2 * Common Block IO controller cgroup interface
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
13 * For policy-specific per-blkcg data:
14 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
15 * Arianna Avanzini <avanzini.arianna@gmail.com>
17 #include <linux/ioprio.h>
18 #include <linux/kdev_t.h>
19 #include <linux/module.h>
20 #include <linux/sched/signal.h>
21 #include <linux/err.h>
22 #include <linux/blkdev.h>
23 #include <linux/backing-dev.h>
24 #include <linux/slab.h>
25 #include <linux/genhd.h>
26 #include <linux/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/blk-cgroup.h>
30 #include <linux/tracehook.h>
33 #define MAX_KEY_LEN 100
36 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
37 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
38 * policy [un]register operations including cgroup file additions /
39 * removals. Putting cgroup file registration outside blkcg_pol_mutex
40 * allows grabbing it from cgroup callbacks.
42 static DEFINE_MUTEX(blkcg_pol_register_mutex);
43 static DEFINE_MUTEX(blkcg_pol_mutex);
45 struct blkcg blkcg_root;
46 EXPORT_SYMBOL_GPL(blkcg_root);
48 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
50 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
52 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
54 static bool blkcg_debug_stats = false;
56 static bool blkcg_policy_enabled(struct request_queue *q,
57 const struct blkcg_policy *pol)
59 return pol && test_bit(pol->plid, q->blkcg_pols);
63 * blkg_free - free a blkg
66 * Free @blkg which may be partially allocated.
68 static void blkg_free(struct blkcg_gq *blkg)
75 for (i = 0; i < BLKCG_MAX_POLS; i++)
77 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
79 if (blkg->blkcg != &blkcg_root)
80 blk_exit_rl(blkg->q, &blkg->rl);
82 blkg_rwstat_exit(&blkg->stat_ios);
83 blkg_rwstat_exit(&blkg->stat_bytes);
88 * blkg_alloc - allocate a blkg
89 * @blkcg: block cgroup the new blkg is associated with
90 * @q: request_queue the new blkg is associated with
91 * @gfp_mask: allocation mask to use
93 * Allocate a new blkg assocating @blkcg and @q.
95 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
98 struct blkcg_gq *blkg;
101 /* alloc and init base part */
102 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
106 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
107 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
111 INIT_LIST_HEAD(&blkg->q_node);
113 atomic_set(&blkg->refcnt, 1);
115 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
116 if (blkcg != &blkcg_root) {
117 if (blk_init_rl(&blkg->rl, q, gfp_mask))
119 blkg->rl.blkg = blkg;
122 for (i = 0; i < BLKCG_MAX_POLS; i++) {
123 struct blkcg_policy *pol = blkcg_policy[i];
124 struct blkg_policy_data *pd;
126 if (!blkcg_policy_enabled(q, pol))
129 /* alloc per-policy data and attach it to blkg */
130 pd = pol->pd_alloc_fn(gfp_mask, q->node);
146 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
147 struct request_queue *q, bool update_hint)
149 struct blkcg_gq *blkg;
152 * Hint didn't match. Look up from the radix tree. Note that the
153 * hint can only be updated under queue_lock as otherwise @blkg
154 * could have already been removed from blkg_tree. The caller is
155 * responsible for grabbing queue_lock if @update_hint.
157 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
158 if (blkg && blkg->q == q) {
160 lockdep_assert_held(q->queue_lock);
161 rcu_assign_pointer(blkcg->blkg_hint, blkg);
168 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
171 * If @new_blkg is %NULL, this function tries to allocate a new one as
172 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
174 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
175 struct request_queue *q,
176 struct blkcg_gq *new_blkg)
178 struct blkcg_gq *blkg;
179 struct bdi_writeback_congested *wb_congested;
182 WARN_ON_ONCE(!rcu_read_lock_held());
183 lockdep_assert_held(q->queue_lock);
185 /* blkg holds a reference to blkcg */
186 if (!css_tryget_online(&blkcg->css)) {
191 wb_congested = wb_congested_get_create(q->backing_dev_info,
193 GFP_NOWAIT | __GFP_NOWARN);
201 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
202 if (unlikely(!new_blkg)) {
204 goto err_put_congested;
208 blkg->wb_congested = wb_congested;
211 if (blkcg_parent(blkcg)) {
212 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
213 if (WARN_ON_ONCE(!blkg->parent)) {
215 goto err_put_congested;
217 blkg_get(blkg->parent);
220 /* invoke per-policy init */
221 for (i = 0; i < BLKCG_MAX_POLS; i++) {
222 struct blkcg_policy *pol = blkcg_policy[i];
224 if (blkg->pd[i] && pol->pd_init_fn)
225 pol->pd_init_fn(blkg->pd[i]);
229 spin_lock(&blkcg->lock);
230 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
232 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
233 list_add(&blkg->q_node, &q->blkg_list);
235 for (i = 0; i < BLKCG_MAX_POLS; i++) {
236 struct blkcg_policy *pol = blkcg_policy[i];
238 if (blkg->pd[i] && pol->pd_online_fn)
239 pol->pd_online_fn(blkg->pd[i]);
243 spin_unlock(&blkcg->lock);
248 /* @blkg failed fully initialized, use the usual release path */
253 wb_congested_put(wb_congested);
255 css_put(&blkcg->css);
262 * blkg_lookup_create - lookup blkg, try to create one if not there
263 * @blkcg: blkcg of interest
264 * @q: request_queue of interest
266 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
267 * create one. blkg creation is performed recursively from blkcg_root such
268 * that all non-root blkg's have access to the parent blkg. This function
269 * should be called under RCU read lock and @q->queue_lock.
271 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
272 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
273 * dead and bypassing, returns ERR_PTR(-EBUSY).
275 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
276 struct request_queue *q)
278 struct blkcg_gq *blkg;
280 WARN_ON_ONCE(!rcu_read_lock_held());
281 lockdep_assert_held(q->queue_lock);
284 * This could be the first entry point of blkcg implementation and
285 * we shouldn't allow anything to go through for a bypassing queue.
287 if (unlikely(blk_queue_bypass(q)))
288 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
290 blkg = __blkg_lookup(blkcg, q, true);
295 * Create blkgs walking down from blkcg_root to @blkcg, so that all
296 * non-root blkgs have access to their parents.
299 struct blkcg *pos = blkcg;
300 struct blkcg *parent = blkcg_parent(blkcg);
302 while (parent && !__blkg_lookup(parent, q, false)) {
304 parent = blkcg_parent(parent);
307 blkg = blkg_create(pos, q, NULL);
308 if (pos == blkcg || IS_ERR(blkg))
313 static void blkg_destroy(struct blkcg_gq *blkg)
315 struct blkcg *blkcg = blkg->blkcg;
316 struct blkcg_gq *parent = blkg->parent;
319 lockdep_assert_held(blkg->q->queue_lock);
320 lockdep_assert_held(&blkcg->lock);
322 /* Something wrong if we are trying to remove same group twice */
323 WARN_ON_ONCE(list_empty(&blkg->q_node));
324 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
326 for (i = 0; i < BLKCG_MAX_POLS; i++) {
327 struct blkcg_policy *pol = blkcg_policy[i];
329 if (blkg->pd[i] && pol->pd_offline_fn)
330 pol->pd_offline_fn(blkg->pd[i]);
334 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
335 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
338 blkg->online = false;
340 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
341 list_del_init(&blkg->q_node);
342 hlist_del_init_rcu(&blkg->blkcg_node);
345 * Both setting lookup hint to and clearing it from @blkg are done
346 * under queue_lock. If it's not pointing to @blkg now, it never
347 * will. Hint assignment itself can race safely.
349 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
350 rcu_assign_pointer(blkcg->blkg_hint, NULL);
353 * Put the reference taken at the time of creation so that when all
354 * queues are gone, group can be destroyed.
360 * blkg_destroy_all - destroy all blkgs associated with a request_queue
361 * @q: request_queue of interest
363 * Destroy all blkgs associated with @q.
365 static void blkg_destroy_all(struct request_queue *q)
367 struct blkcg_gq *blkg, *n;
369 lockdep_assert_held(q->queue_lock);
371 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
372 struct blkcg *blkcg = blkg->blkcg;
374 spin_lock(&blkcg->lock);
376 spin_unlock(&blkcg->lock);
380 q->root_rl.blkg = NULL;
384 * A group is RCU protected, but having an rcu lock does not mean that one
385 * can access all the fields of blkg and assume these are valid. For
386 * example, don't try to follow throtl_data and request queue links.
388 * Having a reference to blkg under an rcu allows accesses to only values
389 * local to groups like group stats and group rate limits.
391 void __blkg_release_rcu(struct rcu_head *rcu_head)
393 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
395 /* release the blkcg and parent blkg refs this blkg has been holding */
396 css_put(&blkg->blkcg->css);
398 blkg_put(blkg->parent);
400 wb_congested_put(blkg->wb_congested);
404 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
407 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
408 * because the root blkg uses @q->root_rl instead of its own rl.
410 struct request_list *__blk_queue_next_rl(struct request_list *rl,
411 struct request_queue *q)
413 struct list_head *ent;
414 struct blkcg_gq *blkg;
417 * Determine the current blkg list_head. The first entry is
418 * root_rl which is off @q->blkg_list and mapped to the head.
420 if (rl == &q->root_rl) {
422 /* There are no more block groups, hence no request lists */
426 blkg = container_of(rl, struct blkcg_gq, rl);
430 /* walk to the next list_head, skip root blkcg */
432 if (ent == &q->root_blkg->q_node)
434 if (ent == &q->blkg_list)
437 blkg = container_of(ent, struct blkcg_gq, q_node);
441 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
442 struct cftype *cftype, u64 val)
444 struct blkcg *blkcg = css_to_blkcg(css);
445 struct blkcg_gq *blkg;
448 mutex_lock(&blkcg_pol_mutex);
449 spin_lock_irq(&blkcg->lock);
452 * Note that stat reset is racy - it doesn't synchronize against
453 * stat updates. This is a debug feature which shouldn't exist
454 * anyway. If you get hit by a race, retry.
456 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
457 blkg_rwstat_reset(&blkg->stat_bytes);
458 blkg_rwstat_reset(&blkg->stat_ios);
460 for (i = 0; i < BLKCG_MAX_POLS; i++) {
461 struct blkcg_policy *pol = blkcg_policy[i];
463 if (blkg->pd[i] && pol->pd_reset_stats_fn)
464 pol->pd_reset_stats_fn(blkg->pd[i]);
468 spin_unlock_irq(&blkcg->lock);
469 mutex_unlock(&blkcg_pol_mutex);
473 const char *blkg_dev_name(struct blkcg_gq *blkg)
475 /* some drivers (floppy) instantiate a queue w/o disk registered */
476 if (blkg->q->backing_dev_info->dev)
477 return dev_name(blkg->q->backing_dev_info->dev);
480 EXPORT_SYMBOL_GPL(blkg_dev_name);
483 * blkcg_print_blkgs - helper for printing per-blkg data
484 * @sf: seq_file to print to
485 * @blkcg: blkcg of interest
486 * @prfill: fill function to print out a blkg
487 * @pol: policy in question
488 * @data: data to be passed to @prfill
489 * @show_total: to print out sum of prfill return values or not
491 * This function invokes @prfill on each blkg of @blkcg if pd for the
492 * policy specified by @pol exists. @prfill is invoked with @sf, the
493 * policy data and @data and the matching queue lock held. If @show_total
494 * is %true, the sum of the return values from @prfill is printed with
495 * "Total" label at the end.
497 * This is to be used to construct print functions for
498 * cftype->read_seq_string method.
500 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
501 u64 (*prfill)(struct seq_file *,
502 struct blkg_policy_data *, int),
503 const struct blkcg_policy *pol, int data,
506 struct blkcg_gq *blkg;
510 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
511 spin_lock_irq(blkg->q->queue_lock);
512 if (blkcg_policy_enabled(blkg->q, pol))
513 total += prfill(sf, blkg->pd[pol->plid], data);
514 spin_unlock_irq(blkg->q->queue_lock);
519 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
521 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
524 * __blkg_prfill_u64 - prfill helper for a single u64 value
525 * @sf: seq_file to print to
526 * @pd: policy private data of interest
529 * Print @v to @sf for the device assocaited with @pd.
531 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
533 const char *dname = blkg_dev_name(pd->blkg);
538 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
541 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
544 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
545 * @sf: seq_file to print to
546 * @pd: policy private data of interest
547 * @rwstat: rwstat to print
549 * Print @rwstat to @sf for the device assocaited with @pd.
551 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
552 const struct blkg_rwstat *rwstat)
554 static const char *rwstr[] = {
555 [BLKG_RWSTAT_READ] = "Read",
556 [BLKG_RWSTAT_WRITE] = "Write",
557 [BLKG_RWSTAT_SYNC] = "Sync",
558 [BLKG_RWSTAT_ASYNC] = "Async",
559 [BLKG_RWSTAT_DISCARD] = "Discard",
561 const char *dname = blkg_dev_name(pd->blkg);
568 for (i = 0; i < BLKG_RWSTAT_NR; i++)
569 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
570 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
572 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
573 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
574 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
575 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
578 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
581 * blkg_prfill_stat - prfill callback for blkg_stat
582 * @sf: seq_file to print to
583 * @pd: policy private data of interest
584 * @off: offset to the blkg_stat in @pd
586 * prfill callback for printing a blkg_stat.
588 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
590 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
592 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
595 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
596 * @sf: seq_file to print to
597 * @pd: policy private data of interest
598 * @off: offset to the blkg_rwstat in @pd
600 * prfill callback for printing a blkg_rwstat.
602 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
605 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
607 return __blkg_prfill_rwstat(sf, pd, &rwstat);
609 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
611 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
612 struct blkg_policy_data *pd, int off)
614 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
616 return __blkg_prfill_rwstat(sf, pd, &rwstat);
620 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
621 * @sf: seq_file to print to
624 * To be used as cftype->seq_show to print blkg->stat_bytes.
625 * cftype->private must be set to the blkcg_policy.
627 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
629 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
630 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
631 offsetof(struct blkcg_gq, stat_bytes), true);
634 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
637 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
638 * @sf: seq_file to print to
641 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
642 * must be set to the blkcg_policy.
644 int blkg_print_stat_ios(struct seq_file *sf, void *v)
646 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
647 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
648 offsetof(struct blkcg_gq, stat_ios), true);
651 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
653 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
654 struct blkg_policy_data *pd,
657 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
659 return __blkg_prfill_rwstat(sf, pd, &rwstat);
663 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
664 * @sf: seq_file to print to
667 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
669 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
670 blkg_prfill_rwstat_field_recursive,
671 (void *)seq_cft(sf)->private,
672 offsetof(struct blkcg_gq, stat_bytes), true);
675 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
678 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
679 * @sf: seq_file to print to
682 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
684 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
685 blkg_prfill_rwstat_field_recursive,
686 (void *)seq_cft(sf)->private,
687 offsetof(struct blkcg_gq, stat_ios), true);
690 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
693 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
694 * @blkg: blkg of interest
695 * @pol: blkcg_policy which contains the blkg_stat
696 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
698 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
699 * online descendants and their aux counts. The caller must be holding the
700 * queue lock for online tests.
702 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
703 * at @off bytes into @blkg's blkg_policy_data of the policy.
705 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
706 struct blkcg_policy *pol, int off)
708 struct blkcg_gq *pos_blkg;
709 struct cgroup_subsys_state *pos_css;
712 lockdep_assert_held(blkg->q->queue_lock);
715 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
716 struct blkg_stat *stat;
718 if (!pos_blkg->online)
722 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
724 stat = (void *)blkg + off;
726 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
732 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
735 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
736 * @blkg: blkg of interest
737 * @pol: blkcg_policy which contains the blkg_rwstat
738 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
740 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
741 * online descendants and their aux counts. The caller must be holding the
742 * queue lock for online tests.
744 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
745 * is at @off bytes into @blkg's blkg_policy_data of the policy.
747 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
748 struct blkcg_policy *pol, int off)
750 struct blkcg_gq *pos_blkg;
751 struct cgroup_subsys_state *pos_css;
752 struct blkg_rwstat sum = { };
755 lockdep_assert_held(blkg->q->queue_lock);
758 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
759 struct blkg_rwstat *rwstat;
761 if (!pos_blkg->online)
765 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
767 rwstat = (void *)pos_blkg + off;
769 for (i = 0; i < BLKG_RWSTAT_NR; i++)
770 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
771 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
778 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
780 /* Performs queue bypass and policy enabled checks then looks up blkg. */
781 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
782 const struct blkcg_policy *pol,
783 struct request_queue *q)
785 WARN_ON_ONCE(!rcu_read_lock_held());
786 lockdep_assert_held(q->queue_lock);
788 if (!blkcg_policy_enabled(q, pol))
789 return ERR_PTR(-EOPNOTSUPP);
792 * This could be the first entry point of blkcg implementation and
793 * we shouldn't allow anything to go through for a bypassing queue.
795 if (unlikely(blk_queue_bypass(q)))
796 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
798 return __blkg_lookup(blkcg, q, true /* update_hint */);
802 * blkg_conf_prep - parse and prepare for per-blkg config update
803 * @blkcg: target block cgroup
804 * @pol: target policy
805 * @input: input string
806 * @ctx: blkg_conf_ctx to be filled
808 * Parse per-blkg config update from @input and initialize @ctx with the
809 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
810 * part of @input following MAJ:MIN. This function returns with RCU read
811 * lock and queue lock held and must be paired with blkg_conf_finish().
813 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
814 char *input, struct blkg_conf_ctx *ctx)
815 __acquires(rcu) __acquires(disk->queue->queue_lock)
817 struct gendisk *disk;
818 struct request_queue *q;
819 struct blkcg_gq *blkg;
820 unsigned int major, minor;
821 int key_len, part, ret;
824 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
827 body = input + key_len;
830 body = skip_spaces(body);
832 disk = get_gendisk(MKDEV(major, minor), &part);
843 spin_lock_irq(q->queue_lock);
845 blkg = blkg_lookup_check(blkcg, pol, q);
855 * Create blkgs walking down from blkcg_root to @blkcg, so that all
856 * non-root blkgs have access to their parents.
859 struct blkcg *pos = blkcg;
860 struct blkcg *parent;
861 struct blkcg_gq *new_blkg;
863 parent = blkcg_parent(blkcg);
864 while (parent && !__blkg_lookup(parent, q, false)) {
866 parent = blkcg_parent(parent);
869 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
870 spin_unlock_irq(q->queue_lock);
873 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
874 if (unlikely(!new_blkg)) {
880 spin_lock_irq(q->queue_lock);
882 blkg = blkg_lookup_check(pos, pol, q);
891 blkg = blkg_create(pos, q, new_blkg);
892 if (unlikely(IS_ERR(blkg))) {
908 spin_unlock_irq(q->queue_lock);
911 put_disk_and_module(disk);
913 * If queue was bypassing, we should retry. Do so after a
914 * short msleep(). It isn't strictly necessary but queue
915 * can be bypassing for some time and it's always nice to
916 * avoid busy looping.
920 ret = restart_syscall();
924 EXPORT_SYMBOL_GPL(blkg_conf_prep);
927 * blkg_conf_finish - finish up per-blkg config update
928 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
930 * Finish up after per-blkg config update. This function must be paired
931 * with blkg_conf_prep().
933 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
934 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
936 spin_unlock_irq(ctx->disk->queue->queue_lock);
938 put_disk_and_module(ctx->disk);
940 EXPORT_SYMBOL_GPL(blkg_conf_finish);
942 static int blkcg_print_stat(struct seq_file *sf, void *v)
944 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
945 struct blkcg_gq *blkg;
949 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
952 struct blkg_rwstat rwstat;
953 u64 rbytes, wbytes, rios, wios, dbytes, dios;
954 size_t size = seq_get_buf(sf, &buf), off = 0;
956 bool has_stats = false;
958 spin_lock_irq(blkg->q->queue_lock);
963 dname = blkg_dev_name(blkg);
968 * Hooray string manipulation, count is the size written NOT
969 * INCLUDING THE \0, so size is now count+1 less than what we
970 * had before, but we want to start writing the next bit from
971 * the \0 so we only add count to buf.
973 off += scnprintf(buf+off, size-off, "%s ", dname);
975 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
976 offsetof(struct blkcg_gq, stat_bytes));
977 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
978 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
979 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
981 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
982 offsetof(struct blkcg_gq, stat_ios));
983 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
984 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
985 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
987 if (rbytes || wbytes || rios || wios) {
989 off += scnprintf(buf+off, size-off,
990 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
991 rbytes, wbytes, rios, wios,
995 if (!blkcg_debug_stats)
998 if (atomic_read(&blkg->use_delay)) {
1000 off += scnprintf(buf+off, size-off,
1001 " use_delay=%d delay_nsec=%llu",
1002 atomic_read(&blkg->use_delay),
1003 (unsigned long long)atomic64_read(&blkg->delay_nsec));
1006 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1007 struct blkcg_policy *pol = blkcg_policy[i];
1010 if (!blkg->pd[i] || !pol->pd_stat_fn)
1013 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
1020 if (off < size - 1) {
1021 off += scnprintf(buf+off, size-off, "\n");
1022 seq_commit(sf, off);
1028 spin_unlock_irq(blkg->q->queue_lock);
1035 static struct cftype blkcg_files[] = {
1038 .flags = CFTYPE_NOT_ON_ROOT,
1039 .seq_show = blkcg_print_stat,
1044 static struct cftype blkcg_legacy_files[] = {
1046 .name = "reset_stats",
1047 .write_u64 = blkcg_reset_stats,
1053 * blkcg destruction is a three-stage process.
1055 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1056 * which offlines writeback. Here we tie the next stage of blkg destruction
1057 * to the completion of writeback associated with the blkcg. This lets us
1058 * avoid punting potentially large amounts of outstanding writeback to root
1059 * while maintaining any ongoing policies. The next stage is triggered when
1060 * the nr_cgwbs count goes to zero.
1062 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1063 * and handles the destruction of blkgs. Here the css reference held by
1064 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1065 * This work may occur in cgwb_release_workfn() on the cgwb_release
1066 * workqueue. Any submitted ios that fail to get the blkg ref will be
1067 * punted to the root_blkg.
1069 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1070 * This finally frees the blkcg.
1074 * blkcg_css_offline - cgroup css_offline callback
1075 * @css: css of interest
1077 * This function is called when @css is about to go away. Here the cgwbs are
1078 * offlined first and only once writeback associated with the blkcg has
1079 * finished do we start step 2 (see above).
1081 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1083 struct blkcg *blkcg = css_to_blkcg(css);
1085 /* this prevents anyone from attaching or migrating to this blkcg */
1086 wb_blkcg_offline(blkcg);
1088 /* put the base cgwb reference allowing step 2 to be triggered */
1089 blkcg_cgwb_put(blkcg);
1093 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1094 * @blkcg: blkcg of interest
1096 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1097 * is nested inside q lock, this function performs reverse double lock dancing.
1098 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1099 * blkcg_css_free to eventually be called.
1101 * This is the blkcg counterpart of ioc_release_fn().
1103 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1105 spin_lock_irq(&blkcg->lock);
1107 while (!hlist_empty(&blkcg->blkg_list)) {
1108 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1109 struct blkcg_gq, blkcg_node);
1110 struct request_queue *q = blkg->q;
1112 if (spin_trylock(q->queue_lock)) {
1114 spin_unlock(q->queue_lock);
1116 spin_unlock_irq(&blkcg->lock);
1118 spin_lock_irq(&blkcg->lock);
1122 spin_unlock_irq(&blkcg->lock);
1125 static void blkcg_css_free(struct cgroup_subsys_state *css)
1127 struct blkcg *blkcg = css_to_blkcg(css);
1130 mutex_lock(&blkcg_pol_mutex);
1132 list_del(&blkcg->all_blkcgs_node);
1134 for (i = 0; i < BLKCG_MAX_POLS; i++)
1136 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1138 mutex_unlock(&blkcg_pol_mutex);
1143 static struct cgroup_subsys_state *
1144 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1146 struct blkcg *blkcg;
1147 struct cgroup_subsys_state *ret;
1150 mutex_lock(&blkcg_pol_mutex);
1153 blkcg = &blkcg_root;
1155 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1157 ret = ERR_PTR(-ENOMEM);
1162 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1163 struct blkcg_policy *pol = blkcg_policy[i];
1164 struct blkcg_policy_data *cpd;
1167 * If the policy hasn't been attached yet, wait for it
1168 * to be attached before doing anything else. Otherwise,
1169 * check if the policy requires any specific per-cgroup
1170 * data: if it does, allocate and initialize it.
1172 if (!pol || !pol->cpd_alloc_fn)
1175 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1177 ret = ERR_PTR(-ENOMEM);
1180 blkcg->cpd[i] = cpd;
1183 if (pol->cpd_init_fn)
1184 pol->cpd_init_fn(cpd);
1187 spin_lock_init(&blkcg->lock);
1188 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1189 INIT_HLIST_HEAD(&blkcg->blkg_list);
1190 #ifdef CONFIG_CGROUP_WRITEBACK
1191 INIT_LIST_HEAD(&blkcg->cgwb_list);
1192 refcount_set(&blkcg->cgwb_refcnt, 1);
1194 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1196 mutex_unlock(&blkcg_pol_mutex);
1200 for (i--; i >= 0; i--)
1202 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1204 if (blkcg != &blkcg_root)
1207 mutex_unlock(&blkcg_pol_mutex);
1212 * blkcg_init_queue - initialize blkcg part of request queue
1213 * @q: request_queue to initialize
1215 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1216 * part of new request_queue @q.
1219 * 0 on success, -errno on failure.
1221 int blkcg_init_queue(struct request_queue *q)
1223 struct blkcg_gq *new_blkg, *blkg;
1227 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1231 preloaded = !radix_tree_preload(GFP_KERNEL);
1233 /* Make sure the root blkg exists. */
1235 spin_lock_irq(q->queue_lock);
1236 blkg = blkg_create(&blkcg_root, q, new_blkg);
1239 q->root_blkg = blkg;
1240 q->root_rl.blkg = blkg;
1241 spin_unlock_irq(q->queue_lock);
1245 radix_tree_preload_end();
1247 ret = blk_iolatency_init(q);
1249 spin_lock_irq(q->queue_lock);
1250 blkg_destroy_all(q);
1251 spin_unlock_irq(q->queue_lock);
1255 ret = blk_throtl_init(q);
1257 spin_lock_irq(q->queue_lock);
1258 blkg_destroy_all(q);
1259 spin_unlock_irq(q->queue_lock);
1264 spin_unlock_irq(q->queue_lock);
1267 radix_tree_preload_end();
1268 return PTR_ERR(blkg);
1272 * blkcg_drain_queue - drain blkcg part of request_queue
1273 * @q: request_queue to drain
1275 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1277 void blkcg_drain_queue(struct request_queue *q)
1279 lockdep_assert_held(q->queue_lock);
1282 * @q could be exiting and already have destroyed all blkgs as
1283 * indicated by NULL root_blkg. If so, don't confuse policies.
1288 blk_throtl_drain(q);
1292 * blkcg_exit_queue - exit and release blkcg part of request_queue
1293 * @q: request_queue being released
1295 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1297 void blkcg_exit_queue(struct request_queue *q)
1299 spin_lock_irq(q->queue_lock);
1300 blkg_destroy_all(q);
1301 spin_unlock_irq(q->queue_lock);
1307 * We cannot support shared io contexts, as we have no mean to support
1308 * two tasks with the same ioc in two different groups without major rework
1309 * of the main cic data structures. For now we allow a task to change
1310 * its cgroup only if it's the only owner of its ioc.
1312 static int blkcg_can_attach(struct cgroup_taskset *tset)
1314 struct task_struct *task;
1315 struct cgroup_subsys_state *dst_css;
1316 struct io_context *ioc;
1319 /* task_lock() is needed to avoid races with exit_io_context() */
1320 cgroup_taskset_for_each(task, dst_css, tset) {
1322 ioc = task->io_context;
1323 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1332 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1336 mutex_lock(&blkcg_pol_mutex);
1338 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1339 struct blkcg_policy *pol = blkcg_policy[i];
1340 struct blkcg *blkcg;
1342 if (!pol || !pol->cpd_bind_fn)
1345 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1346 if (blkcg->cpd[pol->plid])
1347 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1349 mutex_unlock(&blkcg_pol_mutex);
1352 static void blkcg_exit(struct task_struct *tsk)
1354 if (tsk->throttle_queue)
1355 blk_put_queue(tsk->throttle_queue);
1356 tsk->throttle_queue = NULL;
1359 struct cgroup_subsys io_cgrp_subsys = {
1360 .css_alloc = blkcg_css_alloc,
1361 .css_offline = blkcg_css_offline,
1362 .css_free = blkcg_css_free,
1363 .can_attach = blkcg_can_attach,
1365 .dfl_cftypes = blkcg_files,
1366 .legacy_cftypes = blkcg_legacy_files,
1367 .legacy_name = "blkio",
1371 * This ensures that, if available, memcg is automatically enabled
1372 * together on the default hierarchy so that the owner cgroup can
1373 * be retrieved from writeback pages.
1375 .depends_on = 1 << memory_cgrp_id,
1378 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1381 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1382 * @q: request_queue of interest
1383 * @pol: blkcg policy to activate
1385 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1386 * bypass mode to populate its blkgs with policy_data for @pol.
1388 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1389 * from IO path. Update of each blkg is protected by both queue and blkcg
1390 * locks so that holding either lock and testing blkcg_policy_enabled() is
1391 * always enough for dereferencing policy data.
1393 * The caller is responsible for synchronizing [de]activations and policy
1394 * [un]registerations. Returns 0 on success, -errno on failure.
1396 int blkcg_activate_policy(struct request_queue *q,
1397 const struct blkcg_policy *pol)
1399 struct blkg_policy_data *pd_prealloc = NULL;
1400 struct blkcg_gq *blkg;
1403 if (blkcg_policy_enabled(q, pol))
1407 blk_mq_freeze_queue(q);
1409 blk_queue_bypass_start(q);
1412 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1415 goto out_bypass_end;
1419 spin_lock_irq(q->queue_lock);
1421 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1422 struct blkg_policy_data *pd;
1424 if (blkg->pd[pol->plid])
1427 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1429 swap(pd, pd_prealloc);
1431 spin_unlock_irq(q->queue_lock);
1435 blkg->pd[pol->plid] = pd;
1437 pd->plid = pol->plid;
1438 if (pol->pd_init_fn)
1439 pol->pd_init_fn(pd);
1442 __set_bit(pol->plid, q->blkcg_pols);
1445 spin_unlock_irq(q->queue_lock);
1448 blk_mq_unfreeze_queue(q);
1450 blk_queue_bypass_end(q);
1452 pol->pd_free_fn(pd_prealloc);
1455 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1458 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1459 * @q: request_queue of interest
1460 * @pol: blkcg policy to deactivate
1462 * Deactivate @pol on @q. Follows the same synchronization rules as
1463 * blkcg_activate_policy().
1465 void blkcg_deactivate_policy(struct request_queue *q,
1466 const struct blkcg_policy *pol)
1468 struct blkcg_gq *blkg;
1470 if (!blkcg_policy_enabled(q, pol))
1474 blk_mq_freeze_queue(q);
1476 blk_queue_bypass_start(q);
1478 spin_lock_irq(q->queue_lock);
1480 __clear_bit(pol->plid, q->blkcg_pols);
1482 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1483 if (blkg->pd[pol->plid]) {
1484 if (pol->pd_offline_fn)
1485 pol->pd_offline_fn(blkg->pd[pol->plid]);
1486 pol->pd_free_fn(blkg->pd[pol->plid]);
1487 blkg->pd[pol->plid] = NULL;
1491 spin_unlock_irq(q->queue_lock);
1494 blk_mq_unfreeze_queue(q);
1496 blk_queue_bypass_end(q);
1498 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1501 * blkcg_policy_register - register a blkcg policy
1502 * @pol: blkcg policy to register
1504 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1505 * successful registration. Returns 0 on success and -errno on failure.
1507 int blkcg_policy_register(struct blkcg_policy *pol)
1509 struct blkcg *blkcg;
1512 mutex_lock(&blkcg_pol_register_mutex);
1513 mutex_lock(&blkcg_pol_mutex);
1515 /* find an empty slot */
1517 for (i = 0; i < BLKCG_MAX_POLS; i++)
1518 if (!blkcg_policy[i])
1520 if (i >= BLKCG_MAX_POLS) {
1521 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1525 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1526 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1527 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1532 blkcg_policy[pol->plid] = pol;
1534 /* allocate and install cpd's */
1535 if (pol->cpd_alloc_fn) {
1536 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1537 struct blkcg_policy_data *cpd;
1539 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1543 blkcg->cpd[pol->plid] = cpd;
1545 cpd->plid = pol->plid;
1546 pol->cpd_init_fn(cpd);
1550 mutex_unlock(&blkcg_pol_mutex);
1552 /* everything is in place, add intf files for the new policy */
1553 if (pol->dfl_cftypes)
1554 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1556 if (pol->legacy_cftypes)
1557 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1558 pol->legacy_cftypes));
1559 mutex_unlock(&blkcg_pol_register_mutex);
1563 if (pol->cpd_free_fn) {
1564 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1565 if (blkcg->cpd[pol->plid]) {
1566 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1567 blkcg->cpd[pol->plid] = NULL;
1571 blkcg_policy[pol->plid] = NULL;
1573 mutex_unlock(&blkcg_pol_mutex);
1574 mutex_unlock(&blkcg_pol_register_mutex);
1577 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1580 * blkcg_policy_unregister - unregister a blkcg policy
1581 * @pol: blkcg policy to unregister
1583 * Undo blkcg_policy_register(@pol). Might sleep.
1585 void blkcg_policy_unregister(struct blkcg_policy *pol)
1587 struct blkcg *blkcg;
1589 mutex_lock(&blkcg_pol_register_mutex);
1591 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1594 /* kill the intf files first */
1595 if (pol->dfl_cftypes)
1596 cgroup_rm_cftypes(pol->dfl_cftypes);
1597 if (pol->legacy_cftypes)
1598 cgroup_rm_cftypes(pol->legacy_cftypes);
1600 /* remove cpds and unregister */
1601 mutex_lock(&blkcg_pol_mutex);
1603 if (pol->cpd_free_fn) {
1604 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1605 if (blkcg->cpd[pol->plid]) {
1606 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1607 blkcg->cpd[pol->plid] = NULL;
1611 blkcg_policy[pol->plid] = NULL;
1613 mutex_unlock(&blkcg_pol_mutex);
1615 mutex_unlock(&blkcg_pol_register_mutex);
1617 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1620 * Scale the accumulated delay based on how long it has been since we updated
1621 * the delay. We only call this when we are adding delay, in case it's been a
1622 * while since we added delay, and when we are checking to see if we need to
1623 * delay a task, to account for any delays that may have occurred.
1625 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1627 u64 old = atomic64_read(&blkg->delay_start);
1630 * We only want to scale down every second. The idea here is that we
1631 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1632 * time window. We only want to throttle tasks for recent delay that
1633 * has occurred, in 1 second time windows since that's the maximum
1634 * things can be throttled. We save the current delay window in
1635 * blkg->last_delay so we know what amount is still left to be charged
1636 * to the blkg from this point onward. blkg->last_use keeps track of
1637 * the use_delay counter. The idea is if we're unthrottling the blkg we
1638 * are ok with whatever is happening now, and we can take away more of
1639 * the accumulated delay as we've already throttled enough that
1640 * everybody is happy with their IO latencies.
1642 if (time_before64(old + NSEC_PER_SEC, now) &&
1643 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1644 u64 cur = atomic64_read(&blkg->delay_nsec);
1645 u64 sub = min_t(u64, blkg->last_delay, now - old);
1646 int cur_use = atomic_read(&blkg->use_delay);
1649 * We've been unthrottled, subtract a larger chunk of our
1650 * accumulated delay.
1652 if (cur_use < blkg->last_use)
1653 sub = max_t(u64, sub, blkg->last_delay >> 1);
1656 * This shouldn't happen, but handle it anyway. Our delay_nsec
1657 * should only ever be growing except here where we subtract out
1658 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1659 * rather not end up with negative numbers.
1661 if (unlikely(cur < sub)) {
1662 atomic64_set(&blkg->delay_nsec, 0);
1663 blkg->last_delay = 0;
1665 atomic64_sub(sub, &blkg->delay_nsec);
1666 blkg->last_delay = cur - sub;
1668 blkg->last_use = cur_use;
1673 * This is called when we want to actually walk up the hierarchy and check to
1674 * see if we need to throttle, and then actually throttle if there is some
1675 * accumulated delay. This should only be called upon return to user space so
1676 * we're not holding some lock that would induce a priority inversion.
1678 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1680 u64 now = ktime_to_ns(ktime_get());
1685 while (blkg->parent) {
1686 if (atomic_read(&blkg->use_delay)) {
1687 blkcg_scale_delay(blkg, now);
1688 delay_nsec = max_t(u64, delay_nsec,
1689 atomic64_read(&blkg->delay_nsec));
1691 blkg = blkg->parent;
1698 * Let's not sleep for all eternity if we've amassed a huge delay.
1699 * Swapping or metadata IO can accumulate 10's of seconds worth of
1700 * delay, and we want userspace to be able to do _something_ so cap the
1701 * delays at 1 second. If there's 10's of seconds worth of delay then
1702 * the tasks will be delayed for 1 second for every syscall.
1704 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1707 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1708 * that hasn't landed upstream yet. Once that stuff is in place we need
1709 * to do a psi_memstall_enter/leave if memdelay is set.
1712 exp = ktime_add_ns(now, delay_nsec);
1713 tok = io_schedule_prepare();
1715 __set_current_state(TASK_KILLABLE);
1716 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1718 } while (!fatal_signal_pending(current));
1719 io_schedule_finish(tok);
1723 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1725 * This is only called if we've been marked with set_notify_resume(). Obviously
1726 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1727 * check to see if current->throttle_queue is set and if not this doesn't do
1728 * anything. This should only ever be called by the resume code, it's not meant
1729 * to be called by people willy-nilly as it will actually do the work to
1730 * throttle the task if it is setup for throttling.
1732 void blkcg_maybe_throttle_current(void)
1734 struct request_queue *q = current->throttle_queue;
1735 struct cgroup_subsys_state *css;
1736 struct blkcg *blkcg;
1737 struct blkcg_gq *blkg;
1738 bool use_memdelay = current->use_memdelay;
1743 current->throttle_queue = NULL;
1744 current->use_memdelay = false;
1747 css = kthread_blkcg();
1749 blkcg = css_to_blkcg(css);
1751 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1755 blkg = blkg_lookup(blkcg, q);
1758 blkg = blkg_try_get(blkg);
1763 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1771 EXPORT_SYMBOL_GPL(blkcg_maybe_throttle_current);
1774 * blkcg_schedule_throttle - this task needs to check for throttling
1775 * @q - the request queue IO was submitted on
1776 * @use_memdelay - do we charge this to memory delay for PSI
1778 * This is called by the IO controller when we know there's delay accumulated
1779 * for the blkg for this task. We do not pass the blkg because there are places
1780 * we call this that may not have that information, the swapping code for
1781 * instance will only have a request_queue at that point. This set's the
1782 * notify_resume for the task to check and see if it requires throttling before
1783 * returning to user space.
1785 * We will only schedule once per syscall. You can call this over and over
1786 * again and it will only do the check once upon return to user space, and only
1787 * throttle once. If the task needs to be throttled again it'll need to be
1788 * re-set at the next time we see the task.
1790 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1792 if (unlikely(current->flags & PF_KTHREAD))
1795 if (!blk_get_queue(q))
1798 if (current->throttle_queue)
1799 blk_put_queue(current->throttle_queue);
1800 current->throttle_queue = q;
1802 current->use_memdelay = use_memdelay;
1803 set_notify_resume(current);
1805 EXPORT_SYMBOL_GPL(blkcg_schedule_throttle);
1808 * blkcg_add_delay - add delay to this blkg
1809 * @now - the current time in nanoseconds
1810 * @delta - how many nanoseconds of delay to add
1812 * Charge @delta to the blkg's current delay accumulation. This is used to
1813 * throttle tasks if an IO controller thinks we need more throttling.
1815 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1817 blkcg_scale_delay(blkg, now);
1818 atomic64_add(delta, &blkg->delay_nsec);
1820 EXPORT_SYMBOL_GPL(blkcg_add_delay);
1822 module_param(blkcg_debug_stats, bool, 0644);
1823 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");