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"
37 static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu);
40 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
41 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
42 * policy [un]register operations including cgroup file additions /
43 * removals. Putting cgroup file registration outside blkcg_pol_mutex
44 * allows grabbing it from cgroup callbacks.
46 static DEFINE_MUTEX(blkcg_pol_register_mutex);
47 static DEFINE_MUTEX(blkcg_pol_mutex);
49 struct blkcg blkcg_root;
50 EXPORT_SYMBOL_GPL(blkcg_root);
52 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
53 EXPORT_SYMBOL_GPL(blkcg_root_css);
55 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
57 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
59 bool blkcg_debug_stats = false;
61 static DEFINE_RAW_SPINLOCK(blkg_stat_lock);
63 #define BLKG_DESTROY_BATCH_SIZE 64
66 * Lockless lists for tracking IO stats update
68 * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
69 * There are multiple blkg's (one for each block device) attached to each
70 * blkcg. The rstat code keeps track of which cpu has IO stats updated,
71 * but it doesn't know which blkg has the updated stats. If there are many
72 * block devices in a system, the cost of iterating all the blkg's to flush
73 * out the IO stats can be high. To reduce such overhead, a set of percpu
74 * lockless lists (lhead) per blkcg are used to track the set of recently
75 * updated iostat_cpu's since the last flush. An iostat_cpu will be put
76 * onto the lockless list on the update side [blk_cgroup_bio_start()] if
77 * not there yet and then removed when being flushed [blkcg_rstat_flush()].
78 * References to blkg are gotten and then put back in the process to
79 * protect against blkg removal.
81 * Return: 0 if successful or -ENOMEM if allocation fails.
83 static int init_blkcg_llists(struct blkcg *blkcg)
87 blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL);
91 for_each_possible_cpu(cpu)
92 init_llist_head(per_cpu_ptr(blkcg->lhead, cpu));
97 * blkcg_css - find the current css
99 * Find the css associated with either the kthread or the current task.
100 * This may return a dying css, so it is up to the caller to use tryget logic
101 * to confirm it is alive and well.
103 static struct cgroup_subsys_state *blkcg_css(void)
105 struct cgroup_subsys_state *css;
107 css = kthread_blkcg();
110 return task_css(current, io_cgrp_id);
113 static bool blkcg_policy_enabled(struct request_queue *q,
114 const struct blkcg_policy *pol)
116 return pol && test_bit(pol->plid, q->blkcg_pols);
119 static void blkg_free_workfn(struct work_struct *work)
121 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
123 struct request_queue *q = blkg->q;
127 * pd_free_fn() can also be called from blkcg_deactivate_policy(),
128 * in order to make sure pd_free_fn() is called in order, the deletion
129 * of the list blkg->q_node is delayed to here from blkg_destroy(), and
130 * blkcg_mutex is used to synchronize blkg_free_workfn() and
131 * blkcg_deactivate_policy().
133 mutex_lock(&q->blkcg_mutex);
134 for (i = 0; i < BLKCG_MAX_POLS; i++)
136 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
138 blkg_put(blkg->parent);
139 list_del_init(&blkg->q_node);
140 mutex_unlock(&q->blkcg_mutex);
143 free_percpu(blkg->iostat_cpu);
144 percpu_ref_exit(&blkg->refcnt);
149 * blkg_free - free a blkg
150 * @blkg: blkg to free
152 * Free @blkg which may be partially allocated.
154 static void blkg_free(struct blkcg_gq *blkg)
160 * Both ->pd_free_fn() and request queue's release handler may
161 * sleep, so free us by scheduling one work func
163 INIT_WORK(&blkg->free_work, blkg_free_workfn);
164 schedule_work(&blkg->free_work);
167 static void __blkg_release(struct rcu_head *rcu)
169 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
170 struct blkcg *blkcg = blkg->blkcg;
173 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
174 WARN_ON(!bio_list_empty(&blkg->async_bios));
177 * Flush all the non-empty percpu lockless lists before releasing
178 * us, given these stat belongs to us.
180 * blkg_stat_lock is for serializing blkg stat update
182 for_each_possible_cpu(cpu)
183 __blkcg_rstat_flush(blkcg, cpu);
185 /* release the blkcg and parent blkg refs this blkg has been holding */
186 css_put(&blkg->blkcg->css);
191 * A group is RCU protected, but having an rcu lock does not mean that one
192 * can access all the fields of blkg and assume these are valid. For
193 * example, don't try to follow throtl_data and request queue links.
195 * Having a reference to blkg under an rcu allows accesses to only values
196 * local to groups like group stats and group rate limits.
198 static void blkg_release(struct percpu_ref *ref)
200 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
202 call_rcu(&blkg->rcu_head, __blkg_release);
205 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
206 static struct workqueue_struct *blkcg_punt_bio_wq;
208 static void blkg_async_bio_workfn(struct work_struct *work)
210 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
212 struct bio_list bios = BIO_EMPTY_LIST;
214 struct blk_plug plug;
215 bool need_plug = false;
217 /* as long as there are pending bios, @blkg can't go away */
218 spin_lock(&blkg->async_bio_lock);
219 bio_list_merge(&bios, &blkg->async_bios);
220 bio_list_init(&blkg->async_bios);
221 spin_unlock(&blkg->async_bio_lock);
223 /* start plug only when bio_list contains at least 2 bios */
224 if (bios.head && bios.head->bi_next) {
226 blk_start_plug(&plug);
228 while ((bio = bio_list_pop(&bios)))
231 blk_finish_plug(&plug);
235 * When a shared kthread issues a bio for a cgroup, doing so synchronously can
236 * lead to priority inversions as the kthread can be trapped waiting for that
237 * cgroup. Use this helper instead of submit_bio to punt the actual issuing to
238 * a dedicated per-blkcg work item to avoid such priority inversions.
240 void blkcg_punt_bio_submit(struct bio *bio)
242 struct blkcg_gq *blkg = bio->bi_blkg;
245 spin_lock(&blkg->async_bio_lock);
246 bio_list_add(&blkg->async_bios, bio);
247 spin_unlock(&blkg->async_bio_lock);
248 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
250 /* never bounce for the root cgroup */
254 EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit);
256 static int __init blkcg_punt_bio_init(void)
258 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
259 WQ_MEM_RECLAIM | WQ_FREEZABLE |
260 WQ_UNBOUND | WQ_SYSFS, 0);
261 if (!blkcg_punt_bio_wq)
265 subsys_initcall(blkcg_punt_bio_init);
266 #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
269 * bio_blkcg_css - return the blkcg CSS associated with a bio
272 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
273 * associated. Callers are expected to either handle %NULL or know association
274 * has been done prior to calling this.
276 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
278 if (!bio || !bio->bi_blkg)
280 return &bio->bi_blkg->blkcg->css;
282 EXPORT_SYMBOL_GPL(bio_blkcg_css);
285 * blkcg_parent - get the parent of a blkcg
286 * @blkcg: blkcg of interest
288 * Return the parent blkcg of @blkcg. Can be called anytime.
290 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
292 return css_to_blkcg(blkcg->css.parent);
296 * blkg_alloc - allocate a blkg
297 * @blkcg: block cgroup the new blkg is associated with
298 * @disk: gendisk the new blkg is associated with
299 * @gfp_mask: allocation mask to use
301 * Allocate a new blkg assocating @blkcg and @q.
303 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
306 struct blkcg_gq *blkg;
309 /* alloc and init base part */
310 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
313 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
315 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
316 if (!blkg->iostat_cpu)
317 goto out_exit_refcnt;
318 if (!blk_get_queue(disk->queue))
319 goto out_free_iostat;
321 blkg->q = disk->queue;
322 INIT_LIST_HEAD(&blkg->q_node);
324 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
325 spin_lock_init(&blkg->async_bio_lock);
326 bio_list_init(&blkg->async_bios);
327 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
330 u64_stats_init(&blkg->iostat.sync);
331 for_each_possible_cpu(cpu) {
332 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
333 per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
336 for (i = 0; i < BLKCG_MAX_POLS; i++) {
337 struct blkcg_policy *pol = blkcg_policy[i];
338 struct blkg_policy_data *pd;
340 if (!blkcg_policy_enabled(disk->queue, pol))
343 /* alloc per-policy data and attach it to blkg */
344 pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
358 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
359 blk_put_queue(disk->queue);
361 free_percpu(blkg->iostat_cpu);
363 percpu_ref_exit(&blkg->refcnt);
370 * If @new_blkg is %NULL, this function tries to allocate a new one as
371 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
373 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
374 struct blkcg_gq *new_blkg)
376 struct blkcg_gq *blkg;
379 lockdep_assert_held(&disk->queue->queue_lock);
381 /* request_queue is dying, do not create/recreate a blkg */
382 if (blk_queue_dying(disk->queue)) {
387 /* blkg holds a reference to blkcg */
388 if (!css_tryget_online(&blkcg->css)) {
395 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
396 if (unlikely(!new_blkg)) {
404 if (blkcg_parent(blkcg)) {
405 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
406 if (WARN_ON_ONCE(!blkg->parent)) {
410 blkg_get(blkg->parent);
413 /* invoke per-policy init */
414 for (i = 0; i < BLKCG_MAX_POLS; i++) {
415 struct blkcg_policy *pol = blkcg_policy[i];
417 if (blkg->pd[i] && pol->pd_init_fn)
418 pol->pd_init_fn(blkg->pd[i]);
422 spin_lock(&blkcg->lock);
423 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
425 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
426 list_add(&blkg->q_node, &disk->queue->blkg_list);
428 for (i = 0; i < BLKCG_MAX_POLS; i++) {
429 struct blkcg_policy *pol = blkcg_policy[i];
432 if (pol->pd_online_fn)
433 pol->pd_online_fn(blkg->pd[i]);
434 blkg->pd[i]->online = true;
439 spin_unlock(&blkcg->lock);
444 /* @blkg failed fully initialized, use the usual release path */
449 css_put(&blkcg->css);
457 * blkg_lookup_create - lookup blkg, try to create one if not there
458 * @blkcg: blkcg of interest
459 * @disk: gendisk of interest
461 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
462 * create one. blkg creation is performed recursively from blkcg_root such
463 * that all non-root blkg's have access to the parent blkg. This function
464 * should be called under RCU read lock and takes @disk->queue->queue_lock.
466 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
469 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
470 struct gendisk *disk)
472 struct request_queue *q = disk->queue;
473 struct blkcg_gq *blkg;
476 WARN_ON_ONCE(!rcu_read_lock_held());
478 blkg = blkg_lookup(blkcg, q);
482 spin_lock_irqsave(&q->queue_lock, flags);
483 blkg = blkg_lookup(blkcg, q);
485 if (blkcg != &blkcg_root &&
486 blkg != rcu_dereference(blkcg->blkg_hint))
487 rcu_assign_pointer(blkcg->blkg_hint, blkg);
492 * Create blkgs walking down from blkcg_root to @blkcg, so that all
493 * non-root blkgs have access to their parents. Returns the closest
494 * blkg to the intended blkg should blkg_create() fail.
497 struct blkcg *pos = blkcg;
498 struct blkcg *parent = blkcg_parent(blkcg);
499 struct blkcg_gq *ret_blkg = q->root_blkg;
502 blkg = blkg_lookup(parent, q);
504 /* remember closest blkg */
509 parent = blkcg_parent(parent);
512 blkg = blkg_create(pos, disk, NULL);
522 spin_unlock_irqrestore(&q->queue_lock, flags);
526 static void blkg_destroy(struct blkcg_gq *blkg)
528 struct blkcg *blkcg = blkg->blkcg;
531 lockdep_assert_held(&blkg->q->queue_lock);
532 lockdep_assert_held(&blkcg->lock);
535 * blkg stays on the queue list until blkg_free_workfn(), see details in
536 * blkg_free_workfn(), hence this function can be called from
537 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
538 * blkg_free_workfn().
540 if (hlist_unhashed(&blkg->blkcg_node))
543 for (i = 0; i < BLKCG_MAX_POLS; i++) {
544 struct blkcg_policy *pol = blkcg_policy[i];
546 if (blkg->pd[i] && blkg->pd[i]->online) {
547 blkg->pd[i]->online = false;
548 if (pol->pd_offline_fn)
549 pol->pd_offline_fn(blkg->pd[i]);
553 blkg->online = false;
555 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
556 hlist_del_init_rcu(&blkg->blkcg_node);
559 * Both setting lookup hint to and clearing it from @blkg are done
560 * under queue_lock. If it's not pointing to @blkg now, it never
561 * will. Hint assignment itself can race safely.
563 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
564 rcu_assign_pointer(blkcg->blkg_hint, NULL);
567 * Put the reference taken at the time of creation so that when all
568 * queues are gone, group can be destroyed.
570 percpu_ref_kill(&blkg->refcnt);
573 static void blkg_destroy_all(struct gendisk *disk)
575 struct request_queue *q = disk->queue;
576 struct blkcg_gq *blkg, *n;
577 int count = BLKG_DESTROY_BATCH_SIZE;
580 spin_lock_irq(&q->queue_lock);
581 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
582 struct blkcg *blkcg = blkg->blkcg;
584 if (hlist_unhashed(&blkg->blkcg_node))
587 spin_lock(&blkcg->lock);
589 spin_unlock(&blkcg->lock);
592 * in order to avoid holding the spin lock for too long, release
593 * it when a batch of blkgs are destroyed.
596 count = BLKG_DESTROY_BATCH_SIZE;
597 spin_unlock_irq(&q->queue_lock);
604 spin_unlock_irq(&q->queue_lock);
607 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
608 struct cftype *cftype, u64 val)
610 struct blkcg *blkcg = css_to_blkcg(css);
611 struct blkcg_gq *blkg;
614 mutex_lock(&blkcg_pol_mutex);
615 spin_lock_irq(&blkcg->lock);
618 * Note that stat reset is racy - it doesn't synchronize against
619 * stat updates. This is a debug feature which shouldn't exist
620 * anyway. If you get hit by a race, retry.
622 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
623 for_each_possible_cpu(cpu) {
624 struct blkg_iostat_set *bis =
625 per_cpu_ptr(blkg->iostat_cpu, cpu);
626 memset(bis, 0, sizeof(*bis));
628 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
630 for (i = 0; i < BLKCG_MAX_POLS; i++) {
631 struct blkcg_policy *pol = blkcg_policy[i];
633 if (blkg->pd[i] && pol->pd_reset_stats_fn)
634 pol->pd_reset_stats_fn(blkg->pd[i]);
638 spin_unlock_irq(&blkcg->lock);
639 mutex_unlock(&blkcg_pol_mutex);
643 const char *blkg_dev_name(struct blkcg_gq *blkg)
647 return bdi_dev_name(blkg->q->disk->bdi);
651 * blkcg_print_blkgs - helper for printing per-blkg data
652 * @sf: seq_file to print to
653 * @blkcg: blkcg of interest
654 * @prfill: fill function to print out a blkg
655 * @pol: policy in question
656 * @data: data to be passed to @prfill
657 * @show_total: to print out sum of prfill return values or not
659 * This function invokes @prfill on each blkg of @blkcg if pd for the
660 * policy specified by @pol exists. @prfill is invoked with @sf, the
661 * policy data and @data and the matching queue lock held. If @show_total
662 * is %true, the sum of the return values from @prfill is printed with
663 * "Total" label at the end.
665 * This is to be used to construct print functions for
666 * cftype->read_seq_string method.
668 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
669 u64 (*prfill)(struct seq_file *,
670 struct blkg_policy_data *, int),
671 const struct blkcg_policy *pol, int data,
674 struct blkcg_gq *blkg;
678 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
679 spin_lock_irq(&blkg->q->queue_lock);
680 if (blkcg_policy_enabled(blkg->q, pol))
681 total += prfill(sf, blkg->pd[pol->plid], data);
682 spin_unlock_irq(&blkg->q->queue_lock);
687 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
689 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
692 * __blkg_prfill_u64 - prfill helper for a single u64 value
693 * @sf: seq_file to print to
694 * @pd: policy private data of interest
697 * Print @v to @sf for the device associated with @pd.
699 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
701 const char *dname = blkg_dev_name(pd->blkg);
706 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
709 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
712 * blkg_conf_init - initialize a blkg_conf_ctx
713 * @ctx: blkg_conf_ctx to initialize
714 * @input: input string
716 * Initialize @ctx which can be used to parse blkg config input string @input.
717 * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
718 * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
720 void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
722 *ctx = (struct blkg_conf_ctx){ .input = input };
724 EXPORT_SYMBOL_GPL(blkg_conf_init);
727 * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
728 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
730 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
731 * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
732 * set to point past the device node prefix.
734 * This function may be called multiple times on @ctx and the extra calls become
735 * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
736 * explicitly if bdev access is needed without resolving the blkcg / policy part
737 * of @ctx->input. Returns -errno on error.
739 int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
741 char *input = ctx->input;
742 unsigned int major, minor;
743 struct block_device *bdev;
749 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
753 if (!isspace(*input))
755 input = skip_spaces(input);
757 bdev = blkdev_get_no_open(MKDEV(major, minor));
760 if (bdev_is_partition(bdev)) {
761 blkdev_put_no_open(bdev);
771 * blkg_conf_prep - parse and prepare for per-blkg config update
772 * @blkcg: target block cgroup
773 * @pol: target policy
774 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
776 * Parse per-blkg config update from @ctx->input and initialize @ctx
777 * accordingly. On success, @ctx->body points to the part of @ctx->input
778 * following MAJ:MIN, @ctx->bdev points to the target block device and
779 * @ctx->blkg to the blkg being configured.
781 * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
782 * function returns with queue lock held and must be followed by
785 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
786 struct blkg_conf_ctx *ctx)
787 __acquires(&bdev->bd_queue->queue_lock)
789 struct gendisk *disk;
790 struct request_queue *q;
791 struct blkcg_gq *blkg;
794 ret = blkg_conf_open_bdev(ctx);
798 disk = ctx->bdev->bd_disk;
802 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
803 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
805 ret = blk_queue_enter(q, 0);
809 spin_lock_irq(&q->queue_lock);
811 if (!blkcg_policy_enabled(q, pol)) {
816 blkg = blkg_lookup(blkcg, q);
821 * Create blkgs walking down from blkcg_root to @blkcg, so that all
822 * non-root blkgs have access to their parents.
825 struct blkcg *pos = blkcg;
826 struct blkcg *parent;
827 struct blkcg_gq *new_blkg;
829 parent = blkcg_parent(blkcg);
830 while (parent && !blkg_lookup(parent, q)) {
832 parent = blkcg_parent(parent);
835 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
836 spin_unlock_irq(&q->queue_lock);
838 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
839 if (unlikely(!new_blkg)) {
841 goto fail_exit_queue;
844 if (radix_tree_preload(GFP_KERNEL)) {
847 goto fail_exit_queue;
850 spin_lock_irq(&q->queue_lock);
852 if (!blkcg_policy_enabled(q, pol)) {
858 blkg = blkg_lookup(pos, q);
862 blkg = blkg_create(pos, disk, new_blkg);
869 radix_tree_preload_end();
880 radix_tree_preload_end();
882 spin_unlock_irq(&q->queue_lock);
887 * If queue was bypassing, we should retry. Do so after a
888 * short msleep(). It isn't strictly necessary but queue
889 * can be bypassing for some time and it's always nice to
890 * avoid busy looping.
894 ret = restart_syscall();
898 EXPORT_SYMBOL_GPL(blkg_conf_prep);
901 * blkg_conf_exit - clean up per-blkg config update
902 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
904 * Clean up after per-blkg config update. This function must be called on all
905 * blkg_conf_ctx's initialized with blkg_conf_init().
907 void blkg_conf_exit(struct blkg_conf_ctx *ctx)
908 __releases(&ctx->bdev->bd_queue->queue_lock)
911 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
916 blkdev_put_no_open(ctx->bdev);
921 EXPORT_SYMBOL_GPL(blkg_conf_exit);
923 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
927 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
928 dst->bytes[i] = src->bytes[i];
929 dst->ios[i] = src->ios[i];
933 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
937 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
938 dst->bytes[i] += src->bytes[i];
939 dst->ios[i] += src->ios[i];
943 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
947 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
948 dst->bytes[i] -= src->bytes[i];
949 dst->ios[i] -= src->ios[i];
953 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
954 struct blkg_iostat *last)
956 struct blkg_iostat delta;
959 /* propagate percpu delta to global */
960 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
961 blkg_iostat_set(&delta, cur);
962 blkg_iostat_sub(&delta, last);
963 blkg_iostat_add(&blkg->iostat.cur, &delta);
964 blkg_iostat_add(last, &delta);
965 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
968 static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
970 struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
971 struct llist_node *lnode;
972 struct blkg_iostat_set *bisc, *next_bisc;
977 lnode = llist_del_all(lhead);
982 * For covering concurrent parent blkg update from blkg_release().
984 * When flushing from cgroup, cgroup_rstat_lock is always held, so
985 * this lock won't cause contention most of time.
987 raw_spin_lock_irqsave(&blkg_stat_lock, flags);
990 * Iterate only the iostat_cpu's queued in the lockless list.
992 llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
993 struct blkcg_gq *blkg = bisc->blkg;
994 struct blkcg_gq *parent = blkg->parent;
995 struct blkg_iostat cur;
998 WRITE_ONCE(bisc->lqueued, false);
1000 /* fetch the current per-cpu values */
1002 seq = u64_stats_fetch_begin(&bisc->sync);
1003 blkg_iostat_set(&cur, &bisc->cur);
1004 } while (u64_stats_fetch_retry(&bisc->sync, seq));
1006 blkcg_iostat_update(blkg, &cur, &bisc->last);
1008 /* propagate global delta to parent (unless that's root) */
1009 if (parent && parent->parent)
1010 blkcg_iostat_update(parent, &blkg->iostat.cur,
1011 &blkg->iostat.last);
1013 raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
1018 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1020 /* Root-level stats are sourced from system-wide IO stats */
1021 if (cgroup_parent(css->cgroup))
1022 __blkcg_rstat_flush(css_to_blkcg(css), cpu);
1026 * We source root cgroup stats from the system-wide stats to avoid
1027 * tracking the same information twice and incurring overhead when no
1028 * cgroups are defined. For that reason, cgroup_rstat_flush in
1029 * blkcg_print_stat does not actually fill out the iostat in the root
1030 * cgroup's blkcg_gq.
1032 * However, we would like to re-use the printing code between the root and
1033 * non-root cgroups to the extent possible. For that reason, we simulate
1034 * flushing the root cgroup's stats by explicitly filling in the iostat
1035 * with disk level statistics.
1037 static void blkcg_fill_root_iostats(void)
1039 struct class_dev_iter iter;
1042 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
1043 while ((dev = class_dev_iter_next(&iter))) {
1044 struct block_device *bdev = dev_to_bdev(dev);
1045 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
1046 struct blkg_iostat tmp;
1048 unsigned long flags;
1050 memset(&tmp, 0, sizeof(tmp));
1051 for_each_possible_cpu(cpu) {
1052 struct disk_stats *cpu_dkstats;
1054 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1055 tmp.ios[BLKG_IOSTAT_READ] +=
1056 cpu_dkstats->ios[STAT_READ];
1057 tmp.ios[BLKG_IOSTAT_WRITE] +=
1058 cpu_dkstats->ios[STAT_WRITE];
1059 tmp.ios[BLKG_IOSTAT_DISCARD] +=
1060 cpu_dkstats->ios[STAT_DISCARD];
1061 // convert sectors to bytes
1062 tmp.bytes[BLKG_IOSTAT_READ] +=
1063 cpu_dkstats->sectors[STAT_READ] << 9;
1064 tmp.bytes[BLKG_IOSTAT_WRITE] +=
1065 cpu_dkstats->sectors[STAT_WRITE] << 9;
1066 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1067 cpu_dkstats->sectors[STAT_DISCARD] << 9;
1070 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1071 blkg_iostat_set(&blkg->iostat.cur, &tmp);
1072 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1076 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1078 struct blkg_iostat_set *bis = &blkg->iostat;
1079 u64 rbytes, wbytes, rios, wios, dbytes, dios;
1087 dname = blkg_dev_name(blkg);
1091 seq_printf(s, "%s ", dname);
1094 seq = u64_stats_fetch_begin(&bis->sync);
1096 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1097 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1098 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1099 rios = bis->cur.ios[BLKG_IOSTAT_READ];
1100 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1101 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1102 } while (u64_stats_fetch_retry(&bis->sync, seq));
1104 if (rbytes || wbytes || rios || wios) {
1105 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1106 rbytes, wbytes, rios, wios,
1110 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1111 seq_printf(s, " use_delay=%d delay_nsec=%llu",
1112 atomic_read(&blkg->use_delay),
1113 atomic64_read(&blkg->delay_nsec));
1116 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1117 struct blkcg_policy *pol = blkcg_policy[i];
1119 if (!blkg->pd[i] || !pol->pd_stat_fn)
1122 pol->pd_stat_fn(blkg->pd[i], s);
1128 static int blkcg_print_stat(struct seq_file *sf, void *v)
1130 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1131 struct blkcg_gq *blkg;
1133 if (!seq_css(sf)->parent)
1134 blkcg_fill_root_iostats();
1136 cgroup_rstat_flush(blkcg->css.cgroup);
1139 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1140 spin_lock_irq(&blkg->q->queue_lock);
1141 blkcg_print_one_stat(blkg, sf);
1142 spin_unlock_irq(&blkg->q->queue_lock);
1148 static struct cftype blkcg_files[] = {
1151 .seq_show = blkcg_print_stat,
1156 static struct cftype blkcg_legacy_files[] = {
1158 .name = "reset_stats",
1159 .write_u64 = blkcg_reset_stats,
1164 #ifdef CONFIG_CGROUP_WRITEBACK
1165 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1167 return &css_to_blkcg(css)->cgwb_list;
1172 * blkcg destruction is a three-stage process.
1174 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1175 * which offlines writeback. Here we tie the next stage of blkg destruction
1176 * to the completion of writeback associated with the blkcg. This lets us
1177 * avoid punting potentially large amounts of outstanding writeback to root
1178 * while maintaining any ongoing policies. The next stage is triggered when
1179 * the nr_cgwbs count goes to zero.
1181 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1182 * and handles the destruction of blkgs. Here the css reference held by
1183 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1184 * This work may occur in cgwb_release_workfn() on the cgwb_release
1185 * workqueue. Any submitted ios that fail to get the blkg ref will be
1186 * punted to the root_blkg.
1188 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1189 * This finally frees the blkcg.
1193 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1194 * @blkcg: blkcg of interest
1196 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1197 * is nested inside q lock, this function performs reverse double lock dancing.
1198 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1199 * blkcg_css_free to eventually be called.
1201 * This is the blkcg counterpart of ioc_release_fn().
1203 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1207 spin_lock_irq(&blkcg->lock);
1209 while (!hlist_empty(&blkcg->blkg_list)) {
1210 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1211 struct blkcg_gq, blkcg_node);
1212 struct request_queue *q = blkg->q;
1214 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1216 * Given that the system can accumulate a huge number
1217 * of blkgs in pathological cases, check to see if we
1218 * need to rescheduling to avoid softlockup.
1220 spin_unlock_irq(&blkcg->lock);
1222 spin_lock_irq(&blkcg->lock);
1227 spin_unlock(&q->queue_lock);
1230 spin_unlock_irq(&blkcg->lock);
1234 * blkcg_pin_online - pin online state
1235 * @blkcg_css: blkcg of interest
1237 * While pinned, a blkcg is kept online. This is primarily used to
1238 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1239 * while an associated cgwb is still active.
1241 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1243 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1247 * blkcg_unpin_online - unpin online state
1248 * @blkcg_css: blkcg of interest
1250 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1251 * that blkg doesn't go offline while an associated cgwb is still active.
1252 * When this count goes to zero, all active cgwbs have finished so the
1253 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1255 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1257 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1260 if (!refcount_dec_and_test(&blkcg->online_pin))
1262 blkcg_destroy_blkgs(blkcg);
1263 blkcg = blkcg_parent(blkcg);
1268 * blkcg_css_offline - cgroup css_offline callback
1269 * @css: css of interest
1271 * This function is called when @css is about to go away. Here the cgwbs are
1272 * offlined first and only once writeback associated with the blkcg has
1273 * finished do we start step 2 (see above).
1275 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1277 /* this prevents anyone from attaching or migrating to this blkcg */
1278 wb_blkcg_offline(css);
1280 /* put the base online pin allowing step 2 to be triggered */
1281 blkcg_unpin_online(css);
1284 static void blkcg_css_free(struct cgroup_subsys_state *css)
1286 struct blkcg *blkcg = css_to_blkcg(css);
1289 mutex_lock(&blkcg_pol_mutex);
1291 list_del(&blkcg->all_blkcgs_node);
1293 for (i = 0; i < BLKCG_MAX_POLS; i++)
1295 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1297 mutex_unlock(&blkcg_pol_mutex);
1299 free_percpu(blkcg->lhead);
1303 static struct cgroup_subsys_state *
1304 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1306 struct blkcg *blkcg;
1309 mutex_lock(&blkcg_pol_mutex);
1312 blkcg = &blkcg_root;
1314 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1319 if (init_blkcg_llists(blkcg))
1322 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1323 struct blkcg_policy *pol = blkcg_policy[i];
1324 struct blkcg_policy_data *cpd;
1327 * If the policy hasn't been attached yet, wait for it
1328 * to be attached before doing anything else. Otherwise,
1329 * check if the policy requires any specific per-cgroup
1330 * data: if it does, allocate and initialize it.
1332 if (!pol || !pol->cpd_alloc_fn)
1335 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1339 blkcg->cpd[i] = cpd;
1344 spin_lock_init(&blkcg->lock);
1345 refcount_set(&blkcg->online_pin, 1);
1346 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1347 INIT_HLIST_HEAD(&blkcg->blkg_list);
1348 #ifdef CONFIG_CGROUP_WRITEBACK
1349 INIT_LIST_HEAD(&blkcg->cgwb_list);
1351 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1353 mutex_unlock(&blkcg_pol_mutex);
1357 for (i--; i >= 0; i--)
1359 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1360 free_percpu(blkcg->lhead);
1362 if (blkcg != &blkcg_root)
1365 mutex_unlock(&blkcg_pol_mutex);
1366 return ERR_PTR(-ENOMEM);
1369 static int blkcg_css_online(struct cgroup_subsys_state *css)
1371 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1374 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1375 * don't go offline while cgwbs are still active on them. Pin the
1376 * parent so that offline always happens towards the root.
1379 blkcg_pin_online(&parent->css);
1383 int blkcg_init_disk(struct gendisk *disk)
1385 struct request_queue *q = disk->queue;
1386 struct blkcg_gq *new_blkg, *blkg;
1390 INIT_LIST_HEAD(&q->blkg_list);
1391 mutex_init(&q->blkcg_mutex);
1393 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1397 preloaded = !radix_tree_preload(GFP_KERNEL);
1399 /* Make sure the root blkg exists. */
1400 /* spin_lock_irq can serve as RCU read-side critical section. */
1401 spin_lock_irq(&q->queue_lock);
1402 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1405 q->root_blkg = blkg;
1406 spin_unlock_irq(&q->queue_lock);
1409 radix_tree_preload_end();
1411 ret = blk_ioprio_init(disk);
1413 goto err_destroy_all;
1415 ret = blk_throtl_init(disk);
1417 goto err_ioprio_exit;
1422 blk_ioprio_exit(disk);
1424 blkg_destroy_all(disk);
1427 spin_unlock_irq(&q->queue_lock);
1429 radix_tree_preload_end();
1430 return PTR_ERR(blkg);
1433 void blkcg_exit_disk(struct gendisk *disk)
1435 blkg_destroy_all(disk);
1436 blk_throtl_exit(disk);
1439 static void blkcg_exit(struct task_struct *tsk)
1441 if (tsk->throttle_disk)
1442 put_disk(tsk->throttle_disk);
1443 tsk->throttle_disk = NULL;
1446 struct cgroup_subsys io_cgrp_subsys = {
1447 .css_alloc = blkcg_css_alloc,
1448 .css_online = blkcg_css_online,
1449 .css_offline = blkcg_css_offline,
1450 .css_free = blkcg_css_free,
1451 .css_rstat_flush = blkcg_rstat_flush,
1452 .dfl_cftypes = blkcg_files,
1453 .legacy_cftypes = blkcg_legacy_files,
1454 .legacy_name = "blkio",
1458 * This ensures that, if available, memcg is automatically enabled
1459 * together on the default hierarchy so that the owner cgroup can
1460 * be retrieved from writeback pages.
1462 .depends_on = 1 << memory_cgrp_id,
1465 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1468 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1469 * @disk: gendisk of interest
1470 * @pol: blkcg policy to activate
1472 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1473 * bypass mode to populate its blkgs with policy_data for @pol.
1475 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1476 * from IO path. Update of each blkg is protected by both queue and blkcg
1477 * locks so that holding either lock and testing blkcg_policy_enabled() is
1478 * always enough for dereferencing policy data.
1480 * The caller is responsible for synchronizing [de]activations and policy
1481 * [un]registerations. Returns 0 on success, -errno on failure.
1483 int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1485 struct request_queue *q = disk->queue;
1486 struct blkg_policy_data *pd_prealloc = NULL;
1487 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1490 if (blkcg_policy_enabled(q, pol))
1494 blk_mq_freeze_queue(q);
1496 spin_lock_irq(&q->queue_lock);
1498 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1499 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1500 struct blkg_policy_data *pd;
1502 if (blkg->pd[pol->plid])
1505 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1506 if (blkg == pinned_blkg) {
1510 pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1511 GFP_NOWAIT | __GFP_NOWARN);
1516 * GFP_NOWAIT failed. Free the existing one and
1517 * prealloc for @blkg w/ GFP_KERNEL.
1520 blkg_put(pinned_blkg);
1524 spin_unlock_irq(&q->queue_lock);
1527 pol->pd_free_fn(pd_prealloc);
1528 pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1536 blkg->pd[pol->plid] = pd;
1538 pd->plid = pol->plid;
1542 /* all allocated, init in the same order */
1543 if (pol->pd_init_fn)
1544 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1545 pol->pd_init_fn(blkg->pd[pol->plid]);
1547 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1548 if (pol->pd_online_fn)
1549 pol->pd_online_fn(blkg->pd[pol->plid]);
1550 blkg->pd[pol->plid]->online = true;
1553 __set_bit(pol->plid, q->blkcg_pols);
1556 spin_unlock_irq(&q->queue_lock);
1559 blk_mq_unfreeze_queue(q);
1561 blkg_put(pinned_blkg);
1563 pol->pd_free_fn(pd_prealloc);
1567 /* alloc failed, nothing's initialized yet, free everything */
1568 spin_lock_irq(&q->queue_lock);
1569 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1570 struct blkcg *blkcg = blkg->blkcg;
1572 spin_lock(&blkcg->lock);
1573 if (blkg->pd[pol->plid]) {
1574 pol->pd_free_fn(blkg->pd[pol->plid]);
1575 blkg->pd[pol->plid] = NULL;
1577 spin_unlock(&blkcg->lock);
1579 spin_unlock_irq(&q->queue_lock);
1583 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1586 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1587 * @disk: gendisk of interest
1588 * @pol: blkcg policy to deactivate
1590 * Deactivate @pol on @disk. Follows the same synchronization rules as
1591 * blkcg_activate_policy().
1593 void blkcg_deactivate_policy(struct gendisk *disk,
1594 const struct blkcg_policy *pol)
1596 struct request_queue *q = disk->queue;
1597 struct blkcg_gq *blkg;
1599 if (!blkcg_policy_enabled(q, pol))
1603 blk_mq_freeze_queue(q);
1605 mutex_lock(&q->blkcg_mutex);
1606 spin_lock_irq(&q->queue_lock);
1608 __clear_bit(pol->plid, q->blkcg_pols);
1610 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1611 struct blkcg *blkcg = blkg->blkcg;
1613 spin_lock(&blkcg->lock);
1614 if (blkg->pd[pol->plid]) {
1615 if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1616 pol->pd_offline_fn(blkg->pd[pol->plid]);
1617 pol->pd_free_fn(blkg->pd[pol->plid]);
1618 blkg->pd[pol->plid] = NULL;
1620 spin_unlock(&blkcg->lock);
1623 spin_unlock_irq(&q->queue_lock);
1624 mutex_unlock(&q->blkcg_mutex);
1627 blk_mq_unfreeze_queue(q);
1629 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1631 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1633 struct blkcg *blkcg;
1635 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1636 if (blkcg->cpd[pol->plid]) {
1637 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1638 blkcg->cpd[pol->plid] = NULL;
1644 * blkcg_policy_register - register a blkcg policy
1645 * @pol: blkcg policy to register
1647 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1648 * successful registration. Returns 0 on success and -errno on failure.
1650 int blkcg_policy_register(struct blkcg_policy *pol)
1652 struct blkcg *blkcg;
1655 mutex_lock(&blkcg_pol_register_mutex);
1656 mutex_lock(&blkcg_pol_mutex);
1658 /* find an empty slot */
1660 for (i = 0; i < BLKCG_MAX_POLS; i++)
1661 if (!blkcg_policy[i])
1663 if (i >= BLKCG_MAX_POLS) {
1664 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1668 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1669 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1670 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1675 blkcg_policy[pol->plid] = pol;
1677 /* allocate and install cpd's */
1678 if (pol->cpd_alloc_fn) {
1679 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1680 struct blkcg_policy_data *cpd;
1682 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1686 blkcg->cpd[pol->plid] = cpd;
1688 cpd->plid = pol->plid;
1692 mutex_unlock(&blkcg_pol_mutex);
1694 /* everything is in place, add intf files for the new policy */
1695 if (pol->dfl_cftypes)
1696 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1698 if (pol->legacy_cftypes)
1699 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1700 pol->legacy_cftypes));
1701 mutex_unlock(&blkcg_pol_register_mutex);
1705 if (pol->cpd_free_fn)
1706 blkcg_free_all_cpd(pol);
1708 blkcg_policy[pol->plid] = NULL;
1710 mutex_unlock(&blkcg_pol_mutex);
1711 mutex_unlock(&blkcg_pol_register_mutex);
1714 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1717 * blkcg_policy_unregister - unregister a blkcg policy
1718 * @pol: blkcg policy to unregister
1720 * Undo blkcg_policy_register(@pol). Might sleep.
1722 void blkcg_policy_unregister(struct blkcg_policy *pol)
1724 mutex_lock(&blkcg_pol_register_mutex);
1726 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1729 /* kill the intf files first */
1730 if (pol->dfl_cftypes)
1731 cgroup_rm_cftypes(pol->dfl_cftypes);
1732 if (pol->legacy_cftypes)
1733 cgroup_rm_cftypes(pol->legacy_cftypes);
1735 /* remove cpds and unregister */
1736 mutex_lock(&blkcg_pol_mutex);
1738 if (pol->cpd_free_fn)
1739 blkcg_free_all_cpd(pol);
1741 blkcg_policy[pol->plid] = NULL;
1743 mutex_unlock(&blkcg_pol_mutex);
1745 mutex_unlock(&blkcg_pol_register_mutex);
1747 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1750 * Scale the accumulated delay based on how long it has been since we updated
1751 * the delay. We only call this when we are adding delay, in case it's been a
1752 * while since we added delay, and when we are checking to see if we need to
1753 * delay a task, to account for any delays that may have occurred.
1755 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1757 u64 old = atomic64_read(&blkg->delay_start);
1759 /* negative use_delay means no scaling, see blkcg_set_delay() */
1760 if (atomic_read(&blkg->use_delay) < 0)
1764 * We only want to scale down every second. The idea here is that we
1765 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1766 * time window. We only want to throttle tasks for recent delay that
1767 * has occurred, in 1 second time windows since that's the maximum
1768 * things can be throttled. We save the current delay window in
1769 * blkg->last_delay so we know what amount is still left to be charged
1770 * to the blkg from this point onward. blkg->last_use keeps track of
1771 * the use_delay counter. The idea is if we're unthrottling the blkg we
1772 * are ok with whatever is happening now, and we can take away more of
1773 * the accumulated delay as we've already throttled enough that
1774 * everybody is happy with their IO latencies.
1776 if (time_before64(old + NSEC_PER_SEC, now) &&
1777 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1778 u64 cur = atomic64_read(&blkg->delay_nsec);
1779 u64 sub = min_t(u64, blkg->last_delay, now - old);
1780 int cur_use = atomic_read(&blkg->use_delay);
1783 * We've been unthrottled, subtract a larger chunk of our
1784 * accumulated delay.
1786 if (cur_use < blkg->last_use)
1787 sub = max_t(u64, sub, blkg->last_delay >> 1);
1790 * This shouldn't happen, but handle it anyway. Our delay_nsec
1791 * should only ever be growing except here where we subtract out
1792 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1793 * rather not end up with negative numbers.
1795 if (unlikely(cur < sub)) {
1796 atomic64_set(&blkg->delay_nsec, 0);
1797 blkg->last_delay = 0;
1799 atomic64_sub(sub, &blkg->delay_nsec);
1800 blkg->last_delay = cur - sub;
1802 blkg->last_use = cur_use;
1807 * This is called when we want to actually walk up the hierarchy and check to
1808 * see if we need to throttle, and then actually throttle if there is some
1809 * accumulated delay. This should only be called upon return to user space so
1810 * we're not holding some lock that would induce a priority inversion.
1812 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1814 unsigned long pflags;
1816 u64 now = ktime_to_ns(ktime_get());
1821 while (blkg->parent) {
1822 int use_delay = atomic_read(&blkg->use_delay);
1827 blkcg_scale_delay(blkg, now);
1828 this_delay = atomic64_read(&blkg->delay_nsec);
1829 if (this_delay > delay_nsec) {
1830 delay_nsec = this_delay;
1831 clamp = use_delay > 0;
1834 blkg = blkg->parent;
1841 * Let's not sleep for all eternity if we've amassed a huge delay.
1842 * Swapping or metadata IO can accumulate 10's of seconds worth of
1843 * delay, and we want userspace to be able to do _something_ so cap the
1844 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1845 * tasks will be delayed for 0.25 second for every syscall. If
1846 * blkcg_set_delay() was used as indicated by negative use_delay, the
1847 * caller is responsible for regulating the range.
1850 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1853 psi_memstall_enter(&pflags);
1855 exp = ktime_add_ns(now, delay_nsec);
1856 tok = io_schedule_prepare();
1858 __set_current_state(TASK_KILLABLE);
1859 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1861 } while (!fatal_signal_pending(current));
1862 io_schedule_finish(tok);
1865 psi_memstall_leave(&pflags);
1869 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1871 * This is only called if we've been marked with set_notify_resume(). Obviously
1872 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1873 * check to see if current->throttle_disk is set and if not this doesn't do
1874 * anything. This should only ever be called by the resume code, it's not meant
1875 * to be called by people willy-nilly as it will actually do the work to
1876 * throttle the task if it is setup for throttling.
1878 void blkcg_maybe_throttle_current(void)
1880 struct gendisk *disk = current->throttle_disk;
1881 struct blkcg *blkcg;
1882 struct blkcg_gq *blkg;
1883 bool use_memdelay = current->use_memdelay;
1888 current->throttle_disk = NULL;
1889 current->use_memdelay = false;
1892 blkcg = css_to_blkcg(blkcg_css());
1895 blkg = blkg_lookup(blkcg, disk->queue);
1898 if (!blkg_tryget(blkg))
1902 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1911 * blkcg_schedule_throttle - this task needs to check for throttling
1912 * @disk: disk to throttle
1913 * @use_memdelay: do we charge this to memory delay for PSI
1915 * This is called by the IO controller when we know there's delay accumulated
1916 * for the blkg for this task. We do not pass the blkg because there are places
1917 * we call this that may not have that information, the swapping code for
1918 * instance will only have a block_device at that point. This set's the
1919 * notify_resume for the task to check and see if it requires throttling before
1920 * returning to user space.
1922 * We will only schedule once per syscall. You can call this over and over
1923 * again and it will only do the check once upon return to user space, and only
1924 * throttle once. If the task needs to be throttled again it'll need to be
1925 * re-set at the next time we see the task.
1927 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1929 if (unlikely(current->flags & PF_KTHREAD))
1932 if (current->throttle_disk != disk) {
1933 if (test_bit(GD_DEAD, &disk->state))
1935 get_device(disk_to_dev(disk));
1937 if (current->throttle_disk)
1938 put_disk(current->throttle_disk);
1939 current->throttle_disk = disk;
1943 current->use_memdelay = use_memdelay;
1944 set_notify_resume(current);
1948 * blkcg_add_delay - add delay to this blkg
1949 * @blkg: blkg of interest
1950 * @now: the current time in nanoseconds
1951 * @delta: how many nanoseconds of delay to add
1953 * Charge @delta to the blkg's current delay accumulation. This is used to
1954 * throttle tasks if an IO controller thinks we need more throttling.
1956 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1958 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1960 blkcg_scale_delay(blkg, now);
1961 atomic64_add(delta, &blkg->delay_nsec);
1965 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1969 * As the failure mode here is to walk up the blkg tree, this ensure that the
1970 * blkg->parent pointers are always valid. This returns the blkg that it ended
1971 * up taking a reference on or %NULL if no reference was taken.
1973 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1974 struct cgroup_subsys_state *css)
1976 struct blkcg_gq *blkg, *ret_blkg = NULL;
1979 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1981 if (blkg_tryget(blkg)) {
1985 blkg = blkg->parent;
1993 * bio_associate_blkg_from_css - associate a bio with a specified css
1997 * Associate @bio with the blkg found by combining the css's blkg and the
1998 * request_queue of the @bio. An association failure is handled by walking up
1999 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
2000 * and q->root_blkg. This situation only happens when a cgroup is dying and
2001 * then the remaining bios will spill to the closest alive blkg.
2003 * A reference will be taken on the blkg and will be released when @bio is
2006 void bio_associate_blkg_from_css(struct bio *bio,
2007 struct cgroup_subsys_state *css)
2010 blkg_put(bio->bi_blkg);
2012 if (css && css->parent) {
2013 bio->bi_blkg = blkg_tryget_closest(bio, css);
2015 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
2016 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
2019 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2022 * bio_associate_blkg - associate a bio with a blkg
2025 * Associate @bio with the blkg found from the bio's css and request_queue.
2026 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2027 * already associated, the css is reused and association redone as the
2028 * request_queue may have changed.
2030 void bio_associate_blkg(struct bio *bio)
2032 struct cgroup_subsys_state *css;
2037 css = bio_blkcg_css(bio);
2041 bio_associate_blkg_from_css(bio, css);
2045 EXPORT_SYMBOL_GPL(bio_associate_blkg);
2048 * bio_clone_blkg_association - clone blkg association from src to dst bio
2049 * @dst: destination bio
2052 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2055 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2057 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2059 static int blk_cgroup_io_type(struct bio *bio)
2061 if (op_is_discard(bio->bi_opf))
2062 return BLKG_IOSTAT_DISCARD;
2063 if (op_is_write(bio->bi_opf))
2064 return BLKG_IOSTAT_WRITE;
2065 return BLKG_IOSTAT_READ;
2068 void blk_cgroup_bio_start(struct bio *bio)
2070 struct blkcg *blkcg = bio->bi_blkg->blkcg;
2071 int rwd = blk_cgroup_io_type(bio), cpu;
2072 struct blkg_iostat_set *bis;
2073 unsigned long flags;
2075 /* Root-level stats are sourced from system-wide IO stats */
2076 if (!cgroup_parent(blkcg->css.cgroup))
2080 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2081 flags = u64_stats_update_begin_irqsave(&bis->sync);
2084 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2085 * bio and we would have already accounted for the size of the bio.
2087 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2088 bio_set_flag(bio, BIO_CGROUP_ACCT);
2089 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2091 bis->cur.ios[rwd]++;
2094 * If the iostat_cpu isn't in a lockless list, put it into the
2095 * list to indicate that a stat update is pending.
2097 if (!READ_ONCE(bis->lqueued)) {
2098 struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2100 llist_add(&bis->lnode, lhead);
2101 WRITE_ONCE(bis->lqueued, true);
2104 u64_stats_update_end_irqrestore(&bis->sync, flags);
2105 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
2106 cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2110 bool blk_cgroup_congested(void)
2112 struct cgroup_subsys_state *css;
2116 for (css = blkcg_css(); css; css = css->parent) {
2117 if (atomic_read(&css->cgroup->congestion_count)) {
2126 module_param(blkcg_debug_stats, bool, 0644);
2127 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");