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 blkg_rwstat_exit(&blkg->stat_ios);
80 blkg_rwstat_exit(&blkg->stat_bytes);
85 * blkg_alloc - allocate a blkg
86 * @blkcg: block cgroup the new blkg is associated with
87 * @q: request_queue the new blkg is associated with
88 * @gfp_mask: allocation mask to use
90 * Allocate a new blkg assocating @blkcg and @q.
92 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
95 struct blkcg_gq *blkg;
98 /* alloc and init base part */
99 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
103 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
104 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
108 INIT_LIST_HEAD(&blkg->q_node);
110 atomic_set(&blkg->refcnt, 1);
112 for (i = 0; i < BLKCG_MAX_POLS; i++) {
113 struct blkcg_policy *pol = blkcg_policy[i];
114 struct blkg_policy_data *pd;
116 if (!blkcg_policy_enabled(q, pol))
119 /* alloc per-policy data and attach it to blkg */
120 pd = pol->pd_alloc_fn(gfp_mask, q->node);
136 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
137 struct request_queue *q, bool update_hint)
139 struct blkcg_gq *blkg;
142 * Hint didn't match. Look up from the radix tree. Note that the
143 * hint can only be updated under queue_lock as otherwise @blkg
144 * could have already been removed from blkg_tree. The caller is
145 * responsible for grabbing queue_lock if @update_hint.
147 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
148 if (blkg && blkg->q == q) {
150 lockdep_assert_held(&q->queue_lock);
151 rcu_assign_pointer(blkcg->blkg_hint, blkg);
158 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
161 * If @new_blkg is %NULL, this function tries to allocate a new one as
162 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
164 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
165 struct request_queue *q,
166 struct blkcg_gq *new_blkg)
168 struct blkcg_gq *blkg;
169 struct bdi_writeback_congested *wb_congested;
172 WARN_ON_ONCE(!rcu_read_lock_held());
173 lockdep_assert_held(&q->queue_lock);
175 /* blkg holds a reference to blkcg */
176 if (!css_tryget_online(&blkcg->css)) {
181 wb_congested = wb_congested_get_create(q->backing_dev_info,
183 GFP_NOWAIT | __GFP_NOWARN);
191 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
192 if (unlikely(!new_blkg)) {
194 goto err_put_congested;
198 blkg->wb_congested = wb_congested;
201 if (blkcg_parent(blkcg)) {
202 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
203 if (WARN_ON_ONCE(!blkg->parent)) {
205 goto err_put_congested;
207 blkg_get(blkg->parent);
210 /* invoke per-policy init */
211 for (i = 0; i < BLKCG_MAX_POLS; i++) {
212 struct blkcg_policy *pol = blkcg_policy[i];
214 if (blkg->pd[i] && pol->pd_init_fn)
215 pol->pd_init_fn(blkg->pd[i]);
219 spin_lock(&blkcg->lock);
220 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
222 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
223 list_add(&blkg->q_node, &q->blkg_list);
225 for (i = 0; i < BLKCG_MAX_POLS; i++) {
226 struct blkcg_policy *pol = blkcg_policy[i];
228 if (blkg->pd[i] && pol->pd_online_fn)
229 pol->pd_online_fn(blkg->pd[i]);
233 spin_unlock(&blkcg->lock);
238 /* @blkg failed fully initialized, use the usual release path */
243 wb_congested_put(wb_congested);
245 css_put(&blkcg->css);
252 * blkg_lookup_create - lookup blkg, try to create one if not there
253 * @blkcg: blkcg of interest
254 * @q: request_queue of interest
256 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
257 * create one. blkg creation is performed recursively from blkcg_root such
258 * that all non-root blkg's have access to the parent blkg. This function
259 * should be called under RCU read lock and @q->queue_lock.
261 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
262 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
263 * dead and bypassing, returns ERR_PTR(-EBUSY).
265 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
266 struct request_queue *q)
268 struct blkcg_gq *blkg;
270 WARN_ON_ONCE(!rcu_read_lock_held());
271 lockdep_assert_held(&q->queue_lock);
273 blkg = __blkg_lookup(blkcg, q, true);
278 * Create blkgs walking down from blkcg_root to @blkcg, so that all
279 * non-root blkgs have access to their parents.
282 struct blkcg *pos = blkcg;
283 struct blkcg *parent = blkcg_parent(blkcg);
285 while (parent && !__blkg_lookup(parent, q, false)) {
287 parent = blkcg_parent(parent);
290 blkg = blkg_create(pos, q, NULL);
291 if (pos == blkcg || IS_ERR(blkg))
296 static void blkg_destroy(struct blkcg_gq *blkg)
298 struct blkcg *blkcg = blkg->blkcg;
299 struct blkcg_gq *parent = blkg->parent;
302 lockdep_assert_held(&blkg->q->queue_lock);
303 lockdep_assert_held(&blkcg->lock);
305 /* Something wrong if we are trying to remove same group twice */
306 WARN_ON_ONCE(list_empty(&blkg->q_node));
307 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
309 for (i = 0; i < BLKCG_MAX_POLS; i++) {
310 struct blkcg_policy *pol = blkcg_policy[i];
312 if (blkg->pd[i] && pol->pd_offline_fn)
313 pol->pd_offline_fn(blkg->pd[i]);
317 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
318 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
321 blkg->online = false;
323 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
324 list_del_init(&blkg->q_node);
325 hlist_del_init_rcu(&blkg->blkcg_node);
328 * Both setting lookup hint to and clearing it from @blkg are done
329 * under queue_lock. If it's not pointing to @blkg now, it never
330 * will. Hint assignment itself can race safely.
332 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
333 rcu_assign_pointer(blkcg->blkg_hint, NULL);
336 * Put the reference taken at the time of creation so that when all
337 * queues are gone, group can be destroyed.
343 * blkg_destroy_all - destroy all blkgs associated with a request_queue
344 * @q: request_queue of interest
346 * Destroy all blkgs associated with @q.
348 static void blkg_destroy_all(struct request_queue *q)
350 struct blkcg_gq *blkg, *n;
352 spin_lock_irq(&q->queue_lock);
353 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
354 struct blkcg *blkcg = blkg->blkcg;
356 spin_lock(&blkcg->lock);
358 spin_unlock(&blkcg->lock);
362 spin_unlock_irq(&q->queue_lock);
366 * A group is RCU protected, but having an rcu lock does not mean that one
367 * can access all the fields of blkg and assume these are valid. For
368 * example, don't try to follow throtl_data and request queue links.
370 * Having a reference to blkg under an rcu allows accesses to only values
371 * local to groups like group stats and group rate limits.
373 void __blkg_release_rcu(struct rcu_head *rcu_head)
375 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
377 /* release the blkcg and parent blkg refs this blkg has been holding */
378 css_put(&blkg->blkcg->css);
380 blkg_put(blkg->parent);
382 wb_congested_put(blkg->wb_congested);
386 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
388 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
389 struct cftype *cftype, u64 val)
391 struct blkcg *blkcg = css_to_blkcg(css);
392 struct blkcg_gq *blkg;
395 mutex_lock(&blkcg_pol_mutex);
396 spin_lock_irq(&blkcg->lock);
399 * Note that stat reset is racy - it doesn't synchronize against
400 * stat updates. This is a debug feature which shouldn't exist
401 * anyway. If you get hit by a race, retry.
403 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
404 blkg_rwstat_reset(&blkg->stat_bytes);
405 blkg_rwstat_reset(&blkg->stat_ios);
407 for (i = 0; i < BLKCG_MAX_POLS; i++) {
408 struct blkcg_policy *pol = blkcg_policy[i];
410 if (blkg->pd[i] && pol->pd_reset_stats_fn)
411 pol->pd_reset_stats_fn(blkg->pd[i]);
415 spin_unlock_irq(&blkcg->lock);
416 mutex_unlock(&blkcg_pol_mutex);
420 const char *blkg_dev_name(struct blkcg_gq *blkg)
422 /* some drivers (floppy) instantiate a queue w/o disk registered */
423 if (blkg->q->backing_dev_info->dev)
424 return dev_name(blkg->q->backing_dev_info->dev);
429 * blkcg_print_blkgs - helper for printing per-blkg data
430 * @sf: seq_file to print to
431 * @blkcg: blkcg of interest
432 * @prfill: fill function to print out a blkg
433 * @pol: policy in question
434 * @data: data to be passed to @prfill
435 * @show_total: to print out sum of prfill return values or not
437 * This function invokes @prfill on each blkg of @blkcg if pd for the
438 * policy specified by @pol exists. @prfill is invoked with @sf, the
439 * policy data and @data and the matching queue lock held. If @show_total
440 * is %true, the sum of the return values from @prfill is printed with
441 * "Total" label at the end.
443 * This is to be used to construct print functions for
444 * cftype->read_seq_string method.
446 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
447 u64 (*prfill)(struct seq_file *,
448 struct blkg_policy_data *, int),
449 const struct blkcg_policy *pol, int data,
452 struct blkcg_gq *blkg;
456 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
457 spin_lock_irq(&blkg->q->queue_lock);
458 if (blkcg_policy_enabled(blkg->q, pol))
459 total += prfill(sf, blkg->pd[pol->plid], data);
460 spin_unlock_irq(&blkg->q->queue_lock);
465 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
467 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
470 * __blkg_prfill_u64 - prfill helper for a single u64 value
471 * @sf: seq_file to print to
472 * @pd: policy private data of interest
475 * Print @v to @sf for the device assocaited with @pd.
477 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
479 const char *dname = blkg_dev_name(pd->blkg);
484 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
487 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
490 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
491 * @sf: seq_file to print to
492 * @pd: policy private data of interest
493 * @rwstat: rwstat to print
495 * Print @rwstat to @sf for the device assocaited with @pd.
497 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
498 const struct blkg_rwstat *rwstat)
500 static const char *rwstr[] = {
501 [BLKG_RWSTAT_READ] = "Read",
502 [BLKG_RWSTAT_WRITE] = "Write",
503 [BLKG_RWSTAT_SYNC] = "Sync",
504 [BLKG_RWSTAT_ASYNC] = "Async",
505 [BLKG_RWSTAT_DISCARD] = "Discard",
507 const char *dname = blkg_dev_name(pd->blkg);
514 for (i = 0; i < BLKG_RWSTAT_NR; i++)
515 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
516 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
518 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
519 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]) +
520 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_DISCARD]);
521 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
524 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
527 * blkg_prfill_stat - prfill callback for blkg_stat
528 * @sf: seq_file to print to
529 * @pd: policy private data of interest
530 * @off: offset to the blkg_stat in @pd
532 * prfill callback for printing a blkg_stat.
534 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
536 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
538 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
541 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
542 * @sf: seq_file to print to
543 * @pd: policy private data of interest
544 * @off: offset to the blkg_rwstat in @pd
546 * prfill callback for printing a blkg_rwstat.
548 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
551 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
553 return __blkg_prfill_rwstat(sf, pd, &rwstat);
555 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
557 static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
558 struct blkg_policy_data *pd, int off)
560 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
562 return __blkg_prfill_rwstat(sf, pd, &rwstat);
566 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
567 * @sf: seq_file to print to
570 * To be used as cftype->seq_show to print blkg->stat_bytes.
571 * cftype->private must be set to the blkcg_policy.
573 int blkg_print_stat_bytes(struct seq_file *sf, void *v)
575 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
576 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
577 offsetof(struct blkcg_gq, stat_bytes), true);
580 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
583 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
584 * @sf: seq_file to print to
587 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
588 * must be set to the blkcg_policy.
590 int blkg_print_stat_ios(struct seq_file *sf, void *v)
592 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
593 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
594 offsetof(struct blkcg_gq, stat_ios), true);
597 EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
599 static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
600 struct blkg_policy_data *pd,
603 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
605 return __blkg_prfill_rwstat(sf, pd, &rwstat);
609 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
610 * @sf: seq_file to print to
613 int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
615 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
616 blkg_prfill_rwstat_field_recursive,
617 (void *)seq_cft(sf)->private,
618 offsetof(struct blkcg_gq, stat_bytes), true);
621 EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
624 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
625 * @sf: seq_file to print to
628 int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
630 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
631 blkg_prfill_rwstat_field_recursive,
632 (void *)seq_cft(sf)->private,
633 offsetof(struct blkcg_gq, stat_ios), true);
636 EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
639 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
640 * @blkg: blkg of interest
641 * @pol: blkcg_policy which contains the blkg_stat
642 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
644 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
645 * online descendants and their aux counts. The caller must be holding the
646 * queue lock for online tests.
648 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
649 * at @off bytes into @blkg's blkg_policy_data of the policy.
651 u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
652 struct blkcg_policy *pol, int off)
654 struct blkcg_gq *pos_blkg;
655 struct cgroup_subsys_state *pos_css;
658 lockdep_assert_held(&blkg->q->queue_lock);
661 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
662 struct blkg_stat *stat;
664 if (!pos_blkg->online)
668 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
670 stat = (void *)blkg + off;
672 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
678 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
681 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
682 * @blkg: blkg of interest
683 * @pol: blkcg_policy which contains the blkg_rwstat
684 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
686 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
687 * online descendants and their aux counts. The caller must be holding the
688 * queue lock for online tests.
690 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
691 * is at @off bytes into @blkg's blkg_policy_data of the policy.
693 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
694 struct blkcg_policy *pol, int off)
696 struct blkcg_gq *pos_blkg;
697 struct cgroup_subsys_state *pos_css;
698 struct blkg_rwstat sum = { };
701 lockdep_assert_held(&blkg->q->queue_lock);
704 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
705 struct blkg_rwstat *rwstat;
707 if (!pos_blkg->online)
711 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
713 rwstat = (void *)pos_blkg + off;
715 for (i = 0; i < BLKG_RWSTAT_NR; i++)
716 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
717 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
724 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
726 /* Performs queue bypass and policy enabled checks then looks up blkg. */
727 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
728 const struct blkcg_policy *pol,
729 struct request_queue *q)
731 WARN_ON_ONCE(!rcu_read_lock_held());
732 lockdep_assert_held(&q->queue_lock);
734 if (!blkcg_policy_enabled(q, pol))
735 return ERR_PTR(-EOPNOTSUPP);
736 return __blkg_lookup(blkcg, q, true /* update_hint */);
740 * blkg_conf_prep - parse and prepare for per-blkg config update
741 * @blkcg: target block cgroup
742 * @pol: target policy
743 * @input: input string
744 * @ctx: blkg_conf_ctx to be filled
746 * Parse per-blkg config update from @input and initialize @ctx with the
747 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
748 * part of @input following MAJ:MIN. This function returns with RCU read
749 * lock and queue lock held and must be paired with blkg_conf_finish().
751 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
752 char *input, struct blkg_conf_ctx *ctx)
753 __acquires(rcu) __acquires(&disk->queue->queue_lock)
755 struct gendisk *disk;
756 struct request_queue *q;
757 struct blkcg_gq *blkg;
758 unsigned int major, minor;
759 int key_len, part, ret;
762 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
765 body = input + key_len;
768 body = skip_spaces(body);
770 disk = get_gendisk(MKDEV(major, minor), &part);
781 spin_lock_irq(&q->queue_lock);
783 blkg = blkg_lookup_check(blkcg, pol, q);
793 * Create blkgs walking down from blkcg_root to @blkcg, so that all
794 * non-root blkgs have access to their parents.
797 struct blkcg *pos = blkcg;
798 struct blkcg *parent;
799 struct blkcg_gq *new_blkg;
801 parent = blkcg_parent(blkcg);
802 while (parent && !__blkg_lookup(parent, q, false)) {
804 parent = blkcg_parent(parent);
807 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
808 spin_unlock_irq(&q->queue_lock);
811 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
812 if (unlikely(!new_blkg)) {
818 spin_lock_irq(&q->queue_lock);
820 blkg = blkg_lookup_check(pos, pol, q);
829 blkg = blkg_create(pos, q, new_blkg);
830 if (unlikely(IS_ERR(blkg))) {
846 spin_unlock_irq(&q->queue_lock);
849 put_disk_and_module(disk);
851 * If queue was bypassing, we should retry. Do so after a
852 * short msleep(). It isn't strictly necessary but queue
853 * can be bypassing for some time and it's always nice to
854 * avoid busy looping.
858 ret = restart_syscall();
864 * blkg_conf_finish - finish up per-blkg config update
865 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
867 * Finish up after per-blkg config update. This function must be paired
868 * with blkg_conf_prep().
870 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
871 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
873 spin_unlock_irq(&ctx->disk->queue->queue_lock);
875 put_disk_and_module(ctx->disk);
878 static int blkcg_print_stat(struct seq_file *sf, void *v)
880 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
881 struct blkcg_gq *blkg;
885 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
888 struct blkg_rwstat rwstat;
889 u64 rbytes, wbytes, rios, wios, dbytes, dios;
890 size_t size = seq_get_buf(sf, &buf), off = 0;
892 bool has_stats = false;
894 dname = blkg_dev_name(blkg);
899 * Hooray string manipulation, count is the size written NOT
900 * INCLUDING THE \0, so size is now count+1 less than what we
901 * had before, but we want to start writing the next bit from
902 * the \0 so we only add count to buf.
904 off += scnprintf(buf+off, size-off, "%s ", dname);
906 spin_lock_irq(&blkg->q->queue_lock);
908 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
909 offsetof(struct blkcg_gq, stat_bytes));
910 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
911 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
912 dbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
914 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
915 offsetof(struct blkcg_gq, stat_ios));
916 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
917 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
918 dios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_DISCARD]);
920 spin_unlock_irq(&blkg->q->queue_lock);
922 if (rbytes || wbytes || rios || wios) {
924 off += scnprintf(buf+off, size-off,
925 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
926 rbytes, wbytes, rios, wios,
930 if (!blkcg_debug_stats)
933 if (atomic_read(&blkg->use_delay)) {
935 off += scnprintf(buf+off, size-off,
936 " use_delay=%d delay_nsec=%llu",
937 atomic_read(&blkg->use_delay),
938 (unsigned long long)atomic64_read(&blkg->delay_nsec));
941 for (i = 0; i < BLKCG_MAX_POLS; i++) {
942 struct blkcg_policy *pol = blkcg_policy[i];
945 if (!blkg->pd[i] || !pol->pd_stat_fn)
948 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
955 off += scnprintf(buf+off, size-off, "\n");
964 static struct cftype blkcg_files[] = {
967 .flags = CFTYPE_NOT_ON_ROOT,
968 .seq_show = blkcg_print_stat,
973 static struct cftype blkcg_legacy_files[] = {
975 .name = "reset_stats",
976 .write_u64 = blkcg_reset_stats,
982 * blkcg destruction is a three-stage process.
984 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
985 * which offlines writeback. Here we tie the next stage of blkg destruction
986 * to the completion of writeback associated with the blkcg. This lets us
987 * avoid punting potentially large amounts of outstanding writeback to root
988 * while maintaining any ongoing policies. The next stage is triggered when
989 * the nr_cgwbs count goes to zero.
991 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
992 * and handles the destruction of blkgs. Here the css reference held by
993 * the blkg is put back eventually allowing blkcg_css_free() to be called.
994 * This work may occur in cgwb_release_workfn() on the cgwb_release
995 * workqueue. Any submitted ios that fail to get the blkg ref will be
996 * punted to the root_blkg.
998 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
999 * This finally frees the blkcg.
1003 * blkcg_css_offline - cgroup css_offline callback
1004 * @css: css of interest
1006 * This function is called when @css is about to go away. Here the cgwbs are
1007 * offlined first and only once writeback associated with the blkcg has
1008 * finished do we start step 2 (see above).
1010 static void blkcg_css_offline(struct cgroup_subsys_state *css)
1012 struct blkcg *blkcg = css_to_blkcg(css);
1014 /* this prevents anyone from attaching or migrating to this blkcg */
1015 wb_blkcg_offline(blkcg);
1017 /* put the base cgwb reference allowing step 2 to be triggered */
1018 blkcg_cgwb_put(blkcg);
1022 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1023 * @blkcg: blkcg of interest
1025 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1026 * is nested inside q lock, this function performs reverse double lock dancing.
1027 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1028 * blkcg_css_free to eventually be called.
1030 * This is the blkcg counterpart of ioc_release_fn().
1032 void blkcg_destroy_blkgs(struct blkcg *blkcg)
1034 spin_lock_irq(&blkcg->lock);
1036 while (!hlist_empty(&blkcg->blkg_list)) {
1037 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1038 struct blkcg_gq, blkcg_node);
1039 struct request_queue *q = blkg->q;
1041 if (spin_trylock(&q->queue_lock)) {
1043 spin_unlock(&q->queue_lock);
1045 spin_unlock_irq(&blkcg->lock);
1047 spin_lock_irq(&blkcg->lock);
1051 spin_unlock_irq(&blkcg->lock);
1054 static void blkcg_css_free(struct cgroup_subsys_state *css)
1056 struct blkcg *blkcg = css_to_blkcg(css);
1059 mutex_lock(&blkcg_pol_mutex);
1061 list_del(&blkcg->all_blkcgs_node);
1063 for (i = 0; i < BLKCG_MAX_POLS; i++)
1065 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1067 mutex_unlock(&blkcg_pol_mutex);
1072 static struct cgroup_subsys_state *
1073 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1075 struct blkcg *blkcg;
1076 struct cgroup_subsys_state *ret;
1079 mutex_lock(&blkcg_pol_mutex);
1082 blkcg = &blkcg_root;
1084 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1086 ret = ERR_PTR(-ENOMEM);
1091 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1092 struct blkcg_policy *pol = blkcg_policy[i];
1093 struct blkcg_policy_data *cpd;
1096 * If the policy hasn't been attached yet, wait for it
1097 * to be attached before doing anything else. Otherwise,
1098 * check if the policy requires any specific per-cgroup
1099 * data: if it does, allocate and initialize it.
1101 if (!pol || !pol->cpd_alloc_fn)
1104 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1106 ret = ERR_PTR(-ENOMEM);
1109 blkcg->cpd[i] = cpd;
1112 if (pol->cpd_init_fn)
1113 pol->cpd_init_fn(cpd);
1116 spin_lock_init(&blkcg->lock);
1117 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1118 INIT_HLIST_HEAD(&blkcg->blkg_list);
1119 #ifdef CONFIG_CGROUP_WRITEBACK
1120 INIT_LIST_HEAD(&blkcg->cgwb_list);
1121 refcount_set(&blkcg->cgwb_refcnt, 1);
1123 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1125 mutex_unlock(&blkcg_pol_mutex);
1129 for (i--; i >= 0; i--)
1131 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1133 if (blkcg != &blkcg_root)
1136 mutex_unlock(&blkcg_pol_mutex);
1141 * blkcg_init_queue - initialize blkcg part of request queue
1142 * @q: request_queue to initialize
1144 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1145 * part of new request_queue @q.
1148 * 0 on success, -errno on failure.
1150 int blkcg_init_queue(struct request_queue *q)
1152 struct blkcg_gq *new_blkg, *blkg;
1156 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1160 preloaded = !radix_tree_preload(GFP_KERNEL);
1162 /* Make sure the root blkg exists. */
1164 spin_lock_irq(&q->queue_lock);
1165 blkg = blkg_create(&blkcg_root, q, new_blkg);
1168 q->root_blkg = blkg;
1169 spin_unlock_irq(&q->queue_lock);
1173 radix_tree_preload_end();
1175 ret = blk_iolatency_init(q);
1177 goto err_destroy_all;
1179 ret = blk_throtl_init(q);
1181 goto err_destroy_all;
1185 blkg_destroy_all(q);
1188 spin_unlock_irq(&q->queue_lock);
1191 radix_tree_preload_end();
1192 return PTR_ERR(blkg);
1196 * blkcg_drain_queue - drain blkcg part of request_queue
1197 * @q: request_queue to drain
1199 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1201 void blkcg_drain_queue(struct request_queue *q)
1203 lockdep_assert_held(&q->queue_lock);
1206 * @q could be exiting and already have destroyed all blkgs as
1207 * indicated by NULL root_blkg. If so, don't confuse policies.
1212 blk_throtl_drain(q);
1216 * blkcg_exit_queue - exit and release blkcg part of request_queue
1217 * @q: request_queue being released
1219 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1221 void blkcg_exit_queue(struct request_queue *q)
1223 blkg_destroy_all(q);
1228 * We cannot support shared io contexts, as we have no mean to support
1229 * two tasks with the same ioc in two different groups without major rework
1230 * of the main cic data structures. For now we allow a task to change
1231 * its cgroup only if it's the only owner of its ioc.
1233 static int blkcg_can_attach(struct cgroup_taskset *tset)
1235 struct task_struct *task;
1236 struct cgroup_subsys_state *dst_css;
1237 struct io_context *ioc;
1240 /* task_lock() is needed to avoid races with exit_io_context() */
1241 cgroup_taskset_for_each(task, dst_css, tset) {
1243 ioc = task->io_context;
1244 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1253 static void blkcg_bind(struct cgroup_subsys_state *root_css)
1257 mutex_lock(&blkcg_pol_mutex);
1259 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1260 struct blkcg_policy *pol = blkcg_policy[i];
1261 struct blkcg *blkcg;
1263 if (!pol || !pol->cpd_bind_fn)
1266 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1267 if (blkcg->cpd[pol->plid])
1268 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1270 mutex_unlock(&blkcg_pol_mutex);
1273 static void blkcg_exit(struct task_struct *tsk)
1275 if (tsk->throttle_queue)
1276 blk_put_queue(tsk->throttle_queue);
1277 tsk->throttle_queue = NULL;
1280 struct cgroup_subsys io_cgrp_subsys = {
1281 .css_alloc = blkcg_css_alloc,
1282 .css_offline = blkcg_css_offline,
1283 .css_free = blkcg_css_free,
1284 .can_attach = blkcg_can_attach,
1286 .dfl_cftypes = blkcg_files,
1287 .legacy_cftypes = blkcg_legacy_files,
1288 .legacy_name = "blkio",
1292 * This ensures that, if available, memcg is automatically enabled
1293 * together on the default hierarchy so that the owner cgroup can
1294 * be retrieved from writeback pages.
1296 .depends_on = 1 << memory_cgrp_id,
1299 EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1302 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1303 * @q: request_queue of interest
1304 * @pol: blkcg policy to activate
1306 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1307 * bypass mode to populate its blkgs with policy_data for @pol.
1309 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1310 * from IO path. Update of each blkg is protected by both queue and blkcg
1311 * locks so that holding either lock and testing blkcg_policy_enabled() is
1312 * always enough for dereferencing policy data.
1314 * The caller is responsible for synchronizing [de]activations and policy
1315 * [un]registerations. Returns 0 on success, -errno on failure.
1317 int blkcg_activate_policy(struct request_queue *q,
1318 const struct blkcg_policy *pol)
1320 struct blkg_policy_data *pd_prealloc = NULL;
1321 struct blkcg_gq *blkg;
1324 if (blkcg_policy_enabled(q, pol))
1328 blk_mq_freeze_queue(q);
1331 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1334 goto out_bypass_end;
1338 spin_lock_irq(&q->queue_lock);
1340 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1341 struct blkg_policy_data *pd;
1343 if (blkg->pd[pol->plid])
1346 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1348 swap(pd, pd_prealloc);
1350 spin_unlock_irq(&q->queue_lock);
1354 blkg->pd[pol->plid] = pd;
1356 pd->plid = pol->plid;
1357 if (pol->pd_init_fn)
1358 pol->pd_init_fn(pd);
1361 __set_bit(pol->plid, q->blkcg_pols);
1364 spin_unlock_irq(&q->queue_lock);
1367 blk_mq_unfreeze_queue(q);
1369 pol->pd_free_fn(pd_prealloc);
1372 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1375 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1376 * @q: request_queue of interest
1377 * @pol: blkcg policy to deactivate
1379 * Deactivate @pol on @q. Follows the same synchronization rules as
1380 * blkcg_activate_policy().
1382 void blkcg_deactivate_policy(struct request_queue *q,
1383 const struct blkcg_policy *pol)
1385 struct blkcg_gq *blkg;
1387 if (!blkcg_policy_enabled(q, pol))
1391 blk_mq_freeze_queue(q);
1393 spin_lock_irq(&q->queue_lock);
1395 __clear_bit(pol->plid, q->blkcg_pols);
1397 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1398 if (blkg->pd[pol->plid]) {
1399 if (pol->pd_offline_fn)
1400 pol->pd_offline_fn(blkg->pd[pol->plid]);
1401 pol->pd_free_fn(blkg->pd[pol->plid]);
1402 blkg->pd[pol->plid] = NULL;
1406 spin_unlock_irq(&q->queue_lock);
1409 blk_mq_unfreeze_queue(q);
1411 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1414 * blkcg_policy_register - register a blkcg policy
1415 * @pol: blkcg policy to register
1417 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1418 * successful registration. Returns 0 on success and -errno on failure.
1420 int blkcg_policy_register(struct blkcg_policy *pol)
1422 struct blkcg *blkcg;
1425 mutex_lock(&blkcg_pol_register_mutex);
1426 mutex_lock(&blkcg_pol_mutex);
1428 /* find an empty slot */
1430 for (i = 0; i < BLKCG_MAX_POLS; i++)
1431 if (!blkcg_policy[i])
1433 if (i >= BLKCG_MAX_POLS) {
1434 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1438 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1439 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1440 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1445 blkcg_policy[pol->plid] = pol;
1447 /* allocate and install cpd's */
1448 if (pol->cpd_alloc_fn) {
1449 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1450 struct blkcg_policy_data *cpd;
1452 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1456 blkcg->cpd[pol->plid] = cpd;
1458 cpd->plid = pol->plid;
1459 pol->cpd_init_fn(cpd);
1463 mutex_unlock(&blkcg_pol_mutex);
1465 /* everything is in place, add intf files for the new policy */
1466 if (pol->dfl_cftypes)
1467 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1469 if (pol->legacy_cftypes)
1470 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1471 pol->legacy_cftypes));
1472 mutex_unlock(&blkcg_pol_register_mutex);
1476 if (pol->cpd_free_fn) {
1477 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1478 if (blkcg->cpd[pol->plid]) {
1479 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1480 blkcg->cpd[pol->plid] = NULL;
1484 blkcg_policy[pol->plid] = NULL;
1486 mutex_unlock(&blkcg_pol_mutex);
1487 mutex_unlock(&blkcg_pol_register_mutex);
1490 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1493 * blkcg_policy_unregister - unregister a blkcg policy
1494 * @pol: blkcg policy to unregister
1496 * Undo blkcg_policy_register(@pol). Might sleep.
1498 void blkcg_policy_unregister(struct blkcg_policy *pol)
1500 struct blkcg *blkcg;
1502 mutex_lock(&blkcg_pol_register_mutex);
1504 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1507 /* kill the intf files first */
1508 if (pol->dfl_cftypes)
1509 cgroup_rm_cftypes(pol->dfl_cftypes);
1510 if (pol->legacy_cftypes)
1511 cgroup_rm_cftypes(pol->legacy_cftypes);
1513 /* remove cpds and unregister */
1514 mutex_lock(&blkcg_pol_mutex);
1516 if (pol->cpd_free_fn) {
1517 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1518 if (blkcg->cpd[pol->plid]) {
1519 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1520 blkcg->cpd[pol->plid] = NULL;
1524 blkcg_policy[pol->plid] = NULL;
1526 mutex_unlock(&blkcg_pol_mutex);
1528 mutex_unlock(&blkcg_pol_register_mutex);
1530 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1533 * Scale the accumulated delay based on how long it has been since we updated
1534 * the delay. We only call this when we are adding delay, in case it's been a
1535 * while since we added delay, and when we are checking to see if we need to
1536 * delay a task, to account for any delays that may have occurred.
1538 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1540 u64 old = atomic64_read(&blkg->delay_start);
1543 * We only want to scale down every second. The idea here is that we
1544 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1545 * time window. We only want to throttle tasks for recent delay that
1546 * has occurred, in 1 second time windows since that's the maximum
1547 * things can be throttled. We save the current delay window in
1548 * blkg->last_delay so we know what amount is still left to be charged
1549 * to the blkg from this point onward. blkg->last_use keeps track of
1550 * the use_delay counter. The idea is if we're unthrottling the blkg we
1551 * are ok with whatever is happening now, and we can take away more of
1552 * the accumulated delay as we've already throttled enough that
1553 * everybody is happy with their IO latencies.
1555 if (time_before64(old + NSEC_PER_SEC, now) &&
1556 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1557 u64 cur = atomic64_read(&blkg->delay_nsec);
1558 u64 sub = min_t(u64, blkg->last_delay, now - old);
1559 int cur_use = atomic_read(&blkg->use_delay);
1562 * We've been unthrottled, subtract a larger chunk of our
1563 * accumulated delay.
1565 if (cur_use < blkg->last_use)
1566 sub = max_t(u64, sub, blkg->last_delay >> 1);
1569 * This shouldn't happen, but handle it anyway. Our delay_nsec
1570 * should only ever be growing except here where we subtract out
1571 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1572 * rather not end up with negative numbers.
1574 if (unlikely(cur < sub)) {
1575 atomic64_set(&blkg->delay_nsec, 0);
1576 blkg->last_delay = 0;
1578 atomic64_sub(sub, &blkg->delay_nsec);
1579 blkg->last_delay = cur - sub;
1581 blkg->last_use = cur_use;
1586 * This is called when we want to actually walk up the hierarchy and check to
1587 * see if we need to throttle, and then actually throttle if there is some
1588 * accumulated delay. This should only be called upon return to user space so
1589 * we're not holding some lock that would induce a priority inversion.
1591 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1593 u64 now = ktime_to_ns(ktime_get());
1598 while (blkg->parent) {
1599 if (atomic_read(&blkg->use_delay)) {
1600 blkcg_scale_delay(blkg, now);
1601 delay_nsec = max_t(u64, delay_nsec,
1602 atomic64_read(&blkg->delay_nsec));
1604 blkg = blkg->parent;
1611 * Let's not sleep for all eternity if we've amassed a huge delay.
1612 * Swapping or metadata IO can accumulate 10's of seconds worth of
1613 * delay, and we want userspace to be able to do _something_ so cap the
1614 * delays at 1 second. If there's 10's of seconds worth of delay then
1615 * the tasks will be delayed for 1 second for every syscall.
1617 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1620 * TODO: the use_memdelay flag is going to be for the upcoming psi stuff
1621 * that hasn't landed upstream yet. Once that stuff is in place we need
1622 * to do a psi_memstall_enter/leave if memdelay is set.
1625 exp = ktime_add_ns(now, delay_nsec);
1626 tok = io_schedule_prepare();
1628 __set_current_state(TASK_KILLABLE);
1629 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1631 } while (!fatal_signal_pending(current));
1632 io_schedule_finish(tok);
1636 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1638 * This is only called if we've been marked with set_notify_resume(). Obviously
1639 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1640 * check to see if current->throttle_queue is set and if not this doesn't do
1641 * anything. This should only ever be called by the resume code, it's not meant
1642 * to be called by people willy-nilly as it will actually do the work to
1643 * throttle the task if it is setup for throttling.
1645 void blkcg_maybe_throttle_current(void)
1647 struct request_queue *q = current->throttle_queue;
1648 struct cgroup_subsys_state *css;
1649 struct blkcg *blkcg;
1650 struct blkcg_gq *blkg;
1651 bool use_memdelay = current->use_memdelay;
1656 current->throttle_queue = NULL;
1657 current->use_memdelay = false;
1660 css = kthread_blkcg();
1662 blkcg = css_to_blkcg(css);
1664 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1668 blkg = blkg_lookup(blkcg, q);
1671 blkg = blkg_try_get(blkg);
1676 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1686 * blkcg_schedule_throttle - this task needs to check for throttling
1687 * @q - the request queue IO was submitted on
1688 * @use_memdelay - do we charge this to memory delay for PSI
1690 * This is called by the IO controller when we know there's delay accumulated
1691 * for the blkg for this task. We do not pass the blkg because there are places
1692 * we call this that may not have that information, the swapping code for
1693 * instance will only have a request_queue at that point. This set's the
1694 * notify_resume for the task to check and see if it requires throttling before
1695 * returning to user space.
1697 * We will only schedule once per syscall. You can call this over and over
1698 * again and it will only do the check once upon return to user space, and only
1699 * throttle once. If the task needs to be throttled again it'll need to be
1700 * re-set at the next time we see the task.
1702 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1704 if (unlikely(current->flags & PF_KTHREAD))
1707 if (!blk_get_queue(q))
1710 if (current->throttle_queue)
1711 blk_put_queue(current->throttle_queue);
1712 current->throttle_queue = q;
1714 current->use_memdelay = use_memdelay;
1715 set_notify_resume(current);
1719 * blkcg_add_delay - add delay to this blkg
1720 * @now - the current time in nanoseconds
1721 * @delta - how many nanoseconds of delay to add
1723 * Charge @delta to the blkg's current delay accumulation. This is used to
1724 * throttle tasks if an IO controller thinks we need more throttling.
1726 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1728 blkcg_scale_delay(blkg, now);
1729 atomic64_add(delta, &blkg->delay_nsec);
1732 module_param(blkcg_debug_stats, bool, 0644);
1733 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");