1 // SPDX-License-Identifier: GPL-2.0
3 * Common Block IO controller cgroup interface
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
18 #include <linux/ioprio.h>
19 #include <linux/kdev_t.h>
20 #include <linux/module.h>
21 #include <linux/sched/signal.h>
22 #include <linux/err.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/slab.h>
26 #include <linux/delay.h>
27 #include <linux/atomic.h>
28 #include <linux/ctype.h>
29 #include <linux/resume_user_mode.h>
30 #include <linux/psi.h>
31 #include <linux/part_stat.h>
33 #include "blk-cgroup.h"
34 #include "blk-ioprio.h"
35 #include "blk-throttle.h"
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
44 static DEFINE_MUTEX(blkcg_pol_register_mutex
);
45 static DEFINE_MUTEX(blkcg_pol_mutex
);
47 struct blkcg blkcg_root
;
48 EXPORT_SYMBOL_GPL(blkcg_root
);
50 struct cgroup_subsys_state
* const blkcg_root_css
= &blkcg_root
.css
;
51 EXPORT_SYMBOL_GPL(blkcg_root_css
);
53 static struct blkcg_policy
*blkcg_policy
[BLKCG_MAX_POLS
];
55 static LIST_HEAD(all_blkcgs
); /* protected by blkcg_pol_mutex */
57 bool blkcg_debug_stats
= false;
58 static struct workqueue_struct
*blkcg_punt_bio_wq
;
60 #define BLKG_DESTROY_BATCH_SIZE 64
63 * blkcg_css - find the current css
65 * Find the css associated with either the kthread or the current task.
66 * This may return a dying css, so it is up to the caller to use tryget logic
67 * to confirm it is alive and well.
69 static struct cgroup_subsys_state
*blkcg_css(void)
71 struct cgroup_subsys_state
*css
;
73 css
= kthread_blkcg();
76 return task_css(current
, io_cgrp_id
);
79 static bool blkcg_policy_enabled(struct request_queue
*q
,
80 const struct blkcg_policy
*pol
)
82 return pol
&& test_bit(pol
->plid
, q
->blkcg_pols
);
85 static void blkg_free_workfn(struct work_struct
*work
)
87 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
91 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
93 blkcg_policy
[i
]->pd_free_fn(blkg
->pd
[i
]);
96 blk_put_queue(blkg
->q
);
97 free_percpu(blkg
->iostat_cpu
);
98 percpu_ref_exit(&blkg
->refcnt
);
103 * blkg_free - free a blkg
104 * @blkg: blkg to free
106 * Free @blkg which may be partially allocated.
108 static void blkg_free(struct blkcg_gq
*blkg
)
114 * Both ->pd_free_fn() and request queue's release handler may
115 * sleep, so free us by scheduling one work func
117 INIT_WORK(&blkg
->free_work
, blkg_free_workfn
);
118 schedule_work(&blkg
->free_work
);
121 static void __blkg_release(struct rcu_head
*rcu
)
123 struct blkcg_gq
*blkg
= container_of(rcu
, struct blkcg_gq
, rcu_head
);
125 WARN_ON(!bio_list_empty(&blkg
->async_bios
));
127 /* release the blkcg and parent blkg refs this blkg has been holding */
128 css_put(&blkg
->blkcg
->css
);
130 blkg_put(blkg
->parent
);
135 * A group is RCU protected, but having an rcu lock does not mean that one
136 * can access all the fields of blkg and assume these are valid. For
137 * example, don't try to follow throtl_data and request queue links.
139 * Having a reference to blkg under an rcu allows accesses to only values
140 * local to groups like group stats and group rate limits.
142 static void blkg_release(struct percpu_ref
*ref
)
144 struct blkcg_gq
*blkg
= container_of(ref
, struct blkcg_gq
, refcnt
);
146 call_rcu(&blkg
->rcu_head
, __blkg_release
);
149 static void blkg_async_bio_workfn(struct work_struct
*work
)
151 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
153 struct bio_list bios
= BIO_EMPTY_LIST
;
155 struct blk_plug plug
;
156 bool need_plug
= false;
158 /* as long as there are pending bios, @blkg can't go away */
159 spin_lock_bh(&blkg
->async_bio_lock
);
160 bio_list_merge(&bios
, &blkg
->async_bios
);
161 bio_list_init(&blkg
->async_bios
);
162 spin_unlock_bh(&blkg
->async_bio_lock
);
164 /* start plug only when bio_list contains at least 2 bios */
165 if (bios
.head
&& bios
.head
->bi_next
) {
167 blk_start_plug(&plug
);
169 while ((bio
= bio_list_pop(&bios
)))
172 blk_finish_plug(&plug
);
176 * bio_blkcg_css - return the blkcg CSS associated with a bio
179 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
180 * associated. Callers are expected to either handle %NULL or know association
181 * has been done prior to calling this.
183 struct cgroup_subsys_state
*bio_blkcg_css(struct bio
*bio
)
185 if (!bio
|| !bio
->bi_blkg
)
187 return &bio
->bi_blkg
->blkcg
->css
;
189 EXPORT_SYMBOL_GPL(bio_blkcg_css
);
192 * blkcg_parent - get the parent of a blkcg
193 * @blkcg: blkcg of interest
195 * Return the parent blkcg of @blkcg. Can be called anytime.
197 static inline struct blkcg
*blkcg_parent(struct blkcg
*blkcg
)
199 return css_to_blkcg(blkcg
->css
.parent
);
203 * blkg_alloc - allocate a blkg
204 * @blkcg: block cgroup the new blkg is associated with
205 * @q: request_queue the new blkg is associated with
206 * @gfp_mask: allocation mask to use
208 * Allocate a new blkg assocating @blkcg and @q.
210 static struct blkcg_gq
*blkg_alloc(struct blkcg
*blkcg
, struct request_queue
*q
,
213 struct blkcg_gq
*blkg
;
216 /* alloc and init base part */
217 blkg
= kzalloc_node(sizeof(*blkg
), gfp_mask
, q
->node
);
221 if (percpu_ref_init(&blkg
->refcnt
, blkg_release
, 0, gfp_mask
))
224 blkg
->iostat_cpu
= alloc_percpu_gfp(struct blkg_iostat_set
, gfp_mask
);
225 if (!blkg
->iostat_cpu
)
228 if (!blk_get_queue(q
))
232 INIT_LIST_HEAD(&blkg
->q_node
);
233 spin_lock_init(&blkg
->async_bio_lock
);
234 bio_list_init(&blkg
->async_bios
);
235 INIT_WORK(&blkg
->async_bio_work
, blkg_async_bio_workfn
);
238 u64_stats_init(&blkg
->iostat
.sync
);
239 for_each_possible_cpu(cpu
)
240 u64_stats_init(&per_cpu_ptr(blkg
->iostat_cpu
, cpu
)->sync
);
242 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
243 struct blkcg_policy
*pol
= blkcg_policy
[i
];
244 struct blkg_policy_data
*pd
;
246 if (!blkcg_policy_enabled(q
, pol
))
249 /* alloc per-policy data and attach it to blkg */
250 pd
= pol
->pd_alloc_fn(gfp_mask
, q
, blkcg
);
266 struct blkcg_gq
*blkg_lookup_slowpath(struct blkcg
*blkcg
,
267 struct request_queue
*q
, bool update_hint
)
269 struct blkcg_gq
*blkg
;
272 * Hint didn't match. Look up from the radix tree. Note that the
273 * hint can only be updated under queue_lock as otherwise @blkg
274 * could have already been removed from blkg_tree. The caller is
275 * responsible for grabbing queue_lock if @update_hint.
277 blkg
= radix_tree_lookup(&blkcg
->blkg_tree
, q
->id
);
278 if (blkg
&& blkg
->q
== q
) {
280 lockdep_assert_held(&q
->queue_lock
);
281 rcu_assign_pointer(blkcg
->blkg_hint
, blkg
);
288 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath
);
291 * If @new_blkg is %NULL, this function tries to allocate a new one as
292 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
294 static struct blkcg_gq
*blkg_create(struct blkcg
*blkcg
,
295 struct request_queue
*q
,
296 struct blkcg_gq
*new_blkg
)
298 struct blkcg_gq
*blkg
;
301 lockdep_assert_held(&q
->queue_lock
);
303 /* request_queue is dying, do not create/recreate a blkg */
304 if (blk_queue_dying(q
)) {
309 /* blkg holds a reference to blkcg */
310 if (!css_tryget_online(&blkcg
->css
)) {
317 new_blkg
= blkg_alloc(blkcg
, q
, GFP_NOWAIT
| __GFP_NOWARN
);
318 if (unlikely(!new_blkg
)) {
326 if (blkcg_parent(blkcg
)) {
327 blkg
->parent
= __blkg_lookup(blkcg_parent(blkcg
), q
, false);
328 if (WARN_ON_ONCE(!blkg
->parent
)) {
332 blkg_get(blkg
->parent
);
335 /* invoke per-policy init */
336 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
337 struct blkcg_policy
*pol
= blkcg_policy
[i
];
339 if (blkg
->pd
[i
] && pol
->pd_init_fn
)
340 pol
->pd_init_fn(blkg
->pd
[i
]);
344 spin_lock(&blkcg
->lock
);
345 ret
= radix_tree_insert(&blkcg
->blkg_tree
, q
->id
, blkg
);
347 hlist_add_head_rcu(&blkg
->blkcg_node
, &blkcg
->blkg_list
);
348 list_add(&blkg
->q_node
, &q
->blkg_list
);
350 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
351 struct blkcg_policy
*pol
= blkcg_policy
[i
];
353 if (blkg
->pd
[i
] && pol
->pd_online_fn
)
354 pol
->pd_online_fn(blkg
->pd
[i
]);
358 spin_unlock(&blkcg
->lock
);
363 /* @blkg failed fully initialized, use the usual release path */
368 css_put(&blkcg
->css
);
375 * blkg_lookup_create - lookup blkg, try to create one if not there
376 * @blkcg: blkcg of interest
377 * @q: request_queue of interest
379 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
380 * create one. blkg creation is performed recursively from blkcg_root such
381 * that all non-root blkg's have access to the parent blkg. This function
382 * should be called under RCU read lock and takes @q->queue_lock.
384 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
387 static struct blkcg_gq
*blkg_lookup_create(struct blkcg
*blkcg
,
388 struct request_queue
*q
)
390 struct blkcg_gq
*blkg
;
393 WARN_ON_ONCE(!rcu_read_lock_held());
395 blkg
= blkg_lookup(blkcg
, q
);
399 spin_lock_irqsave(&q
->queue_lock
, flags
);
400 blkg
= __blkg_lookup(blkcg
, q
, true);
405 * Create blkgs walking down from blkcg_root to @blkcg, so that all
406 * non-root blkgs have access to their parents. Returns the closest
407 * blkg to the intended blkg should blkg_create() fail.
410 struct blkcg
*pos
= blkcg
;
411 struct blkcg
*parent
= blkcg_parent(blkcg
);
412 struct blkcg_gq
*ret_blkg
= q
->root_blkg
;
415 blkg
= __blkg_lookup(parent
, q
, false);
417 /* remember closest blkg */
422 parent
= blkcg_parent(parent
);
425 blkg
= blkg_create(pos
, q
, NULL
);
435 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
439 static void blkg_destroy(struct blkcg_gq
*blkg
)
441 struct blkcg
*blkcg
= blkg
->blkcg
;
444 lockdep_assert_held(&blkg
->q
->queue_lock
);
445 lockdep_assert_held(&blkcg
->lock
);
447 /* Something wrong if we are trying to remove same group twice */
448 WARN_ON_ONCE(list_empty(&blkg
->q_node
));
449 WARN_ON_ONCE(hlist_unhashed(&blkg
->blkcg_node
));
451 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
452 struct blkcg_policy
*pol
= blkcg_policy
[i
];
454 if (blkg
->pd
[i
] && pol
->pd_offline_fn
)
455 pol
->pd_offline_fn(blkg
->pd
[i
]);
458 blkg
->online
= false;
460 radix_tree_delete(&blkcg
->blkg_tree
, blkg
->q
->id
);
461 list_del_init(&blkg
->q_node
);
462 hlist_del_init_rcu(&blkg
->blkcg_node
);
465 * Both setting lookup hint to and clearing it from @blkg are done
466 * under queue_lock. If it's not pointing to @blkg now, it never
467 * will. Hint assignment itself can race safely.
469 if (rcu_access_pointer(blkcg
->blkg_hint
) == blkg
)
470 rcu_assign_pointer(blkcg
->blkg_hint
, NULL
);
473 * Put the reference taken at the time of creation so that when all
474 * queues are gone, group can be destroyed.
476 percpu_ref_kill(&blkg
->refcnt
);
480 * blkg_destroy_all - destroy all blkgs associated with a request_queue
481 * @q: request_queue of interest
483 * Destroy all blkgs associated with @q.
485 static void blkg_destroy_all(struct request_queue
*q
)
487 struct blkcg_gq
*blkg
, *n
;
488 int count
= BLKG_DESTROY_BATCH_SIZE
;
491 spin_lock_irq(&q
->queue_lock
);
492 list_for_each_entry_safe(blkg
, n
, &q
->blkg_list
, q_node
) {
493 struct blkcg
*blkcg
= blkg
->blkcg
;
495 spin_lock(&blkcg
->lock
);
497 spin_unlock(&blkcg
->lock
);
500 * in order to avoid holding the spin lock for too long, release
501 * it when a batch of blkgs are destroyed.
504 count
= BLKG_DESTROY_BATCH_SIZE
;
505 spin_unlock_irq(&q
->queue_lock
);
512 spin_unlock_irq(&q
->queue_lock
);
515 static int blkcg_reset_stats(struct cgroup_subsys_state
*css
,
516 struct cftype
*cftype
, u64 val
)
518 struct blkcg
*blkcg
= css_to_blkcg(css
);
519 struct blkcg_gq
*blkg
;
522 mutex_lock(&blkcg_pol_mutex
);
523 spin_lock_irq(&blkcg
->lock
);
526 * Note that stat reset is racy - it doesn't synchronize against
527 * stat updates. This is a debug feature which shouldn't exist
528 * anyway. If you get hit by a race, retry.
530 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
531 for_each_possible_cpu(cpu
) {
532 struct blkg_iostat_set
*bis
=
533 per_cpu_ptr(blkg
->iostat_cpu
, cpu
);
534 memset(bis
, 0, sizeof(*bis
));
536 memset(&blkg
->iostat
, 0, sizeof(blkg
->iostat
));
538 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
539 struct blkcg_policy
*pol
= blkcg_policy
[i
];
541 if (blkg
->pd
[i
] && pol
->pd_reset_stats_fn
)
542 pol
->pd_reset_stats_fn(blkg
->pd
[i
]);
546 spin_unlock_irq(&blkcg
->lock
);
547 mutex_unlock(&blkcg_pol_mutex
);
551 const char *blkg_dev_name(struct blkcg_gq
*blkg
)
553 if (!blkg
->q
->disk
|| !blkg
->q
->disk
->bdi
->dev
)
555 return bdi_dev_name(blkg
->q
->disk
->bdi
);
559 * blkcg_print_blkgs - helper for printing per-blkg data
560 * @sf: seq_file to print to
561 * @blkcg: blkcg of interest
562 * @prfill: fill function to print out a blkg
563 * @pol: policy in question
564 * @data: data to be passed to @prfill
565 * @show_total: to print out sum of prfill return values or not
567 * This function invokes @prfill on each blkg of @blkcg if pd for the
568 * policy specified by @pol exists. @prfill is invoked with @sf, the
569 * policy data and @data and the matching queue lock held. If @show_total
570 * is %true, the sum of the return values from @prfill is printed with
571 * "Total" label at the end.
573 * This is to be used to construct print functions for
574 * cftype->read_seq_string method.
576 void blkcg_print_blkgs(struct seq_file
*sf
, struct blkcg
*blkcg
,
577 u64 (*prfill
)(struct seq_file
*,
578 struct blkg_policy_data
*, int),
579 const struct blkcg_policy
*pol
, int data
,
582 struct blkcg_gq
*blkg
;
586 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
587 spin_lock_irq(&blkg
->q
->queue_lock
);
588 if (blkcg_policy_enabled(blkg
->q
, pol
))
589 total
+= prfill(sf
, blkg
->pd
[pol
->plid
], data
);
590 spin_unlock_irq(&blkg
->q
->queue_lock
);
595 seq_printf(sf
, "Total %llu\n", (unsigned long long)total
);
597 EXPORT_SYMBOL_GPL(blkcg_print_blkgs
);
600 * __blkg_prfill_u64 - prfill helper for a single u64 value
601 * @sf: seq_file to print to
602 * @pd: policy private data of interest
605 * Print @v to @sf for the device assocaited with @pd.
607 u64
__blkg_prfill_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
, u64 v
)
609 const char *dname
= blkg_dev_name(pd
->blkg
);
614 seq_printf(sf
, "%s %llu\n", dname
, (unsigned long long)v
);
617 EXPORT_SYMBOL_GPL(__blkg_prfill_u64
);
619 /* Performs queue bypass and policy enabled checks then looks up blkg. */
620 static struct blkcg_gq
*blkg_lookup_check(struct blkcg
*blkcg
,
621 const struct blkcg_policy
*pol
,
622 struct request_queue
*q
)
624 WARN_ON_ONCE(!rcu_read_lock_held());
625 lockdep_assert_held(&q
->queue_lock
);
627 if (!blkcg_policy_enabled(q
, pol
))
628 return ERR_PTR(-EOPNOTSUPP
);
629 return __blkg_lookup(blkcg
, q
, true /* update_hint */);
633 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
634 * @inputp: input string pointer
636 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
637 * from @input and get and return the matching bdev. *@inputp is
638 * updated to point past the device node prefix. Returns an ERR_PTR()
641 * Use this function iff blkg_conf_prep() can't be used for some reason.
643 struct block_device
*blkcg_conf_open_bdev(char **inputp
)
645 char *input
= *inputp
;
646 unsigned int major
, minor
;
647 struct block_device
*bdev
;
650 if (sscanf(input
, "%u:%u%n", &major
, &minor
, &key_len
) != 2)
651 return ERR_PTR(-EINVAL
);
654 if (!isspace(*input
))
655 return ERR_PTR(-EINVAL
);
656 input
= skip_spaces(input
);
658 bdev
= blkdev_get_no_open(MKDEV(major
, minor
));
660 return ERR_PTR(-ENODEV
);
661 if (bdev_is_partition(bdev
)) {
662 blkdev_put_no_open(bdev
);
663 return ERR_PTR(-ENODEV
);
671 * blkg_conf_prep - parse and prepare for per-blkg config update
672 * @blkcg: target block cgroup
673 * @pol: target policy
674 * @input: input string
675 * @ctx: blkg_conf_ctx to be filled
677 * Parse per-blkg config update from @input and initialize @ctx with the
678 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
679 * part of @input following MAJ:MIN. This function returns with RCU read
680 * lock and queue lock held and must be paired with blkg_conf_finish().
682 int blkg_conf_prep(struct blkcg
*blkcg
, const struct blkcg_policy
*pol
,
683 char *input
, struct blkg_conf_ctx
*ctx
)
684 __acquires(rcu
) __acquires(&bdev
->bd_queue
->queue_lock
)
686 struct block_device
*bdev
;
687 struct request_queue
*q
;
688 struct blkcg_gq
*blkg
;
691 bdev
= blkcg_conf_open_bdev(&input
);
693 return PTR_ERR(bdev
);
695 q
= bdev_get_queue(bdev
);
698 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
699 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
701 ret
= blk_queue_enter(q
, 0);
706 spin_lock_irq(&q
->queue_lock
);
708 blkg
= blkg_lookup_check(blkcg
, pol
, q
);
718 * Create blkgs walking down from blkcg_root to @blkcg, so that all
719 * non-root blkgs have access to their parents.
722 struct blkcg
*pos
= blkcg
;
723 struct blkcg
*parent
;
724 struct blkcg_gq
*new_blkg
;
726 parent
= blkcg_parent(blkcg
);
727 while (parent
&& !__blkg_lookup(parent
, q
, false)) {
729 parent
= blkcg_parent(parent
);
732 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
733 spin_unlock_irq(&q
->queue_lock
);
736 new_blkg
= blkg_alloc(pos
, q
, GFP_KERNEL
);
737 if (unlikely(!new_blkg
)) {
739 goto fail_exit_queue
;
742 if (radix_tree_preload(GFP_KERNEL
)) {
745 goto fail_exit_queue
;
749 spin_lock_irq(&q
->queue_lock
);
751 blkg
= blkg_lookup_check(pos
, pol
, q
);
761 blkg
= blkg_create(pos
, q
, new_blkg
);
768 radix_tree_preload_end();
781 radix_tree_preload_end();
783 spin_unlock_irq(&q
->queue_lock
);
788 blkdev_put_no_open(bdev
);
790 * If queue was bypassing, we should retry. Do so after a
791 * short msleep(). It isn't strictly necessary but queue
792 * can be bypassing for some time and it's always nice to
793 * avoid busy looping.
797 ret
= restart_syscall();
801 EXPORT_SYMBOL_GPL(blkg_conf_prep
);
804 * blkg_conf_finish - finish up per-blkg config update
805 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
807 * Finish up after per-blkg config update. This function must be paired
808 * with blkg_conf_prep().
810 void blkg_conf_finish(struct blkg_conf_ctx
*ctx
)
811 __releases(&ctx
->bdev
->bd_queue
->queue_lock
) __releases(rcu
)
813 spin_unlock_irq(&bdev_get_queue(ctx
->bdev
)->queue_lock
);
815 blkdev_put_no_open(ctx
->bdev
);
817 EXPORT_SYMBOL_GPL(blkg_conf_finish
);
819 static void blkg_iostat_set(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
823 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
824 dst
->bytes
[i
] = src
->bytes
[i
];
825 dst
->ios
[i
] = src
->ios
[i
];
829 static void blkg_iostat_add(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
833 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
834 dst
->bytes
[i
] += src
->bytes
[i
];
835 dst
->ios
[i
] += src
->ios
[i
];
839 static void blkg_iostat_sub(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
843 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
844 dst
->bytes
[i
] -= src
->bytes
[i
];
845 dst
->ios
[i
] -= src
->ios
[i
];
849 static void blkcg_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
851 struct blkcg
*blkcg
= css_to_blkcg(css
);
852 struct blkcg_gq
*blkg
;
854 /* Root-level stats are sourced from system-wide IO stats */
855 if (!cgroup_parent(css
->cgroup
))
860 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
861 struct blkcg_gq
*parent
= blkg
->parent
;
862 struct blkg_iostat_set
*bisc
= per_cpu_ptr(blkg
->iostat_cpu
, cpu
);
863 struct blkg_iostat cur
, delta
;
867 /* fetch the current per-cpu values */
869 seq
= u64_stats_fetch_begin(&bisc
->sync
);
870 blkg_iostat_set(&cur
, &bisc
->cur
);
871 } while (u64_stats_fetch_retry(&bisc
->sync
, seq
));
873 /* propagate percpu delta to global */
874 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
875 blkg_iostat_set(&delta
, &cur
);
876 blkg_iostat_sub(&delta
, &bisc
->last
);
877 blkg_iostat_add(&blkg
->iostat
.cur
, &delta
);
878 blkg_iostat_add(&bisc
->last
, &delta
);
879 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
881 /* propagate global delta to parent (unless that's root) */
882 if (parent
&& parent
->parent
) {
883 flags
= u64_stats_update_begin_irqsave(&parent
->iostat
.sync
);
884 blkg_iostat_set(&delta
, &blkg
->iostat
.cur
);
885 blkg_iostat_sub(&delta
, &blkg
->iostat
.last
);
886 blkg_iostat_add(&parent
->iostat
.cur
, &delta
);
887 blkg_iostat_add(&blkg
->iostat
.last
, &delta
);
888 u64_stats_update_end_irqrestore(&parent
->iostat
.sync
, flags
);
896 * We source root cgroup stats from the system-wide stats to avoid
897 * tracking the same information twice and incurring overhead when no
898 * cgroups are defined. For that reason, cgroup_rstat_flush in
899 * blkcg_print_stat does not actually fill out the iostat in the root
902 * However, we would like to re-use the printing code between the root and
903 * non-root cgroups to the extent possible. For that reason, we simulate
904 * flushing the root cgroup's stats by explicitly filling in the iostat
905 * with disk level statistics.
907 static void blkcg_fill_root_iostats(void)
909 struct class_dev_iter iter
;
912 class_dev_iter_init(&iter
, &block_class
, NULL
, &disk_type
);
913 while ((dev
= class_dev_iter_next(&iter
))) {
914 struct block_device
*bdev
= dev_to_bdev(dev
);
915 struct blkcg_gq
*blkg
=
916 blk_queue_root_blkg(bdev_get_queue(bdev
));
917 struct blkg_iostat tmp
;
921 memset(&tmp
, 0, sizeof(tmp
));
922 for_each_possible_cpu(cpu
) {
923 struct disk_stats
*cpu_dkstats
;
925 cpu_dkstats
= per_cpu_ptr(bdev
->bd_stats
, cpu
);
926 tmp
.ios
[BLKG_IOSTAT_READ
] +=
927 cpu_dkstats
->ios
[STAT_READ
];
928 tmp
.ios
[BLKG_IOSTAT_WRITE
] +=
929 cpu_dkstats
->ios
[STAT_WRITE
];
930 tmp
.ios
[BLKG_IOSTAT_DISCARD
] +=
931 cpu_dkstats
->ios
[STAT_DISCARD
];
932 // convert sectors to bytes
933 tmp
.bytes
[BLKG_IOSTAT_READ
] +=
934 cpu_dkstats
->sectors
[STAT_READ
] << 9;
935 tmp
.bytes
[BLKG_IOSTAT_WRITE
] +=
936 cpu_dkstats
->sectors
[STAT_WRITE
] << 9;
937 tmp
.bytes
[BLKG_IOSTAT_DISCARD
] +=
938 cpu_dkstats
->sectors
[STAT_DISCARD
] << 9;
941 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
942 blkg_iostat_set(&blkg
->iostat
.cur
, &tmp
);
943 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
947 static void blkcg_print_one_stat(struct blkcg_gq
*blkg
, struct seq_file
*s
)
949 struct blkg_iostat_set
*bis
= &blkg
->iostat
;
950 u64 rbytes
, wbytes
, rios
, wios
, dbytes
, dios
;
958 dname
= blkg_dev_name(blkg
);
962 seq_printf(s
, "%s ", dname
);
965 seq
= u64_stats_fetch_begin(&bis
->sync
);
967 rbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_READ
];
968 wbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_WRITE
];
969 dbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_DISCARD
];
970 rios
= bis
->cur
.ios
[BLKG_IOSTAT_READ
];
971 wios
= bis
->cur
.ios
[BLKG_IOSTAT_WRITE
];
972 dios
= bis
->cur
.ios
[BLKG_IOSTAT_DISCARD
];
973 } while (u64_stats_fetch_retry(&bis
->sync
, seq
));
975 if (rbytes
|| wbytes
|| rios
|| wios
) {
976 seq_printf(s
, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
977 rbytes
, wbytes
, rios
, wios
,
981 if (blkcg_debug_stats
&& atomic_read(&blkg
->use_delay
)) {
982 seq_printf(s
, " use_delay=%d delay_nsec=%llu",
983 atomic_read(&blkg
->use_delay
),
984 atomic64_read(&blkg
->delay_nsec
));
987 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
988 struct blkcg_policy
*pol
= blkcg_policy
[i
];
990 if (!blkg
->pd
[i
] || !pol
->pd_stat_fn
)
993 pol
->pd_stat_fn(blkg
->pd
[i
], s
);
999 static int blkcg_print_stat(struct seq_file
*sf
, void *v
)
1001 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1002 struct blkcg_gq
*blkg
;
1004 if (!seq_css(sf
)->parent
)
1005 blkcg_fill_root_iostats();
1007 cgroup_rstat_flush(blkcg
->css
.cgroup
);
1010 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
1011 spin_lock_irq(&blkg
->q
->queue_lock
);
1012 blkcg_print_one_stat(blkg
, sf
);
1013 spin_unlock_irq(&blkg
->q
->queue_lock
);
1019 static struct cftype blkcg_files
[] = {
1022 .seq_show
= blkcg_print_stat
,
1027 static struct cftype blkcg_legacy_files
[] = {
1029 .name
= "reset_stats",
1030 .write_u64
= blkcg_reset_stats
,
1035 #ifdef CONFIG_CGROUP_WRITEBACK
1036 struct list_head
*blkcg_get_cgwb_list(struct cgroup_subsys_state
*css
)
1038 return &css_to_blkcg(css
)->cgwb_list
;
1043 * blkcg destruction is a three-stage process.
1045 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1046 * which offlines writeback. Here we tie the next stage of blkg destruction
1047 * to the completion of writeback associated with the blkcg. This lets us
1048 * avoid punting potentially large amounts of outstanding writeback to root
1049 * while maintaining any ongoing policies. The next stage is triggered when
1050 * the nr_cgwbs count goes to zero.
1052 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1053 * and handles the destruction of blkgs. Here the css reference held by
1054 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1055 * This work may occur in cgwb_release_workfn() on the cgwb_release
1056 * workqueue. Any submitted ios that fail to get the blkg ref will be
1057 * punted to the root_blkg.
1059 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1060 * This finally frees the blkcg.
1064 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1065 * @blkcg: blkcg of interest
1067 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1068 * is nested inside q lock, this function performs reverse double lock dancing.
1069 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1070 * blkcg_css_free to eventually be called.
1072 * This is the blkcg counterpart of ioc_release_fn().
1074 static void blkcg_destroy_blkgs(struct blkcg
*blkcg
)
1078 spin_lock_irq(&blkcg
->lock
);
1080 while (!hlist_empty(&blkcg
->blkg_list
)) {
1081 struct blkcg_gq
*blkg
= hlist_entry(blkcg
->blkg_list
.first
,
1082 struct blkcg_gq
, blkcg_node
);
1083 struct request_queue
*q
= blkg
->q
;
1085 if (need_resched() || !spin_trylock(&q
->queue_lock
)) {
1087 * Given that the system can accumulate a huge number
1088 * of blkgs in pathological cases, check to see if we
1089 * need to rescheduling to avoid softlockup.
1091 spin_unlock_irq(&blkcg
->lock
);
1093 spin_lock_irq(&blkcg
->lock
);
1098 spin_unlock(&q
->queue_lock
);
1101 spin_unlock_irq(&blkcg
->lock
);
1105 * blkcg_pin_online - pin online state
1106 * @blkcg_css: blkcg of interest
1108 * While pinned, a blkcg is kept online. This is primarily used to
1109 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1110 * while an associated cgwb is still active.
1112 void blkcg_pin_online(struct cgroup_subsys_state
*blkcg_css
)
1114 refcount_inc(&css_to_blkcg(blkcg_css
)->online_pin
);
1118 * blkcg_unpin_online - unpin online state
1119 * @blkcg_css: blkcg of interest
1121 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1122 * that blkg doesn't go offline while an associated cgwb is still active.
1123 * When this count goes to zero, all active cgwbs have finished so the
1124 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1126 void blkcg_unpin_online(struct cgroup_subsys_state
*blkcg_css
)
1128 struct blkcg
*blkcg
= css_to_blkcg(blkcg_css
);
1131 if (!refcount_dec_and_test(&blkcg
->online_pin
))
1133 blkcg_destroy_blkgs(blkcg
);
1134 blkcg
= blkcg_parent(blkcg
);
1139 * blkcg_css_offline - cgroup css_offline callback
1140 * @css: css of interest
1142 * This function is called when @css is about to go away. Here the cgwbs are
1143 * offlined first and only once writeback associated with the blkcg has
1144 * finished do we start step 2 (see above).
1146 static void blkcg_css_offline(struct cgroup_subsys_state
*css
)
1148 /* this prevents anyone from attaching or migrating to this blkcg */
1149 wb_blkcg_offline(css
);
1151 /* put the base online pin allowing step 2 to be triggered */
1152 blkcg_unpin_online(css
);
1155 static void blkcg_css_free(struct cgroup_subsys_state
*css
)
1157 struct blkcg
*blkcg
= css_to_blkcg(css
);
1160 mutex_lock(&blkcg_pol_mutex
);
1162 list_del(&blkcg
->all_blkcgs_node
);
1164 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1166 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1168 mutex_unlock(&blkcg_pol_mutex
);
1173 static struct cgroup_subsys_state
*
1174 blkcg_css_alloc(struct cgroup_subsys_state
*parent_css
)
1176 struct blkcg
*blkcg
;
1177 struct cgroup_subsys_state
*ret
;
1180 mutex_lock(&blkcg_pol_mutex
);
1183 blkcg
= &blkcg_root
;
1185 blkcg
= kzalloc(sizeof(*blkcg
), GFP_KERNEL
);
1187 ret
= ERR_PTR(-ENOMEM
);
1192 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1193 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1194 struct blkcg_policy_data
*cpd
;
1197 * If the policy hasn't been attached yet, wait for it
1198 * to be attached before doing anything else. Otherwise,
1199 * check if the policy requires any specific per-cgroup
1200 * data: if it does, allocate and initialize it.
1202 if (!pol
|| !pol
->cpd_alloc_fn
)
1205 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1207 ret
= ERR_PTR(-ENOMEM
);
1210 blkcg
->cpd
[i
] = cpd
;
1213 if (pol
->cpd_init_fn
)
1214 pol
->cpd_init_fn(cpd
);
1217 spin_lock_init(&blkcg
->lock
);
1218 refcount_set(&blkcg
->online_pin
, 1);
1219 INIT_RADIX_TREE(&blkcg
->blkg_tree
, GFP_NOWAIT
| __GFP_NOWARN
);
1220 INIT_HLIST_HEAD(&blkcg
->blkg_list
);
1221 #ifdef CONFIG_CGROUP_WRITEBACK
1222 INIT_LIST_HEAD(&blkcg
->cgwb_list
);
1224 list_add_tail(&blkcg
->all_blkcgs_node
, &all_blkcgs
);
1226 mutex_unlock(&blkcg_pol_mutex
);
1230 for (i
--; i
>= 0; i
--)
1232 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1234 if (blkcg
!= &blkcg_root
)
1237 mutex_unlock(&blkcg_pol_mutex
);
1241 static int blkcg_css_online(struct cgroup_subsys_state
*css
)
1243 struct blkcg
*parent
= blkcg_parent(css_to_blkcg(css
));
1246 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1247 * don't go offline while cgwbs are still active on them. Pin the
1248 * parent so that offline always happens towards the root.
1251 blkcg_pin_online(css
);
1256 * blkcg_init_queue - initialize blkcg part of request queue
1257 * @q: request_queue to initialize
1259 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1260 * part of new request_queue @q.
1263 * 0 on success, -errno on failure.
1265 int blkcg_init_queue(struct request_queue
*q
)
1267 struct blkcg_gq
*new_blkg
, *blkg
;
1271 INIT_LIST_HEAD(&q
->blkg_list
);
1273 new_blkg
= blkg_alloc(&blkcg_root
, q
, GFP_KERNEL
);
1277 preloaded
= !radix_tree_preload(GFP_KERNEL
);
1279 /* Make sure the root blkg exists. */
1280 /* spin_lock_irq can serve as RCU read-side critical section. */
1281 spin_lock_irq(&q
->queue_lock
);
1282 blkg
= blkg_create(&blkcg_root
, q
, new_blkg
);
1285 q
->root_blkg
= blkg
;
1286 spin_unlock_irq(&q
->queue_lock
);
1289 radix_tree_preload_end();
1291 ret
= blk_ioprio_init(q
);
1293 goto err_destroy_all
;
1295 ret
= blk_throtl_init(q
);
1297 goto err_destroy_all
;
1299 ret
= blk_iolatency_init(q
);
1302 goto err_destroy_all
;
1308 blkg_destroy_all(q
);
1311 spin_unlock_irq(&q
->queue_lock
);
1313 radix_tree_preload_end();
1314 return PTR_ERR(blkg
);
1318 * blkcg_exit_queue - exit and release blkcg part of request_queue
1319 * @q: request_queue being released
1321 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1323 void blkcg_exit_queue(struct request_queue
*q
)
1325 blkg_destroy_all(q
);
1329 static void blkcg_bind(struct cgroup_subsys_state
*root_css
)
1333 mutex_lock(&blkcg_pol_mutex
);
1335 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1336 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1337 struct blkcg
*blkcg
;
1339 if (!pol
|| !pol
->cpd_bind_fn
)
1342 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
)
1343 if (blkcg
->cpd
[pol
->plid
])
1344 pol
->cpd_bind_fn(blkcg
->cpd
[pol
->plid
]);
1346 mutex_unlock(&blkcg_pol_mutex
);
1349 static void blkcg_exit(struct task_struct
*tsk
)
1351 if (tsk
->throttle_queue
)
1352 blk_put_queue(tsk
->throttle_queue
);
1353 tsk
->throttle_queue
= NULL
;
1356 struct cgroup_subsys io_cgrp_subsys
= {
1357 .css_alloc
= blkcg_css_alloc
,
1358 .css_online
= blkcg_css_online
,
1359 .css_offline
= blkcg_css_offline
,
1360 .css_free
= blkcg_css_free
,
1361 .css_rstat_flush
= blkcg_rstat_flush
,
1363 .dfl_cftypes
= blkcg_files
,
1364 .legacy_cftypes
= blkcg_legacy_files
,
1365 .legacy_name
= "blkio",
1369 * This ensures that, if available, memcg is automatically enabled
1370 * together on the default hierarchy so that the owner cgroup can
1371 * be retrieved from writeback pages.
1373 .depends_on
= 1 << memory_cgrp_id
,
1376 EXPORT_SYMBOL_GPL(io_cgrp_subsys
);
1379 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1380 * @q: request_queue of interest
1381 * @pol: blkcg policy to activate
1383 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1384 * bypass mode to populate its blkgs with policy_data for @pol.
1386 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1387 * from IO path. Update of each blkg is protected by both queue and blkcg
1388 * locks so that holding either lock and testing blkcg_policy_enabled() is
1389 * always enough for dereferencing policy data.
1391 * The caller is responsible for synchronizing [de]activations and policy
1392 * [un]registerations. Returns 0 on success, -errno on failure.
1394 int blkcg_activate_policy(struct request_queue
*q
,
1395 const struct blkcg_policy
*pol
)
1397 struct blkg_policy_data
*pd_prealloc
= NULL
;
1398 struct blkcg_gq
*blkg
, *pinned_blkg
= NULL
;
1401 if (blkcg_policy_enabled(q
, pol
))
1405 blk_mq_freeze_queue(q
);
1407 spin_lock_irq(&q
->queue_lock
);
1409 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1410 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
) {
1411 struct blkg_policy_data
*pd
;
1413 if (blkg
->pd
[pol
->plid
])
1416 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1417 if (blkg
== pinned_blkg
) {
1421 pd
= pol
->pd_alloc_fn(GFP_NOWAIT
| __GFP_NOWARN
, q
,
1427 * GFP_NOWAIT failed. Free the existing one and
1428 * prealloc for @blkg w/ GFP_KERNEL.
1431 blkg_put(pinned_blkg
);
1435 spin_unlock_irq(&q
->queue_lock
);
1438 pol
->pd_free_fn(pd_prealloc
);
1439 pd_prealloc
= pol
->pd_alloc_fn(GFP_KERNEL
, q
,
1447 blkg
->pd
[pol
->plid
] = pd
;
1449 pd
->plid
= pol
->plid
;
1452 /* all allocated, init in the same order */
1453 if (pol
->pd_init_fn
)
1454 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
)
1455 pol
->pd_init_fn(blkg
->pd
[pol
->plid
]);
1457 __set_bit(pol
->plid
, q
->blkcg_pols
);
1460 spin_unlock_irq(&q
->queue_lock
);
1463 blk_mq_unfreeze_queue(q
);
1465 blkg_put(pinned_blkg
);
1467 pol
->pd_free_fn(pd_prealloc
);
1471 /* alloc failed, nothing's initialized yet, free everything */
1472 spin_lock_irq(&q
->queue_lock
);
1473 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1474 struct blkcg
*blkcg
= blkg
->blkcg
;
1476 spin_lock(&blkcg
->lock
);
1477 if (blkg
->pd
[pol
->plid
]) {
1478 pol
->pd_free_fn(blkg
->pd
[pol
->plid
]);
1479 blkg
->pd
[pol
->plid
] = NULL
;
1481 spin_unlock(&blkcg
->lock
);
1483 spin_unlock_irq(&q
->queue_lock
);
1487 EXPORT_SYMBOL_GPL(blkcg_activate_policy
);
1490 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1491 * @q: request_queue of interest
1492 * @pol: blkcg policy to deactivate
1494 * Deactivate @pol on @q. Follows the same synchronization rules as
1495 * blkcg_activate_policy().
1497 void blkcg_deactivate_policy(struct request_queue
*q
,
1498 const struct blkcg_policy
*pol
)
1500 struct blkcg_gq
*blkg
;
1502 if (!blkcg_policy_enabled(q
, pol
))
1506 blk_mq_freeze_queue(q
);
1508 spin_lock_irq(&q
->queue_lock
);
1510 __clear_bit(pol
->plid
, q
->blkcg_pols
);
1512 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1513 struct blkcg
*blkcg
= blkg
->blkcg
;
1515 spin_lock(&blkcg
->lock
);
1516 if (blkg
->pd
[pol
->plid
]) {
1517 if (pol
->pd_offline_fn
)
1518 pol
->pd_offline_fn(blkg
->pd
[pol
->plid
]);
1519 pol
->pd_free_fn(blkg
->pd
[pol
->plid
]);
1520 blkg
->pd
[pol
->plid
] = NULL
;
1522 spin_unlock(&blkcg
->lock
);
1525 spin_unlock_irq(&q
->queue_lock
);
1528 blk_mq_unfreeze_queue(q
);
1530 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy
);
1533 * blkcg_policy_register - register a blkcg policy
1534 * @pol: blkcg policy to register
1536 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1537 * successful registration. Returns 0 on success and -errno on failure.
1539 int blkcg_policy_register(struct blkcg_policy
*pol
)
1541 struct blkcg
*blkcg
;
1544 mutex_lock(&blkcg_pol_register_mutex
);
1545 mutex_lock(&blkcg_pol_mutex
);
1547 /* find an empty slot */
1549 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1550 if (!blkcg_policy
[i
])
1552 if (i
>= BLKCG_MAX_POLS
) {
1553 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1557 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1558 if ((!pol
->cpd_alloc_fn
^ !pol
->cpd_free_fn
) ||
1559 (!pol
->pd_alloc_fn
^ !pol
->pd_free_fn
))
1564 blkcg_policy
[pol
->plid
] = pol
;
1566 /* allocate and install cpd's */
1567 if (pol
->cpd_alloc_fn
) {
1568 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1569 struct blkcg_policy_data
*cpd
;
1571 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1575 blkcg
->cpd
[pol
->plid
] = cpd
;
1577 cpd
->plid
= pol
->plid
;
1578 if (pol
->cpd_init_fn
)
1579 pol
->cpd_init_fn(cpd
);
1583 mutex_unlock(&blkcg_pol_mutex
);
1585 /* everything is in place, add intf files for the new policy */
1586 if (pol
->dfl_cftypes
)
1587 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys
,
1589 if (pol
->legacy_cftypes
)
1590 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys
,
1591 pol
->legacy_cftypes
));
1592 mutex_unlock(&blkcg_pol_register_mutex
);
1596 if (pol
->cpd_free_fn
) {
1597 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1598 if (blkcg
->cpd
[pol
->plid
]) {
1599 pol
->cpd_free_fn(blkcg
->cpd
[pol
->plid
]);
1600 blkcg
->cpd
[pol
->plid
] = NULL
;
1604 blkcg_policy
[pol
->plid
] = NULL
;
1606 mutex_unlock(&blkcg_pol_mutex
);
1607 mutex_unlock(&blkcg_pol_register_mutex
);
1610 EXPORT_SYMBOL_GPL(blkcg_policy_register
);
1613 * blkcg_policy_unregister - unregister a blkcg policy
1614 * @pol: blkcg policy to unregister
1616 * Undo blkcg_policy_register(@pol). Might sleep.
1618 void blkcg_policy_unregister(struct blkcg_policy
*pol
)
1620 struct blkcg
*blkcg
;
1622 mutex_lock(&blkcg_pol_register_mutex
);
1624 if (WARN_ON(blkcg_policy
[pol
->plid
] != pol
))
1627 /* kill the intf files first */
1628 if (pol
->dfl_cftypes
)
1629 cgroup_rm_cftypes(pol
->dfl_cftypes
);
1630 if (pol
->legacy_cftypes
)
1631 cgroup_rm_cftypes(pol
->legacy_cftypes
);
1633 /* remove cpds and unregister */
1634 mutex_lock(&blkcg_pol_mutex
);
1636 if (pol
->cpd_free_fn
) {
1637 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1638 if (blkcg
->cpd
[pol
->plid
]) {
1639 pol
->cpd_free_fn(blkcg
->cpd
[pol
->plid
]);
1640 blkcg
->cpd
[pol
->plid
] = NULL
;
1644 blkcg_policy
[pol
->plid
] = NULL
;
1646 mutex_unlock(&blkcg_pol_mutex
);
1648 mutex_unlock(&blkcg_pol_register_mutex
);
1650 EXPORT_SYMBOL_GPL(blkcg_policy_unregister
);
1652 bool __blkcg_punt_bio_submit(struct bio
*bio
)
1654 struct blkcg_gq
*blkg
= bio
->bi_blkg
;
1656 /* consume the flag first */
1657 bio
->bi_opf
&= ~REQ_CGROUP_PUNT
;
1659 /* never bounce for the root cgroup */
1663 spin_lock_bh(&blkg
->async_bio_lock
);
1664 bio_list_add(&blkg
->async_bios
, bio
);
1665 spin_unlock_bh(&blkg
->async_bio_lock
);
1667 queue_work(blkcg_punt_bio_wq
, &blkg
->async_bio_work
);
1672 * Scale the accumulated delay based on how long it has been since we updated
1673 * the delay. We only call this when we are adding delay, in case it's been a
1674 * while since we added delay, and when we are checking to see if we need to
1675 * delay a task, to account for any delays that may have occurred.
1677 static void blkcg_scale_delay(struct blkcg_gq
*blkg
, u64 now
)
1679 u64 old
= atomic64_read(&blkg
->delay_start
);
1681 /* negative use_delay means no scaling, see blkcg_set_delay() */
1682 if (atomic_read(&blkg
->use_delay
) < 0)
1686 * We only want to scale down every second. The idea here is that we
1687 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1688 * time window. We only want to throttle tasks for recent delay that
1689 * has occurred, in 1 second time windows since that's the maximum
1690 * things can be throttled. We save the current delay window in
1691 * blkg->last_delay so we know what amount is still left to be charged
1692 * to the blkg from this point onward. blkg->last_use keeps track of
1693 * the use_delay counter. The idea is if we're unthrottling the blkg we
1694 * are ok with whatever is happening now, and we can take away more of
1695 * the accumulated delay as we've already throttled enough that
1696 * everybody is happy with their IO latencies.
1698 if (time_before64(old
+ NSEC_PER_SEC
, now
) &&
1699 atomic64_cmpxchg(&blkg
->delay_start
, old
, now
) == old
) {
1700 u64 cur
= atomic64_read(&blkg
->delay_nsec
);
1701 u64 sub
= min_t(u64
, blkg
->last_delay
, now
- old
);
1702 int cur_use
= atomic_read(&blkg
->use_delay
);
1705 * We've been unthrottled, subtract a larger chunk of our
1706 * accumulated delay.
1708 if (cur_use
< blkg
->last_use
)
1709 sub
= max_t(u64
, sub
, blkg
->last_delay
>> 1);
1712 * This shouldn't happen, but handle it anyway. Our delay_nsec
1713 * should only ever be growing except here where we subtract out
1714 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1715 * rather not end up with negative numbers.
1717 if (unlikely(cur
< sub
)) {
1718 atomic64_set(&blkg
->delay_nsec
, 0);
1719 blkg
->last_delay
= 0;
1721 atomic64_sub(sub
, &blkg
->delay_nsec
);
1722 blkg
->last_delay
= cur
- sub
;
1724 blkg
->last_use
= cur_use
;
1729 * This is called when we want to actually walk up the hierarchy and check to
1730 * see if we need to throttle, and then actually throttle if there is some
1731 * accumulated delay. This should only be called upon return to user space so
1732 * we're not holding some lock that would induce a priority inversion.
1734 static void blkcg_maybe_throttle_blkg(struct blkcg_gq
*blkg
, bool use_memdelay
)
1736 unsigned long pflags
;
1738 u64 now
= ktime_to_ns(ktime_get());
1743 while (blkg
->parent
) {
1744 int use_delay
= atomic_read(&blkg
->use_delay
);
1749 blkcg_scale_delay(blkg
, now
);
1750 this_delay
= atomic64_read(&blkg
->delay_nsec
);
1751 if (this_delay
> delay_nsec
) {
1752 delay_nsec
= this_delay
;
1753 clamp
= use_delay
> 0;
1756 blkg
= blkg
->parent
;
1763 * Let's not sleep for all eternity if we've amassed a huge delay.
1764 * Swapping or metadata IO can accumulate 10's of seconds worth of
1765 * delay, and we want userspace to be able to do _something_ so cap the
1766 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1767 * tasks will be delayed for 0.25 second for every syscall. If
1768 * blkcg_set_delay() was used as indicated by negative use_delay, the
1769 * caller is responsible for regulating the range.
1772 delay_nsec
= min_t(u64
, delay_nsec
, 250 * NSEC_PER_MSEC
);
1775 psi_memstall_enter(&pflags
);
1777 exp
= ktime_add_ns(now
, delay_nsec
);
1778 tok
= io_schedule_prepare();
1780 __set_current_state(TASK_KILLABLE
);
1781 if (!schedule_hrtimeout(&exp
, HRTIMER_MODE_ABS
))
1783 } while (!fatal_signal_pending(current
));
1784 io_schedule_finish(tok
);
1787 psi_memstall_leave(&pflags
);
1791 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1793 * This is only called if we've been marked with set_notify_resume(). Obviously
1794 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1795 * check to see if current->throttle_queue is set and if not this doesn't do
1796 * anything. This should only ever be called by the resume code, it's not meant
1797 * to be called by people willy-nilly as it will actually do the work to
1798 * throttle the task if it is setup for throttling.
1800 void blkcg_maybe_throttle_current(void)
1802 struct request_queue
*q
= current
->throttle_queue
;
1803 struct blkcg
*blkcg
;
1804 struct blkcg_gq
*blkg
;
1805 bool use_memdelay
= current
->use_memdelay
;
1810 current
->throttle_queue
= NULL
;
1811 current
->use_memdelay
= false;
1814 blkcg
= css_to_blkcg(blkcg_css());
1817 blkg
= blkg_lookup(blkcg
, q
);
1820 if (!blkg_tryget(blkg
))
1824 blkcg_maybe_throttle_blkg(blkg
, use_memdelay
);
1834 * blkcg_schedule_throttle - this task needs to check for throttling
1835 * @q: the request queue IO was submitted on
1836 * @use_memdelay: do we charge this to memory delay for PSI
1838 * This is called by the IO controller when we know there's delay accumulated
1839 * for the blkg for this task. We do not pass the blkg because there are places
1840 * we call this that may not have that information, the swapping code for
1841 * instance will only have a request_queue at that point. This set's the
1842 * notify_resume for the task to check and see if it requires throttling before
1843 * returning to user space.
1845 * We will only schedule once per syscall. You can call this over and over
1846 * again and it will only do the check once upon return to user space, and only
1847 * throttle once. If the task needs to be throttled again it'll need to be
1848 * re-set at the next time we see the task.
1850 void blkcg_schedule_throttle(struct request_queue
*q
, bool use_memdelay
)
1852 if (unlikely(current
->flags
& PF_KTHREAD
))
1855 if (current
->throttle_queue
!= q
) {
1856 if (!blk_get_queue(q
))
1859 if (current
->throttle_queue
)
1860 blk_put_queue(current
->throttle_queue
);
1861 current
->throttle_queue
= q
;
1865 current
->use_memdelay
= use_memdelay
;
1866 set_notify_resume(current
);
1870 * blkcg_add_delay - add delay to this blkg
1871 * @blkg: blkg of interest
1872 * @now: the current time in nanoseconds
1873 * @delta: how many nanoseconds of delay to add
1875 * Charge @delta to the blkg's current delay accumulation. This is used to
1876 * throttle tasks if an IO controller thinks we need more throttling.
1878 void blkcg_add_delay(struct blkcg_gq
*blkg
, u64 now
, u64 delta
)
1880 if (WARN_ON_ONCE(atomic_read(&blkg
->use_delay
) < 0))
1882 blkcg_scale_delay(blkg
, now
);
1883 atomic64_add(delta
, &blkg
->delay_nsec
);
1887 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1891 * As the failure mode here is to walk up the blkg tree, this ensure that the
1892 * blkg->parent pointers are always valid. This returns the blkg that it ended
1893 * up taking a reference on or %NULL if no reference was taken.
1895 static inline struct blkcg_gq
*blkg_tryget_closest(struct bio
*bio
,
1896 struct cgroup_subsys_state
*css
)
1898 struct blkcg_gq
*blkg
, *ret_blkg
= NULL
;
1901 blkg
= blkg_lookup_create(css_to_blkcg(css
),
1902 bdev_get_queue(bio
->bi_bdev
));
1904 if (blkg_tryget(blkg
)) {
1908 blkg
= blkg
->parent
;
1916 * bio_associate_blkg_from_css - associate a bio with a specified css
1920 * Associate @bio with the blkg found by combining the css's blkg and the
1921 * request_queue of the @bio. An association failure is handled by walking up
1922 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1923 * and q->root_blkg. This situation only happens when a cgroup is dying and
1924 * then the remaining bios will spill to the closest alive blkg.
1926 * A reference will be taken on the blkg and will be released when @bio is
1929 void bio_associate_blkg_from_css(struct bio
*bio
,
1930 struct cgroup_subsys_state
*css
)
1933 blkg_put(bio
->bi_blkg
);
1935 if (css
&& css
->parent
) {
1936 bio
->bi_blkg
= blkg_tryget_closest(bio
, css
);
1938 blkg_get(bdev_get_queue(bio
->bi_bdev
)->root_blkg
);
1939 bio
->bi_blkg
= bdev_get_queue(bio
->bi_bdev
)->root_blkg
;
1942 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css
);
1945 * bio_associate_blkg - associate a bio with a blkg
1948 * Associate @bio with the blkg found from the bio's css and request_queue.
1949 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1950 * already associated, the css is reused and association redone as the
1951 * request_queue may have changed.
1953 void bio_associate_blkg(struct bio
*bio
)
1955 struct cgroup_subsys_state
*css
;
1960 css
= bio_blkcg_css(bio
);
1964 bio_associate_blkg_from_css(bio
, css
);
1968 EXPORT_SYMBOL_GPL(bio_associate_blkg
);
1971 * bio_clone_blkg_association - clone blkg association from src to dst bio
1972 * @dst: destination bio
1975 void bio_clone_blkg_association(struct bio
*dst
, struct bio
*src
)
1978 bio_associate_blkg_from_css(dst
, bio_blkcg_css(src
));
1980 EXPORT_SYMBOL_GPL(bio_clone_blkg_association
);
1982 static int blk_cgroup_io_type(struct bio
*bio
)
1984 if (op_is_discard(bio
->bi_opf
))
1985 return BLKG_IOSTAT_DISCARD
;
1986 if (op_is_write(bio
->bi_opf
))
1987 return BLKG_IOSTAT_WRITE
;
1988 return BLKG_IOSTAT_READ
;
1991 void blk_cgroup_bio_start(struct bio
*bio
)
1993 int rwd
= blk_cgroup_io_type(bio
), cpu
;
1994 struct blkg_iostat_set
*bis
;
1995 unsigned long flags
;
1998 bis
= per_cpu_ptr(bio
->bi_blkg
->iostat_cpu
, cpu
);
1999 flags
= u64_stats_update_begin_irqsave(&bis
->sync
);
2002 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2003 * bio and we would have already accounted for the size of the bio.
2005 if (!bio_flagged(bio
, BIO_CGROUP_ACCT
)) {
2006 bio_set_flag(bio
, BIO_CGROUP_ACCT
);
2007 bis
->cur
.bytes
[rwd
] += bio
->bi_iter
.bi_size
;
2009 bis
->cur
.ios
[rwd
]++;
2011 u64_stats_update_end_irqrestore(&bis
->sync
, flags
);
2012 if (cgroup_subsys_on_dfl(io_cgrp_subsys
))
2013 cgroup_rstat_updated(bio
->bi_blkg
->blkcg
->css
.cgroup
, cpu
);
2017 bool blk_cgroup_congested(void)
2019 struct cgroup_subsys_state
*css
;
2023 for (css
= blkcg_css(); css
; css
= css
->parent
) {
2024 if (atomic_read(&css
->cgroup
->congestion_count
)) {
2033 static int __init
blkcg_init(void)
2035 blkcg_punt_bio_wq
= alloc_workqueue("blkcg_punt_bio",
2036 WQ_MEM_RECLAIM
| WQ_FREEZABLE
|
2037 WQ_UNBOUND
| WQ_SYSFS
, 0);
2038 if (!blkcg_punt_bio_wq
)
2042 subsys_initcall(blkcg_init
);
2044 module_param(blkcg_debug_stats
, bool, 0644);
2045 MODULE_PARM_DESC(blkcg_debug_stats
, "True if you want debug stats, false if not");