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/slab.h>
23 #include <linux/genhd.h>
24 #include <linux/delay.h>
25 #include <linux/atomic.h>
26 #include "blk-cgroup.h"
29 #define MAX_KEY_LEN 100
31 static DEFINE_MUTEX(blkcg_pol_mutex);
33 struct blkcg blkcg_root;
34 EXPORT_SYMBOL_GPL(blkcg_root);
36 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
38 static bool blkcg_policy_enabled(struct request_queue *q,
39 const struct blkcg_policy *pol)
41 return pol && test_bit(pol->plid, q->blkcg_pols);
45 * blkg_free - free a blkg
48 * Free @blkg which may be partially allocated.
50 static void blkg_free(struct blkcg_gq *blkg)
57 for (i = 0; i < BLKCG_MAX_POLS; i++)
60 blk_exit_rl(&blkg->rl);
65 * blkg_alloc - allocate a blkg
66 * @blkcg: block cgroup the new blkg is associated with
67 * @q: request_queue the new blkg is associated with
68 * @gfp_mask: allocation mask to use
70 * Allocate a new blkg assocating @blkcg and @q.
72 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
75 struct blkcg_gq *blkg;
78 /* alloc and init base part */
79 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
84 INIT_LIST_HEAD(&blkg->q_node);
86 atomic_set(&blkg->refcnt, 1);
88 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
89 if (blkcg != &blkcg_root) {
90 if (blk_init_rl(&blkg->rl, q, gfp_mask))
95 for (i = 0; i < BLKCG_MAX_POLS; i++) {
96 struct blkcg_policy *pol = blkcg_policy[i];
97 struct blkg_policy_data *pd;
99 if (!blkcg_policy_enabled(q, pol))
102 /* alloc per-policy data and attach it to blkg */
103 pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
120 * __blkg_lookup - internal version of blkg_lookup()
121 * @blkcg: blkcg of interest
122 * @q: request_queue of interest
123 * @update_hint: whether to update lookup hint with the result or not
125 * This is internal version and shouldn't be used by policy
126 * implementations. Looks up blkgs for the @blkcg - @q pair regardless of
127 * @q's bypass state. If @update_hint is %true, the caller should be
128 * holding @q->queue_lock and lookup hint is updated on success.
130 struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg, struct request_queue *q,
133 struct blkcg_gq *blkg;
135 blkg = rcu_dereference(blkcg->blkg_hint);
136 if (blkg && blkg->q == q)
140 * Hint didn't match. Look up from the radix tree. Note that the
141 * hint can only be updated under queue_lock as otherwise @blkg
142 * could have already been removed from blkg_tree. The caller is
143 * responsible for grabbing queue_lock if @update_hint.
145 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
146 if (blkg && blkg->q == q) {
148 lockdep_assert_held(q->queue_lock);
149 rcu_assign_pointer(blkcg->blkg_hint, blkg);
158 * blkg_lookup - lookup blkg for the specified blkcg - q pair
159 * @blkcg: blkcg of interest
160 * @q: request_queue of interest
162 * Lookup blkg for the @blkcg - @q pair. This function should be called
163 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
164 * - see blk_queue_bypass_start() for details.
166 struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
168 WARN_ON_ONCE(!rcu_read_lock_held());
170 if (unlikely(blk_queue_bypass(q)))
172 return __blkg_lookup(blkcg, q, false);
174 EXPORT_SYMBOL_GPL(blkg_lookup);
177 * If @new_blkg is %NULL, this function tries to allocate a new one as
178 * necessary using %GFP_ATOMIC. @new_blkg is always consumed on return.
180 static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
181 struct request_queue *q,
182 struct blkcg_gq *new_blkg)
184 struct blkcg_gq *blkg;
187 WARN_ON_ONCE(!rcu_read_lock_held());
188 lockdep_assert_held(q->queue_lock);
190 /* blkg holds a reference to blkcg */
191 if (!css_tryget_online(&blkcg->css)) {
198 new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
199 if (unlikely(!new_blkg)) {
207 if (blkcg_parent(blkcg)) {
208 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
209 if (WARN_ON_ONCE(!blkg->parent)) {
213 blkg_get(blkg->parent);
216 /* invoke per-policy init */
217 for (i = 0; i < BLKCG_MAX_POLS; i++) {
218 struct blkcg_policy *pol = blkcg_policy[i];
220 if (blkg->pd[i] && pol->pd_init_fn)
221 pol->pd_init_fn(blkg);
225 spin_lock(&blkcg->lock);
226 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
228 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
229 list_add(&blkg->q_node, &q->blkg_list);
231 for (i = 0; i < BLKCG_MAX_POLS; i++) {
232 struct blkcg_policy *pol = blkcg_policy[i];
234 if (blkg->pd[i] && pol->pd_online_fn)
235 pol->pd_online_fn(blkg);
239 spin_unlock(&blkcg->lock);
242 if (blkcg == &blkcg_root) {
244 q->root_rl.blkg = blkg;
249 /* @blkg failed fully initialized, use the usual release path */
254 css_put(&blkcg->css);
261 * blkg_lookup_create - lookup blkg, try to create one if not there
262 * @blkcg: blkcg of interest
263 * @q: request_queue of interest
265 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
266 * create one. blkg creation is performed recursively from blkcg_root such
267 * that all non-root blkg's have access to the parent blkg. This function
268 * should be called under RCU read lock and @q->queue_lock.
270 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
271 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
272 * dead and bypassing, returns ERR_PTR(-EBUSY).
274 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
275 struct request_queue *q)
277 struct blkcg_gq *blkg;
279 WARN_ON_ONCE(!rcu_read_lock_held());
280 lockdep_assert_held(q->queue_lock);
283 * This could be the first entry point of blkcg implementation and
284 * we shouldn't allow anything to go through for a bypassing queue.
286 if (unlikely(blk_queue_bypass(q)))
287 return ERR_PTR(blk_queue_dying(q) ? -EINVAL : -EBUSY);
289 blkg = __blkg_lookup(blkcg, q, true);
294 * Create blkgs walking down from blkcg_root to @blkcg, so that all
295 * non-root blkgs have access to their parents.
298 struct blkcg *pos = blkcg;
299 struct blkcg *parent = blkcg_parent(blkcg);
301 while (parent && !__blkg_lookup(parent, q, false)) {
303 parent = blkcg_parent(parent);
306 blkg = blkg_create(pos, q, NULL);
307 if (pos == blkcg || IS_ERR(blkg))
311 EXPORT_SYMBOL_GPL(blkg_lookup_create);
313 static void blkg_destroy(struct blkcg_gq *blkg)
315 struct blkcg *blkcg = blkg->blkcg;
318 lockdep_assert_held(blkg->q->queue_lock);
319 lockdep_assert_held(&blkcg->lock);
321 /* Something wrong if we are trying to remove same group twice */
322 WARN_ON_ONCE(list_empty(&blkg->q_node));
323 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
325 for (i = 0; i < BLKCG_MAX_POLS; i++) {
326 struct blkcg_policy *pol = blkcg_policy[i];
328 if (blkg->pd[i] && pol->pd_offline_fn)
329 pol->pd_offline_fn(blkg);
331 blkg->online = false;
333 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
334 list_del_init(&blkg->q_node);
335 hlist_del_init_rcu(&blkg->blkcg_node);
338 * Both setting lookup hint to and clearing it from @blkg are done
339 * under queue_lock. If it's not pointing to @blkg now, it never
340 * will. Hint assignment itself can race safely.
342 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
343 rcu_assign_pointer(blkcg->blkg_hint, NULL);
346 * If root blkg is destroyed. Just clear the pointer since root_rl
347 * does not take reference on root blkg.
349 if (blkcg == &blkcg_root) {
350 blkg->q->root_blkg = NULL;
351 blkg->q->root_rl.blkg = NULL;
355 * Put the reference taken at the time of creation so that when all
356 * queues are gone, group can be destroyed.
362 * blkg_destroy_all - destroy all blkgs associated with a request_queue
363 * @q: request_queue of interest
365 * Destroy all blkgs associated with @q.
367 static void blkg_destroy_all(struct request_queue *q)
369 struct blkcg_gq *blkg, *n;
371 lockdep_assert_held(q->queue_lock);
373 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
374 struct blkcg *blkcg = blkg->blkcg;
376 spin_lock(&blkcg->lock);
378 spin_unlock(&blkcg->lock);
383 * A group is RCU protected, but having an rcu lock does not mean that one
384 * can access all the fields of blkg and assume these are valid. For
385 * example, don't try to follow throtl_data and request queue links.
387 * Having a reference to blkg under an rcu allows accesses to only values
388 * local to groups like group stats and group rate limits.
390 void __blkg_release_rcu(struct rcu_head *rcu_head)
392 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
395 /* tell policies that this one is being freed */
396 for (i = 0; i < BLKCG_MAX_POLS; i++) {
397 struct blkcg_policy *pol = blkcg_policy[i];
399 if (blkg->pd[i] && pol->pd_exit_fn)
400 pol->pd_exit_fn(blkg);
403 /* release the blkcg and parent blkg refs this blkg has been holding */
404 css_put(&blkg->blkcg->css);
406 blkg_put(blkg->parent);
410 EXPORT_SYMBOL_GPL(__blkg_release_rcu);
413 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
414 * because the root blkg uses @q->root_rl instead of its own rl.
416 struct request_list *__blk_queue_next_rl(struct request_list *rl,
417 struct request_queue *q)
419 struct list_head *ent;
420 struct blkcg_gq *blkg;
423 * Determine the current blkg list_head. The first entry is
424 * root_rl which is off @q->blkg_list and mapped to the head.
426 if (rl == &q->root_rl) {
428 /* There are no more block groups, hence no request lists */
432 blkg = container_of(rl, struct blkcg_gq, rl);
436 /* walk to the next list_head, skip root blkcg */
438 if (ent == &q->root_blkg->q_node)
440 if (ent == &q->blkg_list)
443 blkg = container_of(ent, struct blkcg_gq, q_node);
447 static int blkcg_reset_stats(struct cgroup_subsys_state *css,
448 struct cftype *cftype, u64 val)
450 struct blkcg *blkcg = css_to_blkcg(css);
451 struct blkcg_gq *blkg;
455 * XXX: We invoke cgroup_add/rm_cftypes() under blkcg_pol_mutex
456 * which ends up putting cgroup's internal cgroup_tree_mutex under
457 * it; however, cgroup_tree_mutex is nested above cgroup file
458 * active protection and grabbing blkcg_pol_mutex from a cgroup
459 * file operation creates a possible circular dependency. cgroup
460 * internal locking is planned to go through further simplification
461 * and this issue should go away soon. For now, let's trylock
462 * blkcg_pol_mutex and restart the write on failure.
464 * http://lkml.kernel.org/g/5363C04B.4010400@oracle.com
466 if (!mutex_trylock(&blkcg_pol_mutex))
467 return restart_syscall();
468 spin_lock_irq(&blkcg->lock);
471 * Note that stat reset is racy - it doesn't synchronize against
472 * stat updates. This is a debug feature which shouldn't exist
473 * anyway. If you get hit by a race, retry.
475 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
476 for (i = 0; i < BLKCG_MAX_POLS; i++) {
477 struct blkcg_policy *pol = blkcg_policy[i];
479 if (blkcg_policy_enabled(blkg->q, pol) &&
480 pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg);
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
490 static const char *blkg_dev_name(struct blkcg_gq *blkg)
492 /* some drivers (floppy) instantiate a queue w/o disk registered */
493 if (blkg->q->backing_dev_info.dev)
494 return dev_name(blkg->q->backing_dev_info.dev);
499 * blkcg_print_blkgs - helper for printing per-blkg data
500 * @sf: seq_file to print to
501 * @blkcg: blkcg of interest
502 * @prfill: fill function to print out a blkg
503 * @pol: policy in question
504 * @data: data to be passed to @prfill
505 * @show_total: to print out sum of prfill return values or not
507 * This function invokes @prfill on each blkg of @blkcg if pd for the
508 * policy specified by @pol exists. @prfill is invoked with @sf, the
509 * policy data and @data and the matching queue lock held. If @show_total
510 * is %true, the sum of the return values from @prfill is printed with
511 * "Total" label at the end.
513 * This is to be used to construct print functions for
514 * cftype->read_seq_string method.
516 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
517 u64 (*prfill)(struct seq_file *,
518 struct blkg_policy_data *, int),
519 const struct blkcg_policy *pol, int data,
522 struct blkcg_gq *blkg;
526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
527 spin_lock_irq(blkg->q->queue_lock);
528 if (blkcg_policy_enabled(blkg->q, pol))
529 total += prfill(sf, blkg->pd[pol->plid], data);
530 spin_unlock_irq(blkg->q->queue_lock);
535 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
537 EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
540 * __blkg_prfill_u64 - prfill helper for a single u64 value
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
545 * Print @v to @sf for the device assocaited with @pd.
547 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
549 const char *dname = blkg_dev_name(pd->blkg);
554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
557 EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
560 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
561 * @sf: seq_file to print to
562 * @pd: policy private data of interest
563 * @rwstat: rwstat to print
565 * Print @rwstat to @sf for the device assocaited with @pd.
567 u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
568 const struct blkg_rwstat *rwstat)
570 static const char *rwstr[] = {
571 [BLKG_RWSTAT_READ] = "Read",
572 [BLKG_RWSTAT_WRITE] = "Write",
573 [BLKG_RWSTAT_SYNC] = "Sync",
574 [BLKG_RWSTAT_ASYNC] = "Async",
576 const char *dname = blkg_dev_name(pd->blkg);
583 for (i = 0; i < BLKG_RWSTAT_NR; i++)
584 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
585 (unsigned long long)rwstat->cnt[i]);
587 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
588 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
591 EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
594 * blkg_prfill_stat - prfill callback for blkg_stat
595 * @sf: seq_file to print to
596 * @pd: policy private data of interest
597 * @off: offset to the blkg_stat in @pd
599 * prfill callback for printing a blkg_stat.
601 u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
603 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
605 EXPORT_SYMBOL_GPL(blkg_prfill_stat);
608 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
609 * @sf: seq_file to print to
610 * @pd: policy private data of interest
611 * @off: offset to the blkg_rwstat in @pd
613 * prfill callback for printing a blkg_rwstat.
615 u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
618 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
620 return __blkg_prfill_rwstat(sf, pd, &rwstat);
622 EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
625 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
626 * @pd: policy private data of interest
627 * @off: offset to the blkg_stat in @pd
629 * Collect the blkg_stat specified by @off from @pd and all its online
630 * descendants and return the sum. The caller must be holding the queue
631 * lock for online tests.
633 u64 blkg_stat_recursive_sum(struct blkg_policy_data *pd, int off)
635 struct blkcg_policy *pol = blkcg_policy[pd->plid];
636 struct blkcg_gq *pos_blkg;
637 struct cgroup_subsys_state *pos_css;
640 lockdep_assert_held(pd->blkg->q->queue_lock);
643 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
644 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
645 struct blkg_stat *stat = (void *)pos_pd + off;
647 if (pos_blkg->online)
648 sum += blkg_stat_read(stat);
654 EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
657 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
658 * @pd: policy private data of interest
659 * @off: offset to the blkg_stat in @pd
661 * Collect the blkg_rwstat specified by @off from @pd and all its online
662 * descendants and return the sum. The caller must be holding the queue
663 * lock for online tests.
665 struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkg_policy_data *pd,
668 struct blkcg_policy *pol = blkcg_policy[pd->plid];
669 struct blkcg_gq *pos_blkg;
670 struct cgroup_subsys_state *pos_css;
671 struct blkg_rwstat sum = { };
674 lockdep_assert_held(pd->blkg->q->queue_lock);
677 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
678 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
679 struct blkg_rwstat *rwstat = (void *)pos_pd + off;
680 struct blkg_rwstat tmp;
682 if (!pos_blkg->online)
685 tmp = blkg_rwstat_read(rwstat);
687 for (i = 0; i < BLKG_RWSTAT_NR; i++)
688 sum.cnt[i] += tmp.cnt[i];
694 EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
697 * blkg_conf_prep - parse and prepare for per-blkg config update
698 * @blkcg: target block cgroup
699 * @pol: target policy
700 * @input: input string
701 * @ctx: blkg_conf_ctx to be filled
703 * Parse per-blkg config update from @input and initialize @ctx with the
704 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
705 * value. This function returns with RCU read lock and queue lock held and
706 * must be paired with blkg_conf_finish().
708 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
709 const char *input, struct blkg_conf_ctx *ctx)
710 __acquires(rcu) __acquires(disk->queue->queue_lock)
712 struct gendisk *disk;
713 struct blkcg_gq *blkg;
714 unsigned int major, minor;
715 unsigned long long v;
718 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
721 disk = get_gendisk(MKDEV(major, minor), &part);
726 spin_lock_irq(disk->queue->queue_lock);
728 if (blkcg_policy_enabled(disk->queue, pol))
729 blkg = blkg_lookup_create(blkcg, disk->queue);
731 blkg = ERR_PTR(-EINVAL);
736 spin_unlock_irq(disk->queue->queue_lock);
739 * If queue was bypassing, we should retry. Do so after a
740 * short msleep(). It isn't strictly necessary but queue
741 * can be bypassing for some time and it's always nice to
742 * avoid busy looping.
746 ret = restart_syscall();
756 EXPORT_SYMBOL_GPL(blkg_conf_prep);
759 * blkg_conf_finish - finish up per-blkg config update
760 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
762 * Finish up after per-blkg config update. This function must be paired
763 * with blkg_conf_prep().
765 void blkg_conf_finish(struct blkg_conf_ctx *ctx)
766 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
768 spin_unlock_irq(ctx->disk->queue->queue_lock);
772 EXPORT_SYMBOL_GPL(blkg_conf_finish);
774 struct cftype blkcg_files[] = {
776 .name = "reset_stats",
777 .write_u64 = blkcg_reset_stats,
783 * blkcg_css_offline - cgroup css_offline callback
784 * @css: css of interest
786 * This function is called when @css is about to go away and responsible
787 * for shooting down all blkgs associated with @css. blkgs should be
788 * removed while holding both q and blkcg locks. As blkcg lock is nested
789 * inside q lock, this function performs reverse double lock dancing.
791 * This is the blkcg counterpart of ioc_release_fn().
793 static void blkcg_css_offline(struct cgroup_subsys_state *css)
795 struct blkcg *blkcg = css_to_blkcg(css);
797 spin_lock_irq(&blkcg->lock);
799 while (!hlist_empty(&blkcg->blkg_list)) {
800 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
801 struct blkcg_gq, blkcg_node);
802 struct request_queue *q = blkg->q;
804 if (spin_trylock(q->queue_lock)) {
806 spin_unlock(q->queue_lock);
808 spin_unlock_irq(&blkcg->lock);
810 spin_lock_irq(&blkcg->lock);
814 spin_unlock_irq(&blkcg->lock);
817 static void blkcg_css_free(struct cgroup_subsys_state *css)
819 struct blkcg *blkcg = css_to_blkcg(css);
821 if (blkcg != &blkcg_root)
825 static struct cgroup_subsys_state *
826 blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
829 struct cgroup_subsys_state *ret;
837 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
839 ret = ERR_PTR(-ENOMEM);
843 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
844 struct blkcg_policy *pol = blkcg_policy[i];
845 struct blkcg_policy_data *cpd;
848 * If the policy hasn't been attached yet, wait for it
849 * to be attached before doing anything else. Otherwise,
850 * check if the policy requires any specific per-cgroup
851 * data: if it does, allocate and initialize it.
853 if (!pol || !pol->cpd_size)
856 BUG_ON(blkcg->pd[i]);
857 cpd = kzalloc(pol->cpd_size, GFP_KERNEL);
859 ret = ERR_PTR(-ENOMEM);
864 pol->cpd_init_fn(blkcg);
868 spin_lock_init(&blkcg->lock);
869 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
870 INIT_HLIST_HEAD(&blkcg->blkg_list);
875 for (i--; i >= 0; i--)
884 * blkcg_init_queue - initialize blkcg part of request queue
885 * @q: request_queue to initialize
887 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
888 * part of new request_queue @q.
891 * 0 on success, -errno on failure.
893 int blkcg_init_queue(struct request_queue *q)
897 return blk_throtl_init(q);
901 * blkcg_drain_queue - drain blkcg part of request_queue
902 * @q: request_queue to drain
904 * Called from blk_drain_queue(). Responsible for draining blkcg part.
906 void blkcg_drain_queue(struct request_queue *q)
908 lockdep_assert_held(q->queue_lock);
911 * @q could be exiting and already have destroyed all blkgs as
912 * indicated by NULL root_blkg. If so, don't confuse policies.
921 * blkcg_exit_queue - exit and release blkcg part of request_queue
922 * @q: request_queue being released
924 * Called from blk_release_queue(). Responsible for exiting blkcg part.
926 void blkcg_exit_queue(struct request_queue *q)
928 spin_lock_irq(q->queue_lock);
930 spin_unlock_irq(q->queue_lock);
936 * We cannot support shared io contexts, as we have no mean to support
937 * two tasks with the same ioc in two different groups without major rework
938 * of the main cic data structures. For now we allow a task to change
939 * its cgroup only if it's the only owner of its ioc.
941 static int blkcg_can_attach(struct cgroup_subsys_state *css,
942 struct cgroup_taskset *tset)
944 struct task_struct *task;
945 struct io_context *ioc;
948 /* task_lock() is needed to avoid races with exit_io_context() */
949 cgroup_taskset_for_each(task, tset) {
951 ioc = task->io_context;
952 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
961 struct cgroup_subsys blkio_cgrp_subsys = {
962 .css_alloc = blkcg_css_alloc,
963 .css_offline = blkcg_css_offline,
964 .css_free = blkcg_css_free,
965 .can_attach = blkcg_can_attach,
966 .legacy_cftypes = blkcg_files,
969 * This ensures that, if available, memcg is automatically enabled
970 * together on the default hierarchy so that the owner cgroup can
971 * be retrieved from writeback pages.
973 .depends_on = 1 << memory_cgrp_id,
976 EXPORT_SYMBOL_GPL(blkio_cgrp_subsys);
979 * blkcg_activate_policy - activate a blkcg policy on a request_queue
980 * @q: request_queue of interest
981 * @pol: blkcg policy to activate
983 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
984 * bypass mode to populate its blkgs with policy_data for @pol.
986 * Activation happens with @q bypassed, so nobody would be accessing blkgs
987 * from IO path. Update of each blkg is protected by both queue and blkcg
988 * locks so that holding either lock and testing blkcg_policy_enabled() is
989 * always enough for dereferencing policy data.
991 * The caller is responsible for synchronizing [de]activations and policy
992 * [un]registerations. Returns 0 on success, -errno on failure.
994 int blkcg_activate_policy(struct request_queue *q,
995 const struct blkcg_policy *pol)
999 struct blkcg_gq *blkg, *new_blkg;
1000 struct blkg_policy_data *pd, *nd;
1001 struct blkcg_policy_data *cpd, *cnd;
1005 if (blkcg_policy_enabled(q, pol))
1008 /* preallocations for root blkg */
1009 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1013 blk_queue_bypass_start(q);
1015 preloaded = !radix_tree_preload(GFP_KERNEL);
1018 * Make sure the root blkg exists and count the existing blkgs. As
1019 * @q is bypassing at this point, blkg_lookup_create() can't be
1020 * used. Open code it.
1022 spin_lock_irq(q->queue_lock);
1025 blkg = __blkg_lookup(&blkcg_root, q, false);
1027 blkg_free(new_blkg);
1029 blkg = blkg_create(&blkcg_root, q, new_blkg);
1033 radix_tree_preload_end();
1036 ret = PTR_ERR(blkg);
1040 list_for_each_entry(blkg, &q->blkg_list, q_node)
1043 spin_unlock_irq(q->queue_lock);
1046 * Allocate per-blkg and per-blkcg policy data
1047 * for all existing blkgs.
1050 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
1055 list_add_tail(&pd->alloc_node, &pds);
1059 cpd = kzalloc_node(pol->cpd_size, GFP_KERNEL, q->node);
1064 list_add_tail(&cpd->alloc_node, &cpds);
1068 * Install the allocated pds and cpds. With @q bypassing, no new blkg
1069 * should have been created while the queue lock was dropped.
1071 spin_lock_irq(q->queue_lock);
1073 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1074 if (WARN_ON(list_empty(&pds)) ||
1075 WARN_ON(pol->cpd_size && list_empty(&cpds))) {
1076 /* umm... this shouldn't happen, just abort */
1080 cpd = list_first_entry(&cpds, struct blkcg_policy_data,
1082 list_del_init(&cpd->alloc_node);
1083 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
1084 list_del_init(&pd->alloc_node);
1086 /* grab blkcg lock too while installing @pd on @blkg */
1087 spin_lock(&blkg->blkcg->lock);
1091 if (!blkg->blkcg->pd[pol->plid]) {
1092 /* Per-policy per-blkcg data */
1093 blkg->blkcg->pd[pol->plid] = cpd;
1094 cpd->plid = pol->plid;
1095 pol->cpd_init_fn(blkg->blkcg);
1096 } else { /* must free it as it has already been extracted */
1100 blkg->pd[pol->plid] = pd;
1102 pd->plid = pol->plid;
1103 pol->pd_init_fn(blkg);
1105 spin_unlock(&blkg->blkcg->lock);
1108 __set_bit(pol->plid, q->blkcg_pols);
1111 spin_unlock_irq(q->queue_lock);
1113 blk_queue_bypass_end(q);
1114 list_for_each_entry_safe(pd, nd, &pds, alloc_node)
1116 list_for_each_entry_safe(cpd, cnd, &cpds, alloc_node)
1120 EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1123 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1124 * @q: request_queue of interest
1125 * @pol: blkcg policy to deactivate
1127 * Deactivate @pol on @q. Follows the same synchronization rules as
1128 * blkcg_activate_policy().
1130 void blkcg_deactivate_policy(struct request_queue *q,
1131 const struct blkcg_policy *pol)
1133 struct blkcg_gq *blkg;
1135 if (!blkcg_policy_enabled(q, pol))
1138 blk_queue_bypass_start(q);
1139 spin_lock_irq(q->queue_lock);
1141 __clear_bit(pol->plid, q->blkcg_pols);
1143 /* if no policy is left, no need for blkgs - shoot them down */
1144 if (bitmap_empty(q->blkcg_pols, BLKCG_MAX_POLS))
1145 blkg_destroy_all(q);
1147 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1148 /* grab blkcg lock too while removing @pd from @blkg */
1149 spin_lock(&blkg->blkcg->lock);
1151 if (pol->pd_offline_fn)
1152 pol->pd_offline_fn(blkg);
1153 if (pol->pd_exit_fn)
1154 pol->pd_exit_fn(blkg);
1156 kfree(blkg->pd[pol->plid]);
1157 blkg->pd[pol->plid] = NULL;
1158 kfree(blkg->blkcg->pd[pol->plid]);
1159 blkg->blkcg->pd[pol->plid] = NULL;
1161 spin_unlock(&blkg->blkcg->lock);
1164 spin_unlock_irq(q->queue_lock);
1165 blk_queue_bypass_end(q);
1167 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1170 * blkcg_policy_register - register a blkcg policy
1171 * @pol: blkcg policy to register
1173 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1174 * successful registration. Returns 0 on success and -errno on failure.
1176 int blkcg_policy_register(struct blkcg_policy *pol)
1180 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
1183 mutex_lock(&blkcg_pol_mutex);
1185 /* find an empty slot */
1187 for (i = 0; i < BLKCG_MAX_POLS; i++)
1188 if (!blkcg_policy[i])
1190 if (i >= BLKCG_MAX_POLS)
1193 /* register and update blkgs */
1195 blkcg_policy[i] = pol;
1197 /* everything is in place, add intf files for the new policy */
1199 WARN_ON(cgroup_add_legacy_cftypes(&blkio_cgrp_subsys,
1203 mutex_unlock(&blkcg_pol_mutex);
1206 EXPORT_SYMBOL_GPL(blkcg_policy_register);
1209 * blkcg_policy_unregister - unregister a blkcg policy
1210 * @pol: blkcg policy to unregister
1212 * Undo blkcg_policy_register(@pol). Might sleep.
1214 void blkcg_policy_unregister(struct blkcg_policy *pol)
1216 mutex_lock(&blkcg_pol_mutex);
1218 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1221 /* kill the intf files first */
1223 cgroup_rm_cftypes(pol->cftypes);
1225 /* unregister and update blkgs */
1226 blkcg_policy[pol->plid] = NULL;
1228 mutex_unlock(&blkcg_pol_mutex);
1230 EXPORT_SYMBOL_GPL(blkcg_policy_unregister);