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,
122 for (i = 0; i < BLKCG_MAX_POLS; i++)
124 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
127 blk_put_queue(blkg->q);
128 free_percpu(blkg->iostat_cpu);
129 percpu_ref_exit(&blkg->refcnt);
134 * blkg_free - free a blkg
135 * @blkg: blkg to free
137 * Free @blkg which may be partially allocated.
139 static void blkg_free(struct blkcg_gq *blkg)
145 * Both ->pd_free_fn() and request queue's release handler may
146 * sleep, so free us by scheduling one work func
148 INIT_WORK(&blkg->free_work, blkg_free_workfn);
149 schedule_work(&blkg->free_work);
152 static void __blkg_release(struct rcu_head *rcu)
154 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
156 WARN_ON(!bio_list_empty(&blkg->async_bios));
158 /* release the blkcg and parent blkg refs this blkg has been holding */
159 css_put(&blkg->blkcg->css);
161 blkg_put(blkg->parent);
166 * A group is RCU protected, but having an rcu lock does not mean that one
167 * can access all the fields of blkg and assume these are valid. For
168 * example, don't try to follow throtl_data and request queue links.
170 * Having a reference to blkg under an rcu allows accesses to only values
171 * local to groups like group stats and group rate limits.
173 static void blkg_release(struct percpu_ref *ref)
175 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
177 call_rcu(&blkg->rcu_head, __blkg_release);
180 static void blkg_async_bio_workfn(struct work_struct *work)
182 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
184 struct bio_list bios = BIO_EMPTY_LIST;
186 struct blk_plug plug;
187 bool need_plug = false;
189 /* as long as there are pending bios, @blkg can't go away */
190 spin_lock_bh(&blkg->async_bio_lock);
191 bio_list_merge(&bios, &blkg->async_bios);
192 bio_list_init(&blkg->async_bios);
193 spin_unlock_bh(&blkg->async_bio_lock);
195 /* start plug only when bio_list contains at least 2 bios */
196 if (bios.head && bios.head->bi_next) {
198 blk_start_plug(&plug);
200 while ((bio = bio_list_pop(&bios)))
203 blk_finish_plug(&plug);
207 * bio_blkcg_css - return the blkcg CSS associated with a bio
210 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
211 * associated. Callers are expected to either handle %NULL or know association
212 * has been done prior to calling this.
214 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
216 if (!bio || !bio->bi_blkg)
218 return &bio->bi_blkg->blkcg->css;
220 EXPORT_SYMBOL_GPL(bio_blkcg_css);
223 * blkcg_parent - get the parent of a blkcg
224 * @blkcg: blkcg of interest
226 * Return the parent blkcg of @blkcg. Can be called anytime.
228 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
230 return css_to_blkcg(blkcg->css.parent);
234 * blkg_alloc - allocate a blkg
235 * @blkcg: block cgroup the new blkg is associated with
236 * @disk: gendisk the new blkg is associated with
237 * @gfp_mask: allocation mask to use
239 * Allocate a new blkg assocating @blkcg and @q.
241 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
244 struct blkcg_gq *blkg;
247 /* alloc and init base part */
248 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
252 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
255 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
256 if (!blkg->iostat_cpu)
259 if (!blk_get_queue(disk->queue))
262 blkg->q = disk->queue;
263 INIT_LIST_HEAD(&blkg->q_node);
264 spin_lock_init(&blkg->async_bio_lock);
265 bio_list_init(&blkg->async_bios);
266 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
269 u64_stats_init(&blkg->iostat.sync);
270 for_each_possible_cpu(cpu) {
271 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
272 per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
275 for (i = 0; i < BLKCG_MAX_POLS; i++) {
276 struct blkcg_policy *pol = blkcg_policy[i];
277 struct blkg_policy_data *pd;
279 if (!blkcg_policy_enabled(disk->queue, pol))
282 /* alloc per-policy data and attach it to blkg */
283 pd = pol->pd_alloc_fn(gfp_mask, disk->queue, blkcg);
300 * If @new_blkg is %NULL, this function tries to allocate a new one as
301 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
303 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
304 struct blkcg_gq *new_blkg)
306 struct blkcg_gq *blkg;
309 lockdep_assert_held(&disk->queue->queue_lock);
311 /* request_queue is dying, do not create/recreate a blkg */
312 if (blk_queue_dying(disk->queue)) {
317 /* blkg holds a reference to blkcg */
318 if (!css_tryget_online(&blkcg->css)) {
325 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
326 if (unlikely(!new_blkg)) {
334 if (blkcg_parent(blkcg)) {
335 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
336 if (WARN_ON_ONCE(!blkg->parent)) {
340 blkg_get(blkg->parent);
343 /* invoke per-policy init */
344 for (i = 0; i < BLKCG_MAX_POLS; i++) {
345 struct blkcg_policy *pol = blkcg_policy[i];
347 if (blkg->pd[i] && pol->pd_init_fn)
348 pol->pd_init_fn(blkg->pd[i]);
352 spin_lock(&blkcg->lock);
353 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
355 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
356 list_add(&blkg->q_node, &disk->queue->blkg_list);
358 for (i = 0; i < BLKCG_MAX_POLS; i++) {
359 struct blkcg_policy *pol = blkcg_policy[i];
361 if (blkg->pd[i] && pol->pd_online_fn)
362 pol->pd_online_fn(blkg->pd[i]);
366 spin_unlock(&blkcg->lock);
371 /* @blkg failed fully initialized, use the usual release path */
376 css_put(&blkcg->css);
383 * blkg_lookup_create - lookup blkg, try to create one if not there
384 * @blkcg: blkcg of interest
385 * @disk: gendisk of interest
387 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
388 * create one. blkg creation is performed recursively from blkcg_root such
389 * that all non-root blkg's have access to the parent blkg. This function
390 * should be called under RCU read lock and takes @disk->queue->queue_lock.
392 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
395 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
396 struct gendisk *disk)
398 struct request_queue *q = disk->queue;
399 struct blkcg_gq *blkg;
402 WARN_ON_ONCE(!rcu_read_lock_held());
404 blkg = blkg_lookup(blkcg, q);
408 spin_lock_irqsave(&q->queue_lock, flags);
409 blkg = blkg_lookup(blkcg, q);
411 if (blkcg != &blkcg_root &&
412 blkg != rcu_dereference(blkcg->blkg_hint))
413 rcu_assign_pointer(blkcg->blkg_hint, blkg);
418 * Create blkgs walking down from blkcg_root to @blkcg, so that all
419 * non-root blkgs have access to their parents. Returns the closest
420 * blkg to the intended blkg should blkg_create() fail.
423 struct blkcg *pos = blkcg;
424 struct blkcg *parent = blkcg_parent(blkcg);
425 struct blkcg_gq *ret_blkg = q->root_blkg;
428 blkg = blkg_lookup(parent, q);
430 /* remember closest blkg */
435 parent = blkcg_parent(parent);
438 blkg = blkg_create(pos, disk, NULL);
448 spin_unlock_irqrestore(&q->queue_lock, flags);
452 static void blkg_destroy(struct blkcg_gq *blkg)
454 struct blkcg *blkcg = blkg->blkcg;
457 lockdep_assert_held(&blkg->q->queue_lock);
458 lockdep_assert_held(&blkcg->lock);
460 /* Something wrong if we are trying to remove same group twice */
461 WARN_ON_ONCE(list_empty(&blkg->q_node));
462 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
464 for (i = 0; i < BLKCG_MAX_POLS; i++) {
465 struct blkcg_policy *pol = blkcg_policy[i];
467 if (blkg->pd[i] && pol->pd_offline_fn)
468 pol->pd_offline_fn(blkg->pd[i]);
471 blkg->online = false;
473 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
474 list_del_init(&blkg->q_node);
475 hlist_del_init_rcu(&blkg->blkcg_node);
478 * Both setting lookup hint to and clearing it from @blkg are done
479 * under queue_lock. If it's not pointing to @blkg now, it never
480 * will. Hint assignment itself can race safely.
482 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
483 rcu_assign_pointer(blkcg->blkg_hint, NULL);
486 * Put the reference taken at the time of creation so that when all
487 * queues are gone, group can be destroyed.
489 percpu_ref_kill(&blkg->refcnt);
492 static void blkg_destroy_all(struct gendisk *disk)
494 struct request_queue *q = disk->queue;
495 struct blkcg_gq *blkg, *n;
496 int count = BLKG_DESTROY_BATCH_SIZE;
499 spin_lock_irq(&q->queue_lock);
500 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
501 struct blkcg *blkcg = blkg->blkcg;
503 spin_lock(&blkcg->lock);
505 spin_unlock(&blkcg->lock);
508 * in order to avoid holding the spin lock for too long, release
509 * it when a batch of blkgs are destroyed.
512 count = BLKG_DESTROY_BATCH_SIZE;
513 spin_unlock_irq(&q->queue_lock);
520 spin_unlock_irq(&q->queue_lock);
523 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
524 struct cftype *cftype, u64 val)
526 struct blkcg *blkcg = css_to_blkcg(css);
527 struct blkcg_gq *blkg;
530 mutex_lock(&blkcg_pol_mutex);
531 spin_lock_irq(&blkcg->lock);
534 * Note that stat reset is racy - it doesn't synchronize against
535 * stat updates. This is a debug feature which shouldn't exist
536 * anyway. If you get hit by a race, retry.
538 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
539 for_each_possible_cpu(cpu) {
540 struct blkg_iostat_set *bis =
541 per_cpu_ptr(blkg->iostat_cpu, cpu);
542 memset(bis, 0, sizeof(*bis));
544 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
546 for (i = 0; i < BLKCG_MAX_POLS; i++) {
547 struct blkcg_policy *pol = blkcg_policy[i];
549 if (blkg->pd[i] && pol->pd_reset_stats_fn)
550 pol->pd_reset_stats_fn(blkg->pd[i]);
554 spin_unlock_irq(&blkcg->lock);
555 mutex_unlock(&blkcg_pol_mutex);
559 const char *blkg_dev_name(struct blkcg_gq *blkg)
561 if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
563 return bdi_dev_name(blkg->q->disk->bdi);
567 * blkcg_print_blkgs - helper for printing per-blkg data
568 * @sf: seq_file to print to
569 * @blkcg: blkcg of interest
570 * @prfill: fill function to print out a blkg
571 * @pol: policy in question
572 * @data: data to be passed to @prfill
573 * @show_total: to print out sum of prfill return values or not
575 * This function invokes @prfill on each blkg of @blkcg if pd for the
576 * policy specified by @pol exists. @prfill is invoked with @sf, the
577 * policy data and @data and the matching queue lock held. If @show_total
578 * is %true, the sum of the return values from @prfill is printed with
579 * "Total" label at the end.
581 * This is to be used to construct print functions for
582 * cftype->read_seq_string method.
584 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
585 u64 (*prfill)(struct seq_file *,
586 struct blkg_policy_data *, int),
587 const struct blkcg_policy *pol, int data,
590 struct blkcg_gq *blkg;
594 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
595 spin_lock_irq(&blkg->q->queue_lock);
596 if (blkcg_policy_enabled(blkg->q, pol))
597 total += prfill(sf, blkg->pd[pol->plid], data);
598 spin_unlock_irq(&blkg->q->queue_lock);
603 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
605 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
608 * __blkg_prfill_u64 - prfill helper for a single u64 value
609 * @sf: seq_file to print to
610 * @pd: policy private data of interest
613 * Print @v to @sf for the device associated with @pd.
615 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
617 const char *dname = blkg_dev_name(pd->blkg);
622 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
625 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
628 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
629 * @inputp: input string pointer
631 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
632 * from @input and get and return the matching bdev. *@inputp is
633 * updated to point past the device node prefix. Returns an ERR_PTR()
636 * Use this function iff blkg_conf_prep() can't be used for some reason.
638 struct block_device *blkcg_conf_open_bdev(char **inputp)
640 char *input = *inputp;
641 unsigned int major, minor;
642 struct block_device *bdev;
645 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
646 return ERR_PTR(-EINVAL);
649 if (!isspace(*input))
650 return ERR_PTR(-EINVAL);
651 input = skip_spaces(input);
653 bdev = blkdev_get_no_open(MKDEV(major, minor));
655 return ERR_PTR(-ENODEV);
656 if (bdev_is_partition(bdev)) {
657 blkdev_put_no_open(bdev);
658 return ERR_PTR(-ENODEV);
666 * blkg_conf_prep - parse and prepare for per-blkg config update
667 * @blkcg: target block cgroup
668 * @pol: target policy
669 * @input: input string
670 * @ctx: blkg_conf_ctx to be filled
672 * Parse per-blkg config update from @input and initialize @ctx with the
673 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
674 * part of @input following MAJ:MIN. This function returns with RCU read
675 * lock and queue lock held and must be paired with blkg_conf_finish().
677 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
678 char *input, struct blkg_conf_ctx *ctx)
679 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
681 struct block_device *bdev;
682 struct gendisk *disk;
683 struct request_queue *q;
684 struct blkcg_gq *blkg;
687 bdev = blkcg_conf_open_bdev(&input);
689 return PTR_ERR(bdev);
690 disk = bdev->bd_disk;
694 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
695 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
697 ret = blk_queue_enter(q, 0);
702 spin_lock_irq(&q->queue_lock);
704 if (!blkcg_policy_enabled(q, pol)) {
709 blkg = blkg_lookup(blkcg, q);
714 * Create blkgs walking down from blkcg_root to @blkcg, so that all
715 * non-root blkgs have access to their parents.
718 struct blkcg *pos = blkcg;
719 struct blkcg *parent;
720 struct blkcg_gq *new_blkg;
722 parent = blkcg_parent(blkcg);
723 while (parent && !blkg_lookup(parent, q)) {
725 parent = blkcg_parent(parent);
728 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
729 spin_unlock_irq(&q->queue_lock);
732 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
733 if (unlikely(!new_blkg)) {
735 goto fail_exit_queue;
738 if (radix_tree_preload(GFP_KERNEL)) {
741 goto fail_exit_queue;
745 spin_lock_irq(&q->queue_lock);
747 if (!blkcg_policy_enabled(q, pol)) {
753 blkg = blkg_lookup(pos, q);
757 blkg = blkg_create(pos, disk, new_blkg);
764 radix_tree_preload_end();
777 radix_tree_preload_end();
779 spin_unlock_irq(&q->queue_lock);
784 blkdev_put_no_open(bdev);
786 * If queue was bypassing, we should retry. Do so after a
787 * short msleep(). It isn't strictly necessary but queue
788 * can be bypassing for some time and it's always nice to
789 * avoid busy looping.
793 ret = restart_syscall();
797 EXPORT_SYMBOL_GPL(blkg_conf_prep);
800 * blkg_conf_finish - finish up per-blkg config update
801 * @ctx: blkg_conf_ctx initialized by blkg_conf_prep()
803 * Finish up after per-blkg config update. This function must be paired
804 * with blkg_conf_prep().
806 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
807 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
809 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
811 blkdev_put_no_open(ctx->bdev);
813 EXPORT_SYMBOL_GPL(blkg_conf_finish);
815 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
819 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
820 dst->bytes[i] = src->bytes[i];
821 dst->ios[i] = src->ios[i];
825 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
829 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
830 dst->bytes[i] += src->bytes[i];
831 dst->ios[i] += src->ios[i];
835 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
839 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
840 dst->bytes[i] -= src->bytes[i];
841 dst->ios[i] -= src->ios[i];
845 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
846 struct blkg_iostat *last)
848 struct blkg_iostat delta;
851 /* propagate percpu delta to global */
852 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
853 blkg_iostat_set(&delta, cur);
854 blkg_iostat_sub(&delta, last);
855 blkg_iostat_add(&blkg->iostat.cur, &delta);
856 blkg_iostat_add(last, &delta);
857 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
860 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
862 struct blkcg *blkcg = css_to_blkcg(css);
863 struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
864 struct llist_node *lnode;
865 struct blkg_iostat_set *bisc, *next_bisc;
867 /* Root-level stats are sourced from system-wide IO stats */
868 if (!cgroup_parent(css->cgroup))
873 lnode = llist_del_all(lhead);
878 * Iterate only the iostat_cpu's queued in the lockless list.
880 llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
881 struct blkcg_gq *blkg = bisc->blkg;
882 struct blkcg_gq *parent = blkg->parent;
883 struct blkg_iostat cur;
886 WRITE_ONCE(bisc->lqueued, false);
888 /* fetch the current per-cpu values */
890 seq = u64_stats_fetch_begin(&bisc->sync);
891 blkg_iostat_set(&cur, &bisc->cur);
892 } while (u64_stats_fetch_retry(&bisc->sync, seq));
894 blkcg_iostat_update(blkg, &cur, &bisc->last);
896 /* propagate global delta to parent (unless that's root) */
897 if (parent && parent->parent)
898 blkcg_iostat_update(parent, &blkg->iostat.cur,
900 percpu_ref_put(&blkg->refcnt);
908 * We source root cgroup stats from the system-wide stats to avoid
909 * tracking the same information twice and incurring overhead when no
910 * cgroups are defined. For that reason, cgroup_rstat_flush in
911 * blkcg_print_stat does not actually fill out the iostat in the root
914 * However, we would like to re-use the printing code between the root and
915 * non-root cgroups to the extent possible. For that reason, we simulate
916 * flushing the root cgroup's stats by explicitly filling in the iostat
917 * with disk level statistics.
919 static void blkcg_fill_root_iostats(void)
921 struct class_dev_iter iter;
924 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
925 while ((dev = class_dev_iter_next(&iter))) {
926 struct block_device *bdev = dev_to_bdev(dev);
927 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
928 struct blkg_iostat tmp;
932 memset(&tmp, 0, sizeof(tmp));
933 for_each_possible_cpu(cpu) {
934 struct disk_stats *cpu_dkstats;
936 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
937 tmp.ios[BLKG_IOSTAT_READ] +=
938 cpu_dkstats->ios[STAT_READ];
939 tmp.ios[BLKG_IOSTAT_WRITE] +=
940 cpu_dkstats->ios[STAT_WRITE];
941 tmp.ios[BLKG_IOSTAT_DISCARD] +=
942 cpu_dkstats->ios[STAT_DISCARD];
943 // convert sectors to bytes
944 tmp.bytes[BLKG_IOSTAT_READ] +=
945 cpu_dkstats->sectors[STAT_READ] << 9;
946 tmp.bytes[BLKG_IOSTAT_WRITE] +=
947 cpu_dkstats->sectors[STAT_WRITE] << 9;
948 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
949 cpu_dkstats->sectors[STAT_DISCARD] << 9;
952 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
953 blkg_iostat_set(&blkg->iostat.cur, &tmp);
954 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
958 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
960 struct blkg_iostat_set *bis = &blkg->iostat;
961 u64 rbytes, wbytes, rios, wios, dbytes, dios;
969 dname = blkg_dev_name(blkg);
973 seq_printf(s, "%s ", dname);
976 seq = u64_stats_fetch_begin(&bis->sync);
978 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
979 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
980 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
981 rios = bis->cur.ios[BLKG_IOSTAT_READ];
982 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
983 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
984 } while (u64_stats_fetch_retry(&bis->sync, seq));
986 if (rbytes || wbytes || rios || wios) {
987 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
988 rbytes, wbytes, rios, wios,
992 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
993 seq_printf(s, " use_delay=%d delay_nsec=%llu",
994 atomic_read(&blkg->use_delay),
995 atomic64_read(&blkg->delay_nsec));
998 for (i = 0; i < BLKCG_MAX_POLS; i++) {
999 struct blkcg_policy *pol = blkcg_policy[i];
1001 if (!blkg->pd[i] || !pol->pd_stat_fn)
1004 pol->pd_stat_fn(blkg->pd[i], s);
1010 static int blkcg_print_stat(struct seq_file *sf, void *v)
1012 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1013 struct blkcg_gq *blkg;
1015 if (!seq_css(sf)->parent)
1016 blkcg_fill_root_iostats();
1018 cgroup_rstat_flush(blkcg->css.cgroup);
1021 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1022 spin_lock_irq(&blkg->q->queue_lock);
1023 blkcg_print_one_stat(blkg, sf);
1024 spin_unlock_irq(&blkg->q->queue_lock);
1030 static struct cftype blkcg_files[] = {
1033 .seq_show = blkcg_print_stat,
1038 static struct cftype blkcg_legacy_files[] = {
1040 .name = "reset_stats",
1041 .write_u64 = blkcg_reset_stats,
1046 #ifdef CONFIG_CGROUP_WRITEBACK
1047 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1049 return &css_to_blkcg(css)->cgwb_list;
1054 * blkcg destruction is a three-stage process.
1056 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1057 * which offlines writeback. Here we tie the next stage of blkg destruction
1058 * to the completion of writeback associated with the blkcg. This lets us
1059 * avoid punting potentially large amounts of outstanding writeback to root
1060 * while maintaining any ongoing policies. The next stage is triggered when
1061 * the nr_cgwbs count goes to zero.
1063 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1064 * and handles the destruction of blkgs. Here the css reference held by
1065 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1066 * This work may occur in cgwb_release_workfn() on the cgwb_release
1067 * workqueue. Any submitted ios that fail to get the blkg ref will be
1068 * punted to the root_blkg.
1070 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1071 * This finally frees the blkcg.
1075 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1076 * @blkcg: blkcg of interest
1078 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1079 * is nested inside q lock, this function performs reverse double lock dancing.
1080 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1081 * blkcg_css_free to eventually be called.
1083 * This is the blkcg counterpart of ioc_release_fn().
1085 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1089 spin_lock_irq(&blkcg->lock);
1091 while (!hlist_empty(&blkcg->blkg_list)) {
1092 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1093 struct blkcg_gq, blkcg_node);
1094 struct request_queue *q = blkg->q;
1096 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1098 * Given that the system can accumulate a huge number
1099 * of blkgs in pathological cases, check to see if we
1100 * need to rescheduling to avoid softlockup.
1102 spin_unlock_irq(&blkcg->lock);
1104 spin_lock_irq(&blkcg->lock);
1109 spin_unlock(&q->queue_lock);
1112 spin_unlock_irq(&blkcg->lock);
1116 * blkcg_pin_online - pin online state
1117 * @blkcg_css: blkcg of interest
1119 * While pinned, a blkcg is kept online. This is primarily used to
1120 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1121 * while an associated cgwb is still active.
1123 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1125 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1129 * blkcg_unpin_online - unpin online state
1130 * @blkcg_css: blkcg of interest
1132 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1133 * that blkg doesn't go offline while an associated cgwb is still active.
1134 * When this count goes to zero, all active cgwbs have finished so the
1135 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1137 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1139 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1142 if (!refcount_dec_and_test(&blkcg->online_pin))
1144 blkcg_destroy_blkgs(blkcg);
1145 blkcg = blkcg_parent(blkcg);
1150 * blkcg_css_offline - cgroup css_offline callback
1151 * @css: css of interest
1153 * This function is called when @css is about to go away. Here the cgwbs are
1154 * offlined first and only once writeback associated with the blkcg has
1155 * finished do we start step 2 (see above).
1157 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1159 /* this prevents anyone from attaching or migrating to this blkcg */
1160 wb_blkcg_offline(css);
1162 /* put the base online pin allowing step 2 to be triggered */
1163 blkcg_unpin_online(css);
1166 static void blkcg_css_free(struct cgroup_subsys_state *css)
1168 struct blkcg *blkcg = css_to_blkcg(css);
1171 mutex_lock(&blkcg_pol_mutex);
1173 list_del(&blkcg->all_blkcgs_node);
1175 for (i = 0; i < BLKCG_MAX_POLS; i++)
1177 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1179 mutex_unlock(&blkcg_pol_mutex);
1181 free_percpu(blkcg->lhead);
1185 static struct cgroup_subsys_state *
1186 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1188 struct blkcg *blkcg;
1191 mutex_lock(&blkcg_pol_mutex);
1194 blkcg = &blkcg_root;
1196 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1201 if (init_blkcg_llists(blkcg))
1204 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1205 struct blkcg_policy *pol = blkcg_policy[i];
1206 struct blkcg_policy_data *cpd;
1209 * If the policy hasn't been attached yet, wait for it
1210 * to be attached before doing anything else. Otherwise,
1211 * check if the policy requires any specific per-cgroup
1212 * data: if it does, allocate and initialize it.
1214 if (!pol || !pol->cpd_alloc_fn)
1217 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1221 blkcg->cpd[i] = cpd;
1224 if (pol->cpd_init_fn)
1225 pol->cpd_init_fn(cpd);
1228 spin_lock_init(&blkcg->lock);
1229 refcount_set(&blkcg->online_pin, 1);
1230 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1231 INIT_HLIST_HEAD(&blkcg->blkg_list);
1232 #ifdef CONFIG_CGROUP_WRITEBACK
1233 INIT_LIST_HEAD(&blkcg->cgwb_list);
1235 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1237 mutex_unlock(&blkcg_pol_mutex);
1241 for (i--; i >= 0; i--)
1243 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1244 free_percpu(blkcg->lhead);
1246 if (blkcg != &blkcg_root)
1249 mutex_unlock(&blkcg_pol_mutex);
1250 return ERR_PTR(-ENOMEM);
1253 static int blkcg_css_online(struct cgroup_subsys_state *css)
1255 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1258 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1259 * don't go offline while cgwbs are still active on them. Pin the
1260 * parent so that offline always happens towards the root.
1263 blkcg_pin_online(&parent->css);
1267 int blkcg_init_disk(struct gendisk *disk)
1269 struct request_queue *q = disk->queue;
1270 struct blkcg_gq *new_blkg, *blkg;
1274 INIT_LIST_HEAD(&q->blkg_list);
1276 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1280 preloaded = !radix_tree_preload(GFP_KERNEL);
1282 /* Make sure the root blkg exists. */
1283 /* spin_lock_irq can serve as RCU read-side critical section. */
1284 spin_lock_irq(&q->queue_lock);
1285 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1288 q->root_blkg = blkg;
1289 spin_unlock_irq(&q->queue_lock);
1292 radix_tree_preload_end();
1294 ret = blk_ioprio_init(disk);
1296 goto err_destroy_all;
1298 ret = blk_throtl_init(disk);
1300 goto err_ioprio_exit;
1302 ret = blk_iolatency_init(disk);
1304 goto err_throtl_exit;
1309 blk_throtl_exit(disk);
1311 blk_ioprio_exit(disk);
1313 blkg_destroy_all(disk);
1316 spin_unlock_irq(&q->queue_lock);
1318 radix_tree_preload_end();
1319 return PTR_ERR(blkg);
1322 void blkcg_exit_disk(struct gendisk *disk)
1324 blkg_destroy_all(disk);
1325 blk_throtl_exit(disk);
1328 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1332 mutex_lock(&blkcg_pol_mutex);
1334 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1335 struct blkcg_policy *pol = blkcg_policy[i];
1336 struct blkcg *blkcg;
1338 if (!pol || !pol->cpd_bind_fn)
1341 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1342 if (blkcg->cpd[pol->plid])
1343 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1345 mutex_unlock(&blkcg_pol_mutex);
1348 static void blkcg_exit(struct task_struct *tsk)
1350 if (tsk->throttle_queue)
1351 blk_put_queue(tsk->throttle_queue);
1352 tsk->throttle_queue = NULL;
1355 struct cgroup_subsys io_cgrp_subsys = {
1356 .css_alloc = blkcg_css_alloc,
1357 .css_online = blkcg_css_online,
1358 .css_offline = blkcg_css_offline,
1359 .css_free = blkcg_css_free,
1360 .css_rstat_flush = blkcg_rstat_flush,
1362 .dfl_cftypes = blkcg_files,
1363 .legacy_cftypes = blkcg_legacy_files,
1364 .legacy_name = "blkio",
1368 * This ensures that, if available, memcg is automatically enabled
1369 * together on the default hierarchy so that the owner cgroup can
1370 * be retrieved from writeback pages.
1372 .depends_on = 1 << memory_cgrp_id,
1375 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1378 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1379 * @q: request_queue of interest
1380 * @pol: blkcg policy to activate
1382 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1383 * bypass mode to populate its blkgs with policy_data for @pol.
1385 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1386 * from IO path. Update of each blkg is protected by both queue and blkcg
1387 * locks so that holding either lock and testing blkcg_policy_enabled() is
1388 * always enough for dereferencing policy data.
1390 * The caller is responsible for synchronizing [de]activations and policy
1391 * [un]registerations. Returns 0 on success, -errno on failure.
1393 int blkcg_activate_policy(struct request_queue *q,
1394 const struct blkcg_policy *pol)
1396 struct blkg_policy_data *pd_prealloc = NULL;
1397 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1400 if (blkcg_policy_enabled(q, pol))
1404 blk_mq_freeze_queue(q);
1406 spin_lock_irq(&q->queue_lock);
1408 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1409 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1410 struct blkg_policy_data *pd;
1412 if (blkg->pd[pol->plid])
1415 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1416 if (blkg == pinned_blkg) {
1420 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1426 * GFP_NOWAIT failed. Free the existing one and
1427 * prealloc for @blkg w/ GFP_KERNEL.
1430 blkg_put(pinned_blkg);
1434 spin_unlock_irq(&q->queue_lock);
1437 pol->pd_free_fn(pd_prealloc);
1438 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1446 blkg->pd[pol->plid] = pd;
1448 pd->plid = pol->plid;
1451 /* all allocated, init in the same order */
1452 if (pol->pd_init_fn)
1453 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1454 pol->pd_init_fn(blkg->pd[pol->plid]);
1456 __set_bit(pol->plid, q->blkcg_pols);
1459 spin_unlock_irq(&q->queue_lock);
1462 blk_mq_unfreeze_queue(q);
1464 blkg_put(pinned_blkg);
1466 pol->pd_free_fn(pd_prealloc);
1470 /* alloc failed, nothing's initialized yet, free everything */
1471 spin_lock_irq(&q->queue_lock);
1472 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1473 struct blkcg *blkcg = blkg->blkcg;
1475 spin_lock(&blkcg->lock);
1476 if (blkg->pd[pol->plid]) {
1477 pol->pd_free_fn(blkg->pd[pol->plid]);
1478 blkg->pd[pol->plid] = NULL;
1480 spin_unlock(&blkcg->lock);
1482 spin_unlock_irq(&q->queue_lock);
1486 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1489 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1490 * @q: request_queue of interest
1491 * @pol: blkcg policy to deactivate
1493 * Deactivate @pol on @q. Follows the same synchronization rules as
1494 * blkcg_activate_policy().
1496 void blkcg_deactivate_policy(struct request_queue *q,
1497 const struct blkcg_policy *pol)
1499 struct blkcg_gq *blkg;
1501 if (!blkcg_policy_enabled(q, pol))
1505 blk_mq_freeze_queue(q);
1507 spin_lock_irq(&q->queue_lock);
1509 __clear_bit(pol->plid, q->blkcg_pols);
1511 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1512 struct blkcg *blkcg = blkg->blkcg;
1514 spin_lock(&blkcg->lock);
1515 if (blkg->pd[pol->plid]) {
1516 if (pol->pd_offline_fn)
1517 pol->pd_offline_fn(blkg->pd[pol->plid]);
1518 pol->pd_free_fn(blkg->pd[pol->plid]);
1519 blkg->pd[pol->plid] = NULL;
1521 spin_unlock(&blkcg->lock);
1524 spin_unlock_irq(&q->queue_lock);
1527 blk_mq_unfreeze_queue(q);
1529 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1531 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1533 struct blkcg *blkcg;
1535 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1536 if (blkcg->cpd[pol->plid]) {
1537 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1538 blkcg->cpd[pol->plid] = NULL;
1544 * blkcg_policy_register - register a blkcg policy
1545 * @pol: blkcg policy to register
1547 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1548 * successful registration. Returns 0 on success and -errno on failure.
1550 int blkcg_policy_register(struct blkcg_policy *pol)
1552 struct blkcg *blkcg;
1555 mutex_lock(&blkcg_pol_register_mutex);
1556 mutex_lock(&blkcg_pol_mutex);
1558 /* find an empty slot */
1560 for (i = 0; i < BLKCG_MAX_POLS; i++)
1561 if (!blkcg_policy[i])
1563 if (i >= BLKCG_MAX_POLS) {
1564 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1568 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1569 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1570 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1575 blkcg_policy[pol->plid] = pol;
1577 /* allocate and install cpd's */
1578 if (pol->cpd_alloc_fn) {
1579 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1580 struct blkcg_policy_data *cpd;
1582 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1586 blkcg->cpd[pol->plid] = cpd;
1588 cpd->plid = pol->plid;
1589 if (pol->cpd_init_fn)
1590 pol->cpd_init_fn(cpd);
1594 mutex_unlock(&blkcg_pol_mutex);
1596 /* everything is in place, add intf files for the new policy */
1597 if (pol->dfl_cftypes)
1598 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1600 if (pol->legacy_cftypes)
1601 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1602 pol->legacy_cftypes));
1603 mutex_unlock(&blkcg_pol_register_mutex);
1607 if (pol->cpd_free_fn)
1608 blkcg_free_all_cpd(pol);
1610 blkcg_policy[pol->plid] = NULL;
1612 mutex_unlock(&blkcg_pol_mutex);
1613 mutex_unlock(&blkcg_pol_register_mutex);
1616 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1619 * blkcg_policy_unregister - unregister a blkcg policy
1620 * @pol: blkcg policy to unregister
1622 * Undo blkcg_policy_register(@pol). Might sleep.
1624 void blkcg_policy_unregister(struct blkcg_policy *pol)
1626 mutex_lock(&blkcg_pol_register_mutex);
1628 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1631 /* kill the intf files first */
1632 if (pol->dfl_cftypes)
1633 cgroup_rm_cftypes(pol->dfl_cftypes);
1634 if (pol->legacy_cftypes)
1635 cgroup_rm_cftypes(pol->legacy_cftypes);
1637 /* remove cpds and unregister */
1638 mutex_lock(&blkcg_pol_mutex);
1640 if (pol->cpd_free_fn)
1641 blkcg_free_all_cpd(pol);
1643 blkcg_policy[pol->plid] = NULL;
1645 mutex_unlock(&blkcg_pol_mutex);
1647 mutex_unlock(&blkcg_pol_register_mutex);
1649 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1651 bool __blkcg_punt_bio_submit(struct bio *bio)
1653 struct blkcg_gq *blkg = bio->bi_blkg;
1655 /* consume the flag first */
1656 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1658 /* never bounce for the root cgroup */
1662 spin_lock_bh(&blkg->async_bio_lock);
1663 bio_list_add(&blkg->async_bios, bio);
1664 spin_unlock_bh(&blkg->async_bio_lock);
1666 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1671 * Scale the accumulated delay based on how long it has been since we updated
1672 * the delay. We only call this when we are adding delay, in case it's been a
1673 * while since we added delay, and when we are checking to see if we need to
1674 * delay a task, to account for any delays that may have occurred.
1676 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1678 u64 old = atomic64_read(&blkg->delay_start);
1680 /* negative use_delay means no scaling, see blkcg_set_delay() */
1681 if (atomic_read(&blkg->use_delay) < 0)
1685 * We only want to scale down every second. The idea here is that we
1686 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1687 * time window. We only want to throttle tasks for recent delay that
1688 * has occurred, in 1 second time windows since that's the maximum
1689 * things can be throttled. We save the current delay window in
1690 * blkg->last_delay so we know what amount is still left to be charged
1691 * to the blkg from this point onward. blkg->last_use keeps track of
1692 * the use_delay counter. The idea is if we're unthrottling the blkg we
1693 * are ok with whatever is happening now, and we can take away more of
1694 * the accumulated delay as we've already throttled enough that
1695 * everybody is happy with their IO latencies.
1697 if (time_before64(old + NSEC_PER_SEC, now) &&
1698 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1699 u64 cur = atomic64_read(&blkg->delay_nsec);
1700 u64 sub = min_t(u64, blkg->last_delay, now - old);
1701 int cur_use = atomic_read(&blkg->use_delay);
1704 * We've been unthrottled, subtract a larger chunk of our
1705 * accumulated delay.
1707 if (cur_use < blkg->last_use)
1708 sub = max_t(u64, sub, blkg->last_delay >> 1);
1711 * This shouldn't happen, but handle it anyway. Our delay_nsec
1712 * should only ever be growing except here where we subtract out
1713 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1714 * rather not end up with negative numbers.
1716 if (unlikely(cur < sub)) {
1717 atomic64_set(&blkg->delay_nsec, 0);
1718 blkg->last_delay = 0;
1720 atomic64_sub(sub, &blkg->delay_nsec);
1721 blkg->last_delay = cur - sub;
1723 blkg->last_use = cur_use;
1728 * This is called when we want to actually walk up the hierarchy and check to
1729 * see if we need to throttle, and then actually throttle if there is some
1730 * accumulated delay. This should only be called upon return to user space so
1731 * we're not holding some lock that would induce a priority inversion.
1733 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1735 unsigned long pflags;
1737 u64 now = ktime_to_ns(ktime_get());
1742 while (blkg->parent) {
1743 int use_delay = atomic_read(&blkg->use_delay);
1748 blkcg_scale_delay(blkg, now);
1749 this_delay = atomic64_read(&blkg->delay_nsec);
1750 if (this_delay > delay_nsec) {
1751 delay_nsec = this_delay;
1752 clamp = use_delay > 0;
1755 blkg = blkg->parent;
1762 * Let's not sleep for all eternity if we've amassed a huge delay.
1763 * Swapping or metadata IO can accumulate 10's of seconds worth of
1764 * delay, and we want userspace to be able to do _something_ so cap the
1765 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1766 * tasks will be delayed for 0.25 second for every syscall. If
1767 * blkcg_set_delay() was used as indicated by negative use_delay, the
1768 * caller is responsible for regulating the range.
1771 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1774 psi_memstall_enter(&pflags);
1776 exp = ktime_add_ns(now, delay_nsec);
1777 tok = io_schedule_prepare();
1779 __set_current_state(TASK_KILLABLE);
1780 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1782 } while (!fatal_signal_pending(current));
1783 io_schedule_finish(tok);
1786 psi_memstall_leave(&pflags);
1790 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1792 * This is only called if we've been marked with set_notify_resume(). Obviously
1793 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1794 * check to see if current->throttle_queue is set and if not this doesn't do
1795 * anything. This should only ever be called by the resume code, it's not meant
1796 * to be called by people willy-nilly as it will actually do the work to
1797 * throttle the task if it is setup for throttling.
1799 void blkcg_maybe_throttle_current(void)
1801 struct request_queue *q = current->throttle_queue;
1802 struct blkcg *blkcg;
1803 struct blkcg_gq *blkg;
1804 bool use_memdelay = current->use_memdelay;
1809 current->throttle_queue = NULL;
1810 current->use_memdelay = false;
1813 blkcg = css_to_blkcg(blkcg_css());
1816 blkg = blkg_lookup(blkcg, q);
1819 if (!blkg_tryget(blkg))
1823 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1833 * blkcg_schedule_throttle - this task needs to check for throttling
1834 * @disk: disk to throttle
1835 * @use_memdelay: do we charge this to memory delay for PSI
1837 * This is called by the IO controller when we know there's delay accumulated
1838 * for the blkg for this task. We do not pass the blkg because there are places
1839 * we call this that may not have that information, the swapping code for
1840 * instance will only have a block_device at that point. This set's the
1841 * notify_resume for the task to check and see if it requires throttling before
1842 * returning to user space.
1844 * We will only schedule once per syscall. You can call this over and over
1845 * again and it will only do the check once upon return to user space, and only
1846 * throttle once. If the task needs to be throttled again it'll need to be
1847 * re-set at the next time we see the task.
1849 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1851 struct request_queue *q = disk->queue;
1853 if (unlikely(current->flags & PF_KTHREAD))
1856 if (current->throttle_queue != q) {
1857 if (!blk_get_queue(q))
1860 if (current->throttle_queue)
1861 blk_put_queue(current->throttle_queue);
1862 current->throttle_queue = q;
1866 current->use_memdelay = use_memdelay;
1867 set_notify_resume(current);
1871 * blkcg_add_delay - add delay to this blkg
1872 * @blkg: blkg of interest
1873 * @now: the current time in nanoseconds
1874 * @delta: how many nanoseconds of delay to add
1876 * Charge @delta to the blkg's current delay accumulation. This is used to
1877 * throttle tasks if an IO controller thinks we need more throttling.
1879 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1881 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1883 blkcg_scale_delay(blkg, now);
1884 atomic64_add(delta, &blkg->delay_nsec);
1888 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1892 * As the failure mode here is to walk up the blkg tree, this ensure that the
1893 * blkg->parent pointers are always valid. This returns the blkg that it ended
1894 * up taking a reference on or %NULL if no reference was taken.
1896 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1897 struct cgroup_subsys_state *css)
1899 struct blkcg_gq *blkg, *ret_blkg = NULL;
1902 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1904 if (blkg_tryget(blkg)) {
1908 blkg = blkg->parent;
1916 * bio_associate_blkg_from_css - associate a bio with a specified css
1920 * Associate @bio with the blkg found by combining the css's blkg and the
1921 * request_queue of the @bio. An association failure is handled by walking up
1922 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1923 * and q->root_blkg. This situation only happens when a cgroup is dying and
1924 * then the remaining bios will spill to the closest alive blkg.
1926 * A reference will be taken on the blkg and will be released when @bio is
1929 void bio_associate_blkg_from_css(struct bio *bio,
1930 struct cgroup_subsys_state *css)
1933 blkg_put(bio->bi_blkg);
1935 if (css && css->parent) {
1936 bio->bi_blkg = blkg_tryget_closest(bio, css);
1938 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1939 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1942 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1945 * bio_associate_blkg - associate a bio with a blkg
1948 * Associate @bio with the blkg found from the bio's css and request_queue.
1949 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1950 * already associated, the css is reused and association redone as the
1951 * request_queue may have changed.
1953 void bio_associate_blkg(struct bio *bio)
1955 struct cgroup_subsys_state *css;
1960 css = bio_blkcg_css(bio);
1964 bio_associate_blkg_from_css(bio, css);
1968 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1971 * bio_clone_blkg_association - clone blkg association from src to dst bio
1972 * @dst: destination bio
1975 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1978 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
1980 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1982 static int blk_cgroup_io_type(struct bio *bio)
1984 if (op_is_discard(bio->bi_opf))
1985 return BLKG_IOSTAT_DISCARD;
1986 if (op_is_write(bio->bi_opf))
1987 return BLKG_IOSTAT_WRITE;
1988 return BLKG_IOSTAT_READ;
1991 void blk_cgroup_bio_start(struct bio *bio)
1993 struct blkcg *blkcg = bio->bi_blkg->blkcg;
1994 int rwd = blk_cgroup_io_type(bio), cpu;
1995 struct blkg_iostat_set *bis;
1996 unsigned long flags;
1999 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2000 flags = u64_stats_update_begin_irqsave(&bis->sync);
2003 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2004 * bio and we would have already accounted for the size of the bio.
2006 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2007 bio_set_flag(bio, BIO_CGROUP_ACCT);
2008 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2010 bis->cur.ios[rwd]++;
2013 * If the iostat_cpu isn't in a lockless list, put it into the
2014 * list to indicate that a stat update is pending.
2016 if (!READ_ONCE(bis->lqueued)) {
2017 struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2019 llist_add(&bis->lnode, lhead);
2020 WRITE_ONCE(bis->lqueued, true);
2021 percpu_ref_get(&bis->blkg->refcnt);
2024 u64_stats_update_end_irqrestore(&bis->sync, flags);
2025 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
2026 cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2030 bool blk_cgroup_congested(void)
2032 struct cgroup_subsys_state *css;
2036 for (css = blkcg_css(); css; css = css->parent) {
2037 if (atomic_read(&css->cgroup->congestion_count)) {
2046 static int __init blkcg_init(void)
2048 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
2049 WQ_MEM_RECLAIM | WQ_FREEZABLE |
2050 WQ_UNBOUND | WQ_SYSFS, 0);
2051 if (!blkcg_punt_bio_wq)
2055 subsys_initcall(blkcg_init);
2057 module_param(blkcg_debug_stats, bool, 0644);
2058 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");