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>
34 #include "blk-ioprio.h"
37 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
38 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
39 * policy [un]register operations including cgroup file additions /
40 * removals. Putting cgroup file registration outside blkcg_pol_mutex
41 * allows grabbing it from cgroup callbacks.
43 static DEFINE_MUTEX(blkcg_pol_register_mutex);
44 static DEFINE_MUTEX(blkcg_pol_mutex);
46 struct blkcg blkcg_root;
47 EXPORT_SYMBOL_GPL(blkcg_root);
49 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
50 EXPORT_SYMBOL_GPL(blkcg_root_css);
52 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
54 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
56 bool blkcg_debug_stats = false;
57 static struct workqueue_struct *blkcg_punt_bio_wq;
59 #define BLKG_DESTROY_BATCH_SIZE 64
61 static bool blkcg_policy_enabled(struct request_queue *q,
62 const struct blkcg_policy *pol)
64 return pol && test_bit(pol->plid, q->blkcg_pols);
68 * blkg_free - free a blkg
71 * Free @blkg which may be partially allocated.
73 static void blkg_free(struct blkcg_gq *blkg)
80 for (i = 0; i < BLKCG_MAX_POLS; i++)
82 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
84 free_percpu(blkg->iostat_cpu);
85 percpu_ref_exit(&blkg->refcnt);
89 static void __blkg_release(struct rcu_head *rcu)
91 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
93 WARN_ON(!bio_list_empty(&blkg->async_bios));
95 /* release the blkcg and parent blkg refs this blkg has been holding */
96 css_put(&blkg->blkcg->css);
98 blkg_put(blkg->parent);
103 * A group is RCU protected, but having an rcu lock does not mean that one
104 * can access all the fields of blkg and assume these are valid. For
105 * example, don't try to follow throtl_data and request queue links.
107 * Having a reference to blkg under an rcu allows accesses to only values
108 * local to groups like group stats and group rate limits.
110 static void blkg_release(struct percpu_ref *ref)
112 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
114 call_rcu(&blkg->rcu_head, __blkg_release);
117 static void blkg_async_bio_workfn(struct work_struct *work)
119 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
121 struct bio_list bios = BIO_EMPTY_LIST;
123 struct blk_plug plug;
124 bool need_plug = false;
126 /* as long as there are pending bios, @blkg can't go away */
127 spin_lock_bh(&blkg->async_bio_lock);
128 bio_list_merge(&bios, &blkg->async_bios);
129 bio_list_init(&blkg->async_bios);
130 spin_unlock_bh(&blkg->async_bio_lock);
132 /* start plug only when bio_list contains at least 2 bios */
133 if (bios.head && bios.head->bi_next) {
135 blk_start_plug(&plug);
137 while ((bio = bio_list_pop(&bios)))
140 blk_finish_plug(&plug);
144 * blkg_alloc - allocate a blkg
145 * @blkcg: block cgroup the new blkg is associated with
146 * @q: request_queue the new blkg is associated with
147 * @gfp_mask: allocation mask to use
149 * Allocate a new blkg assocating @blkcg and @q.
151 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
154 struct blkcg_gq *blkg;
157 /* alloc and init base part */
158 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
162 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
165 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
166 if (!blkg->iostat_cpu)
170 INIT_LIST_HEAD(&blkg->q_node);
171 spin_lock_init(&blkg->async_bio_lock);
172 bio_list_init(&blkg->async_bios);
173 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
176 u64_stats_init(&blkg->iostat.sync);
177 for_each_possible_cpu(cpu)
178 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
180 for (i = 0; i < BLKCG_MAX_POLS; i++) {
181 struct blkcg_policy *pol = blkcg_policy[i];
182 struct blkg_policy_data *pd;
184 if (!blkcg_policy_enabled(q, pol))
187 /* alloc per-policy data and attach it to blkg */
188 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
204 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
205 struct request_queue *q, bool update_hint)
207 struct blkcg_gq *blkg;
210 * Hint didn't match. Look up from the radix tree. Note that the
211 * hint can only be updated under queue_lock as otherwise @blkg
212 * could have already been removed from blkg_tree. The caller is
213 * responsible for grabbing queue_lock if @update_hint.
215 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
216 if (blkg && blkg->q == q) {
218 lockdep_assert_held(&q->queue_lock);
219 rcu_assign_pointer(blkcg->blkg_hint, blkg);
226 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
229 * If @new_blkg is %NULL, this function tries to allocate a new one as
230 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
232 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
233 struct request_queue *q,
234 struct blkcg_gq *new_blkg)
236 struct blkcg_gq *blkg;
239 WARN_ON_ONCE(!rcu_read_lock_held());
240 lockdep_assert_held(&q->queue_lock);
242 /* request_queue is dying, do not create/recreate a blkg */
243 if (blk_queue_dying(q)) {
248 /* blkg holds a reference to blkcg */
249 if (!css_tryget_online(&blkcg->css)) {
256 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
257 if (unlikely(!new_blkg)) {
265 if (blkcg_parent(blkcg)) {
266 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
267 if (WARN_ON_ONCE(!blkg->parent)) {
271 blkg_get(blkg->parent);
274 /* invoke per-policy init */
275 for (i = 0; i < BLKCG_MAX_POLS; i++) {
276 struct blkcg_policy *pol = blkcg_policy[i];
278 if (blkg->pd[i] && pol->pd_init_fn)
279 pol->pd_init_fn(blkg->pd[i]);
283 spin_lock(&blkcg->lock);
284 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
286 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
287 list_add(&blkg->q_node, &q->blkg_list);
289 for (i = 0; i < BLKCG_MAX_POLS; i++) {
290 struct blkcg_policy *pol = blkcg_policy[i];
292 if (blkg->pd[i] && pol->pd_online_fn)
293 pol->pd_online_fn(blkg->pd[i]);
297 spin_unlock(&blkcg->lock);
302 /* @blkg failed fully initialized, use the usual release path */
307 css_put(&blkcg->css);
314 * blkg_lookup_create - lookup blkg, try to create one if not there
315 * @blkcg: blkcg of interest
316 * @q: request_queue of interest
318 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
319 * create one. blkg creation is performed recursively from blkcg_root such
320 * that all non-root blkg's have access to the parent blkg. This function
321 * should be called under RCU read lock and takes @q->queue_lock.
323 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
326 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
327 struct request_queue *q)
329 struct blkcg_gq *blkg;
332 WARN_ON_ONCE(!rcu_read_lock_held());
334 blkg = blkg_lookup(blkcg, q);
338 spin_lock_irqsave(&q->queue_lock, flags);
339 blkg = __blkg_lookup(blkcg, q, true);
344 * Create blkgs walking down from blkcg_root to @blkcg, so that all
345 * non-root blkgs have access to their parents. Returns the closest
346 * blkg to the intended blkg should blkg_create() fail.
349 struct blkcg *pos = blkcg;
350 struct blkcg *parent = blkcg_parent(blkcg);
351 struct blkcg_gq *ret_blkg = q->root_blkg;
354 blkg = __blkg_lookup(parent, q, false);
356 /* remember closest blkg */
361 parent = blkcg_parent(parent);
364 blkg = blkg_create(pos, q, NULL);
374 spin_unlock_irqrestore(&q->queue_lock, flags);
378 static void blkg_destroy(struct blkcg_gq *blkg)
380 struct blkcg *blkcg = blkg->blkcg;
383 lockdep_assert_held(&blkg->q->queue_lock);
384 lockdep_assert_held(&blkcg->lock);
386 /* Something wrong if we are trying to remove same group twice */
387 WARN_ON_ONCE(list_empty(&blkg->q_node));
388 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
390 for (i = 0; i < BLKCG_MAX_POLS; i++) {
391 struct blkcg_policy *pol = blkcg_policy[i];
393 if (blkg->pd[i] && pol->pd_offline_fn)
394 pol->pd_offline_fn(blkg->pd[i]);
397 blkg->online = false;
399 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
400 list_del_init(&blkg->q_node);
401 hlist_del_init_rcu(&blkg->blkcg_node);
404 * Both setting lookup hint to and clearing it from @blkg are done
405 * under queue_lock. If it's not pointing to @blkg now, it never
406 * will. Hint assignment itself can race safely.
408 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
409 rcu_assign_pointer(blkcg->blkg_hint, NULL);
412 * Put the reference taken at the time of creation so that when all
413 * queues are gone, group can be destroyed.
415 percpu_ref_kill(&blkg->refcnt);
419 * blkg_destroy_all - destroy all blkgs associated with a request_queue
420 * @q: request_queue of interest
422 * Destroy all blkgs associated with @q.
424 static void blkg_destroy_all(struct request_queue *q)
426 struct blkcg_gq *blkg, *n;
427 int count = BLKG_DESTROY_BATCH_SIZE;
430 spin_lock_irq(&q->queue_lock);
431 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
432 struct blkcg *blkcg = blkg->blkcg;
434 spin_lock(&blkcg->lock);
436 spin_unlock(&blkcg->lock);
439 * in order to avoid holding the spin lock for too long, release
440 * it when a batch of blkgs are destroyed.
443 count = BLKG_DESTROY_BATCH_SIZE;
444 spin_unlock_irq(&q->queue_lock);
451 spin_unlock_irq(&q->queue_lock);
454 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
455 struct cftype *cftype, u64 val)
457 struct blkcg *blkcg = css_to_blkcg(css);
458 struct blkcg_gq *blkg;
461 mutex_lock(&blkcg_pol_mutex);
462 spin_lock_irq(&blkcg->lock);
465 * Note that stat reset is racy - it doesn't synchronize against
466 * stat updates. This is a debug feature which shouldn't exist
467 * anyway. If you get hit by a race, retry.
469 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
470 for_each_possible_cpu(cpu) {
471 struct blkg_iostat_set *bis =
472 per_cpu_ptr(blkg->iostat_cpu, cpu);
473 memset(bis, 0, sizeof(*bis));
475 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
477 for (i = 0; i < BLKCG_MAX_POLS; i++) {
478 struct blkcg_policy *pol = blkcg_policy[i];
480 if (blkg->pd[i] && pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg->pd[i]);
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
490 const char *blkg_dev_name(struct blkcg_gq *blkg)
492 if (!queue_has_disk(blkg->q) || !queue_to_disk(blkg->q)->bdi->dev)
494 return bdi_dev_name(queue_to_disk(blkg->q)->bdi);
498 * blkcg_print_blkgs - helper for printing per-blkg data
499 * @sf: seq_file to print to
500 * @blkcg: blkcg of interest
501 * @prfill: fill function to print out a blkg
502 * @pol: policy in question
503 * @data: data to be passed to @prfill
504 * @show_total: to print out sum of prfill return values or not
506 * This function invokes @prfill on each blkg of @blkcg if pd for the
507 * policy specified by @pol exists. @prfill is invoked with @sf, the
508 * policy data and @data and the matching queue lock held. If @show_total
509 * is %true, the sum of the return values from @prfill is printed with
510 * "Total" label at the end.
512 * This is to be used to construct print functions for
513 * cftype->read_seq_string method.
515 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
516 u64 (*prfill)(struct seq_file *,
517 struct blkg_policy_data *, int),
518 const struct blkcg_policy *pol, int data,
521 struct blkcg_gq *blkg;
525 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
526 spin_lock_irq(&blkg->q->queue_lock);
527 if (blkcg_policy_enabled(blkg->q, pol))
528 total += prfill(sf, blkg->pd[pol->plid], data);
529 spin_unlock_irq(&blkg->q->queue_lock);
534 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
536 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
539 * __blkg_prfill_u64 - prfill helper for a single u64 value
540 * @sf: seq_file to print to
541 * @pd: policy private data of interest
544 * Print @v to @sf for the device assocaited with @pd.
546 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
548 const char *dname = blkg_dev_name(pd->blkg);
553 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
556 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
558 /* Performs queue bypass and policy enabled checks then looks up blkg. */
559 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
560 const struct blkcg_policy *pol,
561 struct request_queue *q)
563 WARN_ON_ONCE(!rcu_read_lock_held());
564 lockdep_assert_held(&q->queue_lock);
566 if (!blkcg_policy_enabled(q, pol))
567 return ERR_PTR(-EOPNOTSUPP);
568 return __blkg_lookup(blkcg, q, true /* update_hint */);
572 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
573 * @inputp: input string pointer
575 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
576 * from @input and get and return the matching bdev. *@inputp is
577 * updated to point past the device node prefix. Returns an ERR_PTR()
580 * Use this function iff blkg_conf_prep() can't be used for some reason.
582 struct block_device *blkcg_conf_open_bdev(char **inputp)
584 char *input = *inputp;
585 unsigned int major, minor;
586 struct block_device *bdev;
589 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
590 return ERR_PTR(-EINVAL);
593 if (!isspace(*input))
594 return ERR_PTR(-EINVAL);
595 input = skip_spaces(input);
597 bdev = blkdev_get_no_open(MKDEV(major, minor));
599 return ERR_PTR(-ENODEV);
600 if (bdev_is_partition(bdev)) {
601 blkdev_put_no_open(bdev);
602 return ERR_PTR(-ENODEV);
610 * blkg_conf_prep - parse and prepare for per-blkg config update
611 * @blkcg: target block cgroup
612 * @pol: target policy
613 * @input: input string
614 * @ctx: blkg_conf_ctx to be filled
616 * Parse per-blkg config update from @input and initialize @ctx with the
617 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
618 * part of @input following MAJ:MIN. This function returns with RCU read
619 * lock and queue lock held and must be paired with blkg_conf_finish().
621 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
622 char *input, struct blkg_conf_ctx *ctx)
623 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
625 struct block_device *bdev;
626 struct request_queue *q;
627 struct blkcg_gq *blkg;
630 bdev = blkcg_conf_open_bdev(&input);
632 return PTR_ERR(bdev);
634 q = bdev->bd_disk->queue;
637 spin_lock_irq(&q->queue_lock);
639 blkg = blkg_lookup_check(blkcg, pol, q);
649 * Create blkgs walking down from blkcg_root to @blkcg, so that all
650 * non-root blkgs have access to their parents.
653 struct blkcg *pos = blkcg;
654 struct blkcg *parent;
655 struct blkcg_gq *new_blkg;
657 parent = blkcg_parent(blkcg);
658 while (parent && !__blkg_lookup(parent, q, false)) {
660 parent = blkcg_parent(parent);
663 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
664 spin_unlock_irq(&q->queue_lock);
667 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
668 if (unlikely(!new_blkg)) {
673 if (radix_tree_preload(GFP_KERNEL)) {
680 spin_lock_irq(&q->queue_lock);
682 blkg = blkg_lookup_check(pos, pol, q);
692 blkg = blkg_create(pos, q, new_blkg);
699 radix_tree_preload_end();
711 radix_tree_preload_end();
713 spin_unlock_irq(&q->queue_lock);
716 blkdev_put_no_open(bdev);
718 * If queue was bypassing, we should retry. Do so after a
719 * short msleep(). It isn't strictly necessary but queue
720 * can be bypassing for some time and it's always nice to
721 * avoid busy looping.
725 ret = restart_syscall();
729 EXPORT_SYMBOL_GPL(blkg_conf_prep);
732 * blkg_conf_finish - finish up per-blkg config update
733 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
735 * Finish up after per-blkg config update. This function must be paired
736 * with blkg_conf_prep().
738 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
739 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
741 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
743 blkdev_put_no_open(ctx->bdev);
745 EXPORT_SYMBOL_GPL(blkg_conf_finish);
747 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
751 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
752 dst->bytes[i] = src->bytes[i];
753 dst->ios[i] = src->ios[i];
757 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
761 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
762 dst->bytes[i] += src->bytes[i];
763 dst->ios[i] += src->ios[i];
767 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
771 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
772 dst->bytes[i] -= src->bytes[i];
773 dst->ios[i] -= src->ios[i];
777 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
779 struct blkcg *blkcg = css_to_blkcg(css);
780 struct blkcg_gq *blkg;
782 /* Root-level stats are sourced from system-wide IO stats */
783 if (!cgroup_parent(css->cgroup))
788 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
789 struct blkcg_gq *parent = blkg->parent;
790 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
791 struct blkg_iostat cur, delta;
794 /* fetch the current per-cpu values */
796 seq = u64_stats_fetch_begin(&bisc->sync);
797 blkg_iostat_set(&cur, &bisc->cur);
798 } while (u64_stats_fetch_retry(&bisc->sync, seq));
800 /* propagate percpu delta to global */
801 u64_stats_update_begin(&blkg->iostat.sync);
802 blkg_iostat_set(&delta, &cur);
803 blkg_iostat_sub(&delta, &bisc->last);
804 blkg_iostat_add(&blkg->iostat.cur, &delta);
805 blkg_iostat_add(&bisc->last, &delta);
806 u64_stats_update_end(&blkg->iostat.sync);
808 /* propagate global delta to parent (unless that's root) */
809 if (parent && parent->parent) {
810 u64_stats_update_begin(&parent->iostat.sync);
811 blkg_iostat_set(&delta, &blkg->iostat.cur);
812 blkg_iostat_sub(&delta, &blkg->iostat.last);
813 blkg_iostat_add(&parent->iostat.cur, &delta);
814 blkg_iostat_add(&blkg->iostat.last, &delta);
815 u64_stats_update_end(&parent->iostat.sync);
823 * We source root cgroup stats from the system-wide stats to avoid
824 * tracking the same information twice and incurring overhead when no
825 * cgroups are defined. For that reason, cgroup_rstat_flush in
826 * blkcg_print_stat does not actually fill out the iostat in the root
829 * However, we would like to re-use the printing code between the root and
830 * non-root cgroups to the extent possible. For that reason, we simulate
831 * flushing the root cgroup's stats by explicitly filling in the iostat
832 * with disk level statistics.
834 static void blkcg_fill_root_iostats(void)
836 struct class_dev_iter iter;
839 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
840 while ((dev = class_dev_iter_next(&iter))) {
841 struct block_device *bdev = dev_to_bdev(dev);
842 struct blkcg_gq *blkg =
843 blk_queue_root_blkg(bdev->bd_disk->queue);
844 struct blkg_iostat tmp;
847 memset(&tmp, 0, sizeof(tmp));
848 for_each_possible_cpu(cpu) {
849 struct disk_stats *cpu_dkstats;
851 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
852 tmp.ios[BLKG_IOSTAT_READ] +=
853 cpu_dkstats->ios[STAT_READ];
854 tmp.ios[BLKG_IOSTAT_WRITE] +=
855 cpu_dkstats->ios[STAT_WRITE];
856 tmp.ios[BLKG_IOSTAT_DISCARD] +=
857 cpu_dkstats->ios[STAT_DISCARD];
858 // convert sectors to bytes
859 tmp.bytes[BLKG_IOSTAT_READ] +=
860 cpu_dkstats->sectors[STAT_READ] << 9;
861 tmp.bytes[BLKG_IOSTAT_WRITE] +=
862 cpu_dkstats->sectors[STAT_WRITE] << 9;
863 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
864 cpu_dkstats->sectors[STAT_DISCARD] << 9;
866 u64_stats_update_begin(&blkg->iostat.sync);
867 blkg_iostat_set(&blkg->iostat.cur, &tmp);
868 u64_stats_update_end(&blkg->iostat.sync);
873 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
875 struct blkg_iostat_set *bis = &blkg->iostat;
876 u64 rbytes, wbytes, rios, wios, dbytes, dios;
877 bool has_stats = false;
885 dname = blkg_dev_name(blkg);
889 seq_printf(s, "%s ", dname);
892 seq = u64_stats_fetch_begin(&bis->sync);
894 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
895 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
896 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
897 rios = bis->cur.ios[BLKG_IOSTAT_READ];
898 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
899 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
900 } while (u64_stats_fetch_retry(&bis->sync, seq));
902 if (rbytes || wbytes || rios || wios) {
904 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
905 rbytes, wbytes, rios, wios,
909 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
911 seq_printf(s, " use_delay=%d delay_nsec=%llu",
912 atomic_read(&blkg->use_delay),
913 atomic64_read(&blkg->delay_nsec));
916 for (i = 0; i < BLKCG_MAX_POLS; i++) {
917 struct blkcg_policy *pol = blkcg_policy[i];
919 if (!blkg->pd[i] || !pol->pd_stat_fn)
922 if (pol->pd_stat_fn(blkg->pd[i], s))
930 static int blkcg_print_stat(struct seq_file *sf, void *v)
932 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
933 struct blkcg_gq *blkg;
935 if (!seq_css(sf)->parent)
936 blkcg_fill_root_iostats();
938 cgroup_rstat_flush(blkcg->css.cgroup);
941 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
942 spin_lock_irq(&blkg->q->queue_lock);
943 blkcg_print_one_stat(blkg, sf);
944 spin_unlock_irq(&blkg->q->queue_lock);
950 static struct cftype blkcg_files[] = {
953 .seq_show = blkcg_print_stat,
958 static struct cftype blkcg_legacy_files[] = {
960 .name = "reset_stats",
961 .write_u64 = blkcg_reset_stats,
967 * blkcg destruction is a three-stage process.
969 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
970 * which offlines writeback. Here we tie the next stage of blkg destruction
971 * to the completion of writeback associated with the blkcg. This lets us
972 * avoid punting potentially large amounts of outstanding writeback to root
973 * while maintaining any ongoing policies. The next stage is triggered when
974 * the nr_cgwbs count goes to zero.
976 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
977 * and handles the destruction of blkgs. Here the css reference held by
978 * the blkg is put back eventually allowing blkcg_css_free() to be called.
979 * This work may occur in cgwb_release_workfn() on the cgwb_release
980 * workqueue. Any submitted ios that fail to get the blkg ref will be
981 * punted to the root_blkg.
983 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
984 * This finally frees the blkcg.
988 * blkcg_css_offline - cgroup css_offline callback
989 * @css: css of interest
991 * This function is called when @css is about to go away. Here the cgwbs are
992 * offlined first and only once writeback associated with the blkcg has
993 * finished do we start step 2 (see above).
995 static void blkcg_css_offline(struct cgroup_subsys_state *css)
997 struct blkcg *blkcg = css_to_blkcg(css);
999 /* this prevents anyone from attaching or migrating to this blkcg */
1000 wb_blkcg_offline(blkcg);
1002 /* put the base online pin allowing step 2 to be triggered */
1003 blkcg_unpin_online(blkcg);
1007 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1008 * @blkcg: blkcg of interest
1010 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1011 * is nested inside q lock, this function performs reverse double lock dancing.
1012 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1013 * blkcg_css_free to eventually be called.
1015 * This is the blkcg counterpart of ioc_release_fn().
1017 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1021 spin_lock_irq(&blkcg->lock);
1023 while (!hlist_empty(&blkcg->blkg_list)) {
1024 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1025 struct blkcg_gq, blkcg_node);
1026 struct request_queue *q = blkg->q;
1028 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1030 * Given that the system can accumulate a huge number
1031 * of blkgs in pathological cases, check to see if we
1032 * need to rescheduling to avoid softlockup.
1034 spin_unlock_irq(&blkcg->lock);
1036 spin_lock_irq(&blkcg->lock);
1041 spin_unlock(&q->queue_lock);
1044 spin_unlock_irq(&blkcg->lock);
1047 static void blkcg_css_free(struct cgroup_subsys_state *css)
1049 struct blkcg *blkcg = css_to_blkcg(css);
1052 mutex_lock(&blkcg_pol_mutex);
1054 list_del(&blkcg->all_blkcgs_node);
1056 for (i = 0; i < BLKCG_MAX_POLS; i++)
1058 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1060 mutex_unlock(&blkcg_pol_mutex);
1065 static struct cgroup_subsys_state *
1066 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1068 struct blkcg *blkcg;
1069 struct cgroup_subsys_state *ret;
1072 mutex_lock(&blkcg_pol_mutex);
1075 blkcg = &blkcg_root;
1077 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1079 ret = ERR_PTR(-ENOMEM);
1084 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1085 struct blkcg_policy *pol = blkcg_policy[i];
1086 struct blkcg_policy_data *cpd;
1089 * If the policy hasn't been attached yet, wait for it
1090 * to be attached before doing anything else. Otherwise,
1091 * check if the policy requires any specific per-cgroup
1092 * data: if it does, allocate and initialize it.
1094 if (!pol || !pol->cpd_alloc_fn)
1097 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1099 ret = ERR_PTR(-ENOMEM);
1102 blkcg->cpd[i] = cpd;
1105 if (pol->cpd_init_fn)
1106 pol->cpd_init_fn(cpd);
1109 spin_lock_init(&blkcg->lock);
1110 refcount_set(&blkcg->online_pin, 1);
1111 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1112 INIT_HLIST_HEAD(&blkcg->blkg_list);
1113 #ifdef CONFIG_CGROUP_WRITEBACK
1114 INIT_LIST_HEAD(&blkcg->cgwb_list);
1116 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1118 mutex_unlock(&blkcg_pol_mutex);
1122 for (i--; i >= 0; i--)
1124 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1126 if (blkcg != &blkcg_root)
1129 mutex_unlock(&blkcg_pol_mutex);
1133 static int blkcg_css_online(struct cgroup_subsys_state *css)
1135 struct blkcg *blkcg = css_to_blkcg(css);
1136 struct blkcg *parent = blkcg_parent(blkcg);
1139 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1140 * don't go offline while cgwbs are still active on them. Pin the
1141 * parent so that offline always happens towards the root.
1144 blkcg_pin_online(parent);
1149 * blkcg_init_queue - initialize blkcg part of request queue
1150 * @q: request_queue to initialize
1152 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1153 * part of new request_queue @q.
1156 * 0 on success, -errno on failure.
1158 int blkcg_init_queue(struct request_queue *q)
1160 struct blkcg_gq *new_blkg, *blkg;
1164 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1168 preloaded = !radix_tree_preload(GFP_KERNEL);
1170 /* Make sure the root blkg exists. */
1172 spin_lock_irq(&q->queue_lock);
1173 blkg = blkg_create(&blkcg_root, q, new_blkg);
1176 q->root_blkg = blkg;
1177 spin_unlock_irq(&q->queue_lock);
1181 radix_tree_preload_end();
1183 ret = blk_iolatency_init(q);
1185 goto err_destroy_all;
1187 ret = blk_ioprio_init(q);
1189 goto err_destroy_all;
1191 ret = blk_throtl_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);
1220 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1224 mutex_lock(&blkcg_pol_mutex);
1226 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1227 struct blkcg_policy *pol = blkcg_policy[i];
1228 struct blkcg *blkcg;
1230 if (!pol || !pol->cpd_bind_fn)
1233 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1234 if (blkcg->cpd[pol->plid])
1235 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1237 mutex_unlock(&blkcg_pol_mutex);
1240 static void blkcg_exit(struct task_struct *tsk)
1242 if (tsk->throttle_queue)
1243 blk_put_queue(tsk->throttle_queue);
1244 tsk->throttle_queue = NULL;
1247 struct cgroup_subsys io_cgrp_subsys = {
1248 .css_alloc = blkcg_css_alloc,
1249 .css_online = blkcg_css_online,
1250 .css_offline = blkcg_css_offline,
1251 .css_free = blkcg_css_free,
1252 .css_rstat_flush = blkcg_rstat_flush,
1254 .dfl_cftypes = blkcg_files,
1255 .legacy_cftypes = blkcg_legacy_files,
1256 .legacy_name = "blkio",
1260 * This ensures that, if available, memcg is automatically enabled
1261 * together on the default hierarchy so that the owner cgroup can
1262 * be retrieved from writeback pages.
1264 .depends_on = 1 << memory_cgrp_id,
1267 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1270 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1271 * @q: request_queue of interest
1272 * @pol: blkcg policy to activate
1274 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1275 * bypass mode to populate its blkgs with policy_data for @pol.
1277 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1278 * from IO path. Update of each blkg is protected by both queue and blkcg
1279 * locks so that holding either lock and testing blkcg_policy_enabled() is
1280 * always enough for dereferencing policy data.
1282 * The caller is responsible for synchronizing [de]activations and policy
1283 * [un]registerations. Returns 0 on success, -errno on failure.
1285 int blkcg_activate_policy(struct request_queue *q,
1286 const struct blkcg_policy *pol)
1288 struct blkg_policy_data *pd_prealloc = NULL;
1289 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1292 if (blkcg_policy_enabled(q, pol))
1296 blk_mq_freeze_queue(q);
1298 spin_lock_irq(&q->queue_lock);
1300 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1301 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1302 struct blkg_policy_data *pd;
1304 if (blkg->pd[pol->plid])
1307 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1308 if (blkg == pinned_blkg) {
1312 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1318 * GFP_NOWAIT failed. Free the existing one and
1319 * prealloc for @blkg w/ GFP_KERNEL.
1322 blkg_put(pinned_blkg);
1326 spin_unlock_irq(&q->queue_lock);
1329 pol->pd_free_fn(pd_prealloc);
1330 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1338 blkg->pd[pol->plid] = pd;
1340 pd->plid = pol->plid;
1343 /* all allocated, init in the same order */
1344 if (pol->pd_init_fn)
1345 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1346 pol->pd_init_fn(blkg->pd[pol->plid]);
1348 __set_bit(pol->plid, q->blkcg_pols);
1351 spin_unlock_irq(&q->queue_lock);
1354 blk_mq_unfreeze_queue(q);
1356 blkg_put(pinned_blkg);
1358 pol->pd_free_fn(pd_prealloc);
1362 /* alloc failed, nothing's initialized yet, free everything */
1363 spin_lock_irq(&q->queue_lock);
1364 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1365 if (blkg->pd[pol->plid]) {
1366 pol->pd_free_fn(blkg->pd[pol->plid]);
1367 blkg->pd[pol->plid] = NULL;
1370 spin_unlock_irq(&q->queue_lock);
1374 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1377 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1378 * @q: request_queue of interest
1379 * @pol: blkcg policy to deactivate
1381 * Deactivate @pol on @q. Follows the same synchronization rules as
1382 * blkcg_activate_policy().
1384 void blkcg_deactivate_policy(struct request_queue *q,
1385 const struct blkcg_policy *pol)
1387 struct blkcg_gq *blkg;
1389 if (!blkcg_policy_enabled(q, pol))
1393 blk_mq_freeze_queue(q);
1395 spin_lock_irq(&q->queue_lock);
1397 __clear_bit(pol->plid, q->blkcg_pols);
1399 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1400 if (blkg->pd[pol->plid]) {
1401 if (pol->pd_offline_fn)
1402 pol->pd_offline_fn(blkg->pd[pol->plid]);
1403 pol->pd_free_fn(blkg->pd[pol->plid]);
1404 blkg->pd[pol->plid] = NULL;
1408 spin_unlock_irq(&q->queue_lock);
1411 blk_mq_unfreeze_queue(q);
1413 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1416 * blkcg_policy_register - register a blkcg policy
1417 * @pol: blkcg policy to register
1419 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1420 * successful registration. Returns 0 on success and -errno on failure.
1422 int blkcg_policy_register(struct blkcg_policy *pol)
1424 struct blkcg *blkcg;
1427 mutex_lock(&blkcg_pol_register_mutex);
1428 mutex_lock(&blkcg_pol_mutex);
1430 /* find an empty slot */
1432 for (i = 0; i < BLKCG_MAX_POLS; i++)
1433 if (!blkcg_policy[i])
1435 if (i >= BLKCG_MAX_POLS) {
1436 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1440 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1441 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1442 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1447 blkcg_policy[pol->plid] = pol;
1449 /* allocate and install cpd's */
1450 if (pol->cpd_alloc_fn) {
1451 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1452 struct blkcg_policy_data *cpd;
1454 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1458 blkcg->cpd[pol->plid] = cpd;
1460 cpd->plid = pol->plid;
1461 if (pol->cpd_init_fn)
1462 pol->cpd_init_fn(cpd);
1466 mutex_unlock(&blkcg_pol_mutex);
1468 /* everything is in place, add intf files for the new policy */
1469 if (pol->dfl_cftypes)
1470 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1472 if (pol->legacy_cftypes)
1473 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1474 pol->legacy_cftypes));
1475 mutex_unlock(&blkcg_pol_register_mutex);
1479 if (pol->cpd_free_fn) {
1480 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1481 if (blkcg->cpd[pol->plid]) {
1482 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1483 blkcg->cpd[pol->plid] = NULL;
1487 blkcg_policy[pol->plid] = NULL;
1489 mutex_unlock(&blkcg_pol_mutex);
1490 mutex_unlock(&blkcg_pol_register_mutex);
1493 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1496 * blkcg_policy_unregister - unregister a blkcg policy
1497 * @pol: blkcg policy to unregister
1499 * Undo blkcg_policy_register(@pol). Might sleep.
1501 void blkcg_policy_unregister(struct blkcg_policy *pol)
1503 struct blkcg *blkcg;
1505 mutex_lock(&blkcg_pol_register_mutex);
1507 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1510 /* kill the intf files first */
1511 if (pol->dfl_cftypes)
1512 cgroup_rm_cftypes(pol->dfl_cftypes);
1513 if (pol->legacy_cftypes)
1514 cgroup_rm_cftypes(pol->legacy_cftypes);
1516 /* remove cpds and unregister */
1517 mutex_lock(&blkcg_pol_mutex);
1519 if (pol->cpd_free_fn) {
1520 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1521 if (blkcg->cpd[pol->plid]) {
1522 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1523 blkcg->cpd[pol->plid] = NULL;
1527 blkcg_policy[pol->plid] = NULL;
1529 mutex_unlock(&blkcg_pol_mutex);
1531 mutex_unlock(&blkcg_pol_register_mutex);
1533 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1535 bool __blkcg_punt_bio_submit(struct bio *bio)
1537 struct blkcg_gq *blkg = bio->bi_blkg;
1539 /* consume the flag first */
1540 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1542 /* never bounce for the root cgroup */
1546 spin_lock_bh(&blkg->async_bio_lock);
1547 bio_list_add(&blkg->async_bios, bio);
1548 spin_unlock_bh(&blkg->async_bio_lock);
1550 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1555 * Scale the accumulated delay based on how long it has been since we updated
1556 * the delay. We only call this when we are adding delay, in case it's been a
1557 * while since we added delay, and when we are checking to see if we need to
1558 * delay a task, to account for any delays that may have occurred.
1560 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1562 u64 old = atomic64_read(&blkg->delay_start);
1564 /* negative use_delay means no scaling, see blkcg_set_delay() */
1565 if (atomic_read(&blkg->use_delay) < 0)
1569 * We only want to scale down every second. The idea here is that we
1570 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1571 * time window. We only want to throttle tasks for recent delay that
1572 * has occurred, in 1 second time windows since that's the maximum
1573 * things can be throttled. We save the current delay window in
1574 * blkg->last_delay so we know what amount is still left to be charged
1575 * to the blkg from this point onward. blkg->last_use keeps track of
1576 * the use_delay counter. The idea is if we're unthrottling the blkg we
1577 * are ok with whatever is happening now, and we can take away more of
1578 * the accumulated delay as we've already throttled enough that
1579 * everybody is happy with their IO latencies.
1581 if (time_before64(old + NSEC_PER_SEC, now) &&
1582 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1583 u64 cur = atomic64_read(&blkg->delay_nsec);
1584 u64 sub = min_t(u64, blkg->last_delay, now - old);
1585 int cur_use = atomic_read(&blkg->use_delay);
1588 * We've been unthrottled, subtract a larger chunk of our
1589 * accumulated delay.
1591 if (cur_use < blkg->last_use)
1592 sub = max_t(u64, sub, blkg->last_delay >> 1);
1595 * This shouldn't happen, but handle it anyway. Our delay_nsec
1596 * should only ever be growing except here where we subtract out
1597 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1598 * rather not end up with negative numbers.
1600 if (unlikely(cur < sub)) {
1601 atomic64_set(&blkg->delay_nsec, 0);
1602 blkg->last_delay = 0;
1604 atomic64_sub(sub, &blkg->delay_nsec);
1605 blkg->last_delay = cur - sub;
1607 blkg->last_use = cur_use;
1612 * This is called when we want to actually walk up the hierarchy and check to
1613 * see if we need to throttle, and then actually throttle if there is some
1614 * accumulated delay. This should only be called upon return to user space so
1615 * we're not holding some lock that would induce a priority inversion.
1617 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1619 unsigned long pflags;
1621 u64 now = ktime_to_ns(ktime_get());
1626 while (blkg->parent) {
1627 int use_delay = atomic_read(&blkg->use_delay);
1632 blkcg_scale_delay(blkg, now);
1633 this_delay = atomic64_read(&blkg->delay_nsec);
1634 if (this_delay > delay_nsec) {
1635 delay_nsec = this_delay;
1636 clamp = use_delay > 0;
1639 blkg = blkg->parent;
1646 * Let's not sleep for all eternity if we've amassed a huge delay.
1647 * Swapping or metadata IO can accumulate 10's of seconds worth of
1648 * delay, and we want userspace to be able to do _something_ so cap the
1649 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1650 * tasks will be delayed for 0.25 second for every syscall. If
1651 * blkcg_set_delay() was used as indicated by negative use_delay, the
1652 * caller is responsible for regulating the range.
1655 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1658 psi_memstall_enter(&pflags);
1660 exp = ktime_add_ns(now, delay_nsec);
1661 tok = io_schedule_prepare();
1663 __set_current_state(TASK_KILLABLE);
1664 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1666 } while (!fatal_signal_pending(current));
1667 io_schedule_finish(tok);
1670 psi_memstall_leave(&pflags);
1674 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1676 * This is only called if we've been marked with set_notify_resume(). Obviously
1677 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1678 * check to see if current->throttle_queue is set and if not this doesn't do
1679 * anything. This should only ever be called by the resume code, it's not meant
1680 * to be called by people willy-nilly as it will actually do the work to
1681 * throttle the task if it is setup for throttling.
1683 void blkcg_maybe_throttle_current(void)
1685 struct request_queue *q = current->throttle_queue;
1686 struct cgroup_subsys_state *css;
1687 struct blkcg *blkcg;
1688 struct blkcg_gq *blkg;
1689 bool use_memdelay = current->use_memdelay;
1694 current->throttle_queue = NULL;
1695 current->use_memdelay = false;
1698 css = kthread_blkcg();
1700 blkcg = css_to_blkcg(css);
1702 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1706 blkg = blkg_lookup(blkcg, q);
1709 if (!blkg_tryget(blkg))
1713 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1723 * blkcg_schedule_throttle - this task needs to check for throttling
1724 * @q: the request queue IO was submitted on
1725 * @use_memdelay: do we charge this to memory delay for PSI
1727 * This is called by the IO controller when we know there's delay accumulated
1728 * for the blkg for this task. We do not pass the blkg because there are places
1729 * we call this that may not have that information, the swapping code for
1730 * instance will only have a request_queue at that point. This set's the
1731 * notify_resume for the task to check and see if it requires throttling before
1732 * returning to user space.
1734 * We will only schedule once per syscall. You can call this over and over
1735 * again and it will only do the check once upon return to user space, and only
1736 * throttle once. If the task needs to be throttled again it'll need to be
1737 * re-set at the next time we see the task.
1739 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1741 if (unlikely(current->flags & PF_KTHREAD))
1744 if (current->throttle_queue != q) {
1745 if (!blk_get_queue(q))
1748 if (current->throttle_queue)
1749 blk_put_queue(current->throttle_queue);
1750 current->throttle_queue = q;
1754 current->use_memdelay = use_memdelay;
1755 set_notify_resume(current);
1759 * blkcg_add_delay - add delay to this blkg
1760 * @blkg: blkg of interest
1761 * @now: the current time in nanoseconds
1762 * @delta: how many nanoseconds of delay to add
1764 * Charge @delta to the blkg's current delay accumulation. This is used to
1765 * throttle tasks if an IO controller thinks we need more throttling.
1767 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1769 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1771 blkcg_scale_delay(blkg, now);
1772 atomic64_add(delta, &blkg->delay_nsec);
1776 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1780 * As the failure mode here is to walk up the blkg tree, this ensure that the
1781 * blkg->parent pointers are always valid. This returns the blkg that it ended
1782 * up taking a reference on or %NULL if no reference was taken.
1784 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1785 struct cgroup_subsys_state *css)
1787 struct blkcg_gq *blkg, *ret_blkg = NULL;
1790 blkg = blkg_lookup_create(css_to_blkcg(css),
1791 bio->bi_bdev->bd_disk->queue);
1793 if (blkg_tryget(blkg)) {
1797 blkg = blkg->parent;
1805 * bio_associate_blkg_from_css - associate a bio with a specified css
1809 * Associate @bio with the blkg found by combining the css's blkg and the
1810 * request_queue of the @bio. An association failure is handled by walking up
1811 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1812 * and q->root_blkg. This situation only happens when a cgroup is dying and
1813 * then the remaining bios will spill to the closest alive blkg.
1815 * A reference will be taken on the blkg and will be released when @bio is
1818 void bio_associate_blkg_from_css(struct bio *bio,
1819 struct cgroup_subsys_state *css)
1822 blkg_put(bio->bi_blkg);
1824 if (css && css->parent) {
1825 bio->bi_blkg = blkg_tryget_closest(bio, css);
1827 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg);
1828 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg;
1831 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1834 * bio_associate_blkg - associate a bio with a blkg
1837 * Associate @bio with the blkg found from the bio's css and request_queue.
1838 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1839 * already associated, the css is reused and association redone as the
1840 * request_queue may have changed.
1842 void bio_associate_blkg(struct bio *bio)
1844 struct cgroup_subsys_state *css;
1849 css = &bio_blkcg(bio)->css;
1853 bio_associate_blkg_from_css(bio, css);
1857 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1860 * bio_clone_blkg_association - clone blkg association from src to dst bio
1861 * @dst: destination bio
1864 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1868 blkg_put(dst->bi_blkg);
1869 blkg_get(src->bi_blkg);
1870 dst->bi_blkg = src->bi_blkg;
1873 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1875 static int blk_cgroup_io_type(struct bio *bio)
1877 if (op_is_discard(bio->bi_opf))
1878 return BLKG_IOSTAT_DISCARD;
1879 if (op_is_write(bio->bi_opf))
1880 return BLKG_IOSTAT_WRITE;
1881 return BLKG_IOSTAT_READ;
1884 void blk_cgroup_bio_start(struct bio *bio)
1886 int rwd = blk_cgroup_io_type(bio), cpu;
1887 struct blkg_iostat_set *bis;
1890 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1891 u64_stats_update_begin(&bis->sync);
1894 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1895 * bio and we would have already accounted for the size of the bio.
1897 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1898 bio_set_flag(bio, BIO_CGROUP_ACCT);
1899 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1901 bis->cur.ios[rwd]++;
1903 u64_stats_update_end(&bis->sync);
1904 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1905 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1909 static int __init blkcg_init(void)
1911 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1912 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1913 WQ_UNBOUND | WQ_SYSFS, 0);
1914 if (!blkcg_punt_bio_wq)
1918 subsys_initcall(blkcg_init);
1920 module_param(blkcg_debug_stats, bool, 0644);
1921 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");