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/err.h>
21 #include <linux/blkdev.h>
22 #include <linux/backing-dev.h>
23 #include <linux/slab.h>
24 #include <linux/genhd.h>
25 #include <linux/delay.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-cgroup.h>
30 #define MAX_KEY_LEN 100
33 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
34 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
35 * policy [un]register operations including cgroup file additions /
36 * removals. Putting cgroup file registration outside blkcg_pol_mutex
37 * allows grabbing it from cgroup callbacks.
39 static DEFINE_MUTEX(blkcg_pol_register_mutex);
40 static DEFINE_MUTEX(blkcg_pol_mutex);
42 struct blkcg blkcg_root;
43 EXPORT_SYMBOL_GPL(blkcg_root);
45 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
47 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
49 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
51 static bool blkcg_policy_enabled(struct request_queue *q,
52 const struct blkcg_policy *pol)
54 return pol && test_bit(pol->plid, q->blkcg_pols);
58 * blkg_free - free a blkg
61 * Free @blkg which may be partially allocated.
63 static void blkg_free(struct blkcg_gq *blkg)
70 for (i = 0; i < BLKCG_MAX_POLS; i++)
73 blk_exit_rl(&blkg->rl);
78 * blkg_alloc - allocate a blkg
79 * @blkcg: block cgroup the new blkg is associated with
80 * @q: request_queue the new blkg is associated with
81 * @gfp_mask: allocation mask to use
83 * Allocate a new blkg assocating @blkcg and @q.
85 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
88 struct blkcg_gq *blkg;
91 /* alloc and init base part */
92 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
97 INIT_LIST_HEAD(&blkg->q_node);
99 atomic_set(&blkg->refcnt, 1);
101 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
102 if (blkcg != &blkcg_root) {
103 if (blk_init_rl(&blkg->rl, q, gfp_mask))
105 blkg->rl.blkg = blkg;
108 for (i = 0; i < BLKCG_MAX_POLS; i++) {
109 struct blkcg_policy *pol = blkcg_policy[i];
110 struct blkg_policy_data *pd;
112 if (!blkcg_policy_enabled(q, pol))
115 /* alloc per-policy data and attach it to blkg */
116 pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
133 * __blkg_lookup - internal version of blkg_lookup()
134 * @blkcg: blkcg of interest
135 * @q: request_queue of interest
136 * @update_hint: whether to update lookup hint with the result or not
138 * This is internal version and shouldn't be used by policy
139 * implementations. Looks up blkgs for the @blkcg - @q pair regardless of
140 * @q's bypass state. If @update_hint is %true, the caller should be
141 * holding @q->queue_lock and lookup hint is updated on success.
143 struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg, struct request_queue *q,
146 struct blkcg_gq *blkg;
148 blkg = rcu_dereference(blkcg->blkg_hint);
149 if (blkg && blkg->q == q)
153 * Hint didn't match. Look up from the radix tree. Note that the
154 * hint can only be updated under queue_lock as otherwise @blkg
155 * could have already been removed from blkg_tree. The caller is
156 * responsible for grabbing queue_lock if @update_hint.
158 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
159 if (blkg && blkg->q == q) {
161 lockdep_assert_held(q->queue_lock);
162 rcu_assign_pointer(blkcg->blkg_hint, blkg);
171 * blkg_lookup - lookup blkg for the specified blkcg - q pair
172 * @blkcg: blkcg of interest
173 * @q: request_queue of interest
175 * Lookup blkg for the @blkcg - @q pair. This function should be called
176 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
177 * - see blk_queue_bypass_start() for details.
179 struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
181 WARN_ON_ONCE(!rcu_read_lock_held());
183 if (unlikely(blk_queue_bypass(q)))
185 return __blkg_lookup(blkcg, q, false);
187 EXPORT_SYMBOL_GPL(blkg_lookup);
190 * If @new_blkg is %NULL, this function tries to allocate a new one as
191 * necessary using %GFP_ATOMIC. @new_blkg is always consumed on return.
193 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
194 struct request_queue *q,
195 struct blkcg_gq *new_blkg)
197 struct blkcg_gq *blkg;
198 struct bdi_writeback_congested *wb_congested;
201 WARN_ON_ONCE(!rcu_read_lock_held());
202 lockdep_assert_held(q->queue_lock);
204 /* blkg holds a reference to blkcg */
205 if (!css_tryget_online(&blkcg->css)) {
210 wb_congested = wb_congested_get_create(&q->backing_dev_info,
211 blkcg->css.id, GFP_ATOMIC);
219 new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
220 if (unlikely(!new_blkg)) {
222 goto err_put_congested;
226 blkg->wb_congested = wb_congested;
229 if (blkcg_parent(blkcg)) {
230 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
231 if (WARN_ON_ONCE(!blkg->parent)) {
233 goto err_put_congested;
235 blkg_get(blkg->parent);
238 /* invoke per-policy init */
239 for (i = 0; i < BLKCG_MAX_POLS; i++) {
240 struct blkcg_policy *pol = blkcg_policy[i];
242 if (blkg->pd[i] && pol->pd_init_fn)
243 pol->pd_init_fn(blkg);
247 spin_lock(&blkcg->lock);
248 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
250 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
251 list_add(&blkg->q_node, &q->blkg_list);
253 for (i = 0; i < BLKCG_MAX_POLS; i++) {
254 struct blkcg_policy *pol = blkcg_policy[i];
256 if (blkg->pd[i] && pol->pd_online_fn)
257 pol->pd_online_fn(blkg);
261 spin_unlock(&blkcg->lock);
266 /* @blkg failed fully initialized, use the usual release path */
271 wb_congested_put(wb_congested);
273 css_put(&blkcg->css);
280 * blkg_lookup_create - lookup blkg, try to create one if not there
281 * @blkcg: blkcg of interest
282 * @q: request_queue of interest
284 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
285 * create one. blkg creation is performed recursively from blkcg_root such
286 * that all non-root blkg's have access to the parent blkg. This function
287 * should be called under RCU read lock and @q->queue_lock.
289 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
290 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
291 * dead and bypassing, returns ERR_PTR(-EBUSY).
293 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
294 struct request_queue *q)
296 struct blkcg_gq *blkg;
298 WARN_ON_ONCE(!rcu_read_lock_held());
299 lockdep_assert_held(q->queue_lock);
302 * This could be the first entry point of blkcg implementation and
303 * we shouldn't allow anything to go through for a bypassing queue.
305 if (unlikely(blk_queue_bypass(q)))
306 return ERR_PTR(blk_queue_dying(q) ? -EINVAL : -EBUSY);
308 blkg = __blkg_lookup(blkcg, q, true);
313 * Create blkgs walking down from blkcg_root to @blkcg, so that all
314 * non-root blkgs have access to their parents.
317 struct blkcg *pos = blkcg;
318 struct blkcg *parent = blkcg_parent(blkcg);
320 while (parent && !__blkg_lookup(parent, q, false)) {
322 parent = blkcg_parent(parent);
325 blkg = blkg_create(pos, q, NULL);
326 if (pos == blkcg || IS_ERR(blkg))
330 EXPORT_SYMBOL_GPL(blkg_lookup_create);
332 static void blkg_destroy(struct blkcg_gq *blkg)
334 struct blkcg *blkcg = blkg->blkcg;
337 lockdep_assert_held(blkg->q->queue_lock);
338 lockdep_assert_held(&blkcg->lock);
340 /* Something wrong if we are trying to remove same group twice */
341 WARN_ON_ONCE(list_empty(&blkg->q_node));
342 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
344 for (i = 0; i < BLKCG_MAX_POLS; i++) {
345 struct blkcg_policy *pol = blkcg_policy[i];
347 if (blkg->pd[i] && pol->pd_offline_fn)
348 pol->pd_offline_fn(blkg);
350 blkg->online = false;
352 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
353 list_del_init(&blkg->q_node);
354 hlist_del_init_rcu(&blkg->blkcg_node);
357 * Both setting lookup hint to and clearing it from @blkg are done
358 * under queue_lock. If it's not pointing to @blkg now, it never
359 * will. Hint assignment itself can race safely.
361 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
362 rcu_assign_pointer(blkcg->blkg_hint, NULL);
365 * Put the reference taken at the time of creation so that when all
366 * queues are gone, group can be destroyed.
372 * blkg_destroy_all - destroy all blkgs associated with a request_queue
373 * @q: request_queue of interest
375 * Destroy all blkgs associated with @q.
377 static void blkg_destroy_all(struct request_queue *q)
379 struct blkcg_gq *blkg, *n;
381 lockdep_assert_held(q->queue_lock);
383 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
384 struct blkcg *blkcg = blkg->blkcg;
386 spin_lock(&blkcg->lock);
388 spin_unlock(&blkcg->lock);
393 * A group is RCU protected, but having an rcu lock does not mean that one
394 * can access all the fields of blkg and assume these are valid. For
395 * example, don't try to follow throtl_data and request queue links.
397 * Having a reference to blkg under an rcu allows accesses to only values
398 * local to groups like group stats and group rate limits.
400 void __blkg_release_rcu(struct rcu_head *rcu_head)
402 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
405 /* tell policies that this one is being freed */
406 for (i = 0; i < BLKCG_MAX_POLS; i++) {
407 struct blkcg_policy *pol = blkcg_policy[i];
409 if (blkg->pd[i] && pol->pd_exit_fn)
410 pol->pd_exit_fn(blkg);
413 /* release the blkcg and parent blkg refs this blkg has been holding */
414 css_put(&blkg->blkcg->css);
416 blkg_put(blkg->parent);
418 wb_congested_put(blkg->wb_congested);
422 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
425 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
426 * because the root blkg uses @q->root_rl instead of its own rl.
428 struct request_list *__blk_queue_next_rl(struct request_list *rl,
429 struct request_queue *q)
431 struct list_head *ent;
432 struct blkcg_gq *blkg;
435 * Determine the current blkg list_head. The first entry is
436 * root_rl which is off @q->blkg_list and mapped to the head.
438 if (rl == &q->root_rl) {
440 /* There are no more block groups, hence no request lists */
444 blkg = container_of(rl, struct blkcg_gq, rl);
448 /* walk to the next list_head, skip root blkcg */
450 if (ent == &q->root_blkg->q_node)
452 if (ent == &q->blkg_list)
455 blkg = container_of(ent, struct blkcg_gq, q_node);
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 (i = 0; i < BLKCG_MAX_POLS; i++) {
476 struct blkcg_policy *pol = blkcg_policy[i];
478 if (blkcg_policy_enabled(blkg->q, pol) &&
479 pol->pd_reset_stats_fn)
480 pol->pd_reset_stats_fn(blkg);
484 spin_unlock_irq(&blkcg->lock);
485 mutex_unlock(&blkcg_pol_mutex);
489 static const char *blkg_dev_name(struct blkcg_gq *blkg)
491 /* some drivers (floppy) instantiate a queue w/o disk registered */
492 if (blkg->q->backing_dev_info.dev)
493 return dev_name(blkg->q->backing_dev_info.dev);
498 * blkcg_print_blkgs - helper for printing per-blkg data
499 * @sf: seq_file to print to
500 * @blkcg: blkcg of interest
501 * @prfill: fill function to print out a blkg
502 * @pol: policy in question
503 * @data: data to be passed to @prfill
504 * @show_total: to print out sum of prfill return values or not
506 * This function invokes @prfill on each blkg of @blkcg if pd for the
507 * policy specified by @pol exists. @prfill is invoked with @sf, the
508 * policy data and @data and the matching queue lock held. If @show_total
509 * is %true, the sum of the return values from @prfill is printed with
510 * "Total" label at the end.
512 * This is to be used to construct print functions for
513 * cftype->read_seq_string method.
515 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
516 u64 (*prfill)(struct seq_file *,
517 struct blkg_policy_data *, int),
518 const struct blkcg_policy *pol, int data,
521 struct blkcg_gq *blkg;
525 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
526 spin_lock_irq(blkg->q->queue_lock);
527 if (blkcg_policy_enabled(blkg->q, pol))
528 total += prfill(sf, blkg->pd[pol->plid], data);
529 spin_unlock_irq(blkg->q->queue_lock);
534 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
536 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
539 * __blkg_prfill_u64 - prfill helper for a single u64 value
540 * @sf: seq_file to print to
541 * @pd: policy private data of interest
544 * Print @v to @sf for the device assocaited with @pd.
546 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
548 const char *dname = blkg_dev_name(pd->blkg);
553 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
556 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
559 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
560 * @sf: seq_file to print to
561 * @pd: policy private data of interest
562 * @rwstat: rwstat to print
564 * Print @rwstat to @sf for the device assocaited with @pd.
566 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
567 const struct blkg_rwstat *rwstat)
569 static const char *rwstr[] = {
570 [BLKG_RWSTAT_READ] = "Read",
571 [BLKG_RWSTAT_WRITE] = "Write",
572 [BLKG_RWSTAT_SYNC] = "Sync",
573 [BLKG_RWSTAT_ASYNC] = "Async",
575 const char *dname = blkg_dev_name(pd->blkg);
582 for (i = 0; i < BLKG_RWSTAT_NR; i++)
583 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
584 (unsigned long long)rwstat->cnt[i]);
586 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
587 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
590 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
593 * blkg_prfill_stat - prfill callback for blkg_stat
594 * @sf: seq_file to print to
595 * @pd: policy private data of interest
596 * @off: offset to the blkg_stat in @pd
598 * prfill callback for printing a blkg_stat.
600 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
602 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
604 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
607 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
608 * @sf: seq_file to print to
609 * @pd: policy private data of interest
610 * @off: offset to the blkg_rwstat in @pd
612 * prfill callback for printing a blkg_rwstat.
614 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
617 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
619 return __blkg_prfill_rwstat(sf, pd, &rwstat);
621 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
624 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
625 * @pd: policy private data of interest
626 * @off: offset to the blkg_stat in @pd
628 * Collect the blkg_stat specified by @off from @pd and all its online
629 * descendants and return the sum. The caller must be holding the queue
630 * lock for online tests.
632 u64 blkg_stat_recursive_sum(struct blkg_policy_data *pd, int off)
634 struct blkcg_policy *pol = blkcg_policy[pd->plid];
635 struct blkcg_gq *pos_blkg;
636 struct cgroup_subsys_state *pos_css;
639 lockdep_assert_held(pd->blkg->q->queue_lock);
642 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
643 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
644 struct blkg_stat *stat = (void *)pos_pd + off;
646 if (pos_blkg->online)
647 sum += blkg_stat_read(stat);
653 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
656 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
657 * @pd: policy private data of interest
658 * @off: offset to the blkg_stat in @pd
660 * Collect the blkg_rwstat specified by @off from @pd and all its online
661 * descendants and return the sum. The caller must be holding the queue
662 * lock for online tests.
664 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkg_policy_data *pd,
667 struct blkcg_policy *pol = blkcg_policy[pd->plid];
668 struct blkcg_gq *pos_blkg;
669 struct cgroup_subsys_state *pos_css;
670 struct blkg_rwstat sum = { };
673 lockdep_assert_held(pd->blkg->q->queue_lock);
676 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
677 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
678 struct blkg_rwstat *rwstat = (void *)pos_pd + off;
679 struct blkg_rwstat tmp;
681 if (!pos_blkg->online)
684 tmp = blkg_rwstat_read(rwstat);
686 for (i = 0; i < BLKG_RWSTAT_NR; i++)
687 sum.cnt[i] += tmp.cnt[i];
693 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
696 * blkg_conf_prep - parse and prepare for per-blkg config update
697 * @blkcg: target block cgroup
698 * @pol: target policy
699 * @input: input string
700 * @ctx: blkg_conf_ctx to be filled
702 * Parse per-blkg config update from @input and initialize @ctx with the
703 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
704 * value. This function returns with RCU read lock and queue lock held and
705 * must be paired with blkg_conf_finish().
707 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
708 const char *input, struct blkg_conf_ctx *ctx)
709 __acquires(rcu) __acquires(disk->queue->queue_lock)
711 struct gendisk *disk;
712 struct blkcg_gq *blkg;
713 unsigned int major, minor;
714 unsigned long long v;
717 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
720 disk = get_gendisk(MKDEV(major, minor), &part);
725 spin_lock_irq(disk->queue->queue_lock);
727 if (blkcg_policy_enabled(disk->queue, pol))
728 blkg = blkg_lookup_create(blkcg, disk->queue);
730 blkg = ERR_PTR(-EINVAL);
735 spin_unlock_irq(disk->queue->queue_lock);
738 * If queue was bypassing, we should retry. Do so after a
739 * short msleep(). It isn't strictly necessary but queue
740 * can be bypassing for some time and it's always nice to
741 * avoid busy looping.
745 ret = restart_syscall();
755 EXPORT_SYMBOL_GPL(blkg_conf_prep);
758 * blkg_conf_finish - finish up per-blkg config update
759 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
761 * Finish up after per-blkg config update. This function must be paired
762 * with blkg_conf_prep().
764 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
765 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
767 spin_unlock_irq(ctx->disk->queue->queue_lock);
771 EXPORT_SYMBOL_GPL(blkg_conf_finish);
773 struct cftype blkcg_files[] = {
775 .name = "reset_stats",
776 .write_u64 = blkcg_reset_stats,
782 * blkcg_css_offline - cgroup css_offline callback
783 * @css: css of interest
785 * This function is called when @css is about to go away and responsible
786 * for shooting down all blkgs associated with @css. blkgs should be
787 * removed while holding both q and blkcg locks. As blkcg lock is nested
788 * inside q lock, this function performs reverse double lock dancing.
790 * This is the blkcg counterpart of ioc_release_fn().
792 static void blkcg_css_offline(struct cgroup_subsys_state *css)
794 struct blkcg *blkcg = css_to_blkcg(css);
796 spin_lock_irq(&blkcg->lock);
798 while (!hlist_empty(&blkcg->blkg_list)) {
799 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
800 struct blkcg_gq, blkcg_node);
801 struct request_queue *q = blkg->q;
803 if (spin_trylock(q->queue_lock)) {
805 spin_unlock(q->queue_lock);
807 spin_unlock_irq(&blkcg->lock);
809 spin_lock_irq(&blkcg->lock);
813 spin_unlock_irq(&blkcg->lock);
815 wb_blkcg_offline(blkcg);
818 static void blkcg_css_free(struct cgroup_subsys_state *css)
820 struct blkcg *blkcg = css_to_blkcg(css);
822 mutex_lock(&blkcg_pol_mutex);
823 list_del(&blkcg->all_blkcgs_node);
824 mutex_unlock(&blkcg_pol_mutex);
826 if (blkcg != &blkcg_root) {
829 for (i = 0; i < BLKCG_MAX_POLS; i++)
835 static struct cgroup_subsys_state *
836 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
839 struct cgroup_subsys_state *ret;
842 mutex_lock(&blkcg_pol_mutex);
849 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
851 ret = ERR_PTR(-ENOMEM);
855 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
856 struct blkcg_policy *pol = blkcg_policy[i];
857 struct blkcg_policy_data *cpd;
860 * If the policy hasn't been attached yet, wait for it
861 * to be attached before doing anything else. Otherwise,
862 * check if the policy requires any specific per-cgroup
863 * data: if it does, allocate and initialize it.
865 if (!pol || !pol->cpd_size)
868 BUG_ON(blkcg->pd[i]);
869 cpd = kzalloc(pol->cpd_size, GFP_KERNEL);
871 ret = ERR_PTR(-ENOMEM);
876 pol->cpd_init_fn(blkcg);
880 spin_lock_init(&blkcg->lock);
881 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
882 INIT_HLIST_HEAD(&blkcg->blkg_list);
883 #ifdef CONFIG_CGROUP_WRITEBACK
884 INIT_LIST_HEAD(&blkcg->cgwb_list);
886 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
888 mutex_unlock(&blkcg_pol_mutex);
892 for (i--; i >= 0; i--)
896 mutex_unlock(&blkcg_pol_mutex);
901 * blkcg_init_queue - initialize blkcg part of request queue
902 * @q: request_queue to initialize
904 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
905 * part of new request_queue @q.
908 * 0 on success, -errno on failure.
910 int blkcg_init_queue(struct request_queue *q)
912 struct blkcg_gq *new_blkg, *blkg;
916 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
920 preloaded = !radix_tree_preload(GFP_KERNEL);
923 * Make sure the root blkg exists and count the existing blkgs. As
924 * @q is bypassing at this point, blkg_lookup_create() can't be
925 * used. Open code insertion.
928 spin_lock_irq(q->queue_lock);
929 blkg = blkg_create(&blkcg_root, q, new_blkg);
930 spin_unlock_irq(q->queue_lock);
934 radix_tree_preload_end();
938 return PTR_ERR(blkg);
942 q->root_rl.blkg = blkg;
944 ret = blk_throtl_init(q);
946 spin_lock_irq(q->queue_lock);
948 spin_unlock_irq(q->queue_lock);
954 * blkcg_drain_queue - drain blkcg part of request_queue
955 * @q: request_queue to drain
957 * Called from blk_drain_queue(). Responsible for draining blkcg part.
959 void blkcg_drain_queue(struct request_queue *q)
961 lockdep_assert_held(q->queue_lock);
964 * @q could be exiting and already have destroyed all blkgs as
965 * indicated by NULL root_blkg. If so, don't confuse policies.
974 * blkcg_exit_queue - exit and release blkcg part of request_queue
975 * @q: request_queue being released
977 * Called from blk_release_queue(). Responsible for exiting blkcg part.
979 void blkcg_exit_queue(struct request_queue *q)
981 spin_lock_irq(q->queue_lock);
983 spin_unlock_irq(q->queue_lock);
989 * We cannot support shared io contexts, as we have no mean to support
990 * two tasks with the same ioc in two different groups without major rework
991 * of the main cic data structures. For now we allow a task to change
992 * its cgroup only if it's the only owner of its ioc.
994 static int blkcg_can_attach(struct cgroup_subsys_state *css,
995 struct cgroup_taskset *tset)
997 struct task_struct *task;
998 struct io_context *ioc;
1001 /* task_lock() is needed to avoid races with exit_io_context() */
1002 cgroup_taskset_for_each(task, tset) {
1004 ioc = task->io_context;
1005 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1014 struct cgroup_subsys blkio_cgrp_subsys = {
1015 .css_alloc = blkcg_css_alloc,
1016 .css_offline = blkcg_css_offline,
1017 .css_free = blkcg_css_free,
1018 .can_attach = blkcg_can_attach,
1019 .legacy_cftypes = blkcg_files,
1022 * This ensures that, if available, memcg is automatically enabled
1023 * together on the default hierarchy so that the owner cgroup can
1024 * be retrieved from writeback pages.
1026 .depends_on = 1 << memory_cgrp_id,
1029 EXPORT_SYMBOL_GPL(blkio_cgrp_subsys);
1032 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1033 * @q: request_queue of interest
1034 * @pol: blkcg policy to activate
1036 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1037 * bypass mode to populate its blkgs with policy_data for @pol.
1039 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1040 * from IO path. Update of each blkg is protected by both queue and blkcg
1041 * locks so that holding either lock and testing blkcg_policy_enabled() is
1042 * always enough for dereferencing policy data.
1044 * The caller is responsible for synchronizing [de]activations and policy
1045 * [un]registerations. Returns 0 on success, -errno on failure.
1047 int blkcg_activate_policy(struct request_queue *q,
1048 const struct blkcg_policy *pol)
1052 struct blkcg_gq *blkg;
1053 struct blkg_policy_data *pd, *nd;
1054 struct blkcg_policy_data *cpd, *cnd;
1057 if (blkcg_policy_enabled(q, pol))
1060 /* count and allocate policy_data for all existing blkgs */
1061 blk_queue_bypass_start(q);
1062 spin_lock_irq(q->queue_lock);
1063 list_for_each_entry(blkg, &q->blkg_list, q_node)
1065 spin_unlock_irq(q->queue_lock);
1068 * Allocate per-blkg and per-blkcg policy data
1069 * for all existing blkgs.
1072 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
1077 list_add_tail(&pd->alloc_node, &pds);
1081 cpd = kzalloc_node(pol->cpd_size, GFP_KERNEL, q->node);
1086 list_add_tail(&cpd->alloc_node, &cpds);
1090 * Install the allocated pds and cpds. With @q bypassing, no new blkg
1091 * should have been created while the queue lock was dropped.
1093 spin_lock_irq(q->queue_lock);
1095 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1096 if (WARN_ON(list_empty(&pds)) ||
1097 WARN_ON(pol->cpd_size && list_empty(&cpds))) {
1098 /* umm... this shouldn't happen, just abort */
1102 cpd = list_first_entry(&cpds, struct blkcg_policy_data,
1104 list_del_init(&cpd->alloc_node);
1105 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
1106 list_del_init(&pd->alloc_node);
1108 /* grab blkcg lock too while installing @pd on @blkg */
1109 spin_lock(&blkg->blkcg->lock);
1113 if (!blkg->blkcg->pd[pol->plid]) {
1114 /* Per-policy per-blkcg data */
1115 blkg->blkcg->pd[pol->plid] = cpd;
1116 cpd->plid = pol->plid;
1117 pol->cpd_init_fn(blkg->blkcg);
1118 } else { /* must free it as it has already been extracted */
1122 blkg->pd[pol->plid] = pd;
1124 pd->plid = pol->plid;
1125 pol->pd_init_fn(blkg);
1127 spin_unlock(&blkg->blkcg->lock);
1130 __set_bit(pol->plid, q->blkcg_pols);
1133 spin_unlock_irq(q->queue_lock);
1135 blk_queue_bypass_end(q);
1136 list_for_each_entry_safe(pd, nd, &pds, alloc_node)
1138 list_for_each_entry_safe(cpd, cnd, &cpds, alloc_node)
1142 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1145 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1146 * @q: request_queue of interest
1147 * @pol: blkcg policy to deactivate
1149 * Deactivate @pol on @q. Follows the same synchronization rules as
1150 * blkcg_activate_policy().
1152 void blkcg_deactivate_policy(struct request_queue *q,
1153 const struct blkcg_policy *pol)
1155 struct blkcg_gq *blkg;
1157 if (!blkcg_policy_enabled(q, pol))
1160 blk_queue_bypass_start(q);
1161 spin_lock_irq(q->queue_lock);
1163 __clear_bit(pol->plid, q->blkcg_pols);
1165 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1166 /* grab blkcg lock too while removing @pd from @blkg */
1167 spin_lock(&blkg->blkcg->lock);
1169 if (pol->pd_offline_fn)
1170 pol->pd_offline_fn(blkg);
1171 if (pol->pd_exit_fn)
1172 pol->pd_exit_fn(blkg);
1174 kfree(blkg->pd[pol->plid]);
1175 blkg->pd[pol->plid] = NULL;
1177 spin_unlock(&blkg->blkcg->lock);
1180 spin_unlock_irq(q->queue_lock);
1181 blk_queue_bypass_end(q);
1183 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1186 * blkcg_policy_register - register a blkcg policy
1187 * @pol: blkcg policy to register
1189 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1190 * successful registration. Returns 0 on success and -errno on failure.
1192 int blkcg_policy_register(struct blkcg_policy *pol)
1196 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
1199 mutex_lock(&blkcg_pol_register_mutex);
1200 mutex_lock(&blkcg_pol_mutex);
1202 /* find an empty slot */
1204 for (i = 0; i < BLKCG_MAX_POLS; i++)
1205 if (!blkcg_policy[i])
1207 if (i >= BLKCG_MAX_POLS)
1210 /* register and update blkgs */
1212 blkcg_policy[i] = pol;
1213 mutex_unlock(&blkcg_pol_mutex);
1215 /* everything is in place, add intf files for the new policy */
1217 WARN_ON(cgroup_add_legacy_cftypes(&blkio_cgrp_subsys,
1219 mutex_unlock(&blkcg_pol_register_mutex);
1223 mutex_unlock(&blkcg_pol_mutex);
1224 mutex_unlock(&blkcg_pol_register_mutex);
1227 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1230 * blkcg_policy_unregister - unregister a blkcg policy
1231 * @pol: blkcg policy to unregister
1233 * Undo blkcg_policy_register(@pol). Might sleep.
1235 void blkcg_policy_unregister(struct blkcg_policy *pol)
1237 mutex_lock(&blkcg_pol_register_mutex);
1239 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1242 /* kill the intf files first */
1244 cgroup_rm_cftypes(pol->cftypes);
1246 /* unregister and update blkgs */
1247 mutex_lock(&blkcg_pol_mutex);
1248 blkcg_policy[pol->plid] = NULL;
1249 mutex_unlock(&blkcg_pol_mutex);
1251 mutex_unlock(&blkcg_pol_register_mutex);
1253 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);