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);
99 wb_congested_put(blkg->wb_congested);
105 * A group is RCU protected, but having an rcu lock does not mean that one
106 * can access all the fields of blkg and assume these are valid. For
107 * example, don't try to follow throtl_data and request queue links.
109 * Having a reference to blkg under an rcu allows accesses to only values
110 * local to groups like group stats and group rate limits.
112 static void blkg_release(struct percpu_ref *ref)
114 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
116 call_rcu(&blkg->rcu_head, __blkg_release);
119 static void blkg_async_bio_workfn(struct work_struct *work)
121 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
123 struct bio_list bios = BIO_EMPTY_LIST;
126 /* as long as there are pending bios, @blkg can't go away */
127 spin_lock_bh(&blkg->async_bio_lock);
128 bio_list_merge(&bios, &blkg->async_bios);
129 bio_list_init(&blkg->async_bios);
130 spin_unlock_bh(&blkg->async_bio_lock);
132 while ((bio = bio_list_pop(&bios)))
137 * blkg_alloc - allocate a blkg
138 * @blkcg: block cgroup the new blkg is associated with
139 * @q: request_queue the new blkg is associated with
140 * @gfp_mask: allocation mask to use
142 * Allocate a new blkg assocating @blkcg and @q.
144 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
147 struct blkcg_gq *blkg;
150 /* alloc and init base part */
151 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
155 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
158 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
159 if (!blkg->iostat_cpu)
163 INIT_LIST_HEAD(&blkg->q_node);
164 spin_lock_init(&blkg->async_bio_lock);
165 bio_list_init(&blkg->async_bios);
166 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
169 u64_stats_init(&blkg->iostat.sync);
170 for_each_possible_cpu(cpu)
171 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
173 for (i = 0; i < BLKCG_MAX_POLS; i++) {
174 struct blkcg_policy *pol = blkcg_policy[i];
175 struct blkg_policy_data *pd;
177 if (!blkcg_policy_enabled(q, pol))
180 /* alloc per-policy data and attach it to blkg */
181 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
197 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
198 struct request_queue *q, bool update_hint)
200 struct blkcg_gq *blkg;
203 * Hint didn't match. Look up from the radix tree. Note that the
204 * hint can only be updated under queue_lock as otherwise @blkg
205 * could have already been removed from blkg_tree. The caller is
206 * responsible for grabbing queue_lock if @update_hint.
208 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
209 if (blkg && blkg->q == q) {
211 lockdep_assert_held(&q->queue_lock);
212 rcu_assign_pointer(blkcg->blkg_hint, blkg);
219 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
222 * If @new_blkg is %NULL, this function tries to allocate a new one as
223 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
225 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
226 struct request_queue *q,
227 struct blkcg_gq *new_blkg)
229 struct blkcg_gq *blkg;
230 struct bdi_writeback_congested *wb_congested;
233 WARN_ON_ONCE(!rcu_read_lock_held());
234 lockdep_assert_held(&q->queue_lock);
236 /* request_queue is dying, do not create/recreate a blkg */
237 if (blk_queue_dying(q)) {
242 /* blkg holds a reference to blkcg */
243 if (!css_tryget_online(&blkcg->css)) {
248 wb_congested = wb_congested_get_create(q->backing_dev_info,
250 GFP_NOWAIT | __GFP_NOWARN);
258 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
259 if (unlikely(!new_blkg)) {
261 goto err_put_congested;
265 blkg->wb_congested = wb_congested;
268 if (blkcg_parent(blkcg)) {
269 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
270 if (WARN_ON_ONCE(!blkg->parent)) {
272 goto err_put_congested;
274 blkg_get(blkg->parent);
277 /* invoke per-policy init */
278 for (i = 0; i < BLKCG_MAX_POLS; i++) {
279 struct blkcg_policy *pol = blkcg_policy[i];
281 if (blkg->pd[i] && pol->pd_init_fn)
282 pol->pd_init_fn(blkg->pd[i]);
286 spin_lock(&blkcg->lock);
287 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
289 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
290 list_add(&blkg->q_node, &q->blkg_list);
292 for (i = 0; i < BLKCG_MAX_POLS; i++) {
293 struct blkcg_policy *pol = blkcg_policy[i];
295 if (blkg->pd[i] && pol->pd_online_fn)
296 pol->pd_online_fn(blkg->pd[i]);
300 spin_unlock(&blkcg->lock);
305 /* @blkg failed fully initialized, use the usual release path */
310 wb_congested_put(wb_congested);
312 css_put(&blkcg->css);
319 * __blkg_lookup_create - lookup blkg, try to create one if not there
320 * @blkcg: blkcg of interest
321 * @q: request_queue of interest
323 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
324 * create one. blkg creation is performed recursively from blkcg_root such
325 * that all non-root blkg's have access to the parent blkg. This function
326 * should be called under RCU read lock and @q->queue_lock.
328 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
331 struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
332 struct request_queue *q)
334 struct blkcg_gq *blkg;
336 WARN_ON_ONCE(!rcu_read_lock_held());
337 lockdep_assert_held(&q->queue_lock);
339 blkg = __blkg_lookup(blkcg, q, true);
344 * Create blkgs walking down from blkcg_root to @blkcg, so that all
345 * non-root blkgs have access to their parents. Returns the closest
346 * blkg to the intended blkg should blkg_create() fail.
349 struct blkcg *pos = blkcg;
350 struct blkcg *parent = blkcg_parent(blkcg);
351 struct blkcg_gq *ret_blkg = q->root_blkg;
354 blkg = __blkg_lookup(parent, q, false);
356 /* remember closest blkg */
361 parent = blkcg_parent(parent);
364 blkg = blkg_create(pos, q, NULL);
373 * blkg_lookup_create - find or create a blkg
374 * @blkcg: target block cgroup
375 * @q: target request_queue
377 * This looks up or creates the blkg representing the unique pair
378 * of the blkcg and the request_queue.
380 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
381 struct request_queue *q)
383 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
385 if (unlikely(!blkg)) {
388 spin_lock_irqsave(&q->queue_lock, flags);
389 blkg = __blkg_lookup_create(blkcg, q);
390 spin_unlock_irqrestore(&q->queue_lock, flags);
396 static void blkg_destroy(struct blkcg_gq *blkg)
398 struct blkcg *blkcg = blkg->blkcg;
401 lockdep_assert_held(&blkg->q->queue_lock);
402 lockdep_assert_held(&blkcg->lock);
404 /* Something wrong if we are trying to remove same group twice */
405 WARN_ON_ONCE(list_empty(&blkg->q_node));
406 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
408 for (i = 0; i < BLKCG_MAX_POLS; i++) {
409 struct blkcg_policy *pol = blkcg_policy[i];
411 if (blkg->pd[i] && pol->pd_offline_fn)
412 pol->pd_offline_fn(blkg->pd[i]);
415 blkg->online = false;
417 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
418 list_del_init(&blkg->q_node);
419 hlist_del_init_rcu(&blkg->blkcg_node);
422 * Both setting lookup hint to and clearing it from @blkg are done
423 * under queue_lock. If it's not pointing to @blkg now, it never
424 * will. Hint assignment itself can race safely.
426 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
427 rcu_assign_pointer(blkcg->blkg_hint, NULL);
430 * Put the reference taken at the time of creation so that when all
431 * queues are gone, group can be destroyed.
433 percpu_ref_kill(&blkg->refcnt);
437 * blkg_destroy_all - destroy all blkgs associated with a request_queue
438 * @q: request_queue of interest
440 * Destroy all blkgs associated with @q.
442 static void blkg_destroy_all(struct request_queue *q)
444 struct blkcg_gq *blkg, *n;
446 spin_lock_irq(&q->queue_lock);
447 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
448 struct blkcg *blkcg = blkg->blkcg;
450 spin_lock(&blkcg->lock);
452 spin_unlock(&blkcg->lock);
456 spin_unlock_irq(&q->queue_lock);
459 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
460 struct cftype *cftype, u64 val)
462 struct blkcg *blkcg = css_to_blkcg(css);
463 struct blkcg_gq *blkg;
466 mutex_lock(&blkcg_pol_mutex);
467 spin_lock_irq(&blkcg->lock);
470 * Note that stat reset is racy - it doesn't synchronize against
471 * stat updates. This is a debug feature which shouldn't exist
472 * anyway. If you get hit by a race, retry.
474 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
475 for_each_possible_cpu(cpu) {
476 struct blkg_iostat_set *bis =
477 per_cpu_ptr(blkg->iostat_cpu, cpu);
478 memset(bis, 0, sizeof(*bis));
480 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
482 for (i = 0; i < BLKCG_MAX_POLS; i++) {
483 struct blkcg_policy *pol = blkcg_policy[i];
485 if (blkg->pd[i] && pol->pd_reset_stats_fn)
486 pol->pd_reset_stats_fn(blkg->pd[i]);
490 spin_unlock_irq(&blkcg->lock);
491 mutex_unlock(&blkcg_pol_mutex);
495 const char *blkg_dev_name(struct blkcg_gq *blkg)
497 /* some drivers (floppy) instantiate a queue w/o disk registered */
498 if (blkg->q->backing_dev_info->dev)
499 return dev_name(blkg->q->backing_dev_info->dev);
504 * blkcg_print_blkgs - helper for printing per-blkg data
505 * @sf: seq_file to print to
506 * @blkcg: blkcg of interest
507 * @prfill: fill function to print out a blkg
508 * @pol: policy in question
509 * @data: data to be passed to @prfill
510 * @show_total: to print out sum of prfill return values or not
512 * This function invokes @prfill on each blkg of @blkcg if pd for the
513 * policy specified by @pol exists. @prfill is invoked with @sf, the
514 * policy data and @data and the matching queue lock held. If @show_total
515 * is %true, the sum of the return values from @prfill is printed with
516 * "Total" label at the end.
518 * This is to be used to construct print functions for
519 * cftype->read_seq_string method.
521 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
522 u64 (*prfill)(struct seq_file *,
523 struct blkg_policy_data *, int),
524 const struct blkcg_policy *pol, int data,
527 struct blkcg_gq *blkg;
531 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
532 spin_lock_irq(&blkg->q->queue_lock);
533 if (blkcg_policy_enabled(blkg->q, pol))
534 total += prfill(sf, blkg->pd[pol->plid], data);
535 spin_unlock_irq(&blkg->q->queue_lock);
540 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
542 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
545 * __blkg_prfill_u64 - prfill helper for a single u64 value
546 * @sf: seq_file to print to
547 * @pd: policy private data of interest
550 * Print @v to @sf for the device assocaited with @pd.
552 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
554 const char *dname = blkg_dev_name(pd->blkg);
559 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
562 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
564 /* Performs queue bypass and policy enabled checks then looks up blkg. */
565 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
566 const struct blkcg_policy *pol,
567 struct request_queue *q)
569 WARN_ON_ONCE(!rcu_read_lock_held());
570 lockdep_assert_held(&q->queue_lock);
572 if (!blkcg_policy_enabled(q, pol))
573 return ERR_PTR(-EOPNOTSUPP);
574 return __blkg_lookup(blkcg, q, true /* update_hint */);
578 * blkg_conf_prep - parse and prepare for per-blkg config update
579 * @inputp: input string pointer
581 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
582 * from @input and get and return the matching gendisk. *@inputp is
583 * updated to point past the device node prefix. Returns an ERR_PTR()
586 * Use this function iff blkg_conf_prep() can't be used for some reason.
588 struct gendisk *blkcg_conf_get_disk(char **inputp)
590 char *input = *inputp;
591 unsigned int major, minor;
592 struct gendisk *disk;
595 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
596 return ERR_PTR(-EINVAL);
599 if (!isspace(*input))
600 return ERR_PTR(-EINVAL);
601 input = skip_spaces(input);
603 disk = get_gendisk(MKDEV(major, minor), &part);
605 return ERR_PTR(-ENODEV);
607 put_disk_and_module(disk);
608 return ERR_PTR(-ENODEV);
616 * blkg_conf_prep - parse and prepare for per-blkg config update
617 * @blkcg: target block cgroup
618 * @pol: target policy
619 * @input: input string
620 * @ctx: blkg_conf_ctx to be filled
622 * Parse per-blkg config update from @input and initialize @ctx with the
623 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
624 * part of @input following MAJ:MIN. This function returns with RCU read
625 * lock and queue lock held and must be paired with blkg_conf_finish().
627 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
628 char *input, struct blkg_conf_ctx *ctx)
629 __acquires(rcu) __acquires(&disk->queue->queue_lock)
631 struct gendisk *disk;
632 struct request_queue *q;
633 struct blkcg_gq *blkg;
636 disk = blkcg_conf_get_disk(&input);
638 return PTR_ERR(disk);
643 spin_lock_irq(&q->queue_lock);
645 blkg = blkg_lookup_check(blkcg, pol, q);
655 * Create blkgs walking down from blkcg_root to @blkcg, so that all
656 * non-root blkgs have access to their parents.
659 struct blkcg *pos = blkcg;
660 struct blkcg *parent;
661 struct blkcg_gq *new_blkg;
663 parent = blkcg_parent(blkcg);
664 while (parent && !__blkg_lookup(parent, q, false)) {
666 parent = blkcg_parent(parent);
669 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
670 spin_unlock_irq(&q->queue_lock);
673 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
674 if (unlikely(!new_blkg)) {
680 spin_lock_irq(&q->queue_lock);
682 blkg = blkg_lookup_check(pos, pol, q);
691 blkg = blkg_create(pos, q, new_blkg);
708 spin_unlock_irq(&q->queue_lock);
711 put_disk_and_module(disk);
713 * If queue was bypassing, we should retry. Do so after a
714 * short msleep(). It isn't strictly necessary but queue
715 * can be bypassing for some time and it's always nice to
716 * avoid busy looping.
720 ret = restart_syscall();
724 EXPORT_SYMBOL_GPL(blkg_conf_prep);
727 * blkg_conf_finish - finish up per-blkg config update
728 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
730 * Finish up after per-blkg config update. This function must be paired
731 * with blkg_conf_prep().
733 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
734 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
736 spin_unlock_irq(&ctx->disk->queue->queue_lock);
738 put_disk_and_module(ctx->disk);
740 EXPORT_SYMBOL_GPL(blkg_conf_finish);
742 static int blkcg_print_stat(struct seq_file *sf, void *v)
744 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
745 struct blkcg_gq *blkg;
747 cgroup_rstat_flush(blkcg->css.cgroup);
750 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
751 struct blkg_iostat_set *bis = &blkg->iostat;
754 u64 rbytes, wbytes, rios, wios, dbytes, dios;
755 size_t size = seq_get_buf(sf, &buf), off = 0;
757 bool has_stats = false;
760 spin_lock_irq(&blkg->q->queue_lock);
765 dname = blkg_dev_name(blkg);
770 * Hooray string manipulation, count is the size written NOT
771 * INCLUDING THE \0, so size is now count+1 less than what we
772 * had before, but we want to start writing the next bit from
773 * the \0 so we only add count to buf.
775 off += scnprintf(buf+off, size-off, "%s ", dname);
778 seq = u64_stats_fetch_begin(&bis->sync);
780 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
781 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
782 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
783 rios = bis->cur.ios[BLKG_IOSTAT_READ];
784 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
785 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
786 } while (u64_stats_fetch_retry(&bis->sync, seq));
788 if (rbytes || wbytes || rios || wios) {
790 off += scnprintf(buf+off, size-off,
791 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
792 rbytes, wbytes, rios, wios,
796 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
798 off += scnprintf(buf+off, size-off,
799 " use_delay=%d delay_nsec=%llu",
800 atomic_read(&blkg->use_delay),
801 (unsigned long long)atomic64_read(&blkg->delay_nsec));
804 for (i = 0; i < BLKCG_MAX_POLS; i++) {
805 struct blkcg_policy *pol = blkcg_policy[i];
808 if (!blkg->pd[i] || !pol->pd_stat_fn)
811 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
818 if (off < size - 1) {
819 off += scnprintf(buf+off, size-off, "\n");
826 spin_unlock_irq(&blkg->q->queue_lock);
833 static struct cftype blkcg_files[] = {
836 .flags = CFTYPE_NOT_ON_ROOT,
837 .seq_show = blkcg_print_stat,
842 static struct cftype blkcg_legacy_files[] = {
844 .name = "reset_stats",
845 .write_u64 = blkcg_reset_stats,
851 * blkcg destruction is a three-stage process.
853 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
854 * which offlines writeback. Here we tie the next stage of blkg destruction
855 * to the completion of writeback associated with the blkcg. This lets us
856 * avoid punting potentially large amounts of outstanding writeback to root
857 * while maintaining any ongoing policies. The next stage is triggered when
858 * the nr_cgwbs count goes to zero.
860 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
861 * and handles the destruction of blkgs. Here the css reference held by
862 * the blkg is put back eventually allowing blkcg_css_free() to be called.
863 * This work may occur in cgwb_release_workfn() on the cgwb_release
864 * workqueue. Any submitted ios that fail to get the blkg ref will be
865 * punted to the root_blkg.
867 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
868 * This finally frees the blkcg.
872 * blkcg_css_offline - cgroup css_offline callback
873 * @css: css of interest
875 * This function is called when @css is about to go away. Here the cgwbs are
876 * offlined first and only once writeback associated with the blkcg has
877 * finished do we start step 2 (see above).
879 static void blkcg_css_offline(struct cgroup_subsys_state *css)
881 struct blkcg *blkcg = css_to_blkcg(css);
883 /* this prevents anyone from attaching or migrating to this blkcg */
884 wb_blkcg_offline(blkcg);
886 /* put the base cgwb reference allowing step 2 to be triggered */
887 blkcg_cgwb_put(blkcg);
891 * blkcg_destroy_blkgs - responsible for shooting down blkgs
892 * @blkcg: blkcg of interest
894 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
895 * is nested inside q lock, this function performs reverse double lock dancing.
896 * Destroying the blkgs releases the reference held on the blkcg's css allowing
897 * blkcg_css_free to eventually be called.
899 * This is the blkcg counterpart of ioc_release_fn().
901 void blkcg_destroy_blkgs(struct blkcg *blkcg)
903 spin_lock_irq(&blkcg->lock);
905 while (!hlist_empty(&blkcg->blkg_list)) {
906 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
907 struct blkcg_gq, blkcg_node);
908 struct request_queue *q = blkg->q;
910 if (spin_trylock(&q->queue_lock)) {
912 spin_unlock(&q->queue_lock);
914 spin_unlock_irq(&blkcg->lock);
916 spin_lock_irq(&blkcg->lock);
920 spin_unlock_irq(&blkcg->lock);
923 static void blkcg_css_free(struct cgroup_subsys_state *css)
925 struct blkcg *blkcg = css_to_blkcg(css);
928 mutex_lock(&blkcg_pol_mutex);
930 list_del(&blkcg->all_blkcgs_node);
932 for (i = 0; i < BLKCG_MAX_POLS; i++)
934 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
936 mutex_unlock(&blkcg_pol_mutex);
941 static struct cgroup_subsys_state *
942 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
945 struct cgroup_subsys_state *ret;
948 mutex_lock(&blkcg_pol_mutex);
953 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
955 ret = ERR_PTR(-ENOMEM);
960 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
961 struct blkcg_policy *pol = blkcg_policy[i];
962 struct blkcg_policy_data *cpd;
965 * If the policy hasn't been attached yet, wait for it
966 * to be attached before doing anything else. Otherwise,
967 * check if the policy requires any specific per-cgroup
968 * data: if it does, allocate and initialize it.
970 if (!pol || !pol->cpd_alloc_fn)
973 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
975 ret = ERR_PTR(-ENOMEM);
981 if (pol->cpd_init_fn)
982 pol->cpd_init_fn(cpd);
985 spin_lock_init(&blkcg->lock);
986 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
987 INIT_HLIST_HEAD(&blkcg->blkg_list);
988 #ifdef CONFIG_CGROUP_WRITEBACK
989 INIT_LIST_HEAD(&blkcg->cgwb_list);
990 refcount_set(&blkcg->cgwb_refcnt, 1);
992 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
994 mutex_unlock(&blkcg_pol_mutex);
998 for (i--; i >= 0; i--)
1000 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1002 if (blkcg != &blkcg_root)
1005 mutex_unlock(&blkcg_pol_mutex);
1010 * blkcg_init_queue - initialize blkcg part of request queue
1011 * @q: request_queue to initialize
1013 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1014 * part of new request_queue @q.
1017 * 0 on success, -errno on failure.
1019 int blkcg_init_queue(struct request_queue *q)
1021 struct blkcg_gq *new_blkg, *blkg;
1025 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1029 preloaded = !radix_tree_preload(GFP_KERNEL);
1031 /* Make sure the root blkg exists. */
1033 spin_lock_irq(&q->queue_lock);
1034 blkg = blkg_create(&blkcg_root, q, new_blkg);
1037 q->root_blkg = blkg;
1038 spin_unlock_irq(&q->queue_lock);
1042 radix_tree_preload_end();
1044 ret = blk_iolatency_init(q);
1046 goto err_destroy_all;
1048 ret = blk_throtl_init(q);
1050 goto err_destroy_all;
1054 blkg_destroy_all(q);
1057 spin_unlock_irq(&q->queue_lock);
1060 radix_tree_preload_end();
1061 return PTR_ERR(blkg);
1065 * blkcg_exit_queue - exit and release blkcg part of request_queue
1066 * @q: request_queue being released
1068 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1070 void blkcg_exit_queue(struct request_queue *q)
1072 blkg_destroy_all(q);
1077 * We cannot support shared io contexts, as we have no mean to support
1078 * two tasks with the same ioc in two different groups without major rework
1079 * of the main cic data structures. For now we allow a task to change
1080 * its cgroup only if it's the only owner of its ioc.
1082 static int blkcg_can_attach(struct cgroup_taskset *tset)
1084 struct task_struct *task;
1085 struct cgroup_subsys_state *dst_css;
1086 struct io_context *ioc;
1089 /* task_lock() is needed to avoid races with exit_io_context() */
1090 cgroup_taskset_for_each(task, dst_css, tset) {
1092 ioc = task->io_context;
1093 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1102 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
1106 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
1107 dst->bytes[i] = src->bytes[i];
1108 dst->ios[i] = src->ios[i];
1112 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
1116 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
1117 dst->bytes[i] += src->bytes[i];
1118 dst->ios[i] += src->ios[i];
1122 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
1126 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
1127 dst->bytes[i] -= src->bytes[i];
1128 dst->ios[i] -= src->ios[i];
1132 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1134 struct blkcg *blkcg = css_to_blkcg(css);
1135 struct blkcg_gq *blkg;
1139 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1140 struct blkcg_gq *parent = blkg->parent;
1141 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
1142 struct blkg_iostat cur, delta;
1145 /* fetch the current per-cpu values */
1147 seq = u64_stats_fetch_begin(&bisc->sync);
1148 blkg_iostat_set(&cur, &bisc->cur);
1149 } while (u64_stats_fetch_retry(&bisc->sync, seq));
1151 /* propagate percpu delta to global */
1152 u64_stats_update_begin(&blkg->iostat.sync);
1153 blkg_iostat_set(&delta, &cur);
1154 blkg_iostat_sub(&delta, &bisc->last);
1155 blkg_iostat_add(&blkg->iostat.cur, &delta);
1156 blkg_iostat_add(&bisc->last, &delta);
1157 u64_stats_update_end(&blkg->iostat.sync);
1159 /* propagate global delta to parent */
1161 u64_stats_update_begin(&parent->iostat.sync);
1162 blkg_iostat_set(&delta, &blkg->iostat.cur);
1163 blkg_iostat_sub(&delta, &blkg->iostat.last);
1164 blkg_iostat_add(&parent->iostat.cur, &delta);
1165 blkg_iostat_add(&blkg->iostat.last, &delta);
1166 u64_stats_update_end(&parent->iostat.sync);
1173 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1177 mutex_lock(&blkcg_pol_mutex);
1179 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1180 struct blkcg_policy *pol = blkcg_policy[i];
1181 struct blkcg *blkcg;
1183 if (!pol || !pol->cpd_bind_fn)
1186 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1187 if (blkcg->cpd[pol->plid])
1188 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1190 mutex_unlock(&blkcg_pol_mutex);
1193 static void blkcg_exit(struct task_struct *tsk)
1195 if (tsk->throttle_queue)
1196 blk_put_queue(tsk->throttle_queue);
1197 tsk->throttle_queue = NULL;
1200 struct cgroup_subsys io_cgrp_subsys = {
1201 .css_alloc = blkcg_css_alloc,
1202 .css_offline = blkcg_css_offline,
1203 .css_free = blkcg_css_free,
1204 .can_attach = blkcg_can_attach,
1205 .css_rstat_flush = blkcg_rstat_flush,
1207 .dfl_cftypes = blkcg_files,
1208 .legacy_cftypes = blkcg_legacy_files,
1209 .legacy_name = "blkio",
1213 * This ensures that, if available, memcg is automatically enabled
1214 * together on the default hierarchy so that the owner cgroup can
1215 * be retrieved from writeback pages.
1217 .depends_on = 1 << memory_cgrp_id,
1220 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1223 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1224 * @q: request_queue of interest
1225 * @pol: blkcg policy to activate
1227 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1228 * bypass mode to populate its blkgs with policy_data for @pol.
1230 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1231 * from IO path. Update of each blkg is protected by both queue and blkcg
1232 * locks so that holding either lock and testing blkcg_policy_enabled() is
1233 * always enough for dereferencing policy data.
1235 * The caller is responsible for synchronizing [de]activations and policy
1236 * [un]registerations. Returns 0 on success, -errno on failure.
1238 int blkcg_activate_policy(struct request_queue *q,
1239 const struct blkcg_policy *pol)
1241 struct blkg_policy_data *pd_prealloc = NULL;
1242 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1245 if (blkcg_policy_enabled(q, pol))
1249 blk_mq_freeze_queue(q);
1251 spin_lock_irq(&q->queue_lock);
1253 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1254 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1255 struct blkg_policy_data *pd;
1257 if (blkg->pd[pol->plid])
1260 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1261 if (blkg == pinned_blkg) {
1265 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1271 * GFP_NOWAIT failed. Free the existing one and
1272 * prealloc for @blkg w/ GFP_KERNEL.
1275 blkg_put(pinned_blkg);
1279 spin_unlock_irq(&q->queue_lock);
1282 pol->pd_free_fn(pd_prealloc);
1283 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1291 blkg->pd[pol->plid] = pd;
1293 pd->plid = pol->plid;
1296 /* all allocated, init in the same order */
1297 if (pol->pd_init_fn)
1298 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1299 pol->pd_init_fn(blkg->pd[pol->plid]);
1301 __set_bit(pol->plid, q->blkcg_pols);
1304 spin_unlock_irq(&q->queue_lock);
1307 blk_mq_unfreeze_queue(q);
1309 blkg_put(pinned_blkg);
1311 pol->pd_free_fn(pd_prealloc);
1315 /* alloc failed, nothing's initialized yet, free everything */
1316 spin_lock_irq(&q->queue_lock);
1317 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1318 if (blkg->pd[pol->plid]) {
1319 pol->pd_free_fn(blkg->pd[pol->plid]);
1320 blkg->pd[pol->plid] = NULL;
1323 spin_unlock_irq(&q->queue_lock);
1327 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1330 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1331 * @q: request_queue of interest
1332 * @pol: blkcg policy to deactivate
1334 * Deactivate @pol on @q. Follows the same synchronization rules as
1335 * blkcg_activate_policy().
1337 void blkcg_deactivate_policy(struct request_queue *q,
1338 const struct blkcg_policy *pol)
1340 struct blkcg_gq *blkg;
1342 if (!blkcg_policy_enabled(q, pol))
1346 blk_mq_freeze_queue(q);
1348 spin_lock_irq(&q->queue_lock);
1350 __clear_bit(pol->plid, q->blkcg_pols);
1352 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1353 if (blkg->pd[pol->plid]) {
1354 if (pol->pd_offline_fn)
1355 pol->pd_offline_fn(blkg->pd[pol->plid]);
1356 pol->pd_free_fn(blkg->pd[pol->plid]);
1357 blkg->pd[pol->plid] = NULL;
1361 spin_unlock_irq(&q->queue_lock);
1364 blk_mq_unfreeze_queue(q);
1366 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1369 * blkcg_policy_register - register a blkcg policy
1370 * @pol: blkcg policy to register
1372 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1373 * successful registration. Returns 0 on success and -errno on failure.
1375 int blkcg_policy_register(struct blkcg_policy *pol)
1377 struct blkcg *blkcg;
1380 mutex_lock(&blkcg_pol_register_mutex);
1381 mutex_lock(&blkcg_pol_mutex);
1383 /* find an empty slot */
1385 for (i = 0; i < BLKCG_MAX_POLS; i++)
1386 if (!blkcg_policy[i])
1388 if (i >= BLKCG_MAX_POLS) {
1389 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1393 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1394 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1395 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1400 blkcg_policy[pol->plid] = pol;
1402 /* allocate and install cpd's */
1403 if (pol->cpd_alloc_fn) {
1404 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1405 struct blkcg_policy_data *cpd;
1407 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1411 blkcg->cpd[pol->plid] = cpd;
1413 cpd->plid = pol->plid;
1414 if (pol->cpd_init_fn)
1415 pol->cpd_init_fn(cpd);
1419 mutex_unlock(&blkcg_pol_mutex);
1421 /* everything is in place, add intf files for the new policy */
1422 if (pol->dfl_cftypes)
1423 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1425 if (pol->legacy_cftypes)
1426 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1427 pol->legacy_cftypes));
1428 mutex_unlock(&blkcg_pol_register_mutex);
1432 if (pol->cpd_free_fn) {
1433 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1434 if (blkcg->cpd[pol->plid]) {
1435 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1436 blkcg->cpd[pol->plid] = NULL;
1440 blkcg_policy[pol->plid] = NULL;
1442 mutex_unlock(&blkcg_pol_mutex);
1443 mutex_unlock(&blkcg_pol_register_mutex);
1446 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1449 * blkcg_policy_unregister - unregister a blkcg policy
1450 * @pol: blkcg policy to unregister
1452 * Undo blkcg_policy_register(@pol). Might sleep.
1454 void blkcg_policy_unregister(struct blkcg_policy *pol)
1456 struct blkcg *blkcg;
1458 mutex_lock(&blkcg_pol_register_mutex);
1460 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1463 /* kill the intf files first */
1464 if (pol->dfl_cftypes)
1465 cgroup_rm_cftypes(pol->dfl_cftypes);
1466 if (pol->legacy_cftypes)
1467 cgroup_rm_cftypes(pol->legacy_cftypes);
1469 /* remove cpds and unregister */
1470 mutex_lock(&blkcg_pol_mutex);
1472 if (pol->cpd_free_fn) {
1473 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1474 if (blkcg->cpd[pol->plid]) {
1475 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1476 blkcg->cpd[pol->plid] = NULL;
1480 blkcg_policy[pol->plid] = NULL;
1482 mutex_unlock(&blkcg_pol_mutex);
1484 mutex_unlock(&blkcg_pol_register_mutex);
1486 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1488 bool __blkcg_punt_bio_submit(struct bio *bio)
1490 struct blkcg_gq *blkg = bio->bi_blkg;
1492 /* consume the flag first */
1493 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1495 /* never bounce for the root cgroup */
1499 spin_lock_bh(&blkg->async_bio_lock);
1500 bio_list_add(&blkg->async_bios, bio);
1501 spin_unlock_bh(&blkg->async_bio_lock);
1503 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1508 * Scale the accumulated delay based on how long it has been since we updated
1509 * the delay. We only call this when we are adding delay, in case it's been a
1510 * while since we added delay, and when we are checking to see if we need to
1511 * delay a task, to account for any delays that may have occurred.
1513 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1515 u64 old = atomic64_read(&blkg->delay_start);
1518 * We only want to scale down every second. The idea here is that we
1519 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1520 * time window. We only want to throttle tasks for recent delay that
1521 * has occurred, in 1 second time windows since that's the maximum
1522 * things can be throttled. We save the current delay window in
1523 * blkg->last_delay so we know what amount is still left to be charged
1524 * to the blkg from this point onward. blkg->last_use keeps track of
1525 * the use_delay counter. The idea is if we're unthrottling the blkg we
1526 * are ok with whatever is happening now, and we can take away more of
1527 * the accumulated delay as we've already throttled enough that
1528 * everybody is happy with their IO latencies.
1530 if (time_before64(old + NSEC_PER_SEC, now) &&
1531 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1532 u64 cur = atomic64_read(&blkg->delay_nsec);
1533 u64 sub = min_t(u64, blkg->last_delay, now - old);
1534 int cur_use = atomic_read(&blkg->use_delay);
1537 * We've been unthrottled, subtract a larger chunk of our
1538 * accumulated delay.
1540 if (cur_use < blkg->last_use)
1541 sub = max_t(u64, sub, blkg->last_delay >> 1);
1544 * This shouldn't happen, but handle it anyway. Our delay_nsec
1545 * should only ever be growing except here where we subtract out
1546 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1547 * rather not end up with negative numbers.
1549 if (unlikely(cur < sub)) {
1550 atomic64_set(&blkg->delay_nsec, 0);
1551 blkg->last_delay = 0;
1553 atomic64_sub(sub, &blkg->delay_nsec);
1554 blkg->last_delay = cur - sub;
1556 blkg->last_use = cur_use;
1561 * This is called when we want to actually walk up the hierarchy and check to
1562 * see if we need to throttle, and then actually throttle if there is some
1563 * accumulated delay. This should only be called upon return to user space so
1564 * we're not holding some lock that would induce a priority inversion.
1566 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1568 unsigned long pflags;
1569 u64 now = ktime_to_ns(ktime_get());
1574 while (blkg->parent) {
1575 if (atomic_read(&blkg->use_delay)) {
1576 blkcg_scale_delay(blkg, now);
1577 delay_nsec = max_t(u64, delay_nsec,
1578 atomic64_read(&blkg->delay_nsec));
1580 blkg = blkg->parent;
1587 * Let's not sleep for all eternity if we've amassed a huge delay.
1588 * Swapping or metadata IO can accumulate 10's of seconds worth of
1589 * delay, and we want userspace to be able to do _something_ so cap the
1590 * delays at 1 second. If there's 10's of seconds worth of delay then
1591 * the tasks will be delayed for 1 second for every syscall.
1593 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1596 psi_memstall_enter(&pflags);
1598 exp = ktime_add_ns(now, delay_nsec);
1599 tok = io_schedule_prepare();
1601 __set_current_state(TASK_KILLABLE);
1602 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1604 } while (!fatal_signal_pending(current));
1605 io_schedule_finish(tok);
1608 psi_memstall_leave(&pflags);
1612 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1614 * This is only called if we've been marked with set_notify_resume(). Obviously
1615 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1616 * check to see if current->throttle_queue is set and if not this doesn't do
1617 * anything. This should only ever be called by the resume code, it's not meant
1618 * to be called by people willy-nilly as it will actually do the work to
1619 * throttle the task if it is setup for throttling.
1621 void blkcg_maybe_throttle_current(void)
1623 struct request_queue *q = current->throttle_queue;
1624 struct cgroup_subsys_state *css;
1625 struct blkcg *blkcg;
1626 struct blkcg_gq *blkg;
1627 bool use_memdelay = current->use_memdelay;
1632 current->throttle_queue = NULL;
1633 current->use_memdelay = false;
1636 css = kthread_blkcg();
1638 blkcg = css_to_blkcg(css);
1640 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1644 blkg = blkg_lookup(blkcg, q);
1647 if (!blkg_tryget(blkg))
1651 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1661 * blkcg_schedule_throttle - this task needs to check for throttling
1662 * @q: the request queue IO was submitted on
1663 * @use_memdelay: do we charge this to memory delay for PSI
1665 * This is called by the IO controller when we know there's delay accumulated
1666 * for the blkg for this task. We do not pass the blkg because there are places
1667 * we call this that may not have that information, the swapping code for
1668 * instance will only have a request_queue at that point. This set's the
1669 * notify_resume for the task to check and see if it requires throttling before
1670 * returning to user space.
1672 * We will only schedule once per syscall. You can call this over and over
1673 * again and it will only do the check once upon return to user space, and only
1674 * throttle once. If the task needs to be throttled again it'll need to be
1675 * re-set at the next time we see the task.
1677 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1679 if (unlikely(current->flags & PF_KTHREAD))
1682 if (!blk_get_queue(q))
1685 if (current->throttle_queue)
1686 blk_put_queue(current->throttle_queue);
1687 current->throttle_queue = q;
1689 current->use_memdelay = use_memdelay;
1690 set_notify_resume(current);
1694 * blkcg_add_delay - add delay to this blkg
1695 * @blkg: blkg of interest
1696 * @now: the current time in nanoseconds
1697 * @delta: how many nanoseconds of delay to add
1699 * Charge @delta to the blkg's current delay accumulation. This is used to
1700 * throttle tasks if an IO controller thinks we need more throttling.
1702 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1704 blkcg_scale_delay(blkg, now);
1705 atomic64_add(delta, &blkg->delay_nsec);
1708 static int __init blkcg_init(void)
1710 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1711 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1712 WQ_UNBOUND | WQ_SYSFS, 0);
1713 if (!blkcg_punt_bio_wq)
1717 subsys_initcall(blkcg_init);
1719 module_param(blkcg_debug_stats, bool, 0644);
1720 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");