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 blkg_rwstat_exit(&blkg->stat_ios);
84 blkg_rwstat_exit(&blkg->stat_bytes);
85 percpu_ref_exit(&blkg->refcnt);
89 static void __blkg_release(struct rcu_head *rcu)
91 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
93 WARN_ON(!bio_list_empty(&blkg->async_bios));
95 /* release the blkcg and parent blkg refs this blkg has been holding */
96 css_put(&blkg->blkcg->css);
98 blkg_put(blkg->parent);
100 wb_congested_put(blkg->wb_congested);
106 * A group is RCU protected, but having an rcu lock does not mean that one
107 * can access all the fields of blkg and assume these are valid. For
108 * example, don't try to follow throtl_data and request queue links.
110 * Having a reference to blkg under an rcu allows accesses to only values
111 * local to groups like group stats and group rate limits.
113 static void blkg_release(struct percpu_ref *ref)
115 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
117 call_rcu(&blkg->rcu_head, __blkg_release);
120 static void blkg_async_bio_workfn(struct work_struct *work)
122 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
124 struct bio_list bios = BIO_EMPTY_LIST;
127 /* as long as there are pending bios, @blkg can't go away */
128 spin_lock_bh(&blkg->async_bio_lock);
129 bio_list_merge(&bios, &blkg->async_bios);
130 bio_list_init(&blkg->async_bios);
131 spin_unlock_bh(&blkg->async_bio_lock);
133 while ((bio = bio_list_pop(&bios)))
138 * blkg_alloc - allocate a blkg
139 * @blkcg: block cgroup the new blkg is associated with
140 * @q: request_queue the new blkg is associated with
141 * @gfp_mask: allocation mask to use
143 * Allocate a new blkg assocating @blkcg and @q.
145 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
148 struct blkcg_gq *blkg;
151 /* alloc and init base part */
152 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
156 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
159 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
160 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
164 INIT_LIST_HEAD(&blkg->q_node);
165 spin_lock_init(&blkg->async_bio_lock);
166 bio_list_init(&blkg->async_bios);
167 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
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;
227 struct bdi_writeback_congested *wb_congested;
230 WARN_ON_ONCE(!rcu_read_lock_held());
231 lockdep_assert_held(&q->queue_lock);
233 /* request_queue is dying, do not create/recreate a blkg */
234 if (blk_queue_dying(q)) {
239 /* blkg holds a reference to blkcg */
240 if (!css_tryget_online(&blkcg->css)) {
245 wb_congested = wb_congested_get_create(q->backing_dev_info,
247 GFP_NOWAIT | __GFP_NOWARN);
255 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
256 if (unlikely(!new_blkg)) {
258 goto err_put_congested;
262 blkg->wb_congested = wb_congested;
265 if (blkcg_parent(blkcg)) {
266 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
267 if (WARN_ON_ONCE(!blkg->parent)) {
269 goto err_put_congested;
271 blkg_get(blkg->parent);
274 /* invoke per-policy init */
275 for (i = 0; i < BLKCG_MAX_POLS; i++) {
276 struct blkcg_policy *pol = blkcg_policy[i];
278 if (blkg->pd[i] && pol->pd_init_fn)
279 pol->pd_init_fn(blkg->pd[i]);
283 spin_lock(&blkcg->lock);
284 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
286 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
287 list_add(&blkg->q_node, &q->blkg_list);
289 for (i = 0; i < BLKCG_MAX_POLS; i++) {
290 struct blkcg_policy *pol = blkcg_policy[i];
292 if (blkg->pd[i] && pol->pd_online_fn)
293 pol->pd_online_fn(blkg->pd[i]);
297 spin_unlock(&blkcg->lock);
302 /* @blkg failed fully initialized, use the usual release path */
307 wb_congested_put(wb_congested);
309 css_put(&blkcg->css);
316 * __blkg_lookup_create - lookup blkg, try to create one if not there
317 * @blkcg: blkcg of interest
318 * @q: request_queue of interest
320 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
321 * create one. blkg creation is performed recursively from blkcg_root such
322 * that all non-root blkg's have access to the parent blkg. This function
323 * should be called under RCU read lock and @q->queue_lock.
325 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
328 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
329 struct request_queue *q)
331 struct blkcg_gq *blkg;
333 WARN_ON_ONCE(!rcu_read_lock_held());
334 lockdep_assert_held(&q->queue_lock);
336 blkg = __blkg_lookup(blkcg, q, true);
341 * Create blkgs walking down from blkcg_root to @blkcg, so that all
342 * non-root blkgs have access to their parents. Returns the closest
343 * blkg to the intended blkg should blkg_create() fail.
346 struct blkcg *pos = blkcg;
347 struct blkcg *parent = blkcg_parent(blkcg);
348 struct blkcg_gq *ret_blkg = q->root_blkg;
351 blkg = __blkg_lookup(parent, q, false);
353 /* remember closest blkg */
358 parent = blkcg_parent(parent);
361 blkg = blkg_create(pos, q, NULL);
370 * blkg_lookup_create - find or create a blkg
371 * @blkcg: target block cgroup
372 * @q: target request_queue
374 * This looks up or creates the blkg representing the unique pair
375 * of the blkcg and the request_queue.
377 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
378 struct request_queue *q)
380 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
382 if (unlikely(!blkg)) {
385 spin_lock_irqsave(&q->queue_lock, flags);
386 blkg = __blkg_lookup_create(blkcg, q);
387 spin_unlock_irqrestore(&q->queue_lock, flags);
393 static void blkg_destroy(struct blkcg_gq *blkg)
395 struct blkcg *blkcg = blkg->blkcg;
396 struct blkcg_gq *parent = blkg->parent;
399 lockdep_assert_held(&blkg->q->queue_lock);
400 lockdep_assert_held(&blkcg->lock);
402 /* Something wrong if we are trying to remove same group twice */
403 WARN_ON_ONCE(list_empty(&blkg->q_node));
404 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
406 for (i = 0; i < BLKCG_MAX_POLS; i++) {
407 struct blkcg_policy *pol = blkcg_policy[i];
409 if (blkg->pd[i] && pol->pd_offline_fn)
410 pol->pd_offline_fn(blkg->pd[i]);
414 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
415 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
418 blkg->online = false;
420 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
421 list_del_init(&blkg->q_node);
422 hlist_del_init_rcu(&blkg->blkcg_node);
425 * Both setting lookup hint to and clearing it from @blkg are done
426 * under queue_lock. If it's not pointing to @blkg now, it never
427 * will. Hint assignment itself can race safely.
429 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
430 rcu_assign_pointer(blkcg->blkg_hint, NULL);
433 * Put the reference taken at the time of creation so that when all
434 * queues are gone, group can be destroyed.
436 percpu_ref_kill(&blkg->refcnt);
440 * blkg_destroy_all - destroy all blkgs associated with a request_queue
441 * @q: request_queue of interest
443 * Destroy all blkgs associated with @q.
445 static void blkg_destroy_all(struct request_queue *q)
447 struct blkcg_gq *blkg, *n;
449 spin_lock_irq(&q->queue_lock);
450 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
451 struct blkcg *blkcg = blkg->blkcg;
453 spin_lock(&blkcg->lock);
455 spin_unlock(&blkcg->lock);
459 spin_unlock_irq(&q->queue_lock);
462 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
463 struct cftype *cftype, u64 val)
465 struct blkcg *blkcg = css_to_blkcg(css);
466 struct blkcg_gq *blkg;
469 mutex_lock(&blkcg_pol_mutex);
470 spin_lock_irq(&blkcg->lock);
473 * Note that stat reset is racy - it doesn't synchronize against
474 * stat updates. This is a debug feature which shouldn't exist
475 * anyway. If you get hit by a race, retry.
477 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
478 blkg_rwstat_reset(&blkg->stat_bytes);
479 blkg_rwstat_reset(&blkg->stat_ios);
481 for (i = 0; i < BLKCG_MAX_POLS; i++) {
482 struct blkcg_policy *pol = blkcg_policy[i];
484 if (blkg->pd[i] && pol->pd_reset_stats_fn)
485 pol->pd_reset_stats_fn(blkg->pd[i]);
489 spin_unlock_irq(&blkcg->lock);
490 mutex_unlock(&blkcg_pol_mutex);
494 const char *blkg_dev_name(struct blkcg_gq *blkg)
496 /* some drivers (floppy) instantiate a queue w/o disk registered */
497 if (blkg->q->backing_dev_info->dev)
498 return dev_name(blkg->q->backing_dev_info->dev);
503 * blkcg_print_blkgs - helper for printing per-blkg data
504 * @sf: seq_file to print to
505 * @blkcg: blkcg of interest
506 * @prfill: fill function to print out a blkg
507 * @pol: policy in question
508 * @data: data to be passed to @prfill
509 * @show_total: to print out sum of prfill return values or not
511 * This function invokes @prfill on each blkg of @blkcg if pd for the
512 * policy specified by @pol exists. @prfill is invoked with @sf, the
513 * policy data and @data and the matching queue lock held. If @show_total
514 * is %true, the sum of the return values from @prfill is printed with
515 * "Total" label at the end.
517 * This is to be used to construct print functions for
518 * cftype->read_seq_string method.
520 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
521 u64 (*prfill)(struct seq_file *,
522 struct blkg_policy_data *, int),
523 const struct blkcg_policy *pol, int data,
526 struct blkcg_gq *blkg;
530 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
531 spin_lock_irq(&blkg->q->queue_lock);
532 if (blkcg_policy_enabled(blkg->q, pol))
533 total += prfill(sf, blkg->pd[pol->plid], data);
534 spin_unlock_irq(&blkg->q->queue_lock);
539 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
541 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
544 * __blkg_prfill_u64 - prfill helper for a single u64 value
545 * @sf: seq_file to print to
546 * @pd: policy private data of interest
549 * Print @v to @sf for the device assocaited with @pd.
551 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
553 const char *dname = blkg_dev_name(pd->blkg);
558 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
561 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
564 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
565 * @sf: seq_file to print to
566 * @pd: policy private data of interest
567 * @rwstat: rwstat to print
569 * Print @rwstat to @sf for the device assocaited with @pd.
571 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
572 const struct blkg_rwstat_sample *rwstat)
574 static const char *rwstr[] = {
575 [BLKG_RWSTAT_READ] = "Read",
576 [BLKG_RWSTAT_WRITE] = "Write",
577 [BLKG_RWSTAT_SYNC] = "Sync",
578 [BLKG_RWSTAT_ASYNC] = "Async",
579 [BLKG_RWSTAT_DISCARD] = "Discard",
581 const char *dname = blkg_dev_name(pd->blkg);
588 for (i = 0; i < BLKG_RWSTAT_NR; i++)
589 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
592 v = rwstat->cnt[BLKG_RWSTAT_READ] +
593 rwstat->cnt[BLKG_RWSTAT_WRITE] +
594 rwstat->cnt[BLKG_RWSTAT_DISCARD];
595 seq_printf(sf, "%s Total %llu\n", dname, v);
598 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
601 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
602 * @sf: seq_file to print to
603 * @pd: policy private data of interest
604 * @off: offset to the blkg_rwstat in @pd
606 * prfill callback for printing a blkg_rwstat.
608 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
611 struct blkg_rwstat_sample rwstat = { };
613 blkg_rwstat_read((void *)pd + off, &rwstat);
614 return __blkg_prfill_rwstat(sf, pd, &rwstat);
616 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
618 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
619 struct blkg_policy_data *pd, int off)
621 struct blkg_rwstat_sample rwstat = { };
623 blkg_rwstat_read((void *)pd->blkg + off, &rwstat);
624 return __blkg_prfill_rwstat(sf, pd, &rwstat);
628 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
629 * @sf: seq_file to print to
632 * To be used as cftype->seq_show to print blkg->stat_bytes.
633 * cftype->private must be set to the blkcg_policy.
635 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
637 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
638 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
639 offsetof(struct blkcg_gq, stat_bytes), true);
642 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
645 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
646 * @sf: seq_file to print to
649 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
650 * must be set to the blkcg_policy.
652 int blkg_print_stat_ios(struct seq_file *sf, void *v)
654 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
655 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
656 offsetof(struct blkcg_gq, stat_ios), true);
659 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
661 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
662 struct blkg_policy_data *pd,
665 struct blkg_rwstat_sample rwstat;
667 blkg_rwstat_recursive_sum(pd->blkg, NULL, off, &rwstat);
668 return __blkg_prfill_rwstat(sf, pd, &rwstat);
672 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
673 * @sf: seq_file to print to
676 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
678 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
679 blkg_prfill_rwstat_field_recursive,
680 (void *)seq_cft(sf)->private,
681 offsetof(struct blkcg_gq, stat_bytes), true);
684 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
687 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
688 * @sf: seq_file to print to
691 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
693 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
694 blkg_prfill_rwstat_field_recursive,
695 (void *)seq_cft(sf)->private,
696 offsetof(struct blkcg_gq, stat_ios), true);
699 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
702 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
703 * @blkg: blkg of interest
704 * @pol: blkcg_policy which contains the blkg_rwstat
705 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
706 * @sum: blkg_rwstat_sample structure containing the results
708 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
709 * online descendants and their aux counts. The caller must be holding the
710 * queue lock for online tests.
712 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
713 * is at @off bytes into @blkg's blkg_policy_data of the policy.
715 void blkg_rwstat_recursive_sum(struct blkcg_gq *blkg, struct blkcg_policy *pol,
716 int off, struct blkg_rwstat_sample *sum)
718 struct blkcg_gq *pos_blkg;
719 struct cgroup_subsys_state *pos_css;
722 lockdep_assert_held(&blkg->q->queue_lock);
725 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
726 struct blkg_rwstat *rwstat;
728 if (!pos_blkg->online)
732 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
734 rwstat = (void *)pos_blkg + off;
736 for (i = 0; i < BLKG_RWSTAT_NR; i++)
737 sum->cnt[i] = blkg_rwstat_read_counter(rwstat, i);
741 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
743 /* Performs queue bypass and policy enabled checks then looks up blkg. */
744 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
745 const struct blkcg_policy *pol,
746 struct request_queue *q)
748 WARN_ON_ONCE(!rcu_read_lock_held());
749 lockdep_assert_held(&q->queue_lock);
751 if (!blkcg_policy_enabled(q, pol))
752 return ERR_PTR(-EOPNOTSUPP);
753 return __blkg_lookup(blkcg, q, true /* update_hint */);
757 * blkg_conf_prep - parse and prepare for per-blkg config update
758 * @inputp: input string pointer
760 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
761 * from @input and get and return the matching gendisk. *@inputp is
762 * updated to point past the device node prefix. Returns an ERR_PTR()
765 * Use this function iff blkg_conf_prep() can't be used for some reason.
767 struct gendisk *blkcg_conf_get_disk(char **inputp)
769 char *input = *inputp;
770 unsigned int major, minor;
771 struct gendisk *disk;
774 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
775 return ERR_PTR(-EINVAL);
778 if (!isspace(*input))
779 return ERR_PTR(-EINVAL);
780 input = skip_spaces(input);
782 disk = get_gendisk(MKDEV(major, minor), &part);
784 return ERR_PTR(-ENODEV);
786 put_disk_and_module(disk);
787 return ERR_PTR(-ENODEV);
795 * blkg_conf_prep - parse and prepare for per-blkg config update
796 * @blkcg: target block cgroup
797 * @pol: target policy
798 * @input: input string
799 * @ctx: blkg_conf_ctx to be filled
801 * Parse per-blkg config update from @input and initialize @ctx with the
802 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
803 * part of @input following MAJ:MIN. This function returns with RCU read
804 * lock and queue lock held and must be paired with blkg_conf_finish().
806 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
807 char *input, struct blkg_conf_ctx *ctx)
808 __acquires(rcu) __acquires(&disk->queue->queue_lock)
810 struct gendisk *disk;
811 struct request_queue *q;
812 struct blkcg_gq *blkg;
815 disk = blkcg_conf_get_disk(&input);
817 return PTR_ERR(disk);
822 spin_lock_irq(&q->queue_lock);
824 blkg = blkg_lookup_check(blkcg, pol, q);
834 * Create blkgs walking down from blkcg_root to @blkcg, so that all
835 * non-root blkgs have access to their parents.
838 struct blkcg *pos = blkcg;
839 struct blkcg *parent;
840 struct blkcg_gq *new_blkg;
842 parent = blkcg_parent(blkcg);
843 while (parent && !__blkg_lookup(parent, q, false)) {
845 parent = blkcg_parent(parent);
848 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
849 spin_unlock_irq(&q->queue_lock);
852 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
853 if (unlikely(!new_blkg)) {
858 if (radix_tree_preload(GFP_KERNEL)) {
865 spin_lock_irq(&q->queue_lock);
867 blkg = blkg_lookup_check(pos, pol, q);
877 blkg = blkg_create(pos, q, new_blkg);
884 radix_tree_preload_end();
896 radix_tree_preload_end();
898 spin_unlock_irq(&q->queue_lock);
901 put_disk_and_module(disk);
903 * If queue was bypassing, we should retry. Do so after a
904 * short msleep(). It isn't strictly necessary but queue
905 * can be bypassing for some time and it's always nice to
906 * avoid busy looping.
910 ret = restart_syscall();
914 EXPORT_SYMBOL_GPL(blkg_conf_prep);
917 * blkg_conf_finish - finish up per-blkg config update
918 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
920 * Finish up after per-blkg config update. This function must be paired
921 * with blkg_conf_prep().
923 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
924 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
926 spin_unlock_irq(&ctx->disk->queue->queue_lock);
928 put_disk_and_module(ctx->disk);
930 EXPORT_SYMBOL_GPL(blkg_conf_finish);
932 static int blkcg_print_stat(struct seq_file *sf, void *v)
934 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
935 struct blkcg_gq *blkg;
939 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
942 struct blkg_rwstat_sample rwstat;
943 u64 rbytes, wbytes, rios, wios, dbytes, dios;
944 size_t size = seq_get_buf(sf, &buf), off = 0;
946 bool has_stats = false;
948 spin_lock_irq(&blkg->q->queue_lock);
953 dname = blkg_dev_name(blkg);
958 * Hooray string manipulation, count is the size written NOT
959 * INCLUDING THE \0, so size is now count+1 less than what we
960 * had before, but we want to start writing the next bit from
961 * the \0 so we only add count to buf.
963 off += scnprintf(buf+off, size-off, "%s ", dname);
965 blkg_rwstat_recursive_sum(blkg, NULL,
966 offsetof(struct blkcg_gq, stat_bytes), &rwstat);
967 rbytes = rwstat.cnt[BLKG_RWSTAT_READ];
968 wbytes = rwstat.cnt[BLKG_RWSTAT_WRITE];
969 dbytes = rwstat.cnt[BLKG_RWSTAT_DISCARD];
971 blkg_rwstat_recursive_sum(blkg, NULL,
972 offsetof(struct blkcg_gq, stat_ios), &rwstat);
973 rios = rwstat.cnt[BLKG_RWSTAT_READ];
974 wios = rwstat.cnt[BLKG_RWSTAT_WRITE];
975 dios = rwstat.cnt[BLKG_RWSTAT_DISCARD];
977 if (rbytes || wbytes || rios || wios) {
979 off += scnprintf(buf+off, size-off,
980 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
981 rbytes, wbytes, rios, wios,
985 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
987 off += scnprintf(buf+off, size-off,
988 " use_delay=%d delay_nsec=%llu",
989 atomic_read(&blkg->use_delay),
990 (unsigned long long)atomic64_read(&blkg->delay_nsec));
993 for (i = 0; i < BLKCG_MAX_POLS; i++) {
994 struct blkcg_policy *pol = blkcg_policy[i];
997 if (!blkg->pd[i] || !pol->pd_stat_fn)
1000 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1007 if (off < size - 1) {
1008 off += scnprintf(buf+off, size-off, "\n");
1009 seq_commit(sf, off);
1015 spin_unlock_irq(&blkg->q->queue_lock);
1022 static struct cftype blkcg_files[] = {
1025 .flags = CFTYPE_NOT_ON_ROOT,
1026 .seq_show = blkcg_print_stat,
1031 static struct cftype blkcg_legacy_files[] = {
1033 .name = "reset_stats",
1034 .write_u64 = blkcg_reset_stats,
1040 * blkcg destruction is a three-stage process.
1042 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1043 * which offlines writeback. Here we tie the next stage of blkg destruction
1044 * to the completion of writeback associated with the blkcg. This lets us
1045 * avoid punting potentially large amounts of outstanding writeback to root
1046 * while maintaining any ongoing policies. The next stage is triggered when
1047 * the nr_cgwbs count goes to zero.
1049 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1050 * and handles the destruction of blkgs. Here the css reference held by
1051 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1052 * This work may occur in cgwb_release_workfn() on the cgwb_release
1053 * workqueue. Any submitted ios that fail to get the blkg ref will be
1054 * punted to the root_blkg.
1056 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1057 * This finally frees the blkcg.
1061 * blkcg_css_offline - cgroup css_offline callback
1062 * @css: css of interest
1064 * This function is called when @css is about to go away. Here the cgwbs are
1065 * offlined first and only once writeback associated with the blkcg has
1066 * finished do we start step 2 (see above).
1068 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1070 struct blkcg *blkcg = css_to_blkcg(css);
1072 /* this prevents anyone from attaching or migrating to this blkcg */
1073 wb_blkcg_offline(blkcg);
1075 /* put the base cgwb reference allowing step 2 to be triggered */
1076 blkcg_cgwb_put(blkcg);
1080 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1081 * @blkcg: blkcg of interest
1083 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1084 * is nested inside q lock, this function performs reverse double lock dancing.
1085 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1086 * blkcg_css_free to eventually be called.
1088 * This is the blkcg counterpart of ioc_release_fn().
1090 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1094 spin_lock_irq(&blkcg->lock);
1096 while (!hlist_empty(&blkcg->blkg_list)) {
1097 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1098 struct blkcg_gq, blkcg_node);
1099 struct request_queue *q = blkg->q;
1101 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1103 * Given that the system can accumulate a huge number
1104 * of blkgs in pathological cases, check to see if we
1105 * need to rescheduling to avoid softlockup.
1107 spin_unlock_irq(&blkcg->lock);
1109 spin_lock_irq(&blkcg->lock);
1114 spin_unlock(&q->queue_lock);
1117 spin_unlock_irq(&blkcg->lock);
1120 static void blkcg_css_free(struct cgroup_subsys_state *css)
1122 struct blkcg *blkcg = css_to_blkcg(css);
1125 mutex_lock(&blkcg_pol_mutex);
1127 list_del(&blkcg->all_blkcgs_node);
1129 for (i = 0; i < BLKCG_MAX_POLS; i++)
1131 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1133 mutex_unlock(&blkcg_pol_mutex);
1138 static struct cgroup_subsys_state *
1139 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1141 struct blkcg *blkcg;
1142 struct cgroup_subsys_state *ret;
1145 mutex_lock(&blkcg_pol_mutex);
1148 blkcg = &blkcg_root;
1150 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1152 ret = ERR_PTR(-ENOMEM);
1157 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1158 struct blkcg_policy *pol = blkcg_policy[i];
1159 struct blkcg_policy_data *cpd;
1162 * If the policy hasn't been attached yet, wait for it
1163 * to be attached before doing anything else. Otherwise,
1164 * check if the policy requires any specific per-cgroup
1165 * data: if it does, allocate and initialize it.
1167 if (!pol || !pol->cpd_alloc_fn)
1170 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1172 ret = ERR_PTR(-ENOMEM);
1175 blkcg->cpd[i] = cpd;
1178 if (pol->cpd_init_fn)
1179 pol->cpd_init_fn(cpd);
1182 spin_lock_init(&blkcg->lock);
1183 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1184 INIT_HLIST_HEAD(&blkcg->blkg_list);
1185 #ifdef CONFIG_CGROUP_WRITEBACK
1186 INIT_LIST_HEAD(&blkcg->cgwb_list);
1187 refcount_set(&blkcg->cgwb_refcnt, 1);
1189 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1191 mutex_unlock(&blkcg_pol_mutex);
1195 for (i--; i >= 0; i--)
1197 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1199 if (blkcg != &blkcg_root)
1202 mutex_unlock(&blkcg_pol_mutex);
1207 * blkcg_init_queue - initialize blkcg part of request queue
1208 * @q: request_queue to initialize
1210 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1211 * part of new request_queue @q.
1214 * 0 on success, -errno on failure.
1216 int blkcg_init_queue(struct request_queue *q)
1218 struct blkcg_gq *new_blkg, *blkg;
1222 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1226 preloaded = !radix_tree_preload(GFP_KERNEL);
1228 /* Make sure the root blkg exists. */
1230 spin_lock_irq(&q->queue_lock);
1231 blkg = blkg_create(&blkcg_root, q, new_blkg);
1234 q->root_blkg = blkg;
1235 spin_unlock_irq(&q->queue_lock);
1239 radix_tree_preload_end();
1241 ret = blk_throtl_init(q);
1243 goto err_destroy_all;
1245 ret = blk_iolatency_init(q);
1248 goto err_destroy_all;
1253 blkg_destroy_all(q);
1256 spin_unlock_irq(&q->queue_lock);
1259 radix_tree_preload_end();
1260 return PTR_ERR(blkg);
1264 * blkcg_drain_queue - drain blkcg part of request_queue
1265 * @q: request_queue to drain
1267 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1269 void blkcg_drain_queue(struct request_queue *q)
1271 lockdep_assert_held(&q->queue_lock);
1274 * @q could be exiting and already have destroyed all blkgs as
1275 * indicated by NULL root_blkg. If so, don't confuse policies.
1280 blk_throtl_drain(q);
1284 * blkcg_exit_queue - exit and release blkcg part of request_queue
1285 * @q: request_queue being released
1287 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1289 void blkcg_exit_queue(struct request_queue *q)
1291 blkg_destroy_all(q);
1296 * We cannot support shared io contexts, as we have no mean to support
1297 * two tasks with the same ioc in two different groups without major rework
1298 * of the main cic data structures. For now we allow a task to change
1299 * its cgroup only if it's the only owner of its ioc.
1301 static int blkcg_can_attach(struct cgroup_taskset *tset)
1303 struct task_struct *task;
1304 struct cgroup_subsys_state *dst_css;
1305 struct io_context *ioc;
1308 /* task_lock() is needed to avoid races with exit_io_context() */
1309 cgroup_taskset_for_each(task, dst_css, tset) {
1311 ioc = task->io_context;
1312 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1321 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1325 mutex_lock(&blkcg_pol_mutex);
1327 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1328 struct blkcg_policy *pol = blkcg_policy[i];
1329 struct blkcg *blkcg;
1331 if (!pol || !pol->cpd_bind_fn)
1334 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1335 if (blkcg->cpd[pol->plid])
1336 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1338 mutex_unlock(&blkcg_pol_mutex);
1341 static void blkcg_exit(struct task_struct *tsk)
1343 if (tsk->throttle_queue)
1344 blk_put_queue(tsk->throttle_queue);
1345 tsk->throttle_queue = NULL;
1348 struct cgroup_subsys io_cgrp_subsys = {
1349 .css_alloc = blkcg_css_alloc,
1350 .css_offline = blkcg_css_offline,
1351 .css_free = blkcg_css_free,
1352 .can_attach = blkcg_can_attach,
1354 .dfl_cftypes = blkcg_files,
1355 .legacy_cftypes = blkcg_legacy_files,
1356 .legacy_name = "blkio",
1360 * This ensures that, if available, memcg is automatically enabled
1361 * together on the default hierarchy so that the owner cgroup can
1362 * be retrieved from writeback pages.
1364 .depends_on = 1 << memory_cgrp_id,
1367 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1370 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1371 * @q: request_queue of interest
1372 * @pol: blkcg policy to activate
1374 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1375 * bypass mode to populate its blkgs with policy_data for @pol.
1377 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1378 * from IO path. Update of each blkg is protected by both queue and blkcg
1379 * locks so that holding either lock and testing blkcg_policy_enabled() is
1380 * always enough for dereferencing policy data.
1382 * The caller is responsible for synchronizing [de]activations and policy
1383 * [un]registerations. Returns 0 on success, -errno on failure.
1385 int blkcg_activate_policy(struct request_queue *q,
1386 const struct blkcg_policy *pol)
1388 struct blkg_policy_data *pd_prealloc = NULL;
1389 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1392 if (blkcg_policy_enabled(q, pol))
1396 blk_mq_freeze_queue(q);
1398 spin_lock_irq(&q->queue_lock);
1400 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1401 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1402 struct blkg_policy_data *pd;
1404 if (blkg->pd[pol->plid])
1407 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1408 if (blkg == pinned_blkg) {
1412 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1418 * GFP_NOWAIT failed. Free the existing one and
1419 * prealloc for @blkg w/ GFP_KERNEL.
1422 blkg_put(pinned_blkg);
1426 spin_unlock_irq(&q->queue_lock);
1429 pol->pd_free_fn(pd_prealloc);
1430 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1438 blkg->pd[pol->plid] = pd;
1440 pd->plid = pol->plid;
1443 /* all allocated, init in the same order */
1444 if (pol->pd_init_fn)
1445 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1446 pol->pd_init_fn(blkg->pd[pol->plid]);
1448 __set_bit(pol->plid, q->blkcg_pols);
1451 spin_unlock_irq(&q->queue_lock);
1454 blk_mq_unfreeze_queue(q);
1456 blkg_put(pinned_blkg);
1458 pol->pd_free_fn(pd_prealloc);
1462 /* alloc failed, nothing's initialized yet, free everything */
1463 spin_lock_irq(&q->queue_lock);
1464 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1465 struct blkcg *blkcg = blkg->blkcg;
1467 spin_lock(&blkcg->lock);
1468 if (blkg->pd[pol->plid]) {
1469 pol->pd_free_fn(blkg->pd[pol->plid]);
1470 blkg->pd[pol->plid] = NULL;
1472 spin_unlock(&blkcg->lock);
1474 spin_unlock_irq(&q->queue_lock);
1478 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1481 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1482 * @q: request_queue of interest
1483 * @pol: blkcg policy to deactivate
1485 * Deactivate @pol on @q. Follows the same synchronization rules as
1486 * blkcg_activate_policy().
1488 void blkcg_deactivate_policy(struct request_queue *q,
1489 const struct blkcg_policy *pol)
1491 struct blkcg_gq *blkg;
1493 if (!blkcg_policy_enabled(q, pol))
1497 blk_mq_freeze_queue(q);
1499 spin_lock_irq(&q->queue_lock);
1501 __clear_bit(pol->plid, q->blkcg_pols);
1503 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1504 struct blkcg *blkcg = blkg->blkcg;
1506 spin_lock(&blkcg->lock);
1507 if (blkg->pd[pol->plid]) {
1508 if (pol->pd_offline_fn)
1509 pol->pd_offline_fn(blkg->pd[pol->plid]);
1510 pol->pd_free_fn(blkg->pd[pol->plid]);
1511 blkg->pd[pol->plid] = NULL;
1513 spin_unlock(&blkcg->lock);
1516 spin_unlock_irq(&q->queue_lock);
1519 blk_mq_unfreeze_queue(q);
1521 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1524 * blkcg_policy_register - register a blkcg policy
1525 * @pol: blkcg policy to register
1527 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1528 * successful registration. Returns 0 on success and -errno on failure.
1530 int blkcg_policy_register(struct blkcg_policy *pol)
1532 struct blkcg *blkcg;
1535 mutex_lock(&blkcg_pol_register_mutex);
1536 mutex_lock(&blkcg_pol_mutex);
1538 /* find an empty slot */
1540 for (i = 0; i < BLKCG_MAX_POLS; i++)
1541 if (!blkcg_policy[i])
1543 if (i >= BLKCG_MAX_POLS) {
1544 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1548 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1549 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1550 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1555 blkcg_policy[pol->plid] = pol;
1557 /* allocate and install cpd's */
1558 if (pol->cpd_alloc_fn) {
1559 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1560 struct blkcg_policy_data *cpd;
1562 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1566 blkcg->cpd[pol->plid] = cpd;
1568 cpd->plid = pol->plid;
1569 if (pol->cpd_init_fn)
1570 pol->cpd_init_fn(cpd);
1574 mutex_unlock(&blkcg_pol_mutex);
1576 /* everything is in place, add intf files for the new policy */
1577 if (pol->dfl_cftypes)
1578 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1580 if (pol->legacy_cftypes)
1581 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1582 pol->legacy_cftypes));
1583 mutex_unlock(&blkcg_pol_register_mutex);
1587 if (pol->cpd_free_fn) {
1588 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1589 if (blkcg->cpd[pol->plid]) {
1590 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1591 blkcg->cpd[pol->plid] = NULL;
1595 blkcg_policy[pol->plid] = NULL;
1597 mutex_unlock(&blkcg_pol_mutex);
1598 mutex_unlock(&blkcg_pol_register_mutex);
1601 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1604 * blkcg_policy_unregister - unregister a blkcg policy
1605 * @pol: blkcg policy to unregister
1607 * Undo blkcg_policy_register(@pol). Might sleep.
1609 void blkcg_policy_unregister(struct blkcg_policy *pol)
1611 struct blkcg *blkcg;
1613 mutex_lock(&blkcg_pol_register_mutex);
1615 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1618 /* kill the intf files first */
1619 if (pol->dfl_cftypes)
1620 cgroup_rm_cftypes(pol->dfl_cftypes);
1621 if (pol->legacy_cftypes)
1622 cgroup_rm_cftypes(pol->legacy_cftypes);
1624 /* remove cpds and unregister */
1625 mutex_lock(&blkcg_pol_mutex);
1627 if (pol->cpd_free_fn) {
1628 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1629 if (blkcg->cpd[pol->plid]) {
1630 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1631 blkcg->cpd[pol->plid] = NULL;
1635 blkcg_policy[pol->plid] = NULL;
1637 mutex_unlock(&blkcg_pol_mutex);
1639 mutex_unlock(&blkcg_pol_register_mutex);
1641 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1643 bool __blkcg_punt_bio_submit(struct bio *bio)
1645 struct blkcg_gq *blkg = bio->bi_blkg;
1647 /* consume the flag first */
1648 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1650 /* never bounce for the root cgroup */
1654 spin_lock_bh(&blkg->async_bio_lock);
1655 bio_list_add(&blkg->async_bios, bio);
1656 spin_unlock_bh(&blkg->async_bio_lock);
1658 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1663 * Scale the accumulated delay based on how long it has been since we updated
1664 * the delay. We only call this when we are adding delay, in case it's been a
1665 * while since we added delay, and when we are checking to see if we need to
1666 * delay a task, to account for any delays that may have occurred.
1668 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1670 u64 old = atomic64_read(&blkg->delay_start);
1673 * We only want to scale down every second. The idea here is that we
1674 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1675 * time window. We only want to throttle tasks for recent delay that
1676 * has occurred, in 1 second time windows since that's the maximum
1677 * things can be throttled. We save the current delay window in
1678 * blkg->last_delay so we know what amount is still left to be charged
1679 * to the blkg from this point onward. blkg->last_use keeps track of
1680 * the use_delay counter. The idea is if we're unthrottling the blkg we
1681 * are ok with whatever is happening now, and we can take away more of
1682 * the accumulated delay as we've already throttled enough that
1683 * everybody is happy with their IO latencies.
1685 if (time_before64(old + NSEC_PER_SEC, now) &&
1686 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1687 u64 cur = atomic64_read(&blkg->delay_nsec);
1688 u64 sub = min_t(u64, blkg->last_delay, now - old);
1689 int cur_use = atomic_read(&blkg->use_delay);
1692 * We've been unthrottled, subtract a larger chunk of our
1693 * accumulated delay.
1695 if (cur_use < blkg->last_use)
1696 sub = max_t(u64, sub, blkg->last_delay >> 1);
1699 * This shouldn't happen, but handle it anyway. Our delay_nsec
1700 * should only ever be growing except here where we subtract out
1701 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1702 * rather not end up with negative numbers.
1704 if (unlikely(cur < sub)) {
1705 atomic64_set(&blkg->delay_nsec, 0);
1706 blkg->last_delay = 0;
1708 atomic64_sub(sub, &blkg->delay_nsec);
1709 blkg->last_delay = cur - sub;
1711 blkg->last_use = cur_use;
1716 * This is called when we want to actually walk up the hierarchy and check to
1717 * see if we need to throttle, and then actually throttle if there is some
1718 * accumulated delay. This should only be called upon return to user space so
1719 * we're not holding some lock that would induce a priority inversion.
1721 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1723 unsigned long pflags;
1724 u64 now = ktime_to_ns(ktime_get());
1729 while (blkg->parent) {
1730 if (atomic_read(&blkg->use_delay)) {
1731 blkcg_scale_delay(blkg, now);
1732 delay_nsec = max_t(u64, delay_nsec,
1733 atomic64_read(&blkg->delay_nsec));
1735 blkg = blkg->parent;
1742 * Let's not sleep for all eternity if we've amassed a huge delay.
1743 * Swapping or metadata IO can accumulate 10's of seconds worth of
1744 * delay, and we want userspace to be able to do _something_ so cap the
1745 * delays at 1 second. If there's 10's of seconds worth of delay then
1746 * the tasks will be delayed for 1 second for every syscall.
1748 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1751 psi_memstall_enter(&pflags);
1753 exp = ktime_add_ns(now, delay_nsec);
1754 tok = io_schedule_prepare();
1756 __set_current_state(TASK_KILLABLE);
1757 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1759 } while (!fatal_signal_pending(current));
1760 io_schedule_finish(tok);
1763 psi_memstall_leave(&pflags);
1767 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1769 * This is only called if we've been marked with set_notify_resume(). Obviously
1770 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1771 * check to see if current->throttle_queue is set and if not this doesn't do
1772 * anything. This should only ever be called by the resume code, it's not meant
1773 * to be called by people willy-nilly as it will actually do the work to
1774 * throttle the task if it is setup for throttling.
1776 void blkcg_maybe_throttle_current(void)
1778 struct request_queue *q = current->throttle_queue;
1779 struct cgroup_subsys_state *css;
1780 struct blkcg *blkcg;
1781 struct blkcg_gq *blkg;
1782 bool use_memdelay = current->use_memdelay;
1787 current->throttle_queue = NULL;
1788 current->use_memdelay = false;
1791 css = kthread_blkcg();
1793 blkcg = css_to_blkcg(css);
1795 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1799 blkg = blkg_lookup(blkcg, q);
1802 if (!blkg_tryget(blkg))
1806 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1816 * blkcg_schedule_throttle - this task needs to check for throttling
1817 * @q: the request queue IO was submitted on
1818 * @use_memdelay: do we charge this to memory delay for PSI
1820 * This is called by the IO controller when we know there's delay accumulated
1821 * for the blkg for this task. We do not pass the blkg because there are places
1822 * we call this that may not have that information, the swapping code for
1823 * instance will only have a request_queue at that point. This set's the
1824 * notify_resume for the task to check and see if it requires throttling before
1825 * returning to user space.
1827 * We will only schedule once per syscall. You can call this over and over
1828 * again and it will only do the check once upon return to user space, and only
1829 * throttle once. If the task needs to be throttled again it'll need to be
1830 * re-set at the next time we see the task.
1832 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1834 if (unlikely(current->flags & PF_KTHREAD))
1837 if (!blk_get_queue(q))
1840 if (current->throttle_queue)
1841 blk_put_queue(current->throttle_queue);
1842 current->throttle_queue = q;
1844 current->use_memdelay = use_memdelay;
1845 set_notify_resume(current);
1849 * blkcg_add_delay - add delay to this blkg
1850 * @blkg: blkg of interest
1851 * @now: the current time in nanoseconds
1852 * @delta: how many nanoseconds of delay to add
1854 * Charge @delta to the blkg's current delay accumulation. This is used to
1855 * throttle tasks if an IO controller thinks we need more throttling.
1857 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1859 blkcg_scale_delay(blkg, now);
1860 atomic64_add(delta, &blkg->delay_nsec);
1863 static int __init blkcg_init(void)
1865 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1866 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1867 WQ_UNBOUND | WQ_SYSFS, 0);
1868 if (!blkcg_punt_bio_wq)
1872 subsys_initcall(blkcg_init);
1874 module_param(blkcg_debug_stats, bool, 0644);
1875 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");