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"
37 static void __blkcg_rstat_flush(struct blkcg
*blkcg
, int cpu
);
40 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
41 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
42 * policy [un]register operations including cgroup file additions /
43 * removals. Putting cgroup file registration outside blkcg_pol_mutex
44 * allows grabbing it from cgroup callbacks.
46 static DEFINE_MUTEX(blkcg_pol_register_mutex
);
47 static DEFINE_MUTEX(blkcg_pol_mutex
);
49 struct blkcg blkcg_root
;
50 EXPORT_SYMBOL_GPL(blkcg_root
);
52 struct cgroup_subsys_state
* const blkcg_root_css
= &blkcg_root
.css
;
53 EXPORT_SYMBOL_GPL(blkcg_root_css
);
55 static struct blkcg_policy
*blkcg_policy
[BLKCG_MAX_POLS
];
57 static LIST_HEAD(all_blkcgs
); /* protected by blkcg_pol_mutex */
59 bool blkcg_debug_stats
= false;
61 static DEFINE_RAW_SPINLOCK(blkg_stat_lock
);
63 #define BLKG_DESTROY_BATCH_SIZE 64
66 * Lockless lists for tracking IO stats update
68 * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
69 * There are multiple blkg's (one for each block device) attached to each
70 * blkcg. The rstat code keeps track of which cpu has IO stats updated,
71 * but it doesn't know which blkg has the updated stats. If there are many
72 * block devices in a system, the cost of iterating all the blkg's to flush
73 * out the IO stats can be high. To reduce such overhead, a set of percpu
74 * lockless lists (lhead) per blkcg are used to track the set of recently
75 * updated iostat_cpu's since the last flush. An iostat_cpu will be put
76 * onto the lockless list on the update side [blk_cgroup_bio_start()] if
77 * not there yet and then removed when being flushed [blkcg_rstat_flush()].
78 * References to blkg are gotten and then put back in the process to
79 * protect against blkg removal.
81 * Return: 0 if successful or -ENOMEM if allocation fails.
83 static int init_blkcg_llists(struct blkcg
*blkcg
)
87 blkcg
->lhead
= alloc_percpu_gfp(struct llist_head
, GFP_KERNEL
);
91 for_each_possible_cpu(cpu
)
92 init_llist_head(per_cpu_ptr(blkcg
->lhead
, cpu
));
97 * blkcg_css - find the current css
99 * Find the css associated with either the kthread or the current task.
100 * This may return a dying css, so it is up to the caller to use tryget logic
101 * to confirm it is alive and well.
103 static struct cgroup_subsys_state
*blkcg_css(void)
105 struct cgroup_subsys_state
*css
;
107 css
= kthread_blkcg();
110 return task_css(current
, io_cgrp_id
);
113 static bool blkcg_policy_enabled(struct request_queue
*q
,
114 const struct blkcg_policy
*pol
)
116 return pol
&& test_bit(pol
->plid
, q
->blkcg_pols
);
119 static void blkg_free_workfn(struct work_struct
*work
)
121 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
123 struct request_queue
*q
= blkg
->q
;
127 * pd_free_fn() can also be called from blkcg_deactivate_policy(),
128 * in order to make sure pd_free_fn() is called in order, the deletion
129 * of the list blkg->q_node is delayed to here from blkg_destroy(), and
130 * blkcg_mutex is used to synchronize blkg_free_workfn() and
131 * blkcg_deactivate_policy().
133 mutex_lock(&q
->blkcg_mutex
);
134 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
136 blkcg_policy
[i
]->pd_free_fn(blkg
->pd
[i
]);
138 blkg_put(blkg
->parent
);
139 spin_lock_irq(&q
->queue_lock
);
140 list_del_init(&blkg
->q_node
);
141 spin_unlock_irq(&q
->queue_lock
);
142 mutex_unlock(&q
->blkcg_mutex
);
145 free_percpu(blkg
->iostat_cpu
);
146 percpu_ref_exit(&blkg
->refcnt
);
151 * blkg_free - free a blkg
152 * @blkg: blkg to free
154 * Free @blkg which may be partially allocated.
156 static void blkg_free(struct blkcg_gq
*blkg
)
162 * Both ->pd_free_fn() and request queue's release handler may
163 * sleep, so free us by scheduling one work func
165 INIT_WORK(&blkg
->free_work
, blkg_free_workfn
);
166 schedule_work(&blkg
->free_work
);
169 static void __blkg_release(struct rcu_head
*rcu
)
171 struct blkcg_gq
*blkg
= container_of(rcu
, struct blkcg_gq
, rcu_head
);
172 struct blkcg
*blkcg
= blkg
->blkcg
;
175 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
176 WARN_ON(!bio_list_empty(&blkg
->async_bios
));
179 * Flush all the non-empty percpu lockless lists before releasing
180 * us, given these stat belongs to us.
182 * blkg_stat_lock is for serializing blkg stat update
184 for_each_possible_cpu(cpu
)
185 __blkcg_rstat_flush(blkcg
, cpu
);
187 /* release the blkcg and parent blkg refs this blkg has been holding */
188 css_put(&blkg
->blkcg
->css
);
193 * A group is RCU protected, but having an rcu lock does not mean that one
194 * can access all the fields of blkg and assume these are valid. For
195 * example, don't try to follow throtl_data and request queue links.
197 * Having a reference to blkg under an rcu allows accesses to only values
198 * local to groups like group stats and group rate limits.
200 static void blkg_release(struct percpu_ref
*ref
)
202 struct blkcg_gq
*blkg
= container_of(ref
, struct blkcg_gq
, refcnt
);
204 call_rcu(&blkg
->rcu_head
, __blkg_release
);
207 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
208 static struct workqueue_struct
*blkcg_punt_bio_wq
;
210 static void blkg_async_bio_workfn(struct work_struct
*work
)
212 struct blkcg_gq
*blkg
= container_of(work
, struct blkcg_gq
,
214 struct bio_list bios
= BIO_EMPTY_LIST
;
216 struct blk_plug plug
;
217 bool need_plug
= false;
219 /* as long as there are pending bios, @blkg can't go away */
220 spin_lock(&blkg
->async_bio_lock
);
221 bio_list_merge(&bios
, &blkg
->async_bios
);
222 bio_list_init(&blkg
->async_bios
);
223 spin_unlock(&blkg
->async_bio_lock
);
225 /* start plug only when bio_list contains at least 2 bios */
226 if (bios
.head
&& bios
.head
->bi_next
) {
228 blk_start_plug(&plug
);
230 while ((bio
= bio_list_pop(&bios
)))
233 blk_finish_plug(&plug
);
237 * When a shared kthread issues a bio for a cgroup, doing so synchronously can
238 * lead to priority inversions as the kthread can be trapped waiting for that
239 * cgroup. Use this helper instead of submit_bio to punt the actual issuing to
240 * a dedicated per-blkcg work item to avoid such priority inversions.
242 void blkcg_punt_bio_submit(struct bio
*bio
)
244 struct blkcg_gq
*blkg
= bio
->bi_blkg
;
247 spin_lock(&blkg
->async_bio_lock
);
248 bio_list_add(&blkg
->async_bios
, bio
);
249 spin_unlock(&blkg
->async_bio_lock
);
250 queue_work(blkcg_punt_bio_wq
, &blkg
->async_bio_work
);
252 /* never bounce for the root cgroup */
256 EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit
);
258 static int __init
blkcg_punt_bio_init(void)
260 blkcg_punt_bio_wq
= alloc_workqueue("blkcg_punt_bio",
261 WQ_MEM_RECLAIM
| WQ_FREEZABLE
|
262 WQ_UNBOUND
| WQ_SYSFS
, 0);
263 if (!blkcg_punt_bio_wq
)
267 subsys_initcall(blkcg_punt_bio_init
);
268 #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
271 * bio_blkcg_css - return the blkcg CSS associated with a bio
274 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
275 * associated. Callers are expected to either handle %NULL or know association
276 * has been done prior to calling this.
278 struct cgroup_subsys_state
*bio_blkcg_css(struct bio
*bio
)
280 if (!bio
|| !bio
->bi_blkg
)
282 return &bio
->bi_blkg
->blkcg
->css
;
284 EXPORT_SYMBOL_GPL(bio_blkcg_css
);
287 * blkcg_parent - get the parent of a blkcg
288 * @blkcg: blkcg of interest
290 * Return the parent blkcg of @blkcg. Can be called anytime.
292 static inline struct blkcg
*blkcg_parent(struct blkcg
*blkcg
)
294 return css_to_blkcg(blkcg
->css
.parent
);
298 * blkg_alloc - allocate a blkg
299 * @blkcg: block cgroup the new blkg is associated with
300 * @disk: gendisk the new blkg is associated with
301 * @gfp_mask: allocation mask to use
303 * Allocate a new blkg assocating @blkcg and @q.
305 static struct blkcg_gq
*blkg_alloc(struct blkcg
*blkcg
, struct gendisk
*disk
,
308 struct blkcg_gq
*blkg
;
311 /* alloc and init base part */
312 blkg
= kzalloc_node(sizeof(*blkg
), gfp_mask
, disk
->queue
->node
);
315 if (percpu_ref_init(&blkg
->refcnt
, blkg_release
, 0, gfp_mask
))
317 blkg
->iostat_cpu
= alloc_percpu_gfp(struct blkg_iostat_set
, gfp_mask
);
318 if (!blkg
->iostat_cpu
)
319 goto out_exit_refcnt
;
320 if (!blk_get_queue(disk
->queue
))
321 goto out_free_iostat
;
323 blkg
->q
= disk
->queue
;
324 INIT_LIST_HEAD(&blkg
->q_node
);
326 #ifdef CONFIG_BLK_CGROUP_PUNT_BIO
327 spin_lock_init(&blkg
->async_bio_lock
);
328 bio_list_init(&blkg
->async_bios
);
329 INIT_WORK(&blkg
->async_bio_work
, blkg_async_bio_workfn
);
332 u64_stats_init(&blkg
->iostat
.sync
);
333 for_each_possible_cpu(cpu
) {
334 u64_stats_init(&per_cpu_ptr(blkg
->iostat_cpu
, cpu
)->sync
);
335 per_cpu_ptr(blkg
->iostat_cpu
, cpu
)->blkg
= blkg
;
338 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
339 struct blkcg_policy
*pol
= blkcg_policy
[i
];
340 struct blkg_policy_data
*pd
;
342 if (!blkcg_policy_enabled(disk
->queue
, pol
))
345 /* alloc per-policy data and attach it to blkg */
346 pd
= pol
->pd_alloc_fn(disk
, blkcg
, gfp_mask
);
360 blkcg_policy
[i
]->pd_free_fn(blkg
->pd
[i
]);
361 blk_put_queue(disk
->queue
);
363 free_percpu(blkg
->iostat_cpu
);
365 percpu_ref_exit(&blkg
->refcnt
);
372 * If @new_blkg is %NULL, this function tries to allocate a new one as
373 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
375 static struct blkcg_gq
*blkg_create(struct blkcg
*blkcg
, struct gendisk
*disk
,
376 struct blkcg_gq
*new_blkg
)
378 struct blkcg_gq
*blkg
;
381 lockdep_assert_held(&disk
->queue
->queue_lock
);
383 /* request_queue is dying, do not create/recreate a blkg */
384 if (blk_queue_dying(disk
->queue
)) {
389 /* blkg holds a reference to blkcg */
390 if (!css_tryget_online(&blkcg
->css
)) {
397 new_blkg
= blkg_alloc(blkcg
, disk
, GFP_NOWAIT
| __GFP_NOWARN
);
398 if (unlikely(!new_blkg
)) {
406 if (blkcg_parent(blkcg
)) {
407 blkg
->parent
= blkg_lookup(blkcg_parent(blkcg
), disk
->queue
);
408 if (WARN_ON_ONCE(!blkg
->parent
)) {
412 blkg_get(blkg
->parent
);
415 /* invoke per-policy init */
416 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
417 struct blkcg_policy
*pol
= blkcg_policy
[i
];
419 if (blkg
->pd
[i
] && pol
->pd_init_fn
)
420 pol
->pd_init_fn(blkg
->pd
[i
]);
424 spin_lock(&blkcg
->lock
);
425 ret
= radix_tree_insert(&blkcg
->blkg_tree
, disk
->queue
->id
, blkg
);
427 hlist_add_head_rcu(&blkg
->blkcg_node
, &blkcg
->blkg_list
);
428 list_add(&blkg
->q_node
, &disk
->queue
->blkg_list
);
430 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
431 struct blkcg_policy
*pol
= blkcg_policy
[i
];
434 if (pol
->pd_online_fn
)
435 pol
->pd_online_fn(blkg
->pd
[i
]);
436 blkg
->pd
[i
]->online
= true;
441 spin_unlock(&blkcg
->lock
);
446 /* @blkg failed fully initialized, use the usual release path */
451 css_put(&blkcg
->css
);
459 * blkg_lookup_create - lookup blkg, try to create one if not there
460 * @blkcg: blkcg of interest
461 * @disk: gendisk of interest
463 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
464 * create one. blkg creation is performed recursively from blkcg_root such
465 * that all non-root blkg's have access to the parent blkg. This function
466 * should be called under RCU read lock and takes @disk->queue->queue_lock.
468 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
471 static struct blkcg_gq
*blkg_lookup_create(struct blkcg
*blkcg
,
472 struct gendisk
*disk
)
474 struct request_queue
*q
= disk
->queue
;
475 struct blkcg_gq
*blkg
;
478 WARN_ON_ONCE(!rcu_read_lock_held());
480 blkg
= blkg_lookup(blkcg
, q
);
484 spin_lock_irqsave(&q
->queue_lock
, flags
);
485 blkg
= blkg_lookup(blkcg
, q
);
487 if (blkcg
!= &blkcg_root
&&
488 blkg
!= rcu_dereference(blkcg
->blkg_hint
))
489 rcu_assign_pointer(blkcg
->blkg_hint
, blkg
);
494 * Create blkgs walking down from blkcg_root to @blkcg, so that all
495 * non-root blkgs have access to their parents. Returns the closest
496 * blkg to the intended blkg should blkg_create() fail.
499 struct blkcg
*pos
= blkcg
;
500 struct blkcg
*parent
= blkcg_parent(blkcg
);
501 struct blkcg_gq
*ret_blkg
= q
->root_blkg
;
504 blkg
= blkg_lookup(parent
, q
);
506 /* remember closest blkg */
511 parent
= blkcg_parent(parent
);
514 blkg
= blkg_create(pos
, disk
, NULL
);
524 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
528 static void blkg_destroy(struct blkcg_gq
*blkg
)
530 struct blkcg
*blkcg
= blkg
->blkcg
;
533 lockdep_assert_held(&blkg
->q
->queue_lock
);
534 lockdep_assert_held(&blkcg
->lock
);
537 * blkg stays on the queue list until blkg_free_workfn(), see details in
538 * blkg_free_workfn(), hence this function can be called from
539 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
540 * blkg_free_workfn().
542 if (hlist_unhashed(&blkg
->blkcg_node
))
545 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
546 struct blkcg_policy
*pol
= blkcg_policy
[i
];
548 if (blkg
->pd
[i
] && blkg
->pd
[i
]->online
) {
549 blkg
->pd
[i
]->online
= false;
550 if (pol
->pd_offline_fn
)
551 pol
->pd_offline_fn(blkg
->pd
[i
]);
555 blkg
->online
= false;
557 radix_tree_delete(&blkcg
->blkg_tree
, blkg
->q
->id
);
558 hlist_del_init_rcu(&blkg
->blkcg_node
);
561 * Both setting lookup hint to and clearing it from @blkg are done
562 * under queue_lock. If it's not pointing to @blkg now, it never
563 * will. Hint assignment itself can race safely.
565 if (rcu_access_pointer(blkcg
->blkg_hint
) == blkg
)
566 rcu_assign_pointer(blkcg
->blkg_hint
, NULL
);
569 * Put the reference taken at the time of creation so that when all
570 * queues are gone, group can be destroyed.
572 percpu_ref_kill(&blkg
->refcnt
);
575 static void blkg_destroy_all(struct gendisk
*disk
)
577 struct request_queue
*q
= disk
->queue
;
578 struct blkcg_gq
*blkg
, *n
;
579 int count
= BLKG_DESTROY_BATCH_SIZE
;
583 spin_lock_irq(&q
->queue_lock
);
584 list_for_each_entry_safe(blkg
, n
, &q
->blkg_list
, q_node
) {
585 struct blkcg
*blkcg
= blkg
->blkcg
;
587 if (hlist_unhashed(&blkg
->blkcg_node
))
590 spin_lock(&blkcg
->lock
);
592 spin_unlock(&blkcg
->lock
);
595 * in order to avoid holding the spin lock for too long, release
596 * it when a batch of blkgs are destroyed.
599 count
= BLKG_DESTROY_BATCH_SIZE
;
600 spin_unlock_irq(&q
->queue_lock
);
607 * Mark policy deactivated since policy offline has been done, and
608 * the free is scheduled, so future blkcg_deactivate_policy() can
611 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
612 struct blkcg_policy
*pol
= blkcg_policy
[i
];
615 __clear_bit(pol
->plid
, q
->blkcg_pols
);
619 spin_unlock_irq(&q
->queue_lock
);
622 static int blkcg_reset_stats(struct cgroup_subsys_state
*css
,
623 struct cftype
*cftype
, u64 val
)
625 struct blkcg
*blkcg
= css_to_blkcg(css
);
626 struct blkcg_gq
*blkg
;
629 mutex_lock(&blkcg_pol_mutex
);
630 spin_lock_irq(&blkcg
->lock
);
633 * Note that stat reset is racy - it doesn't synchronize against
634 * stat updates. This is a debug feature which shouldn't exist
635 * anyway. If you get hit by a race, retry.
637 hlist_for_each_entry(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
638 for_each_possible_cpu(cpu
) {
639 struct blkg_iostat_set
*bis
=
640 per_cpu_ptr(blkg
->iostat_cpu
, cpu
);
641 memset(bis
, 0, sizeof(*bis
));
643 /* Re-initialize the cleared blkg_iostat_set */
644 u64_stats_init(&bis
->sync
);
647 memset(&blkg
->iostat
, 0, sizeof(blkg
->iostat
));
648 u64_stats_init(&blkg
->iostat
.sync
);
650 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
651 struct blkcg_policy
*pol
= blkcg_policy
[i
];
653 if (blkg
->pd
[i
] && pol
->pd_reset_stats_fn
)
654 pol
->pd_reset_stats_fn(blkg
->pd
[i
]);
658 spin_unlock_irq(&blkcg
->lock
);
659 mutex_unlock(&blkcg_pol_mutex
);
663 const char *blkg_dev_name(struct blkcg_gq
*blkg
)
667 return bdi_dev_name(blkg
->q
->disk
->bdi
);
671 * blkcg_print_blkgs - helper for printing per-blkg data
672 * @sf: seq_file to print to
673 * @blkcg: blkcg of interest
674 * @prfill: fill function to print out a blkg
675 * @pol: policy in question
676 * @data: data to be passed to @prfill
677 * @show_total: to print out sum of prfill return values or not
679 * This function invokes @prfill on each blkg of @blkcg if pd for the
680 * policy specified by @pol exists. @prfill is invoked with @sf, the
681 * policy data and @data and the matching queue lock held. If @show_total
682 * is %true, the sum of the return values from @prfill is printed with
683 * "Total" label at the end.
685 * This is to be used to construct print functions for
686 * cftype->read_seq_string method.
688 void blkcg_print_blkgs(struct seq_file
*sf
, struct blkcg
*blkcg
,
689 u64 (*prfill
)(struct seq_file
*,
690 struct blkg_policy_data
*, int),
691 const struct blkcg_policy
*pol
, int data
,
694 struct blkcg_gq
*blkg
;
698 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
699 spin_lock_irq(&blkg
->q
->queue_lock
);
700 if (blkcg_policy_enabled(blkg
->q
, pol
))
701 total
+= prfill(sf
, blkg
->pd
[pol
->plid
], data
);
702 spin_unlock_irq(&blkg
->q
->queue_lock
);
707 seq_printf(sf
, "Total %llu\n", (unsigned long long)total
);
709 EXPORT_SYMBOL_GPL(blkcg_print_blkgs
);
712 * __blkg_prfill_u64 - prfill helper for a single u64 value
713 * @sf: seq_file to print to
714 * @pd: policy private data of interest
717 * Print @v to @sf for the device associated with @pd.
719 u64
__blkg_prfill_u64(struct seq_file
*sf
, struct blkg_policy_data
*pd
, u64 v
)
721 const char *dname
= blkg_dev_name(pd
->blkg
);
726 seq_printf(sf
, "%s %llu\n", dname
, (unsigned long long)v
);
729 EXPORT_SYMBOL_GPL(__blkg_prfill_u64
);
732 * blkg_conf_init - initialize a blkg_conf_ctx
733 * @ctx: blkg_conf_ctx to initialize
734 * @input: input string
736 * Initialize @ctx which can be used to parse blkg config input string @input.
737 * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
738 * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
740 void blkg_conf_init(struct blkg_conf_ctx
*ctx
, char *input
)
742 *ctx
= (struct blkg_conf_ctx
){ .input
= input
};
744 EXPORT_SYMBOL_GPL(blkg_conf_init
);
747 * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
748 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
750 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
751 * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
752 * set to point past the device node prefix.
754 * This function may be called multiple times on @ctx and the extra calls become
755 * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
756 * explicitly if bdev access is needed without resolving the blkcg / policy part
757 * of @ctx->input. Returns -errno on error.
759 int blkg_conf_open_bdev(struct blkg_conf_ctx
*ctx
)
761 char *input
= ctx
->input
;
762 unsigned int major
, minor
;
763 struct block_device
*bdev
;
769 if (sscanf(input
, "%u:%u%n", &major
, &minor
, &key_len
) != 2)
773 if (!isspace(*input
))
775 input
= skip_spaces(input
);
777 bdev
= blkdev_get_no_open(MKDEV(major
, minor
));
780 if (bdev_is_partition(bdev
)) {
781 blkdev_put_no_open(bdev
);
785 mutex_lock(&bdev
->bd_queue
->rq_qos_mutex
);
786 if (!disk_live(bdev
->bd_disk
)) {
787 blkdev_put_no_open(bdev
);
788 mutex_unlock(&bdev
->bd_queue
->rq_qos_mutex
);
798 * blkg_conf_prep - parse and prepare for per-blkg config update
799 * @blkcg: target block cgroup
800 * @pol: target policy
801 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
803 * Parse per-blkg config update from @ctx->input and initialize @ctx
804 * accordingly. On success, @ctx->body points to the part of @ctx->input
805 * following MAJ:MIN, @ctx->bdev points to the target block device and
806 * @ctx->blkg to the blkg being configured.
808 * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
809 * function returns with queue lock held and must be followed by
812 int blkg_conf_prep(struct blkcg
*blkcg
, const struct blkcg_policy
*pol
,
813 struct blkg_conf_ctx
*ctx
)
814 __acquires(&bdev
->bd_queue
->queue_lock
)
816 struct gendisk
*disk
;
817 struct request_queue
*q
;
818 struct blkcg_gq
*blkg
;
821 ret
= blkg_conf_open_bdev(ctx
);
825 disk
= ctx
->bdev
->bd_disk
;
829 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
830 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
832 ret
= blk_queue_enter(q
, 0);
836 spin_lock_irq(&q
->queue_lock
);
838 if (!blkcg_policy_enabled(q
, pol
)) {
843 blkg
= blkg_lookup(blkcg
, q
);
848 * Create blkgs walking down from blkcg_root to @blkcg, so that all
849 * non-root blkgs have access to their parents.
852 struct blkcg
*pos
= blkcg
;
853 struct blkcg
*parent
;
854 struct blkcg_gq
*new_blkg
;
856 parent
= blkcg_parent(blkcg
);
857 while (parent
&& !blkg_lookup(parent
, q
)) {
859 parent
= blkcg_parent(parent
);
862 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
863 spin_unlock_irq(&q
->queue_lock
);
865 new_blkg
= blkg_alloc(pos
, disk
, GFP_KERNEL
);
866 if (unlikely(!new_blkg
)) {
868 goto fail_exit_queue
;
871 if (radix_tree_preload(GFP_KERNEL
)) {
874 goto fail_exit_queue
;
877 spin_lock_irq(&q
->queue_lock
);
879 if (!blkcg_policy_enabled(q
, pol
)) {
885 blkg
= blkg_lookup(pos
, q
);
889 blkg
= blkg_create(pos
, disk
, new_blkg
);
896 radix_tree_preload_end();
907 radix_tree_preload_end();
909 spin_unlock_irq(&q
->queue_lock
);
914 * If queue was bypassing, we should retry. Do so after a
915 * short msleep(). It isn't strictly necessary but queue
916 * can be bypassing for some time and it's always nice to
917 * avoid busy looping.
921 ret
= restart_syscall();
925 EXPORT_SYMBOL_GPL(blkg_conf_prep
);
928 * blkg_conf_exit - clean up per-blkg config update
929 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
931 * Clean up after per-blkg config update. This function must be called on all
932 * blkg_conf_ctx's initialized with blkg_conf_init().
934 void blkg_conf_exit(struct blkg_conf_ctx
*ctx
)
935 __releases(&ctx
->bdev
->bd_queue
->queue_lock
)
936 __releases(&ctx
->bdev
->bd_queue
->rq_qos_mutex
)
939 spin_unlock_irq(&bdev_get_queue(ctx
->bdev
)->queue_lock
);
944 mutex_unlock(&ctx
->bdev
->bd_queue
->rq_qos_mutex
);
945 blkdev_put_no_open(ctx
->bdev
);
950 EXPORT_SYMBOL_GPL(blkg_conf_exit
);
952 static void blkg_iostat_set(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
956 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
957 dst
->bytes
[i
] = src
->bytes
[i
];
958 dst
->ios
[i
] = src
->ios
[i
];
962 static void blkg_iostat_add(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
966 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
967 dst
->bytes
[i
] += src
->bytes
[i
];
968 dst
->ios
[i
] += src
->ios
[i
];
972 static void blkg_iostat_sub(struct blkg_iostat
*dst
, struct blkg_iostat
*src
)
976 for (i
= 0; i
< BLKG_IOSTAT_NR
; i
++) {
977 dst
->bytes
[i
] -= src
->bytes
[i
];
978 dst
->ios
[i
] -= src
->ios
[i
];
982 static void blkcg_iostat_update(struct blkcg_gq
*blkg
, struct blkg_iostat
*cur
,
983 struct blkg_iostat
*last
)
985 struct blkg_iostat delta
;
988 /* propagate percpu delta to global */
989 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
990 blkg_iostat_set(&delta
, cur
);
991 blkg_iostat_sub(&delta
, last
);
992 blkg_iostat_add(&blkg
->iostat
.cur
, &delta
);
993 blkg_iostat_add(last
, &delta
);
994 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
997 static void __blkcg_rstat_flush(struct blkcg
*blkcg
, int cpu
)
999 struct llist_head
*lhead
= per_cpu_ptr(blkcg
->lhead
, cpu
);
1000 struct llist_node
*lnode
;
1001 struct blkg_iostat_set
*bisc
, *next_bisc
;
1002 unsigned long flags
;
1006 lnode
= llist_del_all(lhead
);
1011 * For covering concurrent parent blkg update from blkg_release().
1013 * When flushing from cgroup, cgroup_rstat_lock is always held, so
1014 * this lock won't cause contention most of time.
1016 raw_spin_lock_irqsave(&blkg_stat_lock
, flags
);
1019 * Iterate only the iostat_cpu's queued in the lockless list.
1021 llist_for_each_entry_safe(bisc
, next_bisc
, lnode
, lnode
) {
1022 struct blkcg_gq
*blkg
= bisc
->blkg
;
1023 struct blkcg_gq
*parent
= blkg
->parent
;
1024 struct blkg_iostat cur
;
1027 WRITE_ONCE(bisc
->lqueued
, false);
1029 /* fetch the current per-cpu values */
1031 seq
= u64_stats_fetch_begin(&bisc
->sync
);
1032 blkg_iostat_set(&cur
, &bisc
->cur
);
1033 } while (u64_stats_fetch_retry(&bisc
->sync
, seq
));
1035 blkcg_iostat_update(blkg
, &cur
, &bisc
->last
);
1037 /* propagate global delta to parent (unless that's root) */
1038 if (parent
&& parent
->parent
)
1039 blkcg_iostat_update(parent
, &blkg
->iostat
.cur
,
1040 &blkg
->iostat
.last
);
1042 raw_spin_unlock_irqrestore(&blkg_stat_lock
, flags
);
1047 static void blkcg_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
1049 /* Root-level stats are sourced from system-wide IO stats */
1050 if (cgroup_parent(css
->cgroup
))
1051 __blkcg_rstat_flush(css_to_blkcg(css
), cpu
);
1055 * We source root cgroup stats from the system-wide stats to avoid
1056 * tracking the same information twice and incurring overhead when no
1057 * cgroups are defined. For that reason, cgroup_rstat_flush in
1058 * blkcg_print_stat does not actually fill out the iostat in the root
1059 * cgroup's blkcg_gq.
1061 * However, we would like to re-use the printing code between the root and
1062 * non-root cgroups to the extent possible. For that reason, we simulate
1063 * flushing the root cgroup's stats by explicitly filling in the iostat
1064 * with disk level statistics.
1066 static void blkcg_fill_root_iostats(void)
1068 struct class_dev_iter iter
;
1071 class_dev_iter_init(&iter
, &block_class
, NULL
, &disk_type
);
1072 while ((dev
= class_dev_iter_next(&iter
))) {
1073 struct block_device
*bdev
= dev_to_bdev(dev
);
1074 struct blkcg_gq
*blkg
= bdev
->bd_disk
->queue
->root_blkg
;
1075 struct blkg_iostat tmp
;
1077 unsigned long flags
;
1079 memset(&tmp
, 0, sizeof(tmp
));
1080 for_each_possible_cpu(cpu
) {
1081 struct disk_stats
*cpu_dkstats
;
1083 cpu_dkstats
= per_cpu_ptr(bdev
->bd_stats
, cpu
);
1084 tmp
.ios
[BLKG_IOSTAT_READ
] +=
1085 cpu_dkstats
->ios
[STAT_READ
];
1086 tmp
.ios
[BLKG_IOSTAT_WRITE
] +=
1087 cpu_dkstats
->ios
[STAT_WRITE
];
1088 tmp
.ios
[BLKG_IOSTAT_DISCARD
] +=
1089 cpu_dkstats
->ios
[STAT_DISCARD
];
1090 // convert sectors to bytes
1091 tmp
.bytes
[BLKG_IOSTAT_READ
] +=
1092 cpu_dkstats
->sectors
[STAT_READ
] << 9;
1093 tmp
.bytes
[BLKG_IOSTAT_WRITE
] +=
1094 cpu_dkstats
->sectors
[STAT_WRITE
] << 9;
1095 tmp
.bytes
[BLKG_IOSTAT_DISCARD
] +=
1096 cpu_dkstats
->sectors
[STAT_DISCARD
] << 9;
1099 flags
= u64_stats_update_begin_irqsave(&blkg
->iostat
.sync
);
1100 blkg_iostat_set(&blkg
->iostat
.cur
, &tmp
);
1101 u64_stats_update_end_irqrestore(&blkg
->iostat
.sync
, flags
);
1105 static void blkcg_print_one_stat(struct blkcg_gq
*blkg
, struct seq_file
*s
)
1107 struct blkg_iostat_set
*bis
= &blkg
->iostat
;
1108 u64 rbytes
, wbytes
, rios
, wios
, dbytes
, dios
;
1116 dname
= blkg_dev_name(blkg
);
1120 seq_printf(s
, "%s ", dname
);
1123 seq
= u64_stats_fetch_begin(&bis
->sync
);
1125 rbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_READ
];
1126 wbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_WRITE
];
1127 dbytes
= bis
->cur
.bytes
[BLKG_IOSTAT_DISCARD
];
1128 rios
= bis
->cur
.ios
[BLKG_IOSTAT_READ
];
1129 wios
= bis
->cur
.ios
[BLKG_IOSTAT_WRITE
];
1130 dios
= bis
->cur
.ios
[BLKG_IOSTAT_DISCARD
];
1131 } while (u64_stats_fetch_retry(&bis
->sync
, seq
));
1133 if (rbytes
|| wbytes
|| rios
|| wios
) {
1134 seq_printf(s
, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1135 rbytes
, wbytes
, rios
, wios
,
1139 if (blkcg_debug_stats
&& atomic_read(&blkg
->use_delay
)) {
1140 seq_printf(s
, " use_delay=%d delay_nsec=%llu",
1141 atomic_read(&blkg
->use_delay
),
1142 atomic64_read(&blkg
->delay_nsec
));
1145 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1146 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1148 if (!blkg
->pd
[i
] || !pol
->pd_stat_fn
)
1151 pol
->pd_stat_fn(blkg
->pd
[i
], s
);
1157 static int blkcg_print_stat(struct seq_file
*sf
, void *v
)
1159 struct blkcg
*blkcg
= css_to_blkcg(seq_css(sf
));
1160 struct blkcg_gq
*blkg
;
1162 if (!seq_css(sf
)->parent
)
1163 blkcg_fill_root_iostats();
1165 cgroup_rstat_flush(blkcg
->css
.cgroup
);
1168 hlist_for_each_entry_rcu(blkg
, &blkcg
->blkg_list
, blkcg_node
) {
1169 spin_lock_irq(&blkg
->q
->queue_lock
);
1170 blkcg_print_one_stat(blkg
, sf
);
1171 spin_unlock_irq(&blkg
->q
->queue_lock
);
1177 static struct cftype blkcg_files
[] = {
1180 .seq_show
= blkcg_print_stat
,
1185 static struct cftype blkcg_legacy_files
[] = {
1187 .name
= "reset_stats",
1188 .write_u64
= blkcg_reset_stats
,
1193 #ifdef CONFIG_CGROUP_WRITEBACK
1194 struct list_head
*blkcg_get_cgwb_list(struct cgroup_subsys_state
*css
)
1196 return &css_to_blkcg(css
)->cgwb_list
;
1201 * blkcg destruction is a three-stage process.
1203 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1204 * which offlines writeback. Here we tie the next stage of blkg destruction
1205 * to the completion of writeback associated with the blkcg. This lets us
1206 * avoid punting potentially large amounts of outstanding writeback to root
1207 * while maintaining any ongoing policies. The next stage is triggered when
1208 * the nr_cgwbs count goes to zero.
1210 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1211 * and handles the destruction of blkgs. Here the css reference held by
1212 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1213 * This work may occur in cgwb_release_workfn() on the cgwb_release
1214 * workqueue. Any submitted ios that fail to get the blkg ref will be
1215 * punted to the root_blkg.
1217 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1218 * This finally frees the blkcg.
1222 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1223 * @blkcg: blkcg of interest
1225 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1226 * is nested inside q lock, this function performs reverse double lock dancing.
1227 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1228 * blkcg_css_free to eventually be called.
1230 * This is the blkcg counterpart of ioc_release_fn().
1232 static void blkcg_destroy_blkgs(struct blkcg
*blkcg
)
1236 spin_lock_irq(&blkcg
->lock
);
1238 while (!hlist_empty(&blkcg
->blkg_list
)) {
1239 struct blkcg_gq
*blkg
= hlist_entry(blkcg
->blkg_list
.first
,
1240 struct blkcg_gq
, blkcg_node
);
1241 struct request_queue
*q
= blkg
->q
;
1243 if (need_resched() || !spin_trylock(&q
->queue_lock
)) {
1245 * Given that the system can accumulate a huge number
1246 * of blkgs in pathological cases, check to see if we
1247 * need to rescheduling to avoid softlockup.
1249 spin_unlock_irq(&blkcg
->lock
);
1251 spin_lock_irq(&blkcg
->lock
);
1256 spin_unlock(&q
->queue_lock
);
1259 spin_unlock_irq(&blkcg
->lock
);
1263 * blkcg_pin_online - pin online state
1264 * @blkcg_css: blkcg of interest
1266 * While pinned, a blkcg is kept online. This is primarily used to
1267 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1268 * while an associated cgwb is still active.
1270 void blkcg_pin_online(struct cgroup_subsys_state
*blkcg_css
)
1272 refcount_inc(&css_to_blkcg(blkcg_css
)->online_pin
);
1276 * blkcg_unpin_online - unpin online state
1277 * @blkcg_css: blkcg of interest
1279 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1280 * that blkg doesn't go offline while an associated cgwb is still active.
1281 * When this count goes to zero, all active cgwbs have finished so the
1282 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1284 void blkcg_unpin_online(struct cgroup_subsys_state
*blkcg_css
)
1286 struct blkcg
*blkcg
= css_to_blkcg(blkcg_css
);
1289 if (!refcount_dec_and_test(&blkcg
->online_pin
))
1291 blkcg_destroy_blkgs(blkcg
);
1292 blkcg
= blkcg_parent(blkcg
);
1297 * blkcg_css_offline - cgroup css_offline callback
1298 * @css: css of interest
1300 * This function is called when @css is about to go away. Here the cgwbs are
1301 * offlined first and only once writeback associated with the blkcg has
1302 * finished do we start step 2 (see above).
1304 static void blkcg_css_offline(struct cgroup_subsys_state
*css
)
1306 /* this prevents anyone from attaching or migrating to this blkcg */
1307 wb_blkcg_offline(css
);
1309 /* put the base online pin allowing step 2 to be triggered */
1310 blkcg_unpin_online(css
);
1313 static void blkcg_css_free(struct cgroup_subsys_state
*css
)
1315 struct blkcg
*blkcg
= css_to_blkcg(css
);
1318 mutex_lock(&blkcg_pol_mutex
);
1320 list_del(&blkcg
->all_blkcgs_node
);
1322 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1324 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1326 mutex_unlock(&blkcg_pol_mutex
);
1328 free_percpu(blkcg
->lhead
);
1332 static struct cgroup_subsys_state
*
1333 blkcg_css_alloc(struct cgroup_subsys_state
*parent_css
)
1335 struct blkcg
*blkcg
;
1338 mutex_lock(&blkcg_pol_mutex
);
1341 blkcg
= &blkcg_root
;
1343 blkcg
= kzalloc(sizeof(*blkcg
), GFP_KERNEL
);
1348 if (init_blkcg_llists(blkcg
))
1351 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++) {
1352 struct blkcg_policy
*pol
= blkcg_policy
[i
];
1353 struct blkcg_policy_data
*cpd
;
1356 * If the policy hasn't been attached yet, wait for it
1357 * to be attached before doing anything else. Otherwise,
1358 * check if the policy requires any specific per-cgroup
1359 * data: if it does, allocate and initialize it.
1361 if (!pol
|| !pol
->cpd_alloc_fn
)
1364 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1368 blkcg
->cpd
[i
] = cpd
;
1373 spin_lock_init(&blkcg
->lock
);
1374 refcount_set(&blkcg
->online_pin
, 1);
1375 INIT_RADIX_TREE(&blkcg
->blkg_tree
, GFP_NOWAIT
| __GFP_NOWARN
);
1376 INIT_HLIST_HEAD(&blkcg
->blkg_list
);
1377 #ifdef CONFIG_CGROUP_WRITEBACK
1378 INIT_LIST_HEAD(&blkcg
->cgwb_list
);
1380 list_add_tail(&blkcg
->all_blkcgs_node
, &all_blkcgs
);
1382 mutex_unlock(&blkcg_pol_mutex
);
1386 for (i
--; i
>= 0; i
--)
1388 blkcg_policy
[i
]->cpd_free_fn(blkcg
->cpd
[i
]);
1389 free_percpu(blkcg
->lhead
);
1391 if (blkcg
!= &blkcg_root
)
1394 mutex_unlock(&blkcg_pol_mutex
);
1395 return ERR_PTR(-ENOMEM
);
1398 static int blkcg_css_online(struct cgroup_subsys_state
*css
)
1400 struct blkcg
*parent
= blkcg_parent(css_to_blkcg(css
));
1403 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1404 * don't go offline while cgwbs are still active on them. Pin the
1405 * parent so that offline always happens towards the root.
1408 blkcg_pin_online(&parent
->css
);
1412 int blkcg_init_disk(struct gendisk
*disk
)
1414 struct request_queue
*q
= disk
->queue
;
1415 struct blkcg_gq
*new_blkg
, *blkg
;
1419 INIT_LIST_HEAD(&q
->blkg_list
);
1420 mutex_init(&q
->blkcg_mutex
);
1422 new_blkg
= blkg_alloc(&blkcg_root
, disk
, GFP_KERNEL
);
1426 preloaded
= !radix_tree_preload(GFP_KERNEL
);
1428 /* Make sure the root blkg exists. */
1429 /* spin_lock_irq can serve as RCU read-side critical section. */
1430 spin_lock_irq(&q
->queue_lock
);
1431 blkg
= blkg_create(&blkcg_root
, disk
, new_blkg
);
1434 q
->root_blkg
= blkg
;
1435 spin_unlock_irq(&q
->queue_lock
);
1438 radix_tree_preload_end();
1440 ret
= blk_ioprio_init(disk
);
1442 goto err_destroy_all
;
1444 ret
= blk_throtl_init(disk
);
1446 goto err_ioprio_exit
;
1451 blk_ioprio_exit(disk
);
1453 blkg_destroy_all(disk
);
1456 spin_unlock_irq(&q
->queue_lock
);
1458 radix_tree_preload_end();
1459 return PTR_ERR(blkg
);
1462 void blkcg_exit_disk(struct gendisk
*disk
)
1464 blkg_destroy_all(disk
);
1465 blk_throtl_exit(disk
);
1468 static void blkcg_exit(struct task_struct
*tsk
)
1470 if (tsk
->throttle_disk
)
1471 put_disk(tsk
->throttle_disk
);
1472 tsk
->throttle_disk
= NULL
;
1475 struct cgroup_subsys io_cgrp_subsys
= {
1476 .css_alloc
= blkcg_css_alloc
,
1477 .css_online
= blkcg_css_online
,
1478 .css_offline
= blkcg_css_offline
,
1479 .css_free
= blkcg_css_free
,
1480 .css_rstat_flush
= blkcg_rstat_flush
,
1481 .dfl_cftypes
= blkcg_files
,
1482 .legacy_cftypes
= blkcg_legacy_files
,
1483 .legacy_name
= "blkio",
1487 * This ensures that, if available, memcg is automatically enabled
1488 * together on the default hierarchy so that the owner cgroup can
1489 * be retrieved from writeback pages.
1491 .depends_on
= 1 << memory_cgrp_id
,
1494 EXPORT_SYMBOL_GPL(io_cgrp_subsys
);
1497 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1498 * @disk: gendisk of interest
1499 * @pol: blkcg policy to activate
1501 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1502 * bypass mode to populate its blkgs with policy_data for @pol.
1504 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1505 * from IO path. Update of each blkg is protected by both queue and blkcg
1506 * locks so that holding either lock and testing blkcg_policy_enabled() is
1507 * always enough for dereferencing policy data.
1509 * The caller is responsible for synchronizing [de]activations and policy
1510 * [un]registerations. Returns 0 on success, -errno on failure.
1512 int blkcg_activate_policy(struct gendisk
*disk
, const struct blkcg_policy
*pol
)
1514 struct request_queue
*q
= disk
->queue
;
1515 struct blkg_policy_data
*pd_prealloc
= NULL
;
1516 struct blkcg_gq
*blkg
, *pinned_blkg
= NULL
;
1519 if (blkcg_policy_enabled(q
, pol
))
1523 blk_mq_freeze_queue(q
);
1525 spin_lock_irq(&q
->queue_lock
);
1527 /* blkg_list is pushed at the head, reverse walk to initialize parents first */
1528 list_for_each_entry_reverse(blkg
, &q
->blkg_list
, q_node
) {
1529 struct blkg_policy_data
*pd
;
1531 if (blkg
->pd
[pol
->plid
])
1534 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1535 if (blkg
== pinned_blkg
) {
1539 pd
= pol
->pd_alloc_fn(disk
, blkg
->blkcg
,
1540 GFP_NOWAIT
| __GFP_NOWARN
);
1545 * GFP_NOWAIT failed. Free the existing one and
1546 * prealloc for @blkg w/ GFP_KERNEL.
1549 blkg_put(pinned_blkg
);
1553 spin_unlock_irq(&q
->queue_lock
);
1556 pol
->pd_free_fn(pd_prealloc
);
1557 pd_prealloc
= pol
->pd_alloc_fn(disk
, blkg
->blkcg
,
1565 spin_lock(&blkg
->blkcg
->lock
);
1568 pd
->plid
= pol
->plid
;
1569 blkg
->pd
[pol
->plid
] = pd
;
1571 if (pol
->pd_init_fn
)
1572 pol
->pd_init_fn(pd
);
1574 if (pol
->pd_online_fn
)
1575 pol
->pd_online_fn(pd
);
1578 spin_unlock(&blkg
->blkcg
->lock
);
1581 __set_bit(pol
->plid
, q
->blkcg_pols
);
1584 spin_unlock_irq(&q
->queue_lock
);
1587 blk_mq_unfreeze_queue(q
);
1589 blkg_put(pinned_blkg
);
1591 pol
->pd_free_fn(pd_prealloc
);
1595 /* alloc failed, take down everything */
1596 spin_lock_irq(&q
->queue_lock
);
1597 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1598 struct blkcg
*blkcg
= blkg
->blkcg
;
1599 struct blkg_policy_data
*pd
;
1601 spin_lock(&blkcg
->lock
);
1602 pd
= blkg
->pd
[pol
->plid
];
1604 if (pd
->online
&& pol
->pd_offline_fn
)
1605 pol
->pd_offline_fn(pd
);
1607 pol
->pd_free_fn(pd
);
1608 blkg
->pd
[pol
->plid
] = NULL
;
1610 spin_unlock(&blkcg
->lock
);
1612 spin_unlock_irq(&q
->queue_lock
);
1616 EXPORT_SYMBOL_GPL(blkcg_activate_policy
);
1619 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1620 * @disk: gendisk of interest
1621 * @pol: blkcg policy to deactivate
1623 * Deactivate @pol on @disk. Follows the same synchronization rules as
1624 * blkcg_activate_policy().
1626 void blkcg_deactivate_policy(struct gendisk
*disk
,
1627 const struct blkcg_policy
*pol
)
1629 struct request_queue
*q
= disk
->queue
;
1630 struct blkcg_gq
*blkg
;
1632 if (!blkcg_policy_enabled(q
, pol
))
1636 blk_mq_freeze_queue(q
);
1638 mutex_lock(&q
->blkcg_mutex
);
1639 spin_lock_irq(&q
->queue_lock
);
1641 __clear_bit(pol
->plid
, q
->blkcg_pols
);
1643 list_for_each_entry(blkg
, &q
->blkg_list
, q_node
) {
1644 struct blkcg
*blkcg
= blkg
->blkcg
;
1646 spin_lock(&blkcg
->lock
);
1647 if (blkg
->pd
[pol
->plid
]) {
1648 if (blkg
->pd
[pol
->plid
]->online
&& pol
->pd_offline_fn
)
1649 pol
->pd_offline_fn(blkg
->pd
[pol
->plid
]);
1650 pol
->pd_free_fn(blkg
->pd
[pol
->plid
]);
1651 blkg
->pd
[pol
->plid
] = NULL
;
1653 spin_unlock(&blkcg
->lock
);
1656 spin_unlock_irq(&q
->queue_lock
);
1657 mutex_unlock(&q
->blkcg_mutex
);
1660 blk_mq_unfreeze_queue(q
);
1662 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy
);
1664 static void blkcg_free_all_cpd(struct blkcg_policy
*pol
)
1666 struct blkcg
*blkcg
;
1668 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1669 if (blkcg
->cpd
[pol
->plid
]) {
1670 pol
->cpd_free_fn(blkcg
->cpd
[pol
->plid
]);
1671 blkcg
->cpd
[pol
->plid
] = NULL
;
1677 * blkcg_policy_register - register a blkcg policy
1678 * @pol: blkcg policy to register
1680 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1681 * successful registration. Returns 0 on success and -errno on failure.
1683 int blkcg_policy_register(struct blkcg_policy
*pol
)
1685 struct blkcg
*blkcg
;
1688 mutex_lock(&blkcg_pol_register_mutex
);
1689 mutex_lock(&blkcg_pol_mutex
);
1691 /* find an empty slot */
1693 for (i
= 0; i
< BLKCG_MAX_POLS
; i
++)
1694 if (!blkcg_policy
[i
])
1696 if (i
>= BLKCG_MAX_POLS
) {
1697 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1701 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1702 if ((!pol
->cpd_alloc_fn
^ !pol
->cpd_free_fn
) ||
1703 (!pol
->pd_alloc_fn
^ !pol
->pd_free_fn
))
1708 blkcg_policy
[pol
->plid
] = pol
;
1710 /* allocate and install cpd's */
1711 if (pol
->cpd_alloc_fn
) {
1712 list_for_each_entry(blkcg
, &all_blkcgs
, all_blkcgs_node
) {
1713 struct blkcg_policy_data
*cpd
;
1715 cpd
= pol
->cpd_alloc_fn(GFP_KERNEL
);
1719 blkcg
->cpd
[pol
->plid
] = cpd
;
1721 cpd
->plid
= pol
->plid
;
1725 mutex_unlock(&blkcg_pol_mutex
);
1727 /* everything is in place, add intf files for the new policy */
1728 if (pol
->dfl_cftypes
)
1729 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys
,
1731 if (pol
->legacy_cftypes
)
1732 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys
,
1733 pol
->legacy_cftypes
));
1734 mutex_unlock(&blkcg_pol_register_mutex
);
1738 if (pol
->cpd_free_fn
)
1739 blkcg_free_all_cpd(pol
);
1741 blkcg_policy
[pol
->plid
] = NULL
;
1743 mutex_unlock(&blkcg_pol_mutex
);
1744 mutex_unlock(&blkcg_pol_register_mutex
);
1747 EXPORT_SYMBOL_GPL(blkcg_policy_register
);
1750 * blkcg_policy_unregister - unregister a blkcg policy
1751 * @pol: blkcg policy to unregister
1753 * Undo blkcg_policy_register(@pol). Might sleep.
1755 void blkcg_policy_unregister(struct blkcg_policy
*pol
)
1757 mutex_lock(&blkcg_pol_register_mutex
);
1759 if (WARN_ON(blkcg_policy
[pol
->plid
] != pol
))
1762 /* kill the intf files first */
1763 if (pol
->dfl_cftypes
)
1764 cgroup_rm_cftypes(pol
->dfl_cftypes
);
1765 if (pol
->legacy_cftypes
)
1766 cgroup_rm_cftypes(pol
->legacy_cftypes
);
1768 /* remove cpds and unregister */
1769 mutex_lock(&blkcg_pol_mutex
);
1771 if (pol
->cpd_free_fn
)
1772 blkcg_free_all_cpd(pol
);
1774 blkcg_policy
[pol
->plid
] = NULL
;
1776 mutex_unlock(&blkcg_pol_mutex
);
1778 mutex_unlock(&blkcg_pol_register_mutex
);
1780 EXPORT_SYMBOL_GPL(blkcg_policy_unregister
);
1783 * Scale the accumulated delay based on how long it has been since we updated
1784 * the delay. We only call this when we are adding delay, in case it's been a
1785 * while since we added delay, and when we are checking to see if we need to
1786 * delay a task, to account for any delays that may have occurred.
1788 static void blkcg_scale_delay(struct blkcg_gq
*blkg
, u64 now
)
1790 u64 old
= atomic64_read(&blkg
->delay_start
);
1792 /* negative use_delay means no scaling, see blkcg_set_delay() */
1793 if (atomic_read(&blkg
->use_delay
) < 0)
1797 * We only want to scale down every second. The idea here is that we
1798 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1799 * time window. We only want to throttle tasks for recent delay that
1800 * has occurred, in 1 second time windows since that's the maximum
1801 * things can be throttled. We save the current delay window in
1802 * blkg->last_delay so we know what amount is still left to be charged
1803 * to the blkg from this point onward. blkg->last_use keeps track of
1804 * the use_delay counter. The idea is if we're unthrottling the blkg we
1805 * are ok with whatever is happening now, and we can take away more of
1806 * the accumulated delay as we've already throttled enough that
1807 * everybody is happy with their IO latencies.
1809 if (time_before64(old
+ NSEC_PER_SEC
, now
) &&
1810 atomic64_try_cmpxchg(&blkg
->delay_start
, &old
, now
)) {
1811 u64 cur
= atomic64_read(&blkg
->delay_nsec
);
1812 u64 sub
= min_t(u64
, blkg
->last_delay
, now
- old
);
1813 int cur_use
= atomic_read(&blkg
->use_delay
);
1816 * We've been unthrottled, subtract a larger chunk of our
1817 * accumulated delay.
1819 if (cur_use
< blkg
->last_use
)
1820 sub
= max_t(u64
, sub
, blkg
->last_delay
>> 1);
1823 * This shouldn't happen, but handle it anyway. Our delay_nsec
1824 * should only ever be growing except here where we subtract out
1825 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1826 * rather not end up with negative numbers.
1828 if (unlikely(cur
< sub
)) {
1829 atomic64_set(&blkg
->delay_nsec
, 0);
1830 blkg
->last_delay
= 0;
1832 atomic64_sub(sub
, &blkg
->delay_nsec
);
1833 blkg
->last_delay
= cur
- sub
;
1835 blkg
->last_use
= cur_use
;
1840 * This is called when we want to actually walk up the hierarchy and check to
1841 * see if we need to throttle, and then actually throttle if there is some
1842 * accumulated delay. This should only be called upon return to user space so
1843 * we're not holding some lock that would induce a priority inversion.
1845 static void blkcg_maybe_throttle_blkg(struct blkcg_gq
*blkg
, bool use_memdelay
)
1847 unsigned long pflags
;
1849 u64 now
= ktime_to_ns(ktime_get());
1854 while (blkg
->parent
) {
1855 int use_delay
= atomic_read(&blkg
->use_delay
);
1860 blkcg_scale_delay(blkg
, now
);
1861 this_delay
= atomic64_read(&blkg
->delay_nsec
);
1862 if (this_delay
> delay_nsec
) {
1863 delay_nsec
= this_delay
;
1864 clamp
= use_delay
> 0;
1867 blkg
= blkg
->parent
;
1874 * Let's not sleep for all eternity if we've amassed a huge delay.
1875 * Swapping or metadata IO can accumulate 10's of seconds worth of
1876 * delay, and we want userspace to be able to do _something_ so cap the
1877 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1878 * tasks will be delayed for 0.25 second for every syscall. If
1879 * blkcg_set_delay() was used as indicated by negative use_delay, the
1880 * caller is responsible for regulating the range.
1883 delay_nsec
= min_t(u64
, delay_nsec
, 250 * NSEC_PER_MSEC
);
1886 psi_memstall_enter(&pflags
);
1888 exp
= ktime_add_ns(now
, delay_nsec
);
1889 tok
= io_schedule_prepare();
1891 __set_current_state(TASK_KILLABLE
);
1892 if (!schedule_hrtimeout(&exp
, HRTIMER_MODE_ABS
))
1894 } while (!fatal_signal_pending(current
));
1895 io_schedule_finish(tok
);
1898 psi_memstall_leave(&pflags
);
1902 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1904 * This is only called if we've been marked with set_notify_resume(). Obviously
1905 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1906 * check to see if current->throttle_disk is set and if not this doesn't do
1907 * anything. This should only ever be called by the resume code, it's not meant
1908 * to be called by people willy-nilly as it will actually do the work to
1909 * throttle the task if it is setup for throttling.
1911 void blkcg_maybe_throttle_current(void)
1913 struct gendisk
*disk
= current
->throttle_disk
;
1914 struct blkcg
*blkcg
;
1915 struct blkcg_gq
*blkg
;
1916 bool use_memdelay
= current
->use_memdelay
;
1921 current
->throttle_disk
= NULL
;
1922 current
->use_memdelay
= false;
1925 blkcg
= css_to_blkcg(blkcg_css());
1928 blkg
= blkg_lookup(blkcg
, disk
->queue
);
1931 if (!blkg_tryget(blkg
))
1935 blkcg_maybe_throttle_blkg(blkg
, use_memdelay
);
1944 * blkcg_schedule_throttle - this task needs to check for throttling
1945 * @disk: disk to throttle
1946 * @use_memdelay: do we charge this to memory delay for PSI
1948 * This is called by the IO controller when we know there's delay accumulated
1949 * for the blkg for this task. We do not pass the blkg because there are places
1950 * we call this that may not have that information, the swapping code for
1951 * instance will only have a block_device at that point. This set's the
1952 * notify_resume for the task to check and see if it requires throttling before
1953 * returning to user space.
1955 * We will only schedule once per syscall. You can call this over and over
1956 * again and it will only do the check once upon return to user space, and only
1957 * throttle once. If the task needs to be throttled again it'll need to be
1958 * re-set at the next time we see the task.
1960 void blkcg_schedule_throttle(struct gendisk
*disk
, bool use_memdelay
)
1962 if (unlikely(current
->flags
& PF_KTHREAD
))
1965 if (current
->throttle_disk
!= disk
) {
1966 if (test_bit(GD_DEAD
, &disk
->state
))
1968 get_device(disk_to_dev(disk
));
1970 if (current
->throttle_disk
)
1971 put_disk(current
->throttle_disk
);
1972 current
->throttle_disk
= disk
;
1976 current
->use_memdelay
= use_memdelay
;
1977 set_notify_resume(current
);
1981 * blkcg_add_delay - add delay to this blkg
1982 * @blkg: blkg of interest
1983 * @now: the current time in nanoseconds
1984 * @delta: how many nanoseconds of delay to add
1986 * Charge @delta to the blkg's current delay accumulation. This is used to
1987 * throttle tasks if an IO controller thinks we need more throttling.
1989 void blkcg_add_delay(struct blkcg_gq
*blkg
, u64 now
, u64 delta
)
1991 if (WARN_ON_ONCE(atomic_read(&blkg
->use_delay
) < 0))
1993 blkcg_scale_delay(blkg
, now
);
1994 atomic64_add(delta
, &blkg
->delay_nsec
);
1998 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
2002 * As the failure mode here is to walk up the blkg tree, this ensure that the
2003 * blkg->parent pointers are always valid. This returns the blkg that it ended
2004 * up taking a reference on or %NULL if no reference was taken.
2006 static inline struct blkcg_gq
*blkg_tryget_closest(struct bio
*bio
,
2007 struct cgroup_subsys_state
*css
)
2009 struct blkcg_gq
*blkg
, *ret_blkg
= NULL
;
2012 blkg
= blkg_lookup_create(css_to_blkcg(css
), bio
->bi_bdev
->bd_disk
);
2014 if (blkg_tryget(blkg
)) {
2018 blkg
= blkg
->parent
;
2026 * bio_associate_blkg_from_css - associate a bio with a specified css
2030 * Associate @bio with the blkg found by combining the css's blkg and the
2031 * request_queue of the @bio. An association failure is handled by walking up
2032 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
2033 * and q->root_blkg. This situation only happens when a cgroup is dying and
2034 * then the remaining bios will spill to the closest alive blkg.
2036 * A reference will be taken on the blkg and will be released when @bio is
2039 void bio_associate_blkg_from_css(struct bio
*bio
,
2040 struct cgroup_subsys_state
*css
)
2043 blkg_put(bio
->bi_blkg
);
2045 if (css
&& css
->parent
) {
2046 bio
->bi_blkg
= blkg_tryget_closest(bio
, css
);
2048 blkg_get(bdev_get_queue(bio
->bi_bdev
)->root_blkg
);
2049 bio
->bi_blkg
= bdev_get_queue(bio
->bi_bdev
)->root_blkg
;
2052 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css
);
2055 * bio_associate_blkg - associate a bio with a blkg
2058 * Associate @bio with the blkg found from the bio's css and request_queue.
2059 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2060 * already associated, the css is reused and association redone as the
2061 * request_queue may have changed.
2063 void bio_associate_blkg(struct bio
*bio
)
2065 struct cgroup_subsys_state
*css
;
2070 css
= bio_blkcg_css(bio
);
2074 bio_associate_blkg_from_css(bio
, css
);
2078 EXPORT_SYMBOL_GPL(bio_associate_blkg
);
2081 * bio_clone_blkg_association - clone blkg association from src to dst bio
2082 * @dst: destination bio
2085 void bio_clone_blkg_association(struct bio
*dst
, struct bio
*src
)
2088 bio_associate_blkg_from_css(dst
, bio_blkcg_css(src
));
2090 EXPORT_SYMBOL_GPL(bio_clone_blkg_association
);
2092 static int blk_cgroup_io_type(struct bio
*bio
)
2094 if (op_is_discard(bio
->bi_opf
))
2095 return BLKG_IOSTAT_DISCARD
;
2096 if (op_is_write(bio
->bi_opf
))
2097 return BLKG_IOSTAT_WRITE
;
2098 return BLKG_IOSTAT_READ
;
2101 void blk_cgroup_bio_start(struct bio
*bio
)
2103 struct blkcg
*blkcg
= bio
->bi_blkg
->blkcg
;
2104 int rwd
= blk_cgroup_io_type(bio
), cpu
;
2105 struct blkg_iostat_set
*bis
;
2106 unsigned long flags
;
2108 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
))
2111 /* Root-level stats are sourced from system-wide IO stats */
2112 if (!cgroup_parent(blkcg
->css
.cgroup
))
2116 bis
= per_cpu_ptr(bio
->bi_blkg
->iostat_cpu
, cpu
);
2117 flags
= u64_stats_update_begin_irqsave(&bis
->sync
);
2120 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2121 * bio and we would have already accounted for the size of the bio.
2123 if (!bio_flagged(bio
, BIO_CGROUP_ACCT
)) {
2124 bio_set_flag(bio
, BIO_CGROUP_ACCT
);
2125 bis
->cur
.bytes
[rwd
] += bio
->bi_iter
.bi_size
;
2127 bis
->cur
.ios
[rwd
]++;
2130 * If the iostat_cpu isn't in a lockless list, put it into the
2131 * list to indicate that a stat update is pending.
2133 if (!READ_ONCE(bis
->lqueued
)) {
2134 struct llist_head
*lhead
= this_cpu_ptr(blkcg
->lhead
);
2136 llist_add(&bis
->lnode
, lhead
);
2137 WRITE_ONCE(bis
->lqueued
, true);
2140 u64_stats_update_end_irqrestore(&bis
->sync
, flags
);
2141 cgroup_rstat_updated(blkcg
->css
.cgroup
, cpu
);
2145 bool blk_cgroup_congested(void)
2147 struct cgroup_subsys_state
*css
;
2151 for (css
= blkcg_css(); css
; css
= css
->parent
) {
2152 if (atomic_read(&css
->cgroup
->congestion_count
)) {
2161 module_param(blkcg_debug_stats
, bool, 0644);
2162 MODULE_PARM_DESC(blkcg_debug_stats
, "True if you want debug stats, false if not");