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/genhd.h>
27 #include <linux/delay.h>
28 #include <linux/atomic.h>
29 #include <linux/ctype.h>
30 #include <linux/blk-cgroup.h>
31 #include <linux/tracehook.h>
32 #include <linux/psi.h>
33 #include <linux/part_stat.h>
35 #include "blk-ioprio.h"
36 #include "blk-throttle.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
63 static bool blkcg_policy_enabled(struct request_queue *q,
64 const struct blkcg_policy *pol)
66 return pol && test_bit(pol->plid, q->blkcg_pols);
70 * blkg_free - free a blkg
73 * Free @blkg which may be partially allocated.
75 static void blkg_free(struct blkcg_gq *blkg)
82 for (i = 0; i < BLKCG_MAX_POLS; i++)
84 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
86 free_percpu(blkg->iostat_cpu);
87 percpu_ref_exit(&blkg->refcnt);
91 static void __blkg_release(struct rcu_head *rcu)
93 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
95 WARN_ON(!bio_list_empty(&blkg->async_bios));
97 /* release the blkcg and parent blkg refs this blkg has been holding */
98 css_put(&blkg->blkcg->css);
100 blkg_put(blkg->parent);
105 * A group is RCU protected, but having an rcu lock does not mean that one
106 * can access all the fields of blkg and assume these are valid. For
107 * example, don't try to follow throtl_data and request queue links.
109 * Having a reference to blkg under an rcu allows accesses to only values
110 * local to groups like group stats and group rate limits.
112 static void blkg_release(struct percpu_ref *ref)
114 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
116 call_rcu(&blkg->rcu_head, __blkg_release);
119 static void blkg_async_bio_workfn(struct work_struct *work)
121 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
123 struct bio_list bios = BIO_EMPTY_LIST;
125 struct blk_plug plug;
126 bool need_plug = false;
128 /* as long as there are pending bios, @blkg can't go away */
129 spin_lock_bh(&blkg->async_bio_lock);
130 bio_list_merge(&bios, &blkg->async_bios);
131 bio_list_init(&blkg->async_bios);
132 spin_unlock_bh(&blkg->async_bio_lock);
134 /* start plug only when bio_list contains at least 2 bios */
135 if (bios.head && bios.head->bi_next) {
137 blk_start_plug(&plug);
139 while ((bio = bio_list_pop(&bios)))
142 blk_finish_plug(&plug);
146 * blkg_alloc - allocate a blkg
147 * @blkcg: block cgroup the new blkg is associated with
148 * @q: request_queue the new blkg is associated with
149 * @gfp_mask: allocation mask to use
151 * Allocate a new blkg assocating @blkcg and @q.
153 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
156 struct blkcg_gq *blkg;
159 /* alloc and init base part */
160 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
164 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
167 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
168 if (!blkg->iostat_cpu)
172 INIT_LIST_HEAD(&blkg->q_node);
173 spin_lock_init(&blkg->async_bio_lock);
174 bio_list_init(&blkg->async_bios);
175 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
178 u64_stats_init(&blkg->iostat.sync);
179 for_each_possible_cpu(cpu)
180 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
182 for (i = 0; i < BLKCG_MAX_POLS; i++) {
183 struct blkcg_policy *pol = blkcg_policy[i];
184 struct blkg_policy_data *pd;
186 if (!blkcg_policy_enabled(q, pol))
189 /* alloc per-policy data and attach it to blkg */
190 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
206 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
207 struct request_queue *q, bool update_hint)
209 struct blkcg_gq *blkg;
212 * Hint didn't match. Look up from the radix tree. Note that the
213 * hint can only be updated under queue_lock as otherwise @blkg
214 * could have already been removed from blkg_tree. The caller is
215 * responsible for grabbing queue_lock if @update_hint.
217 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
218 if (blkg && blkg->q == q) {
220 lockdep_assert_held(&q->queue_lock);
221 rcu_assign_pointer(blkcg->blkg_hint, blkg);
228 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
231 * If @new_blkg is %NULL, this function tries to allocate a new one as
232 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
234 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
235 struct request_queue *q,
236 struct blkcg_gq *new_blkg)
238 struct blkcg_gq *blkg;
241 WARN_ON_ONCE(!rcu_read_lock_held());
242 lockdep_assert_held(&q->queue_lock);
244 /* request_queue is dying, do not create/recreate a blkg */
245 if (blk_queue_dying(q)) {
250 /* blkg holds a reference to blkcg */
251 if (!css_tryget_online(&blkcg->css)) {
258 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
259 if (unlikely(!new_blkg)) {
267 if (blkcg_parent(blkcg)) {
268 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
269 if (WARN_ON_ONCE(!blkg->parent)) {
273 blkg_get(blkg->parent);
276 /* invoke per-policy init */
277 for (i = 0; i < BLKCG_MAX_POLS; i++) {
278 struct blkcg_policy *pol = blkcg_policy[i];
280 if (blkg->pd[i] && pol->pd_init_fn)
281 pol->pd_init_fn(blkg->pd[i]);
285 spin_lock(&blkcg->lock);
286 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
288 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
289 list_add(&blkg->q_node, &q->blkg_list);
291 for (i = 0; i < BLKCG_MAX_POLS; i++) {
292 struct blkcg_policy *pol = blkcg_policy[i];
294 if (blkg->pd[i] && pol->pd_online_fn)
295 pol->pd_online_fn(blkg->pd[i]);
299 spin_unlock(&blkcg->lock);
304 /* @blkg failed fully initialized, use the usual release path */
309 css_put(&blkcg->css);
316 * blkg_lookup_create - lookup blkg, try to create one if not there
317 * @blkcg: blkcg of interest
318 * @q: request_queue of interest
320 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
321 * create one. blkg creation is performed recursively from blkcg_root such
322 * that all non-root blkg's have access to the parent blkg. This function
323 * should be called under RCU read lock and takes @q->queue_lock.
325 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
328 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
329 struct request_queue *q)
331 struct blkcg_gq *blkg;
334 WARN_ON_ONCE(!rcu_read_lock_held());
336 blkg = blkg_lookup(blkcg, q);
340 spin_lock_irqsave(&q->queue_lock, flags);
341 blkg = __blkg_lookup(blkcg, q, true);
346 * Create blkgs walking down from blkcg_root to @blkcg, so that all
347 * non-root blkgs have access to their parents. Returns the closest
348 * blkg to the intended blkg should blkg_create() fail.
351 struct blkcg *pos = blkcg;
352 struct blkcg *parent = blkcg_parent(blkcg);
353 struct blkcg_gq *ret_blkg = q->root_blkg;
356 blkg = __blkg_lookup(parent, q, false);
358 /* remember closest blkg */
363 parent = blkcg_parent(parent);
366 blkg = blkg_create(pos, q, NULL);
376 spin_unlock_irqrestore(&q->queue_lock, flags);
380 static void blkg_destroy(struct blkcg_gq *blkg)
382 struct blkcg *blkcg = blkg->blkcg;
385 lockdep_assert_held(&blkg->q->queue_lock);
386 lockdep_assert_held(&blkcg->lock);
388 /* Something wrong if we are trying to remove same group twice */
389 WARN_ON_ONCE(list_empty(&blkg->q_node));
390 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
392 for (i = 0; i < BLKCG_MAX_POLS; i++) {
393 struct blkcg_policy *pol = blkcg_policy[i];
395 if (blkg->pd[i] && pol->pd_offline_fn)
396 pol->pd_offline_fn(blkg->pd[i]);
399 blkg->online = false;
401 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
402 list_del_init(&blkg->q_node);
403 hlist_del_init_rcu(&blkg->blkcg_node);
406 * Both setting lookup hint to and clearing it from @blkg are done
407 * under queue_lock. If it's not pointing to @blkg now, it never
408 * will. Hint assignment itself can race safely.
410 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
411 rcu_assign_pointer(blkcg->blkg_hint, NULL);
414 * Put the reference taken at the time of creation so that when all
415 * queues are gone, group can be destroyed.
417 percpu_ref_kill(&blkg->refcnt);
421 * blkg_destroy_all - destroy all blkgs associated with a request_queue
422 * @q: request_queue of interest
424 * Destroy all blkgs associated with @q.
426 static void blkg_destroy_all(struct request_queue *q)
428 struct blkcg_gq *blkg, *n;
429 int count = BLKG_DESTROY_BATCH_SIZE;
432 spin_lock_irq(&q->queue_lock);
433 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
434 struct blkcg *blkcg = blkg->blkcg;
436 spin_lock(&blkcg->lock);
438 spin_unlock(&blkcg->lock);
441 * in order to avoid holding the spin lock for too long, release
442 * it when a batch of blkgs are destroyed.
445 count = BLKG_DESTROY_BATCH_SIZE;
446 spin_unlock_irq(&q->queue_lock);
453 spin_unlock_irq(&q->queue_lock);
456 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
457 struct cftype *cftype, u64 val)
459 struct blkcg *blkcg = css_to_blkcg(css);
460 struct blkcg_gq *blkg;
463 mutex_lock(&blkcg_pol_mutex);
464 spin_lock_irq(&blkcg->lock);
467 * Note that stat reset is racy - it doesn't synchronize against
468 * stat updates. This is a debug feature which shouldn't exist
469 * anyway. If you get hit by a race, retry.
471 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
472 for_each_possible_cpu(cpu) {
473 struct blkg_iostat_set *bis =
474 per_cpu_ptr(blkg->iostat_cpu, cpu);
475 memset(bis, 0, sizeof(*bis));
477 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
479 for (i = 0; i < BLKCG_MAX_POLS; i++) {
480 struct blkcg_policy *pol = blkcg_policy[i];
482 if (blkg->pd[i] && pol->pd_reset_stats_fn)
483 pol->pd_reset_stats_fn(blkg->pd[i]);
487 spin_unlock_irq(&blkcg->lock);
488 mutex_unlock(&blkcg_pol_mutex);
492 const char *blkg_dev_name(struct blkcg_gq *blkg)
494 if (!blkg->q->disk || !blkg->q->disk->bdi->dev)
496 return bdi_dev_name(blkg->q->disk->bdi);
500 * blkcg_print_blkgs - helper for printing per-blkg data
501 * @sf: seq_file to print to
502 * @blkcg: blkcg of interest
503 * @prfill: fill function to print out a blkg
504 * @pol: policy in question
505 * @data: data to be passed to @prfill
506 * @show_total: to print out sum of prfill return values or not
508 * This function invokes @prfill on each blkg of @blkcg if pd for the
509 * policy specified by @pol exists. @prfill is invoked with @sf, the
510 * policy data and @data and the matching queue lock held. If @show_total
511 * is %true, the sum of the return values from @prfill is printed with
512 * "Total" label at the end.
514 * This is to be used to construct print functions for
515 * cftype->read_seq_string method.
517 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
518 u64 (*prfill)(struct seq_file *,
519 struct blkg_policy_data *, int),
520 const struct blkcg_policy *pol, int data,
523 struct blkcg_gq *blkg;
527 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
528 spin_lock_irq(&blkg->q->queue_lock);
529 if (blkcg_policy_enabled(blkg->q, pol))
530 total += prfill(sf, blkg->pd[pol->plid], data);
531 spin_unlock_irq(&blkg->q->queue_lock);
536 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
538 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
541 * __blkg_prfill_u64 - prfill helper for a single u64 value
542 * @sf: seq_file to print to
543 * @pd: policy private data of interest
546 * Print @v to @sf for the device assocaited with @pd.
548 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
550 const char *dname = blkg_dev_name(pd->blkg);
555 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
558 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
560 /* Performs queue bypass and policy enabled checks then looks up blkg. */
561 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
562 const struct blkcg_policy *pol,
563 struct request_queue *q)
565 WARN_ON_ONCE(!rcu_read_lock_held());
566 lockdep_assert_held(&q->queue_lock);
568 if (!blkcg_policy_enabled(q, pol))
569 return ERR_PTR(-EOPNOTSUPP);
570 return __blkg_lookup(blkcg, q, true /* update_hint */);
574 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
575 * @inputp: input string pointer
577 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
578 * from @input and get and return the matching bdev. *@inputp is
579 * updated to point past the device node prefix. Returns an ERR_PTR()
582 * Use this function iff blkg_conf_prep() can't be used for some reason.
584 struct block_device *blkcg_conf_open_bdev(char **inputp)
586 char *input = *inputp;
587 unsigned int major, minor;
588 struct block_device *bdev;
591 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
592 return ERR_PTR(-EINVAL);
595 if (!isspace(*input))
596 return ERR_PTR(-EINVAL);
597 input = skip_spaces(input);
599 bdev = blkdev_get_no_open(MKDEV(major, minor));
601 return ERR_PTR(-ENODEV);
602 if (bdev_is_partition(bdev)) {
603 blkdev_put_no_open(bdev);
604 return ERR_PTR(-ENODEV);
612 * blkg_conf_prep - parse and prepare for per-blkg config update
613 * @blkcg: target block cgroup
614 * @pol: target policy
615 * @input: input string
616 * @ctx: blkg_conf_ctx to be filled
618 * Parse per-blkg config update from @input and initialize @ctx with the
619 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
620 * part of @input following MAJ:MIN. This function returns with RCU read
621 * lock and queue lock held and must be paired with blkg_conf_finish().
623 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
624 char *input, struct blkg_conf_ctx *ctx)
625 __acquires(rcu) __acquires(&bdev->bd_queue->queue_lock)
627 struct block_device *bdev;
628 struct request_queue *q;
629 struct blkcg_gq *blkg;
632 bdev = blkcg_conf_open_bdev(&input);
634 return PTR_ERR(bdev);
636 q = bdev_get_queue(bdev);
639 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
640 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
642 ret = blk_queue_enter(q, 0);
647 spin_lock_irq(&q->queue_lock);
649 blkg = blkg_lookup_check(blkcg, pol, q);
659 * Create blkgs walking down from blkcg_root to @blkcg, so that all
660 * non-root blkgs have access to their parents.
663 struct blkcg *pos = blkcg;
664 struct blkcg *parent;
665 struct blkcg_gq *new_blkg;
667 parent = blkcg_parent(blkcg);
668 while (parent && !__blkg_lookup(parent, q, false)) {
670 parent = blkcg_parent(parent);
673 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
674 spin_unlock_irq(&q->queue_lock);
677 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
678 if (unlikely(!new_blkg)) {
680 goto fail_exit_queue;
683 if (radix_tree_preload(GFP_KERNEL)) {
686 goto fail_exit_queue;
690 spin_lock_irq(&q->queue_lock);
692 blkg = blkg_lookup_check(pos, pol, q);
702 blkg = blkg_create(pos, q, new_blkg);
709 radix_tree_preload_end();
722 radix_tree_preload_end();
724 spin_unlock_irq(&q->queue_lock);
729 blkdev_put_no_open(bdev);
731 * If queue was bypassing, we should retry. Do so after a
732 * short msleep(). It isn't strictly necessary but queue
733 * can be bypassing for some time and it's always nice to
734 * avoid busy looping.
738 ret = restart_syscall();
742 EXPORT_SYMBOL_GPL(blkg_conf_prep);
745 * blkg_conf_finish - finish up per-blkg config update
746 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
748 * Finish up after per-blkg config update. This function must be paired
749 * with blkg_conf_prep().
751 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
752 __releases(&ctx->bdev->bd_queue->queue_lock) __releases(rcu)
754 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
756 blkdev_put_no_open(ctx->bdev);
758 EXPORT_SYMBOL_GPL(blkg_conf_finish);
760 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
764 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
765 dst->bytes[i] = src->bytes[i];
766 dst->ios[i] = src->ios[i];
770 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
774 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
775 dst->bytes[i] += src->bytes[i];
776 dst->ios[i] += src->ios[i];
780 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
784 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
785 dst->bytes[i] -= src->bytes[i];
786 dst->ios[i] -= src->ios[i];
790 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
792 struct blkcg *blkcg = css_to_blkcg(css);
793 struct blkcg_gq *blkg;
795 /* Root-level stats are sourced from system-wide IO stats */
796 if (!cgroup_parent(css->cgroup))
801 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
802 struct blkcg_gq *parent = blkg->parent;
803 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
804 struct blkg_iostat cur, delta;
808 /* fetch the current per-cpu values */
810 seq = u64_stats_fetch_begin(&bisc->sync);
811 blkg_iostat_set(&cur, &bisc->cur);
812 } while (u64_stats_fetch_retry(&bisc->sync, seq));
814 /* propagate percpu delta to global */
815 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
816 blkg_iostat_set(&delta, &cur);
817 blkg_iostat_sub(&delta, &bisc->last);
818 blkg_iostat_add(&blkg->iostat.cur, &delta);
819 blkg_iostat_add(&bisc->last, &delta);
820 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
822 /* propagate global delta to parent (unless that's root) */
823 if (parent && parent->parent) {
824 flags = u64_stats_update_begin_irqsave(&parent->iostat.sync);
825 blkg_iostat_set(&delta, &blkg->iostat.cur);
826 blkg_iostat_sub(&delta, &blkg->iostat.last);
827 blkg_iostat_add(&parent->iostat.cur, &delta);
828 blkg_iostat_add(&blkg->iostat.last, &delta);
829 u64_stats_update_end_irqrestore(&parent->iostat.sync, flags);
837 * We source root cgroup stats from the system-wide stats to avoid
838 * tracking the same information twice and incurring overhead when no
839 * cgroups are defined. For that reason, cgroup_rstat_flush in
840 * blkcg_print_stat does not actually fill out the iostat in the root
843 * However, we would like to re-use the printing code between the root and
844 * non-root cgroups to the extent possible. For that reason, we simulate
845 * flushing the root cgroup's stats by explicitly filling in the iostat
846 * with disk level statistics.
848 static void blkcg_fill_root_iostats(void)
850 struct class_dev_iter iter;
853 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
854 while ((dev = class_dev_iter_next(&iter))) {
855 struct block_device *bdev = dev_to_bdev(dev);
856 struct blkcg_gq *blkg =
857 blk_queue_root_blkg(bdev_get_queue(bdev));
858 struct blkg_iostat tmp;
861 memset(&tmp, 0, sizeof(tmp));
862 for_each_possible_cpu(cpu) {
863 struct disk_stats *cpu_dkstats;
866 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
867 tmp.ios[BLKG_IOSTAT_READ] +=
868 cpu_dkstats->ios[STAT_READ];
869 tmp.ios[BLKG_IOSTAT_WRITE] +=
870 cpu_dkstats->ios[STAT_WRITE];
871 tmp.ios[BLKG_IOSTAT_DISCARD] +=
872 cpu_dkstats->ios[STAT_DISCARD];
873 // convert sectors to bytes
874 tmp.bytes[BLKG_IOSTAT_READ] +=
875 cpu_dkstats->sectors[STAT_READ] << 9;
876 tmp.bytes[BLKG_IOSTAT_WRITE] +=
877 cpu_dkstats->sectors[STAT_WRITE] << 9;
878 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
879 cpu_dkstats->sectors[STAT_DISCARD] << 9;
881 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
882 blkg_iostat_set(&blkg->iostat.cur, &tmp);
883 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
888 static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
890 struct blkg_iostat_set *bis = &blkg->iostat;
891 u64 rbytes, wbytes, rios, wios, dbytes, dios;
892 bool has_stats = false;
900 dname = blkg_dev_name(blkg);
904 seq_printf(s, "%s ", dname);
907 seq = u64_stats_fetch_begin(&bis->sync);
909 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
910 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
911 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
912 rios = bis->cur.ios[BLKG_IOSTAT_READ];
913 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
914 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
915 } while (u64_stats_fetch_retry(&bis->sync, seq));
917 if (rbytes || wbytes || rios || wios) {
919 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
920 rbytes, wbytes, rios, wios,
924 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
926 seq_printf(s, " use_delay=%d delay_nsec=%llu",
927 atomic_read(&blkg->use_delay),
928 atomic64_read(&blkg->delay_nsec));
931 for (i = 0; i < BLKCG_MAX_POLS; i++) {
932 struct blkcg_policy *pol = blkcg_policy[i];
934 if (!blkg->pd[i] || !pol->pd_stat_fn)
937 if (pol->pd_stat_fn(blkg->pd[i], s))
945 static int blkcg_print_stat(struct seq_file *sf, void *v)
947 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
948 struct blkcg_gq *blkg;
950 if (!seq_css(sf)->parent)
951 blkcg_fill_root_iostats();
953 cgroup_rstat_flush(blkcg->css.cgroup);
956 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
957 spin_lock_irq(&blkg->q->queue_lock);
958 blkcg_print_one_stat(blkg, sf);
959 spin_unlock_irq(&blkg->q->queue_lock);
965 static struct cftype blkcg_files[] = {
968 .seq_show = blkcg_print_stat,
973 static struct cftype blkcg_legacy_files[] = {
975 .name = "reset_stats",
976 .write_u64 = blkcg_reset_stats,
982 * blkcg destruction is a three-stage process.
984 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
985 * which offlines writeback. Here we tie the next stage of blkg destruction
986 * to the completion of writeback associated with the blkcg. This lets us
987 * avoid punting potentially large amounts of outstanding writeback to root
988 * while maintaining any ongoing policies. The next stage is triggered when
989 * the nr_cgwbs count goes to zero.
991 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
992 * and handles the destruction of blkgs. Here the css reference held by
993 * the blkg is put back eventually allowing blkcg_css_free() to be called.
994 * This work may occur in cgwb_release_workfn() on the cgwb_release
995 * workqueue. Any submitted ios that fail to get the blkg ref will be
996 * punted to the root_blkg.
998 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
999 * This finally frees the blkcg.
1003 * blkcg_css_offline - cgroup css_offline callback
1004 * @css: css of interest
1006 * This function is called when @css is about to go away. Here the cgwbs are
1007 * offlined first and only once writeback associated with the blkcg has
1008 * finished do we start step 2 (see above).
1010 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1012 struct blkcg *blkcg = css_to_blkcg(css);
1014 /* this prevents anyone from attaching or migrating to this blkcg */
1015 wb_blkcg_offline(blkcg);
1017 /* put the base online pin allowing step 2 to be triggered */
1018 blkcg_unpin_online(blkcg);
1022 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1023 * @blkcg: blkcg of interest
1025 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1026 * is nested inside q lock, this function performs reverse double lock dancing.
1027 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1028 * blkcg_css_free to eventually be called.
1030 * This is the blkcg counterpart of ioc_release_fn().
1032 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1036 spin_lock_irq(&blkcg->lock);
1038 while (!hlist_empty(&blkcg->blkg_list)) {
1039 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1040 struct blkcg_gq, blkcg_node);
1041 struct request_queue *q = blkg->q;
1043 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1045 * Given that the system can accumulate a huge number
1046 * of blkgs in pathological cases, check to see if we
1047 * need to rescheduling to avoid softlockup.
1049 spin_unlock_irq(&blkcg->lock);
1051 spin_lock_irq(&blkcg->lock);
1056 spin_unlock(&q->queue_lock);
1059 spin_unlock_irq(&blkcg->lock);
1062 static void blkcg_css_free(struct cgroup_subsys_state *css)
1064 struct blkcg *blkcg = css_to_blkcg(css);
1067 mutex_lock(&blkcg_pol_mutex);
1069 list_del(&blkcg->all_blkcgs_node);
1071 for (i = 0; i < BLKCG_MAX_POLS; i++)
1073 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1075 mutex_unlock(&blkcg_pol_mutex);
1080 static struct cgroup_subsys_state *
1081 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1083 struct blkcg *blkcg;
1084 struct cgroup_subsys_state *ret;
1087 mutex_lock(&blkcg_pol_mutex);
1090 blkcg = &blkcg_root;
1092 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1094 ret = ERR_PTR(-ENOMEM);
1099 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1100 struct blkcg_policy *pol = blkcg_policy[i];
1101 struct blkcg_policy_data *cpd;
1104 * If the policy hasn't been attached yet, wait for it
1105 * to be attached before doing anything else. Otherwise,
1106 * check if the policy requires any specific per-cgroup
1107 * data: if it does, allocate and initialize it.
1109 if (!pol || !pol->cpd_alloc_fn)
1112 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1114 ret = ERR_PTR(-ENOMEM);
1117 blkcg->cpd[i] = cpd;
1120 if (pol->cpd_init_fn)
1121 pol->cpd_init_fn(cpd);
1124 spin_lock_init(&blkcg->lock);
1125 refcount_set(&blkcg->online_pin, 1);
1126 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1127 INIT_HLIST_HEAD(&blkcg->blkg_list);
1128 #ifdef CONFIG_CGROUP_WRITEBACK
1129 INIT_LIST_HEAD(&blkcg->cgwb_list);
1131 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1133 mutex_unlock(&blkcg_pol_mutex);
1137 for (i--; i >= 0; i--)
1139 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1141 if (blkcg != &blkcg_root)
1144 mutex_unlock(&blkcg_pol_mutex);
1148 static int blkcg_css_online(struct cgroup_subsys_state *css)
1150 struct blkcg *blkcg = css_to_blkcg(css);
1151 struct blkcg *parent = blkcg_parent(blkcg);
1154 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1155 * don't go offline while cgwbs are still active on them. Pin the
1156 * parent so that offline always happens towards the root.
1159 blkcg_pin_online(parent);
1164 * blkcg_init_queue - initialize blkcg part of request queue
1165 * @q: request_queue to initialize
1167 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1168 * part of new request_queue @q.
1171 * 0 on success, -errno on failure.
1173 int blkcg_init_queue(struct request_queue *q)
1175 struct blkcg_gq *new_blkg, *blkg;
1179 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1183 preloaded = !radix_tree_preload(GFP_KERNEL);
1185 /* Make sure the root blkg exists. */
1187 spin_lock_irq(&q->queue_lock);
1188 blkg = blkg_create(&blkcg_root, q, new_blkg);
1191 q->root_blkg = blkg;
1192 spin_unlock_irq(&q->queue_lock);
1196 radix_tree_preload_end();
1198 ret = blk_ioprio_init(q);
1200 goto err_destroy_all;
1202 ret = blk_throtl_init(q);
1204 goto err_destroy_all;
1206 ret = blk_iolatency_init(q);
1209 goto err_destroy_all;
1215 blkg_destroy_all(q);
1218 spin_unlock_irq(&q->queue_lock);
1221 radix_tree_preload_end();
1222 return PTR_ERR(blkg);
1226 * blkcg_exit_queue - exit and release blkcg part of request_queue
1227 * @q: request_queue being released
1229 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1231 void blkcg_exit_queue(struct request_queue *q)
1233 blkg_destroy_all(q);
1237 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1241 mutex_lock(&blkcg_pol_mutex);
1243 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1244 struct blkcg_policy *pol = blkcg_policy[i];
1245 struct blkcg *blkcg;
1247 if (!pol || !pol->cpd_bind_fn)
1250 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1251 if (blkcg->cpd[pol->plid])
1252 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1254 mutex_unlock(&blkcg_pol_mutex);
1257 static void blkcg_exit(struct task_struct *tsk)
1259 if (tsk->throttle_queue)
1260 blk_put_queue(tsk->throttle_queue);
1261 tsk->throttle_queue = NULL;
1264 struct cgroup_subsys io_cgrp_subsys = {
1265 .css_alloc = blkcg_css_alloc,
1266 .css_online = blkcg_css_online,
1267 .css_offline = blkcg_css_offline,
1268 .css_free = blkcg_css_free,
1269 .css_rstat_flush = blkcg_rstat_flush,
1271 .dfl_cftypes = blkcg_files,
1272 .legacy_cftypes = blkcg_legacy_files,
1273 .legacy_name = "blkio",
1277 * This ensures that, if available, memcg is automatically enabled
1278 * together on the default hierarchy so that the owner cgroup can
1279 * be retrieved from writeback pages.
1281 .depends_on = 1 << memory_cgrp_id,
1284 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1287 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1288 * @q: request_queue of interest
1289 * @pol: blkcg policy to activate
1291 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1292 * bypass mode to populate its blkgs with policy_data for @pol.
1294 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1295 * from IO path. Update of each blkg is protected by both queue and blkcg
1296 * locks so that holding either lock and testing blkcg_policy_enabled() is
1297 * always enough for dereferencing policy data.
1299 * The caller is responsible for synchronizing [de]activations and policy
1300 * [un]registerations. Returns 0 on success, -errno on failure.
1302 int blkcg_activate_policy(struct request_queue *q,
1303 const struct blkcg_policy *pol)
1305 struct blkg_policy_data *pd_prealloc = NULL;
1306 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1309 if (blkcg_policy_enabled(q, pol))
1313 blk_mq_freeze_queue(q);
1315 spin_lock_irq(&q->queue_lock);
1317 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1318 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1319 struct blkg_policy_data *pd;
1321 if (blkg->pd[pol->plid])
1324 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1325 if (blkg == pinned_blkg) {
1329 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1335 * GFP_NOWAIT failed. Free the existing one and
1336 * prealloc for @blkg w/ GFP_KERNEL.
1339 blkg_put(pinned_blkg);
1343 spin_unlock_irq(&q->queue_lock);
1346 pol->pd_free_fn(pd_prealloc);
1347 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1355 blkg->pd[pol->plid] = pd;
1357 pd->plid = pol->plid;
1360 /* all allocated, init in the same order */
1361 if (pol->pd_init_fn)
1362 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1363 pol->pd_init_fn(blkg->pd[pol->plid]);
1365 __set_bit(pol->plid, q->blkcg_pols);
1368 spin_unlock_irq(&q->queue_lock);
1371 blk_mq_unfreeze_queue(q);
1373 blkg_put(pinned_blkg);
1375 pol->pd_free_fn(pd_prealloc);
1379 /* alloc failed, nothing's initialized yet, free everything */
1380 spin_lock_irq(&q->queue_lock);
1381 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1382 struct blkcg *blkcg = blkg->blkcg;
1384 spin_lock(&blkcg->lock);
1385 if (blkg->pd[pol->plid]) {
1386 pol->pd_free_fn(blkg->pd[pol->plid]);
1387 blkg->pd[pol->plid] = NULL;
1389 spin_unlock(&blkcg->lock);
1391 spin_unlock_irq(&q->queue_lock);
1395 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1398 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1399 * @q: request_queue of interest
1400 * @pol: blkcg policy to deactivate
1402 * Deactivate @pol on @q. Follows the same synchronization rules as
1403 * blkcg_activate_policy().
1405 void blkcg_deactivate_policy(struct request_queue *q,
1406 const struct blkcg_policy *pol)
1408 struct blkcg_gq *blkg;
1410 if (!blkcg_policy_enabled(q, pol))
1414 blk_mq_freeze_queue(q);
1416 spin_lock_irq(&q->queue_lock);
1418 __clear_bit(pol->plid, q->blkcg_pols);
1420 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1421 struct blkcg *blkcg = blkg->blkcg;
1423 spin_lock(&blkcg->lock);
1424 if (blkg->pd[pol->plid]) {
1425 if (pol->pd_offline_fn)
1426 pol->pd_offline_fn(blkg->pd[pol->plid]);
1427 pol->pd_free_fn(blkg->pd[pol->plid]);
1428 blkg->pd[pol->plid] = NULL;
1430 spin_unlock(&blkcg->lock);
1433 spin_unlock_irq(&q->queue_lock);
1436 blk_mq_unfreeze_queue(q);
1438 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1441 * blkcg_policy_register - register a blkcg policy
1442 * @pol: blkcg policy to register
1444 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1445 * successful registration. Returns 0 on success and -errno on failure.
1447 int blkcg_policy_register(struct blkcg_policy *pol)
1449 struct blkcg *blkcg;
1452 mutex_lock(&blkcg_pol_register_mutex);
1453 mutex_lock(&blkcg_pol_mutex);
1455 /* find an empty slot */
1457 for (i = 0; i < BLKCG_MAX_POLS; i++)
1458 if (!blkcg_policy[i])
1460 if (i >= BLKCG_MAX_POLS) {
1461 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1465 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1466 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1467 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1472 blkcg_policy[pol->plid] = pol;
1474 /* allocate and install cpd's */
1475 if (pol->cpd_alloc_fn) {
1476 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1477 struct blkcg_policy_data *cpd;
1479 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1483 blkcg->cpd[pol->plid] = cpd;
1485 cpd->plid = pol->plid;
1486 if (pol->cpd_init_fn)
1487 pol->cpd_init_fn(cpd);
1491 mutex_unlock(&blkcg_pol_mutex);
1493 /* everything is in place, add intf files for the new policy */
1494 if (pol->dfl_cftypes)
1495 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1497 if (pol->legacy_cftypes)
1498 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1499 pol->legacy_cftypes));
1500 mutex_unlock(&blkcg_pol_register_mutex);
1504 if (pol->cpd_free_fn) {
1505 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1506 if (blkcg->cpd[pol->plid]) {
1507 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1508 blkcg->cpd[pol->plid] = NULL;
1512 blkcg_policy[pol->plid] = NULL;
1514 mutex_unlock(&blkcg_pol_mutex);
1515 mutex_unlock(&blkcg_pol_register_mutex);
1518 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1521 * blkcg_policy_unregister - unregister a blkcg policy
1522 * @pol: blkcg policy to unregister
1524 * Undo blkcg_policy_register(@pol). Might sleep.
1526 void blkcg_policy_unregister(struct blkcg_policy *pol)
1528 struct blkcg *blkcg;
1530 mutex_lock(&blkcg_pol_register_mutex);
1532 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1535 /* kill the intf files first */
1536 if (pol->dfl_cftypes)
1537 cgroup_rm_cftypes(pol->dfl_cftypes);
1538 if (pol->legacy_cftypes)
1539 cgroup_rm_cftypes(pol->legacy_cftypes);
1541 /* remove cpds and unregister */
1542 mutex_lock(&blkcg_pol_mutex);
1544 if (pol->cpd_free_fn) {
1545 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1546 if (blkcg->cpd[pol->plid]) {
1547 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1548 blkcg->cpd[pol->plid] = NULL;
1552 blkcg_policy[pol->plid] = NULL;
1554 mutex_unlock(&blkcg_pol_mutex);
1556 mutex_unlock(&blkcg_pol_register_mutex);
1558 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1560 bool __blkcg_punt_bio_submit(struct bio *bio)
1562 struct blkcg_gq *blkg = bio->bi_blkg;
1564 /* consume the flag first */
1565 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1567 /* never bounce for the root cgroup */
1571 spin_lock_bh(&blkg->async_bio_lock);
1572 bio_list_add(&blkg->async_bios, bio);
1573 spin_unlock_bh(&blkg->async_bio_lock);
1575 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1580 * Scale the accumulated delay based on how long it has been since we updated
1581 * the delay. We only call this when we are adding delay, in case it's been a
1582 * while since we added delay, and when we are checking to see if we need to
1583 * delay a task, to account for any delays that may have occurred.
1585 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1587 u64 old = atomic64_read(&blkg->delay_start);
1589 /* negative use_delay means no scaling, see blkcg_set_delay() */
1590 if (atomic_read(&blkg->use_delay) < 0)
1594 * We only want to scale down every second. The idea here is that we
1595 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1596 * time window. We only want to throttle tasks for recent delay that
1597 * has occurred, in 1 second time windows since that's the maximum
1598 * things can be throttled. We save the current delay window in
1599 * blkg->last_delay so we know what amount is still left to be charged
1600 * to the blkg from this point onward. blkg->last_use keeps track of
1601 * the use_delay counter. The idea is if we're unthrottling the blkg we
1602 * are ok with whatever is happening now, and we can take away more of
1603 * the accumulated delay as we've already throttled enough that
1604 * everybody is happy with their IO latencies.
1606 if (time_before64(old + NSEC_PER_SEC, now) &&
1607 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1608 u64 cur = atomic64_read(&blkg->delay_nsec);
1609 u64 sub = min_t(u64, blkg->last_delay, now - old);
1610 int cur_use = atomic_read(&blkg->use_delay);
1613 * We've been unthrottled, subtract a larger chunk of our
1614 * accumulated delay.
1616 if (cur_use < blkg->last_use)
1617 sub = max_t(u64, sub, blkg->last_delay >> 1);
1620 * This shouldn't happen, but handle it anyway. Our delay_nsec
1621 * should only ever be growing except here where we subtract out
1622 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1623 * rather not end up with negative numbers.
1625 if (unlikely(cur < sub)) {
1626 atomic64_set(&blkg->delay_nsec, 0);
1627 blkg->last_delay = 0;
1629 atomic64_sub(sub, &blkg->delay_nsec);
1630 blkg->last_delay = cur - sub;
1632 blkg->last_use = cur_use;
1637 * This is called when we want to actually walk up the hierarchy and check to
1638 * see if we need to throttle, and then actually throttle if there is some
1639 * accumulated delay. This should only be called upon return to user space so
1640 * we're not holding some lock that would induce a priority inversion.
1642 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1644 unsigned long pflags;
1646 u64 now = ktime_to_ns(ktime_get());
1651 while (blkg->parent) {
1652 int use_delay = atomic_read(&blkg->use_delay);
1657 blkcg_scale_delay(blkg, now);
1658 this_delay = atomic64_read(&blkg->delay_nsec);
1659 if (this_delay > delay_nsec) {
1660 delay_nsec = this_delay;
1661 clamp = use_delay > 0;
1664 blkg = blkg->parent;
1671 * Let's not sleep for all eternity if we've amassed a huge delay.
1672 * Swapping or metadata IO can accumulate 10's of seconds worth of
1673 * delay, and we want userspace to be able to do _something_ so cap the
1674 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1675 * tasks will be delayed for 0.25 second for every syscall. If
1676 * blkcg_set_delay() was used as indicated by negative use_delay, the
1677 * caller is responsible for regulating the range.
1680 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1683 psi_memstall_enter(&pflags);
1685 exp = ktime_add_ns(now, delay_nsec);
1686 tok = io_schedule_prepare();
1688 __set_current_state(TASK_KILLABLE);
1689 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1691 } while (!fatal_signal_pending(current));
1692 io_schedule_finish(tok);
1695 psi_memstall_leave(&pflags);
1699 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1701 * This is only called if we've been marked with set_notify_resume(). Obviously
1702 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1703 * check to see if current->throttle_queue is set and if not this doesn't do
1704 * anything. This should only ever be called by the resume code, it's not meant
1705 * to be called by people willy-nilly as it will actually do the work to
1706 * throttle the task if it is setup for throttling.
1708 void blkcg_maybe_throttle_current(void)
1710 struct request_queue *q = current->throttle_queue;
1711 struct cgroup_subsys_state *css;
1712 struct blkcg *blkcg;
1713 struct blkcg_gq *blkg;
1714 bool use_memdelay = current->use_memdelay;
1719 current->throttle_queue = NULL;
1720 current->use_memdelay = false;
1723 css = kthread_blkcg();
1725 blkcg = css_to_blkcg(css);
1727 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1731 blkg = blkg_lookup(blkcg, q);
1734 if (!blkg_tryget(blkg))
1738 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1748 * blkcg_schedule_throttle - this task needs to check for throttling
1749 * @q: the request queue IO was submitted on
1750 * @use_memdelay: do we charge this to memory delay for PSI
1752 * This is called by the IO controller when we know there's delay accumulated
1753 * for the blkg for this task. We do not pass the blkg because there are places
1754 * we call this that may not have that information, the swapping code for
1755 * instance will only have a request_queue at that point. This set's the
1756 * notify_resume for the task to check and see if it requires throttling before
1757 * returning to user space.
1759 * We will only schedule once per syscall. You can call this over and over
1760 * again and it will only do the check once upon return to user space, and only
1761 * throttle once. If the task needs to be throttled again it'll need to be
1762 * re-set at the next time we see the task.
1764 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1766 if (unlikely(current->flags & PF_KTHREAD))
1769 if (current->throttle_queue != q) {
1770 if (!blk_get_queue(q))
1773 if (current->throttle_queue)
1774 blk_put_queue(current->throttle_queue);
1775 current->throttle_queue = q;
1779 current->use_memdelay = use_memdelay;
1780 set_notify_resume(current);
1784 * blkcg_add_delay - add delay to this blkg
1785 * @blkg: blkg of interest
1786 * @now: the current time in nanoseconds
1787 * @delta: how many nanoseconds of delay to add
1789 * Charge @delta to the blkg's current delay accumulation. This is used to
1790 * throttle tasks if an IO controller thinks we need more throttling.
1792 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1794 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1796 blkcg_scale_delay(blkg, now);
1797 atomic64_add(delta, &blkg->delay_nsec);
1801 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1805 * As the failure mode here is to walk up the blkg tree, this ensure that the
1806 * blkg->parent pointers are always valid. This returns the blkg that it ended
1807 * up taking a reference on or %NULL if no reference was taken.
1809 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1810 struct cgroup_subsys_state *css)
1812 struct blkcg_gq *blkg, *ret_blkg = NULL;
1815 blkg = blkg_lookup_create(css_to_blkcg(css),
1816 bdev_get_queue(bio->bi_bdev));
1818 if (blkg_tryget(blkg)) {
1822 blkg = blkg->parent;
1830 * bio_associate_blkg_from_css - associate a bio with a specified css
1834 * Associate @bio with the blkg found by combining the css's blkg and the
1835 * request_queue of the @bio. An association failure is handled by walking up
1836 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1837 * and q->root_blkg. This situation only happens when a cgroup is dying and
1838 * then the remaining bios will spill to the closest alive blkg.
1840 * A reference will be taken on the blkg and will be released when @bio is
1843 void bio_associate_blkg_from_css(struct bio *bio,
1844 struct cgroup_subsys_state *css)
1847 blkg_put(bio->bi_blkg);
1849 if (css && css->parent) {
1850 bio->bi_blkg = blkg_tryget_closest(bio, css);
1852 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
1853 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
1856 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1859 * bio_associate_blkg - associate a bio with a blkg
1862 * Associate @bio with the blkg found from the bio's css and request_queue.
1863 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1864 * already associated, the css is reused and association redone as the
1865 * request_queue may have changed.
1867 void bio_associate_blkg(struct bio *bio)
1869 struct cgroup_subsys_state *css;
1874 css = &bio_blkcg(bio)->css;
1878 bio_associate_blkg_from_css(bio, css);
1882 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1885 * bio_clone_blkg_association - clone blkg association from src to dst bio
1886 * @dst: destination bio
1889 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1893 blkg_put(dst->bi_blkg);
1894 blkg_get(src->bi_blkg);
1895 dst->bi_blkg = src->bi_blkg;
1898 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1900 static int blk_cgroup_io_type(struct bio *bio)
1902 if (op_is_discard(bio->bi_opf))
1903 return BLKG_IOSTAT_DISCARD;
1904 if (op_is_write(bio->bi_opf))
1905 return BLKG_IOSTAT_WRITE;
1906 return BLKG_IOSTAT_READ;
1909 void blk_cgroup_bio_start(struct bio *bio)
1911 int rwd = blk_cgroup_io_type(bio), cpu;
1912 struct blkg_iostat_set *bis;
1913 unsigned long flags;
1916 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1917 flags = u64_stats_update_begin_irqsave(&bis->sync);
1920 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1921 * bio and we would have already accounted for the size of the bio.
1923 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1924 bio_set_flag(bio, BIO_CGROUP_ACCT);
1925 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1927 bis->cur.ios[rwd]++;
1929 u64_stats_update_end_irqrestore(&bis->sync, flags);
1930 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1931 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1935 static int __init blkcg_init(void)
1937 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1938 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1939 WQ_UNBOUND | WQ_SYSFS, 0);
1940 if (!blkcg_punt_bio_wq)
1944 subsys_initcall(blkcg_init);
1946 module_param(blkcg_debug_stats, bool, 0644);
1947 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");