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;
123 /* as long as there are pending bios, @blkg can't go away */
124 spin_lock_bh(&blkg->async_bio_lock);
125 bio_list_merge(&bios, &blkg->async_bios);
126 bio_list_init(&blkg->async_bios);
127 spin_unlock_bh(&blkg->async_bio_lock);
129 while ((bio = bio_list_pop(&bios)))
134 * blkg_alloc - allocate a blkg
135 * @blkcg: block cgroup the new blkg is associated with
136 * @q: request_queue the new blkg is associated with
137 * @gfp_mask: allocation mask to use
139 * Allocate a new blkg assocating @blkcg and @q.
141 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
144 struct blkcg_gq *blkg;
147 /* alloc and init base part */
148 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
152 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
155 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
156 if (!blkg->iostat_cpu)
160 INIT_LIST_HEAD(&blkg->q_node);
161 spin_lock_init(&blkg->async_bio_lock);
162 bio_list_init(&blkg->async_bios);
163 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
166 u64_stats_init(&blkg->iostat.sync);
167 for_each_possible_cpu(cpu)
168 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
170 for (i = 0; i < BLKCG_MAX_POLS; i++) {
171 struct blkcg_policy *pol = blkcg_policy[i];
172 struct blkg_policy_data *pd;
174 if (!blkcg_policy_enabled(q, pol))
177 /* alloc per-policy data and attach it to blkg */
178 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
194 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
195 struct request_queue *q, bool update_hint)
197 struct blkcg_gq *blkg;
200 * Hint didn't match. Look up from the radix tree. Note that the
201 * hint can only be updated under queue_lock as otherwise @blkg
202 * could have already been removed from blkg_tree. The caller is
203 * responsible for grabbing queue_lock if @update_hint.
205 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
206 if (blkg && blkg->q == q) {
208 lockdep_assert_held(&q->queue_lock);
209 rcu_assign_pointer(blkcg->blkg_hint, blkg);
216 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
219 * If @new_blkg is %NULL, this function tries to allocate a new one as
220 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
222 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
223 struct request_queue *q,
224 struct blkcg_gq *new_blkg)
226 struct blkcg_gq *blkg;
229 WARN_ON_ONCE(!rcu_read_lock_held());
230 lockdep_assert_held(&q->queue_lock);
232 /* request_queue is dying, do not create/recreate a blkg */
233 if (blk_queue_dying(q)) {
238 /* blkg holds a reference to blkcg */
239 if (!css_tryget_online(&blkcg->css)) {
246 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
247 if (unlikely(!new_blkg)) {
255 if (blkcg_parent(blkcg)) {
256 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
257 if (WARN_ON_ONCE(!blkg->parent)) {
261 blkg_get(blkg->parent);
264 /* invoke per-policy init */
265 for (i = 0; i < BLKCG_MAX_POLS; i++) {
266 struct blkcg_policy *pol = blkcg_policy[i];
268 if (blkg->pd[i] && pol->pd_init_fn)
269 pol->pd_init_fn(blkg->pd[i]);
273 spin_lock(&blkcg->lock);
274 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
276 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
277 list_add(&blkg->q_node, &q->blkg_list);
279 for (i = 0; i < BLKCG_MAX_POLS; i++) {
280 struct blkcg_policy *pol = blkcg_policy[i];
282 if (blkg->pd[i] && pol->pd_online_fn)
283 pol->pd_online_fn(blkg->pd[i]);
287 spin_unlock(&blkcg->lock);
292 /* @blkg failed fully initialized, use the usual release path */
297 css_put(&blkcg->css);
304 * blkg_lookup_create - lookup blkg, try to create one if not there
305 * @blkcg: blkcg of interest
306 * @q: request_queue of interest
308 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
309 * create one. blkg creation is performed recursively from blkcg_root such
310 * that all non-root blkg's have access to the parent blkg. This function
311 * should be called under RCU read lock and takes @q->queue_lock.
313 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
316 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
317 struct request_queue *q)
319 struct blkcg_gq *blkg;
322 WARN_ON_ONCE(!rcu_read_lock_held());
324 blkg = blkg_lookup(blkcg, q);
328 spin_lock_irqsave(&q->queue_lock, flags);
329 blkg = __blkg_lookup(blkcg, q, true);
334 * Create blkgs walking down from blkcg_root to @blkcg, so that all
335 * non-root blkgs have access to their parents. Returns the closest
336 * blkg to the intended blkg should blkg_create() fail.
339 struct blkcg *pos = blkcg;
340 struct blkcg *parent = blkcg_parent(blkcg);
341 struct blkcg_gq *ret_blkg = q->root_blkg;
344 blkg = __blkg_lookup(parent, q, false);
346 /* remember closest blkg */
351 parent = blkcg_parent(parent);
354 blkg = blkg_create(pos, q, NULL);
364 spin_unlock_irqrestore(&q->queue_lock, flags);
368 static void blkg_destroy(struct blkcg_gq *blkg)
370 struct blkcg *blkcg = blkg->blkcg;
373 lockdep_assert_held(&blkg->q->queue_lock);
374 lockdep_assert_held(&blkcg->lock);
376 /* Something wrong if we are trying to remove same group twice */
377 WARN_ON_ONCE(list_empty(&blkg->q_node));
378 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
380 for (i = 0; i < BLKCG_MAX_POLS; i++) {
381 struct blkcg_policy *pol = blkcg_policy[i];
383 if (blkg->pd[i] && pol->pd_offline_fn)
384 pol->pd_offline_fn(blkg->pd[i]);
387 blkg->online = false;
389 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
390 list_del_init(&blkg->q_node);
391 hlist_del_init_rcu(&blkg->blkcg_node);
394 * Both setting lookup hint to and clearing it from @blkg are done
395 * under queue_lock. If it's not pointing to @blkg now, it never
396 * will. Hint assignment itself can race safely.
398 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
399 rcu_assign_pointer(blkcg->blkg_hint, NULL);
402 * Put the reference taken at the time of creation so that when all
403 * queues are gone, group can be destroyed.
405 percpu_ref_kill(&blkg->refcnt);
409 * blkg_destroy_all - destroy all blkgs associated with a request_queue
410 * @q: request_queue of interest
412 * Destroy all blkgs associated with @q.
414 static void blkg_destroy_all(struct request_queue *q)
416 struct blkcg_gq *blkg, *n;
418 spin_lock_irq(&q->queue_lock);
419 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
420 struct blkcg *blkcg = blkg->blkcg;
422 spin_lock(&blkcg->lock);
424 spin_unlock(&blkcg->lock);
428 spin_unlock_irq(&q->queue_lock);
431 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
432 struct cftype *cftype, u64 val)
434 struct blkcg *blkcg = css_to_blkcg(css);
435 struct blkcg_gq *blkg;
438 mutex_lock(&blkcg_pol_mutex);
439 spin_lock_irq(&blkcg->lock);
442 * Note that stat reset is racy - it doesn't synchronize against
443 * stat updates. This is a debug feature which shouldn't exist
444 * anyway. If you get hit by a race, retry.
446 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
447 for_each_possible_cpu(cpu) {
448 struct blkg_iostat_set *bis =
449 per_cpu_ptr(blkg->iostat_cpu, cpu);
450 memset(bis, 0, sizeof(*bis));
452 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
454 for (i = 0; i < BLKCG_MAX_POLS; i++) {
455 struct blkcg_policy *pol = blkcg_policy[i];
457 if (blkg->pd[i] && pol->pd_reset_stats_fn)
458 pol->pd_reset_stats_fn(blkg->pd[i]);
462 spin_unlock_irq(&blkcg->lock);
463 mutex_unlock(&blkcg_pol_mutex);
467 const char *blkg_dev_name(struct blkcg_gq *blkg)
469 /* some drivers (floppy) instantiate a queue w/o disk registered */
470 if (blkg->q->backing_dev_info->dev)
471 return bdi_dev_name(blkg->q->backing_dev_info);
476 * blkcg_print_blkgs - helper for printing per-blkg data
477 * @sf: seq_file to print to
478 * @blkcg: blkcg of interest
479 * @prfill: fill function to print out a blkg
480 * @pol: policy in question
481 * @data: data to be passed to @prfill
482 * @show_total: to print out sum of prfill return values or not
484 * This function invokes @prfill on each blkg of @blkcg if pd for the
485 * policy specified by @pol exists. @prfill is invoked with @sf, the
486 * policy data and @data and the matching queue lock held. If @show_total
487 * is %true, the sum of the return values from @prfill is printed with
488 * "Total" label at the end.
490 * This is to be used to construct print functions for
491 * cftype->read_seq_string method.
493 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
494 u64 (*prfill)(struct seq_file *,
495 struct blkg_policy_data *, int),
496 const struct blkcg_policy *pol, int data,
499 struct blkcg_gq *blkg;
503 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
504 spin_lock_irq(&blkg->q->queue_lock);
505 if (blkcg_policy_enabled(blkg->q, pol))
506 total += prfill(sf, blkg->pd[pol->plid], data);
507 spin_unlock_irq(&blkg->q->queue_lock);
512 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
514 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
517 * __blkg_prfill_u64 - prfill helper for a single u64 value
518 * @sf: seq_file to print to
519 * @pd: policy private data of interest
522 * Print @v to @sf for the device assocaited with @pd.
524 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
526 const char *dname = blkg_dev_name(pd->blkg);
531 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
534 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
536 /* Performs queue bypass and policy enabled checks then looks up blkg. */
537 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
538 const struct blkcg_policy *pol,
539 struct request_queue *q)
541 WARN_ON_ONCE(!rcu_read_lock_held());
542 lockdep_assert_held(&q->queue_lock);
544 if (!blkcg_policy_enabled(q, pol))
545 return ERR_PTR(-EOPNOTSUPP);
546 return __blkg_lookup(blkcg, q, true /* update_hint */);
550 * blkg_conf_prep - parse and prepare for per-blkg config update
551 * @inputp: input string pointer
553 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
554 * from @input and get and return the matching gendisk. *@inputp is
555 * updated to point past the device node prefix. Returns an ERR_PTR()
558 * Use this function iff blkg_conf_prep() can't be used for some reason.
560 struct gendisk *blkcg_conf_get_disk(char **inputp)
562 char *input = *inputp;
563 unsigned int major, minor;
564 struct gendisk *disk;
567 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
568 return ERR_PTR(-EINVAL);
571 if (!isspace(*input))
572 return ERR_PTR(-EINVAL);
573 input = skip_spaces(input);
575 disk = get_gendisk(MKDEV(major, minor), &part);
577 return ERR_PTR(-ENODEV);
579 put_disk_and_module(disk);
580 return ERR_PTR(-ENODEV);
588 * blkg_conf_prep - parse and prepare for per-blkg config update
589 * @blkcg: target block cgroup
590 * @pol: target policy
591 * @input: input string
592 * @ctx: blkg_conf_ctx to be filled
594 * Parse per-blkg config update from @input and initialize @ctx with the
595 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
596 * part of @input following MAJ:MIN. This function returns with RCU read
597 * lock and queue lock held and must be paired with blkg_conf_finish().
599 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
600 char *input, struct blkg_conf_ctx *ctx)
601 __acquires(rcu) __acquires(&disk->queue->queue_lock)
603 struct gendisk *disk;
604 struct request_queue *q;
605 struct blkcg_gq *blkg;
608 disk = blkcg_conf_get_disk(&input);
610 return PTR_ERR(disk);
615 spin_lock_irq(&q->queue_lock);
617 blkg = blkg_lookup_check(blkcg, pol, q);
627 * Create blkgs walking down from blkcg_root to @blkcg, so that all
628 * non-root blkgs have access to their parents.
631 struct blkcg *pos = blkcg;
632 struct blkcg *parent;
633 struct blkcg_gq *new_blkg;
635 parent = blkcg_parent(blkcg);
636 while (parent && !__blkg_lookup(parent, q, false)) {
638 parent = blkcg_parent(parent);
641 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
642 spin_unlock_irq(&q->queue_lock);
645 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
646 if (unlikely(!new_blkg)) {
652 spin_lock_irq(&q->queue_lock);
654 blkg = blkg_lookup_check(pos, pol, q);
663 blkg = blkg_create(pos, q, new_blkg);
680 spin_unlock_irq(&q->queue_lock);
683 put_disk_and_module(disk);
685 * If queue was bypassing, we should retry. Do so after a
686 * short msleep(). It isn't strictly necessary but queue
687 * can be bypassing for some time and it's always nice to
688 * avoid busy looping.
692 ret = restart_syscall();
696 EXPORT_SYMBOL_GPL(blkg_conf_prep);
699 * blkg_conf_finish - finish up per-blkg config update
700 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
702 * Finish up after per-blkg config update. This function must be paired
703 * with blkg_conf_prep().
705 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
706 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
708 spin_unlock_irq(&ctx->disk->queue->queue_lock);
710 put_disk_and_module(ctx->disk);
712 EXPORT_SYMBOL_GPL(blkg_conf_finish);
714 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
718 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
719 dst->bytes[i] = src->bytes[i];
720 dst->ios[i] = src->ios[i];
724 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
728 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
729 dst->bytes[i] += src->bytes[i];
730 dst->ios[i] += src->ios[i];
734 static void blkg_iostat_sub(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 blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
746 struct blkcg *blkcg = css_to_blkcg(css);
747 struct blkcg_gq *blkg;
751 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
752 struct blkcg_gq *parent = blkg->parent;
753 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
754 struct blkg_iostat cur, delta;
757 /* fetch the current per-cpu values */
759 seq = u64_stats_fetch_begin(&bisc->sync);
760 blkg_iostat_set(&cur, &bisc->cur);
761 } while (u64_stats_fetch_retry(&bisc->sync, seq));
763 /* propagate percpu delta to global */
764 u64_stats_update_begin(&blkg->iostat.sync);
765 blkg_iostat_set(&delta, &cur);
766 blkg_iostat_sub(&delta, &bisc->last);
767 blkg_iostat_add(&blkg->iostat.cur, &delta);
768 blkg_iostat_add(&bisc->last, &delta);
769 u64_stats_update_end(&blkg->iostat.sync);
771 /* propagate global delta to parent */
773 u64_stats_update_begin(&parent->iostat.sync);
774 blkg_iostat_set(&delta, &blkg->iostat.cur);
775 blkg_iostat_sub(&delta, &blkg->iostat.last);
776 blkg_iostat_add(&parent->iostat.cur, &delta);
777 blkg_iostat_add(&blkg->iostat.last, &delta);
778 u64_stats_update_end(&parent->iostat.sync);
786 * The rstat algorithms intentionally don't handle the root cgroup to avoid
787 * incurring overhead when no cgroups are defined. For that reason,
788 * cgroup_rstat_flush in blkcg_print_stat does not actually fill out the
789 * iostat in the root cgroup's blkcg_gq.
791 * However, we would like to re-use the printing code between the root and
792 * non-root cgroups to the extent possible. For that reason, we simulate
793 * flushing the root cgroup's stats by explicitly filling in the iostat
794 * with disk level statistics.
796 static void blkcg_fill_root_iostats(void)
798 struct class_dev_iter iter;
801 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
802 while ((dev = class_dev_iter_next(&iter))) {
803 struct gendisk *disk = dev_to_disk(dev);
804 struct hd_struct *part = disk_get_part(disk, 0);
805 struct blkcg_gq *blkg = blk_queue_root_blkg(disk->queue);
806 struct blkg_iostat tmp;
809 memset(&tmp, 0, sizeof(tmp));
810 for_each_possible_cpu(cpu) {
811 struct disk_stats *cpu_dkstats;
813 cpu_dkstats = per_cpu_ptr(part->dkstats, cpu);
814 tmp.ios[BLKG_IOSTAT_READ] +=
815 cpu_dkstats->ios[STAT_READ];
816 tmp.ios[BLKG_IOSTAT_WRITE] +=
817 cpu_dkstats->ios[STAT_WRITE];
818 tmp.ios[BLKG_IOSTAT_DISCARD] +=
819 cpu_dkstats->ios[STAT_DISCARD];
820 // convert sectors to bytes
821 tmp.bytes[BLKG_IOSTAT_READ] +=
822 cpu_dkstats->sectors[STAT_READ] << 9;
823 tmp.bytes[BLKG_IOSTAT_WRITE] +=
824 cpu_dkstats->sectors[STAT_WRITE] << 9;
825 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
826 cpu_dkstats->sectors[STAT_DISCARD] << 9;
828 u64_stats_update_begin(&blkg->iostat.sync);
829 blkg_iostat_set(&blkg->iostat.cur, &tmp);
830 u64_stats_update_end(&blkg->iostat.sync);
835 static int blkcg_print_stat(struct seq_file *sf, void *v)
837 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
838 struct blkcg_gq *blkg;
840 if (!seq_css(sf)->parent)
841 blkcg_fill_root_iostats();
843 cgroup_rstat_flush(blkcg->css.cgroup);
847 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
848 struct blkg_iostat_set *bis = &blkg->iostat;
851 u64 rbytes, wbytes, rios, wios, dbytes, dios;
852 size_t size = seq_get_buf(sf, &buf), off = 0;
854 bool has_stats = false;
857 spin_lock_irq(&blkg->q->queue_lock);
862 dname = blkg_dev_name(blkg);
867 * Hooray string manipulation, count is the size written NOT
868 * INCLUDING THE \0, so size is now count+1 less than what we
869 * had before, but we want to start writing the next bit from
870 * the \0 so we only add count to buf.
872 off += scnprintf(buf+off, size-off, "%s ", dname);
875 seq = u64_stats_fetch_begin(&bis->sync);
877 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
878 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
879 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
880 rios = bis->cur.ios[BLKG_IOSTAT_READ];
881 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
882 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
883 } while (u64_stats_fetch_retry(&bis->sync, seq));
885 if (rbytes || wbytes || rios || wios) {
887 off += scnprintf(buf+off, size-off,
888 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
889 rbytes, wbytes, rios, wios,
893 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
895 off += scnprintf(buf+off, size-off,
896 " use_delay=%d delay_nsec=%llu",
897 atomic_read(&blkg->use_delay),
898 (unsigned long long)atomic64_read(&blkg->delay_nsec));
901 for (i = 0; i < BLKCG_MAX_POLS; i++) {
902 struct blkcg_policy *pol = blkcg_policy[i];
905 if (!blkg->pd[i] || !pol->pd_stat_fn)
908 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
915 if (off < size - 1) {
916 off += scnprintf(buf+off, size-off, "\n");
923 spin_unlock_irq(&blkg->q->queue_lock);
930 static struct cftype blkcg_files[] = {
933 .seq_show = blkcg_print_stat,
938 static struct cftype blkcg_legacy_files[] = {
940 .name = "reset_stats",
941 .write_u64 = blkcg_reset_stats,
947 * blkcg destruction is a three-stage process.
949 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
950 * which offlines writeback. Here we tie the next stage of blkg destruction
951 * to the completion of writeback associated with the blkcg. This lets us
952 * avoid punting potentially large amounts of outstanding writeback to root
953 * while maintaining any ongoing policies. The next stage is triggered when
954 * the nr_cgwbs count goes to zero.
956 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
957 * and handles the destruction of blkgs. Here the css reference held by
958 * the blkg is put back eventually allowing blkcg_css_free() to be called.
959 * This work may occur in cgwb_release_workfn() on the cgwb_release
960 * workqueue. Any submitted ios that fail to get the blkg ref will be
961 * punted to the root_blkg.
963 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
964 * This finally frees the blkcg.
968 * blkcg_css_offline - cgroup css_offline callback
969 * @css: css of interest
971 * This function is called when @css is about to go away. Here the cgwbs are
972 * offlined first and only once writeback associated with the blkcg has
973 * finished do we start step 2 (see above).
975 static void blkcg_css_offline(struct cgroup_subsys_state *css)
977 struct blkcg *blkcg = css_to_blkcg(css);
979 /* this prevents anyone from attaching or migrating to this blkcg */
980 wb_blkcg_offline(blkcg);
982 /* put the base online pin allowing step 2 to be triggered */
983 blkcg_unpin_online(blkcg);
987 * blkcg_destroy_blkgs - responsible for shooting down blkgs
988 * @blkcg: blkcg of interest
990 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
991 * is nested inside q lock, this function performs reverse double lock dancing.
992 * Destroying the blkgs releases the reference held on the blkcg's css allowing
993 * blkcg_css_free to eventually be called.
995 * This is the blkcg counterpart of ioc_release_fn().
997 void blkcg_destroy_blkgs(struct blkcg *blkcg)
999 spin_lock_irq(&blkcg->lock);
1001 while (!hlist_empty(&blkcg->blkg_list)) {
1002 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1003 struct blkcg_gq, blkcg_node);
1004 struct request_queue *q = blkg->q;
1006 if (spin_trylock(&q->queue_lock)) {
1008 spin_unlock(&q->queue_lock);
1010 spin_unlock_irq(&blkcg->lock);
1012 spin_lock_irq(&blkcg->lock);
1016 spin_unlock_irq(&blkcg->lock);
1019 static void blkcg_css_free(struct cgroup_subsys_state *css)
1021 struct blkcg *blkcg = css_to_blkcg(css);
1024 mutex_lock(&blkcg_pol_mutex);
1026 list_del(&blkcg->all_blkcgs_node);
1028 for (i = 0; i < BLKCG_MAX_POLS; i++)
1030 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1032 mutex_unlock(&blkcg_pol_mutex);
1037 static struct cgroup_subsys_state *
1038 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1040 struct blkcg *blkcg;
1041 struct cgroup_subsys_state *ret;
1044 mutex_lock(&blkcg_pol_mutex);
1047 blkcg = &blkcg_root;
1049 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1051 ret = ERR_PTR(-ENOMEM);
1056 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1057 struct blkcg_policy *pol = blkcg_policy[i];
1058 struct blkcg_policy_data *cpd;
1061 * If the policy hasn't been attached yet, wait for it
1062 * to be attached before doing anything else. Otherwise,
1063 * check if the policy requires any specific per-cgroup
1064 * data: if it does, allocate and initialize it.
1066 if (!pol || !pol->cpd_alloc_fn)
1069 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1071 ret = ERR_PTR(-ENOMEM);
1074 blkcg->cpd[i] = cpd;
1077 if (pol->cpd_init_fn)
1078 pol->cpd_init_fn(cpd);
1081 spin_lock_init(&blkcg->lock);
1082 refcount_set(&blkcg->online_pin, 1);
1083 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1084 INIT_HLIST_HEAD(&blkcg->blkg_list);
1085 #ifdef CONFIG_CGROUP_WRITEBACK
1086 INIT_LIST_HEAD(&blkcg->cgwb_list);
1088 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1090 mutex_unlock(&blkcg_pol_mutex);
1094 for (i--; i >= 0; i--)
1096 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1098 if (blkcg != &blkcg_root)
1101 mutex_unlock(&blkcg_pol_mutex);
1105 static int blkcg_css_online(struct cgroup_subsys_state *css)
1107 struct blkcg *blkcg = css_to_blkcg(css);
1108 struct blkcg *parent = blkcg_parent(blkcg);
1111 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1112 * don't go offline while cgwbs are still active on them. Pin the
1113 * parent so that offline always happens towards the root.
1116 blkcg_pin_online(parent);
1121 * blkcg_init_queue - initialize blkcg part of request queue
1122 * @q: request_queue to initialize
1124 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1125 * part of new request_queue @q.
1128 * 0 on success, -errno on failure.
1130 int blkcg_init_queue(struct request_queue *q)
1132 struct blkcg_gq *new_blkg, *blkg;
1136 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1140 preloaded = !radix_tree_preload(GFP_KERNEL);
1142 /* Make sure the root blkg exists. */
1144 spin_lock_irq(&q->queue_lock);
1145 blkg = blkg_create(&blkcg_root, q, new_blkg);
1148 q->root_blkg = blkg;
1149 spin_unlock_irq(&q->queue_lock);
1153 radix_tree_preload_end();
1155 ret = blk_throtl_init(q);
1157 goto err_destroy_all;
1159 ret = blk_iolatency_init(q);
1162 goto err_destroy_all;
1167 blkg_destroy_all(q);
1170 spin_unlock_irq(&q->queue_lock);
1173 radix_tree_preload_end();
1174 return PTR_ERR(blkg);
1178 * blkcg_exit_queue - exit and release blkcg part of request_queue
1179 * @q: request_queue being released
1181 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1183 void blkcg_exit_queue(struct request_queue *q)
1185 blkg_destroy_all(q);
1190 * We cannot support shared io contexts, as we have no mean to support
1191 * two tasks with the same ioc in two different groups without major rework
1192 * of the main cic data structures. For now we allow a task to change
1193 * its cgroup only if it's the only owner of its ioc.
1195 static int blkcg_can_attach(struct cgroup_taskset *tset)
1197 struct task_struct *task;
1198 struct cgroup_subsys_state *dst_css;
1199 struct io_context *ioc;
1202 /* task_lock() is needed to avoid races with exit_io_context() */
1203 cgroup_taskset_for_each(task, dst_css, tset) {
1205 ioc = task->io_context;
1206 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1215 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1219 mutex_lock(&blkcg_pol_mutex);
1221 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1222 struct blkcg_policy *pol = blkcg_policy[i];
1223 struct blkcg *blkcg;
1225 if (!pol || !pol->cpd_bind_fn)
1228 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1229 if (blkcg->cpd[pol->plid])
1230 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1232 mutex_unlock(&blkcg_pol_mutex);
1235 static void blkcg_exit(struct task_struct *tsk)
1237 if (tsk->throttle_queue)
1238 blk_put_queue(tsk->throttle_queue);
1239 tsk->throttle_queue = NULL;
1242 struct cgroup_subsys io_cgrp_subsys = {
1243 .css_alloc = blkcg_css_alloc,
1244 .css_online = blkcg_css_online,
1245 .css_offline = blkcg_css_offline,
1246 .css_free = blkcg_css_free,
1247 .can_attach = blkcg_can_attach,
1248 .css_rstat_flush = blkcg_rstat_flush,
1250 .dfl_cftypes = blkcg_files,
1251 .legacy_cftypes = blkcg_legacy_files,
1252 .legacy_name = "blkio",
1256 * This ensures that, if available, memcg is automatically enabled
1257 * together on the default hierarchy so that the owner cgroup can
1258 * be retrieved from writeback pages.
1260 .depends_on = 1 << memory_cgrp_id,
1263 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1266 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1267 * @q: request_queue of interest
1268 * @pol: blkcg policy to activate
1270 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1271 * bypass mode to populate its blkgs with policy_data for @pol.
1273 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1274 * from IO path. Update of each blkg is protected by both queue and blkcg
1275 * locks so that holding either lock and testing blkcg_policy_enabled() is
1276 * always enough for dereferencing policy data.
1278 * The caller is responsible for synchronizing [de]activations and policy
1279 * [un]registerations. Returns 0 on success, -errno on failure.
1281 int blkcg_activate_policy(struct request_queue *q,
1282 const struct blkcg_policy *pol)
1284 struct blkg_policy_data *pd_prealloc = NULL;
1285 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1288 if (blkcg_policy_enabled(q, pol))
1292 blk_mq_freeze_queue(q);
1294 spin_lock_irq(&q->queue_lock);
1296 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1297 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1298 struct blkg_policy_data *pd;
1300 if (blkg->pd[pol->plid])
1303 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1304 if (blkg == pinned_blkg) {
1308 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1314 * GFP_NOWAIT failed. Free the existing one and
1315 * prealloc for @blkg w/ GFP_KERNEL.
1318 blkg_put(pinned_blkg);
1322 spin_unlock_irq(&q->queue_lock);
1325 pol->pd_free_fn(pd_prealloc);
1326 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1334 blkg->pd[pol->plid] = pd;
1336 pd->plid = pol->plid;
1339 /* all allocated, init in the same order */
1340 if (pol->pd_init_fn)
1341 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1342 pol->pd_init_fn(blkg->pd[pol->plid]);
1344 __set_bit(pol->plid, q->blkcg_pols);
1347 spin_unlock_irq(&q->queue_lock);
1350 blk_mq_unfreeze_queue(q);
1352 blkg_put(pinned_blkg);
1354 pol->pd_free_fn(pd_prealloc);
1358 /* alloc failed, nothing's initialized yet, free everything */
1359 spin_lock_irq(&q->queue_lock);
1360 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1361 if (blkg->pd[pol->plid]) {
1362 pol->pd_free_fn(blkg->pd[pol->plid]);
1363 blkg->pd[pol->plid] = NULL;
1366 spin_unlock_irq(&q->queue_lock);
1370 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1373 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1374 * @q: request_queue of interest
1375 * @pol: blkcg policy to deactivate
1377 * Deactivate @pol on @q. Follows the same synchronization rules as
1378 * blkcg_activate_policy().
1380 void blkcg_deactivate_policy(struct request_queue *q,
1381 const struct blkcg_policy *pol)
1383 struct blkcg_gq *blkg;
1385 if (!blkcg_policy_enabled(q, pol))
1389 blk_mq_freeze_queue(q);
1391 spin_lock_irq(&q->queue_lock);
1393 __clear_bit(pol->plid, q->blkcg_pols);
1395 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1396 if (blkg->pd[pol->plid]) {
1397 if (pol->pd_offline_fn)
1398 pol->pd_offline_fn(blkg->pd[pol->plid]);
1399 pol->pd_free_fn(blkg->pd[pol->plid]);
1400 blkg->pd[pol->plid] = NULL;
1404 spin_unlock_irq(&q->queue_lock);
1407 blk_mq_unfreeze_queue(q);
1409 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1412 * blkcg_policy_register - register a blkcg policy
1413 * @pol: blkcg policy to register
1415 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1416 * successful registration. Returns 0 on success and -errno on failure.
1418 int blkcg_policy_register(struct blkcg_policy *pol)
1420 struct blkcg *blkcg;
1423 mutex_lock(&blkcg_pol_register_mutex);
1424 mutex_lock(&blkcg_pol_mutex);
1426 /* find an empty slot */
1428 for (i = 0; i < BLKCG_MAX_POLS; i++)
1429 if (!blkcg_policy[i])
1431 if (i >= BLKCG_MAX_POLS) {
1432 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1436 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1437 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1438 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1443 blkcg_policy[pol->plid] = pol;
1445 /* allocate and install cpd's */
1446 if (pol->cpd_alloc_fn) {
1447 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1448 struct blkcg_policy_data *cpd;
1450 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1454 blkcg->cpd[pol->plid] = cpd;
1456 cpd->plid = pol->plid;
1457 if (pol->cpd_init_fn)
1458 pol->cpd_init_fn(cpd);
1462 mutex_unlock(&blkcg_pol_mutex);
1464 /* everything is in place, add intf files for the new policy */
1465 if (pol->dfl_cftypes)
1466 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1468 if (pol->legacy_cftypes)
1469 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1470 pol->legacy_cftypes));
1471 mutex_unlock(&blkcg_pol_register_mutex);
1475 if (pol->cpd_free_fn) {
1476 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1477 if (blkcg->cpd[pol->plid]) {
1478 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1479 blkcg->cpd[pol->plid] = NULL;
1483 blkcg_policy[pol->plid] = NULL;
1485 mutex_unlock(&blkcg_pol_mutex);
1486 mutex_unlock(&blkcg_pol_register_mutex);
1489 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1492 * blkcg_policy_unregister - unregister a blkcg policy
1493 * @pol: blkcg policy to unregister
1495 * Undo blkcg_policy_register(@pol). Might sleep.
1497 void blkcg_policy_unregister(struct blkcg_policy *pol)
1499 struct blkcg *blkcg;
1501 mutex_lock(&blkcg_pol_register_mutex);
1503 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1506 /* kill the intf files first */
1507 if (pol->dfl_cftypes)
1508 cgroup_rm_cftypes(pol->dfl_cftypes);
1509 if (pol->legacy_cftypes)
1510 cgroup_rm_cftypes(pol->legacy_cftypes);
1512 /* remove cpds and unregister */
1513 mutex_lock(&blkcg_pol_mutex);
1515 if (pol->cpd_free_fn) {
1516 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1517 if (blkcg->cpd[pol->plid]) {
1518 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1519 blkcg->cpd[pol->plid] = NULL;
1523 blkcg_policy[pol->plid] = NULL;
1525 mutex_unlock(&blkcg_pol_mutex);
1527 mutex_unlock(&blkcg_pol_register_mutex);
1529 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1531 bool __blkcg_punt_bio_submit(struct bio *bio)
1533 struct blkcg_gq *blkg = bio->bi_blkg;
1535 /* consume the flag first */
1536 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1538 /* never bounce for the root cgroup */
1542 spin_lock_bh(&blkg->async_bio_lock);
1543 bio_list_add(&blkg->async_bios, bio);
1544 spin_unlock_bh(&blkg->async_bio_lock);
1546 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1551 * Scale the accumulated delay based on how long it has been since we updated
1552 * the delay. We only call this when we are adding delay, in case it's been a
1553 * while since we added delay, and when we are checking to see if we need to
1554 * delay a task, to account for any delays that may have occurred.
1556 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1558 u64 old = atomic64_read(&blkg->delay_start);
1560 /* negative use_delay means no scaling, see blkcg_set_delay() */
1561 if (atomic_read(&blkg->use_delay) < 0)
1565 * We only want to scale down every second. The idea here is that we
1566 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1567 * time window. We only want to throttle tasks for recent delay that
1568 * has occurred, in 1 second time windows since that's the maximum
1569 * things can be throttled. We save the current delay window in
1570 * blkg->last_delay so we know what amount is still left to be charged
1571 * to the blkg from this point onward. blkg->last_use keeps track of
1572 * the use_delay counter. The idea is if we're unthrottling the blkg we
1573 * are ok with whatever is happening now, and we can take away more of
1574 * the accumulated delay as we've already throttled enough that
1575 * everybody is happy with their IO latencies.
1577 if (time_before64(old + NSEC_PER_SEC, now) &&
1578 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1579 u64 cur = atomic64_read(&blkg->delay_nsec);
1580 u64 sub = min_t(u64, blkg->last_delay, now - old);
1581 int cur_use = atomic_read(&blkg->use_delay);
1584 * We've been unthrottled, subtract a larger chunk of our
1585 * accumulated delay.
1587 if (cur_use < blkg->last_use)
1588 sub = max_t(u64, sub, blkg->last_delay >> 1);
1591 * This shouldn't happen, but handle it anyway. Our delay_nsec
1592 * should only ever be growing except here where we subtract out
1593 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1594 * rather not end up with negative numbers.
1596 if (unlikely(cur < sub)) {
1597 atomic64_set(&blkg->delay_nsec, 0);
1598 blkg->last_delay = 0;
1600 atomic64_sub(sub, &blkg->delay_nsec);
1601 blkg->last_delay = cur - sub;
1603 blkg->last_use = cur_use;
1608 * This is called when we want to actually walk up the hierarchy and check to
1609 * see if we need to throttle, and then actually throttle if there is some
1610 * accumulated delay. This should only be called upon return to user space so
1611 * we're not holding some lock that would induce a priority inversion.
1613 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1615 unsigned long pflags;
1616 u64 now = ktime_to_ns(ktime_get());
1621 while (blkg->parent) {
1622 if (atomic_read(&blkg->use_delay)) {
1623 blkcg_scale_delay(blkg, now);
1624 delay_nsec = max_t(u64, delay_nsec,
1625 atomic64_read(&blkg->delay_nsec));
1627 blkg = blkg->parent;
1634 * Let's not sleep for all eternity if we've amassed a huge delay.
1635 * Swapping or metadata IO can accumulate 10's of seconds worth of
1636 * delay, and we want userspace to be able to do _something_ so cap the
1637 * delays at 1 second. If there's 10's of seconds worth of delay then
1638 * the tasks will be delayed for 1 second for every syscall.
1640 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1643 psi_memstall_enter(&pflags);
1645 exp = ktime_add_ns(now, delay_nsec);
1646 tok = io_schedule_prepare();
1648 __set_current_state(TASK_KILLABLE);
1649 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1651 } while (!fatal_signal_pending(current));
1652 io_schedule_finish(tok);
1655 psi_memstall_leave(&pflags);
1659 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1661 * This is only called if we've been marked with set_notify_resume(). Obviously
1662 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1663 * check to see if current->throttle_queue is set and if not this doesn't do
1664 * anything. This should only ever be called by the resume code, it's not meant
1665 * to be called by people willy-nilly as it will actually do the work to
1666 * throttle the task if it is setup for throttling.
1668 void blkcg_maybe_throttle_current(void)
1670 struct request_queue *q = current->throttle_queue;
1671 struct cgroup_subsys_state *css;
1672 struct blkcg *blkcg;
1673 struct blkcg_gq *blkg;
1674 bool use_memdelay = current->use_memdelay;
1679 current->throttle_queue = NULL;
1680 current->use_memdelay = false;
1683 css = kthread_blkcg();
1685 blkcg = css_to_blkcg(css);
1687 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1691 blkg = blkg_lookup(blkcg, q);
1694 if (!blkg_tryget(blkg))
1698 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1708 * blkcg_schedule_throttle - this task needs to check for throttling
1709 * @q: the request queue IO was submitted on
1710 * @use_memdelay: do we charge this to memory delay for PSI
1712 * This is called by the IO controller when we know there's delay accumulated
1713 * for the blkg for this task. We do not pass the blkg because there are places
1714 * we call this that may not have that information, the swapping code for
1715 * instance will only have a request_queue at that point. This set's the
1716 * notify_resume for the task to check and see if it requires throttling before
1717 * returning to user space.
1719 * We will only schedule once per syscall. You can call this over and over
1720 * again and it will only do the check once upon return to user space, and only
1721 * throttle once. If the task needs to be throttled again it'll need to be
1722 * re-set at the next time we see the task.
1724 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1726 if (unlikely(current->flags & PF_KTHREAD))
1729 if (!blk_get_queue(q))
1732 if (current->throttle_queue)
1733 blk_put_queue(current->throttle_queue);
1734 current->throttle_queue = q;
1736 current->use_memdelay = use_memdelay;
1737 set_notify_resume(current);
1741 * blkcg_add_delay - add delay to this blkg
1742 * @blkg: blkg of interest
1743 * @now: the current time in nanoseconds
1744 * @delta: how many nanoseconds of delay to add
1746 * Charge @delta to the blkg's current delay accumulation. This is used to
1747 * throttle tasks if an IO controller thinks we need more throttling.
1749 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1751 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1753 blkcg_scale_delay(blkg, now);
1754 atomic64_add(delta, &blkg->delay_nsec);
1758 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1762 * As the failure mode here is to walk up the blkg tree, this ensure that the
1763 * blkg->parent pointers are always valid. This returns the blkg that it ended
1764 * up taking a reference on or %NULL if no reference was taken.
1766 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1767 struct cgroup_subsys_state *css)
1769 struct blkcg_gq *blkg, *ret_blkg = NULL;
1772 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_disk->queue);
1774 if (blkg_tryget(blkg)) {
1778 blkg = blkg->parent;
1786 * bio_associate_blkg_from_css - associate a bio with a specified css
1790 * Associate @bio with the blkg found by combining the css's blkg and the
1791 * request_queue of the @bio. An association failure is handled by walking up
1792 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1793 * and q->root_blkg. This situation only happens when a cgroup is dying and
1794 * then the remaining bios will spill to the closest alive blkg.
1796 * A reference will be taken on the blkg and will be released when @bio is
1799 void bio_associate_blkg_from_css(struct bio *bio,
1800 struct cgroup_subsys_state *css)
1803 blkg_put(bio->bi_blkg);
1805 if (css && css->parent) {
1806 bio->bi_blkg = blkg_tryget_closest(bio, css);
1808 blkg_get(bio->bi_disk->queue->root_blkg);
1809 bio->bi_blkg = bio->bi_disk->queue->root_blkg;
1812 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1815 * bio_associate_blkg - associate a bio with a blkg
1818 * Associate @bio with the blkg found from the bio's css and request_queue.
1819 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1820 * already associated, the css is reused and association redone as the
1821 * request_queue may have changed.
1823 void bio_associate_blkg(struct bio *bio)
1825 struct cgroup_subsys_state *css;
1830 css = &bio_blkcg(bio)->css;
1834 bio_associate_blkg_from_css(bio, css);
1838 EXPORT_SYMBOL_GPL(bio_associate_blkg);
1841 * bio_clone_blkg_association - clone blkg association from src to dst bio
1842 * @dst: destination bio
1845 void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1849 blkg_put(dst->bi_blkg);
1850 blkg_get(src->bi_blkg);
1851 dst->bi_blkg = src->bi_blkg;
1854 EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1856 static int blk_cgroup_io_type(struct bio *bio)
1858 if (op_is_discard(bio->bi_opf))
1859 return BLKG_IOSTAT_DISCARD;
1860 if (op_is_write(bio->bi_opf))
1861 return BLKG_IOSTAT_WRITE;
1862 return BLKG_IOSTAT_READ;
1865 void blk_cgroup_bio_start(struct bio *bio)
1867 int rwd = blk_cgroup_io_type(bio), cpu;
1868 struct blkg_iostat_set *bis;
1871 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1872 u64_stats_update_begin(&bis->sync);
1875 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1876 * bio and we would have already accounted for the size of the bio.
1878 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1879 bio_set_flag(bio, BIO_CGROUP_ACCT);
1880 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1882 bis->cur.ios[rwd]++;
1884 u64_stats_update_end(&bis->sync);
1885 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1886 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1890 static int __init blkcg_init(void)
1892 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1893 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1894 WQ_UNBOUND | WQ_SYSFS, 0);
1895 if (!blkcg_punt_bio_wq)
1899 subsys_initcall(blkcg_init);
1901 module_param(blkcg_debug_stats, bool, 0644);
1902 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");