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;
59 #define BLKG_DESTROY_BATCH_SIZE 64
62 * Lockless lists for tracking IO stats update
64 * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
65 * There are multiple blkg's (one for each block device) attached to each
66 * blkcg. The rstat code keeps track of which cpu has IO stats updated,
67 * but it doesn't know which blkg has the updated stats. If there are many
68 * block devices in a system, the cost of iterating all the blkg's to flush
69 * out the IO stats can be high. To reduce such overhead, a set of percpu
70 * lockless lists (lhead) per blkcg are used to track the set of recently
71 * updated iostat_cpu's since the last flush. An iostat_cpu will be put
72 * onto the lockless list on the update side [blk_cgroup_bio_start()] if
73 * not there yet and then removed when being flushed [blkcg_rstat_flush()].
74 * References to blkg are gotten and then put back in the process to
75 * protect against blkg removal.
77 * Return: 0 if successful or -ENOMEM if allocation fails.
79 static int init_blkcg_llists(struct blkcg
*blkcg
)
83 blkcg
->lhead
= alloc_percpu_gfp(struct llist_head
, GFP_KERNEL
);
87 for_each_possible_cpu(cpu
)
88 init_llist_head(per_cpu_ptr(blkcg
->lhead
, cpu
));
93 * blkcg_css - find the current css
95 * Find the css associated with either the kthread or the current task.
96 * This may return a dying css, so it is up to the caller to use tryget logic
97 * to confirm it is alive and well.
99 static struct cgroup_subsys_state
*blkcg_css(void)
101 struct cgroup_subsys_state
*css
;
103 css
= kthread_blkcg();
106 return task_css(current
, io_cgrp_id
);
109 static bool blkcg_policy_enabled(struct request_queue
*q
,
110 const struct blkcg_policy
*pol
)
112 return pol
&& test_bit(pol
->plid
, q
->blkcg_pols
);
115 static void blkg_free_workfn(struct work_struct
*work
)
117 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
119 struct request_queue
*q
= blkg
->q
;
123 * pd_free_fn() can also be called from blkcg_deactivate_policy(),
124 * in order to make sure pd_free_fn() is called in order, the deletion
125 * of the list blkg->q_node is delayed to here from blkg_destroy(), and
126 * blkcg_mutex is used to synchronize blkg_free_workfn() and
127 * blkcg_deactivate_policy().
129 mutex_lock(&q
->blkcg_mutex
);
130 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
132 blkcg_policy
[i
]->pd_free_fn(blkg
->pd
[i
]);
134 blkg_put(blkg
->parent
);
135 list_del_init(&blkg
->q_node
);
136 mutex_unlock(&q
->blkcg_mutex
);
139 free_percpu(blkg
->iostat_cpu
);
140 percpu_ref_exit(&blkg
->refcnt
);
145 * blkg_free - free a blkg
146 * @blkg: blkg to free
148 * Free @blkg which may be partially allocated.
150 static void blkg_free(struct blkcg_gq
*blkg
)
156 * Both ->pd_free_fn() and request queue's release handler may
157 * sleep, so free us by scheduling one work func
159 INIT_WORK(&blkg
->free_work
, blkg_free_workfn
);
160 schedule_work(&blkg
->free_work
);
163 static void __blkg_release(struct rcu_head
*rcu
)
165 struct blkcg_gq
*blkg
= container_of(rcu
, struct blkcg_gq
, rcu_head
);
167 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
168 WARN_ON(!bio_list_empty(&blkg
->async_bios
));
171 /* release the blkcg and parent blkg refs this blkg has been holding */
172 css_put(&blkg
->blkcg
->css
);
177 * A group is RCU protected, but having an rcu lock does not mean that one
178 * can access all the fields of blkg and assume these are valid. For
179 * example, don't try to follow throtl_data and request queue links.
181 * Having a reference to blkg under an rcu allows accesses to only values
182 * local to groups like group stats and group rate limits.
184 static void blkg_release(struct percpu_ref
*ref
)
186 struct blkcg_gq
*blkg
= container_of(ref
, struct blkcg_gq
, refcnt
);
188 call_rcu(&blkg
->rcu_head
, __blkg_release
);
191 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
192 static struct workqueue_struct
*blkcg_punt_bio_wq
;
194 static void blkg_async_bio_workfn(struct work_struct
*work
)
196 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
198 struct bio_list bios
= BIO_EMPTY_LIST
;
200 struct blk_plug plug
;
201 bool need_plug
= false;
203 /* as long as there are pending bios, @blkg can't go away */
204 spin_lock(&blkg
->async_bio_lock
);
205 bio_list_merge(&bios
, &blkg
->async_bios
);
206 bio_list_init(&blkg
->async_bios
);
207 spin_unlock(&blkg
->async_bio_lock
);
209 /* start plug only when bio_list contains at least 2 bios */
210 if (bios
.head
&& bios
.head
->bi_next
) {
212 blk_start_plug(&plug
);
214 while ((bio
= bio_list_pop(&bios
)))
217 blk_finish_plug(&plug
);
221 * When a shared kthread issues a bio for a cgroup, doing so synchronously can
222 * lead to priority inversions as the kthread can be trapped waiting for that
223 * cgroup. Use this helper instead of submit_bio to punt the actual issuing to
224 * a dedicated per-blkcg work item to avoid such priority inversions.
226 void blkcg_punt_bio_submit(struct bio
*bio
)
228 struct blkcg_gq
*blkg
= bio
->bi_blkg
;
231 spin_lock(&blkg
->async_bio_lock
);
232 bio_list_add(&blkg
->async_bios
, bio
);
233 spin_unlock(&blkg
->async_bio_lock
);
234 queue_work(blkcg_punt_bio_wq
, &blkg
->async_bio_work
);
236 /* never bounce for the root cgroup */
240 EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit
);
242 static int __init
blkcg_punt_bio_init(void)
244 blkcg_punt_bio_wq
= alloc_workqueue("blkcg_punt_bio",
245 WQ_MEM_RECLAIM
| WQ_FREEZABLE
|
246 WQ_UNBOUND
| WQ_SYSFS
, 0);
247 if (!blkcg_punt_bio_wq
)
251 subsys_initcall(blkcg_punt_bio_init
);
252 #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
255 * bio_blkcg_css - return the blkcg CSS associated with a bio
258 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
259 * associated. Callers are expected to either handle %NULL or know association
260 * has been done prior to calling this.
262 struct cgroup_subsys_state
*bio_blkcg_css(struct bio
*bio
)
264 if (!bio
|| !bio
->bi_blkg
)
266 return &bio
->bi_blkg
->blkcg
->css
;
268 EXPORT_SYMBOL_GPL(bio_blkcg_css
);
271 * blkcg_parent - get the parent of a blkcg
272 * @blkcg: blkcg of interest
274 * Return the parent blkcg of @blkcg. Can be called anytime.
276 static inline struct blkcg
*blkcg_parent(struct blkcg
*blkcg
)
278 return css_to_blkcg(blkcg
->css
.parent
);
282 * blkg_alloc - allocate a blkg
283 * @blkcg: block cgroup the new blkg is associated with
284 * @disk: gendisk the new blkg is associated with
285 * @gfp_mask: allocation mask to use
287 * Allocate a new blkg assocating @blkcg and @q.
289 static struct blkcg_gq
*blkg_alloc(struct blkcg
*blkcg
, struct gendisk
*disk
,
292 struct blkcg_gq
*blkg
;
295 /* alloc and init base part */
296 blkg
= kzalloc_node(sizeof(*blkg
), gfp_mask
, disk
->queue
->node
);
299 if (percpu_ref_init(&blkg
->refcnt
, blkg_release
, 0, gfp_mask
))
301 blkg
->iostat_cpu
= alloc_percpu_gfp(struct blkg_iostat_set
, gfp_mask
);
302 if (!blkg
->iostat_cpu
)
303 goto out_exit_refcnt
;
304 if (!blk_get_queue(disk
->queue
))
305 goto out_free_iostat
;
307 blkg
->q
= disk
->queue
;
308 INIT_LIST_HEAD(&blkg
->q_node
);
310 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
311 spin_lock_init(&blkg
->async_bio_lock
);
312 bio_list_init(&blkg
->async_bios
);
313 INIT_WORK(&blkg
->async_bio_work
, blkg_async_bio_workfn
);
316 u64_stats_init(&blkg
->iostat
.sync
);
317 for_each_possible_cpu(cpu
) {
318 u64_stats_init(&per_cpu_ptr(blkg
->iostat_cpu
, cpu
)->sync
);
319 per_cpu_ptr(blkg
->iostat_cpu
, cpu
)->blkg
= blkg
;
322 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
323 struct blkcg_policy
*pol
= blkcg_policy
[i
];
324 struct blkg_policy_data
*pd
;
326 if (!blkcg_policy_enabled(disk
->queue
, pol
))
329 /* alloc per-policy data and attach it to blkg */
330 pd
= pol
->pd_alloc_fn(disk
, blkcg
, gfp_mask
);
344 blkcg_policy
[i
]->pd_free_fn(blkg
->pd
[i
]);
345 blk_put_queue(disk
->queue
);
347 free_percpu(blkg
->iostat_cpu
);
349 percpu_ref_exit(&blkg
->refcnt
);
356 * If @new_blkg is %NULL, this function tries to allocate a new one as
357 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
359 static struct blkcg_gq
*blkg_create(struct blkcg
*blkcg
, struct gendisk
*disk
,
360 struct blkcg_gq
*new_blkg
)
362 struct blkcg_gq
*blkg
;
365 lockdep_assert_held(&disk
->queue
->queue_lock
);
367 /* request_queue is dying, do not create/recreate a blkg */
368 if (blk_queue_dying(disk
->queue
)) {
373 /* blkg holds a reference to blkcg */
374 if (!css_tryget_online(&blkcg
->css
)) {
381 new_blkg
= blkg_alloc(blkcg
, disk
, GFP_NOWAIT
| __GFP_NOWARN
);
382 if (unlikely(!new_blkg
)) {
390 if (blkcg_parent(blkcg
)) {
391 blkg
->parent
= blkg_lookup(blkcg_parent(blkcg
), disk
->queue
);
392 if (WARN_ON_ONCE(!blkg
->parent
)) {
396 blkg_get(blkg
->parent
);
399 /* invoke per-policy init */
400 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
401 struct blkcg_policy
*pol
= blkcg_policy
[i
];
403 if (blkg
->pd
[i
] && pol
->pd_init_fn
)
404 pol
->pd_init_fn(blkg
->pd
[i
]);
408 spin_lock(&blkcg
->lock
);
409 ret
= radix_tree_insert(&blkcg
->blkg_tree
, disk
->queue
->id
, blkg
);
411 hlist_add_head_rcu(&blkg
->blkcg_node
, &blkcg
->blkg_list
);
412 list_add(&blkg
->q_node
, &disk
->queue
->blkg_list
);
414 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
415 struct blkcg_policy
*pol
= blkcg_policy
[i
];
418 if (pol
->pd_online_fn
)
419 pol
->pd_online_fn(blkg
->pd
[i
]);
420 blkg
->pd
[i
]->online
= true;
425 spin_unlock(&blkcg
->lock
);
430 /* @blkg failed fully initialized, use the usual release path */
435 css_put(&blkcg
->css
);
443 * blkg_lookup_create - lookup blkg, try to create one if not there
444 * @blkcg: blkcg of interest
445 * @disk: gendisk of interest
447 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
448 * create one. blkg creation is performed recursively from blkcg_root such
449 * that all non-root blkg's have access to the parent blkg. This function
450 * should be called under RCU read lock and takes @disk->queue->queue_lock.
452 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
455 static struct blkcg_gq
*blkg_lookup_create(struct blkcg
*blkcg
,
456 struct gendisk
*disk
)
458 struct request_queue
*q
= disk
->queue
;
459 struct blkcg_gq
*blkg
;
462 WARN_ON_ONCE(!rcu_read_lock_held());
464 blkg
= blkg_lookup(blkcg
, q
);
468 spin_lock_irqsave(&q
->queue_lock
, flags
);
469 blkg
= blkg_lookup(blkcg
, q
);
471 if (blkcg
!= &blkcg_root
&&
472 blkg
!= rcu_dereference(blkcg
->blkg_hint
))
473 rcu_assign_pointer(blkcg
->blkg_hint
, blkg
);
478 * Create blkgs walking down from blkcg_root to @blkcg, so that all
479 * non-root blkgs have access to their parents. Returns the closest
480 * blkg to the intended blkg should blkg_create() fail.
483 struct blkcg
*pos
= blkcg
;
484 struct blkcg
*parent
= blkcg_parent(blkcg
);
485 struct blkcg_gq
*ret_blkg
= q
->root_blkg
;
488 blkg
= blkg_lookup(parent
, q
);
490 /* remember closest blkg */
495 parent
= blkcg_parent(parent
);
498 blkg
= blkg_create(pos
, disk
, NULL
);
508 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
512 static void blkg_destroy(struct blkcg_gq
*blkg
)
514 struct blkcg
*blkcg
= blkg
->blkcg
;
517 lockdep_assert_held(&blkg
->q
->queue_lock
);
518 lockdep_assert_held(&blkcg
->lock
);
521 * blkg stays on the queue list until blkg_free_workfn(), see details in
522 * blkg_free_workfn(), hence this function can be called from
523 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
524 * blkg_free_workfn().
526 if (hlist_unhashed(&blkg
->blkcg_node
))
529 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
530 struct blkcg_policy
*pol
= blkcg_policy
[i
];
532 if (blkg
->pd
[i
] && blkg
->pd
[i
]->online
) {
533 blkg
->pd
[i
]->online
= false;
534 if (pol
->pd_offline_fn
)
535 pol
->pd_offline_fn(blkg
->pd
[i
]);
539 blkg
->online
= false;
541 radix_tree_delete(&blkcg
->blkg_tree
, blkg
->q
->id
);
542 hlist_del_init_rcu(&blkg
->blkcg_node
);
545 * Both setting lookup hint to and clearing it from @blkg are done
546 * under queue_lock. If it's not pointing to @blkg now, it never
547 * will. Hint assignment itself can race safely.
549 if (rcu_access_pointer(blkcg
->blkg_hint
) == blkg
)
550 rcu_assign_pointer(blkcg
->blkg_hint
, NULL
);
553 * Put the reference taken at the time of creation so that when all
554 * queues are gone, group can be destroyed.
556 percpu_ref_kill(&blkg
->refcnt
);
559 static void blkg_destroy_all(struct gendisk
*disk
)
561 struct request_queue
*q
= disk
->queue
;
562 struct blkcg_gq
*blkg
, *n
;
563 int count
= BLKG_DESTROY_BATCH_SIZE
;
566 spin_lock_irq(&q
->queue_lock
);
567 list_for_each_entry_safe(blkg
, n
, &q
->blkg_list
, q_node
) {
568 struct blkcg
*blkcg
= blkg
->blkcg
;
570 if (hlist_unhashed(&blkg
->blkcg_node
))
573 spin_lock(&blkcg
->lock
);
575 spin_unlock(&blkcg
->lock
);
578 * in order to avoid holding the spin lock for too long, release
579 * it when a batch of blkgs are destroyed.
582 count
= BLKG_DESTROY_BATCH_SIZE
;
583 spin_unlock_irq(&q
->queue_lock
);
590 spin_unlock_irq(&q
->queue_lock
);
593 static int blkcg_reset_stats(struct cgroup_subsys_state
*css
,
594 struct cftype
*cftype
, u64 val
)
596 struct blkcg
*blkcg
= css_to_blkcg(css
);
597 struct blkcg_gq
*blkg
;
600 mutex_lock(&blkcg_pol_mutex
);
601 spin_lock_irq(&blkcg
->lock
);
604 * Note that stat reset is racy - it doesn't synchronize against
605 * stat updates. This is a debug feature which shouldn't exist
606 * anyway. If you get hit by a race, retry.
608 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
609 for_each_possible_cpu(cpu
) {
610 struct blkg_iostat_set
*bis
=
611 per_cpu_ptr(blkg
->iostat_cpu
, cpu
);
612 memset(bis
, 0, sizeof(*bis
));
614 memset(&blkg
->iostat
, 0, sizeof(blkg
->iostat
));
616 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
617 struct blkcg_policy
*pol
= blkcg_policy
[i
];
619 if (blkg
->pd
[i
] && pol
->pd_reset_stats_fn
)
620 pol
->pd_reset_stats_fn(blkg
->pd
[i
]);
624 spin_unlock_irq(&blkcg
->lock
);
625 mutex_unlock(&blkcg_pol_mutex
);
629 const char *blkg_dev_name(struct blkcg_gq
*blkg
)
633 return bdi_dev_name(blkg
->q
->disk
->bdi
);
637 * blkcg_print_blkgs - helper for printing per-blkg data
638 * @sf: seq_file to print to
639 * @blkcg: blkcg of interest
640 * @prfill: fill function to print out a blkg
641 * @pol: policy in question
642 * @data: data to be passed to @prfill
643 * @show_total: to print out sum of prfill return values or not
645 * This function invokes @prfill on each blkg of @blkcg if pd for the
646 * policy specified by @pol exists. @prfill is invoked with @sf, the
647 * policy data and @data and the matching queue lock held. If @show_total
648 * is %true, the sum of the return values from @prfill is printed with
649 * "Total" label at the end.
651 * This is to be used to construct print functions for
652 * cftype->read_seq_string method.
654 void blkcg_print_blkgs(struct seq_file
*sf
, struct blkcg
*blkcg
,
655 u64 (*prfill
)(struct seq_file
*,
656 struct blkg_policy_data
*, int),
657 const struct blkcg_policy
*pol
, int data
,
660 struct blkcg_gq
*blkg
;
664 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
665 spin_lock_irq(&blkg
->q
->queue_lock
);
666 if (blkcg_policy_enabled(blkg
->q
, pol
))
667 total
+= prfill(sf
, blkg
->pd
[pol
->plid
], data
);
668 spin_unlock_irq(&blkg
->q
->queue_lock
);
673 seq_printf(sf
, "Total %llu\n", (unsigned long long)total
);
675 EXPORT_SYMBOL_GPL(blkcg_print_blkgs
);
678 * __blkg_prfill_u64 - prfill helper for a single u64 value
679 * @sf: seq_file to print to
680 * @pd: policy private data of interest
683 * Print @v to @sf for the device associated with @pd.
685 u64
__blkg_prfill_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
, u64 v
)
687 const char *dname
= blkg_dev_name(pd
->blkg
);
692 seq_printf(sf
, "%s %llu\n", dname
, (unsigned long long)v
);
695 EXPORT_SYMBOL_GPL(__blkg_prfill_u64
);
698 * blkg_conf_init - initialize a blkg_conf_ctx
699 * @ctx: blkg_conf_ctx to initialize
700 * @input: input string
702 * Initialize @ctx which can be used to parse blkg config input string @input.
703 * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
704 * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
706 void blkg_conf_init(struct blkg_conf_ctx
*ctx
, char *input
)
708 *ctx
= (struct blkg_conf_ctx
){ .input
= input
};
710 EXPORT_SYMBOL_GPL(blkg_conf_init
);
713 * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
714 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
716 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
717 * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
718 * set to point past the device node prefix.
720 * This function may be called multiple times on @ctx and the extra calls become
721 * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
722 * explicitly if bdev access is needed without resolving the blkcg / policy part
723 * of @ctx->input. Returns -errno on error.
725 int blkg_conf_open_bdev(struct blkg_conf_ctx
*ctx
)
727 char *input
= ctx
->input
;
728 unsigned int major
, minor
;
729 struct block_device
*bdev
;
735 if (sscanf(input
, "%u:%u%n", &major
, &minor
, &key_len
) != 2)
739 if (!isspace(*input
))
741 input
= skip_spaces(input
);
743 bdev
= blkdev_get_no_open(MKDEV(major
, minor
));
746 if (bdev_is_partition(bdev
)) {
747 blkdev_put_no_open(bdev
);
757 * blkg_conf_prep - parse and prepare for per-blkg config update
758 * @blkcg: target block cgroup
759 * @pol: target policy
760 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
762 * Parse per-blkg config update from @ctx->input and initialize @ctx
763 * accordingly. On success, @ctx->body points to the part of @ctx->input
764 * following MAJ:MIN, @ctx->bdev points to the target block device and
765 * @ctx->blkg to the blkg being configured.
767 * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
768 * function returns with queue lock held and must be followed by
771 int blkg_conf_prep(struct blkcg
*blkcg
, const struct blkcg_policy
*pol
,
772 struct blkg_conf_ctx
*ctx
)
773 __acquires(&bdev
->bd_queue
->queue_lock
)
775 struct gendisk
*disk
;
776 struct request_queue
*q
;
777 struct blkcg_gq
*blkg
;
780 ret
= blkg_conf_open_bdev(ctx
);
784 disk
= ctx
->bdev
->bd_disk
;
788 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
789 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
791 ret
= blk_queue_enter(q
, 0);
795 spin_lock_irq(&q
->queue_lock
);
797 if (!blkcg_policy_enabled(q
, pol
)) {
802 blkg
= blkg_lookup(blkcg
, q
);
807 * Create blkgs walking down from blkcg_root to @blkcg, so that all
808 * non-root blkgs have access to their parents.
811 struct blkcg
*pos
= blkcg
;
812 struct blkcg
*parent
;
813 struct blkcg_gq
*new_blkg
;
815 parent
= blkcg_parent(blkcg
);
816 while (parent
&& !blkg_lookup(parent
, q
)) {
818 parent
= blkcg_parent(parent
);
821 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
822 spin_unlock_irq(&q
->queue_lock
);
824 new_blkg
= blkg_alloc(pos
, disk
, GFP_KERNEL
);
825 if (unlikely(!new_blkg
)) {
827 goto fail_exit_queue
;
830 if (radix_tree_preload(GFP_KERNEL
)) {
833 goto fail_exit_queue
;
836 spin_lock_irq(&q
->queue_lock
);
838 if (!blkcg_policy_enabled(q
, pol
)) {
844 blkg
= blkg_lookup(pos
, q
);
848 blkg
= blkg_create(pos
, disk
, new_blkg
);
855 radix_tree_preload_end();
866 radix_tree_preload_end();
868 spin_unlock_irq(&q
->queue_lock
);
873 * If queue was bypassing, we should retry. Do so after a
874 * short msleep(). It isn't strictly necessary but queue
875 * can be bypassing for some time and it's always nice to
876 * avoid busy looping.
880 ret
= restart_syscall();
884 EXPORT_SYMBOL_GPL(blkg_conf_prep
);
887 * blkg_conf_exit - clean up per-blkg config update
888 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
890 * Clean up after per-blkg config update. This function must be called on all
891 * blkg_conf_ctx's initialized with blkg_conf_init().
893 void blkg_conf_exit(struct blkg_conf_ctx
*ctx
)
894 __releases(&ctx
->bdev
->bd_queue
->queue_lock
)
897 spin_unlock_irq(&bdev_get_queue(ctx
->bdev
)->queue_lock
);
902 blkdev_put_no_open(ctx
->bdev
);
907 EXPORT_SYMBOL_GPL(blkg_conf_exit
);
909 static void blkg_iostat_set(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
913 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
914 dst
->bytes
[i
] = src
->bytes
[i
];
915 dst
->ios
[i
] = src
->ios
[i
];
919 static void blkg_iostat_add(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
923 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
924 dst
->bytes
[i
] += src
->bytes
[i
];
925 dst
->ios
[i
] += src
->ios
[i
];
929 static void blkg_iostat_sub(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
933 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
934 dst
->bytes
[i
] -= src
->bytes
[i
];
935 dst
->ios
[i
] -= src
->ios
[i
];
939 static void blkcg_iostat_update(struct blkcg_gq
*blkg
, struct blkg_iostat
*cur
,
940 struct blkg_iostat
*last
)
942 struct blkg_iostat delta
;
945 /* propagate percpu delta to global */
946 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
947 blkg_iostat_set(&delta
, cur
);
948 blkg_iostat_sub(&delta
, last
);
949 blkg_iostat_add(&blkg
->iostat
.cur
, &delta
);
950 blkg_iostat_add(last
, &delta
);
951 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
954 static void blkcg_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
956 struct blkcg
*blkcg
= css_to_blkcg(css
);
957 struct llist_head
*lhead
= per_cpu_ptr(blkcg
->lhead
, cpu
);
958 struct llist_node
*lnode
;
959 struct blkg_iostat_set
*bisc
, *next_bisc
;
961 /* Root-level stats are sourced from system-wide IO stats */
962 if (!cgroup_parent(css
->cgroup
))
967 lnode
= llist_del_all(lhead
);
972 * Iterate only the iostat_cpu's queued in the lockless list.
974 llist_for_each_entry_safe(bisc
, next_bisc
, lnode
, lnode
) {
975 struct blkcg_gq
*blkg
= bisc
->blkg
;
976 struct blkcg_gq
*parent
= blkg
->parent
;
977 struct blkg_iostat cur
;
980 WRITE_ONCE(bisc
->lqueued
, false);
982 /* fetch the current per-cpu values */
984 seq
= u64_stats_fetch_begin(&bisc
->sync
);
985 blkg_iostat_set(&cur
, &bisc
->cur
);
986 } while (u64_stats_fetch_retry(&bisc
->sync
, seq
));
988 blkcg_iostat_update(blkg
, &cur
, &bisc
->last
);
990 /* propagate global delta to parent (unless that's root) */
991 if (parent
&& parent
->parent
)
992 blkcg_iostat_update(parent
, &blkg
->iostat
.cur
,
994 percpu_ref_put(&blkg
->refcnt
);
1002 * We source root cgroup stats from the system-wide stats to avoid
1003 * tracking the same information twice and incurring overhead when no
1004 * cgroups are defined. For that reason, cgroup_rstat_flush in
1005 * blkcg_print_stat does not actually fill out the iostat in the root
1006 * cgroup's blkcg_gq.
1008 * However, we would like to re-use the printing code between the root and
1009 * non-root cgroups to the extent possible. For that reason, we simulate
1010 * flushing the root cgroup's stats by explicitly filling in the iostat
1011 * with disk level statistics.
1013 static void blkcg_fill_root_iostats(void)
1015 struct class_dev_iter iter
;
1018 class_dev_iter_init(&iter
, &block_class
, NULL
, &disk_type
);
1019 while ((dev
= class_dev_iter_next(&iter
))) {
1020 struct block_device
*bdev
= dev_to_bdev(dev
);
1021 struct blkcg_gq
*blkg
= bdev
->bd_disk
->queue
->root_blkg
;
1022 struct blkg_iostat tmp
;
1024 unsigned long flags
;
1026 memset(&tmp
, 0, sizeof(tmp
));
1027 for_each_possible_cpu(cpu
) {
1028 struct disk_stats
*cpu_dkstats
;
1030 cpu_dkstats
= per_cpu_ptr(bdev
->bd_stats
, cpu
);
1031 tmp
.ios
[BLKG_IOSTAT_READ
] +=
1032 cpu_dkstats
->ios
[STAT_READ
];
1033 tmp
.ios
[BLKG_IOSTAT_WRITE
] +=
1034 cpu_dkstats
->ios
[STAT_WRITE
];
1035 tmp
.ios
[BLKG_IOSTAT_DISCARD
] +=
1036 cpu_dkstats
->ios
[STAT_DISCARD
];
1037 // convert sectors to bytes
1038 tmp
.bytes
[BLKG_IOSTAT_READ
] +=
1039 cpu_dkstats
->sectors
[STAT_READ
] << 9;
1040 tmp
.bytes
[BLKG_IOSTAT_WRITE
] +=
1041 cpu_dkstats
->sectors
[STAT_WRITE
] << 9;
1042 tmp
.bytes
[BLKG_IOSTAT_DISCARD
] +=
1043 cpu_dkstats
->sectors
[STAT_DISCARD
] << 9;
1046 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
1047 blkg_iostat_set(&blkg
->iostat
.cur
, &tmp
);
1048 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
1052 static void blkcg_print_one_stat(struct blkcg_gq
*blkg
, struct seq_file
*s
)
1054 struct blkg_iostat_set
*bis
= &blkg
->iostat
;
1055 u64 rbytes
, wbytes
, rios
, wios
, dbytes
, dios
;
1063 dname
= blkg_dev_name(blkg
);
1067 seq_printf(s
, "%s ", dname
);
1070 seq
= u64_stats_fetch_begin(&bis
->sync
);
1072 rbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_READ
];
1073 wbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_WRITE
];
1074 dbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_DISCARD
];
1075 rios
= bis
->cur
.ios
[BLKG_IOSTAT_READ
];
1076 wios
= bis
->cur
.ios
[BLKG_IOSTAT_WRITE
];
1077 dios
= bis
->cur
.ios
[BLKG_IOSTAT_DISCARD
];
1078 } while (u64_stats_fetch_retry(&bis
->sync
, seq
));
1080 if (rbytes
|| wbytes
|| rios
|| wios
) {
1081 seq_printf(s
, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1082 rbytes
, wbytes
, rios
, wios
,
1086 if (blkcg_debug_stats
&& atomic_read(&blkg
->use_delay
)) {
1087 seq_printf(s
, " use_delay=%d delay_nsec=%llu",
1088 atomic_read(&blkg
->use_delay
),
1089 atomic64_read(&blkg
->delay_nsec
));
1092 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1093 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1095 if (!blkg
->pd
[i
] || !pol
->pd_stat_fn
)
1098 pol
->pd_stat_fn(blkg
->pd
[i
], s
);
1104 static int blkcg_print_stat(struct seq_file
*sf
, void *v
)
1106 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1107 struct blkcg_gq
*blkg
;
1109 if (!seq_css(sf
)->parent
)
1110 blkcg_fill_root_iostats();
1112 cgroup_rstat_flush(blkcg
->css
.cgroup
);
1115 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
1116 spin_lock_irq(&blkg
->q
->queue_lock
);
1117 blkcg_print_one_stat(blkg
, sf
);
1118 spin_unlock_irq(&blkg
->q
->queue_lock
);
1124 static struct cftype blkcg_files
[] = {
1127 .seq_show
= blkcg_print_stat
,
1132 static struct cftype blkcg_legacy_files
[] = {
1134 .name
= "reset_stats",
1135 .write_u64
= blkcg_reset_stats
,
1140 #ifdef CONFIG_CGROUP_WRITEBACK
1141 struct list_head
*blkcg_get_cgwb_list(struct cgroup_subsys_state
*css
)
1143 return &css_to_blkcg(css
)->cgwb_list
;
1148 * blkcg destruction is a three-stage process.
1150 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1151 * which offlines writeback. Here we tie the next stage of blkg destruction
1152 * to the completion of writeback associated with the blkcg. This lets us
1153 * avoid punting potentially large amounts of outstanding writeback to root
1154 * while maintaining any ongoing policies. The next stage is triggered when
1155 * the nr_cgwbs count goes to zero.
1157 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1158 * and handles the destruction of blkgs. Here the css reference held by
1159 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1160 * This work may occur in cgwb_release_workfn() on the cgwb_release
1161 * workqueue. Any submitted ios that fail to get the blkg ref will be
1162 * punted to the root_blkg.
1164 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1165 * This finally frees the blkcg.
1169 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1170 * @blkcg: blkcg of interest
1172 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1173 * is nested inside q lock, this function performs reverse double lock dancing.
1174 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1175 * blkcg_css_free to eventually be called.
1177 * This is the blkcg counterpart of ioc_release_fn().
1179 static void blkcg_destroy_blkgs(struct blkcg
*blkcg
)
1183 spin_lock_irq(&blkcg
->lock
);
1185 while (!hlist_empty(&blkcg
->blkg_list
)) {
1186 struct blkcg_gq
*blkg
= hlist_entry(blkcg
->blkg_list
.first
,
1187 struct blkcg_gq
, blkcg_node
);
1188 struct request_queue
*q
= blkg
->q
;
1190 if (need_resched() || !spin_trylock(&q
->queue_lock
)) {
1192 * Given that the system can accumulate a huge number
1193 * of blkgs in pathological cases, check to see if we
1194 * need to rescheduling to avoid softlockup.
1196 spin_unlock_irq(&blkcg
->lock
);
1198 spin_lock_irq(&blkcg
->lock
);
1203 spin_unlock(&q
->queue_lock
);
1206 spin_unlock_irq(&blkcg
->lock
);
1210 * blkcg_pin_online - pin online state
1211 * @blkcg_css: blkcg of interest
1213 * While pinned, a blkcg is kept online. This is primarily used to
1214 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1215 * while an associated cgwb is still active.
1217 void blkcg_pin_online(struct cgroup_subsys_state
*blkcg_css
)
1219 refcount_inc(&css_to_blkcg(blkcg_css
)->online_pin
);
1223 * blkcg_unpin_online - unpin online state
1224 * @blkcg_css: blkcg of interest
1226 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1227 * that blkg doesn't go offline while an associated cgwb is still active.
1228 * When this count goes to zero, all active cgwbs have finished so the
1229 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1231 void blkcg_unpin_online(struct cgroup_subsys_state
*blkcg_css
)
1233 struct blkcg
*blkcg
= css_to_blkcg(blkcg_css
);
1236 if (!refcount_dec_and_test(&blkcg
->online_pin
))
1238 blkcg_destroy_blkgs(blkcg
);
1239 blkcg
= blkcg_parent(blkcg
);
1244 * blkcg_css_offline - cgroup css_offline callback
1245 * @css: css of interest
1247 * This function is called when @css is about to go away. Here the cgwbs are
1248 * offlined first and only once writeback associated with the blkcg has
1249 * finished do we start step 2 (see above).
1251 static void blkcg_css_offline(struct cgroup_subsys_state
*css
)
1253 /* this prevents anyone from attaching or migrating to this blkcg */
1254 wb_blkcg_offline(css
);
1256 /* put the base online pin allowing step 2 to be triggered */
1257 blkcg_unpin_online(css
);
1260 static void blkcg_css_free(struct cgroup_subsys_state
*css
)
1262 struct blkcg
*blkcg
= css_to_blkcg(css
);
1265 mutex_lock(&blkcg_pol_mutex
);
1267 list_del(&blkcg
->all_blkcgs_node
);
1269 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1271 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1273 mutex_unlock(&blkcg_pol_mutex
);
1275 free_percpu(blkcg
->lhead
);
1279 static struct cgroup_subsys_state
*
1280 blkcg_css_alloc(struct cgroup_subsys_state
*parent_css
)
1282 struct blkcg
*blkcg
;
1285 mutex_lock(&blkcg_pol_mutex
);
1288 blkcg
= &blkcg_root
;
1290 blkcg
= kzalloc(sizeof(*blkcg
), GFP_KERNEL
);
1295 if (init_blkcg_llists(blkcg
))
1298 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1299 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1300 struct blkcg_policy_data
*cpd
;
1303 * If the policy hasn't been attached yet, wait for it
1304 * to be attached before doing anything else. Otherwise,
1305 * check if the policy requires any specific per-cgroup
1306 * data: if it does, allocate and initialize it.
1308 if (!pol
|| !pol
->cpd_alloc_fn
)
1311 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1315 blkcg
->cpd
[i
] = cpd
;
1320 spin_lock_init(&blkcg
->lock
);
1321 refcount_set(&blkcg
->online_pin
, 1);
1322 INIT_RADIX_TREE(&blkcg
->blkg_tree
, GFP_NOWAIT
| __GFP_NOWARN
);
1323 INIT_HLIST_HEAD(&blkcg
->blkg_list
);
1324 #ifdef CONFIG_CGROUP_WRITEBACK
1325 INIT_LIST_HEAD(&blkcg
->cgwb_list
);
1327 list_add_tail(&blkcg
->all_blkcgs_node
, &all_blkcgs
);
1329 mutex_unlock(&blkcg_pol_mutex
);
1333 for (i
--; i
>= 0; i
--)
1335 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1336 free_percpu(blkcg
->lhead
);
1338 if (blkcg
!= &blkcg_root
)
1341 mutex_unlock(&blkcg_pol_mutex
);
1342 return ERR_PTR(-ENOMEM
);
1345 static int blkcg_css_online(struct cgroup_subsys_state
*css
)
1347 struct blkcg
*parent
= blkcg_parent(css_to_blkcg(css
));
1350 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1351 * don't go offline while cgwbs are still active on them. Pin the
1352 * parent so that offline always happens towards the root.
1355 blkcg_pin_online(&parent
->css
);
1359 int blkcg_init_disk(struct gendisk
*disk
)
1361 struct request_queue
*q
= disk
->queue
;
1362 struct blkcg_gq
*new_blkg
, *blkg
;
1366 INIT_LIST_HEAD(&q
->blkg_list
);
1367 mutex_init(&q
->blkcg_mutex
);
1369 new_blkg
= blkg_alloc(&blkcg_root
, disk
, GFP_KERNEL
);
1373 preloaded
= !radix_tree_preload(GFP_KERNEL
);
1375 /* Make sure the root blkg exists. */
1376 /* spin_lock_irq can serve as RCU read-side critical section. */
1377 spin_lock_irq(&q
->queue_lock
);
1378 blkg
= blkg_create(&blkcg_root
, disk
, new_blkg
);
1381 q
->root_blkg
= blkg
;
1382 spin_unlock_irq(&q
->queue_lock
);
1385 radix_tree_preload_end();
1387 ret
= blk_ioprio_init(disk
);
1389 goto err_destroy_all
;
1391 ret
= blk_throtl_init(disk
);
1393 goto err_ioprio_exit
;
1398 blk_ioprio_exit(disk
);
1400 blkg_destroy_all(disk
);
1403 spin_unlock_irq(&q
->queue_lock
);
1405 radix_tree_preload_end();
1406 return PTR_ERR(blkg
);
1409 void blkcg_exit_disk(struct gendisk
*disk
)
1411 blkg_destroy_all(disk
);
1412 blk_throtl_exit(disk
);
1415 static void blkcg_exit(struct task_struct
*tsk
)
1417 if (tsk
->throttle_disk
)
1418 put_disk(tsk
->throttle_disk
);
1419 tsk
->throttle_disk
= NULL
;
1422 struct cgroup_subsys io_cgrp_subsys
= {
1423 .css_alloc
= blkcg_css_alloc
,
1424 .css_online
= blkcg_css_online
,
1425 .css_offline
= blkcg_css_offline
,
1426 .css_free
= blkcg_css_free
,
1427 .css_rstat_flush
= blkcg_rstat_flush
,
1428 .dfl_cftypes
= blkcg_files
,
1429 .legacy_cftypes
= blkcg_legacy_files
,
1430 .legacy_name
= "blkio",
1434 * This ensures that, if available, memcg is automatically enabled
1435 * together on the default hierarchy so that the owner cgroup can
1436 * be retrieved from writeback pages.
1438 .depends_on
= 1 << memory_cgrp_id
,
1441 EXPORT_SYMBOL_GPL(io_cgrp_subsys
);
1444 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1445 * @disk: gendisk of interest
1446 * @pol: blkcg policy to activate
1448 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1449 * bypass mode to populate its blkgs with policy_data for @pol.
1451 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1452 * from IO path. Update of each blkg is protected by both queue and blkcg
1453 * locks so that holding either lock and testing blkcg_policy_enabled() is
1454 * always enough for dereferencing policy data.
1456 * The caller is responsible for synchronizing [de]activations and policy
1457 * [un]registerations. Returns 0 on success, -errno on failure.
1459 int blkcg_activate_policy(struct gendisk
*disk
, const struct blkcg_policy
*pol
)
1461 struct request_queue
*q
= disk
->queue
;
1462 struct blkg_policy_data
*pd_prealloc
= NULL
;
1463 struct blkcg_gq
*blkg
, *pinned_blkg
= NULL
;
1466 if (blkcg_policy_enabled(q
, pol
))
1470 blk_mq_freeze_queue(q
);
1472 spin_lock_irq(&q
->queue_lock
);
1474 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1475 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
) {
1476 struct blkg_policy_data
*pd
;
1478 if (blkg
->pd
[pol
->plid
])
1481 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1482 if (blkg
== pinned_blkg
) {
1486 pd
= pol
->pd_alloc_fn(disk
, blkg
->blkcg
,
1487 GFP_NOWAIT
| __GFP_NOWARN
);
1492 * GFP_NOWAIT failed. Free the existing one and
1493 * prealloc for @blkg w/ GFP_KERNEL.
1496 blkg_put(pinned_blkg
);
1500 spin_unlock_irq(&q
->queue_lock
);
1503 pol
->pd_free_fn(pd_prealloc
);
1504 pd_prealloc
= pol
->pd_alloc_fn(disk
, blkg
->blkcg
,
1512 blkg
->pd
[pol
->plid
] = pd
;
1514 pd
->plid
= pol
->plid
;
1518 /* all allocated, init in the same order */
1519 if (pol
->pd_init_fn
)
1520 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
)
1521 pol
->pd_init_fn(blkg
->pd
[pol
->plid
]);
1523 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
) {
1524 if (pol
->pd_online_fn
)
1525 pol
->pd_online_fn(blkg
->pd
[pol
->plid
]);
1526 blkg
->pd
[pol
->plid
]->online
= true;
1529 __set_bit(pol
->plid
, q
->blkcg_pols
);
1532 spin_unlock_irq(&q
->queue_lock
);
1535 blk_mq_unfreeze_queue(q
);
1537 blkg_put(pinned_blkg
);
1539 pol
->pd_free_fn(pd_prealloc
);
1543 /* alloc failed, nothing's initialized yet, free everything */
1544 spin_lock_irq(&q
->queue_lock
);
1545 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1546 struct blkcg
*blkcg
= blkg
->blkcg
;
1548 spin_lock(&blkcg
->lock
);
1549 if (blkg
->pd
[pol
->plid
]) {
1550 pol
->pd_free_fn(blkg
->pd
[pol
->plid
]);
1551 blkg
->pd
[pol
->plid
] = NULL
;
1553 spin_unlock(&blkcg
->lock
);
1555 spin_unlock_irq(&q
->queue_lock
);
1559 EXPORT_SYMBOL_GPL(blkcg_activate_policy
);
1562 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1563 * @disk: gendisk of interest
1564 * @pol: blkcg policy to deactivate
1566 * Deactivate @pol on @disk. Follows the same synchronization rules as
1567 * blkcg_activate_policy().
1569 void blkcg_deactivate_policy(struct gendisk
*disk
,
1570 const struct blkcg_policy
*pol
)
1572 struct request_queue
*q
= disk
->queue
;
1573 struct blkcg_gq
*blkg
;
1575 if (!blkcg_policy_enabled(q
, pol
))
1579 blk_mq_freeze_queue(q
);
1581 mutex_lock(&q
->blkcg_mutex
);
1582 spin_lock_irq(&q
->queue_lock
);
1584 __clear_bit(pol
->plid
, q
->blkcg_pols
);
1586 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1587 struct blkcg
*blkcg
= blkg
->blkcg
;
1589 spin_lock(&blkcg
->lock
);
1590 if (blkg
->pd
[pol
->plid
]) {
1591 if (blkg
->pd
[pol
->plid
]->online
&& pol
->pd_offline_fn
)
1592 pol
->pd_offline_fn(blkg
->pd
[pol
->plid
]);
1593 pol
->pd_free_fn(blkg
->pd
[pol
->plid
]);
1594 blkg
->pd
[pol
->plid
] = NULL
;
1596 spin_unlock(&blkcg
->lock
);
1599 spin_unlock_irq(&q
->queue_lock
);
1600 mutex_unlock(&q
->blkcg_mutex
);
1603 blk_mq_unfreeze_queue(q
);
1605 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy
);
1607 static void blkcg_free_all_cpd(struct blkcg_policy
*pol
)
1609 struct blkcg
*blkcg
;
1611 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1612 if (blkcg
->cpd
[pol
->plid
]) {
1613 pol
->cpd_free_fn(blkcg
->cpd
[pol
->plid
]);
1614 blkcg
->cpd
[pol
->plid
] = NULL
;
1620 * blkcg_policy_register - register a blkcg policy
1621 * @pol: blkcg policy to register
1623 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1624 * successful registration. Returns 0 on success and -errno on failure.
1626 int blkcg_policy_register(struct blkcg_policy
*pol
)
1628 struct blkcg
*blkcg
;
1631 mutex_lock(&blkcg_pol_register_mutex
);
1632 mutex_lock(&blkcg_pol_mutex
);
1634 /* find an empty slot */
1636 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1637 if (!blkcg_policy
[i
])
1639 if (i
>= BLKCG_MAX_POLS
) {
1640 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1644 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1645 if ((!pol
->cpd_alloc_fn
^ !pol
->cpd_free_fn
) ||
1646 (!pol
->pd_alloc_fn
^ !pol
->pd_free_fn
))
1651 blkcg_policy
[pol
->plid
] = pol
;
1653 /* allocate and install cpd's */
1654 if (pol
->cpd_alloc_fn
) {
1655 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1656 struct blkcg_policy_data
*cpd
;
1658 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1662 blkcg
->cpd
[pol
->plid
] = cpd
;
1664 cpd
->plid
= pol
->plid
;
1668 mutex_unlock(&blkcg_pol_mutex
);
1670 /* everything is in place, add intf files for the new policy */
1671 if (pol
->dfl_cftypes
)
1672 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys
,
1674 if (pol
->legacy_cftypes
)
1675 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys
,
1676 pol
->legacy_cftypes
));
1677 mutex_unlock(&blkcg_pol_register_mutex
);
1681 if (pol
->cpd_free_fn
)
1682 blkcg_free_all_cpd(pol
);
1684 blkcg_policy
[pol
->plid
] = NULL
;
1686 mutex_unlock(&blkcg_pol_mutex
);
1687 mutex_unlock(&blkcg_pol_register_mutex
);
1690 EXPORT_SYMBOL_GPL(blkcg_policy_register
);
1693 * blkcg_policy_unregister - unregister a blkcg policy
1694 * @pol: blkcg policy to unregister
1696 * Undo blkcg_policy_register(@pol). Might sleep.
1698 void blkcg_policy_unregister(struct blkcg_policy
*pol
)
1700 mutex_lock(&blkcg_pol_register_mutex
);
1702 if (WARN_ON(blkcg_policy
[pol
->plid
] != pol
))
1705 /* kill the intf files first */
1706 if (pol
->dfl_cftypes
)
1707 cgroup_rm_cftypes(pol
->dfl_cftypes
);
1708 if (pol
->legacy_cftypes
)
1709 cgroup_rm_cftypes(pol
->legacy_cftypes
);
1711 /* remove cpds and unregister */
1712 mutex_lock(&blkcg_pol_mutex
);
1714 if (pol
->cpd_free_fn
)
1715 blkcg_free_all_cpd(pol
);
1717 blkcg_policy
[pol
->plid
] = NULL
;
1719 mutex_unlock(&blkcg_pol_mutex
);
1721 mutex_unlock(&blkcg_pol_register_mutex
);
1723 EXPORT_SYMBOL_GPL(blkcg_policy_unregister
);
1726 * Scale the accumulated delay based on how long it has been since we updated
1727 * the delay. We only call this when we are adding delay, in case it's been a
1728 * while since we added delay, and when we are checking to see if we need to
1729 * delay a task, to account for any delays that may have occurred.
1731 static void blkcg_scale_delay(struct blkcg_gq
*blkg
, u64 now
)
1733 u64 old
= atomic64_read(&blkg
->delay_start
);
1735 /* negative use_delay means no scaling, see blkcg_set_delay() */
1736 if (atomic_read(&blkg
->use_delay
) < 0)
1740 * We only want to scale down every second. The idea here is that we
1741 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1742 * time window. We only want to throttle tasks for recent delay that
1743 * has occurred, in 1 second time windows since that's the maximum
1744 * things can be throttled. We save the current delay window in
1745 * blkg->last_delay so we know what amount is still left to be charged
1746 * to the blkg from this point onward. blkg->last_use keeps track of
1747 * the use_delay counter. The idea is if we're unthrottling the blkg we
1748 * are ok with whatever is happening now, and we can take away more of
1749 * the accumulated delay as we've already throttled enough that
1750 * everybody is happy with their IO latencies.
1752 if (time_before64(old
+ NSEC_PER_SEC
, now
) &&
1753 atomic64_try_cmpxchg(&blkg
->delay_start
, &old
, now
)) {
1754 u64 cur
= atomic64_read(&blkg
->delay_nsec
);
1755 u64 sub
= min_t(u64
, blkg
->last_delay
, now
- old
);
1756 int cur_use
= atomic_read(&blkg
->use_delay
);
1759 * We've been unthrottled, subtract a larger chunk of our
1760 * accumulated delay.
1762 if (cur_use
< blkg
->last_use
)
1763 sub
= max_t(u64
, sub
, blkg
->last_delay
>> 1);
1766 * This shouldn't happen, but handle it anyway. Our delay_nsec
1767 * should only ever be growing except here where we subtract out
1768 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1769 * rather not end up with negative numbers.
1771 if (unlikely(cur
< sub
)) {
1772 atomic64_set(&blkg
->delay_nsec
, 0);
1773 blkg
->last_delay
= 0;
1775 atomic64_sub(sub
, &blkg
->delay_nsec
);
1776 blkg
->last_delay
= cur
- sub
;
1778 blkg
->last_use
= cur_use
;
1783 * This is called when we want to actually walk up the hierarchy and check to
1784 * see if we need to throttle, and then actually throttle if there is some
1785 * accumulated delay. This should only be called upon return to user space so
1786 * we're not holding some lock that would induce a priority inversion.
1788 static void blkcg_maybe_throttle_blkg(struct blkcg_gq
*blkg
, bool use_memdelay
)
1790 unsigned long pflags
;
1792 u64 now
= ktime_to_ns(ktime_get());
1797 while (blkg
->parent
) {
1798 int use_delay
= atomic_read(&blkg
->use_delay
);
1803 blkcg_scale_delay(blkg
, now
);
1804 this_delay
= atomic64_read(&blkg
->delay_nsec
);
1805 if (this_delay
> delay_nsec
) {
1806 delay_nsec
= this_delay
;
1807 clamp
= use_delay
> 0;
1810 blkg
= blkg
->parent
;
1817 * Let's not sleep for all eternity if we've amassed a huge delay.
1818 * Swapping or metadata IO can accumulate 10's of seconds worth of
1819 * delay, and we want userspace to be able to do _something_ so cap the
1820 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1821 * tasks will be delayed for 0.25 second for every syscall. If
1822 * blkcg_set_delay() was used as indicated by negative use_delay, the
1823 * caller is responsible for regulating the range.
1826 delay_nsec
= min_t(u64
, delay_nsec
, 250 * NSEC_PER_MSEC
);
1829 psi_memstall_enter(&pflags
);
1831 exp
= ktime_add_ns(now
, delay_nsec
);
1832 tok
= io_schedule_prepare();
1834 __set_current_state(TASK_KILLABLE
);
1835 if (!schedule_hrtimeout(&exp
, HRTIMER_MODE_ABS
))
1837 } while (!fatal_signal_pending(current
));
1838 io_schedule_finish(tok
);
1841 psi_memstall_leave(&pflags
);
1845 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1847 * This is only called if we've been marked with set_notify_resume(). Obviously
1848 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1849 * check to see if current->throttle_disk is set and if not this doesn't do
1850 * anything. This should only ever be called by the resume code, it's not meant
1851 * to be called by people willy-nilly as it will actually do the work to
1852 * throttle the task if it is setup for throttling.
1854 void blkcg_maybe_throttle_current(void)
1856 struct gendisk
*disk
= current
->throttle_disk
;
1857 struct blkcg
*blkcg
;
1858 struct blkcg_gq
*blkg
;
1859 bool use_memdelay
= current
->use_memdelay
;
1864 current
->throttle_disk
= NULL
;
1865 current
->use_memdelay
= false;
1868 blkcg
= css_to_blkcg(blkcg_css());
1871 blkg
= blkg_lookup(blkcg
, disk
->queue
);
1874 if (!blkg_tryget(blkg
))
1878 blkcg_maybe_throttle_blkg(blkg
, use_memdelay
);
1887 * blkcg_schedule_throttle - this task needs to check for throttling
1888 * @disk: disk to throttle
1889 * @use_memdelay: do we charge this to memory delay for PSI
1891 * This is called by the IO controller when we know there's delay accumulated
1892 * for the blkg for this task. We do not pass the blkg because there are places
1893 * we call this that may not have that information, the swapping code for
1894 * instance will only have a block_device at that point. This set's the
1895 * notify_resume for the task to check and see if it requires throttling before
1896 * returning to user space.
1898 * We will only schedule once per syscall. You can call this over and over
1899 * again and it will only do the check once upon return to user space, and only
1900 * throttle once. If the task needs to be throttled again it'll need to be
1901 * re-set at the next time we see the task.
1903 void blkcg_schedule_throttle(struct gendisk
*disk
, bool use_memdelay
)
1905 if (unlikely(current
->flags
& PF_KTHREAD
))
1908 if (current
->throttle_disk
!= disk
) {
1909 if (test_bit(GD_DEAD
, &disk
->state
))
1911 get_device(disk_to_dev(disk
));
1913 if (current
->throttle_disk
)
1914 put_disk(current
->throttle_disk
);
1915 current
->throttle_disk
= disk
;
1919 current
->use_memdelay
= use_memdelay
;
1920 set_notify_resume(current
);
1924 * blkcg_add_delay - add delay to this blkg
1925 * @blkg: blkg of interest
1926 * @now: the current time in nanoseconds
1927 * @delta: how many nanoseconds of delay to add
1929 * Charge @delta to the blkg's current delay accumulation. This is used to
1930 * throttle tasks if an IO controller thinks we need more throttling.
1932 void blkcg_add_delay(struct blkcg_gq
*blkg
, u64 now
, u64 delta
)
1934 if (WARN_ON_ONCE(atomic_read(&blkg
->use_delay
) < 0))
1936 blkcg_scale_delay(blkg
, now
);
1937 atomic64_add(delta
, &blkg
->delay_nsec
);
1941 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1945 * As the failure mode here is to walk up the blkg tree, this ensure that the
1946 * blkg->parent pointers are always valid. This returns the blkg that it ended
1947 * up taking a reference on or %NULL if no reference was taken.
1949 static inline struct blkcg_gq
*blkg_tryget_closest(struct bio
*bio
,
1950 struct cgroup_subsys_state
*css
)
1952 struct blkcg_gq
*blkg
, *ret_blkg
= NULL
;
1955 blkg
= blkg_lookup_create(css_to_blkcg(css
), bio
->bi_bdev
->bd_disk
);
1957 if (blkg_tryget(blkg
)) {
1961 blkg
= blkg
->parent
;
1969 * bio_associate_blkg_from_css - associate a bio with a specified css
1973 * Associate @bio with the blkg found by combining the css's blkg and the
1974 * request_queue of the @bio. An association failure is handled by walking up
1975 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1976 * and q->root_blkg. This situation only happens when a cgroup is dying and
1977 * then the remaining bios will spill to the closest alive blkg.
1979 * A reference will be taken on the blkg and will be released when @bio is
1982 void bio_associate_blkg_from_css(struct bio
*bio
,
1983 struct cgroup_subsys_state
*css
)
1986 blkg_put(bio
->bi_blkg
);
1988 if (css
&& css
->parent
) {
1989 bio
->bi_blkg
= blkg_tryget_closest(bio
, css
);
1991 blkg_get(bdev_get_queue(bio
->bi_bdev
)->root_blkg
);
1992 bio
->bi_blkg
= bdev_get_queue(bio
->bi_bdev
)->root_blkg
;
1995 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css
);
1998 * bio_associate_blkg - associate a bio with a blkg
2001 * Associate @bio with the blkg found from the bio's css and request_queue.
2002 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2003 * already associated, the css is reused and association redone as the
2004 * request_queue may have changed.
2006 void bio_associate_blkg(struct bio
*bio
)
2008 struct cgroup_subsys_state
*css
;
2013 css
= bio_blkcg_css(bio
);
2017 bio_associate_blkg_from_css(bio
, css
);
2021 EXPORT_SYMBOL_GPL(bio_associate_blkg
);
2024 * bio_clone_blkg_association - clone blkg association from src to dst bio
2025 * @dst: destination bio
2028 void bio_clone_blkg_association(struct bio
*dst
, struct bio
*src
)
2031 bio_associate_blkg_from_css(dst
, bio_blkcg_css(src
));
2033 EXPORT_SYMBOL_GPL(bio_clone_blkg_association
);
2035 static int blk_cgroup_io_type(struct bio
*bio
)
2037 if (op_is_discard(bio
->bi_opf
))
2038 return BLKG_IOSTAT_DISCARD
;
2039 if (op_is_write(bio
->bi_opf
))
2040 return BLKG_IOSTAT_WRITE
;
2041 return BLKG_IOSTAT_READ
;
2044 void blk_cgroup_bio_start(struct bio
*bio
)
2046 struct blkcg
*blkcg
= bio
->bi_blkg
->blkcg
;
2047 int rwd
= blk_cgroup_io_type(bio
), cpu
;
2048 struct blkg_iostat_set
*bis
;
2049 unsigned long flags
;
2051 /* Root-level stats are sourced from system-wide IO stats */
2052 if (!cgroup_parent(blkcg
->css
.cgroup
))
2056 bis
= per_cpu_ptr(bio
->bi_blkg
->iostat_cpu
, cpu
);
2057 flags
= u64_stats_update_begin_irqsave(&bis
->sync
);
2060 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2061 * bio and we would have already accounted for the size of the bio.
2063 if (!bio_flagged(bio
, BIO_CGROUP_ACCT
)) {
2064 bio_set_flag(bio
, BIO_CGROUP_ACCT
);
2065 bis
->cur
.bytes
[rwd
] += bio
->bi_iter
.bi_size
;
2067 bis
->cur
.ios
[rwd
]++;
2070 * If the iostat_cpu isn't in a lockless list, put it into the
2071 * list to indicate that a stat update is pending.
2073 if (!READ_ONCE(bis
->lqueued
)) {
2074 struct llist_head
*lhead
= this_cpu_ptr(blkcg
->lhead
);
2076 llist_add(&bis
->lnode
, lhead
);
2077 WRITE_ONCE(bis
->lqueued
, true);
2078 percpu_ref_get(&bis
->blkg
->refcnt
);
2081 u64_stats_update_end_irqrestore(&bis
->sync
, flags
);
2082 if (cgroup_subsys_on_dfl(io_cgrp_subsys
))
2083 cgroup_rstat_updated(blkcg
->css
.cgroup
, cpu
);
2087 bool blk_cgroup_congested(void)
2089 struct cgroup_subsys_state
*css
;
2093 for (css
= blkcg_css(); css
; css
= css
->parent
) {
2094 if (atomic_read(&css
->cgroup
->congestion_count
)) {
2103 module_param(blkcg_debug_stats
, bool, 0644);
2104 MODULE_PARM_DESC(blkcg_debug_stats
, "True if you want debug stats, false if not");