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;
468 spin_lock_irq(&q->queue_lock);
469 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
470 struct blkcg *blkcg = blkg->blkcg;
472 if (hlist_unhashed(&blkg->blkcg_node))
475 spin_lock(&blkcg->lock);
477 spin_unlock(&blkcg->lock);
480 * in order to avoid holding the spin lock for too long, release
481 * it when a batch of blkgs are destroyed.
484 count = BLKG_DESTROY_BATCH_SIZE;
485 spin_unlock_irq(&q->queue_lock);
492 * Mark policy deactivated since policy offline has been done, and
493 * the free is scheduled, so future blkcg_deactivate_policy() can
496 for (i = 0; i < BLKCG_MAX_POLS; i++) {
497 struct blkcg_policy *pol = blkcg_policy[i];
500 __clear_bit(pol->plid, q->blkcg_pols);
504 spin_unlock_irq(&q->queue_lock);
507 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
508 struct cftype *cftype, u64 val)
510 struct blkcg *blkcg = css_to_blkcg(css);
511 struct blkcg_gq *blkg;
514 mutex_lock(&blkcg_pol_mutex);
515 spin_lock_irq(&blkcg->lock);
518 * Note that stat reset is racy - it doesn't synchronize against
519 * stat updates. This is a debug feature which shouldn't exist
520 * anyway. If you get hit by a race, retry.
522 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
523 for_each_possible_cpu(cpu) {
524 struct blkg_iostat_set *bis =
525 per_cpu_ptr(blkg->iostat_cpu, cpu);
526 memset(bis, 0, sizeof(*bis));
528 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
530 for (i = 0; i < BLKCG_MAX_POLS; i++) {
531 struct blkcg_policy *pol = blkcg_policy[i];
533 if (blkg->pd[i] && pol->pd_reset_stats_fn)
534 pol->pd_reset_stats_fn(blkg->pd[i]);
538 spin_unlock_irq(&blkcg->lock);
539 mutex_unlock(&blkcg_pol_mutex);
543 const char *blkg_dev_name(struct blkcg_gq *blkg)
545 if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
547 return bdi_dev_name(blkg->q->disk->bdi);
551 * blkcg_print_blkgs - helper for printing per-blkg data
552 * @sf: seq_file to print to
553 * @blkcg: blkcg of interest
554 * @prfill: fill function to print out a blkg
555 * @pol: policy in question
556 * @data: data to be passed to @prfill
557 * @show_total: to print out sum of prfill return values or not
559 * This function invokes @prfill on each blkg of @blkcg if pd for the
560 * policy specified by @pol exists. @prfill is invoked with @sf, the
561 * policy data and @data and the matching queue lock held. If @show_total
562 * is %true, the sum of the return values from @prfill is printed with
563 * "Total" label at the end.
565 * This is to be used to construct print functions for
566 * cftype->read_seq_string method.
568 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
569 u64 (*prfill)(struct seq_file *,
570 struct blkg_policy_data *, int),
571 const struct blkcg_policy *pol, int data,
574 struct blkcg_gq *blkg;
578 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
579 spin_lock_irq(&blkg->q->queue_lock);
580 if (blkcg_policy_enabled(blkg->q, pol))
581 total += prfill(sf, blkg->pd[pol->plid], data);
582 spin_unlock_irq(&blkg->q->queue_lock);
587 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
589 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
592 * __blkg_prfill_u64 - prfill helper for a single u64 value
593 * @sf: seq_file to print to
594 * @pd: policy private data of interest
597 * Print @v to @sf for the device assocaited with @pd.
599 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
601 const char *dname = blkg_dev_name(pd->blkg);
606 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
609 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
612 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
613 * @inputp: input string pointer
615 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
616 * from @input and get and return the matching bdev. *@inputp is
617 * updated to point past the device node prefix. Returns an ERR_PTR()
620 * Use this function iff blkg_conf_prep() can't be used for some reason.
622 struct block_device *blkcg_conf_open_bdev(char **inputp)
624 char *input = *inputp;
625 unsigned int major, minor;
626 struct block_device *bdev;
629 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
630 return ERR_PTR(-EINVAL);
633 if (!isspace(*input))
634 return ERR_PTR(-EINVAL);
635 input = skip_spaces(input);
637 bdev = blkdev_get_no_open(MKDEV(major, minor));
639 return ERR_PTR(-ENODEV);
640 if (bdev_is_partition(bdev)) {
641 blkdev_put_no_open(bdev);
642 return ERR_PTR(-ENODEV);
650 * blkg_conf_prep - parse and prepare for per-blkg config update
651 * @blkcg: target block cgroup
652 * @pol: target policy
653 * @input: input string
654 * @ctx: blkg_conf_ctx to be filled
656 * Parse per-blkg config update from @input and initialize @ctx with the
657 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
658 * part of @input following MAJ:MIN. This function returns with RCU read
659 * lock and queue lock held and must be paired with blkg_conf_finish().
661 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
662 char *input, struct blkg_conf_ctx *ctx)
663 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
665 struct block_device *bdev;
666 struct gendisk *disk;
667 struct request_queue *q;
668 struct blkcg_gq *blkg;
671 bdev = blkcg_conf_open_bdev(&input);
673 return PTR_ERR(bdev);
674 disk = bdev->bd_disk;
678 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
679 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
681 ret = blk_queue_enter(q, 0);
686 spin_lock_irq(&q->queue_lock);
688 if (!blkcg_policy_enabled(q, pol)) {
693 blkg = blkg_lookup(blkcg, q);
698 * Create blkgs walking down from blkcg_root to @blkcg, so that all
699 * non-root blkgs have access to their parents.
702 struct blkcg *pos = blkcg;
703 struct blkcg *parent;
704 struct blkcg_gq *new_blkg;
706 parent = blkcg_parent(blkcg);
707 while (parent && !blkg_lookup(parent, q)) {
709 parent = blkcg_parent(parent);
712 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
713 spin_unlock_irq(&q->queue_lock);
716 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
717 if (unlikely(!new_blkg)) {
719 goto fail_exit_queue;
722 if (radix_tree_preload(GFP_KERNEL)) {
725 goto fail_exit_queue;
729 spin_lock_irq(&q->queue_lock);
731 if (!blkcg_policy_enabled(q, pol)) {
737 blkg = blkg_lookup(pos, q);
741 blkg = blkg_create(pos, disk, new_blkg);
748 radix_tree_preload_end();
761 radix_tree_preload_end();
763 spin_unlock_irq(&q->queue_lock);
768 blkdev_put_no_open(bdev);
770 * If queue was bypassing, we should retry. Do so after a
771 * short msleep(). It isn't strictly necessary but queue
772 * can be bypassing for some time and it's always nice to
773 * avoid busy looping.
777 ret = restart_syscall();
781 EXPORT_SYMBOL_GPL(blkg_conf_prep);
784 * blkg_conf_finish - finish up per-blkg config update
785 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
787 * Finish up after per-blkg config update. This function must be paired
788 * with blkg_conf_prep().
790 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
791 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
793 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
795 blkdev_put_no_open(ctx->bdev);
797 EXPORT_SYMBOL_GPL(blkg_conf_finish);
799 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
803 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
804 dst->bytes[i] = src->bytes[i];
805 dst->ios[i] = src->ios[i];
809 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
813 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
814 dst->bytes[i] += src->bytes[i];
815 dst->ios[i] += src->ios[i];
819 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
823 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
824 dst->bytes[i] -= src->bytes[i];
825 dst->ios[i] -= src->ios[i];
829 static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
830 struct blkg_iostat *last)
832 struct blkg_iostat delta;
835 /* propagate percpu delta to global */
836 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
837 blkg_iostat_set(&delta, cur);
838 blkg_iostat_sub(&delta, last);
839 blkg_iostat_add(&blkg->iostat.cur, &delta);
840 blkg_iostat_add(last, &delta);
841 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
844 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
846 struct blkcg *blkcg = css_to_blkcg(css);
847 struct blkcg_gq *blkg;
849 /* Root-level stats are sourced from system-wide IO stats */
850 if (!cgroup_parent(css->cgroup))
855 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
856 struct blkcg_gq *parent = blkg->parent;
857 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
858 struct blkg_iostat cur;
861 /* fetch the current per-cpu values */
863 seq = u64_stats_fetch_begin(&bisc->sync);
864 blkg_iostat_set(&cur, &bisc->cur);
865 } while (u64_stats_fetch_retry(&bisc->sync, seq));
867 blkcg_iostat_update(blkg, &cur, &bisc->last);
869 /* propagate global delta to parent (unless that's root) */
870 if (parent && parent->parent)
871 blkcg_iostat_update(parent, &blkg->iostat.cur,
879 * We source root cgroup stats from the system-wide stats to avoid
880 * tracking the same information twice and incurring overhead when no
881 * cgroups are defined. For that reason, cgroup_rstat_flush in
882 * blkcg_print_stat does not actually fill out the iostat in the root
885 * However, we would like to re-use the printing code between the root and
886 * non-root cgroups to the extent possible. For that reason, we simulate
887 * flushing the root cgroup's stats by explicitly filling in the iostat
888 * with disk level statistics.
890 static void blkcg_fill_root_iostats(void)
892 struct class_dev_iter iter;
895 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
896 while ((dev = class_dev_iter_next(&iter))) {
897 struct block_device *bdev = dev_to_bdev(dev);
898 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
899 struct blkg_iostat tmp;
903 memset(&tmp, 0, sizeof(tmp));
904 for_each_possible_cpu(cpu) {
905 struct disk_stats *cpu_dkstats;
907 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
908 tmp.ios[BLKG_IOSTAT_READ] +=
909 cpu_dkstats->ios[STAT_READ];
910 tmp.ios[BLKG_IOSTAT_WRITE] +=
911 cpu_dkstats->ios[STAT_WRITE];
912 tmp.ios[BLKG_IOSTAT_DISCARD] +=
913 cpu_dkstats->ios[STAT_DISCARD];
914 // convert sectors to bytes
915 tmp.bytes[BLKG_IOSTAT_READ] +=
916 cpu_dkstats->sectors[STAT_READ] << 9;
917 tmp.bytes[BLKG_IOSTAT_WRITE] +=
918 cpu_dkstats->sectors[STAT_WRITE] << 9;
919 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
920 cpu_dkstats->sectors[STAT_DISCARD] << 9;
923 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
924 blkg_iostat_set(&blkg->iostat.cur, &tmp);
925 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
929 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
931 struct blkg_iostat_set *bis = &blkg->iostat;
932 u64 rbytes, wbytes, rios, wios, dbytes, dios;
940 dname = blkg_dev_name(blkg);
944 seq_printf(s, "%s ", dname);
947 seq = u64_stats_fetch_begin(&bis->sync);
949 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
950 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
951 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
952 rios = bis->cur.ios[BLKG_IOSTAT_READ];
953 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
954 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
955 } while (u64_stats_fetch_retry(&bis->sync, seq));
957 if (rbytes || wbytes || rios || wios) {
958 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
959 rbytes, wbytes, rios, wios,
963 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
964 seq_printf(s, " use_delay=%d delay_nsec=%llu",
965 atomic_read(&blkg->use_delay),
966 atomic64_read(&blkg->delay_nsec));
969 for (i = 0; i < BLKCG_MAX_POLS; i++) {
970 struct blkcg_policy *pol = blkcg_policy[i];
972 if (!blkg->pd[i] || !pol->pd_stat_fn)
975 pol->pd_stat_fn(blkg->pd[i], s);
981 static int blkcg_print_stat(struct seq_file *sf, void *v)
983 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
984 struct blkcg_gq *blkg;
986 if (!seq_css(sf)->parent)
987 blkcg_fill_root_iostats();
989 cgroup_rstat_flush(blkcg->css.cgroup);
992 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
993 spin_lock_irq(&blkg->q->queue_lock);
994 blkcg_print_one_stat(blkg, sf);
995 spin_unlock_irq(&blkg->q->queue_lock);
1001 static struct cftype blkcg_files[] = {
1004 .seq_show = blkcg_print_stat,
1009 static struct cftype blkcg_legacy_files[] = {
1011 .name = "reset_stats",
1012 .write_u64 = blkcg_reset_stats,
1017 #ifdef CONFIG_CGROUP_WRITEBACK
1018 struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1020 return &css_to_blkcg(css)->cgwb_list;
1025 * blkcg destruction is a three-stage process.
1027 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1028 * which offlines writeback. Here we tie the next stage of blkg destruction
1029 * to the completion of writeback associated with the blkcg. This lets us
1030 * avoid punting potentially large amounts of outstanding writeback to root
1031 * while maintaining any ongoing policies. The next stage is triggered when
1032 * the nr_cgwbs count goes to zero.
1034 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1035 * and handles the destruction of blkgs. Here the css reference held by
1036 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1037 * This work may occur in cgwb_release_workfn() on the cgwb_release
1038 * workqueue. Any submitted ios that fail to get the blkg ref will be
1039 * punted to the root_blkg.
1041 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1042 * This finally frees the blkcg.
1046 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1047 * @blkcg: blkcg of interest
1049 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1050 * is nested inside q lock, this function performs reverse double lock dancing.
1051 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1052 * blkcg_css_free to eventually be called.
1054 * This is the blkcg counterpart of ioc_release_fn().
1056 static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1060 spin_lock_irq(&blkcg->lock);
1062 while (!hlist_empty(&blkcg->blkg_list)) {
1063 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1064 struct blkcg_gq, blkcg_node);
1065 struct request_queue *q = blkg->q;
1067 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1069 * Given that the system can accumulate a huge number
1070 * of blkgs in pathological cases, check to see if we
1071 * need to rescheduling to avoid softlockup.
1073 spin_unlock_irq(&blkcg->lock);
1075 spin_lock_irq(&blkcg->lock);
1080 spin_unlock(&q->queue_lock);
1083 spin_unlock_irq(&blkcg->lock);
1087 * blkcg_pin_online - pin online state
1088 * @blkcg_css: blkcg of interest
1090 * While pinned, a blkcg is kept online. This is primarily used to
1091 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1092 * while an associated cgwb is still active.
1094 void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1096 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1100 * blkcg_unpin_online - unpin online state
1101 * @blkcg_css: blkcg of interest
1103 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1104 * that blkg doesn't go offline while an associated cgwb is still active.
1105 * When this count goes to zero, all active cgwbs have finished so the
1106 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1108 void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1110 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1113 if (!refcount_dec_and_test(&blkcg->online_pin))
1115 blkcg_destroy_blkgs(blkcg);
1116 blkcg = blkcg_parent(blkcg);
1121 * blkcg_css_offline - cgroup css_offline callback
1122 * @css: css of interest
1124 * This function is called when @css is about to go away. Here the cgwbs are
1125 * offlined first and only once writeback associated with the blkcg has
1126 * finished do we start step 2 (see above).
1128 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1130 /* this prevents anyone from attaching or migrating to this blkcg */
1131 wb_blkcg_offline(css);
1133 /* put the base online pin allowing step 2 to be triggered */
1134 blkcg_unpin_online(css);
1137 static void blkcg_css_free(struct cgroup_subsys_state *css)
1139 struct blkcg *blkcg = css_to_blkcg(css);
1142 mutex_lock(&blkcg_pol_mutex);
1144 list_del(&blkcg->all_blkcgs_node);
1146 for (i = 0; i < BLKCG_MAX_POLS; i++)
1148 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1150 mutex_unlock(&blkcg_pol_mutex);
1155 static struct cgroup_subsys_state *
1156 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1158 struct blkcg *blkcg;
1159 struct cgroup_subsys_state *ret;
1162 mutex_lock(&blkcg_pol_mutex);
1165 blkcg = &blkcg_root;
1167 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1169 ret = ERR_PTR(-ENOMEM);
1174 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1175 struct blkcg_policy *pol = blkcg_policy[i];
1176 struct blkcg_policy_data *cpd;
1179 * If the policy hasn't been attached yet, wait for it
1180 * to be attached before doing anything else. Otherwise,
1181 * check if the policy requires any specific per-cgroup
1182 * data: if it does, allocate and initialize it.
1184 if (!pol || !pol->cpd_alloc_fn)
1187 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1189 ret = ERR_PTR(-ENOMEM);
1192 blkcg->cpd[i] = cpd;
1195 if (pol->cpd_init_fn)
1196 pol->cpd_init_fn(cpd);
1199 spin_lock_init(&blkcg->lock);
1200 refcount_set(&blkcg->online_pin, 1);
1201 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1202 INIT_HLIST_HEAD(&blkcg->blkg_list);
1203 #ifdef CONFIG_CGROUP_WRITEBACK
1204 INIT_LIST_HEAD(&blkcg->cgwb_list);
1206 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1208 mutex_unlock(&blkcg_pol_mutex);
1212 for (i--; i >= 0; i--)
1214 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1216 if (blkcg != &blkcg_root)
1219 mutex_unlock(&blkcg_pol_mutex);
1223 static int blkcg_css_online(struct cgroup_subsys_state *css)
1225 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1228 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1229 * don't go offline while cgwbs are still active on them. Pin the
1230 * parent so that offline always happens towards the root.
1233 blkcg_pin_online(&parent->css);
1237 int blkcg_init_disk(struct gendisk *disk)
1239 struct request_queue *q = disk->queue;
1240 struct blkcg_gq *new_blkg, *blkg;
1244 INIT_LIST_HEAD(&q->blkg_list);
1246 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1250 preloaded = !radix_tree_preload(GFP_KERNEL);
1252 /* Make sure the root blkg exists. */
1253 /* spin_lock_irq can serve as RCU read-side critical section. */
1254 spin_lock_irq(&q->queue_lock);
1255 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1258 q->root_blkg = blkg;
1259 spin_unlock_irq(&q->queue_lock);
1262 radix_tree_preload_end();
1264 ret = blk_ioprio_init(disk);
1266 goto err_destroy_all;
1268 ret = blk_throtl_init(disk);
1270 goto err_ioprio_exit;
1272 ret = blk_iolatency_init(disk);
1274 goto err_throtl_exit;
1279 blk_throtl_exit(disk);
1281 blk_ioprio_exit(disk);
1283 blkg_destroy_all(disk);
1286 spin_unlock_irq(&q->queue_lock);
1288 radix_tree_preload_end();
1289 return PTR_ERR(blkg);
1292 void blkcg_exit_disk(struct gendisk *disk)
1294 blkg_destroy_all(disk);
1295 rq_qos_exit(disk->queue);
1296 blk_throtl_exit(disk);
1299 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1303 mutex_lock(&blkcg_pol_mutex);
1305 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1306 struct blkcg_policy *pol = blkcg_policy[i];
1307 struct blkcg *blkcg;
1309 if (!pol || !pol->cpd_bind_fn)
1312 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1313 if (blkcg->cpd[pol->plid])
1314 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1316 mutex_unlock(&blkcg_pol_mutex);
1319 static void blkcg_exit(struct task_struct *tsk)
1321 if (tsk->throttle_queue)
1322 blk_put_queue(tsk->throttle_queue);
1323 tsk->throttle_queue = NULL;
1326 struct cgroup_subsys io_cgrp_subsys = {
1327 .css_alloc = blkcg_css_alloc,
1328 .css_online = blkcg_css_online,
1329 .css_offline = blkcg_css_offline,
1330 .css_free = blkcg_css_free,
1331 .css_rstat_flush = blkcg_rstat_flush,
1333 .dfl_cftypes = blkcg_files,
1334 .legacy_cftypes = blkcg_legacy_files,
1335 .legacy_name = "blkio",
1339 * This ensures that, if available, memcg is automatically enabled
1340 * together on the default hierarchy so that the owner cgroup can
1341 * be retrieved from writeback pages.
1343 .depends_on = 1 << memory_cgrp_id,
1346 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1349 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1350 * @q: request_queue of interest
1351 * @pol: blkcg policy to activate
1353 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1354 * bypass mode to populate its blkgs with policy_data for @pol.
1356 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1357 * from IO path. Update of each blkg is protected by both queue and blkcg
1358 * locks so that holding either lock and testing blkcg_policy_enabled() is
1359 * always enough for dereferencing policy data.
1361 * The caller is responsible for synchronizing [de]activations and policy
1362 * [un]registerations. Returns 0 on success, -errno on failure.
1364 int blkcg_activate_policy(struct request_queue *q,
1365 const struct blkcg_policy *pol)
1367 struct blkg_policy_data *pd_prealloc = NULL;
1368 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1371 if (blkcg_policy_enabled(q, pol))
1375 blk_mq_freeze_queue(q);
1377 spin_lock_irq(&q->queue_lock);
1379 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1380 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1381 struct blkg_policy_data *pd;
1383 if (blkg->pd[pol->plid])
1386 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1387 if (blkg == pinned_blkg) {
1391 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1397 * GFP_NOWAIT failed. Free the existing one and
1398 * prealloc for @blkg w/ GFP_KERNEL.
1401 blkg_put(pinned_blkg);
1405 spin_unlock_irq(&q->queue_lock);
1408 pol->pd_free_fn(pd_prealloc);
1409 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1417 blkg->pd[pol->plid] = pd;
1419 pd->plid = pol->plid;
1422 /* all allocated, init in the same order */
1423 if (pol->pd_init_fn)
1424 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1425 pol->pd_init_fn(blkg->pd[pol->plid]);
1427 if (pol->pd_online_fn)
1428 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1429 pol->pd_online_fn(blkg->pd[pol->plid]);
1431 __set_bit(pol->plid, q->blkcg_pols);
1434 spin_unlock_irq(&q->queue_lock);
1437 blk_mq_unfreeze_queue(q);
1439 blkg_put(pinned_blkg);
1441 pol->pd_free_fn(pd_prealloc);
1445 /* alloc failed, nothing's initialized yet, free everything */
1446 spin_lock_irq(&q->queue_lock);
1447 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1448 struct blkcg *blkcg = blkg->blkcg;
1450 spin_lock(&blkcg->lock);
1451 if (blkg->pd[pol->plid]) {
1452 pol->pd_free_fn(blkg->pd[pol->plid]);
1453 blkg->pd[pol->plid] = NULL;
1455 spin_unlock(&blkcg->lock);
1457 spin_unlock_irq(&q->queue_lock);
1461 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1464 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1465 * @q: request_queue of interest
1466 * @pol: blkcg policy to deactivate
1468 * Deactivate @pol on @q. Follows the same synchronization rules as
1469 * blkcg_activate_policy().
1471 void blkcg_deactivate_policy(struct request_queue *q,
1472 const struct blkcg_policy *pol)
1474 struct blkcg_gq *blkg;
1476 if (!blkcg_policy_enabled(q, pol))
1480 blk_mq_freeze_queue(q);
1482 spin_lock_irq(&q->queue_lock);
1484 __clear_bit(pol->plid, q->blkcg_pols);
1486 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1487 struct blkcg *blkcg = blkg->blkcg;
1489 spin_lock(&blkcg->lock);
1490 if (blkg->pd[pol->plid]) {
1491 if (pol->pd_offline_fn)
1492 pol->pd_offline_fn(blkg->pd[pol->plid]);
1493 pol->pd_free_fn(blkg->pd[pol->plid]);
1494 blkg->pd[pol->plid] = NULL;
1496 spin_unlock(&blkcg->lock);
1499 spin_unlock_irq(&q->queue_lock);
1502 blk_mq_unfreeze_queue(q);
1504 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1506 static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1508 struct blkcg *blkcg;
1510 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1511 if (blkcg->cpd[pol->plid]) {
1512 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1513 blkcg->cpd[pol->plid] = NULL;
1519 * blkcg_policy_register - register a blkcg policy
1520 * @pol: blkcg policy to register
1522 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1523 * successful registration. Returns 0 on success and -errno on failure.
1525 int blkcg_policy_register(struct blkcg_policy *pol)
1527 struct blkcg *blkcg;
1530 mutex_lock(&blkcg_pol_register_mutex);
1531 mutex_lock(&blkcg_pol_mutex);
1533 /* find an empty slot */
1535 for (i = 0; i < BLKCG_MAX_POLS; i++)
1536 if (!blkcg_policy[i])
1538 if (i >= BLKCG_MAX_POLS) {
1539 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1543 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1544 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1545 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1550 blkcg_policy[pol->plid] = pol;
1552 /* allocate and install cpd's */
1553 if (pol->cpd_alloc_fn) {
1554 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1555 struct blkcg_policy_data *cpd;
1557 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1561 blkcg->cpd[pol->plid] = cpd;
1563 cpd->plid = pol->plid;
1564 if (pol->cpd_init_fn)
1565 pol->cpd_init_fn(cpd);
1569 mutex_unlock(&blkcg_pol_mutex);
1571 /* everything is in place, add intf files for the new policy */
1572 if (pol->dfl_cftypes)
1573 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1575 if (pol->legacy_cftypes)
1576 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1577 pol->legacy_cftypes));
1578 mutex_unlock(&blkcg_pol_register_mutex);
1582 if (pol->cpd_free_fn)
1583 blkcg_free_all_cpd(pol);
1585 blkcg_policy[pol->plid] = NULL;
1587 mutex_unlock(&blkcg_pol_mutex);
1588 mutex_unlock(&blkcg_pol_register_mutex);
1591 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1594 * blkcg_policy_unregister - unregister a blkcg policy
1595 * @pol: blkcg policy to unregister
1597 * Undo blkcg_policy_register(@pol). Might sleep.
1599 void blkcg_policy_unregister(struct blkcg_policy *pol)
1601 mutex_lock(&blkcg_pol_register_mutex);
1603 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1606 /* kill the intf files first */
1607 if (pol->dfl_cftypes)
1608 cgroup_rm_cftypes(pol->dfl_cftypes);
1609 if (pol->legacy_cftypes)
1610 cgroup_rm_cftypes(pol->legacy_cftypes);
1612 /* remove cpds and unregister */
1613 mutex_lock(&blkcg_pol_mutex);
1615 if (pol->cpd_free_fn)
1616 blkcg_free_all_cpd(pol);
1618 blkcg_policy[pol->plid] = NULL;
1620 mutex_unlock(&blkcg_pol_mutex);
1622 mutex_unlock(&blkcg_pol_register_mutex);
1624 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1626 bool __blkcg_punt_bio_submit(struct bio *bio)
1628 struct blkcg_gq *blkg = bio->bi_blkg;
1630 /* consume the flag first */
1631 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1633 /* never bounce for the root cgroup */
1637 spin_lock_bh(&blkg->async_bio_lock);
1638 bio_list_add(&blkg->async_bios, bio);
1639 spin_unlock_bh(&blkg->async_bio_lock);
1641 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1646 * Scale the accumulated delay based on how long it has been since we updated
1647 * the delay. We only call this when we are adding delay, in case it's been a
1648 * while since we added delay, and when we are checking to see if we need to
1649 * delay a task, to account for any delays that may have occurred.
1651 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1653 u64 old = atomic64_read(&blkg->delay_start);
1655 /* negative use_delay means no scaling, see blkcg_set_delay() */
1656 if (atomic_read(&blkg->use_delay) < 0)
1660 * We only want to scale down every second. The idea here is that we
1661 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1662 * time window. We only want to throttle tasks for recent delay that
1663 * has occurred, in 1 second time windows since that's the maximum
1664 * things can be throttled. We save the current delay window in
1665 * blkg->last_delay so we know what amount is still left to be charged
1666 * to the blkg from this point onward. blkg->last_use keeps track of
1667 * the use_delay counter. The idea is if we're unthrottling the blkg we
1668 * are ok with whatever is happening now, and we can take away more of
1669 * the accumulated delay as we've already throttled enough that
1670 * everybody is happy with their IO latencies.
1672 if (time_before64(old + NSEC_PER_SEC, now) &&
1673 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1674 u64 cur = atomic64_read(&blkg->delay_nsec);
1675 u64 sub = min_t(u64, blkg->last_delay, now - old);
1676 int cur_use = atomic_read(&blkg->use_delay);
1679 * We've been unthrottled, subtract a larger chunk of our
1680 * accumulated delay.
1682 if (cur_use < blkg->last_use)
1683 sub = max_t(u64, sub, blkg->last_delay >> 1);
1686 * This shouldn't happen, but handle it anyway. Our delay_nsec
1687 * should only ever be growing except here where we subtract out
1688 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1689 * rather not end up with negative numbers.
1691 if (unlikely(cur < sub)) {
1692 atomic64_set(&blkg->delay_nsec, 0);
1693 blkg->last_delay = 0;
1695 atomic64_sub(sub, &blkg->delay_nsec);
1696 blkg->last_delay = cur - sub;
1698 blkg->last_use = cur_use;
1703 * This is called when we want to actually walk up the hierarchy and check to
1704 * see if we need to throttle, and then actually throttle if there is some
1705 * accumulated delay. This should only be called upon return to user space so
1706 * we're not holding some lock that would induce a priority inversion.
1708 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1710 unsigned long pflags;
1712 u64 now = ktime_to_ns(ktime_get());
1717 while (blkg->parent) {
1718 int use_delay = atomic_read(&blkg->use_delay);
1723 blkcg_scale_delay(blkg, now);
1724 this_delay = atomic64_read(&blkg->delay_nsec);
1725 if (this_delay > delay_nsec) {
1726 delay_nsec = this_delay;
1727 clamp = use_delay > 0;
1730 blkg = blkg->parent;
1737 * Let's not sleep for all eternity if we've amassed a huge delay.
1738 * Swapping or metadata IO can accumulate 10's of seconds worth of
1739 * delay, and we want userspace to be able to do _something_ so cap the
1740 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1741 * tasks will be delayed for 0.25 second for every syscall. If
1742 * blkcg_set_delay() was used as indicated by negative use_delay, the
1743 * caller is responsible for regulating the range.
1746 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1749 psi_memstall_enter(&pflags);
1751 exp = ktime_add_ns(now, delay_nsec);
1752 tok = io_schedule_prepare();
1754 __set_current_state(TASK_KILLABLE);
1755 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1757 } while (!fatal_signal_pending(current));
1758 io_schedule_finish(tok);
1761 psi_memstall_leave(&pflags);
1765 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1767 * This is only called if we've been marked with set_notify_resume(). Obviously
1768 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1769 * check to see if current->throttle_queue is set and if not this doesn't do
1770 * anything. This should only ever be called by the resume code, it's not meant
1771 * to be called by people willy-nilly as it will actually do the work to
1772 * throttle the task if it is setup for throttling.
1774 void blkcg_maybe_throttle_current(void)
1776 struct request_queue *q = current->throttle_queue;
1777 struct blkcg *blkcg;
1778 struct blkcg_gq *blkg;
1779 bool use_memdelay = current->use_memdelay;
1784 current->throttle_queue = NULL;
1785 current->use_memdelay = false;
1788 blkcg = css_to_blkcg(blkcg_css());
1791 blkg = blkg_lookup(blkcg, q);
1794 if (!blkg_tryget(blkg))
1798 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1808 * blkcg_schedule_throttle - this task needs to check for throttling
1809 * @gendisk: disk to throttle
1810 * @use_memdelay: do we charge this to memory delay for PSI
1812 * This is called by the IO controller when we know there's delay accumulated
1813 * for the blkg for this task. We do not pass the blkg because there are places
1814 * we call this that may not have that information, the swapping code for
1815 * instance will only have a block_device at that point. This set's the
1816 * notify_resume for the task to check and see if it requires throttling before
1817 * returning to user space.
1819 * We will only schedule once per syscall. You can call this over and over
1820 * again and it will only do the check once upon return to user space, and only
1821 * throttle once. If the task needs to be throttled again it'll need to be
1822 * re-set at the next time we see the task.
1824 void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
1826 struct request_queue *q = disk->queue;
1828 if (unlikely(current->flags & PF_KTHREAD))
1831 if (current->throttle_queue != q) {
1832 if (!blk_get_queue(q))
1835 if (current->throttle_queue)
1836 blk_put_queue(current->throttle_queue);
1837 current->throttle_queue = q;
1841 current->use_memdelay = use_memdelay;
1842 set_notify_resume(current);
1846 * blkcg_add_delay - add delay to this blkg
1847 * @blkg: blkg of interest
1848 * @now: the current time in nanoseconds
1849 * @delta: how many nanoseconds of delay to add
1851 * Charge @delta to the blkg's current delay accumulation. This is used to
1852 * throttle tasks if an IO controller thinks we need more throttling.
1854 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1856 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1858 blkcg_scale_delay(blkg, now);
1859 atomic64_add(delta, &blkg->delay_nsec);
1863 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1867 * As the failure mode here is to walk up the blkg tree, this ensure that the
1868 * blkg->parent pointers are always valid. This returns the blkg that it ended
1869 * up taking a reference on or %NULL if no reference was taken.
1871 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1872 struct cgroup_subsys_state *css)
1874 struct blkcg_gq *blkg, *ret_blkg = NULL;
1877 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
1879 if (blkg_tryget(blkg)) {
1883 blkg = blkg->parent;
1891 * bio_associate_blkg_from_css - associate a bio with a specified css
1895 * Associate @bio with the blkg found by combining the css's blkg and the
1896 * request_queue of the @bio. An association failure is handled by walking up
1897 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1898 * and q->root_blkg. This situation only happens when a cgroup is dying and
1899 * then the remaining bios will spill to the closest alive blkg.
1901 * A reference will be taken on the blkg and will be released when @bio is
1904 void bio_associate_blkg_from_css(struct bio *bio,
1905 struct cgroup_subsys_state *css)
1908 blkg_put(bio->bi_blkg);
1910 if (css && css->parent) {
1911 bio->bi_blkg = blkg_tryget_closest(bio, css);
1913 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1914 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1917 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1920 * bio_associate_blkg - associate a bio with a blkg
1923 * Associate @bio with the blkg found from the bio's css and request_queue.
1924 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1925 * already associated, the css is reused and association redone as the
1926 * request_queue may have changed.
1928 void bio_associate_blkg(struct bio *bio)
1930 struct cgroup_subsys_state *css;
1935 css = bio_blkcg_css(bio);
1939 bio_associate_blkg_from_css(bio, css);
1943 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1946 * bio_clone_blkg_association - clone blkg association from src to dst bio
1947 * @dst: destination bio
1950 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1953 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
1955 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1957 static int blk_cgroup_io_type(struct bio *bio)
1959 if (op_is_discard(bio->bi_opf))
1960 return BLKG_IOSTAT_DISCARD;
1961 if (op_is_write(bio->bi_opf))
1962 return BLKG_IOSTAT_WRITE;
1963 return BLKG_IOSTAT_READ;
1966 void blk_cgroup_bio_start(struct bio *bio)
1968 int rwd = blk_cgroup_io_type(bio), cpu;
1969 struct blkg_iostat_set *bis;
1970 unsigned long flags;
1973 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1974 flags = u64_stats_update_begin_irqsave(&bis->sync);
1977 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1978 * bio and we would have already accounted for the size of the bio.
1980 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1981 bio_set_flag(bio, BIO_CGROUP_ACCT);
1982 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1984 bis->cur.ios[rwd]++;
1986 u64_stats_update_end_irqrestore(&bis->sync, flags);
1987 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1988 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1992 bool blk_cgroup_congested(void)
1994 struct cgroup_subsys_state *css;
1998 for (css = blkcg_css(); css; css = css->parent) {
1999 if (atomic_read(&css->cgroup->congestion_count)) {
2008 static int __init blkcg_init(void)
2010 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
2011 WQ_MEM_RECLAIM | WQ_FREEZABLE |
2012 WQ_UNBOUND | WQ_SYSFS, 0);
2013 if (!blkcg_punt_bio_wq)
2017 subsys_initcall(blkcg_init);
2019 module_param(blkcg_debug_stats, bool, 0644);
2020 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");