(bio->bi_bdev));
1824 if (blkg_tryget(blkg)) {
1828 blkg = blkg->parent;
1836 * bio_associate_blkg_from_css - associate a bio with a specified css
1840 * Associate @bio with the blkg found by combining the css's blkg and the
1841 * request_queue of the @bio. An association failure is handled by walking up
1842 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1843 * and q->root_blkg. This situation only happens when a cgroup is dying and
1844 * then the remaining bios will spill to the closest alive blkg.
1846 * A reference will be taken on the blkg and will be released when @bio is
1849 void bio_associate_blkg_from_css(struct bio *bio,
1850 struct cgroup_subsys_state *css)
1853 blkg_put(bio->bi_blkg);
1855 if (css && css->parent) {
1856 bio->bi_blkg = blkg_tryget_closest(bio, css);
1858 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1859 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1862 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1865 * bio_associate_blkg - associate a bio with a blkg
1868 * Associate @bio with the blkg found from the bio's css and request_queue.
1869 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1870 * already associated, the css is reused and association redone as the
1871 * request_queue may have changed.
1873 void bio_associate_blkg(struct bio *bio)
1875 struct cgroup_subsys_state *css;
1880 css = &bio_blkcg(bio)->css;
1884 bio_associate_blkg_from_css(bio, css);
1888 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1891 * bio_clone_blkg_association - clone blkg association from src to dst bio
1892 * @dst: destination bio
1895 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1899 blkg_put(dst->bi_blkg);
1900 blkg_get(src->bi_blkg);
1901 dst->bi_blkg = src->bi_blkg;
1904 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1906 static int blk_cgroup_io_type(struct bio *bio)
1908 if (op_is_discard(bio->bi_opf))
1909 return BLKG_IOSTAT_DISCARD;
1910 if (op_is_write(bio->bi_opf))
1911 return BLKG_IOSTAT_WRITE;
1912 return BLKG_IOSTAT_READ;
1915 void blk_cgroup_bio_start(struct bio *bio)
1917 int rwd = blk_cgroup_io_type(bio), cpu;
1918 struct blkg_iostat_set *bis;
1919 unsigned long flags;
1922 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1923 flags = u64_stats_update_begin_irqsave(&bis->sync);
1926 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1927 * bio and we would have already accounted for the size of the bio.
1929 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1930 bio_set_flag(bio, BIO_CGROUP_ACCT);
1931 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1933 bis->cur.ios[rwd]++;
1935 u64_stats_update_end_irqrestore(&bis->sync, flags);
1936 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1937 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1941 static int __init blkcg_init(void)
1943 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1944 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1945 WQ_UNBOUND | WQ_SYSFS, 0);
1946 if (!blkcg_punt_bio_wq)
1950 subsys_initcall(blkcg_init);
1952 module_param(blkcg_debug_stats, bool, 0644);
1953 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");