1 // SPDX-License-Identifier: GPL-2.0
3 * Common Block IO controller cgroup interface
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
18 #include <linux/ioprio.h>
19 #include <linux/kdev_t.h>
20 #include <linux/module.h>
21 #include <linux/sched/signal.h>
22 #include <linux/err.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/slab.h>
26 #include <linux/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/resume_user_mode.h>
30 #include <linux/psi.h>
31 #include <linux/part_stat.h>
33 #include "blk-cgroup.h"
34 #include "blk-ioprio.h"
35 #include "blk-throttle.h"
36 #include "blk-rq-qos.h"
39 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
40 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
41 * policy [un]register operations including cgroup file additions /
42 * removals. Putting cgroup file registration outside blkcg_pol_mutex
43 * allows grabbing it from cgroup callbacks.
45 static DEFINE_MUTEX(blkcg_pol_register_mutex);
46 static DEFINE_MUTEX(blkcg_pol_mutex);
48 struct blkcg blkcg_root;
49 EXPORT_SYMBOL_GPL(blkcg_root);
51 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
52 EXPORT_SYMBOL_GPL(blkcg_root_css);
54 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
56 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
58 bool blkcg_debug_stats = false;
60 #define BLKG_DESTROY_BATCH_SIZE 64
63 * Lockless lists for tracking IO stats update
65 * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
66 * There are multiple blkg's (one for each block device) attached to each
67 * blkcg. The rstat code keeps track of which cpu has IO stats updated,
68 * but it doesn't know which blkg has the updated stats. If there are many
69 * block devices in a system, the cost of iterating all the blkg's to flush
70 * out the IO stats can be high. To reduce such overhead, a set of percpu
71 * lockless lists (lhead) per blkcg are used to track the set of recently
72 * updated iostat_cpu's since the last flush. An iostat_cpu will be put
73 * onto the lockless list on the update side [blk_cgroup_bio_start()] if
74 * not there yet and then removed when being flushed [blkcg_rstat_flush()].
75 * References to blkg are gotten and then put back in the process to
76 * protect against blkg removal.
78 * Return: 0 if successful or -ENOMEM if allocation fails.
80 static int init_blkcg_llists(struct blkcg *blkcg)
84 blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL);
88 for_each_possible_cpu(cpu)
89 init_llist_head(per_cpu_ptr(blkcg->lhead, cpu));
94 * blkcg_css - find the current css
96 * Find the css associated with either the kthread or the current task.
97 * This may return a dying css, so it is up to the caller to use tryget logic
98 * to confirm it is alive and well.
100 static struct cgroup_subsys_state *blkcg_css(void)
102 struct cgroup_subsys_state *css;
104 css = kthread_blkcg();
107 return task_css(current, io_cgrp_id);
110 static bool blkcg_policy_enabled(struct request_queue *q,
111 const struct blkcg_policy *pol)
113 return pol && test_bit(pol->plid, q->blkcg_pols);
116 static void blkg_free_workfn(struct work_struct *work)
118 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
120 struct request_queue *q = blkg->q;
124 * pd_free_fn() can also be called from blkcg_deactivate_policy(),
125 * in order to make sure pd_free_fn() is called in order, the deletion
126 * of the list blkg->q_node is delayed to here from blkg_destroy(), and
127 * blkcg_mutex is used to synchronize blkg_free_workfn() and
128 * blkcg_deactivate_policy().
130 mutex_lock(&q->blkcg_mutex);
131 for (i = 0; i < BLKCG_MAX_POLS; i++)
133 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
135 blkg_put(blkg->parent);
136 list_del_init(&blkg->q_node);
137 mutex_unlock(&q->blkcg_mutex);
140 free_percpu(blkg->iostat_cpu);
141 percpu_ref_exit(&blkg->refcnt);
146 * blkg_free - free a blkg
147 * @blkg: blkg to free
149 * Free @blkg which may be partially allocated.
151 static void blkg_free(struct blkcg_gq *blkg)
157 * Both ->pd_free_fn() and request queue's release handler may
158 * sleep, so free us by scheduling one work func
160 INIT_WORK(&blkg->free_work, blkg_free_workfn);
161 schedule_work(&blkg->free_work);
164 static void __blkg_release(struct rcu_head *rcu)
166 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
168 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
169 WARN_ON(!bio_list_empty(&blkg->async_bios));
172 /* release the blkcg and parent blkg refs this blkg has been holding */
173 css_put(&blkg->blkcg->css);
178 * A group is RCU protected, but having an rcu lock does not mean that one
179 * can access all the fields of blkg and assume these are valid. For
180 * example, don't try to follow throtl_data and request queue links.
182 * Having a reference to blkg under an rcu allows accesses to only values
183 * local to groups like group stats and group rate limits.
185 static void blkg_release(struct percpu_ref *ref)
187 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
189 call_rcu(&blkg->rcu_head, __blkg_release);
192 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
193 static struct workqueue_struct *blkcg_punt_bio_wq;
195 static void blkg_async_bio_workfn(struct work_struct *work)
197 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
199 struct bio_list bios = BIO_EMPTY_LIST;
201 struct blk_plug plug;
202 bool need_plug = false;
204 /* as long as there are pending bios, @blkg can't go away */
205 spin_lock(&blkg->async_bio_lock);
206 bio_list_merge(&bios, &blkg->async_bios);
207 bio_list_init(&blkg->async_bios);
208 spin_unlock(&blkg->async_bio_lock);
210 /* start plug only when bio_list contains at least 2 bios */
211 if (bios.head && bios.head->bi_next) {
213 blk_start_plug(&plug);
215 while ((bio = bio_list_pop(&bios)))
218 blk_finish_plug(&plug);
222 * When a shared kthread issues a bio for a cgroup, doing so synchronously can
223 * lead to priority inversions as the kthread can be trapped waiting for that
224 * cgroup. Use this helper instead of submit_bio to punt the actual issuing to
225 * a dedicated per-blkcg work item to avoid such priority inversions.
227 void blkcg_punt_bio_submit(struct bio *bio)
229 struct blkcg_gq *blkg = bio->bi_blkg;
232 spin_lock(&blkg->async_bio_lock);
233 bio_list_add(&blkg->async_bios, bio);
234 spin_unlock(&blkg->async_bio_lock);
235 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
237 /* never bounce for the root cgroup */
241 EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit);
243 static int __init blkcg_punt_bio_init(void)
245 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
246 WQ_MEM_RECLAIM | WQ_FREEZABLE |
247 WQ_UNBOUND | WQ_SYSFS, 0);
248 if (!blkcg_punt_bio_wq)
252 subsys_initcall(blkcg_punt_bio_init);
253 #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
256 * bio_blkcg_css - return the blkcg CSS associated with a bio
259 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
260 * associated. Callers are expected to either handle %NULL or know association
261 * has been done prior to calling this.
263 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
265 if (!bio || !bio->bi_blkg)
267 return &bio->bi_blkg->blkcg->css;
269 EXPORT_SYMBOL_GPL(bio_blkcg_css);
272 * blkcg_parent - get the parent of a blkcg
273 * @blkcg: blkcg of interest
275 * Return the parent blkcg of @blkcg. Can be called anytime.
277 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
279 return css_to_blkcg(blkcg->css.parent);
283 * blkg_alloc - allocate a blkg
284 * @blkcg: block cgroup the new blkg is associated with
285 * @disk: gendisk the new blkg is associated with
286 * @gfp_mask: allocation mask to use
288 * Allocate a new blkg assocating @blkcg and @q.
290 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
293 struct blkcg_gq *blkg;
296 /* alloc and init base part */
297 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
300 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
302 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
303 if (!blkg->iostat_cpu)
304 goto out_exit_refcnt;
305 if (!blk_get_queue(disk->queue))
306 goto out_free_iostat;
308 blkg->q = disk->queue;
309 INIT_LIST_HEAD(&blkg->q_node);
311 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
312 spin_lock_init(&blkg->async_bio_lock);
313 bio_list_init(&blkg->async_bios);
314 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
317 u64_stats_init(&blkg->iostat.sync);
318 for_each_possible_cpu(cpu) {
319 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
320 per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
323 for (i = 0; i < BLKCG_MAX_POLS; i++) {
324 struct blkcg_policy *pol = blkcg_policy[i];
325 struct blkg_policy_data *pd;
327 if (!blkcg_policy_enabled(disk->queue, pol))
330 /* alloc per-policy data and attach it to blkg */
331 pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
345 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
346 blk_put_queue(disk->queue);
348 free_percpu(blkg->iostat_cpu);
350 percpu_ref_exit(&blkg->refcnt);
357 * If @new_blkg is %NULL, this function tries to allocate a new one as
358 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
360 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
361 struct blkcg_gq *new_blkg)
363 struct blkcg_gq *blkg;
366 lockdep_assert_held(&disk->queue->queue_lock);
368 /* request_queue is dying, do not create/recreate a blkg */
369 if (blk_queue_dying(disk->queue)) {
374 /* blkg holds a reference to blkcg */
375 if (!css_tryget_online(&blkcg->css)) {
382 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
383 if (unlikely(!new_blkg)) {
391 if (blkcg_parent(blkcg)) {
392 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
393 if (WARN_ON_ONCE(!blkg->parent)) {
397 blkg_get(blkg->parent);
400 /* invoke per-policy init */
401 for (i = 0; i < BLKCG_MAX_POLS; i++) {
402 struct blkcg_policy *pol = blkcg_policy[i];
404 if (blkg->pd[i] && pol->pd_init_fn)
405 pol->pd_init_fn(blkg->pd[i]);
409 spin_lock(&blkcg->lock);
410 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
412 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
413 list_add(&blkg->q_node, &disk->queue->blkg_list);
415 for (i = 0; i < BLKCG_MAX_POLS; i++) {
416 struct blkcg_policy *pol = blkcg_policy[i];
419 if (pol->pd_online_fn)
420 pol->pd_online_fn(blkg->pd[i]);
421 blkg->pd[i]->online = true;
426 spin_unlock(&blkcg->lock);
431 /* @blkg failed fully initialized, use the usual release path */
436 css_put(&blkcg->css);
444 * blkg_lookup_create - lookup blkg, try to create one if not there
445 * @blkcg: blkcg of interest
446 * @disk: gendisk of interest
448 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
449 * create one. blkg creation is performed recursively from blkcg_root such
450 * that all non-root blkg's have access to the parent blkg. This function
451 * should be called under RCU read lock and takes @disk->queue->queue_lock.
453 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
456 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
457 struct gendisk *disk)
459 struct request_queue *q = disk->queue;
460 struct blkcg_gq *blkg;
463 WARN_ON_ONCE(!rcu_read_lock_held());
465 blkg = blkg_lookup(blkcg, q);
469 spin_lock_irqsave(&q->queue_lock, flags);
470 blkg = blkg_lookup(blkcg, q);
472 if (blkcg != &blkcg_root &&
473 blkg != rcu_dereference(blkcg->blkg_hint))
474 rcu_assign_pointer(blkcg->blkg_hint, blkg);
479 * Create blkgs walking down from blkcg_root to @blkcg, so that all
480 * non-root blkgs have access to their parents. Returns the closest
481 * blkg to the intended blkg should blkg_create() fail.
484 struct blkcg *pos = blkcg;
485 struct blkcg *parent = blkcg_parent(blkcg);
486 struct blkcg_gq *ret_blkg = q->root_blkg;
489 blkg = blkg_lookup(parent, q);
491 /* remember closest blkg */
496 parent = blkcg_parent(parent);
499 blkg = blkg_create(pos, disk, NULL);
509 spin_unlock_irqrestore(&q->queue_lock, flags);
513 static void blkg_destroy(struct blkcg_gq *blkg)
515 struct blkcg *blkcg = blkg->blkcg;
518 lockdep_assert_held(&blkg->q->queue_lock);
519 lockdep_assert_held(&blkcg->lock);
522 * blkg stays on the queue list until blkg_free_workfn(), see details in
523 * blkg_free_workfn(), hence this function can be called from
524 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
525 * blkg_free_workfn().
527 if (hlist_unhashed(&blkg->blkcg_node))
530 for (i = 0; i < BLKCG_MAX_POLS; i++) {
531 struct blkcg_policy *pol = blkcg_policy[i];
533 if (blkg->pd[i] && blkg->pd[i]->online) {
534 blkg->pd[i]->online = false;
535 if (pol->pd_offline_fn)
536 pol->pd_offline_fn(blkg->pd[i]);
540 blkg->online = false;
542 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
543 hlist_del_init_rcu(&blkg->blkcg_node);
546 * Both setting lookup hint to and clearing it from @blkg are done
547 * under queue_lock. If it's not pointing to @blkg now, it never
548 * will. Hint assignment itself can race safely.
550 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
551 rcu_assign_pointer(blkcg->blkg_hint, NULL);
554 * Put the reference taken at the time of creation so that when all
555 * queues are gone, group can be destroyed.
557 percpu_ref_kill(&blkg->refcnt);
560 static void blkg_destroy_all(struct gendisk *disk)
562 struct request_queue *q = disk->queue;
563 struct blkcg_gq *blkg, *n;
564 int count = BLKG_DESTROY_BATCH_SIZE;
567 spin_lock_irq(&q->queue_lock);
568 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
569 struct blkcg *blkcg = blkg->blkcg;
571 spin_lock(&blkcg->lock);
573 spin_unlock(&blkcg->lock);
576 * in order to avoid holding the spin lock for too long, release
577 * it when a batch of blkgs are destroyed.
580 count = BLKG_DESTROY_BATCH_SIZE;
581 spin_unlock_irq(&q->queue_lock);
588 spin_unlock_irq(&q->queue_lock);
591 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
592 struct cftype *cftype, u64 val)
594 struct blkcg *blkcg = css_to_blkcg(css);
595 struct blkcg_gq *blkg;
598 mutex_lock(&blkcg_pol_mutex);
599 spin_lock_irq(&blkcg->lock);
602 * Note that stat reset is racy - it doesn't synchronize against
603 * stat updates. This is a debug feature which shouldn't exist
604 * anyway. If you get hit by a race, retry.
606 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
607 for_each_possible_cpu(cpu) {
608 struct blkg_iostat_set *bis =
609 per_cpu_ptr(blkg->iostat_cpu, cpu);
610 memset(bis, 0, sizeof(*bis));
612 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
614 for (i = 0; i < BLKCG_MAX_POLS; i++) {
615 struct blkcg_policy *pol = blkcg_policy[i];
617 if (blkg->pd[i] && pol->pd_reset_stats_fn)
618 pol->pd_reset_stats_fn(blkg->pd[i]);
622 spin_unlock_irq(&blkcg->lock);
623 mutex_unlock(&blkcg_pol_mutex);
627 const char *blkg_dev_name(struct blkcg_gq *blkg)
631 return bdi_dev_name(blkg->q->disk->bdi);
635 * blkcg_print_blkgs - helper for printing per-blkg data
636 * @sf: seq_file to print to
637 * @blkcg: blkcg of interest
638 * @prfill: fill function to print out a blkg
639 * @pol: policy in question
640 * @data: data to be passed to @prfill
641 * @show_total: to print out sum of prfill return values or not
643 * This function invokes @prfill on each blkg of @blkcg if pd for the
644 * policy specified by @pol exists. @prfill is invoked with @sf, the
645 * policy data and @data and the matching queue lock held. If @show_total
646 * is %true, the sum of the return values from @prfill is printed with
647 * "Total" label at the end.
649 * This is to be used to construct print functions for
650 * cftype->read_seq_string method.
652 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
653 u64 (*prfill)(struct seq_file *,
654 struct blkg_policy_data *, int),
655 const struct blkcg_policy *pol, int data,
658 struct blkcg_gq *blkg;
662 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
663 spin_lock_irq(&blkg->q->queue_lock);
664 if (blkcg_policy_enabled(blkg->q, pol))
665 total += prfill(sf, blkg->pd[pol->plid], data);
666 spin_unlock_irq(&blkg->q->queue_lock);
671 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
673 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
676 * __blkg_prfill_u64 - prfill helper for a single u64 value
677 * @sf: seq_file to print to
678 * @pd: policy private data of interest
681 * Print @v to @sf for the device associated with @pd.
683 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
685 const char *dname = blkg_dev_name(pd->blkg);
690 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
693 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
696 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
697 * @inputp: input string pointer
699 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
700 * from @input and get and return the matching bdev. *@inputp is
701 * updated to point past the device node prefix. Returns an ERR_PTR()
704 * Use this function iff blkg_conf_prep() can't be used for some reason.
706 struct block_device *blkcg_conf_open_bdev(char **inputp)
708 char *input = *inputp;
709 unsigned int major, minor;
710 struct block_device *bdev;
713 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
714 return ERR_PTR(-EINVAL);
717 if (!isspace(*input))
718 return ERR_PTR(-EINVAL);
719 input = skip_spaces(input);
721 bdev = blkdev_get_no_open(MKDEV(major, minor));
723 return ERR_PTR(-ENODEV);
724 if (bdev_is_partition(bdev)) {
725 blkdev_put_no_open(bdev);
726 return ERR_PTR(-ENODEV);
734 * blkg_conf_prep - parse and prepare for per-blkg config update
735 * @blkcg: target block cgroup
736 * @pol: target policy
737 * @input: input string
738 * @ctx: blkg_conf_ctx to be filled
740 * Parse per-blkg config update from @input and initialize @ctx with the
741 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
742 * part of @input following MAJ:MIN. This function returns with RCU read
743 * lock and queue lock held and must be paired with blkg_conf_finish().
745 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
746 char *input, struct blkg_conf_ctx *ctx)
747 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
749 struct block_device *bdev;
750 struct gendisk *disk;
751 struct request_queue *q;
752 struct blkcg_gq *blkg;
755 bdev = blkcg_conf_open_bdev(&input);
757 return PTR_ERR(bdev);
758 disk = bdev->bd_disk;
762 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
763 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
765 ret = blk_queue_enter(q, 0);
770 spin_lock_irq(&q->queue_lock);
772 if (!blkcg_policy_enabled(q, pol)) {
777 blkg = blkg_lookup(blkcg, q);
782 * Create blkgs walking down from blkcg_root to @blkcg, so that all
783 * non-root blkgs have access to their parents.
786 struct blkcg *pos = blkcg;
787 struct blkcg *parent;
788 struct blkcg_gq *new_blkg;
790 parent = blkcg_parent(blkcg);
791 while (parent && !blkg_lookup(parent, q)) {
793 parent = blkcg_parent(parent);
796 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
797 spin_unlock_irq(&q->queue_lock);
800 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
801 if (unlikely(!new_blkg)) {
803 goto fail_exit_queue;
806 if (radix_tree_preload(GFP_KERNEL)) {
809 goto fail_exit_queue;
813 spin_lock_irq(&q->queue_lock);
815 if (!blkcg_policy_enabled(q, pol)) {
821 blkg = blkg_lookup(pos, q);
825 blkg = blkg_create(pos, disk, new_blkg);
832 radix_tree_preload_end();
845 radix_tree_preload_end();
847 spin_unlock_irq(&q->queue_lock);
852 blkdev_put_no_open(bdev);
854 * If queue was bypassing, we should retry. Do so after a
855 * short msleep(). It isn't strictly necessary but queue
856 * can be bypassing for some time and it's always nice to
857 * avoid busy looping.
861 ret = restart_syscall();
865 EXPORT_SYMBOL_GPL(blkg_conf_prep);
868 * blkg_conf_finish - finish up per-blkg config update
869 * @ctx: blkg_conf_ctx initialized by blkg_conf_prep()
871 * Finish up after per-blkg config update. This function must be paired
872 * with blkg_conf_prep().
874 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
875 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
877 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
879 blkdev_put_no_open(ctx->bdev);
881 EXPORT_SYMBOL_GPL(blkg_conf_finish);
883 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
887 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
888 dst->bytes[i] = src->bytes[i];
889 dst->ios[i] = src->ios[i];
893 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
897 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
898 dst->bytes[i] += src->bytes[i];
899 dst->ios[i] += src->ios[i];
903 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
907 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
908 dst->bytes[i] -= src->bytes[i];
909 dst->ios[i] -= src->ios[i];
913 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
914 struct blkg_iostat *last)
916 struct blkg_iostat delta;
919 /* propagate percpu delta to global */
920 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
921 blkg_iostat_set(&delta, cur);
922 blkg_iostat_sub(&delta, last);
923 blkg_iostat_add(&blkg->iostat.cur, &delta);
924 blkg_iostat_add(last, &delta);
925 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
928 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
930 struct blkcg *blkcg = css_to_blkcg(css);
931 struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
932 struct llist_node *lnode;
933 struct blkg_iostat_set *bisc, *next_bisc;
935 /* Root-level stats are sourced from system-wide IO stats */
936 if (!cgroup_parent(css->cgroup))
941 lnode = llist_del_all(lhead);
946 * Iterate only the iostat_cpu's queued in the lockless list.
948 llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
949 struct blkcg_gq *blkg = bisc->blkg;
950 struct blkcg_gq *parent = blkg->parent;
951 struct blkg_iostat cur;
954 WRITE_ONCE(bisc->lqueued, false);
956 /* fetch the current per-cpu values */
958 seq = u64_stats_fetch_begin(&bisc->sync);
959 blkg_iostat_set(&cur, &bisc->cur);
960 } while (u64_stats_fetch_retry(&bisc->sync, seq));
962 blkcg_iostat_update(blkg, &cur, &bisc->last);
964 /* propagate global delta to parent (unless that's root) */
965 if (parent && parent->parent)
966 blkcg_iostat_update(parent, &blkg->iostat.cur,
968 percpu_ref_put(&blkg->refcnt);
976 * We source root cgroup stats from the system-wide stats to avoid
977 * tracking the same information twice and incurring overhead when no
978 * cgroups are defined. For that reason, cgroup_rstat_flush in
979 * blkcg_print_stat does not actually fill out the iostat in the root
982 * However, we would like to re-use the printing code between the root and
983 * non-root cgroups to the extent possible. For that reason, we simulate
984 * flushing the root cgroup's stats by explicitly filling in the iostat
985 * with disk level statistics.
987 static void blkcg_fill_root_iostats(void)
989 struct class_dev_iter iter;
992 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
993 while ((dev = class_dev_iter_next(&iter))) {
994 struct block_device *bdev = dev_to_bdev(dev);
995 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
996 struct blkg_iostat tmp;
1000 memset(&tmp, 0, sizeof(tmp));
1001 for_each_possible_cpu(cpu) {
1002 struct disk_stats *cpu_dkstats;
1004 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1005 tmp.ios[BLKG_IOSTAT_READ] +=
1006 cpu_dkstats->ios[STAT_READ];
1007 tmp.ios[BLKG_IOSTAT_WRITE] +=
1008 cpu_dkstats->ios[STAT_WRITE];
1009 tmp.ios[BLKG_IOSTAT_DISCARD] +=
1010 cpu_dkstats->ios[STAT_DISCARD];
1011 // convert sectors to bytes
1012 tmp.bytes[BLKG_IOSTAT_READ] +=
1013 cpu_dkstats->sectors[STAT_READ] << 9;
1014 tmp.bytes[BLKG_IOSTAT_WRITE] +=
1015 cpu_dkstats->sectors[STAT_WRITE] << 9;
1016 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1017 cpu_dkstats->sectors[STAT_DISCARD] << 9;
1020 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1021 blkg_iostat_set(&blkg->iostat.cur, &tmp);
1022 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1026 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1028 struct blkg_iostat_set *bis = &blkg->iostat;
1029 u64 rbytes, wbytes, rios, wios, dbytes, dios;
1037 dname = blkg_dev_name(blkg);
1041 seq_printf(s, "%s ", dname);
1044 seq = u64_stats_fetch_begin(&bis->sync);
1046 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1047 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1048 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1049 rios = bis->cur.ios[BLKG_IOSTAT_READ];
1050 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1051 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1052 } while (u64_stats_fetch_retry(&bis->sync, seq));
1054 if (rbytes || wbytes || rios || wios) {
1055 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1056 rbytes, wbytes, rios, wios,
1060 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1061 seq_printf(s, " use_delay=%d delay_nsec=%llu",
1062 atomic_read(&blkg->use_delay),
1063 atomic64_read(&blkg->delay_nsec));
1066 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1067 struct blkcg_policy *pol = blkcg_policy[i];
1069 if (!blkg->pd[i] || !pol->pd_stat_fn)
1072 pol->pd_stat_fn(blkg->pd[i], s);
1078 static int blkcg_print_stat(struct seq_file *sf, void *v)
1080 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1081 struct blkcg_gq *blkg;
1083 if (!seq_css(sf)->parent)
1084 blkcg_fill_root_iostats();
1086 cgroup_rstat_flush(blkcg->css.cgroup);
1089 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1090 spin_lock_irq(&blkg->q->queue_lock);
1091 blkcg_print_one_stat(blkg, sf);
1092 spin_unlock_irq(&blkg->q->queue_lock);
1098 static struct cftype blkcg_files[] = {
1101 .seq_show = blkcg_print_stat,
1106 static struct cftype blkcg_legacy_files[] = {
1108 .name = "reset_stats",
1109 .write_u64 = blkcg_reset_stats,
1114 #ifdef CONFIG_CGROUP_WRITEBACK
1115 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1117 return &css_to_blkcg(css)->cgwb_list;
1122 * blkcg destruction is a three-stage process.
1124 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1125 * which offlines writeback. Here we tie the next stage of blkg destruction
1126 * to the completion of writeback associated with the blkcg. This lets us
1127 * avoid punting potentially large amounts of outstanding writeback to root
1128 * while maintaining any ongoing policies. The next stage is triggered when
1129 * the nr_cgwbs count goes to zero.
1131 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1132 * and handles the destruction of blkgs. Here the css reference held by
1133 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1134 * This work may occur in cgwb_release_workfn() on the cgwb_release
1135 * workqueue. Any submitted ios that fail to get the blkg ref will be
1136 * punted to the root_blkg.
1138 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1139 * This finally frees the blkcg.
1143 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1144 * @blkcg: blkcg of interest
1146 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1147 * is nested inside q lock, this function performs reverse double lock dancing.
1148 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1149 * blkcg_css_free to eventually be called.
1151 * This is the blkcg counterpart of ioc_release_fn().
1153 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1157 spin_lock_irq(&blkcg->lock);
1159 while (!hlist_empty(&blkcg->blkg_list)) {
1160 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1161 struct blkcg_gq, blkcg_node);
1162 struct request_queue *q = blkg->q;
1164 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1166 * Given that the system can accumulate a huge number
1167 * of blkgs in pathological cases, check to see if we
1168 * need to rescheduling to avoid softlockup.
1170 spin_unlock_irq(&blkcg->lock);
1172 spin_lock_irq(&blkcg->lock);
1177 spin_unlock(&q->queue_lock);
1180 spin_unlock_irq(&blkcg->lock);
1184 * blkcg_pin_online - pin online state
1185 * @blkcg_css: blkcg of interest
1187 * While pinned, a blkcg is kept online. This is primarily used to
1188 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1189 * while an associated cgwb is still active.
1191 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1193 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1197 * blkcg_unpin_online - unpin online state
1198 * @blkcg_css: blkcg of interest
1200 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1201 * that blkg doesn't go offline while an associated cgwb is still active.
1202 * When this count goes to zero, all active cgwbs have finished so the
1203 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1205 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1207 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1210 if (!refcount_dec_and_test(&blkcg->online_pin))
1212 blkcg_destroy_blkgs(blkcg);
1213 blkcg = blkcg_parent(blkcg);
1218 * blkcg_css_offline - cgroup css_offline callback
1219 * @css: css of interest
1221 * This function is called when @css is about to go away. Here the cgwbs are
1222 * offlined first and only once writeback associated with the blkcg has
1223 * finished do we start step 2 (see above).
1225 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1227 /* this prevents anyone from attaching or migrating to this blkcg */
1228 wb_blkcg_offline(css);
1230 /* put the base online pin allowing step 2 to be triggered */
1231 blkcg_unpin_online(css);
1234 static void blkcg_css_free(struct cgroup_subsys_state *css)
1236 struct blkcg *blkcg = css_to_blkcg(css);
1239 mutex_lock(&blkcg_pol_mutex);
1241 list_del(&blkcg->all_blkcgs_node);
1243 for (i = 0; i < BLKCG_MAX_POLS; i++)
1245 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1247 mutex_unlock(&blkcg_pol_mutex);
1249 free_percpu(blkcg->lhead);
1253 static struct cgroup_subsys_state *
1254 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1256 struct blkcg *blkcg;
1259 mutex_lock(&blkcg_pol_mutex);
1262 blkcg = &blkcg_root;
1264 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1269 if (init_blkcg_llists(blkcg))
1272 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1273 struct blkcg_policy *pol = blkcg_policy[i];
1274 struct blkcg_policy_data *cpd;
1277 * If the policy hasn't been attached yet, wait for it
1278 * to be attached before doing anything else. Otherwise,
1279 * check if the policy requires any specific per-cgroup
1280 * data: if it does, allocate and initialize it.
1282 if (!pol || !pol->cpd_alloc_fn)
1285 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1289 blkcg->cpd[i] = cpd;
1292 if (pol->cpd_init_fn)
1293 pol->cpd_init_fn(cpd);
1296 spin_lock_init(&blkcg->lock);
1297 refcount_set(&blkcg->online_pin, 1);
1298 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1299 INIT_HLIST_HEAD(&blkcg->blkg_list);
1300 #ifdef CONFIG_CGROUP_WRITEBACK
1301 INIT_LIST_HEAD(&blkcg->cgwb_list);
1303 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1305 mutex_unlock(&blkcg_pol_mutex);
1309 for (i--; i >= 0; i--)
1311 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1312 free_percpu(blkcg->lhead);
1314 if (blkcg != &blkcg_root)
1317 mutex_unlock(&blkcg_pol_mutex);
1318 return ERR_PTR(-ENOMEM);
1321 static int blkcg_css_online(struct cgroup_subsys_state *css)
1323 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1326 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1327 * don't go offline while cgwbs are still active on them. Pin the
1328 * parent so that offline always happens towards the root.
1331 blkcg_pin_online(&parent->css);
1335 int blkcg_init_disk(struct gendisk *disk)
1337 struct request_queue *q = disk->queue;
1338 struct blkcg_gq *new_blkg, *blkg;
1342 INIT_LIST_HEAD(&q->blkg_list);
1343 mutex_init(&q->blkcg_mutex);
1345 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1349 preloaded = !radix_tree_preload(GFP_KERNEL);
1351 /* Make sure the root blkg exists. */
1352 /* spin_lock_irq can serve as RCU read-side critical section. */
1353 spin_lock_irq(&q->queue_lock);
1354 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1357 q->root_blkg = blkg;
1358 spin_unlock_irq(&q->queue_lock);
1361 radix_tree_preload_end();
1363 ret = blk_ioprio_init(disk);
1365 goto err_destroy_all;
1367 ret = blk_throtl_init(disk);
1369 goto err_ioprio_exit;
1371 ret = blk_iolatency_init(disk);
1373 goto err_throtl_exit;
1378 blk_throtl_exit(disk);
1380 blk_ioprio_exit(disk);
1382 blkg_destroy_all(disk);
1385 spin_unlock_irq(&q->queue_lock);
1387 radix_tree_preload_end();
1388 return PTR_ERR(blkg);
1391 void blkcg_exit_disk(struct gendisk *disk)
1393 blkg_destroy_all(disk);
1394 rq_qos_exit(disk->queue);
1395 blk_throtl_exit(disk);
1398 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1402 mutex_lock(&blkcg_pol_mutex);
1404 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1405 struct blkcg_policy *pol = blkcg_policy[i];
1406 struct blkcg *blkcg;
1408 if (!pol || !pol->cpd_bind_fn)
1411 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1412 if (blkcg->cpd[pol->plid])
1413 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1415 mutex_unlock(&blkcg_pol_mutex);
1418 static void blkcg_exit(struct task_struct *tsk)
1420 if (tsk->throttle_disk)
1421 put_disk(tsk->throttle_disk);
1422 tsk->throttle_disk = NULL;
1425 struct cgroup_subsys io_cgrp_subsys = {
1426 .css_alloc = blkcg_css_alloc,
1427 .css_online = blkcg_css_online,
1428 .css_offline = blkcg_css_offline,
1429 .css_free = blkcg_css_free,
1430 .css_rstat_flush = blkcg_rstat_flush,
1432 .dfl_cftypes = blkcg_files,
1433 .legacy_cftypes = blkcg_legacy_files,
1434 .legacy_name = "blkio",
1438 * This ensures that, if available, memcg is automatically enabled
1439 * together on the default hierarchy so that the owner cgroup can
1440 * be retrieved from writeback pages.
1442 .depends_on = 1 << memory_cgrp_id,
1445 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1448 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1449 * @disk: gendisk of interest
1450 * @pol: blkcg policy to activate
1452 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1453 * bypass mode to populate its blkgs with policy_data for @pol.
1455 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1456 * from IO path. Update of each blkg is protected by both queue and blkcg
1457 * locks so that holding either lock and testing blkcg_policy_enabled() is
1458 * always enough for dereferencing policy data.
1460 * The caller is responsible for synchronizing [de]activations and policy
1461 * [un]registerations. Returns 0 on success, -errno on failure.
1463 int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1465 struct request_queue *q = disk->queue;
1466 struct blkg_policy_data *pd_prealloc = NULL;
1467 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1470 if (blkcg_policy_enabled(q, pol))
1474 blk_mq_freeze_queue(q);
1476 spin_lock_irq(&q->queue_lock);
1478 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1479 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1480 struct blkg_policy_data *pd;
1482 if (blkg->pd[pol->plid])
1485 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1486 if (blkg == pinned_blkg) {
1490 pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1491 GFP_NOWAIT | __GFP_NOWARN);
1496 * GFP_NOWAIT failed. Free the existing one and
1497 * prealloc for @blkg w/ GFP_KERNEL.
1500 blkg_put(pinned_blkg);
1504 spin_unlock_irq(&q->queue_lock);
1507 pol->pd_free_fn(pd_prealloc);
1508 pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1516 blkg->pd[pol->plid] = pd;
1518 pd->plid = pol->plid;
1522 /* all allocated, init in the same order */
1523 if (pol->pd_init_fn)
1524 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1525 pol->pd_init_fn(blkg->pd[pol->plid]);
1527 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1528 if (pol->pd_online_fn)
1529 pol->pd_online_fn(blkg->pd[pol->plid]);
1530 blkg->pd[pol->plid]->online = true;
1533 __set_bit(pol->plid, q->blkcg_pols);
1536 spin_unlock_irq(&q->queue_lock);
1539 blk_mq_unfreeze_queue(q);
1541 blkg_put(pinned_blkg);
1543 pol->pd_free_fn(pd_prealloc);
1547 /* alloc failed, nothing's initialized yet, free everything */
1548 spin_lock_irq(&q->queue_lock);
1549 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1550 struct blkcg *blkcg = blkg->blkcg;
1552 spin_lock(&blkcg->lock);
1553 if (blkg->pd[pol->plid]) {
1554 pol->pd_free_fn(blkg->pd[pol->plid]);
1555 blkg->pd[pol->plid] = NULL;
1557 spin_unlock(&blkcg->lock);
1559 spin_unlock_irq(&q->queue_lock);
1563 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1566 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1567 * @disk: gendisk of interest
1568 * @pol: blkcg policy to deactivate
1570 * Deactivate @pol on @disk. Follows the same synchronization rules as
1571 * blkcg_activate_policy().
1573 void blkcg_deactivate_policy(struct gendisk *disk,
1574 const struct blkcg_policy *pol)
1576 struct request_queue *q = disk->queue;
1577 struct blkcg_gq *blkg;
1579 if (!blkcg_policy_enabled(q, pol))
1583 blk_mq_freeze_queue(q);
1585 mutex_lock(&q->blkcg_mutex);
1586 spin_lock_irq(&q->queue_lock);
1588 __clear_bit(pol->plid, q->blkcg_pols);
1590 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1591 struct blkcg *blkcg = blkg->blkcg;
1593 spin_lock(&blkcg->lock);
1594 if (blkg->pd[pol->plid]) {
1595 if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1596 pol->pd_offline_fn(blkg->pd[pol->plid]);
1597 pol->pd_free_fn(blkg->pd[pol->plid]);
1598 blkg->pd[pol->plid] = NULL;
1600 spin_unlock(&blkcg->lock);
1603 spin_unlock_irq(&q->queue_lock);
1604 mutex_unlock(&q->blkcg_mutex);
1607 blk_mq_unfreeze_queue(q);
1609 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1611 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1613 struct blkcg *blkcg;
1615 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1616 if (blkcg->cpd[pol->plid]) {
1617 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1618 blkcg->cpd[pol->plid] = NULL;
1624 * blkcg_policy_register - register a blkcg policy
1625 * @pol: blkcg policy to register
1627 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1628 * successful registration. Returns 0 on success and -errno on failure.
1630 int blkcg_policy_register(struct blkcg_policy *pol)
1632 struct blkcg *blkcg;
1635 mutex_lock(&blkcg_pol_register_mutex);
1636 mutex_lock(&blkcg_pol_mutex);
1638 /* find an empty slot */
1640 for (i = 0; i < BLKCG_MAX_POLS; i++)
1641 if (!blkcg_policy[i])
1643 if (i >= BLKCG_MAX_POLS) {
1644 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1648 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1649 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1650 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1655 blkcg_policy[pol->plid] = pol;
1657 /* allocate and install cpd's */
1658 if (pol->cpd_alloc_fn) {
1659 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1660 struct blkcg_policy_data *cpd;
1662 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1666 blkcg->cpd[pol->plid] = cpd;
1668 cpd->plid = pol->plid;
1669 if (pol->cpd_init_fn)
1670 pol->cpd_init_fn(cpd);
1674 mutex_unlock(&blkcg_pol_mutex);
1676 /* everything is in place, add intf files for the new policy */
1677 if (pol->dfl_cftypes)
1678 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1680 if (pol->legacy_cftypes)
1681 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1682 pol->legacy_cftypes));
1683 mutex_unlock(&blkcg_pol_register_mutex);
1687 if (pol->cpd_free_fn)
1688 blkcg_free_all_cpd(pol);
1690 blkcg_policy[pol->plid] = NULL;
1692 mutex_unlock(&blkcg_pol_mutex);
1693 mutex_unlock(&blkcg_pol_register_mutex);
1696 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1699 * blkcg_policy_unregister - unregister a blkcg policy
1700 * @pol: blkcg policy to unregister
1702 * Undo blkcg_policy_register(@pol). Might sleep.
1704 void blkcg_policy_unregister(struct blkcg_policy *pol)
1706 mutex_lock(&blkcg_pol_register_mutex);
1708 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1711 /* kill the intf files first */
1712 if (pol->dfl_cftypes)
1713 cgroup_rm_cftypes(pol->dfl_cftypes);
1714 if (pol->legacy_cftypes)
1715 cgroup_rm_cftypes(pol->legacy_cftypes);
1717 /* remove cpds and unregister */
1718 mutex_lock(&blkcg_pol_mutex);
1720 if (pol->cpd_free_fn)
1721 blkcg_free_all_cpd(pol);
1723 blkcg_policy[pol->plid] = NULL;
1725 mutex_unlock(&blkcg_pol_mutex);
1727 mutex_unlock(&blkcg_pol_register_mutex);
1729 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1732 * Scale the accumulated delay based on how long it has been since we updated
1733 * the delay. We only call this when we are adding delay, in case it's been a
1734 * while since we added delay, and when we are checking to see if we need to
1735 * delay a task, to account for any delays that may have occurred.
1737 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1739 u64 old = atomic64_read(&blkg->delay_start);
1741 /* negative use_delay means no scaling, see blkcg_set_delay() */
1742 if (atomic_read(&blkg->use_delay) < 0)
1746 * We only want to scale down every second. The idea here is that we
1747 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1748 * time window. We only want to throttle tasks for recent delay that
1749 * has occurred, in 1 second time windows since that's the maximum
1750 * things can be throttled. We save the current delay window in
1751 * blkg->last_delay so we know what amount is still left to be charged
1752 * to the blkg from this point onward. blkg->last_use keeps track of
1753 * the use_delay counter. The idea is if we're unthrottling the blkg we
1754 * are ok with whatever is happening now, and we can take away more of
1755 * the accumulated delay as we've already throttled enough that
1756 * everybody is happy with their IO latencies.
1758 if (time_before64(old + NSEC_PER_SEC, now) &&
1759 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1760 u64 cur = atomic64_read(&blkg->delay_nsec);
1761 u64 sub = min_t(u64, blkg->last_delay, now - old);
1762 int cur_use = atomic_read(&blkg->use_delay);
1765 * We've been unthrottled, subtract a larger chunk of our
1766 * accumulated delay.
1768 if (cur_use < blkg->last_use)
1769 sub = max_t(u64, sub, blkg->last_delay >> 1);
1772 * This shouldn't happen, but handle it anyway. Our delay_nsec
1773 * should only ever be growing except here where we subtract out
1774 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1775 * rather not end up with negative numbers.
1777 if (unlikely(cur < sub)) {
1778 atomic64_set(&blkg->delay_nsec, 0);
1779 blkg->last_delay = 0;
1781 atomic64_sub(sub, &blkg->delay_nsec);
1782 blkg->last_delay = cur - sub;
1784 blkg->last_use = cur_use;
1789 * This is called when we want to actually walk up the hierarchy and check to
1790 * see if we need to throttle, and then actually throttle if there is some
1791 * accumulated delay. This should only be called upon return to user space so
1792 * we're not holding some lock that would induce a priority inversion.
1794 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1796 unsigned long pflags;
1798 u64 now = ktime_to_ns(ktime_get());
1803 while (blkg->parent) {
1804 int use_delay = atomic_read(&blkg->use_delay);
1809 blkcg_scale_delay(blkg, now);
1810 this_delay = atomic64_read(&blkg->delay_nsec);
1811 if (this_delay > delay_nsec) {
1812 delay_nsec = this_delay;
1813 clamp = use_delay > 0;
1816 blkg = blkg->parent;
1823 * Let's not sleep for all eternity if we've amassed a huge delay.
1824 * Swapping or metadata IO can accumulate 10's of seconds worth of
1825 * delay, and we want userspace to be able to do _something_ so cap the
1826 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1827 * tasks will be delayed for 0.25 second for every syscall. If
1828 * blkcg_set_delay() was used as indicated by negative use_delay, the
1829 * caller is responsible for regulating the range.
1832 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1835 psi_memstall_enter(&pflags);
1837 exp = ktime_add_ns(now, delay_nsec);
1838 tok = io_schedule_prepare();
1840 __set_current_state(TASK_KILLABLE);
1841 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1843 } while (!fatal_signal_pending(current));
1844 io_schedule_finish(tok);
1847 psi_memstall_leave(&pflags);
1851 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1853 * This is only called if we've been marked with set_notify_resume(). Obviously
1854 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1855 * check to see if current->throttle_disk is set and if not this doesn't do
1856 * anything. This should only ever be called by the resume code, it's not meant
1857 * to be called by people willy-nilly as it will actually do the work to
1858 * throttle the task if it is setup for throttling.
1860 void blkcg_maybe_throttle_current(void)
1862 struct gendisk *disk = current->throttle_disk;
1863 struct blkcg *blkcg;
1864 struct blkcg_gq *blkg;
1865 bool use_memdelay = current->use_memdelay;
1870 current->throttle_disk = NULL;
1871 current->use_memdelay = false;
1874 blkcg = css_to_blkcg(blkcg_css());
1877 blkg = blkg_lookup(blkcg, disk->queue);
1880 if (!blkg_tryget(blkg))
1884 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1893 * blkcg_schedule_throttle - this task needs to check for throttling
1894 * @disk: disk to throttle
1895 * @use_memdelay: do we charge this to memory delay for PSI
1897 * This is called by the IO controller when we know there's delay accumulated
1898 * for the blkg for this task. We do not pass the blkg because there are places
1899 * we call this that may not have that information, the swapping code for
1900 * instance will only have a block_device at that point. This set's the
1901 * notify_resume for the task to check and see if it requires throttling before
1902 * returning to user space.
1904 * We will only schedule once per syscall. You can call this over and over
1905 * again and it will only do the check once upon return to user space, and only
1906 * throttle once. If the task needs to be throttled again it'll need to be
1907 * re-set at the next time we see the task.
1909 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1911 if (unlikely(current->flags & PF_KTHREAD))
1914 if (current->throttle_disk != disk) {
1915 if (test_bit(GD_DEAD, &disk->state))
1917 get_device(disk_to_dev(disk));
1919 if (current->throttle_disk)
1920 put_disk(current->throttle_disk);
1921 current->throttle_disk = disk;
1925 current->use_memdelay = use_memdelay;
1926 set_notify_resume(current);
1930 * blkcg_add_delay - add delay to this blkg
1931 * @blkg: blkg of interest
1932 * @now: the current time in nanoseconds
1933 * @delta: how many nanoseconds of delay to add
1935 * Charge @delta to the blkg's current delay accumulation. This is used to
1936 * throttle tasks if an IO controller thinks we need more throttling.
1938 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1940 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1942 blkcg_scale_delay(blkg, now);
1943 atomic64_add(delta, &blkg->delay_nsec);
1947 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1951 * As the failure mode here is to walk up the blkg tree, this ensure that the
1952 * blkg->parent pointers are always valid. This returns the blkg that it ended
1953 * up taking a reference on or %NULL if no reference was taken.
1955 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1956 struct cgroup_subsys_state *css)
1958 struct blkcg_gq *blkg, *ret_blkg = NULL;
1961 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1963 if (blkg_tryget(blkg)) {
1967 blkg = blkg->parent;
1975 * bio_associate_blkg_from_css - associate a bio with a specified css
1979 * Associate @bio with the blkg found by combining the css's blkg and the
1980 * request_queue of the @bio. An association failure is handled by walking up
1981 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1982 * and q->root_blkg. This situation only happens when a cgroup is dying and
1983 * then the remaining bios will spill to the closest alive blkg.
1985 * A reference will be taken on the blkg and will be released when @bio is
1988 void bio_associate_blkg_from_css(struct bio *bio,
1989 struct cgroup_subsys_state *css)
1992 blkg_put(bio->bi_blkg);
1994 if (css && css->parent) {
1995 bio->bi_blkg = blkg_tryget_closest(bio, css);
1997 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1998 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
2001 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2004 * bio_associate_blkg - associate a bio with a blkg
2007 * Associate @bio with the blkg found from the bio's css and request_queue.
2008 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2009 * already associated, the css is reused and association redone as the
2010 * request_queue may have changed.
2012 void bio_associate_blkg(struct bio *bio)
2014 struct cgroup_subsys_state *css;
2019 css = bio_blkcg_css(bio);
2023 bio_associate_blkg_from_css(bio, css);
2027 EXPORT_SYMBOL_GPL(bio_associate_blkg);
2030 * bio_clone_blkg_association - clone blkg association from src to dst bio
2031 * @dst: destination bio
2034 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2037 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2039 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2041 static int blk_cgroup_io_type(struct bio *bio)
2043 if (op_is_discard(bio->bi_opf))
2044 return BLKG_IOSTAT_DISCARD;
2045 if (op_is_write(bio->bi_opf))
2046 return BLKG_IOSTAT_WRITE;
2047 return BLKG_IOSTAT_READ;
2050 void blk_cgroup_bio_start(struct bio *bio)
2052 struct blkcg *blkcg = bio->bi_blkg->blkcg;
2053 int rwd = blk_cgroup_io_type(bio), cpu;
2054 struct blkg_iostat_set *bis;
2055 unsigned long flags;
2057 /* Root-level stats are sourced from system-wide IO stats */
2058 if (!cgroup_parent(blkcg->css.cgroup))
2062 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2063 flags = u64_stats_update_begin_irqsave(&bis->sync);
2066 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2067 * bio and we would have already accounted for the size of the bio.
2069 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2070 bio_set_flag(bio, BIO_CGROUP_ACCT);
2071 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2073 bis->cur.ios[rwd]++;
2076 * If the iostat_cpu isn't in a lockless list, put it into the
2077 * list to indicate that a stat update is pending.
2079 if (!READ_ONCE(bis->lqueued)) {
2080 struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2082 llist_add(&bis->lnode, lhead);
2083 WRITE_ONCE(bis->lqueued, true);
2084 percpu_ref_get(&bis->blkg->refcnt);
2087 u64_stats_update_end_irqrestore(&bis->sync, flags);
2088 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
2089 cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2093 bool blk_cgroup_congested(void)
2095 struct cgroup_subsys_state *css;
2099 for (css = blkcg_css(); css; css = css->parent) {
2100 if (atomic_read(&css->cgroup->congestion_count)) {
2109 module_param(blkcg_debug_stats, bool, 0644);
2110 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");