1 // SPDX-License-Identifier: GPL-2.0
3 * Common Block IO controller cgroup interface
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
18 #include <linux/ioprio.h>
19 #include <linux/kdev_t.h>
20 #include <linux/module.h>
21 #include <linux/sched/signal.h>
22 #include <linux/err.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/slab.h>
26 #include <linux/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/resume_user_mode.h>
30 #include <linux/psi.h>
31 #include <linux/part_stat.h>
33 #include "blk-cgroup.h"
34 #include "blk-ioprio.h"
35 #include "blk-throttle.h"
36 #include "blk-rq-qos.h"
39 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
40 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
41 * policy [un]register operations including cgroup file additions /
42 * removals. Putting cgroup file registration outside blkcg_pol_mutex
43 * allows grabbing it from cgroup callbacks.
45 static DEFINE_MUTEX(blkcg_pol_register_mutex);
46 static DEFINE_MUTEX(blkcg_pol_mutex);
48 struct blkcg blkcg_root;
49 EXPORT_SYMBOL_GPL(blkcg_root);
51 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
52 EXPORT_SYMBOL_GPL(blkcg_root_css);
54 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
56 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
58 bool blkcg_debug_stats = false;
59 static struct workqueue_struct *blkcg_punt_bio_wq;
61 #define BLKG_DESTROY_BATCH_SIZE 64
64 * blkcg_css - find the current css
66 * Find the css associated with either the kthread or the current task.
67 * This may return a dying css, so it is up to the caller to use tryget logic
68 * to confirm it is alive and well.
70 static struct cgroup_subsys_state *blkcg_css(void)
72 struct cgroup_subsys_state *css;
74 css = kthread_blkcg();
77 return task_css(current, io_cgrp_id);
80 static bool blkcg_policy_enabled(struct request_queue *q,
81 const struct blkcg_policy *pol)
83 return pol && test_bit(pol->plid, q->blkcg_pols);
86 static void blkg_free_workfn(struct work_struct *work)
88 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
92 for (i = 0; i < BLKCG_MAX_POLS; i++)
94 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
97 blk_put_queue(blkg->q);
98 free_percpu(blkg->iostat_cpu);
99 percpu_ref_exit(&blkg->refcnt);
104 * blkg_free - free a blkg
105 * @blkg: blkg to free
107 * Free @blkg which may be partially allocated.
109 static void blkg_free(struct blkcg_gq *blkg)
115 * Both ->pd_free_fn() and request queue's release handler may
116 * sleep, so free us by scheduling one work func
118 INIT_WORK(&blkg->free_work, blkg_free_workfn);
119 schedule_work(&blkg->free_work);
122 static void __blkg_release(struct rcu_head *rcu)
124 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
126 WARN_ON(!bio_list_empty(&blkg->async_bios));
128 /* release the blkcg and parent blkg refs this blkg has been holding */
129 css_put(&blkg->blkcg->css);
131 blkg_put(blkg->parent);
136 * A group is RCU protected, but having an rcu lock does not mean that one
137 * can access all the fields of blkg and assume these are valid. For
138 * example, don't try to follow throtl_data and request queue links.
140 * Having a reference to blkg under an rcu allows accesses to only values
141 * local to groups like group stats and group rate limits.
143 static void blkg_release(struct percpu_ref *ref)
145 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
147 call_rcu(&blkg->rcu_head, __blkg_release);
150 static void blkg_async_bio_workfn(struct work_struct *work)
152 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
154 struct bio_list bios = BIO_EMPTY_LIST;
156 struct blk_plug plug;
157 bool need_plug = false;
159 /* as long as there are pending bios, @blkg can't go away */
160 spin_lock_bh(&blkg->async_bio_lock);
161 bio_list_merge(&bios, &blkg->async_bios);
162 bio_list_init(&blkg->async_bios);
163 spin_unlock_bh(&blkg->async_bio_lock);
165 /* start plug only when bio_list contains at least 2 bios */
166 if (bios.head && bios.head->bi_next) {
168 blk_start_plug(&plug);
170 while ((bio = bio_list_pop(&bios)))
173 blk_finish_plug(&plug);
177 * bio_blkcg_css - return the blkcg CSS associated with a bio
180 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
181 * associated. Callers are expected to either handle %NULL or know association
182 * has been done prior to calling this.
184 struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
186 if (!bio || !bio->bi_blkg)
188 return &bio->bi_blkg->blkcg->css;
190 EXPORT_SYMBOL_GPL(bio_blkcg_css);
193 * blkcg_parent - get the parent of a blkcg
194 * @blkcg: blkcg of interest
196 * Return the parent blkcg of @blkcg. Can be called anytime.
198 static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
200 return css_to_blkcg(blkcg->css.parent);
204 * blkg_alloc - allocate a blkg
205 * @blkcg: block cgroup the new blkg is associated with
206 * @disk: gendisk the new blkg is associated with
207 * @gfp_mask: allocation mask to use
209 * Allocate a new blkg assocating @blkcg and @q.
211 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
214 struct blkcg_gq *blkg;
217 /* alloc and init base part */
218 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
222 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
225 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
226 if (!blkg->iostat_cpu)
229 if (!blk_get_queue(disk->queue))
232 blkg->q = disk->queue;
233 INIT_LIST_HEAD(&blkg->q_node);
234 spin_lock_init(&blkg->async_bio_lock);
235 bio_list_init(&blkg->async_bios);
236 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
239 u64_stats_init(&blkg->iostat.sync);
240 for_each_possible_cpu(cpu)
241 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
243 for (i = 0; i < BLKCG_MAX_POLS; i++) {
244 struct blkcg_policy *pol = blkcg_policy[i];
245 struct blkg_policy_data *pd;
247 if (!blkcg_policy_enabled(disk->queue, pol))
250 /* alloc per-policy data and attach it to blkg */
251 pd = pol->pd_alloc_fn(gfp_mask, disk->queue, blkcg);
268 * If @new_blkg is %NULL, this function tries to allocate a new one as
269 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
271 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
272 struct blkcg_gq *new_blkg)
274 struct blkcg_gq *blkg;
277 lockdep_assert_held(&disk->queue->queue_lock);
279 /* request_queue is dying, do not create/recreate a blkg */
280 if (blk_queue_dying(disk->queue)) {
285 /* blkg holds a reference to blkcg */
286 if (!css_tryget_online(&blkcg->css)) {
293 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
294 if (unlikely(!new_blkg)) {
302 if (blkcg_parent(blkcg)) {
303 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
304 if (WARN_ON_ONCE(!blkg->parent)) {
308 blkg_get(blkg->parent);
311 /* invoke per-policy init */
312 for (i = 0; i < BLKCG_MAX_POLS; i++) {
313 struct blkcg_policy *pol = blkcg_policy[i];
315 if (blkg->pd[i] && pol->pd_init_fn)
316 pol->pd_init_fn(blkg->pd[i]);
320 spin_lock(&blkcg->lock);
321 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
323 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
324 list_add(&blkg->q_node, &disk->queue->blkg_list);
326 for (i = 0; i < BLKCG_MAX_POLS; i++) {
327 struct blkcg_policy *pol = blkcg_policy[i];
329 if (blkg->pd[i] && pol->pd_online_fn)
330 pol->pd_online_fn(blkg->pd[i]);
334 spin_unlock(&blkcg->lock);
339 /* @blkg failed fully initialized, use the usual release path */
344 css_put(&blkcg->css);
351 * blkg_lookup_create - lookup blkg, try to create one if not there
352 * @blkcg: blkcg of interest
353 * @disk: gendisk of interest
355 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
356 * create one. blkg creation is performed recursively from blkcg_root such
357 * that all non-root blkg's have access to the parent blkg. This function
358 * should be called under RCU read lock and takes @disk->queue->queue_lock.
360 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
363 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
364 struct gendisk *disk)
366 struct request_queue *q = disk->queue;
367 struct blkcg_gq *blkg;
370 WARN_ON_ONCE(!rcu_read_lock_held());
372 blkg = blkg_lookup(blkcg, q);
376 spin_lock_irqsave(&q->queue_lock, flags);
377 blkg = blkg_lookup(blkcg, q);
379 if (blkcg != &blkcg_root &&
380 blkg != rcu_dereference(blkcg->blkg_hint))
381 rcu_assign_pointer(blkcg->blkg_hint, blkg);
386 * Create blkgs walking down from blkcg_root to @blkcg, so that all
387 * non-root blkgs have access to their parents. Returns the closest
388 * blkg to the intended blkg should blkg_create() fail.
391 struct blkcg *pos = blkcg;
392 struct blkcg *parent = blkcg_parent(blkcg);
393 struct blkcg_gq *ret_blkg = q->root_blkg;
396 blkg = blkg_lookup(parent, q);
398 /* remember closest blkg */
403 parent = blkcg_parent(parent);
406 blkg = blkg_create(pos, disk, NULL);
416 spin_unlock_irqrestore(&q->queue_lock, flags);
420 static void blkg_destroy(struct blkcg_gq *blkg)
422 struct blkcg *blkcg = blkg->blkcg;
425 lockdep_assert_held(&blkg->q->queue_lock);
426 lockdep_assert_held(&blkcg->lock);
428 /* Something wrong if we are trying to remove same group twice */
429 WARN_ON_ONCE(list_empty(&blkg->q_node));
430 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
432 for (i = 0; i < BLKCG_MAX_POLS; i++) {
433 struct blkcg_policy *pol = blkcg_policy[i];
435 if (blkg->pd[i] && pol->pd_offline_fn)
436 pol->pd_offline_fn(blkg->pd[i]);
439 blkg->online = false;
441 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
442 list_del_init(&blkg->q_node);
443 hlist_del_init_rcu(&blkg->blkcg_node);
446 * Both setting lookup hint to and clearing it from @blkg are done
447 * under queue_lock. If it's not pointing to @blkg now, it never
448 * will. Hint assignment itself can race safely.
450 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
451 rcu_assign_pointer(blkcg->blkg_hint, NULL);
454 * Put the reference taken at the time of creation so that when all
455 * queues are gone, group can be destroyed.
457 percpu_ref_kill(&blkg->refcnt);
460 static void blkg_destroy_all(struct gendisk *disk)
462 struct request_queue *q = disk->queue;
463 struct blkcg_gq *blkg, *n;
464 int count = BLKG_DESTROY_BATCH_SIZE;
467 spin_lock_irq(&q->queue_lock);
468 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
469 struct blkcg *blkcg = blkg->blkcg;
471 spin_lock(&blkcg->lock);
473 spin_unlock(&blkcg->lock);
476 * in order to avoid holding the spin lock for too long, release
477 * it when a batch of blkgs are destroyed.
480 count = BLKG_DESTROY_BATCH_SIZE;
481 spin_unlock_irq(&q->queue_lock);
488 spin_unlock_irq(&q->queue_lock);
491 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
492 struct cftype *cftype, u64 val)
494 struct blkcg *blkcg = css_to_blkcg(css);
495 struct blkcg_gq *blkg;
498 mutex_lock(&blkcg_pol_mutex);
499 spin_lock_irq(&blkcg->lock);
502 * Note that stat reset is racy - it doesn't synchronize against
503 * stat updates. This is a debug feature which shouldn't exist
504 * anyway. If you get hit by a race, retry.
506 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
507 for_each_possible_cpu(cpu) {
508 struct blkg_iostat_set *bis =
509 per_cpu_ptr(blkg->iostat_cpu, cpu);
510 memset(bis, 0, sizeof(*bis));
512 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
514 for (i = 0; i < BLKCG_MAX_POLS; i++) {
515 struct blkcg_policy *pol = blkcg_policy[i];
517 if (blkg->pd[i] && pol->pd_reset_stats_fn)
518 pol->pd_reset_stats_fn(blkg->pd[i]);
522 spin_unlock_irq(&blkcg->lock);
523 mutex_unlock(&blkcg_pol_mutex);
527 const char *blkg_dev_name(struct blkcg_gq *blkg)
529 if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
531 return bdi_dev_name(blkg->q->disk->bdi);
535 * blkcg_print_blkgs - helper for printing per-blkg data
536 * @sf: seq_file to print to
537 * @blkcg: blkcg of interest
538 * @prfill: fill function to print out a blkg
539 * @pol: policy in question
540 * @data: data to be passed to @prfill
541 * @show_total: to print out sum of prfill return values or not
543 * This function invokes @prfill on each blkg of @blkcg if pd for the
544 * policy specified by @pol exists. @prfill is invoked with @sf, the
545 * policy data and @data and the matching queue lock held. If @show_total
546 * is %true, the sum of the return values from @prfill is printed with
547 * "Total" label at the end.
549 * This is to be used to construct print functions for
550 * cftype->read_seq_string method.
552 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
553 u64 (*prfill)(struct seq_file *,
554 struct blkg_policy_data *, int),
555 const struct blkcg_policy *pol, int data,
558 struct blkcg_gq *blkg;
562 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
563 spin_lock_irq(&blkg->q->queue_lock);
564 if (blkcg_policy_enabled(blkg->q, pol))
565 total += prfill(sf, blkg->pd[pol->plid], data);
566 spin_unlock_irq(&blkg->q->queue_lock);
571 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
573 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
576 * __blkg_prfill_u64 - prfill helper for a single u64 value
577 * @sf: seq_file to print to
578 * @pd: policy private data of interest
581 * Print @v to @sf for the device assocaited with @pd.
583 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
585 const char *dname = blkg_dev_name(pd->blkg);
590 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
593 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
596 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
597 * @inputp: input string pointer
599 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
600 * from @input and get and return the matching bdev. *@inputp is
601 * updated to point past the device node prefix. Returns an ERR_PTR()
604 * Use this function iff blkg_conf_prep() can't be used for some reason.
606 struct block_device *blkcg_conf_open_bdev(char **inputp)
608 char *input = *inputp;
609 unsigned int major, minor;
610 struct block_device *bdev;
613 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
614 return ERR_PTR(-EINVAL);
617 if (!isspace(*input))
618 return ERR_PTR(-EINVAL);
619 input = skip_spaces(input);
621 bdev = blkdev_get_no_open(MKDEV(major, minor));
623 return ERR_PTR(-ENODEV);
624 if (bdev_is_partition(bdev)) {
625 blkdev_put_no_open(bdev);
626 return ERR_PTR(-ENODEV);
634 * blkg_conf_prep - parse and prepare for per-blkg config update
635 * @blkcg: target block cgroup
636 * @pol: target policy
637 * @input: input string
638 * @ctx: blkg_conf_ctx to be filled
640 * Parse per-blkg config update from @input and initialize @ctx with the
641 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
642 * part of @input following MAJ:MIN. This function returns with RCU read
643 * lock and queue lock held and must be paired with blkg_conf_finish().
645 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
646 char *input, struct blkg_conf_ctx *ctx)
647 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
649 struct block_device *bdev;
650 struct gendisk *disk;
651 struct request_queue *q;
652 struct blkcg_gq *blkg;
655 bdev = blkcg_conf_open_bdev(&input);
657 return PTR_ERR(bdev);
658 disk = bdev->bd_disk;
662 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
663 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
665 ret = blk_queue_enter(q, 0);
670 spin_lock_irq(&q->queue_lock);
672 if (!blkcg_policy_enabled(q, pol)) {
677 blkg = blkg_lookup(blkcg, q);
682 * Create blkgs walking down from blkcg_root to @blkcg, so that all
683 * non-root blkgs have access to their parents.
686 struct blkcg *pos = blkcg;
687 struct blkcg *parent;
688 struct blkcg_gq *new_blkg;
690 parent = blkcg_parent(blkcg);
691 while (parent && !blkg_lookup(parent, q)) {
693 parent = blkcg_parent(parent);
696 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
697 spin_unlock_irq(&q->queue_lock);
700 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
701 if (unlikely(!new_blkg)) {
703 goto fail_exit_queue;
706 if (radix_tree_preload(GFP_KERNEL)) {
709 goto fail_exit_queue;
713 spin_lock_irq(&q->queue_lock);
715 if (!blkcg_policy_enabled(q, pol)) {
721 blkg = blkg_lookup(pos, q);
725 blkg = blkg_create(pos, disk, new_blkg);
732 radix_tree_preload_end();
745 radix_tree_preload_end();
747 spin_unlock_irq(&q->queue_lock);
752 blkdev_put_no_open(bdev);
754 * If queue was bypassing, we should retry. Do so after a
755 * short msleep(). It isn't strictly necessary but queue
756 * can be bypassing for some time and it's always nice to
757 * avoid busy looping.
761 ret = restart_syscall();
765 EXPORT_SYMBOL_GPL(blkg_conf_prep);
768 * blkg_conf_finish - finish up per-blkg config update
769 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
771 * Finish up after per-blkg config update. This function must be paired
772 * with blkg_conf_prep().
774 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
775 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
777 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
779 blkdev_put_no_open(ctx->bdev);
781 EXPORT_SYMBOL_GPL(blkg_conf_finish);
783 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
787 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
788 dst->bytes[i] = src->bytes[i];
789 dst->ios[i] = src->ios[i];
793 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
797 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
798 dst->bytes[i] += src->bytes[i];
799 dst->ios[i] += src->ios[i];
803 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
807 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
808 dst->bytes[i] -= src->bytes[i];
809 dst->ios[i] -= src->ios[i];
813 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
814 struct blkg_iostat *last)
816 struct blkg_iostat delta;
819 /* propagate percpu delta to global */
820 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
821 blkg_iostat_set(&delta, cur);
822 blkg_iostat_sub(&delta, last);
823 blkg_iostat_add(&blkg->iostat.cur, &delta);
824 blkg_iostat_add(last, &delta);
825 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
828 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
830 struct blkcg *blkcg = css_to_blkcg(css);
831 struct blkcg_gq *blkg;
833 /* Root-level stats are sourced from system-wide IO stats */
834 if (!cgroup_parent(css->cgroup))
839 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
840 struct blkcg_gq *parent = blkg->parent;
841 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
842 struct blkg_iostat cur;
845 /* fetch the current per-cpu values */
847 seq = u64_stats_fetch_begin(&bisc->sync);
848 blkg_iostat_set(&cur, &bisc->cur);
849 } while (u64_stats_fetch_retry(&bisc->sync, seq));
851 blkcg_iostat_update(blkg, &cur, &bisc->last);
853 /* propagate global delta to parent (unless that's root) */
854 if (parent && parent->parent)
855 blkcg_iostat_update(parent, &blkg->iostat.cur,
863 * We source root cgroup stats from the system-wide stats to avoid
864 * tracking the same information twice and incurring overhead when no
865 * cgroups are defined. For that reason, cgroup_rstat_flush in
866 * blkcg_print_stat does not actually fill out the iostat in the root
869 * However, we would like to re-use the printing code between the root and
870 * non-root cgroups to the extent possible. For that reason, we simulate
871 * flushing the root cgroup's stats by explicitly filling in the iostat
872 * with disk level statistics.
874 static void blkcg_fill_root_iostats(void)
876 struct class_dev_iter iter;
879 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
880 while ((dev = class_dev_iter_next(&iter))) {
881 struct block_device *bdev = dev_to_bdev(dev);
882 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
883 struct blkg_iostat tmp;
887 memset(&tmp, 0, sizeof(tmp));
888 for_each_possible_cpu(cpu) {
889 struct disk_stats *cpu_dkstats;
891 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
892 tmp.ios[BLKG_IOSTAT_READ] +=
893 cpu_dkstats->ios[STAT_READ];
894 tmp.ios[BLKG_IOSTAT_WRITE] +=
895 cpu_dkstats->ios[STAT_WRITE];
896 tmp.ios[BLKG_IOSTAT_DISCARD] +=
897 cpu_dkstats->ios[STAT_DISCARD];
898 // convert sectors to bytes
899 tmp.bytes[BLKG_IOSTAT_READ] +=
900 cpu_dkstats->sectors[STAT_READ] << 9;
901 tmp.bytes[BLKG_IOSTAT_WRITE] +=
902 cpu_dkstats->sectors[STAT_WRITE] << 9;
903 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
904 cpu_dkstats->sectors[STAT_DISCARD] << 9;
907 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
908 blkg_iostat_set(&blkg->iostat.cur, &tmp);
909 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
913 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
915 struct blkg_iostat_set *bis = &blkg->iostat;
916 u64 rbytes, wbytes, rios, wios, dbytes, dios;
924 dname = blkg_dev_name(blkg);
928 seq_printf(s, "%s ", dname);
931 seq = u64_stats_fetch_begin(&bis->sync);
933 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
934 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
935 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
936 rios = bis->cur.ios[BLKG_IOSTAT_READ];
937 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
938 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
939 } while (u64_stats_fetch_retry(&bis->sync, seq));
941 if (rbytes || wbytes || rios || wios) {
942 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
943 rbytes, wbytes, rios, wios,
947 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
948 seq_printf(s, " use_delay=%d delay_nsec=%llu",
949 atomic_read(&blkg->use_delay),
950 atomic64_read(&blkg->delay_nsec));
953 for (i = 0; i < BLKCG_MAX_POLS; i++) {
954 struct blkcg_policy *pol = blkcg_policy[i];
956 if (!blkg->pd[i] || !pol->pd_stat_fn)
959 pol->pd_stat_fn(blkg->pd[i], s);
965 static int blkcg_print_stat(struct seq_file *sf, void *v)
967 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
968 struct blkcg_gq *blkg;
970 if (!seq_css(sf)->parent)
971 blkcg_fill_root_iostats();
973 cgroup_rstat_flush(blkcg->css.cgroup);
976 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
977 spin_lock_irq(&blkg->q->queue_lock);
978 blkcg_print_one_stat(blkg, sf);
979 spin_unlock_irq(&blkg->q->queue_lock);
985 static struct cftype blkcg_files[] = {
988 .seq_show = blkcg_print_stat,
993 static struct cftype blkcg_legacy_files[] = {
995 .name = "reset_stats",
996 .write_u64 = blkcg_reset_stats,
1001 #ifdef CONFIG_CGROUP_WRITEBACK
1002 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1004 return &css_to_blkcg(css)->cgwb_list;
1009 * blkcg destruction is a three-stage process.
1011 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1012 * which offlines writeback. Here we tie the next stage of blkg destruction
1013 * to the completion of writeback associated with the blkcg. This lets us
1014 * avoid punting potentially large amounts of outstanding writeback to root
1015 * while maintaining any ongoing policies. The next stage is triggered when
1016 * the nr_cgwbs count goes to zero.
1018 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1019 * and handles the destruction of blkgs. Here the css reference held by
1020 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1021 * This work may occur in cgwb_release_workfn() on the cgwb_release
1022 * workqueue. Any submitted ios that fail to get the blkg ref will be
1023 * punted to the root_blkg.
1025 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1026 * This finally frees the blkcg.
1030 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1031 * @blkcg: blkcg of interest
1033 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1034 * is nested inside q lock, this function performs reverse double lock dancing.
1035 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1036 * blkcg_css_free to eventually be called.
1038 * This is the blkcg counterpart of ioc_release_fn().
1040 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1044 spin_lock_irq(&blkcg->lock);
1046 while (!hlist_empty(&blkcg->blkg_list)) {
1047 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1048 struct blkcg_gq, blkcg_node);
1049 struct request_queue *q = blkg->q;
1051 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1053 * Given that the system can accumulate a huge number
1054 * of blkgs in pathological cases, check to see if we
1055 * need to rescheduling to avoid softlockup.
1057 spin_unlock_irq(&blkcg->lock);
1059 spin_lock_irq(&blkcg->lock);
1064 spin_unlock(&q->queue_lock);
1067 spin_unlock_irq(&blkcg->lock);
1071 * blkcg_pin_online - pin online state
1072 * @blkcg_css: blkcg of interest
1074 * While pinned, a blkcg is kept online. This is primarily used to
1075 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1076 * while an associated cgwb is still active.
1078 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1080 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1084 * blkcg_unpin_online - unpin online state
1085 * @blkcg_css: blkcg of interest
1087 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1088 * that blkg doesn't go offline while an associated cgwb is still active.
1089 * When this count goes to zero, all active cgwbs have finished so the
1090 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1092 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1094 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1097 if (!refcount_dec_and_test(&blkcg->online_pin))
1099 blkcg_destroy_blkgs(blkcg);
1100 blkcg = blkcg_parent(blkcg);
1105 * blkcg_css_offline - cgroup css_offline callback
1106 * @css: css of interest
1108 * This function is called when @css is about to go away. Here the cgwbs are
1109 * offlined first and only once writeback associated with the blkcg has
1110 * finished do we start step 2 (see above).
1112 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1114 /* this prevents anyone from attaching or migrating to this blkcg */
1115 wb_blkcg_offline(css);
1117 /* put the base online pin allowing step 2 to be triggered */
1118 blkcg_unpin_online(css);
1121 static void blkcg_css_free(struct cgroup_subsys_state *css)
1123 struct blkcg *blkcg = css_to_blkcg(css);
1126 mutex_lock(&blkcg_pol_mutex);
1128 list_del(&blkcg->all_blkcgs_node);
1130 for (i = 0; i < BLKCG_MAX_POLS; i++)
1132 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1134 mutex_unlock(&blkcg_pol_mutex);
1139 static struct cgroup_subsys_state *
1140 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1142 struct blkcg *blkcg;
1143 struct cgroup_subsys_state *ret;
1146 mutex_lock(&blkcg_pol_mutex);
1149 blkcg = &blkcg_root;
1151 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1153 ret = ERR_PTR(-ENOMEM);
1158 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1159 struct blkcg_policy *pol = blkcg_policy[i];
1160 struct blkcg_policy_data *cpd;
1163 * If the policy hasn't been attached yet, wait for it
1164 * to be attached before doing anything else. Otherwise,
1165 * check if the policy requires any specific per-cgroup
1166 * data: if it does, allocate and initialize it.
1168 if (!pol || !pol->cpd_alloc_fn)
1171 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1173 ret = ERR_PTR(-ENOMEM);
1176 blkcg->cpd[i] = cpd;
1179 if (pol->cpd_init_fn)
1180 pol->cpd_init_fn(cpd);
1183 spin_lock_init(&blkcg->lock);
1184 refcount_set(&blkcg->online_pin, 1);
1185 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1186 INIT_HLIST_HEAD(&blkcg->blkg_list);
1187 #ifdef CONFIG_CGROUP_WRITEBACK
1188 INIT_LIST_HEAD(&blkcg->cgwb_list);
1190 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1192 mutex_unlock(&blkcg_pol_mutex);
1196 for (i--; i >= 0; i--)
1198 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1200 if (blkcg != &blkcg_root)
1203 mutex_unlock(&blkcg_pol_mutex);
1207 static int blkcg_css_online(struct cgroup_subsys_state *css)
1209 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1212 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1213 * don't go offline while cgwbs are still active on them. Pin the
1214 * parent so that offline always happens towards the root.
1217 blkcg_pin_online(&parent->css);
1221 int blkcg_init_disk(struct gendisk *disk)
1223 struct request_queue *q = disk->queue;
1224 struct blkcg_gq *new_blkg, *blkg;
1228 INIT_LIST_HEAD(&q->blkg_list);
1230 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1234 preloaded = !radix_tree_preload(GFP_KERNEL);
1236 /* Make sure the root blkg exists. */
1237 /* spin_lock_irq can serve as RCU read-side critical section. */
1238 spin_lock_irq(&q->queue_lock);
1239 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1242 q->root_blkg = blkg;
1243 spin_unlock_irq(&q->queue_lock);
1246 radix_tree_preload_end();
1248 ret = blk_ioprio_init(disk);
1250 goto err_destroy_all;
1252 ret = blk_throtl_init(disk);
1254 goto err_ioprio_exit;
1256 ret = blk_iolatency_init(disk);
1258 goto err_throtl_exit;
1263 blk_throtl_exit(disk);
1265 blk_ioprio_exit(disk);
1267 blkg_destroy_all(disk);
1270 spin_unlock_irq(&q->queue_lock);
1272 radix_tree_preload_end();
1273 return PTR_ERR(blkg);
1276 void blkcg_exit_disk(struct gendisk *disk)
1278 blkg_destroy_all(disk);
1279 rq_qos_exit(disk->queue);
1280 blk_throtl_exit(disk);
1283 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1287 mutex_lock(&blkcg_pol_mutex);
1289 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1290 struct blkcg_policy *pol = blkcg_policy[i];
1291 struct blkcg *blkcg;
1293 if (!pol || !pol->cpd_bind_fn)
1296 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1297 if (blkcg->cpd[pol->plid])
1298 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1300 mutex_unlock(&blkcg_pol_mutex);
1303 static void blkcg_exit(struct task_struct *tsk)
1305 if (tsk->throttle_queue)
1306 blk_put_queue(tsk->throttle_queue);
1307 tsk->throttle_queue = NULL;
1310 struct cgroup_subsys io_cgrp_subsys = {
1311 .css_alloc = blkcg_css_alloc,
1312 .css_online = blkcg_css_online,
1313 .css_offline = blkcg_css_offline,
1314 .css_free = blkcg_css_free,
1315 .css_rstat_flush = blkcg_rstat_flush,
1317 .dfl_cftypes = blkcg_files,
1318 .legacy_cftypes = blkcg_legacy_files,
1319 .legacy_name = "blkio",
1323 * This ensures that, if available, memcg is automatically enabled
1324 * together on the default hierarchy so that the owner cgroup can
1325 * be retrieved from writeback pages.
1327 .depends_on = 1 << memory_cgrp_id,
1330 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1333 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1334 * @q: request_queue of interest
1335 * @pol: blkcg policy to activate
1337 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1338 * bypass mode to populate its blkgs with policy_data for @pol.
1340 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1341 * from IO path. Update of each blkg is protected by both queue and blkcg
1342 * locks so that holding either lock and testing blkcg_policy_enabled() is
1343 * always enough for dereferencing policy data.
1345 * The caller is responsible for synchronizing [de]activations and policy
1346 * [un]registerations. Returns 0 on success, -errno on failure.
1348 int blkcg_activate_policy(struct request_queue *q,
1349 const struct blkcg_policy *pol)
1351 struct blkg_policy_data *pd_prealloc = NULL;
1352 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1355 if (blkcg_policy_enabled(q, pol))
1359 blk_mq_freeze_queue(q);
1361 spin_lock_irq(&q->queue_lock);
1363 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1364 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1365 struct blkg_policy_data *pd;
1367 if (blkg->pd[pol->plid])
1370 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1371 if (blkg == pinned_blkg) {
1375 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1381 * GFP_NOWAIT failed. Free the existing one and
1382 * prealloc for @blkg w/ GFP_KERNEL.
1385 blkg_put(pinned_blkg);
1389 spin_unlock_irq(&q->queue_lock);
1392 pol->pd_free_fn(pd_prealloc);
1393 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1401 blkg->pd[pol->plid] = pd;
1403 pd->plid = pol->plid;
1406 /* all allocated, init in the same order */
1407 if (pol->pd_init_fn)
1408 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1409 pol->pd_init_fn(blkg->pd[pol->plid]);
1411 if (pol->pd_online_fn)
1412 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1413 pol->pd_online_fn(blkg->pd[pol->plid]);
1415 __set_bit(pol->plid, q->blkcg_pols);
1418 spin_unlock_irq(&q->queue_lock);
1421 blk_mq_unfreeze_queue(q);
1423 blkg_put(pinned_blkg);
1425 pol->pd_free_fn(pd_prealloc);
1429 /* alloc failed, nothing's initialized yet, free everything */
1430 spin_lock_irq(&q->queue_lock);
1431 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1432 struct blkcg *blkcg = blkg->blkcg;
1434 spin_lock(&blkcg->lock);
1435 if (blkg->pd[pol->plid]) {
1436 pol->pd_free_fn(blkg->pd[pol->plid]);
1437 blkg->pd[pol->plid] = NULL;
1439 spin_unlock(&blkcg->lock);
1441 spin_unlock_irq(&q->queue_lock);
1445 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1448 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1449 * @q: request_queue of interest
1450 * @pol: blkcg policy to deactivate
1452 * Deactivate @pol on @q. Follows the same synchronization rules as
1453 * blkcg_activate_policy().
1455 void blkcg_deactivate_policy(struct request_queue *q,
1456 const struct blkcg_policy *pol)
1458 struct blkcg_gq *blkg;
1460 if (!blkcg_policy_enabled(q, pol))
1464 blk_mq_freeze_queue(q);
1466 spin_lock_irq(&q->queue_lock);
1468 __clear_bit(pol->plid, q->blkcg_pols);
1470 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1471 struct blkcg *blkcg = blkg->blkcg;
1473 spin_lock(&blkcg->lock);
1474 if (blkg->pd[pol->plid]) {
1475 if (pol->pd_offline_fn)
1476 pol->pd_offline_fn(blkg->pd[pol->plid]);
1477 pol->pd_free_fn(blkg->pd[pol->plid]);
1478 blkg->pd[pol->plid] = NULL;
1480 spin_unlock(&blkcg->lock);
1483 spin_unlock_irq(&q->queue_lock);
1486 blk_mq_unfreeze_queue(q);
1488 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1490 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1492 struct blkcg *blkcg;
1494 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1495 if (blkcg->cpd[pol->plid]) {
1496 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1497 blkcg->cpd[pol->plid] = NULL;
1503 * blkcg_policy_register - register a blkcg policy
1504 * @pol: blkcg policy to register
1506 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1507 * successful registration. Returns 0 on success and -errno on failure.
1509 int blkcg_policy_register(struct blkcg_policy *pol)
1511 struct blkcg *blkcg;
1514 mutex_lock(&blkcg_pol_register_mutex);
1515 mutex_lock(&blkcg_pol_mutex);
1517 /* find an empty slot */
1519 for (i = 0; i < BLKCG_MAX_POLS; i++)
1520 if (!blkcg_policy[i])
1522 if (i >= BLKCG_MAX_POLS) {
1523 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1527 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1528 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1529 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1534 blkcg_policy[pol->plid] = pol;
1536 /* allocate and install cpd's */
1537 if (pol->cpd_alloc_fn) {
1538 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1539 struct blkcg_policy_data *cpd;
1541 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1545 blkcg->cpd[pol->plid] = cpd;
1547 cpd->plid = pol->plid;
1548 if (pol->cpd_init_fn)
1549 pol->cpd_init_fn(cpd);
1553 mutex_unlock(&blkcg_pol_mutex);
1555 /* everything is in place, add intf files for the new policy */
1556 if (pol->dfl_cftypes)
1557 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1559 if (pol->legacy_cftypes)
1560 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1561 pol->legacy_cftypes));
1562 mutex_unlock(&blkcg_pol_register_mutex);
1566 if (pol->cpd_free_fn)
1567 blkcg_free_all_cpd(pol);
1569 blkcg_policy[pol->plid] = NULL;
1571 mutex_unlock(&blkcg_pol_mutex);
1572 mutex_unlock(&blkcg_pol_register_mutex);
1575 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1578 * blkcg_policy_unregister - unregister a blkcg policy
1579 * @pol: blkcg policy to unregister
1581 * Undo blkcg_policy_register(@pol). Might sleep.
1583 void blkcg_policy_unregister(struct blkcg_policy *pol)
1585 mutex_lock(&blkcg_pol_register_mutex);
1587 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1590 /* kill the intf files first */
1591 if (pol->dfl_cftypes)
1592 cgroup_rm_cftypes(pol->dfl_cftypes);
1593 if (pol->legacy_cftypes)
1594 cgroup_rm_cftypes(pol->legacy_cftypes);
1596 /* remove cpds and unregister */
1597 mutex_lock(&blkcg_pol_mutex);
1599 if (pol->cpd_free_fn)
1600 blkcg_free_all_cpd(pol);
1602 blkcg_policy[pol->plid] = NULL;
1604 mutex_unlock(&blkcg_pol_mutex);
1606 mutex_unlock(&blkcg_pol_register_mutex);
1608 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1610 bool __blkcg_punt_bio_submit(struct bio *bio)
1612 struct blkcg_gq *blkg = bio->bi_blkg;
1614 /* consume the flag first */
1615 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1617 /* never bounce for the root cgroup */
1621 spin_lock_bh(&blkg->async_bio_lock);
1622 bio_list_add(&blkg->async_bios, bio);
1623 spin_unlock_bh(&blkg->async_bio_lock);
1625 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1630 * Scale the accumulated delay based on how long it has been since we updated
1631 * the delay. We only call this when we are adding delay, in case it's been a
1632 * while since we added delay, and when we are checking to see if we need to
1633 * delay a task, to account for any delays that may have occurred.
1635 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1637 u64 old = atomic64_read(&blkg->delay_start);
1639 /* negative use_delay means no scaling, see blkcg_set_delay() */
1640 if (atomic_read(&blkg->use_delay) < 0)
1644 * We only want to scale down every second. The idea here is that we
1645 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1646 * time window. We only want to throttle tasks for recent delay that
1647 * has occurred, in 1 second time windows since that's the maximum
1648 * things can be throttled. We save the current delay window in
1649 * blkg->last_delay so we know what amount is still left to be charged
1650 * to the blkg from this point onward. blkg->last_use keeps track of
1651 * the use_delay counter. The idea is if we're unthrottling the blkg we
1652 * are ok with whatever is happening now, and we can take away more of
1653 * the accumulated delay as we've already throttled enough that
1654 * everybody is happy with their IO latencies.
1656 if (time_before64(old + NSEC_PER_SEC, now) &&
1657 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1658 u64 cur = atomic64_read(&blkg->delay_nsec);
1659 u64 sub = min_t(u64, blkg->last_delay, now - old);
1660 int cur_use = atomic_read(&blkg->use_delay);
1663 * We've been unthrottled, subtract a larger chunk of our
1664 * accumulated delay.
1666 if (cur_use < blkg->last_use)
1667 sub = max_t(u64, sub, blkg->last_delay >> 1);
1670 * This shouldn't happen, but handle it anyway. Our delay_nsec
1671 * should only ever be growing except here where we subtract out
1672 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1673 * rather not end up with negative numbers.
1675 if (unlikely(cur < sub)) {
1676 atomic64_set(&blkg->delay_nsec, 0);
1677 blkg->last_delay = 0;
1679 atomic64_sub(sub, &blkg->delay_nsec);
1680 blkg->last_delay = cur - sub;
1682 blkg->last_use = cur_use;
1687 * This is called when we want to actually walk up the hierarchy and check to
1688 * see if we need to throttle, and then actually throttle if there is some
1689 * accumulated delay. This should only be called upon return to user space so
1690 * we're not holding some lock that would induce a priority inversion.
1692 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1694 unsigned long pflags;
1696 u64 now = ktime_to_ns(ktime_get());
1701 while (blkg->parent) {
1702 int use_delay = atomic_read(&blkg->use_delay);
1707 blkcg_scale_delay(blkg, now);
1708 this_delay = atomic64_read(&blkg->delay_nsec);
1709 if (this_delay > delay_nsec) {
1710 delay_nsec = this_delay;
1711 clamp = use_delay > 0;
1714 blkg = blkg->parent;
1721 * Let's not sleep for all eternity if we've amassed a huge delay.
1722 * Swapping or metadata IO can accumulate 10's of seconds worth of
1723 * delay, and we want userspace to be able to do _something_ so cap the
1724 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1725 * tasks will be delayed for 0.25 second for every syscall. If
1726 * blkcg_set_delay() was used as indicated by negative use_delay, the
1727 * caller is responsible for regulating the range.
1730 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1733 psi_memstall_enter(&pflags);
1735 exp = ktime_add_ns(now, delay_nsec);
1736 tok = io_schedule_prepare();
1738 __set_current_state(TASK_KILLABLE);
1739 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1741 } while (!fatal_signal_pending(current));
1742 io_schedule_finish(tok);
1745 psi_memstall_leave(&pflags);
1749 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1751 * This is only called if we've been marked with set_notify_resume(). Obviously
1752 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1753 * check to see if current->throttle_queue is set and if not this doesn't do
1754 * anything. This should only ever be called by the resume code, it's not meant
1755 * to be called by people willy-nilly as it will actually do the work to
1756 * throttle the task if it is setup for throttling.
1758 void blkcg_maybe_throttle_current(void)
1760 struct request_queue *q = current->throttle_queue;
1761 struct blkcg *blkcg;
1762 struct blkcg_gq *blkg;
1763 bool use_memdelay = current->use_memdelay;
1768 current->throttle_queue = NULL;
1769 current->use_memdelay = false;
1772 blkcg = css_to_blkcg(blkcg_css());
1775 blkg = blkg_lookup(blkcg, q);
1778 if (!blkg_tryget(blkg))
1782 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1792 * blkcg_schedule_throttle - this task needs to check for throttling
1793 * @gendisk: disk to throttle
1794 * @use_memdelay: do we charge this to memory delay for PSI
1796 * This is called by the IO controller when we know there's delay accumulated
1797 * for the blkg for this task. We do not pass the blkg because there are places
1798 * we call this that may not have that information, the swapping code for
1799 * instance will only have a block_device at that point. This set's the
1800 * notify_resume for the task to check and see if it requires throttling before
1801 * returning to user space.
1803 * We will only schedule once per syscall. You can call this over and over
1804 * again and it will only do the check once upon return to user space, and only
1805 * throttle once. If the task needs to be throttled again it'll need to be
1806 * re-set at the next time we see the task.
1808 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1810 struct request_queue *q = disk->queue;
1812 if (unlikely(current->flags & PF_KTHREAD))
1815 if (current->throttle_queue != q) {
1816 if (!blk_get_queue(q))
1819 if (current->throttle_queue)
1820 blk_put_queue(current->throttle_queue);
1821 current->throttle_queue = q;
1825 current->use_memdelay = use_memdelay;
1826 set_notify_resume(current);
1830 * blkcg_add_delay - add delay to this blkg
1831 * @blkg: blkg of interest
1832 * @now: the current time in nanoseconds
1833 * @delta: how many nanoseconds of delay to add
1835 * Charge @delta to the blkg's current delay accumulation. This is used to
1836 * throttle tasks if an IO controller thinks we need more throttling.
1838 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1840 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1842 blkcg_scale_delay(blkg, now);
1843 atomic64_add(delta, &blkg->delay_nsec);
1847 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1851 * As the failure mode here is to walk up the blkg tree, this ensure that the
1852 * blkg->parent pointers are always valid. This returns the blkg that it ended
1853 * up taking a reference on or %NULL if no reference was taken.
1855 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1856 struct cgroup_subsys_state *css)
1858 struct blkcg_gq *blkg, *ret_blkg = NULL;
1861 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1863 if (blkg_tryget(blkg)) {
1867 blkg = blkg->parent;
1875 * bio_associate_blkg_from_css - associate a bio with a specified css
1879 * Associate @bio with the blkg found by combining the css's blkg and the
1880 * request_queue of the @bio. An association failure is handled by walking up
1881 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1882 * and q->root_blkg. This situation only happens when a cgroup is dying and
1883 * then the remaining bios will spill to the closest alive blkg.
1885 * A reference will be taken on the blkg and will be released when @bio is
1888 void bio_associate_blkg_from_css(struct bio *bio,
1889 struct cgroup_subsys_state *css)
1892 blkg_put(bio->bi_blkg);
1894 if (css && css->parent) {
1895 bio->bi_blkg = blkg_tryget_closest(bio, css);
1897 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1898 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1901 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1904 * bio_associate_blkg - associate a bio with a blkg
1907 * Associate @bio with the blkg found from the bio's css and request_queue.
1908 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1909 * already associated, the css is reused and association redone as the
1910 * request_queue may have changed.
1912 void bio_associate_blkg(struct bio *bio)
1914 struct cgroup_subsys_state *css;
1919 css = bio_blkcg_css(bio);
1923 bio_associate_blkg_from_css(bio, css);
1927 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1930 * bio_clone_blkg_association - clone blkg association from src to dst bio
1931 * @dst: destination bio
1934 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1937 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
1939 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1941 static int blk_cgroup_io_type(struct bio *bio)
1943 if (op_is_discard(bio->bi_opf))
1944 return BLKG_IOSTAT_DISCARD;
1945 if (op_is_write(bio->bi_opf))
1946 return BLKG_IOSTAT_WRITE;
1947 return BLKG_IOSTAT_READ;
1950 void blk_cgroup_bio_start(struct bio *bio)
1952 int rwd = blk_cgroup_io_type(bio), cpu;
1953 struct blkg_iostat_set *bis;
1954 unsigned long flags;
1957 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1958 flags = u64_stats_update_begin_irqsave(&bis->sync);
1961 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1962 * bio and we would have already accounted for the size of the bio.
1964 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1965 bio_set_flag(bio, BIO_CGROUP_ACCT);
1966 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1968 bis->cur.ios[rwd]++;
1970 u64_stats_update_end_irqrestore(&bis->sync, flags);
1971 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1972 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1976 bool blk_cgroup_congested(void)
1978 struct cgroup_subsys_state *css;
1982 for (css = blkcg_css(); css; css = css->parent) {
1983 if (atomic_read(&css->cgroup->congestion_count)) {
1992 static int __init blkcg_init(void)
1994 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1995 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1996 WQ_UNBOUND | WQ_SYSFS, 0);
1997 if (!blkcg_punt_bio_wq)
2001 subsys_initcall(blkcg_init);
2003 module_param(blkcg_debug_stats, bool, 0644);
2004 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");