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"
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
44 static DEFINE_MUTEX(blkcg_pol_register_mutex);
45 static DEFINE_MUTEX(blkcg_pol_mutex);
47 struct blkcg blkcg_root;
48 EXPORT_SYMBOL_GPL(blkcg_root);
50 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51 EXPORT_SYMBOL_GPL(blkcg_root_css);
53 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
55 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
57 bool blkcg_debug_stats = false;
58 static struct workqueue_struct *blkcg_punt_bio_wq;
60 #define BLKG_DESTROY_BATCH_SIZE 64
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 WARN_ON(!bio_list_empty(&blkg->async_bios));
170 /* release the blkcg and parent blkg refs this blkg has been holding */
171 css_put(&blkg->blkcg->css);
176 * A group is RCU protected, but having an rcu lock does not mean that one
177 * can access all the fields of blkg and assume these are valid. For
178 * example, don't try to follow throtl_data and request queue links.
180 * Having a reference to blkg under an rcu allows accesses to only values
181 * local to groups like group stats and group rate limits.
183 static void blkg_release(struct percpu_ref *ref)
185 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
187 call_rcu(&blkg->rcu_head, __blkg_release);
190 static void blkg_async_bio_workfn(struct work_struct *work)
192 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
194 struct bio_list bios = BIO_EMPTY_LIST;
196 struct blk_plug plug;
197 bool need_plug = false;
199 /* as long as there are pending bios, @blkg can't go away */
200 spin_lock_bh(&blkg->async_bio_lock);
201 bio_list_merge(&bios, &blkg->async_bios);
202 bio_list_init(&blkg->async_bios);
203 spin_unlock_bh(&blkg->async_bio_lock);
205 /* start plug only when bio_list contains at least 2 bios */
206 if (bios.head && bios.head->bi_next) {
208 blk_start_plug(&plug);
210 while ((bio = bio_list_pop(&bios)))
213 blk_finish_plug(&plug);
217 * bio_blkcg_css - return the blkcg CSS associated with a bio
220 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
221 * associated. Callers are expected to either handle %NULL or know association
222 * has been done prior to calling this.
224 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
226 if (!bio || !bio->bi_blkg)
228 return &bio->bi_blkg->blkcg->css;
230 EXPORT_SYMBOL_GPL(bio_blkcg_css);
233 * blkcg_parent - get the parent of a blkcg
234 * @blkcg: blkcg of interest
236 * Return the parent blkcg of @blkcg. Can be called anytime.
238 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
240 return css_to_blkcg(blkcg->css.parent);
244 * blkg_alloc - allocate a blkg
245 * @blkcg: block cgroup the new blkg is associated with
246 * @disk: gendisk the new blkg is associated with
247 * @gfp_mask: allocation mask to use
249 * Allocate a new blkg assocating @blkcg and @q.
251 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
254 struct blkcg_gq *blkg;
257 /* alloc and init base part */
258 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
261 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
263 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
264 if (!blkg->iostat_cpu)
265 goto out_exit_refcnt;
266 if (!blk_get_queue(disk->queue))
267 goto out_free_iostat;
269 blkg->q = disk->queue;
270 INIT_LIST_HEAD(&blkg->q_node);
271 spin_lock_init(&blkg->async_bio_lock);
272 bio_list_init(&blkg->async_bios);
273 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
276 u64_stats_init(&blkg->iostat.sync);
277 for_each_possible_cpu(cpu) {
278 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
279 per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
282 for (i = 0; i < BLKCG_MAX_POLS; i++) {
283 struct blkcg_policy *pol = blkcg_policy[i];
284 struct blkg_policy_data *pd;
286 if (!blkcg_policy_enabled(disk->queue, pol))
289 /* alloc per-policy data and attach it to blkg */
290 pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
304 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
305 blk_put_queue(disk->queue);
307 free_percpu(blkg->iostat_cpu);
309 percpu_ref_exit(&blkg->refcnt);
316 * If @new_blkg is %NULL, this function tries to allocate a new one as
317 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
319 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
320 struct blkcg_gq *new_blkg)
322 struct blkcg_gq *blkg;
325 lockdep_assert_held(&disk->queue->queue_lock);
327 /* request_queue is dying, do not create/recreate a blkg */
328 if (blk_queue_dying(disk->queue)) {
333 /* blkg holds a reference to blkcg */
334 if (!css_tryget_online(&blkcg->css)) {
341 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
342 if (unlikely(!new_blkg)) {
350 if (blkcg_parent(blkcg)) {
351 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
352 if (WARN_ON_ONCE(!blkg->parent)) {
356 blkg_get(blkg->parent);
359 /* invoke per-policy init */
360 for (i = 0; i < BLKCG_MAX_POLS; i++) {
361 struct blkcg_policy *pol = blkcg_policy[i];
363 if (blkg->pd[i] && pol->pd_init_fn)
364 pol->pd_init_fn(blkg->pd[i]);
368 spin_lock(&blkcg->lock);
369 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
371 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
372 list_add(&blkg->q_node, &disk->queue->blkg_list);
374 for (i = 0; i < BLKCG_MAX_POLS; i++) {
375 struct blkcg_policy *pol = blkcg_policy[i];
378 if (pol->pd_online_fn)
379 pol->pd_online_fn(blkg->pd[i]);
380 blkg->pd[i]->online = true;
385 spin_unlock(&blkcg->lock);
390 /* @blkg failed fully initialized, use the usual release path */
395 css_put(&blkcg->css);
403 * blkg_lookup_create - lookup blkg, try to create one if not there
404 * @blkcg: blkcg of interest
405 * @disk: gendisk of interest
407 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
408 * create one. blkg creation is performed recursively from blkcg_root such
409 * that all non-root blkg's have access to the parent blkg. This function
410 * should be called under RCU read lock and takes @disk->queue->queue_lock.
412 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
415 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
416 struct gendisk *disk)
418 struct request_queue *q = disk->queue;
419 struct blkcg_gq *blkg;
422 WARN_ON_ONCE(!rcu_read_lock_held());
424 blkg = blkg_lookup(blkcg, q);
428 spin_lock_irqsave(&q->queue_lock, flags);
429 blkg = blkg_lookup(blkcg, q);
431 if (blkcg != &blkcg_root &&
432 blkg != rcu_dereference(blkcg->blkg_hint))
433 rcu_assign_pointer(blkcg->blkg_hint, blkg);
438 * Create blkgs walking down from blkcg_root to @blkcg, so that all
439 * non-root blkgs have access to their parents. Returns the closest
440 * blkg to the intended blkg should blkg_create() fail.
443 struct blkcg *pos = blkcg;
444 struct blkcg *parent = blkcg_parent(blkcg);
445 struct blkcg_gq *ret_blkg = q->root_blkg;
448 blkg = blkg_lookup(parent, q);
450 /* remember closest blkg */
455 parent = blkcg_parent(parent);
458 blkg = blkg_create(pos, disk, NULL);
468 spin_unlock_irqrestore(&q->queue_lock, flags);
472 static void blkg_destroy(struct blkcg_gq *blkg)
474 struct blkcg *blkcg = blkg->blkcg;
477 lockdep_assert_held(&blkg->q->queue_lock);
478 lockdep_assert_held(&blkcg->lock);
481 * blkg stays on the queue list until blkg_free_workfn(), see details in
482 * blkg_free_workfn(), hence this function can be called from
483 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
484 * blkg_free_workfn().
486 if (hlist_unhashed(&blkg->blkcg_node))
489 for (i = 0; i < BLKCG_MAX_POLS; i++) {
490 struct blkcg_policy *pol = blkcg_policy[i];
492 if (blkg->pd[i] && blkg->pd[i]->online) {
493 blkg->pd[i]->online = false;
494 if (pol->pd_offline_fn)
495 pol->pd_offline_fn(blkg->pd[i]);
499 blkg->online = false;
501 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
502 hlist_del_init_rcu(&blkg->blkcg_node);
505 * Both setting lookup hint to and clearing it from @blkg are done
506 * under queue_lock. If it's not pointing to @blkg now, it never
507 * will. Hint assignment itself can race safely.
509 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
510 rcu_assign_pointer(blkcg->blkg_hint, NULL);
513 * Put the reference taken at the time of creation so that when all
514 * queues are gone, group can be destroyed.
516 percpu_ref_kill(&blkg->refcnt);
519 static void blkg_destroy_all(struct gendisk *disk)
521 struct request_queue *q = disk->queue;
522 struct blkcg_gq *blkg, *n;
523 int count = BLKG_DESTROY_BATCH_SIZE;
526 spin_lock_irq(&q->queue_lock);
527 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
528 struct blkcg *blkcg = blkg->blkcg;
530 spin_lock(&blkcg->lock);
532 spin_unlock(&blkcg->lock);
535 * in order to avoid holding the spin lock for too long, release
536 * it when a batch of blkgs are destroyed.
539 count = BLKG_DESTROY_BATCH_SIZE;
540 spin_unlock_irq(&q->queue_lock);
547 spin_unlock_irq(&q->queue_lock);
550 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
551 struct cftype *cftype, u64 val)
553 struct blkcg *blkcg = css_to_blkcg(css);
554 struct blkcg_gq *blkg;
557 mutex_lock(&blkcg_pol_mutex);
558 spin_lock_irq(&blkcg->lock);
561 * Note that stat reset is racy - it doesn't synchronize against
562 * stat updates. This is a debug feature which shouldn't exist
563 * anyway. If you get hit by a race, retry.
565 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
566 for_each_possible_cpu(cpu) {
567 struct blkg_iostat_set *bis =
568 per_cpu_ptr(blkg->iostat_cpu, cpu);
569 memset(bis, 0, sizeof(*bis));
571 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
573 for (i = 0; i < BLKCG_MAX_POLS; i++) {
574 struct blkcg_policy *pol = blkcg_policy[i];
576 if (blkg->pd[i] && pol->pd_reset_stats_fn)
577 pol->pd_reset_stats_fn(blkg->pd[i]);
581 spin_unlock_irq(&blkcg->lock);
582 mutex_unlock(&blkcg_pol_mutex);
586 const char *blkg_dev_name(struct blkcg_gq *blkg)
590 return bdi_dev_name(blkg->q->disk->bdi);
594 * blkcg_print_blkgs - helper for printing per-blkg data
595 * @sf: seq_file to print to
596 * @blkcg: blkcg of interest
597 * @prfill: fill function to print out a blkg
598 * @pol: policy in question
599 * @data: data to be passed to @prfill
600 * @show_total: to print out sum of prfill return values or not
602 * This function invokes @prfill on each blkg of @blkcg if pd for the
603 * policy specified by @pol exists. @prfill is invoked with @sf, the
604 * policy data and @data and the matching queue lock held. If @show_total
605 * is %true, the sum of the return values from @prfill is printed with
606 * "Total" label at the end.
608 * This is to be used to construct print functions for
609 * cftype->read_seq_string method.
611 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
612 u64 (*prfill)(struct seq_file *,
613 struct blkg_policy_data *, int),
614 const struct blkcg_policy *pol, int data,
617 struct blkcg_gq *blkg;
621 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
622 spin_lock_irq(&blkg->q->queue_lock);
623 if (blkcg_policy_enabled(blkg->q, pol))
624 total += prfill(sf, blkg->pd[pol->plid], data);
625 spin_unlock_irq(&blkg->q->queue_lock);
630 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
632 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
635 * __blkg_prfill_u64 - prfill helper for a single u64 value
636 * @sf: seq_file to print to
637 * @pd: policy private data of interest
640 * Print @v to @sf for the device associated with @pd.
642 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
644 const char *dname = blkg_dev_name(pd->blkg);
649 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
652 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
655 * blkg_conf_init - initialize a blkg_conf_ctx
656 * @ctx: blkg_conf_ctx to initialize
657 * @input: input string
659 * Initialize @ctx which can be used to parse blkg config input string @input.
660 * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
661 * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
663 void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
665 *ctx = (struct blkg_conf_ctx){ .input = input };
667 EXPORT_SYMBOL_GPL(blkg_conf_init);
670 * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
671 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
673 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
674 * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
675 * set to point past the device node prefix.
677 * This function may be called multiple times on @ctx and the extra calls become
678 * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
679 * explicitly if bdev access is needed without resolving the blkcg / policy part
680 * of @ctx->input. Returns -errno on error.
682 int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
684 char *input = ctx->input;
685 unsigned int major, minor;
686 struct block_device *bdev;
692 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
696 if (!isspace(*input))
698 input = skip_spaces(input);
700 bdev = blkdev_get_no_open(MKDEV(major, minor));
703 if (bdev_is_partition(bdev)) {
704 blkdev_put_no_open(bdev);
714 * blkg_conf_prep - parse and prepare for per-blkg config update
715 * @blkcg: target block cgroup
716 * @pol: target policy
717 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
719 * Parse per-blkg config update from @ctx->input and initialize @ctx
720 * accordingly. On success, @ctx->body points to the part of @ctx->input
721 * following MAJ:MIN, @ctx->bdev points to the target block device and
722 * @ctx->blkg to the blkg being configured.
724 * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
725 * function returns with queue lock held and must be followed by
728 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
729 struct blkg_conf_ctx *ctx)
730 __acquires(&bdev->bd_queue->queue_lock)
732 struct gendisk *disk;
733 struct request_queue *q;
734 struct blkcg_gq *blkg;
737 ret = blkg_conf_open_bdev(ctx);
741 disk = ctx->bdev->bd_disk;
745 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
746 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
748 ret = blk_queue_enter(q, 0);
752 spin_lock_irq(&q->queue_lock);
754 if (!blkcg_policy_enabled(q, pol)) {
759 blkg = blkg_lookup(blkcg, q);
764 * Create blkgs walking down from blkcg_root to @blkcg, so that all
765 * non-root blkgs have access to their parents.
768 struct blkcg *pos = blkcg;
769 struct blkcg *parent;
770 struct blkcg_gq *new_blkg;
772 parent = blkcg_parent(blkcg);
773 while (parent && !blkg_lookup(parent, q)) {
775 parent = blkcg_parent(parent);
778 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
779 spin_unlock_irq(&q->queue_lock);
781 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
782 if (unlikely(!new_blkg)) {
784 goto fail_exit_queue;
787 if (radix_tree_preload(GFP_KERNEL)) {
790 goto fail_exit_queue;
793 spin_lock_irq(&q->queue_lock);
795 if (!blkcg_policy_enabled(q, pol)) {
801 blkg = blkg_lookup(pos, q);
805 blkg = blkg_create(pos, disk, new_blkg);
812 radix_tree_preload_end();
823 radix_tree_preload_end();
825 spin_unlock_irq(&q->queue_lock);
830 * If queue was bypassing, we should retry. Do so after a
831 * short msleep(). It isn't strictly necessary but queue
832 * can be bypassing for some time and it's always nice to
833 * avoid busy looping.
837 ret = restart_syscall();
841 EXPORT_SYMBOL_GPL(blkg_conf_prep);
844 * blkg_conf_exit - clean up per-blkg config update
845 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
847 * Clean up after per-blkg config update. This function must be called on all
848 * blkg_conf_ctx's initialized with blkg_conf_init().
850 void blkg_conf_exit(struct blkg_conf_ctx *ctx)
851 __releases(&ctx->bdev->bd_queue->queue_lock)
854 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
859 blkdev_put_no_open(ctx->bdev);
864 EXPORT_SYMBOL_GPL(blkg_conf_exit);
866 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
870 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
871 dst->bytes[i] = src->bytes[i];
872 dst->ios[i] = src->ios[i];
876 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
880 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
881 dst->bytes[i] += src->bytes[i];
882 dst->ios[i] += src->ios[i];
886 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
890 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
891 dst->bytes[i] -= src->bytes[i];
892 dst->ios[i] -= src->ios[i];
896 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
897 struct blkg_iostat *last)
899 struct blkg_iostat delta;
902 /* propagate percpu delta to global */
903 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
904 blkg_iostat_set(&delta, cur);
905 blkg_iostat_sub(&delta, last);
906 blkg_iostat_add(&blkg->iostat.cur, &delta);
907 blkg_iostat_add(last, &delta);
908 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
911 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
913 struct blkcg *blkcg = css_to_blkcg(css);
914 struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
915 struct llist_node *lnode;
916 struct blkg_iostat_set *bisc, *next_bisc;
918 /* Root-level stats are sourced from system-wide IO stats */
919 if (!cgroup_parent(css->cgroup))
924 lnode = llist_del_all(lhead);
929 * Iterate only the iostat_cpu's queued in the lockless list.
931 llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
932 struct blkcg_gq *blkg = bisc->blkg;
933 struct blkcg_gq *parent = blkg->parent;
934 struct blkg_iostat cur;
937 WRITE_ONCE(bisc->lqueued, false);
939 /* fetch the current per-cpu values */
941 seq = u64_stats_fetch_begin(&bisc->sync);
942 blkg_iostat_set(&cur, &bisc->cur);
943 } while (u64_stats_fetch_retry(&bisc->sync, seq));
945 blkcg_iostat_update(blkg, &cur, &bisc->last);
947 /* propagate global delta to parent (unless that's root) */
948 if (parent && parent->parent)
949 blkcg_iostat_update(parent, &blkg->iostat.cur,
951 percpu_ref_put(&blkg->refcnt);
959 * We source root cgroup stats from the system-wide stats to avoid
960 * tracking the same information twice and incurring overhead when no
961 * cgroups are defined. For that reason, cgroup_rstat_flush in
962 * blkcg_print_stat does not actually fill out the iostat in the root
965 * However, we would like to re-use the printing code between the root and
966 * non-root cgroups to the extent possible. For that reason, we simulate
967 * flushing the root cgroup's stats by explicitly filling in the iostat
968 * with disk level statistics.
970 static void blkcg_fill_root_iostats(void)
972 struct class_dev_iter iter;
975 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
976 while ((dev = class_dev_iter_next(&iter))) {
977 struct block_device *bdev = dev_to_bdev(dev);
978 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
979 struct blkg_iostat tmp;
983 memset(&tmp, 0, sizeof(tmp));
984 for_each_possible_cpu(cpu) {
985 struct disk_stats *cpu_dkstats;
987 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
988 tmp.ios[BLKG_IOSTAT_READ] +=
989 cpu_dkstats->ios[STAT_READ];
990 tmp.ios[BLKG_IOSTAT_WRITE] +=
991 cpu_dkstats->ios[STAT_WRITE];
992 tmp.ios[BLKG_IOSTAT_DISCARD] +=
993 cpu_dkstats->ios[STAT_DISCARD];
994 // convert sectors to bytes
995 tmp.bytes[BLKG_IOSTAT_READ] +=
996 cpu_dkstats->sectors[STAT_READ] << 9;
997 tmp.bytes[BLKG_IOSTAT_WRITE] +=
998 cpu_dkstats->sectors[STAT_WRITE] << 9;
999 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1000 cpu_dkstats->sectors[STAT_DISCARD] << 9;
1003 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1004 blkg_iostat_set(&blkg->iostat.cur, &tmp);
1005 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1009 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1011 struct blkg_iostat_set *bis = &blkg->iostat;
1012 u64 rbytes, wbytes, rios, wios, dbytes, dios;
1020 dname = blkg_dev_name(blkg);
1024 seq_printf(s, "%s ", dname);
1027 seq = u64_stats_fetch_begin(&bis->sync);
1029 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1030 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1031 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1032 rios = bis->cur.ios[BLKG_IOSTAT_READ];
1033 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1034 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1035 } while (u64_stats_fetch_retry(&bis->sync, seq));
1037 if (rbytes || wbytes || rios || wios) {
1038 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1039 rbytes, wbytes, rios, wios,
1043 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1044 seq_printf(s, " use_delay=%d delay_nsec=%llu",
1045 atomic_read(&blkg->use_delay),
1046 atomic64_read(&blkg->delay_nsec));
1049 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1050 struct blkcg_policy *pol = blkcg_policy[i];
1052 if (!blkg->pd[i] || !pol->pd_stat_fn)
1055 pol->pd_stat_fn(blkg->pd[i], s);
1061 static int blkcg_print_stat(struct seq_file *sf, void *v)
1063 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1064 struct blkcg_gq *blkg;
1066 if (!seq_css(sf)->parent)
1067 blkcg_fill_root_iostats();
1069 cgroup_rstat_flush(blkcg->css.cgroup);
1072 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1073 spin_lock_irq(&blkg->q->queue_lock);
1074 blkcg_print_one_stat(blkg, sf);
1075 spin_unlock_irq(&blkg->q->queue_lock);
1081 static struct cftype blkcg_files[] = {
1084 .seq_show = blkcg_print_stat,
1089 static struct cftype blkcg_legacy_files[] = {
1091 .name = "reset_stats",
1092 .write_u64 = blkcg_reset_stats,
1097 #ifdef CONFIG_CGROUP_WRITEBACK
1098 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1100 return &css_to_blkcg(css)->cgwb_list;
1105 * blkcg destruction is a three-stage process.
1107 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1108 * which offlines writeback. Here we tie the next stage of blkg destruction
1109 * to the completion of writeback associated with the blkcg. This lets us
1110 * avoid punting potentially large amounts of outstanding writeback to root
1111 * while maintaining any ongoing policies. The next stage is triggered when
1112 * the nr_cgwbs count goes to zero.
1114 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1115 * and handles the destruction of blkgs. Here the css reference held by
1116 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1117 * This work may occur in cgwb_release_workfn() on the cgwb_release
1118 * workqueue. Any submitted ios that fail to get the blkg ref will be
1119 * punted to the root_blkg.
1121 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1122 * This finally frees the blkcg.
1126 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1127 * @blkcg: blkcg of interest
1129 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1130 * is nested inside q lock, this function performs reverse double lock dancing.
1131 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1132 * blkcg_css_free to eventually be called.
1134 * This is the blkcg counterpart of ioc_release_fn().
1136 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1140 spin_lock_irq(&blkcg->lock);
1142 while (!hlist_empty(&blkcg->blkg_list)) {
1143 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1144 struct blkcg_gq, blkcg_node);
1145 struct request_queue *q = blkg->q;
1147 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1149 * Given that the system can accumulate a huge number
1150 * of blkgs in pathological cases, check to see if we
1151 * need to rescheduling to avoid softlockup.
1153 spin_unlock_irq(&blkcg->lock);
1155 spin_lock_irq(&blkcg->lock);
1160 spin_unlock(&q->queue_lock);
1163 spin_unlock_irq(&blkcg->lock);
1167 * blkcg_pin_online - pin online state
1168 * @blkcg_css: blkcg of interest
1170 * While pinned, a blkcg is kept online. This is primarily used to
1171 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1172 * while an associated cgwb is still active.
1174 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1176 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1180 * blkcg_unpin_online - unpin online state
1181 * @blkcg_css: blkcg of interest
1183 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1184 * that blkg doesn't go offline while an associated cgwb is still active.
1185 * When this count goes to zero, all active cgwbs have finished so the
1186 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1188 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1190 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1193 if (!refcount_dec_and_test(&blkcg->online_pin))
1195 blkcg_destroy_blkgs(blkcg);
1196 blkcg = blkcg_parent(blkcg);
1201 * blkcg_css_offline - cgroup css_offline callback
1202 * @css: css of interest
1204 * This function is called when @css is about to go away. Here the cgwbs are
1205 * offlined first and only once writeback associated with the blkcg has
1206 * finished do we start step 2 (see above).
1208 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1210 /* this prevents anyone from attaching or migrating to this blkcg */
1211 wb_blkcg_offline(css);
1213 /* put the base online pin allowing step 2 to be triggered */
1214 blkcg_unpin_online(css);
1217 static void blkcg_css_free(struct cgroup_subsys_state *css)
1219 struct blkcg *blkcg = css_to_blkcg(css);
1222 mutex_lock(&blkcg_pol_mutex);
1224 list_del(&blkcg->all_blkcgs_node);
1226 for (i = 0; i < BLKCG_MAX_POLS; i++)
1228 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1230 mutex_unlock(&blkcg_pol_mutex);
1232 free_percpu(blkcg->lhead);
1236 static struct cgroup_subsys_state *
1237 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1239 struct blkcg *blkcg;
1242 mutex_lock(&blkcg_pol_mutex);
1245 blkcg = &blkcg_root;
1247 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1252 if (init_blkcg_llists(blkcg))
1255 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1256 struct blkcg_policy *pol = blkcg_policy[i];
1257 struct blkcg_policy_data *cpd;
1260 * If the policy hasn't been attached yet, wait for it
1261 * to be attached before doing anything else. Otherwise,
1262 * check if the policy requires any specific per-cgroup
1263 * data: if it does, allocate and initialize it.
1265 if (!pol || !pol->cpd_alloc_fn)
1268 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1272 blkcg->cpd[i] = cpd;
1277 spin_lock_init(&blkcg->lock);
1278 refcount_set(&blkcg->online_pin, 1);
1279 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1280 INIT_HLIST_HEAD(&blkcg->blkg_list);
1281 #ifdef CONFIG_CGROUP_WRITEBACK
1282 INIT_LIST_HEAD(&blkcg->cgwb_list);
1284 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1286 mutex_unlock(&blkcg_pol_mutex);
1290 for (i--; i >= 0; i--)
1292 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1293 free_percpu(blkcg->lhead);
1295 if (blkcg != &blkcg_root)
1298 mutex_unlock(&blkcg_pol_mutex);
1299 return ERR_PTR(-ENOMEM);
1302 static int blkcg_css_online(struct cgroup_subsys_state *css)
1304 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1307 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1308 * don't go offline while cgwbs are still active on them. Pin the
1309 * parent so that offline always happens towards the root.
1312 blkcg_pin_online(&parent->css);
1316 int blkcg_init_disk(struct gendisk *disk)
1318 struct request_queue *q = disk->queue;
1319 struct blkcg_gq *new_blkg, *blkg;
1323 INIT_LIST_HEAD(&q->blkg_list);
1324 mutex_init(&q->blkcg_mutex);
1326 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1330 preloaded = !radix_tree_preload(GFP_KERNEL);
1332 /* Make sure the root blkg exists. */
1333 /* spin_lock_irq can serve as RCU read-side critical section. */
1334 spin_lock_irq(&q->queue_lock);
1335 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1338 q->root_blkg = blkg;
1339 spin_unlock_irq(&q->queue_lock);
1342 radix_tree_preload_end();
1344 ret = blk_ioprio_init(disk);
1346 goto err_destroy_all;
1348 ret = blk_throtl_init(disk);
1350 goto err_ioprio_exit;
1355 blk_ioprio_exit(disk);
1357 blkg_destroy_all(disk);
1360 spin_unlock_irq(&q->queue_lock);
1362 radix_tree_preload_end();
1363 return PTR_ERR(blkg);
1366 void blkcg_exit_disk(struct gendisk *disk)
1368 blkg_destroy_all(disk);
1369 blk_throtl_exit(disk);
1372 static void blkcg_exit(struct task_struct *tsk)
1374 if (tsk->throttle_disk)
1375 put_disk(tsk->throttle_disk);
1376 tsk->throttle_disk = NULL;
1379 struct cgroup_subsys io_cgrp_subsys = {
1380 .css_alloc = blkcg_css_alloc,
1381 .css_online = blkcg_css_online,
1382 .css_offline = blkcg_css_offline,
1383 .css_free = blkcg_css_free,
1384 .css_rstat_flush = blkcg_rstat_flush,
1385 .dfl_cftypes = blkcg_files,
1386 .legacy_cftypes = blkcg_legacy_files,
1387 .legacy_name = "blkio",
1391 * This ensures that, if available, memcg is automatically enabled
1392 * together on the default hierarchy so that the owner cgroup can
1393 * be retrieved from writeback pages.
1395 .depends_on = 1 << memory_cgrp_id,
1398 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1401 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1402 * @disk: gendisk of interest
1403 * @pol: blkcg policy to activate
1405 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1406 * bypass mode to populate its blkgs with policy_data for @pol.
1408 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1409 * from IO path. Update of each blkg is protected by both queue and blkcg
1410 * locks so that holding either lock and testing blkcg_policy_enabled() is
1411 * always enough for dereferencing policy data.
1413 * The caller is responsible for synchronizing [de]activations and policy
1414 * [un]registerations. Returns 0 on success, -errno on failure.
1416 int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1418 struct request_queue *q = disk->queue;
1419 struct blkg_policy_data *pd_prealloc = NULL;
1420 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1423 if (blkcg_policy_enabled(q, pol))
1427 blk_mq_freeze_queue(q);
1429 spin_lock_irq(&q->queue_lock);
1431 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1432 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1433 struct blkg_policy_data *pd;
1435 if (blkg->pd[pol->plid])
1438 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1439 if (blkg == pinned_blkg) {
1443 pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1444 GFP_NOWAIT | __GFP_NOWARN);
1449 * GFP_NOWAIT failed. Free the existing one and
1450 * prealloc for @blkg w/ GFP_KERNEL.
1453 blkg_put(pinned_blkg);
1457 spin_unlock_irq(&q->queue_lock);
1460 pol->pd_free_fn(pd_prealloc);
1461 pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1469 blkg->pd[pol->plid] = pd;
1471 pd->plid = pol->plid;
1475 /* all allocated, init in the same order */
1476 if (pol->pd_init_fn)
1477 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1478 pol->pd_init_fn(blkg->pd[pol->plid]);
1480 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1481 if (pol->pd_online_fn)
1482 pol->pd_online_fn(blkg->pd[pol->plid]);
1483 blkg->pd[pol->plid]->online = true;
1486 __set_bit(pol->plid, q->blkcg_pols);
1489 spin_unlock_irq(&q->queue_lock);
1492 blk_mq_unfreeze_queue(q);
1494 blkg_put(pinned_blkg);
1496 pol->pd_free_fn(pd_prealloc);
1500 /* alloc failed, nothing's initialized yet, free everything */
1501 spin_lock_irq(&q->queue_lock);
1502 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1503 struct blkcg *blkcg = blkg->blkcg;
1505 spin_lock(&blkcg->lock);
1506 if (blkg->pd[pol->plid]) {
1507 pol->pd_free_fn(blkg->pd[pol->plid]);
1508 blkg->pd[pol->plid] = NULL;
1510 spin_unlock(&blkcg->lock);
1512 spin_unlock_irq(&q->queue_lock);
1516 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1519 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1520 * @disk: gendisk of interest
1521 * @pol: blkcg policy to deactivate
1523 * Deactivate @pol on @disk. Follows the same synchronization rules as
1524 * blkcg_activate_policy().
1526 void blkcg_deactivate_policy(struct gendisk *disk,
1527 const struct blkcg_policy *pol)
1529 struct request_queue *q = disk->queue;
1530 struct blkcg_gq *blkg;
1532 if (!blkcg_policy_enabled(q, pol))
1536 blk_mq_freeze_queue(q);
1538 mutex_lock(&q->blkcg_mutex);
1539 spin_lock_irq(&q->queue_lock);
1541 __clear_bit(pol->plid, q->blkcg_pols);
1543 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1544 struct blkcg *blkcg = blkg->blkcg;
1546 spin_lock(&blkcg->lock);
1547 if (blkg->pd[pol->plid]) {
1548 if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1549 pol->pd_offline_fn(blkg->pd[pol->plid]);
1550 pol->pd_free_fn(blkg->pd[pol->plid]);
1551 blkg->pd[pol->plid] = NULL;
1553 spin_unlock(&blkcg->lock);
1556 spin_unlock_irq(&q->queue_lock);
1557 mutex_unlock(&q->blkcg_mutex);
1560 blk_mq_unfreeze_queue(q);
1562 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1564 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1566 struct blkcg *blkcg;
1568 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1569 if (blkcg->cpd[pol->plid]) {
1570 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1571 blkcg->cpd[pol->plid] = NULL;
1577 * blkcg_policy_register - register a blkcg policy
1578 * @pol: blkcg policy to register
1580 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1581 * successful registration. Returns 0 on success and -errno on failure.
1583 int blkcg_policy_register(struct blkcg_policy *pol)
1585 struct blkcg *blkcg;
1588 mutex_lock(&blkcg_pol_register_mutex);
1589 mutex_lock(&blkcg_pol_mutex);
1591 /* find an empty slot */
1593 for (i = 0; i < BLKCG_MAX_POLS; i++)
1594 if (!blkcg_policy[i])
1596 if (i >= BLKCG_MAX_POLS) {
1597 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1601 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1602 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1603 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1608 blkcg_policy[pol->plid] = pol;
1610 /* allocate and install cpd's */
1611 if (pol->cpd_alloc_fn) {
1612 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1613 struct blkcg_policy_data *cpd;
1615 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1619 blkcg->cpd[pol->plid] = cpd;
1621 cpd->plid = pol->plid;
1625 mutex_unlock(&blkcg_pol_mutex);
1627 /* everything is in place, add intf files for the new policy */
1628 if (pol->dfl_cftypes)
1629 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1631 if (pol->legacy_cftypes)
1632 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1633 pol->legacy_cftypes));
1634 mutex_unlock(&blkcg_pol_register_mutex);
1638 if (pol->cpd_free_fn)
1639 blkcg_free_all_cpd(pol);
1641 blkcg_policy[pol->plid] = NULL;
1643 mutex_unlock(&blkcg_pol_mutex);
1644 mutex_unlock(&blkcg_pol_register_mutex);
1647 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1650 * blkcg_policy_unregister - unregister a blkcg policy
1651 * @pol: blkcg policy to unregister
1653 * Undo blkcg_policy_register(@pol). Might sleep.
1655 void blkcg_policy_unregister(struct blkcg_policy *pol)
1657 mutex_lock(&blkcg_pol_register_mutex);
1659 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1662 /* kill the intf files first */
1663 if (pol->dfl_cftypes)
1664 cgroup_rm_cftypes(pol->dfl_cftypes);
1665 if (pol->legacy_cftypes)
1666 cgroup_rm_cftypes(pol->legacy_cftypes);
1668 /* remove cpds and unregister */
1669 mutex_lock(&blkcg_pol_mutex);
1671 if (pol->cpd_free_fn)
1672 blkcg_free_all_cpd(pol);
1674 blkcg_policy[pol->plid] = NULL;
1676 mutex_unlock(&blkcg_pol_mutex);
1678 mutex_unlock(&blkcg_pol_register_mutex);
1680 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1682 bool __blkcg_punt_bio_submit(struct bio *bio)
1684 struct blkcg_gq *blkg = bio->bi_blkg;
1686 /* consume the flag first */
1687 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1689 /* never bounce for the root cgroup */
1693 spin_lock_bh(&blkg->async_bio_lock);
1694 bio_list_add(&blkg->async_bios, bio);
1695 spin_unlock_bh(&blkg->async_bio_lock);
1697 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1702 * Scale the accumulated delay based on how long it has been since we updated
1703 * the delay. We only call this when we are adding delay, in case it's been a
1704 * while since we added delay, and when we are checking to see if we need to
1705 * delay a task, to account for any delays that may have occurred.
1707 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1709 u64 old = atomic64_read(&blkg->delay_start);
1711 /* negative use_delay means no scaling, see blkcg_set_delay() */
1712 if (atomic_read(&blkg->use_delay) < 0)
1716 * We only want to scale down every second. The idea here is that we
1717 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1718 * time window. We only want to throttle tasks for recent delay that
1719 * has occurred, in 1 second time windows since that's the maximum
1720 * things can be throttled. We save the current delay window in
1721 * blkg->last_delay so we know what amount is still left to be charged
1722 * to the blkg from this point onward. blkg->last_use keeps track of
1723 * the use_delay counter. The idea is if we're unthrottling the blkg we
1724 * are ok with whatever is happening now, and we can take away more of
1725 * the accumulated delay as we've already throttled enough that
1726 * everybody is happy with their IO latencies.
1728 if (time_before64(old + NSEC_PER_SEC, now) &&
1729 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1730 u64 cur = atomic64_read(&blkg->delay_nsec);
1731 u64 sub = min_t(u64, blkg->last_delay, now - old);
1732 int cur_use = atomic_read(&blkg->use_delay);
1735 * We've been unthrottled, subtract a larger chunk of our
1736 * accumulated delay.
1738 if (cur_use < blkg->last_use)
1739 sub = max_t(u64, sub, blkg->last_delay >> 1);
1742 * This shouldn't happen, but handle it anyway. Our delay_nsec
1743 * should only ever be growing except here where we subtract out
1744 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1745 * rather not end up with negative numbers.
1747 if (unlikely(cur < sub)) {
1748 atomic64_set(&blkg->delay_nsec, 0);
1749 blkg->last_delay = 0;
1751 atomic64_sub(sub, &blkg->delay_nsec);
1752 blkg->last_delay = cur - sub;
1754 blkg->last_use = cur_use;
1759 * This is called when we want to actually walk up the hierarchy and check to
1760 * see if we need to throttle, and then actually throttle if there is some
1761 * accumulated delay. This should only be called upon return to user space so
1762 * we're not holding some lock that would induce a priority inversion.
1764 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1766 unsigned long pflags;
1768 u64 now = ktime_to_ns(ktime_get());
1773 while (blkg->parent) {
1774 int use_delay = atomic_read(&blkg->use_delay);
1779 blkcg_scale_delay(blkg, now);
1780 this_delay = atomic64_read(&blkg->delay_nsec);
1781 if (this_delay > delay_nsec) {
1782 delay_nsec = this_delay;
1783 clamp = use_delay > 0;
1786 blkg = blkg->parent;
1793 * Let's not sleep for all eternity if we've amassed a huge delay.
1794 * Swapping or metadata IO can accumulate 10's of seconds worth of
1795 * delay, and we want userspace to be able to do _something_ so cap the
1796 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1797 * tasks will be delayed for 0.25 second for every syscall. If
1798 * blkcg_set_delay() was used as indicated by negative use_delay, the
1799 * caller is responsible for regulating the range.
1802 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1805 psi_memstall_enter(&pflags);
1807 exp = ktime_add_ns(now, delay_nsec);
1808 tok = io_schedule_prepare();
1810 __set_current_state(TASK_KILLABLE);
1811 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1813 } while (!fatal_signal_pending(current));
1814 io_schedule_finish(tok);
1817 psi_memstall_leave(&pflags);
1821 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1823 * This is only called if we've been marked with set_notify_resume(). Obviously
1824 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1825 * check to see if current->throttle_disk is set and if not this doesn't do
1826 * anything. This should only ever be called by the resume code, it's not meant
1827 * to be called by people willy-nilly as it will actually do the work to
1828 * throttle the task if it is setup for throttling.
1830 void blkcg_maybe_throttle_current(void)
1832 struct gendisk *disk = current->throttle_disk;
1833 struct blkcg *blkcg;
1834 struct blkcg_gq *blkg;
1835 bool use_memdelay = current->use_memdelay;
1840 current->throttle_disk = NULL;
1841 current->use_memdelay = false;
1844 blkcg = css_to_blkcg(blkcg_css());
1847 blkg = blkg_lookup(blkcg, disk->queue);
1850 if (!blkg_tryget(blkg))
1854 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1863 * blkcg_schedule_throttle - this task needs to check for throttling
1864 * @disk: disk to throttle
1865 * @use_memdelay: do we charge this to memory delay for PSI
1867 * This is called by the IO controller when we know there's delay accumulated
1868 * for the blkg for this task. We do not pass the blkg because there are places
1869 * we call this that may not have that information, the swapping code for
1870 * instance will only have a block_device at that point. This set's the
1871 * notify_resume for the task to check and see if it requires throttling before
1872 * returning to user space.
1874 * We will only schedule once per syscall. You can call this over and over
1875 * again and it will only do the check once upon return to user space, and only
1876 * throttle once. If the task needs to be throttled again it'll need to be
1877 * re-set at the next time we see the task.
1879 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1881 if (unlikely(current->flags & PF_KTHREAD))
1884 if (current->throttle_disk != disk) {
1885 if (test_bit(GD_DEAD, &disk->state))
1887 get_device(disk_to_dev(disk));
1889 if (current->throttle_disk)
1890 put_disk(current->throttle_disk);
1891 current->throttle_disk = disk;
1895 current->use_memdelay = use_memdelay;
1896 set_notify_resume(current);
1900 * blkcg_add_delay - add delay to this blkg
1901 * @blkg: blkg of interest
1902 * @now: the current time in nanoseconds
1903 * @delta: how many nanoseconds of delay to add
1905 * Charge @delta to the blkg's current delay accumulation. This is used to
1906 * throttle tasks if an IO controller thinks we need more throttling.
1908 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1910 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1912 blkcg_scale_delay(blkg, now);
1913 atomic64_add(delta, &blkg->delay_nsec);
1917 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1921 * As the failure mode here is to walk up the blkg tree, this ensure that the
1922 * blkg->parent pointers are always valid. This returns the blkg that it ended
1923 * up taking a reference on or %NULL if no reference was taken.
1925 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1926 struct cgroup_subsys_state *css)
1928 struct blkcg_gq *blkg, *ret_blkg = NULL;
1931 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1933 if (blkg_tryget(blkg)) {
1937 blkg = blkg->parent;
1945 * bio_associate_blkg_from_css - associate a bio with a specified css
1949 * Associate @bio with the blkg found by combining the css's blkg and the
1950 * request_queue of the @bio. An association failure is handled by walking up
1951 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1952 * and q->root_blkg. This situation only happens when a cgroup is dying and
1953 * then the remaining bios will spill to the closest alive blkg.
1955 * A reference will be taken on the blkg and will be released when @bio is
1958 void bio_associate_blkg_from_css(struct bio *bio,
1959 struct cgroup_subsys_state *css)
1962 blkg_put(bio->bi_blkg);
1964 if (css && css->parent) {
1965 bio->bi_blkg = blkg_tryget_closest(bio, css);
1967 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1968 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1971 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1974 * bio_associate_blkg - associate a bio with a blkg
1977 * Associate @bio with the blkg found from the bio's css and request_queue.
1978 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1979 * already associated, the css is reused and association redone as the
1980 * request_queue may have changed.
1982 void bio_associate_blkg(struct bio *bio)
1984 struct cgroup_subsys_state *css;
1989 css = bio_blkcg_css(bio);
1993 bio_associate_blkg_from_css(bio, css);
1997 EXPORT_SYMBOL_GPL(bio_associate_blkg);
2000 * bio_clone_blkg_association - clone blkg association from src to dst bio
2001 * @dst: destination bio
2004 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2007 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2009 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2011 static int blk_cgroup_io_type(struct bio *bio)
2013 if (op_is_discard(bio->bi_opf))
2014 return BLKG_IOSTAT_DISCARD;
2015 if (op_is_write(bio->bi_opf))
2016 return BLKG_IOSTAT_WRITE;
2017 return BLKG_IOSTAT_READ;
2020 void blk_cgroup_bio_start(struct bio *bio)
2022 struct blkcg *blkcg = bio->bi_blkg->blkcg;
2023 int rwd = blk_cgroup_io_type(bio), cpu;
2024 struct blkg_iostat_set *bis;
2025 unsigned long flags;
2027 /* Root-level stats are sourced from system-wide IO stats */
2028 if (!cgroup_parent(blkcg->css.cgroup))
2032 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2033 flags = u64_stats_update_begin_irqsave(&bis->sync);
2036 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2037 * bio and we would have already accounted for the size of the bio.
2039 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2040 bio_set_flag(bio, BIO_CGROUP_ACCT);
2041 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2043 bis->cur.ios[rwd]++;
2046 * If the iostat_cpu isn't in a lockless list, put it into the
2047 * list to indicate that a stat update is pending.
2049 if (!READ_ONCE(bis->lqueued)) {
2050 struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2052 llist_add(&bis->lnode, lhead);
2053 WRITE_ONCE(bis->lqueued, true);
2054 percpu_ref_get(&bis->blkg->refcnt);
2057 u64_stats_update_end_irqrestore(&bis->sync, flags);
2058 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
2059 cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2063 bool blk_cgroup_congested(void)
2065 struct cgroup_subsys_state *css;
2069 for (css = blkcg_css(); css; css = css->parent) {
2070 if (atomic_read(&css->cgroup->congestion_count)) {
2079 static int __init blkcg_init(void)
2081 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
2082 WQ_MEM_RECLAIM | WQ_FREEZABLE |
2083 WQ_UNBOUND | WQ_SYSFS, 0);
2084 if (!blkcg_punt_bio_wq)
2088 subsys_initcall(blkcg_init);
2090 module_param(blkcg_debug_stats, bool, 0644);
2091 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");