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>
35 #define MAX_KEY_LEN 100
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
44 static DEFINE_MUTEX(blkcg_pol_register_mutex);
45 static DEFINE_MUTEX(blkcg_pol_mutex);
47 struct blkcg blkcg_root;
48 EXPORT_SYMBOL_GPL(blkcg_root);
50 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51 EXPORT_SYMBOL_GPL(blkcg_root_css);
53 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
55 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
57 bool blkcg_debug_stats = false;
58 static struct workqueue_struct *blkcg_punt_bio_wq;
60 static bool blkcg_policy_enabled(struct request_queue *q,
61 const struct blkcg_policy *pol)
63 return pol && test_bit(pol->plid, q->blkcg_pols);
67 * blkg_free - free a blkg
70 * Free @blkg which may be partially allocated.
72 static void blkg_free(struct blkcg_gq *blkg)
79 for (i = 0; i < BLKCG_MAX_POLS; i++)
81 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
83 free_percpu(blkg->iostat_cpu);
84 percpu_ref_exit(&blkg->refcnt);
88 static void __blkg_release(struct rcu_head *rcu)
90 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
92 WARN_ON(!bio_list_empty(&blkg->async_bios));
94 /* release the blkcg and parent blkg refs this blkg has been holding */
95 css_put(&blkg->blkcg->css);
97 blkg_put(blkg->parent);
102 * A group is RCU protected, but having an rcu lock does not mean that one
103 * can access all the fields of blkg and assume these are valid. For
104 * example, don't try to follow throtl_data and request queue links.
106 * Having a reference to blkg under an rcu allows accesses to only values
107 * local to groups like group stats and group rate limits.
109 static void blkg_release(struct percpu_ref *ref)
111 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
113 call_rcu(&blkg->rcu_head, __blkg_release);
116 static void blkg_async_bio_workfn(struct work_struct *work)
118 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
120 struct bio_list bios = BIO_EMPTY_LIST;
122 struct blk_plug plug;
123 bool need_plug = false;
125 /* as long as there are pending bios, @blkg can't go away */
126 spin_lock_bh(&blkg->async_bio_lock);
127 bio_list_merge(&bios, &blkg->async_bios);
128 bio_list_init(&blkg->async_bios);
129 spin_unlock_bh(&blkg->async_bio_lock);
131 /* start plug only when bio_list contains at least 2 bios */
132 if (bios.head && bios.head->bi_next) {
134 blk_start_plug(&plug);
136 while ((bio = bio_list_pop(&bios)))
139 blk_finish_plug(&plug);
143 * blkg_alloc - allocate a blkg
144 * @blkcg: block cgroup the new blkg is associated with
145 * @q: request_queue the new blkg is associated with
146 * @gfp_mask: allocation mask to use
148 * Allocate a new blkg assocating @blkcg and @q.
150 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
153 struct blkcg_gq *blkg;
156 /* alloc and init base part */
157 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
161 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
164 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
165 if (!blkg->iostat_cpu)
169 INIT_LIST_HEAD(&blkg->q_node);
170 spin_lock_init(&blkg->async_bio_lock);
171 bio_list_init(&blkg->async_bios);
172 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
175 u64_stats_init(&blkg->iostat.sync);
176 for_each_possible_cpu(cpu)
177 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
179 for (i = 0; i < BLKCG_MAX_POLS; i++) {
180 struct blkcg_policy *pol = blkcg_policy[i];
181 struct blkg_policy_data *pd;
183 if (!blkcg_policy_enabled(q, pol))
186 /* alloc per-policy data and attach it to blkg */
187 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
203 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
204 struct request_queue *q, bool update_hint)
206 struct blkcg_gq *blkg;
209 * Hint didn't match. Look up from the radix tree. Note that the
210 * hint can only be updated under queue_lock as otherwise @blkg
211 * could have already been removed from blkg_tree. The caller is
212 * responsible for grabbing queue_lock if @update_hint.
214 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
215 if (blkg && blkg->q == q) {
217 lockdep_assert_held(&q->queue_lock);
218 rcu_assign_pointer(blkcg->blkg_hint, blkg);
225 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
228 * If @new_blkg is %NULL, this function tries to allocate a new one as
229 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
231 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
232 struct request_queue *q,
233 struct blkcg_gq *new_blkg)
235 struct blkcg_gq *blkg;
238 WARN_ON_ONCE(!rcu_read_lock_held());
239 lockdep_assert_held(&q->queue_lock);
241 /* request_queue is dying, do not create/recreate a blkg */
242 if (blk_queue_dying(q)) {
247 /* blkg holds a reference to blkcg */
248 if (!css_tryget_online(&blkcg->css)) {
255 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
256 if (unlikely(!new_blkg)) {
264 if (blkcg_parent(blkcg)) {
265 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
266 if (WARN_ON_ONCE(!blkg->parent)) {
270 blkg_get(blkg->parent);
273 /* invoke per-policy init */
274 for (i = 0; i < BLKCG_MAX_POLS; i++) {
275 struct blkcg_policy *pol = blkcg_policy[i];
277 if (blkg->pd[i] && pol->pd_init_fn)
278 pol->pd_init_fn(blkg->pd[i]);
282 spin_lock(&blkcg->lock);
283 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
285 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
286 list_add(&blkg->q_node, &q->blkg_list);
288 for (i = 0; i < BLKCG_MAX_POLS; i++) {
289 struct blkcg_policy *pol = blkcg_policy[i];
291 if (blkg->pd[i] && pol->pd_online_fn)
292 pol->pd_online_fn(blkg->pd[i]);
296 spin_unlock(&blkcg->lock);
301 /* @blkg failed fully initialized, use the usual release path */
306 css_put(&blkcg->css);
313 * blkg_lookup_create - lookup blkg, try to create one if not there
314 * @blkcg: blkcg of interest
315 * @q: request_queue of interest
317 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
318 * create one. blkg creation is performed recursively from blkcg_root such
319 * that all non-root blkg's have access to the parent blkg. This function
320 * should be called under RCU read lock and takes @q->queue_lock.
322 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
325 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
326 struct request_queue *q)
328 struct blkcg_gq *blkg;
331 WARN_ON_ONCE(!rcu_read_lock_held());
333 blkg = blkg_lookup(blkcg, q);
337 spin_lock_irqsave(&q->queue_lock, flags);
338 blkg = __blkg_lookup(blkcg, q, true);
343 * Create blkgs walking down from blkcg_root to @blkcg, so that all
344 * non-root blkgs have access to their parents. Returns the closest
345 * blkg to the intended blkg should blkg_create() fail.
348 struct blkcg *pos = blkcg;
349 struct blkcg *parent = blkcg_parent(blkcg);
350 struct blkcg_gq *ret_blkg = q->root_blkg;
353 blkg = __blkg_lookup(parent, q, false);
355 /* remember closest blkg */
360 parent = blkcg_parent(parent);
363 blkg = blkg_create(pos, q, NULL);
373 spin_unlock_irqrestore(&q->queue_lock, flags);
377 static void blkg_destroy(struct blkcg_gq *blkg)
379 struct blkcg *blkcg = blkg->blkcg;
382 lockdep_assert_held(&blkg->q->queue_lock);
383 lockdep_assert_held(&blkcg->lock);
385 /* Something wrong if we are trying to remove same group twice */
386 WARN_ON_ONCE(list_empty(&blkg->q_node));
387 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
389 for (i = 0; i < BLKCG_MAX_POLS; i++) {
390 struct blkcg_policy *pol = blkcg_policy[i];
392 if (blkg->pd[i] && pol->pd_offline_fn)
393 pol->pd_offline_fn(blkg->pd[i]);
396 blkg->online = false;
398 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
399 list_del_init(&blkg->q_node);
400 hlist_del_init_rcu(&blkg->blkcg_node);
403 * Both setting lookup hint to and clearing it from @blkg are done
404 * under queue_lock. If it's not pointing to @blkg now, it never
405 * will. Hint assignment itself can race safely.
407 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
408 rcu_assign_pointer(blkcg->blkg_hint, NULL);
411 * Put the reference taken at the time of creation so that when all
412 * queues are gone, group can be destroyed.
414 percpu_ref_kill(&blkg->refcnt);
418 * blkg_destroy_all - destroy all blkgs associated with a request_queue
419 * @q: request_queue of interest
421 * Destroy all blkgs associated with @q.
423 static void blkg_destroy_all(struct request_queue *q)
425 struct blkcg_gq *blkg, *n;
427 spin_lock_irq(&q->queue_lock);
428 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
429 struct blkcg *blkcg = blkg->blkcg;
431 spin_lock(&blkcg->lock);
433 spin_unlock(&blkcg->lock);
437 spin_unlock_irq(&q->queue_lock);
440 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
441 struct cftype *cftype, u64 val)
443 struct blkcg *blkcg = css_to_blkcg(css);
444 struct blkcg_gq *blkg;
447 mutex_lock(&blkcg_pol_mutex);
448 spin_lock_irq(&blkcg->lock);
451 * Note that stat reset is racy - it doesn't synchronize against
452 * stat updates. This is a debug feature which shouldn't exist
453 * anyway. If you get hit by a race, retry.
455 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
456 for_each_possible_cpu(cpu) {
457 struct blkg_iostat_set *bis =
458 per_cpu_ptr(blkg->iostat_cpu, cpu);
459 memset(bis, 0, sizeof(*bis));
461 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
463 for (i = 0; i < BLKCG_MAX_POLS; i++) {
464 struct blkcg_policy *pol = blkcg_policy[i];
466 if (blkg->pd[i] && pol->pd_reset_stats_fn)
467 pol->pd_reset_stats_fn(blkg->pd[i]);
471 spin_unlock_irq(&blkcg->lock);
472 mutex_unlock(&blkcg_pol_mutex);
476 const char *blkg_dev_name(struct blkcg_gq *blkg)
478 /* some drivers (floppy) instantiate a queue w/o disk registered */
479 if (blkg->q->backing_dev_info->dev)
480 return bdi_dev_name(blkg->q->backing_dev_info);
485 * blkcg_print_blkgs - helper for printing per-blkg data
486 * @sf: seq_file to print to
487 * @blkcg: blkcg of interest
488 * @prfill: fill function to print out a blkg
489 * @pol: policy in question
490 * @data: data to be passed to @prfill
491 * @show_total: to print out sum of prfill return values or not
493 * This function invokes @prfill on each blkg of @blkcg if pd for the
494 * policy specified by @pol exists. @prfill is invoked with @sf, the
495 * policy data and @data and the matching queue lock held. If @show_total
496 * is %true, the sum of the return values from @prfill is printed with
497 * "Total" label at the end.
499 * This is to be used to construct print functions for
500 * cftype->read_seq_string method.
502 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
503 u64 (*prfill)(struct seq_file *,
504 struct blkg_policy_data *, int),
505 const struct blkcg_policy *pol, int data,
508 struct blkcg_gq *blkg;
512 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
513 spin_lock_irq(&blkg->q->queue_lock);
514 if (blkcg_policy_enabled(blkg->q, pol))
515 total += prfill(sf, blkg->pd[pol->plid], data);
516 spin_unlock_irq(&blkg->q->queue_lock);
521 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
523 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
526 * __blkg_prfill_u64 - prfill helper for a single u64 value
527 * @sf: seq_file to print to
528 * @pd: policy private data of interest
531 * Print @v to @sf for the device assocaited with @pd.
533 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
535 const char *dname = blkg_dev_name(pd->blkg);
540 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
543 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
545 /* Performs queue bypass and policy enabled checks then looks up blkg. */
546 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
547 const struct blkcg_policy *pol,
548 struct request_queue *q)
550 WARN_ON_ONCE(!rcu_read_lock_held());
551 lockdep_assert_held(&q->queue_lock);
553 if (!blkcg_policy_enabled(q, pol))
554 return ERR_PTR(-EOPNOTSUPP);
555 return __blkg_lookup(blkcg, q, true /* update_hint */);
559 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
560 * @inputp: input string pointer
562 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
563 * from @input and get and return the matching bdev. *@inputp is
564 * updated to point past the device node prefix. Returns an ERR_PTR()
567 * Use this function iff blkg_conf_prep() can't be used for some reason.
569 struct block_device *blkcg_conf_open_bdev(char **inputp)
571 char *input = *inputp;
572 unsigned int major, minor;
573 struct block_device *bdev;
576 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
577 return ERR_PTR(-EINVAL);
580 if (!isspace(*input))
581 return ERR_PTR(-EINVAL);
582 input = skip_spaces(input);
584 bdev = blkdev_get_no_open(MKDEV(major, minor));
586 return ERR_PTR(-ENODEV);
587 if (bdev_is_partition(bdev)) {
588 blkdev_put_no_open(bdev);
589 return ERR_PTR(-ENODEV);
597 * blkg_conf_prep - parse and prepare for per-blkg config update
598 * @blkcg: target block cgroup
599 * @pol: target policy
600 * @input: input string
601 * @ctx: blkg_conf_ctx to be filled
603 * Parse per-blkg config update from @input and initialize @ctx with the
604 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
605 * part of @input following MAJ:MIN. This function returns with RCU read
606 * lock and queue lock held and must be paired with blkg_conf_finish().
608 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
609 char *input, struct blkg_conf_ctx *ctx)
610 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
612 struct block_device *bdev;
613 struct request_queue *q;
614 struct blkcg_gq *blkg;
617 bdev = blkcg_conf_open_bdev(&input);
619 return PTR_ERR(bdev);
621 q = bdev->bd_disk->queue;
624 spin_lock_irq(&q->queue_lock);
626 blkg = blkg_lookup_check(blkcg, pol, q);
636 * Create blkgs walking down from blkcg_root to @blkcg, so that all
637 * non-root blkgs have access to their parents.
640 struct blkcg *pos = blkcg;
641 struct blkcg *parent;
642 struct blkcg_gq *new_blkg;
644 parent = blkcg_parent(blkcg);
645 while (parent && !__blkg_lookup(parent, q, false)) {
647 parent = blkcg_parent(parent);
650 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
651 spin_unlock_irq(&q->queue_lock);
654 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
655 if (unlikely(!new_blkg)) {
660 if (radix_tree_preload(GFP_KERNEL)) {
667 spin_lock_irq(&q->queue_lock);
669 blkg = blkg_lookup_check(pos, pol, q);
679 blkg = blkg_create(pos, q, new_blkg);
686 radix_tree_preload_end();
698 radix_tree_preload_end();
700 spin_unlock_irq(&q->queue_lock);
703 blkdev_put_no_open(bdev);
705 * If queue was bypassing, we should retry. Do so after a
706 * short msleep(). It isn't strictly necessary but queue
707 * can be bypassing for some time and it's always nice to
708 * avoid busy looping.
712 ret = restart_syscall();
716 EXPORT_SYMBOL_GPL(blkg_conf_prep);
719 * blkg_conf_finish - finish up per-blkg config update
720 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
722 * Finish up after per-blkg config update. This function must be paired
723 * with blkg_conf_prep().
725 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
726 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
728 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
730 blkdev_put_no_open(ctx->bdev);
732 EXPORT_SYMBOL_GPL(blkg_conf_finish);
734 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
738 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
739 dst->bytes[i] = src->bytes[i];
740 dst->ios[i] = src->ios[i];
744 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
748 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
749 dst->bytes[i] += src->bytes[i];
750 dst->ios[i] += src->ios[i];
754 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
758 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
759 dst->bytes[i] -= src->bytes[i];
760 dst->ios[i] -= src->ios[i];
764 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
766 struct blkcg *blkcg = css_to_blkcg(css);
767 struct blkcg_gq *blkg;
771 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
772 struct blkcg_gq *parent = blkg->parent;
773 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
774 struct blkg_iostat cur, delta;
777 /* fetch the current per-cpu values */
779 seq = u64_stats_fetch_begin(&bisc->sync);
780 blkg_iostat_set(&cur, &bisc->cur);
781 } while (u64_stats_fetch_retry(&bisc->sync, seq));
783 /* propagate percpu delta to global */
784 u64_stats_update_begin(&blkg->iostat.sync);
785 blkg_iostat_set(&delta, &cur);
786 blkg_iostat_sub(&delta, &bisc->last);
787 blkg_iostat_add(&blkg->iostat.cur, &delta);
788 blkg_iostat_add(&bisc->last, &delta);
789 u64_stats_update_end(&blkg->iostat.sync);
791 /* propagate global delta to parent */
793 u64_stats_update_begin(&parent->iostat.sync);
794 blkg_iostat_set(&delta, &blkg->iostat.cur);
795 blkg_iostat_sub(&delta, &blkg->iostat.last);
796 blkg_iostat_add(&parent->iostat.cur, &delta);
797 blkg_iostat_add(&blkg->iostat.last, &delta);
798 u64_stats_update_end(&parent->iostat.sync);
806 * The rstat algorithms intentionally don't handle the root cgroup to avoid
807 * incurring overhead when no cgroups are defined. For that reason,
808 * cgroup_rstat_flush in blkcg_print_stat does not actually fill out the
809 * iostat in the root cgroup's blkcg_gq.
811 * However, we would like to re-use the printing code between the root and
812 * non-root cgroups to the extent possible. For that reason, we simulate
813 * flushing the root cgroup's stats by explicitly filling in the iostat
814 * with disk level statistics.
816 static void blkcg_fill_root_iostats(void)
818 struct class_dev_iter iter;
821 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
822 while ((dev = class_dev_iter_next(&iter))) {
823 struct block_device *bdev = dev_to_bdev(dev);
824 struct blkcg_gq *blkg =
825 blk_queue_root_blkg(bdev->bd_disk->queue);
826 struct blkg_iostat tmp;
829 memset(&tmp, 0, sizeof(tmp));
830 for_each_possible_cpu(cpu) {
831 struct disk_stats *cpu_dkstats;
833 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
834 tmp.ios[BLKG_IOSTAT_READ] +=
835 cpu_dkstats->ios[STAT_READ];
836 tmp.ios[BLKG_IOSTAT_WRITE] +=
837 cpu_dkstats->ios[STAT_WRITE];
838 tmp.ios[BLKG_IOSTAT_DISCARD] +=
839 cpu_dkstats->ios[STAT_DISCARD];
840 // convert sectors to bytes
841 tmp.bytes[BLKG_IOSTAT_READ] +=
842 cpu_dkstats->sectors[STAT_READ] << 9;
843 tmp.bytes[BLKG_IOSTAT_WRITE] +=
844 cpu_dkstats->sectors[STAT_WRITE] << 9;
845 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
846 cpu_dkstats->sectors[STAT_DISCARD] << 9;
848 u64_stats_update_begin(&blkg->iostat.sync);
849 blkg_iostat_set(&blkg->iostat.cur, &tmp);
850 u64_stats_update_end(&blkg->iostat.sync);
855 static int blkcg_print_stat(struct seq_file *sf, void *v)
857 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
858 struct blkcg_gq *blkg;
860 if (!seq_css(sf)->parent)
861 blkcg_fill_root_iostats();
863 cgroup_rstat_flush(blkcg->css.cgroup);
867 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
868 struct blkg_iostat_set *bis = &blkg->iostat;
871 u64 rbytes, wbytes, rios, wios, dbytes, dios;
872 size_t size = seq_get_buf(sf, &buf), off = 0;
874 bool has_stats = false;
877 spin_lock_irq(&blkg->q->queue_lock);
882 dname = blkg_dev_name(blkg);
887 * Hooray string manipulation, count is the size written NOT
888 * INCLUDING THE \0, so size is now count+1 less than what we
889 * had before, but we want to start writing the next bit from
890 * the \0 so we only add count to buf.
892 off += scnprintf(buf+off, size-off, "%s ", dname);
895 seq = u64_stats_fetch_begin(&bis->sync);
897 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
898 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
899 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
900 rios = bis->cur.ios[BLKG_IOSTAT_READ];
901 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
902 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
903 } while (u64_stats_fetch_retry(&bis->sync, seq));
905 if (rbytes || wbytes || rios || wios) {
907 off += scnprintf(buf+off, size-off,
908 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
909 rbytes, wbytes, rios, wios,
913 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
915 off += scnprintf(buf+off, size-off,
916 " use_delay=%d delay_nsec=%llu",
917 atomic_read(&blkg->use_delay),
918 (unsigned long long)atomic64_read(&blkg->delay_nsec));
921 for (i = 0; i < BLKCG_MAX_POLS; i++) {
922 struct blkcg_policy *pol = blkcg_policy[i];
925 if (!blkg->pd[i] || !pol->pd_stat_fn)
928 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
935 if (off < size - 1) {
936 off += scnprintf(buf+off, size-off, "\n");
943 spin_unlock_irq(&blkg->q->queue_lock);
950 static struct cftype blkcg_files[] = {
953 .seq_show = blkcg_print_stat,
958 static struct cftype blkcg_legacy_files[] = {
960 .name = "reset_stats",
961 .write_u64 = blkcg_reset_stats,
967 * blkcg destruction is a three-stage process.
969 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
970 * which offlines writeback. Here we tie the next stage of blkg destruction
971 * to the completion of writeback associated with the blkcg. This lets us
972 * avoid punting potentially large amounts of outstanding writeback to root
973 * while maintaining any ongoing policies. The next stage is triggered when
974 * the nr_cgwbs count goes to zero.
976 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
977 * and handles the destruction of blkgs. Here the css reference held by
978 * the blkg is put back eventually allowing blkcg_css_free() to be called.
979 * This work may occur in cgwb_release_workfn() on the cgwb_release
980 * workqueue. Any submitted ios that fail to get the blkg ref will be
981 * punted to the root_blkg.
983 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
984 * This finally frees the blkcg.
988 * blkcg_css_offline - cgroup css_offline callback
989 * @css: css of interest
991 * This function is called when @css is about to go away. Here the cgwbs are
992 * offlined first and only once writeback associated with the blkcg has
993 * finished do we start step 2 (see above).
995 static void blkcg_css_offline(struct cgroup_subsys_state *css)
997 struct blkcg *blkcg = css_to_blkcg(css);
999 /* this prevents anyone from attaching or migrating to this blkcg */
1000 wb_blkcg_offline(blkcg);
1002 /* put the base online pin allowing step 2 to be triggered */
1003 blkcg_unpin_online(blkcg);
1007 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1008 * @blkcg: blkcg of interest
1010 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1011 * is nested inside q lock, this function performs reverse double lock dancing.
1012 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1013 * blkcg_css_free to eventually be called.
1015 * This is the blkcg counterpart of ioc_release_fn().
1017 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1019 spin_lock_irq(&blkcg->lock);
1021 while (!hlist_empty(&blkcg->blkg_list)) {
1022 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1023 struct blkcg_gq, blkcg_node);
1024 struct request_queue *q = blkg->q;
1026 if (spin_trylock(&q->queue_lock)) {
1028 spin_unlock(&q->queue_lock);
1030 spin_unlock_irq(&blkcg->lock);
1032 spin_lock_irq(&blkcg->lock);
1036 spin_unlock_irq(&blkcg->lock);
1039 static void blkcg_css_free(struct cgroup_subsys_state *css)
1041 struct blkcg *blkcg = css_to_blkcg(css);
1044 mutex_lock(&blkcg_pol_mutex);
1046 list_del(&blkcg->all_blkcgs_node);
1048 for (i = 0; i < BLKCG_MAX_POLS; i++)
1050 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1052 mutex_unlock(&blkcg_pol_mutex);
1057 static struct cgroup_subsys_state *
1058 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1060 struct blkcg *blkcg;
1061 struct cgroup_subsys_state *ret;
1064 mutex_lock(&blkcg_pol_mutex);
1067 blkcg = &blkcg_root;
1069 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1071 ret = ERR_PTR(-ENOMEM);
1076 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1077 struct blkcg_policy *pol = blkcg_policy[i];
1078 struct blkcg_policy_data *cpd;
1081 * If the policy hasn't been attached yet, wait for it
1082 * to be attached before doing anything else. Otherwise,
1083 * check if the policy requires any specific per-cgroup
1084 * data: if it does, allocate and initialize it.
1086 if (!pol || !pol->cpd_alloc_fn)
1089 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1091 ret = ERR_PTR(-ENOMEM);
1094 blkcg->cpd[i] = cpd;
1097 if (pol->cpd_init_fn)
1098 pol->cpd_init_fn(cpd);
1101 spin_lock_init(&blkcg->lock);
1102 refcount_set(&blkcg->online_pin, 1);
1103 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1104 INIT_HLIST_HEAD(&blkcg->blkg_list);
1105 #ifdef CONFIG_CGROUP_WRITEBACK
1106 INIT_LIST_HEAD(&blkcg->cgwb_list);
1108 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1110 mutex_unlock(&blkcg_pol_mutex);
1114 for (i--; i >= 0; i--)
1116 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1118 if (blkcg != &blkcg_root)
1121 mutex_unlock(&blkcg_pol_mutex);
1125 static int blkcg_css_online(struct cgroup_subsys_state *css)
1127 struct blkcg *blkcg = css_to_blkcg(css);
1128 struct blkcg *parent = blkcg_parent(blkcg);
1131 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1132 * don't go offline while cgwbs are still active on them. Pin the
1133 * parent so that offline always happens towards the root.
1136 blkcg_pin_online(parent);
1141 * blkcg_init_queue - initialize blkcg part of request queue
1142 * @q: request_queue to initialize
1144 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1145 * part of new request_queue @q.
1148 * 0 on success, -errno on failure.
1150 int blkcg_init_queue(struct request_queue *q)
1152 struct blkcg_gq *new_blkg, *blkg;
1156 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1160 preloaded = !radix_tree_preload(GFP_KERNEL);
1162 /* Make sure the root blkg exists. */
1164 spin_lock_irq(&q->queue_lock);
1165 blkg = blkg_create(&blkcg_root, q, new_blkg);
1168 q->root_blkg = blkg;
1169 spin_unlock_irq(&q->queue_lock);
1173 radix_tree_preload_end();
1175 ret = blk_throtl_init(q);
1177 goto err_destroy_all;
1179 ret = blk_iolatency_init(q);
1182 goto err_destroy_all;
1187 blkg_destroy_all(q);
1190 spin_unlock_irq(&q->queue_lock);
1193 radix_tree_preload_end();
1194 return PTR_ERR(blkg);
1198 * blkcg_exit_queue - exit and release blkcg part of request_queue
1199 * @q: request_queue being released
1201 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1203 void blkcg_exit_queue(struct request_queue *q)
1205 blkg_destroy_all(q);
1210 * We cannot support shared io contexts, as we have no mean to support
1211 * two tasks with the same ioc in two different groups without major rework
1212 * of the main cic data structures. For now we allow a task to change
1213 * its cgroup only if it's the only owner of its ioc.
1215 static int blkcg_can_attach(struct cgroup_taskset *tset)
1217 struct task_struct *task;
1218 struct cgroup_subsys_state *dst_css;
1219 struct io_context *ioc;
1222 /* task_lock() is needed to avoid races with exit_io_context() */
1223 cgroup_taskset_for_each(task, dst_css, tset) {
1225 ioc = task->io_context;
1226 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1235 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1239 mutex_lock(&blkcg_pol_mutex);
1241 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1242 struct blkcg_policy *pol = blkcg_policy[i];
1243 struct blkcg *blkcg;
1245 if (!pol || !pol->cpd_bind_fn)
1248 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1249 if (blkcg->cpd[pol->plid])
1250 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1252 mutex_unlock(&blkcg_pol_mutex);
1255 static void blkcg_exit(struct task_struct *tsk)
1257 if (tsk->throttle_queue)
1258 blk_put_queue(tsk->throttle_queue);
1259 tsk->throttle_queue = NULL;
1262 struct cgroup_subsys io_cgrp_subsys = {
1263 .css_alloc = blkcg_css_alloc,
1264 .css_online = blkcg_css_online,
1265 .css_offline = blkcg_css_offline,
1266 .css_free = blkcg_css_free,
1267 .can_attach = blkcg_can_attach,
1268 .css_rstat_flush = blkcg_rstat_flush,
1270 .dfl_cftypes = blkcg_files,
1271 .legacy_cftypes = blkcg_legacy_files,
1272 .legacy_name = "blkio",
1276 * This ensures that, if available, memcg is automatically enabled
1277 * together on the default hierarchy so that the owner cgroup can
1278 * be retrieved from writeback pages.
1280 .depends_on = 1 << memory_cgrp_id,
1283 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1286 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1287 * @q: request_queue of interest
1288 * @pol: blkcg policy to activate
1290 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1291 * bypass mode to populate its blkgs with policy_data for @pol.
1293 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1294 * from IO path. Update of each blkg is protected by both queue and blkcg
1295 * locks so that holding either lock and testing blkcg_policy_enabled() is
1296 * always enough for dereferencing policy data.
1298 * The caller is responsible for synchronizing [de]activations and policy
1299 * [un]registerations. Returns 0 on success, -errno on failure.
1301 int blkcg_activate_policy(struct request_queue *q,
1302 const struct blkcg_policy *pol)
1304 struct blkg_policy_data *pd_prealloc = NULL;
1305 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1308 if (blkcg_policy_enabled(q, pol))
1312 blk_mq_freeze_queue(q);
1314 spin_lock_irq(&q->queue_lock);
1316 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1317 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1318 struct blkg_policy_data *pd;
1320 if (blkg->pd[pol->plid])
1323 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1324 if (blkg == pinned_blkg) {
1328 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1334 * GFP_NOWAIT failed. Free the existing one and
1335 * prealloc for @blkg w/ GFP_KERNEL.
1338 blkg_put(pinned_blkg);
1342 spin_unlock_irq(&q->queue_lock);
1345 pol->pd_free_fn(pd_prealloc);
1346 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1354 blkg->pd[pol->plid] = pd;
1356 pd->plid = pol->plid;
1359 /* all allocated, init in the same order */
1360 if (pol->pd_init_fn)
1361 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1362 pol->pd_init_fn(blkg->pd[pol->plid]);
1364 __set_bit(pol->plid, q->blkcg_pols);
1367 spin_unlock_irq(&q->queue_lock);
1370 blk_mq_unfreeze_queue(q);
1372 blkg_put(pinned_blkg);
1374 pol->pd_free_fn(pd_prealloc);
1378 /* alloc failed, nothing's initialized yet, free everything */
1379 spin_lock_irq(&q->queue_lock);
1380 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1381 if (blkg->pd[pol->plid]) {
1382 pol->pd_free_fn(blkg->pd[pol->plid]);
1383 blkg->pd[pol->plid] = NULL;
1386 spin_unlock_irq(&q->queue_lock);
1390 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1393 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1394 * @q: request_queue of interest
1395 * @pol: blkcg policy to deactivate
1397 * Deactivate @pol on @q. Follows the same synchronization rules as
1398 * blkcg_activate_policy().
1400 void blkcg_deactivate_policy(struct request_queue *q,
1401 const struct blkcg_policy *pol)
1403 struct blkcg_gq *blkg;
1405 if (!blkcg_policy_enabled(q, pol))
1409 blk_mq_freeze_queue(q);
1411 spin_lock_irq(&q->queue_lock);
1413 __clear_bit(pol->plid, q->blkcg_pols);
1415 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1416 if (blkg->pd[pol->plid]) {
1417 if (pol->pd_offline_fn)
1418 pol->pd_offline_fn(blkg->pd[pol->plid]);
1419 pol->pd_free_fn(blkg->pd[pol->plid]);
1420 blkg->pd[pol->plid] = NULL;
1424 spin_unlock_irq(&q->queue_lock);
1427 blk_mq_unfreeze_queue(q);
1429 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1432 * blkcg_policy_register - register a blkcg policy
1433 * @pol: blkcg policy to register
1435 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1436 * successful registration. Returns 0 on success and -errno on failure.
1438 int blkcg_policy_register(struct blkcg_policy *pol)
1440 struct blkcg *blkcg;
1443 mutex_lock(&blkcg_pol_register_mutex);
1444 mutex_lock(&blkcg_pol_mutex);
1446 /* find an empty slot */
1448 for (i = 0; i < BLKCG_MAX_POLS; i++)
1449 if (!blkcg_policy[i])
1451 if (i >= BLKCG_MAX_POLS) {
1452 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1456 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1457 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1458 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1463 blkcg_policy[pol->plid] = pol;
1465 /* allocate and install cpd's */
1466 if (pol->cpd_alloc_fn) {
1467 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1468 struct blkcg_policy_data *cpd;
1470 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1474 blkcg->cpd[pol->plid] = cpd;
1476 cpd->plid = pol->plid;
1477 if (pol->cpd_init_fn)
1478 pol->cpd_init_fn(cpd);
1482 mutex_unlock(&blkcg_pol_mutex);
1484 /* everything is in place, add intf files for the new policy */
1485 if (pol->dfl_cftypes)
1486 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1488 if (pol->legacy_cftypes)
1489 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1490 pol->legacy_cftypes));
1491 mutex_unlock(&blkcg_pol_register_mutex);
1495 if (pol->cpd_free_fn) {
1496 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1497 if (blkcg->cpd[pol->plid]) {
1498 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1499 blkcg->cpd[pol->plid] = NULL;
1503 blkcg_policy[pol->plid] = NULL;
1505 mutex_unlock(&blkcg_pol_mutex);
1506 mutex_unlock(&blkcg_pol_register_mutex);
1509 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1512 * blkcg_policy_unregister - unregister a blkcg policy
1513 * @pol: blkcg policy to unregister
1515 * Undo blkcg_policy_register(@pol). Might sleep.
1517 void blkcg_policy_unregister(struct blkcg_policy *pol)
1519 struct blkcg *blkcg;
1521 mutex_lock(&blkcg_pol_register_mutex);
1523 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1526 /* kill the intf files first */
1527 if (pol->dfl_cftypes)
1528 cgroup_rm_cftypes(pol->dfl_cftypes);
1529 if (pol->legacy_cftypes)
1530 cgroup_rm_cftypes(pol->legacy_cftypes);
1532 /* remove cpds and unregister */
1533 mutex_lock(&blkcg_pol_mutex);
1535 if (pol->cpd_free_fn) {
1536 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1537 if (blkcg->cpd[pol->plid]) {
1538 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1539 blkcg->cpd[pol->plid] = NULL;
1543 blkcg_policy[pol->plid] = NULL;
1545 mutex_unlock(&blkcg_pol_mutex);
1547 mutex_unlock(&blkcg_pol_register_mutex);
1549 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1551 bool __blkcg_punt_bio_submit(struct bio *bio)
1553 struct blkcg_gq *blkg = bio->bi_blkg;
1555 /* consume the flag first */
1556 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1558 /* never bounce for the root cgroup */
1562 spin_lock_bh(&blkg->async_bio_lock);
1563 bio_list_add(&blkg->async_bios, bio);
1564 spin_unlock_bh(&blkg->async_bio_lock);
1566 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1571 * Scale the accumulated delay based on how long it has been since we updated
1572 * the delay. We only call this when we are adding delay, in case it's been a
1573 * while since we added delay, and when we are checking to see if we need to
1574 * delay a task, to account for any delays that may have occurred.
1576 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1578 u64 old = atomic64_read(&blkg->delay_start);
1580 /* negative use_delay means no scaling, see blkcg_set_delay() */
1581 if (atomic_read(&blkg->use_delay) < 0)
1585 * We only want to scale down every second. The idea here is that we
1586 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1587 * time window. We only want to throttle tasks for recent delay that
1588 * has occurred, in 1 second time windows since that's the maximum
1589 * things can be throttled. We save the current delay window in
1590 * blkg->last_delay so we know what amount is still left to be charged
1591 * to the blkg from this point onward. blkg->last_use keeps track of
1592 * the use_delay counter. The idea is if we're unthrottling the blkg we
1593 * are ok with whatever is happening now, and we can take away more of
1594 * the accumulated delay as we've already throttled enough that
1595 * everybody is happy with their IO latencies.
1597 if (time_before64(old + NSEC_PER_SEC, now) &&
1598 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1599 u64 cur = atomic64_read(&blkg->delay_nsec);
1600 u64 sub = min_t(u64, blkg->last_delay, now - old);
1601 int cur_use = atomic_read(&blkg->use_delay);
1604 * We've been unthrottled, subtract a larger chunk of our
1605 * accumulated delay.
1607 if (cur_use < blkg->last_use)
1608 sub = max_t(u64, sub, blkg->last_delay >> 1);
1611 * This shouldn't happen, but handle it anyway. Our delay_nsec
1612 * should only ever be growing except here where we subtract out
1613 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1614 * rather not end up with negative numbers.
1616 if (unlikely(cur < sub)) {
1617 atomic64_set(&blkg->delay_nsec, 0);
1618 blkg->last_delay = 0;
1620 atomic64_sub(sub, &blkg->delay_nsec);
1621 blkg->last_delay = cur - sub;
1623 blkg->last_use = cur_use;
1628 * This is called when we want to actually walk up the hierarchy and check to
1629 * see if we need to throttle, and then actually throttle if there is some
1630 * accumulated delay. This should only be called upon return to user space so
1631 * we're not holding some lock that would induce a priority inversion.
1633 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1635 unsigned long pflags;
1637 u64 now = ktime_to_ns(ktime_get());
1642 while (blkg->parent) {
1643 int use_delay = atomic_read(&blkg->use_delay);
1648 blkcg_scale_delay(blkg, now);
1649 this_delay = atomic64_read(&blkg->delay_nsec);
1650 if (this_delay > delay_nsec) {
1651 delay_nsec = this_delay;
1652 clamp = use_delay > 0;
1655 blkg = blkg->parent;
1662 * Let's not sleep for all eternity if we've amassed a huge delay.
1663 * Swapping or metadata IO can accumulate 10's of seconds worth of
1664 * delay, and we want userspace to be able to do _something_ so cap the
1665 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1666 * tasks will be delayed for 0.25 second for every syscall. If
1667 * blkcg_set_delay() was used as indicated by negative use_delay, the
1668 * caller is responsible for regulating the range.
1671 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1674 psi_memstall_enter(&pflags);
1676 exp = ktime_add_ns(now, delay_nsec);
1677 tok = io_schedule_prepare();
1679 __set_current_state(TASK_KILLABLE);
1680 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1682 } while (!fatal_signal_pending(current));
1683 io_schedule_finish(tok);
1686 psi_memstall_leave(&pflags);
1690 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1692 * This is only called if we've been marked with set_notify_resume(). Obviously
1693 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1694 * check to see if current->throttle_queue is set and if not this doesn't do
1695 * anything. This should only ever be called by the resume code, it's not meant
1696 * to be called by people willy-nilly as it will actually do the work to
1697 * throttle the task if it is setup for throttling.
1699 void blkcg_maybe_throttle_current(void)
1701 struct request_queue *q = current->throttle_queue;
1702 struct cgroup_subsys_state *css;
1703 struct blkcg *blkcg;
1704 struct blkcg_gq *blkg;
1705 bool use_memdelay = current->use_memdelay;
1710 current->throttle_queue = NULL;
1711 current->use_memdelay = false;
1714 css = kthread_blkcg();
1716 blkcg = css_to_blkcg(css);
1718 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1722 blkg = blkg_lookup(blkcg, q);
1725 if (!blkg_tryget(blkg))
1729 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1739 * blkcg_schedule_throttle - this task needs to check for throttling
1740 * @q: the request queue IO was submitted on
1741 * @use_memdelay: do we charge this to memory delay for PSI
1743 * This is called by the IO controller when we know there's delay accumulated
1744 * for the blkg for this task. We do not pass the blkg because there are places
1745 * we call this that may not have that information, the swapping code for
1746 * instance will only have a request_queue at that point. This set's the
1747 * notify_resume for the task to check and see if it requires throttling before
1748 * returning to user space.
1750 * We will only schedule once per syscall. You can call this over and over
1751 * again and it will only do the check once upon return to user space, and only
1752 * throttle once. If the task needs to be throttled again it'll need to be
1753 * re-set at the next time we see the task.
1755 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1757 if (unlikely(current->flags & PF_KTHREAD))
1760 if (!blk_get_queue(q))
1763 if (current->throttle_queue)
1764 blk_put_queue(current->throttle_queue);
1765 current->throttle_queue = q;
1767 current->use_memdelay = use_memdelay;
1768 set_notify_resume(current);
1772 * blkcg_add_delay - add delay to this blkg
1773 * @blkg: blkg of interest
1774 * @now: the current time in nanoseconds
1775 * @delta: how many nanoseconds of delay to add
1777 * Charge @delta to the blkg's current delay accumulation. This is used to
1778 * throttle tasks if an IO controller thinks we need more throttling.
1780 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1782 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1784 blkcg_scale_delay(blkg, now);
1785 atomic64_add(delta, &blkg->delay_nsec);
1789 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1793 * As the failure mode here is to walk up the blkg tree, this ensure that the
1794 * blkg->parent pointers are always valid. This returns the blkg that it ended
1795 * up taking a reference on or %NULL if no reference was taken.
1797 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1798 struct cgroup_subsys_state *css)
1800 struct blkcg_gq *blkg, *ret_blkg = NULL;
1803 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_disk->queue);
1805 if (blkg_tryget(blkg)) {
1809 blkg = blkg->parent;
1817 * bio_associate_blkg_from_css - associate a bio with a specified css
1821 * Associate @bio with the blkg found by combining the css's blkg and the
1822 * request_queue of the @bio. An association failure is handled by walking up
1823 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1824 * and q->root_blkg. This situation only happens when a cgroup is dying and
1825 * then the remaining bios will spill to the closest alive blkg.
1827 * A reference will be taken on the blkg and will be released when @bio is
1830 void bio_associate_blkg_from_css(struct bio *bio,
1831 struct cgroup_subsys_state *css)
1834 blkg_put(bio->bi_blkg);
1836 if (css && css->parent) {
1837 bio->bi_blkg = blkg_tryget_closest(bio, css);
1839 blkg_get(bio->bi_disk->queue->root_blkg);
1840 bio->bi_blkg = bio->bi_disk->queue->root_blkg;
1843 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1846 * bio_associate_blkg - associate a bio with a blkg
1849 * Associate @bio with the blkg found from the bio's css and request_queue.
1850 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1851 * already associated, the css is reused and association redone as the
1852 * request_queue may have changed.
1854 void bio_associate_blkg(struct bio *bio)
1856 struct cgroup_subsys_state *css;
1861 css = &bio_blkcg(bio)->css;
1865 bio_associate_blkg_from_css(bio, css);
1869 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1872 * bio_clone_blkg_association - clone blkg association from src to dst bio
1873 * @dst: destination bio
1876 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1880 blkg_put(dst->bi_blkg);
1881 blkg_get(src->bi_blkg);
1882 dst->bi_blkg = src->bi_blkg;
1885 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1887 static int blk_cgroup_io_type(struct bio *bio)
1889 if (op_is_discard(bio->bi_opf))
1890 return BLKG_IOSTAT_DISCARD;
1891 if (op_is_write(bio->bi_opf))
1892 return BLKG_IOSTAT_WRITE;
1893 return BLKG_IOSTAT_READ;
1896 void blk_cgroup_bio_start(struct bio *bio)
1898 int rwd = blk_cgroup_io_type(bio), cpu;
1899 struct blkg_iostat_set *bis;
1902 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1903 u64_stats_update_begin(&bis->sync);
1906 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1907 * bio and we would have already accounted for the size of the bio.
1909 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1910 bio_set_flag(bio, BIO_CGROUP_ACCT);
1911 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1913 bis->cur.ios[rwd]++;
1915 u64_stats_update_end(&bis->sync);
1916 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1917 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1921 static int __init blkcg_init(void)
1923 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1924 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1925 WQ_UNBOUND | WQ_SYSFS, 0);
1926 if (!blkcg_punt_bio_wq)
1930 subsys_initcall(blkcg_init);
1932 module_param(blkcg_debug_stats, bool, 0644);
1933 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");