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 /* some drivers (floppy) instantiate a queue w/o disk registered */
493 if (blkg->q->backing_dev_info->dev)
494 return bdi_dev_name(blkg->q->backing_dev_info);
499 * blkcg_print_blkgs - helper for printing per-blkg data
500 * @sf: seq_file to print to
501 * @blkcg: blkcg of interest
502 * @prfill: fill function to print out a blkg
503 * @pol: policy in question
504 * @data: data to be passed to @prfill
505 * @show_total: to print out sum of prfill return values or not
507 * This function invokes @prfill on each blkg of @blkcg if pd for the
508 * policy specified by @pol exists. @prfill is invoked with @sf, the
509 * policy data and @data and the matching queue lock held. If @show_total
510 * is %true, the sum of the return values from @prfill is printed with
511 * "Total" label at the end.
513 * This is to be used to construct print functions for
514 * cftype->read_seq_string method.
516 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
517 u64 (*prfill)(struct seq_file *,
518 struct blkg_policy_data *, int),
519 const struct blkcg_policy *pol, int data,
522 struct blkcg_gq *blkg;
526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
527 spin_lock_irq(&blkg->q->queue_lock);
528 if (blkcg_policy_enabled(blkg->q, pol))
529 total += prfill(sf, blkg->pd[pol->plid], data);
530 spin_unlock_irq(&blkg->q->queue_lock);
535 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
537 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
540 * __blkg_prfill_u64 - prfill helper for a single u64 value
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
545 * Print @v to @sf for the device assocaited with @pd.
547 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
549 const char *dname = blkg_dev_name(pd->blkg);
554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
557 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
559 /* Performs queue bypass and policy enabled checks then looks up blkg. */
560 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
561 const struct blkcg_policy *pol,
562 struct request_queue *q)
564 WARN_ON_ONCE(!rcu_read_lock_held());
565 lockdep_assert_held(&q->queue_lock);
567 if (!blkcg_policy_enabled(q, pol))
568 return ERR_PTR(-EOPNOTSUPP);
569 return __blkg_lookup(blkcg, q, true /* update_hint */);
573 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
574 * @inputp: input string pointer
576 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
577 * from @input and get and return the matching bdev. *@inputp is
578 * updated to point past the device node prefix. Returns an ERR_PTR()
581 * Use this function iff blkg_conf_prep() can't be used for some reason.
583 struct block_device *blkcg_conf_open_bdev(char **inputp)
585 char *input = *inputp;
586 unsigned int major, minor;
587 struct block_device *bdev;
590 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
591 return ERR_PTR(-EINVAL);
594 if (!isspace(*input))
595 return ERR_PTR(-EINVAL);
596 input = skip_spaces(input);
598 bdev = blkdev_get_no_open(MKDEV(major, minor));
600 return ERR_PTR(-ENODEV);
601 if (bdev_is_partition(bdev)) {
602 blkdev_put_no_open(bdev);
603 return ERR_PTR(-ENODEV);
611 * blkg_conf_prep - parse and prepare for per-blkg config update
612 * @blkcg: target block cgroup
613 * @pol: target policy
614 * @input: input string
615 * @ctx: blkg_conf_ctx to be filled
617 * Parse per-blkg config update from @input and initialize @ctx with the
618 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
619 * part of @input following MAJ:MIN. This function returns with RCU read
620 * lock and queue lock held and must be paired with blkg_conf_finish().
622 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
623 char *input, struct blkg_conf_ctx *ctx)
624 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
626 struct block_device *bdev;
627 struct request_queue *q;
628 struct blkcg_gq *blkg;
631 bdev = blkcg_conf_open_bdev(&input);
633 return PTR_ERR(bdev);
635 q = bdev->bd_disk->queue;
638 spin_lock_irq(&q->queue_lock);
640 blkg = blkg_lookup_check(blkcg, pol, q);
650 * Create blkgs walking down from blkcg_root to @blkcg, so that all
651 * non-root blkgs have access to their parents.
654 struct blkcg *pos = blkcg;
655 struct blkcg *parent;
656 struct blkcg_gq *new_blkg;
658 parent = blkcg_parent(blkcg);
659 while (parent && !__blkg_lookup(parent, q, false)) {
661 parent = blkcg_parent(parent);
664 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
665 spin_unlock_irq(&q->queue_lock);
668 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
669 if (unlikely(!new_blkg)) {
674 if (radix_tree_preload(GFP_KERNEL)) {
681 spin_lock_irq(&q->queue_lock);
683 blkg = blkg_lookup_check(pos, pol, q);
693 blkg = blkg_create(pos, q, new_blkg);
700 radix_tree_preload_end();
712 radix_tree_preload_end();
714 spin_unlock_irq(&q->queue_lock);
717 blkdev_put_no_open(bdev);
719 * If queue was bypassing, we should retry. Do so after a
720 * short msleep(). It isn't strictly necessary but queue
721 * can be bypassing for some time and it's always nice to
722 * avoid busy looping.
726 ret = restart_syscall();
730 EXPORT_SYMBOL_GPL(blkg_conf_prep);
733 * blkg_conf_finish - finish up per-blkg config update
734 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
736 * Finish up after per-blkg config update. This function must be paired
737 * with blkg_conf_prep().
739 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
740 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
742 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
744 blkdev_put_no_open(ctx->bdev);
746 EXPORT_SYMBOL_GPL(blkg_conf_finish);
748 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
752 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
753 dst->bytes[i] = src->bytes[i];
754 dst->ios[i] = src->ios[i];
758 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
762 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
763 dst->bytes[i] += src->bytes[i];
764 dst->ios[i] += src->ios[i];
768 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
772 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
773 dst->bytes[i] -= src->bytes[i];
774 dst->ios[i] -= src->ios[i];
778 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
780 struct blkcg *blkcg = css_to_blkcg(css);
781 struct blkcg_gq *blkg;
783 /* Root-level stats are sourced from system-wide IO stats */
784 if (!cgroup_parent(css->cgroup))
789 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
790 struct blkcg_gq *parent = blkg->parent;
791 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
792 struct blkg_iostat cur, delta;
795 /* fetch the current per-cpu values */
797 seq = u64_stats_fetch_begin(&bisc->sync);
798 blkg_iostat_set(&cur, &bisc->cur);
799 } while (u64_stats_fetch_retry(&bisc->sync, seq));
801 /* propagate percpu delta to global */
802 u64_stats_update_begin(&blkg->iostat.sync);
803 blkg_iostat_set(&delta, &cur);
804 blkg_iostat_sub(&delta, &bisc->last);
805 blkg_iostat_add(&blkg->iostat.cur, &delta);
806 blkg_iostat_add(&bisc->last, &delta);
807 u64_stats_update_end(&blkg->iostat.sync);
809 /* propagate global delta to parent (unless that's root) */
810 if (parent && parent->parent) {
811 u64_stats_update_begin(&parent->iostat.sync);
812 blkg_iostat_set(&delta, &blkg->iostat.cur);
813 blkg_iostat_sub(&delta, &blkg->iostat.last);
814 blkg_iostat_add(&parent->iostat.cur, &delta);
815 blkg_iostat_add(&blkg->iostat.last, &delta);
816 u64_stats_update_end(&parent->iostat.sync);
824 * We source root cgroup stats from the system-wide stats to avoid
825 * tracking the same information twice and incurring overhead when no
826 * cgroups are defined. For that reason, cgroup_rstat_flush in
827 * blkcg_print_stat does not actually fill out the iostat in the root
830 * However, we would like to re-use the printing code between the root and
831 * non-root cgroups to the extent possible. For that reason, we simulate
832 * flushing the root cgroup's stats by explicitly filling in the iostat
833 * with disk level statistics.
835 static void blkcg_fill_root_iostats(void)
837 struct class_dev_iter iter;
840 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
841 while ((dev = class_dev_iter_next(&iter))) {
842 struct block_device *bdev = dev_to_bdev(dev);
843 struct blkcg_gq *blkg =
844 blk_queue_root_blkg(bdev->bd_disk->queue);
845 struct blkg_iostat tmp;
848 memset(&tmp, 0, sizeof(tmp));
849 for_each_possible_cpu(cpu) {
850 struct disk_stats *cpu_dkstats;
852 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
853 tmp.ios[BLKG_IOSTAT_READ] +=
854 cpu_dkstats->ios[STAT_READ];
855 tmp.ios[BLKG_IOSTAT_WRITE] +=
856 cpu_dkstats->ios[STAT_WRITE];
857 tmp.ios[BLKG_IOSTAT_DISCARD] +=
858 cpu_dkstats->ios[STAT_DISCARD];
859 // convert sectors to bytes
860 tmp.bytes[BLKG_IOSTAT_READ] +=
861 cpu_dkstats->sectors[STAT_READ] << 9;
862 tmp.bytes[BLKG_IOSTAT_WRITE] +=
863 cpu_dkstats->sectors[STAT_WRITE] << 9;
864 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
865 cpu_dkstats->sectors[STAT_DISCARD] << 9;
867 u64_stats_update_begin(&blkg->iostat.sync);
868 blkg_iostat_set(&blkg->iostat.cur, &tmp);
869 u64_stats_update_end(&blkg->iostat.sync);
874 static int blkcg_print_stat(struct seq_file *sf, void *v)
876 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
877 struct blkcg_gq *blkg;
879 if (!seq_css(sf)->parent)
880 blkcg_fill_root_iostats();
882 cgroup_rstat_flush(blkcg->css.cgroup);
886 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
887 struct blkg_iostat_set *bis = &blkg->iostat;
890 u64 rbytes, wbytes, rios, wios, dbytes, dios;
891 size_t size = seq_get_buf(sf, &buf), off = 0;
893 bool has_stats = false;
896 spin_lock_irq(&blkg->q->queue_lock);
901 dname = blkg_dev_name(blkg);
906 * Hooray string manipulation, count is the size written NOT
907 * INCLUDING THE \0, so size is now count+1 less than what we
908 * had before, but we want to start writing the next bit from
909 * the \0 so we only add count to buf.
911 off += scnprintf(buf+off, size-off, "%s ", dname);
914 seq = u64_stats_fetch_begin(&bis->sync);
916 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
917 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
918 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
919 rios = bis->cur.ios[BLKG_IOSTAT_READ];
920 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
921 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
922 } while (u64_stats_fetch_retry(&bis->sync, seq));
924 if (rbytes || wbytes || rios || wios) {
926 off += scnprintf(buf+off, size-off,
927 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
928 rbytes, wbytes, rios, wios,
932 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
934 off += scnprintf(buf+off, size-off,
935 " use_delay=%d delay_nsec=%llu",
936 atomic_read(&blkg->use_delay),
937 (unsigned long long)atomic64_read(&blkg->delay_nsec));
940 for (i = 0; i < BLKCG_MAX_POLS; i++) {
941 struct blkcg_policy *pol = blkcg_policy[i];
944 if (!blkg->pd[i] || !pol->pd_stat_fn)
947 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
954 if (off < size - 1) {
955 off += scnprintf(buf+off, size-off, "\n");
962 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 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1187 preloaded = !radix_tree_preload(GFP_KERNEL);
1189 /* Make sure the root blkg exists. */
1191 spin_lock_irq(&q->queue_lock);
1192 blkg = blkg_create(&blkcg_root, q, new_blkg);
1195 q->root_blkg = blkg;
1196 spin_unlock_irq(&q->queue_lock);
1200 radix_tree_preload_end();
1202 ret = blk_iolatency_init(q);
1204 goto err_destroy_all;
1206 ret = blk_ioprio_init(q);
1208 goto err_destroy_all;
1210 ret = blk_throtl_init(q);
1212 goto err_destroy_all;
1217 blkg_destroy_all(q);
1220 spin_unlock_irq(&q->queue_lock);
1223 radix_tree_preload_end();
1224 return PTR_ERR(blkg);
1228 * blkcg_exit_queue - exit and release blkcg part of request_queue
1229 * @q: request_queue being released
1231 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1233 void blkcg_exit_queue(struct request_queue *q)
1235 blkg_destroy_all(q);
1239 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1243 mutex_lock(&blkcg_pol_mutex);
1245 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1246 struct blkcg_policy *pol = blkcg_policy[i];
1247 struct blkcg *blkcg;
1249 if (!pol || !pol->cpd_bind_fn)
1252 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1253 if (blkcg->cpd[pol->plid])
1254 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1256 mutex_unlock(&blkcg_pol_mutex);
1259 static void blkcg_exit(struct task_struct *tsk)
1261 if (tsk->throttle_queue)
1262 blk_put_queue(tsk->throttle_queue);
1263 tsk->throttle_queue = NULL;
1266 struct cgroup_subsys io_cgrp_subsys = {
1267 .css_alloc = blkcg_css_alloc,
1268 .css_online = blkcg_css_online,
1269 .css_offline = blkcg_css_offline,
1270 .css_free = blkcg_css_free,
1271 .css_rstat_flush = blkcg_rstat_flush,
1273 .dfl_cftypes = blkcg_files,
1274 .legacy_cftypes = blkcg_legacy_files,
1275 .legacy_name = "blkio",
1279 * This ensures that, if available, memcg is automatically enabled
1280 * together on the default hierarchy so that the owner cgroup can
1281 * be retrieved from writeback pages.
1283 .depends_on = 1 << memory_cgrp_id,
1286 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1289 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1290 * @q: request_queue of interest
1291 * @pol: blkcg policy to activate
1293 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1294 * bypass mode to populate its blkgs with policy_data for @pol.
1296 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1297 * from IO path. Update of each blkg is protected by both queue and blkcg
1298 * locks so that holding either lock and testing blkcg_policy_enabled() is
1299 * always enough for dereferencing policy data.
1301 * The caller is responsible for synchronizing [de]activations and policy
1302 * [un]registerations. Returns 0 on success, -errno on failure.
1304 int blkcg_activate_policy(struct request_queue *q,
1305 const struct blkcg_policy *pol)
1307 struct blkg_policy_data *pd_prealloc = NULL;
1308 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1311 if (blkcg_policy_enabled(q, pol))
1315 blk_mq_freeze_queue(q);
1317 spin_lock_irq(&q->queue_lock);
1319 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1320 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1321 struct blkg_policy_data *pd;
1323 if (blkg->pd[pol->plid])
1326 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1327 if (blkg == pinned_blkg) {
1331 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1337 * GFP_NOWAIT failed. Free the existing one and
1338 * prealloc for @blkg w/ GFP_KERNEL.
1341 blkg_put(pinned_blkg);
1345 spin_unlock_irq(&q->queue_lock);
1348 pol->pd_free_fn(pd_prealloc);
1349 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1357 blkg->pd[pol->plid] = pd;
1359 pd->plid = pol->plid;
1362 /* all allocated, init in the same order */
1363 if (pol->pd_init_fn)
1364 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1365 pol->pd_init_fn(blkg->pd[pol->plid]);
1367 __set_bit(pol->plid, q->blkcg_pols);
1370 spin_unlock_irq(&q->queue_lock);
1373 blk_mq_unfreeze_queue(q);
1375 blkg_put(pinned_blkg);
1377 pol->pd_free_fn(pd_prealloc);
1381 /* alloc failed, nothing's initialized yet, free everything */
1382 spin_lock_irq(&q->queue_lock);
1383 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1384 if (blkg->pd[pol->plid]) {
1385 pol->pd_free_fn(blkg->pd[pol->plid]);
1386 blkg->pd[pol->plid] = NULL;
1389 spin_unlock_irq(&q->queue_lock);
1393 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1396 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1397 * @q: request_queue of interest
1398 * @pol: blkcg policy to deactivate
1400 * Deactivate @pol on @q. Follows the same synchronization rules as
1401 * blkcg_activate_policy().
1403 void blkcg_deactivate_policy(struct request_queue *q,
1404 const struct blkcg_policy *pol)
1406 struct blkcg_gq *blkg;
1408 if (!blkcg_policy_enabled(q, pol))
1412 blk_mq_freeze_queue(q);
1414 spin_lock_irq(&q->queue_lock);
1416 __clear_bit(pol->plid, q->blkcg_pols);
1418 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1419 if (blkg->pd[pol->plid]) {
1420 if (pol->pd_offline_fn)
1421 pol->pd_offline_fn(blkg->pd[pol->plid]);
1422 pol->pd_free_fn(blkg->pd[pol->plid]);
1423 blkg->pd[pol->plid] = NULL;
1427 spin_unlock_irq(&q->queue_lock);
1430 blk_mq_unfreeze_queue(q);
1432 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1435 * blkcg_policy_register - register a blkcg policy
1436 * @pol: blkcg policy to register
1438 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1439 * successful registration. Returns 0 on success and -errno on failure.
1441 int blkcg_policy_register(struct blkcg_policy *pol)
1443 struct blkcg *blkcg;
1446 mutex_lock(&blkcg_pol_register_mutex);
1447 mutex_lock(&blkcg_pol_mutex);
1449 /* find an empty slot */
1451 for (i = 0; i < BLKCG_MAX_POLS; i++)
1452 if (!blkcg_policy[i])
1454 if (i >= BLKCG_MAX_POLS) {
1455 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1459 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1460 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1461 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1466 blkcg_policy[pol->plid] = pol;
1468 /* allocate and install cpd's */
1469 if (pol->cpd_alloc_fn) {
1470 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1471 struct blkcg_policy_data *cpd;
1473 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1477 blkcg->cpd[pol->plid] = cpd;
1479 cpd->plid = pol->plid;
1480 if (pol->cpd_init_fn)
1481 pol->cpd_init_fn(cpd);
1485 mutex_unlock(&blkcg_pol_mutex);
1487 /* everything is in place, add intf files for the new policy */
1488 if (pol->dfl_cftypes)
1489 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1491 if (pol->legacy_cftypes)
1492 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1493 pol->legacy_cftypes));
1494 mutex_unlock(&blkcg_pol_register_mutex);
1498 if (pol->cpd_free_fn) {
1499 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1500 if (blkcg->cpd[pol->plid]) {
1501 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1502 blkcg->cpd[pol->plid] = NULL;
1506 blkcg_policy[pol->plid] = NULL;
1508 mutex_unlock(&blkcg_pol_mutex);
1509 mutex_unlock(&blkcg_pol_register_mutex);
1512 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1515 * blkcg_policy_unregister - unregister a blkcg policy
1516 * @pol: blkcg policy to unregister
1518 * Undo blkcg_policy_register(@pol). Might sleep.
1520 void blkcg_policy_unregister(struct blkcg_policy *pol)
1522 struct blkcg *blkcg;
1524 mutex_lock(&blkcg_pol_register_mutex);
1526 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1529 /* kill the intf files first */
1530 if (pol->dfl_cftypes)
1531 cgroup_rm_cftypes(pol->dfl_cftypes);
1532 if (pol->legacy_cftypes)
1533 cgroup_rm_cftypes(pol->legacy_cftypes);
1535 /* remove cpds and unregister */
1536 mutex_lock(&blkcg_pol_mutex);
1538 if (pol->cpd_free_fn) {
1539 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1540 if (blkcg->cpd[pol->plid]) {
1541 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1542 blkcg->cpd[pol->plid] = NULL;
1546 blkcg_policy[pol->plid] = NULL;
1548 mutex_unlock(&blkcg_pol_mutex);
1550 mutex_unlock(&blkcg_pol_register_mutex);
1552 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1554 bool __blkcg_punt_bio_submit(struct bio *bio)
1556 struct blkcg_gq *blkg = bio->bi_blkg;
1558 /* consume the flag first */
1559 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1561 /* never bounce for the root cgroup */
1565 spin_lock_bh(&blkg->async_bio_lock);
1566 bio_list_add(&blkg->async_bios, bio);
1567 spin_unlock_bh(&blkg->async_bio_lock);
1569 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1574 * Scale the accumulated delay based on how long it has been since we updated
1575 * the delay. We only call this when we are adding delay, in case it's been a
1576 * while since we added delay, and when we are checking to see if we need to
1577 * delay a task, to account for any delays that may have occurred.
1579 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1581 u64 old = atomic64_read(&blkg->delay_start);
1583 /* negative use_delay means no scaling, see blkcg_set_delay() */
1584 if (atomic_read(&blkg->use_delay) < 0)
1588 * We only want to scale down every second. The idea here is that we
1589 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1590 * time window. We only want to throttle tasks for recent delay that
1591 * has occurred, in 1 second time windows since that's the maximum
1592 * things can be throttled. We save the current delay window in
1593 * blkg->last_delay so we know what amount is still left to be charged
1594 * to the blkg from this point onward. blkg->last_use keeps track of
1595 * the use_delay counter. The idea is if we're unthrottling the blkg we
1596 * are ok with whatever is happening now, and we can take away more of
1597 * the accumulated delay as we've already throttled enough that
1598 * everybody is happy with their IO latencies.
1600 if (time_before64(old + NSEC_PER_SEC, now) &&
1601 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1602 u64 cur = atomic64_read(&blkg->delay_nsec);
1603 u64 sub = min_t(u64, blkg->last_delay, now - old);
1604 int cur_use = atomic_read(&blkg->use_delay);
1607 * We've been unthrottled, subtract a larger chunk of our
1608 * accumulated delay.
1610 if (cur_use < blkg->last_use)
1611 sub = max_t(u64, sub, blkg->last_delay >> 1);
1614 * This shouldn't happen, but handle it anyway. Our delay_nsec
1615 * should only ever be growing except here where we subtract out
1616 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1617 * rather not end up with negative numbers.
1619 if (unlikely(cur < sub)) {
1620 atomic64_set(&blkg->delay_nsec, 0);
1621 blkg->last_delay = 0;
1623 atomic64_sub(sub, &blkg->delay_nsec);
1624 blkg->last_delay = cur - sub;
1626 blkg->last_use = cur_use;
1631 * This is called when we want to actually walk up the hierarchy and check to
1632 * see if we need to throttle, and then actually throttle if there is some
1633 * accumulated delay. This should only be called upon return to user space so
1634 * we're not holding some lock that would induce a priority inversion.
1636 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1638 unsigned long pflags;
1640 u64 now = ktime_to_ns(ktime_get());
1645 while (blkg->parent) {
1646 int use_delay = atomic_read(&blkg->use_delay);
1651 blkcg_scale_delay(blkg, now);
1652 this_delay = atomic64_read(&blkg->delay_nsec);
1653 if (this_delay > delay_nsec) {
1654 delay_nsec = this_delay;
1655 clamp = use_delay > 0;
1658 blkg = blkg->parent;
1665 * Let's not sleep for all eternity if we've amassed a huge delay.
1666 * Swapping or metadata IO can accumulate 10's of seconds worth of
1667 * delay, and we want userspace to be able to do _something_ so cap the
1668 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1669 * tasks will be delayed for 0.25 second for every syscall. If
1670 * blkcg_set_delay() was used as indicated by negative use_delay, the
1671 * caller is responsible for regulating the range.
1674 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1677 psi_memstall_enter(&pflags);
1679 exp = ktime_add_ns(now, delay_nsec);
1680 tok = io_schedule_prepare();
1682 __set_current_state(TASK_KILLABLE);
1683 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1685 } while (!fatal_signal_pending(current));
1686 io_schedule_finish(tok);
1689 psi_memstall_leave(&pflags);
1693 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1695 * This is only called if we've been marked with set_notify_resume(). Obviously
1696 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1697 * check to see if current->throttle_queue is set and if not this doesn't do
1698 * anything. This should only ever be called by the resume code, it's not meant
1699 * to be called by people willy-nilly as it will actually do the work to
1700 * throttle the task if it is setup for throttling.
1702 void blkcg_maybe_throttle_current(void)
1704 struct request_queue *q = current->throttle_queue;
1705 struct cgroup_subsys_state *css;
1706 struct blkcg *blkcg;
1707 struct blkcg_gq *blkg;
1708 bool use_memdelay = current->use_memdelay;
1713 current->throttle_queue = NULL;
1714 current->use_memdelay = false;
1717 css = kthread_blkcg();
1719 blkcg = css_to_blkcg(css);
1721 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1725 blkg = blkg_lookup(blkcg, q);
1728 if (!blkg_tryget(blkg))
1732 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1742 * blkcg_schedule_throttle - this task needs to check for throttling
1743 * @q: the request queue IO was submitted on
1744 * @use_memdelay: do we charge this to memory delay for PSI
1746 * This is called by the IO controller when we know there's delay accumulated
1747 * for the blkg for this task. We do not pass the blkg because there are places
1748 * we call this that may not have that information, the swapping code for
1749 * instance will only have a request_queue at that point. This set's the
1750 * notify_resume for the task to check and see if it requires throttling before
1751 * returning to user space.
1753 * We will only schedule once per syscall. You can call this over and over
1754 * again and it will only do the check once upon return to user space, and only
1755 * throttle once. If the task needs to be throttled again it'll need to be
1756 * re-set at the next time we see the task.
1758 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1760 if (unlikely(current->flags & PF_KTHREAD))
1763 if (current->throttle_queue != q) {
1764 if (!blk_get_queue(q))
1767 if (current->throttle_queue)
1768 blk_put_queue(current->throttle_queue);
1769 current->throttle_queue = q;
1773 current->use_memdelay = use_memdelay;
1774 set_notify_resume(current);
1778 * blkcg_add_delay - add delay to this blkg
1779 * @blkg: blkg of interest
1780 * @now: the current time in nanoseconds
1781 * @delta: how many nanoseconds of delay to add
1783 * Charge @delta to the blkg's current delay accumulation. This is used to
1784 * throttle tasks if an IO controller thinks we need more throttling.
1786 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1788 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1790 blkcg_scale_delay(blkg, now);
1791 atomic64_add(delta, &blkg->delay_nsec);
1795 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1799 * As the failure mode here is to walk up the blkg tree, this ensure that the
1800 * blkg->parent pointers are always valid. This returns the blkg that it ended
1801 * up taking a reference on or %NULL if no reference was taken.
1803 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1804 struct cgroup_subsys_state *css)
1806 struct blkcg_gq *blkg, *ret_blkg = NULL;
1809 blkg = blkg_lookup_create(css_to_blkcg(css),
1810 bio->bi_bdev->bd_disk->queue);
1812 if (blkg_tryget(blkg)) {
1816 blkg = blkg->parent;
1824 * bio_associate_blkg_from_css - associate a bio with a specified css
1828 * Associate @bio with the blkg found by combining the css's blkg and the
1829 * request_queue of the @bio. An association failure is handled by walking up
1830 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1831 * and q->root_blkg. This situation only happens when a cgroup is dying and
1832 * then the remaining bios will spill to the closest alive blkg.
1834 * A reference will be taken on the blkg and will be released when @bio is
1837 void bio_associate_blkg_from_css(struct bio *bio,
1838 struct cgroup_subsys_state *css)
1841 blkg_put(bio->bi_blkg);
1843 if (css && css->parent) {
1844 bio->bi_blkg = blkg_tryget_closest(bio, css);
1846 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg);
1847 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg;
1850 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1853 * bio_associate_blkg - associate a bio with a blkg
1856 * Associate @bio with the blkg found from the bio's css and request_queue.
1857 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1858 * already associated, the css is reused and association redone as the
1859 * request_queue may have changed.
1861 void bio_associate_blkg(struct bio *bio)
1863 struct cgroup_subsys_state *css;
1868 css = &bio_blkcg(bio)->css;
1872 bio_associate_blkg_from_css(bio, css);
1876 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1879 * bio_clone_blkg_association - clone blkg association from src to dst bio
1880 * @dst: destination bio
1883 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1887 blkg_put(dst->bi_blkg);
1888 blkg_get(src->bi_blkg);
1889 dst->bi_blkg = src->bi_blkg;
1892 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1894 static int blk_cgroup_io_type(struct bio *bio)
1896 if (op_is_discard(bio->bi_opf))
1897 return BLKG_IOSTAT_DISCARD;
1898 if (op_is_write(bio->bi_opf))
1899 return BLKG_IOSTAT_WRITE;
1900 return BLKG_IOSTAT_READ;
1903 void blk_cgroup_bio_start(struct bio *bio)
1905 int rwd = blk_cgroup_io_type(bio), cpu;
1906 struct blkg_iostat_set *bis;
1909 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1910 u64_stats_update_begin(&bis->sync);
1913 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1914 * bio and we would have already accounted for the size of the bio.
1916 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1917 bio_set_flag(bio, BIO_CGROUP_ACCT);
1918 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1920 bis->cur.ios[rwd]++;
1922 u64_stats_update_end(&bis->sync);
1923 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1924 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1928 static int __init blkcg_init(void)
1930 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1931 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1932 WQ_UNBOUND | WQ_SYSFS, 0);
1933 if (!blkcg_punt_bio_wq)
1937 subsys_initcall(blkcg_init);
1939 module_param(blkcg_debug_stats, bool, 0644);
1940 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");