blob: 582d2f18717ee3c1cd767332471a3612b19f35bc [file] [log] [blame]
Christoph Hellwig3dcf60bc2019-04-30 14:42:43 -04001// SPDX-License-Identifier: GPL-2.0
Vivek Goyal31e4c282009-12-03 12:59:42 -05002/*
3 * Common Block IO controller cgroup interface
4 *
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
Arianna Avanzinie48453c2015-06-05 23:38:42 +020013 *
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
Vivek Goyal31e4c282009-12-03 12:59:42 -050017 */
18#include <linux/ioprio.h>
Vivek Goyal22084192009-12-03 12:59:49 -050019#include <linux/kdev_t.h>
Vivek Goyal9d6a9862009-12-04 10:36:41 -050020#include <linux/module.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010021#include <linux/sched/signal.h>
Stephen Rothwellaccee782009-12-07 19:29:39 +110022#include <linux/err.h>
Divyesh Shah91952912010-04-01 15:01:41 -070023#include <linux/blkdev.h>
Tejun Heo52ebea72015-05-22 17:13:37 -040024#include <linux/backing-dev.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090025#include <linux/slab.h>
Gui Jianfeng34d0f172010-04-13 16:05:49 +080026#include <linux/genhd.h>
Tejun Heo72e06c22012-03-05 13:15:00 -080027#include <linux/delay.h>
Tejun Heo9a9e8a22012-03-19 15:10:56 -070028#include <linux/atomic.h>
Tejun Heo36aa9e52015-08-18 14:55:31 -070029#include <linux/ctype.h>
Tejun Heoeea8f412015-05-22 17:13:17 -040030#include <linux/blk-cgroup.h>
Josef Bacikd09d8df2018-07-03 11:14:55 -040031#include <linux/tracehook.h>
Josef Bacikfd112c72019-07-09 14:41:29 -070032#include <linux/psi.h>
Tejun Heo5efd6112012-03-05 13:15:12 -080033#include "blk.h"
Vivek Goyal3e252062009-12-04 10:36:42 -050034
Tejun Heo838f13b2015-07-09 16:39:47 -040035/*
36 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
37 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
38 * policy [un]register operations including cgroup file additions /
39 * removals. Putting cgroup file registration outside blkcg_pol_mutex
40 * allows grabbing it from cgroup callbacks.
41 */
42static DEFINE_MUTEX(blkcg_pol_register_mutex);
Tejun Heobc0d6502012-04-13 13:11:26 -070043static DEFINE_MUTEX(blkcg_pol_mutex);
Tejun Heo923adde2012-03-05 13:15:13 -080044
Arianna Avanzinie48453c2015-06-05 23:38:42 +020045struct blkcg blkcg_root;
Tejun Heo3c798392012-04-16 13:57:25 -070046EXPORT_SYMBOL_GPL(blkcg_root);
Vivek Goyal9d6a9862009-12-04 10:36:41 -050047
Tejun Heo496d5e72015-05-22 17:13:21 -040048struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
Tejun Heo9b0eb692019-06-27 13:39:48 -070049EXPORT_SYMBOL_GPL(blkcg_root_css);
Tejun Heo496d5e72015-05-22 17:13:21 -040050
Tejun Heo3c798392012-04-16 13:57:25 -070051static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
Tejun Heo035d10b2012-03-05 13:15:04 -080052
Tejun Heo7876f932015-07-09 16:39:49 -040053static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
54
Tejun Heo07b0fde2019-07-16 07:58:31 -070055bool blkcg_debug_stats = false;
Tejun Heod3f77df2019-06-27 13:39:52 -070056static struct workqueue_struct *blkcg_punt_bio_wq;
Josef Bacik903d23f2018-07-03 11:14:52 -040057
Tejun Heoa2b16932012-04-13 13:11:33 -070058static bool blkcg_policy_enabled(struct request_queue *q,
Tejun Heo3c798392012-04-16 13:57:25 -070059 const struct blkcg_policy *pol)
Tejun Heoa2b16932012-04-13 13:11:33 -070060{
61 return pol && test_bit(pol->plid, q->blkcg_pols);
62}
63
Tejun Heo03814112012-03-05 13:15:14 -080064/**
65 * blkg_free - free a blkg
66 * @blkg: blkg to free
67 *
68 * Free @blkg which may be partially allocated.
69 */
Tejun Heo3c798392012-04-16 13:57:25 -070070static void blkg_free(struct blkcg_gq *blkg)
Tejun Heo03814112012-03-05 13:15:14 -080071{
Tejun Heoe8989fa2012-03-05 13:15:20 -080072 int i;
Tejun Heo549d3aa2012-03-05 13:15:16 -080073
74 if (!blkg)
75 return;
76
Tejun Heodb613672013-05-14 13:52:31 -070077 for (i = 0; i < BLKCG_MAX_POLS; i++)
Tejun Heo001bea72015-08-18 14:55:11 -070078 if (blkg->pd[i])
79 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
Tejun Heoe8989fa2012-03-05 13:15:20 -080080
Tejun Heof7331642019-11-07 11:18:03 -080081 free_percpu(blkg->iostat_cpu);
Tejun Heoef069b92019-06-13 15:30:39 -070082 percpu_ref_exit(&blkg->refcnt);
Tejun Heo549d3aa2012-03-05 13:15:16 -080083 kfree(blkg);
Tejun Heo03814112012-03-05 13:15:14 -080084}
85
Dennis Zhou7fcf2b02018-12-05 12:10:38 -050086static void __blkg_release(struct rcu_head *rcu)
87{
88 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
89
Tejun Heod3f77df2019-06-27 13:39:52 -070090 WARN_ON(!bio_list_empty(&blkg->async_bios));
91
Dennis Zhou7fcf2b02018-12-05 12:10:38 -050092 /* release the blkcg and parent blkg refs this blkg has been holding */
93 css_put(&blkg->blkcg->css);
94 if (blkg->parent)
95 blkg_put(blkg->parent);
Dennis Zhou7fcf2b02018-12-05 12:10:38 -050096 blkg_free(blkg);
97}
98
99/*
100 * A group is RCU protected, but having an rcu lock does not mean that one
101 * can access all the fields of blkg and assume these are valid. For
102 * example, don't try to follow throtl_data and request queue links.
103 *
104 * Having a reference to blkg under an rcu allows accesses to only values
105 * local to groups like group stats and group rate limits.
106 */
107static void blkg_release(struct percpu_ref *ref)
108{
109 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
110
111 call_rcu(&blkg->rcu_head, __blkg_release);
112}
113
Tejun Heod3f77df2019-06-27 13:39:52 -0700114static void blkg_async_bio_workfn(struct work_struct *work)
115{
116 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
117 async_bio_work);
118 struct bio_list bios = BIO_EMPTY_LIST;
119 struct bio *bio;
Xianting Tian192f1c62020-09-10 14:15:06 +0800120 struct blk_plug plug;
121 bool need_plug = false;
Tejun Heod3f77df2019-06-27 13:39:52 -0700122
123 /* as long as there are pending bios, @blkg can't go away */
124 spin_lock_bh(&blkg->async_bio_lock);
125 bio_list_merge(&bios, &blkg->async_bios);
126 bio_list_init(&blkg->async_bios);
127 spin_unlock_bh(&blkg->async_bio_lock);
128
Xianting Tian192f1c62020-09-10 14:15:06 +0800129 /* start plug only when bio_list contains at least 2 bios */
130 if (bios.head && bios.head->bi_next) {
131 need_plug = true;
132 blk_start_plug(&plug);
133 }
Tejun Heod3f77df2019-06-27 13:39:52 -0700134 while ((bio = bio_list_pop(&bios)))
135 submit_bio(bio);
Xianting Tian192f1c62020-09-10 14:15:06 +0800136 if (need_plug)
137 blk_finish_plug(&plug);
Tejun Heod3f77df2019-06-27 13:39:52 -0700138}
139
Tejun Heo03814112012-03-05 13:15:14 -0800140/**
141 * blkg_alloc - allocate a blkg
142 * @blkcg: block cgroup the new blkg is associated with
143 * @q: request_queue the new blkg is associated with
Tejun Heo15974992012-06-04 20:40:52 -0700144 * @gfp_mask: allocation mask to use
Tejun Heo03814112012-03-05 13:15:14 -0800145 *
Tejun Heoe8989fa2012-03-05 13:15:20 -0800146 * Allocate a new blkg assocating @blkcg and @q.
Tejun Heo03814112012-03-05 13:15:14 -0800147 */
Tejun Heo15974992012-06-04 20:40:52 -0700148static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
149 gfp_t gfp_mask)
Tejun Heo03814112012-03-05 13:15:14 -0800150{
Tejun Heo3c798392012-04-16 13:57:25 -0700151 struct blkcg_gq *blkg;
Tejun Heof7331642019-11-07 11:18:03 -0800152 int i, cpu;
Tejun Heo03814112012-03-05 13:15:14 -0800153
154 /* alloc and init base part */
Tejun Heo15974992012-06-04 20:40:52 -0700155 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
Tejun Heo03814112012-03-05 13:15:14 -0800156 if (!blkg)
157 return NULL;
158
Tejun Heoef069b92019-06-13 15:30:39 -0700159 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
160 goto err_free;
161
Tejun Heof7331642019-11-07 11:18:03 -0800162 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
163 if (!blkg->iostat_cpu)
Tejun Heo77ea7332015-08-18 14:55:24 -0700164 goto err_free;
165
Tejun Heoc875f4d2012-03-05 13:15:22 -0800166 blkg->q = q;
Tejun Heoe8989fa2012-03-05 13:15:20 -0800167 INIT_LIST_HEAD(&blkg->q_node);
Tejun Heod3f77df2019-06-27 13:39:52 -0700168 spin_lock_init(&blkg->async_bio_lock);
169 bio_list_init(&blkg->async_bios);
170 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
Tejun Heo03814112012-03-05 13:15:14 -0800171 blkg->blkcg = blkcg;
Tejun Heo03814112012-03-05 13:15:14 -0800172
Tejun Heof7331642019-11-07 11:18:03 -0800173 u64_stats_init(&blkg->iostat.sync);
174 for_each_possible_cpu(cpu)
175 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
176
Tejun Heo8bd435b2012-04-13 13:11:28 -0700177 for (i = 0; i < BLKCG_MAX_POLS; i++) {
Tejun Heo3c798392012-04-16 13:57:25 -0700178 struct blkcg_policy *pol = blkcg_policy[i];
Tejun Heoe8989fa2012-03-05 13:15:20 -0800179 struct blkg_policy_data *pd;
Tejun Heo03814112012-03-05 13:15:14 -0800180
Tejun Heoa2b16932012-04-13 13:11:33 -0700181 if (!blkcg_policy_enabled(q, pol))
Tejun Heoe8989fa2012-03-05 13:15:20 -0800182 continue;
Tejun Heo549d3aa2012-03-05 13:15:16 -0800183
Tejun Heoe8989fa2012-03-05 13:15:20 -0800184 /* alloc per-policy data and attach it to blkg */
Tejun Heocf09a8e2019-08-28 15:05:51 -0700185 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
Tejun Heoa0516612012-06-26 15:05:44 -0700186 if (!pd)
187 goto err_free;
Tejun Heo549d3aa2012-03-05 13:15:16 -0800188
Tejun Heoe8989fa2012-03-05 13:15:20 -0800189 blkg->pd[i] = pd;
190 pd->blkg = blkg;
Tejun Heob276a872013-01-09 08:05:12 -0800191 pd->plid = i;
Tejun Heoe8989fa2012-03-05 13:15:20 -0800192 }
193
Tejun Heo03814112012-03-05 13:15:14 -0800194 return blkg;
Tejun Heoa0516612012-06-26 15:05:44 -0700195
196err_free:
197 blkg_free(blkg);
198 return NULL;
Tejun Heo03814112012-03-05 13:15:14 -0800199}
200
Tejun Heo24f29042015-08-18 14:55:17 -0700201struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
202 struct request_queue *q, bool update_hint)
Tejun Heo80fd9972012-04-13 14:50:53 -0700203{
Tejun Heo3c798392012-04-16 13:57:25 -0700204 struct blkcg_gq *blkg;
Tejun Heo80fd9972012-04-13 14:50:53 -0700205
Tejun Heoa6371202012-04-19 16:29:24 -0700206 /*
Tejun Heo86cde6b2013-01-09 08:05:10 -0800207 * Hint didn't match. Look up from the radix tree. Note that the
208 * hint can only be updated under queue_lock as otherwise @blkg
209 * could have already been removed from blkg_tree. The caller is
210 * responsible for grabbing queue_lock if @update_hint.
Tejun Heoa6371202012-04-19 16:29:24 -0700211 */
212 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
Tejun Heo86cde6b2013-01-09 08:05:10 -0800213 if (blkg && blkg->q == q) {
214 if (update_hint) {
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700215 lockdep_assert_held(&q->queue_lock);
Tejun Heo86cde6b2013-01-09 08:05:10 -0800216 rcu_assign_pointer(blkcg->blkg_hint, blkg);
217 }
Tejun Heoa6371202012-04-19 16:29:24 -0700218 return blkg;
Tejun Heo86cde6b2013-01-09 08:05:10 -0800219 }
Tejun Heoa6371202012-04-19 16:29:24 -0700220
Tejun Heo80fd9972012-04-13 14:50:53 -0700221 return NULL;
222}
Tejun Heoae118892015-08-18 14:55:20 -0700223EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
Tejun Heo80fd9972012-04-13 14:50:53 -0700224
Tejun Heo15974992012-06-04 20:40:52 -0700225/*
Jens Axboed708f0d2017-03-29 11:25:48 -0600226 * If @new_blkg is %NULL, this function tries to allocate a new one as
227 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
Tejun Heo15974992012-06-04 20:40:52 -0700228 */
Tejun Heo86cde6b2013-01-09 08:05:10 -0800229static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
Jens Axboed708f0d2017-03-29 11:25:48 -0600230 struct request_queue *q,
231 struct blkcg_gq *new_blkg)
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400232{
Jens Axboed708f0d2017-03-29 11:25:48 -0600233 struct blkcg_gq *blkg;
Tejun Heof427d902013-01-09 08:05:12 -0800234 int i, ret;
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400235
Tejun Heocd1604f2012-03-05 13:15:06 -0800236 WARN_ON_ONCE(!rcu_read_lock_held());
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700237 lockdep_assert_held(&q->queue_lock);
Vivek Goyal31e4c282009-12-03 12:59:42 -0500238
Dennis Zhou0273ac32018-12-11 18:03:08 -0500239 /* request_queue is dying, do not create/recreate a blkg */
240 if (blk_queue_dying(q)) {
241 ret = -ENODEV;
242 goto err_free_blkg;
243 }
244
Tejun Heo7ee9c562012-03-05 13:15:11 -0800245 /* blkg holds a reference to blkcg */
Tejun Heoec903c02014-05-13 12:11:01 -0400246 if (!css_tryget_online(&blkcg->css)) {
Tejun Heo20386ce2015-08-18 14:55:28 -0700247 ret = -ENODEV;
Tejun Heo93e6d5d2013-01-09 08:05:10 -0800248 goto err_free_blkg;
Tejun Heo15974992012-06-04 20:40:52 -0700249 }
Tejun Heocd1604f2012-03-05 13:15:06 -0800250
Jens Axboed708f0d2017-03-29 11:25:48 -0600251 /* allocate */
252 if (!new_blkg) {
253 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
254 if (unlikely(!new_blkg)) {
255 ret = -ENOMEM;
Christoph Hellwig8c911f32020-07-01 11:06:21 +0200256 goto err_put_css;
Jens Axboed708f0d2017-03-29 11:25:48 -0600257 }
258 }
259 blkg = new_blkg;
Tahsin Erdogan7fc6b872017-03-09 00:05:31 -0800260
Tejun Heodb613672013-05-14 13:52:31 -0700261 /* link parent */
Tejun Heo3c547862013-01-09 08:05:10 -0800262 if (blkcg_parent(blkcg)) {
263 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
264 if (WARN_ON_ONCE(!blkg->parent)) {
Tejun Heo20386ce2015-08-18 14:55:28 -0700265 ret = -ENODEV;
Christoph Hellwig8c911f32020-07-01 11:06:21 +0200266 goto err_put_css;
Tejun Heo3c547862013-01-09 08:05:10 -0800267 }
268 blkg_get(blkg->parent);
269 }
270
Tejun Heodb613672013-05-14 13:52:31 -0700271 /* invoke per-policy init */
272 for (i = 0; i < BLKCG_MAX_POLS; i++) {
273 struct blkcg_policy *pol = blkcg_policy[i];
274
275 if (blkg->pd[i] && pol->pd_init_fn)
Tejun Heoa9520cd2015-08-18 14:55:14 -0700276 pol->pd_init_fn(blkg->pd[i]);
Tejun Heodb613672013-05-14 13:52:31 -0700277 }
278
279 /* insert */
Tejun Heoa6371202012-04-19 16:29:24 -0700280 spin_lock(&blkcg->lock);
281 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
282 if (likely(!ret)) {
283 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
284 list_add(&blkg->q_node, &q->blkg_list);
Tejun Heof427d902013-01-09 08:05:12 -0800285
286 for (i = 0; i < BLKCG_MAX_POLS; i++) {
287 struct blkcg_policy *pol = blkcg_policy[i];
288
289 if (blkg->pd[i] && pol->pd_online_fn)
Tejun Heoa9520cd2015-08-18 14:55:14 -0700290 pol->pd_online_fn(blkg->pd[i]);
Tejun Heof427d902013-01-09 08:05:12 -0800291 }
Tejun Heoa6371202012-04-19 16:29:24 -0700292 }
Tejun Heof427d902013-01-09 08:05:12 -0800293 blkg->online = true;
Tejun Heoa6371202012-04-19 16:29:24 -0700294 spin_unlock(&blkcg->lock);
295
Tejun Heoec13b1d2015-05-22 17:13:19 -0400296 if (!ret)
Tejun Heoa6371202012-04-19 16:29:24 -0700297 return blkg;
Tejun Heo15974992012-06-04 20:40:52 -0700298
Tejun Heo3c547862013-01-09 08:05:10 -0800299 /* @blkg failed fully initialized, use the usual release path */
300 blkg_put(blkg);
301 return ERR_PTR(ret);
302
Jens Axboed708f0d2017-03-29 11:25:48 -0600303err_put_css:
Tejun Heo496fb782012-04-19 16:29:23 -0700304 css_put(&blkcg->css);
Tejun Heo93e6d5d2013-01-09 08:05:10 -0800305err_free_blkg:
Jens Axboed708f0d2017-03-29 11:25:48 -0600306 blkg_free(new_blkg);
Tejun Heo93e6d5d2013-01-09 08:05:10 -0800307 return ERR_PTR(ret);
Vivek Goyal31e4c282009-12-03 12:59:42 -0500308}
Tejun Heo3c96cb32012-04-13 13:11:34 -0700309
Tejun Heo86cde6b2013-01-09 08:05:10 -0800310/**
Christoph Hellwig8c546282020-06-27 09:31:52 +0200311 * blkg_lookup_create - lookup blkg, try to create one if not there
Tejun Heo86cde6b2013-01-09 08:05:10 -0800312 * @blkcg: blkcg of interest
313 * @q: request_queue of interest
314 *
315 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
Tejun Heo3c547862013-01-09 08:05:10 -0800316 * create one. blkg creation is performed recursively from blkcg_root such
317 * that all non-root blkg's have access to the parent blkg. This function
Christoph Hellwig8c546282020-06-27 09:31:52 +0200318 * should be called under RCU read lock and takes @q->queue_lock.
Tejun Heo86cde6b2013-01-09 08:05:10 -0800319 *
Dennis Zhoubeea9da2018-12-05 12:10:28 -0500320 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
321 * down from root.
Tejun Heo86cde6b2013-01-09 08:05:10 -0800322 */
Christoph Hellwig8c546282020-06-27 09:31:52 +0200323static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
324 struct request_queue *q)
Tejun Heo3c96cb32012-04-13 13:11:34 -0700325{
Tejun Heo86cde6b2013-01-09 08:05:10 -0800326 struct blkcg_gq *blkg;
Christoph Hellwig8c546282020-06-27 09:31:52 +0200327 unsigned long flags;
Tejun Heo86cde6b2013-01-09 08:05:10 -0800328
329 WARN_ON_ONCE(!rcu_read_lock_held());
Tejun Heo86cde6b2013-01-09 08:05:10 -0800330
Christoph Hellwig8c546282020-06-27 09:31:52 +0200331 blkg = blkg_lookup(blkcg, q);
Tejun Heo86cde6b2013-01-09 08:05:10 -0800332 if (blkg)
333 return blkg;
334
Christoph Hellwig8c546282020-06-27 09:31:52 +0200335 spin_lock_irqsave(&q->queue_lock, flags);
336 blkg = __blkg_lookup(blkcg, q, true);
337 if (blkg)
338 goto found;
339
Tejun Heo3c547862013-01-09 08:05:10 -0800340 /*
341 * Create blkgs walking down from blkcg_root to @blkcg, so that all
Dennis Zhoubeea9da2018-12-05 12:10:28 -0500342 * non-root blkgs have access to their parents. Returns the closest
343 * blkg to the intended blkg should blkg_create() fail.
Tejun Heo3c547862013-01-09 08:05:10 -0800344 */
345 while (true) {
346 struct blkcg *pos = blkcg;
347 struct blkcg *parent = blkcg_parent(blkcg);
Dennis Zhoubeea9da2018-12-05 12:10:28 -0500348 struct blkcg_gq *ret_blkg = q->root_blkg;
Tejun Heo3c547862013-01-09 08:05:10 -0800349
Dennis Zhoubeea9da2018-12-05 12:10:28 -0500350 while (parent) {
351 blkg = __blkg_lookup(parent, q, false);
352 if (blkg) {
353 /* remember closest blkg */
354 ret_blkg = blkg;
355 break;
356 }
Tejun Heo3c547862013-01-09 08:05:10 -0800357 pos = parent;
358 parent = blkcg_parent(parent);
359 }
360
Jens Axboed708f0d2017-03-29 11:25:48 -0600361 blkg = blkg_create(pos, q, NULL);
Christoph Hellwig8c546282020-06-27 09:31:52 +0200362 if (IS_ERR(blkg)) {
363 blkg = ret_blkg;
364 break;
365 }
Dennis Zhoubeea9da2018-12-05 12:10:28 -0500366 if (pos == blkcg)
Christoph Hellwig8c546282020-06-27 09:31:52 +0200367 break;
Dennis Zhoub9789622018-12-05 12:10:27 -0500368 }
369
Christoph Hellwig8c546282020-06-27 09:31:52 +0200370found:
371 spin_unlock_irqrestore(&q->queue_lock, flags);
Dennis Zhoub9789622018-12-05 12:10:27 -0500372 return blkg;
373}
374
Tejun Heo3c798392012-04-16 13:57:25 -0700375static void blkg_destroy(struct blkcg_gq *blkg)
Tejun Heo72e06c22012-03-05 13:15:00 -0800376{
Tejun Heo3c798392012-04-16 13:57:25 -0700377 struct blkcg *blkcg = blkg->blkcg;
Dennis Zhou (Facebook)6b065462018-08-31 16:22:42 -0400378 int i;
Tejun Heo03aa2642012-03-05 13:15:19 -0800379
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700380 lockdep_assert_held(&blkg->q->queue_lock);
Tejun Heo9f13ef62012-03-05 13:15:21 -0800381 lockdep_assert_held(&blkcg->lock);
Tejun Heo03aa2642012-03-05 13:15:19 -0800382
383 /* Something wrong if we are trying to remove same group twice */
Tejun Heoe8989fa2012-03-05 13:15:20 -0800384 WARN_ON_ONCE(list_empty(&blkg->q_node));
Tejun Heo9f13ef62012-03-05 13:15:21 -0800385 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
Tejun Heoa6371202012-04-19 16:29:24 -0700386
Dennis Zhou (Facebook)6b065462018-08-31 16:22:42 -0400387 for (i = 0; i < BLKCG_MAX_POLS; i++) {
388 struct blkcg_policy *pol = blkcg_policy[i];
389
390 if (blkg->pd[i] && pol->pd_offline_fn)
391 pol->pd_offline_fn(blkg->pd[i]);
392 }
393
Tejun Heof427d902013-01-09 08:05:12 -0800394 blkg->online = false;
395
Tejun Heoa6371202012-04-19 16:29:24 -0700396 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
Tejun Heoe8989fa2012-03-05 13:15:20 -0800397 list_del_init(&blkg->q_node);
Tejun Heo9f13ef62012-03-05 13:15:21 -0800398 hlist_del_init_rcu(&blkg->blkcg_node);
Tejun Heo03aa2642012-03-05 13:15:19 -0800399
Tejun Heo03aa2642012-03-05 13:15:19 -0800400 /*
Tejun Heoa6371202012-04-19 16:29:24 -0700401 * Both setting lookup hint to and clearing it from @blkg are done
402 * under queue_lock. If it's not pointing to @blkg now, it never
403 * will. Hint assignment itself can race safely.
404 */
Paul E. McKenneyec6c676a2014-02-17 13:35:57 -0800405 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
Tejun Heoa6371202012-04-19 16:29:24 -0700406 rcu_assign_pointer(blkcg->blkg_hint, NULL);
407
408 /*
Tejun Heo03aa2642012-03-05 13:15:19 -0800409 * Put the reference taken at the time of creation so that when all
410 * queues are gone, group can be destroyed.
411 */
Dennis Zhou7fcf2b02018-12-05 12:10:38 -0500412 percpu_ref_kill(&blkg->refcnt);
Tejun Heo03aa2642012-03-05 13:15:19 -0800413}
414
Tejun Heo9f13ef62012-03-05 13:15:21 -0800415/**
416 * blkg_destroy_all - destroy all blkgs associated with a request_queue
417 * @q: request_queue of interest
Tejun Heo9f13ef62012-03-05 13:15:21 -0800418 *
Tejun Heo3c96cb32012-04-13 13:11:34 -0700419 * Destroy all blkgs associated with @q.
Tejun Heo9f13ef62012-03-05 13:15:21 -0800420 */
Tejun Heo3c96cb32012-04-13 13:11:34 -0700421static void blkg_destroy_all(struct request_queue *q)
Tejun Heo03aa2642012-03-05 13:15:19 -0800422{
Tejun Heo3c798392012-04-16 13:57:25 -0700423 struct blkcg_gq *blkg, *n;
Tejun Heo72e06c22012-03-05 13:15:00 -0800424
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700425 spin_lock_irq(&q->queue_lock);
Tejun Heo9f13ef62012-03-05 13:15:21 -0800426 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
Tejun Heo3c798392012-04-16 13:57:25 -0700427 struct blkcg *blkcg = blkg->blkcg;
Tejun Heo72e06c22012-03-05 13:15:00 -0800428
Tejun Heo9f13ef62012-03-05 13:15:21 -0800429 spin_lock(&blkcg->lock);
430 blkg_destroy(blkg);
431 spin_unlock(&blkcg->lock);
Tejun Heo72e06c22012-03-05 13:15:00 -0800432 }
Tejun Heo6fe810b2015-09-05 15:47:36 -0400433
434 q->root_blkg = NULL;
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700435 spin_unlock_irq(&q->queue_lock);
Tejun Heo72e06c22012-03-05 13:15:00 -0800436}
437
Tejun Heo182446d2013-08-08 20:11:24 -0400438static int blkcg_reset_stats(struct cgroup_subsys_state *css,
439 struct cftype *cftype, u64 val)
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700440{
Tejun Heo182446d2013-08-08 20:11:24 -0400441 struct blkcg *blkcg = css_to_blkcg(css);
Tejun Heo3c798392012-04-16 13:57:25 -0700442 struct blkcg_gq *blkg;
Tejun Heof7331642019-11-07 11:18:03 -0800443 int i, cpu;
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700444
Tejun Heo838f13b2015-07-09 16:39:47 -0400445 mutex_lock(&blkcg_pol_mutex);
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700446 spin_lock_irq(&blkcg->lock);
Tejun Heo997a0262012-03-08 10:53:58 -0800447
448 /*
449 * Note that stat reset is racy - it doesn't synchronize against
450 * stat updates. This is a debug feature which shouldn't exist
451 * anyway. If you get hit by a race, retry.
452 */
Sasha Levinb67bfe02013-02-27 17:06:00 -0800453 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
Tejun Heof7331642019-11-07 11:18:03 -0800454 for_each_possible_cpu(cpu) {
455 struct blkg_iostat_set *bis =
456 per_cpu_ptr(blkg->iostat_cpu, cpu);
457 memset(bis, 0, sizeof(*bis));
458 }
459 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
Tejun Heo77ea7332015-08-18 14:55:24 -0700460
Tejun Heo8bd435b2012-04-13 13:11:28 -0700461 for (i = 0; i < BLKCG_MAX_POLS; i++) {
Tejun Heo3c798392012-04-16 13:57:25 -0700462 struct blkcg_policy *pol = blkcg_policy[i];
Tejun Heo549d3aa2012-03-05 13:15:16 -0800463
Tejun Heoa9520cd2015-08-18 14:55:14 -0700464 if (blkg->pd[i] && pol->pd_reset_stats_fn)
465 pol->pd_reset_stats_fn(blkg->pd[i]);
Tejun Heobc0d6502012-04-13 13:11:26 -0700466 }
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700467 }
Vivek Goyalf0bdc8c2011-05-19 15:38:30 -0400468
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700469 spin_unlock_irq(&blkcg->lock);
Tejun Heobc0d6502012-04-13 13:11:26 -0700470 mutex_unlock(&blkcg_pol_mutex);
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700471 return 0;
472}
473
Tejun Heodd165eb2015-08-18 14:55:33 -0700474const char *blkg_dev_name(struct blkcg_gq *blkg)
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700475{
Tejun Heod3d32e62012-04-01 14:38:42 -0700476 /* some drivers (floppy) instantiate a queue w/o disk registered */
Jan Karadc3b17c2017-02-02 15:56:50 +0100477 if (blkg->q->backing_dev_info->dev)
Yufen Yud51cfc52020-05-04 14:47:55 +0200478 return bdi_dev_name(blkg->q->backing_dev_info);
Tejun Heod3d32e62012-04-01 14:38:42 -0700479 return NULL;
Divyesh Shah303a3ac2010-04-01 15:01:24 -0700480}
481
Tejun Heod3d32e62012-04-01 14:38:42 -0700482/**
483 * blkcg_print_blkgs - helper for printing per-blkg data
484 * @sf: seq_file to print to
485 * @blkcg: blkcg of interest
486 * @prfill: fill function to print out a blkg
487 * @pol: policy in question
488 * @data: data to be passed to @prfill
489 * @show_total: to print out sum of prfill return values or not
490 *
491 * This function invokes @prfill on each blkg of @blkcg if pd for the
492 * policy specified by @pol exists. @prfill is invoked with @sf, the
Tejun Heo810ecfa2013-01-09 08:05:13 -0800493 * policy data and @data and the matching queue lock held. If @show_total
494 * is %true, the sum of the return values from @prfill is printed with
495 * "Total" label at the end.
Tejun Heod3d32e62012-04-01 14:38:42 -0700496 *
497 * This is to be used to construct print functions for
498 * cftype->read_seq_string method.
499 */
Tejun Heo3c798392012-04-16 13:57:25 -0700500void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
Tejun Heof95a04a2012-04-16 13:57:26 -0700501 u64 (*prfill)(struct seq_file *,
502 struct blkg_policy_data *, int),
Tejun Heo3c798392012-04-16 13:57:25 -0700503 const struct blkcg_policy *pol, int data,
Tejun Heoec399342012-04-13 13:11:27 -0700504 bool show_total)
Vivek Goyal5624a4e2011-05-19 15:38:28 -0400505{
Tejun Heo3c798392012-04-16 13:57:25 -0700506 struct blkcg_gq *blkg;
Tejun Heod3d32e62012-04-01 14:38:42 -0700507 u64 total = 0;
508
Tejun Heo810ecfa2013-01-09 08:05:13 -0800509 rcu_read_lock();
Linus Torvaldsee89f812013-02-28 12:52:24 -0800510 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700511 spin_lock_irq(&blkg->q->queue_lock);
Tejun Heoa2b16932012-04-13 13:11:33 -0700512 if (blkcg_policy_enabled(blkg->q, pol))
Tejun Heof95a04a2012-04-16 13:57:26 -0700513 total += prfill(sf, blkg->pd[pol->plid], data);
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700514 spin_unlock_irq(&blkg->q->queue_lock);
Tejun Heo810ecfa2013-01-09 08:05:13 -0800515 }
516 rcu_read_unlock();
Tejun Heod3d32e62012-04-01 14:38:42 -0700517
518 if (show_total)
519 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
520}
Tejun Heo829fdb52012-04-01 14:38:43 -0700521EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
Tejun Heod3d32e62012-04-01 14:38:42 -0700522
523/**
524 * __blkg_prfill_u64 - prfill helper for a single u64 value
525 * @sf: seq_file to print to
Tejun Heof95a04a2012-04-16 13:57:26 -0700526 * @pd: policy private data of interest
Tejun Heod3d32e62012-04-01 14:38:42 -0700527 * @v: value to print
528 *
Tejun Heof95a04a2012-04-16 13:57:26 -0700529 * Print @v to @sf for the device assocaited with @pd.
Tejun Heod3d32e62012-04-01 14:38:42 -0700530 */
Tejun Heof95a04a2012-04-16 13:57:26 -0700531u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
Tejun Heod3d32e62012-04-01 14:38:42 -0700532{
Tejun Heof95a04a2012-04-16 13:57:26 -0700533 const char *dname = blkg_dev_name(pd->blkg);
Tejun Heod3d32e62012-04-01 14:38:42 -0700534
535 if (!dname)
536 return 0;
537
538 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
539 return v;
540}
Tejun Heo829fdb52012-04-01 14:38:43 -0700541EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
Tejun Heod3d32e62012-04-01 14:38:42 -0700542
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600543/* Performs queue bypass and policy enabled checks then looks up blkg. */
544static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
545 const struct blkcg_policy *pol,
546 struct request_queue *q)
547{
548 WARN_ON_ONCE(!rcu_read_lock_held());
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700549 lockdep_assert_held(&q->queue_lock);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600550
551 if (!blkcg_policy_enabled(q, pol))
552 return ERR_PTR(-EOPNOTSUPP);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600553 return __blkg_lookup(blkcg, q, true /* update_hint */);
554}
555
Tejun Heo16b3de62013-01-09 08:05:12 -0800556/**
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100557 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
Tejun Heo015d2542019-08-28 15:05:53 -0700558 * @inputp: input string pointer
559 *
560 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100561 * from @input and get and return the matching bdev. *@inputp is
Tejun Heo015d2542019-08-28 15:05:53 -0700562 * updated to point past the device node prefix. Returns an ERR_PTR()
563 * value on error.
564 *
565 * Use this function iff blkg_conf_prep() can't be used for some reason.
566 */
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100567struct block_device *blkcg_conf_open_bdev(char **inputp)
Tejun Heo015d2542019-08-28 15:05:53 -0700568{
569 char *input = *inputp;
570 unsigned int major, minor;
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100571 struct block_device *bdev;
572 int key_len;
Tejun Heo015d2542019-08-28 15:05:53 -0700573
574 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
575 return ERR_PTR(-EINVAL);
576
577 input += key_len;
578 if (!isspace(*input))
579 return ERR_PTR(-EINVAL);
580 input = skip_spaces(input);
581
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100582 bdev = blkdev_get_no_open(MKDEV(major, minor));
583 if (!bdev)
Tejun Heo015d2542019-08-28 15:05:53 -0700584 return ERR_PTR(-ENODEV);
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100585 if (bdev_is_partition(bdev)) {
586 blkdev_put_no_open(bdev);
Tejun Heo015d2542019-08-28 15:05:53 -0700587 return ERR_PTR(-ENODEV);
588 }
589
590 *inputp = input;
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100591 return bdev;
Tejun Heo015d2542019-08-28 15:05:53 -0700592}
593
594/**
595 * blkg_conf_prep - parse and prepare for per-blkg config update
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700596 * @blkcg: target block cgroup
Tejun Heoda8b0662012-04-13 13:11:29 -0700597 * @pol: target policy
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700598 * @input: input string
599 * @ctx: blkg_conf_ctx to be filled
600 *
601 * Parse per-blkg config update from @input and initialize @ctx with the
Tejun Heo36aa9e52015-08-18 14:55:31 -0700602 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
603 * part of @input following MAJ:MIN. This function returns with RCU read
604 * lock and queue lock held and must be paired with blkg_conf_finish().
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700605 */
Tejun Heo3c798392012-04-16 13:57:25 -0700606int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
Tejun Heo36aa9e52015-08-18 14:55:31 -0700607 char *input, struct blkg_conf_ctx *ctx)
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100608 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800609{
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100610 struct block_device *bdev;
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600611 struct request_queue *q;
Tejun Heo3c798392012-04-16 13:57:25 -0700612 struct blkcg_gq *blkg;
Tejun Heo015d2542019-08-28 15:05:53 -0700613 int ret;
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800614
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100615 bdev = blkcg_conf_open_bdev(&input);
616 if (IS_ERR(bdev))
617 return PTR_ERR(bdev);
Tejun Heoe56da7e2012-03-05 13:15:07 -0800618
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100619 q = bdev->bd_disk->queue;
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600620
Tejun Heoe56da7e2012-03-05 13:15:07 -0800621 rcu_read_lock();
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700622 spin_lock_irq(&q->queue_lock);
Tejun Heoda8b0662012-04-13 13:11:29 -0700623
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600624 blkg = blkg_lookup_check(blkcg, pol, q);
Tejun Heo4bfd4822012-03-05 13:15:08 -0800625 if (IS_ERR(blkg)) {
626 ret = PTR_ERR(blkg);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600627 goto fail_unlock;
Vivek Goyal062a6442010-09-15 17:06:33 -0400628 }
Tejun Heoe56da7e2012-03-05 13:15:07 -0800629
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600630 if (blkg)
631 goto success;
632
633 /*
634 * Create blkgs walking down from blkcg_root to @blkcg, so that all
635 * non-root blkgs have access to their parents.
636 */
637 while (true) {
638 struct blkcg *pos = blkcg;
639 struct blkcg *parent;
640 struct blkcg_gq *new_blkg;
641
642 parent = blkcg_parent(blkcg);
643 while (parent && !__blkg_lookup(parent, q, false)) {
644 pos = parent;
645 parent = blkcg_parent(parent);
646 }
647
648 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700649 spin_unlock_irq(&q->queue_lock);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600650 rcu_read_unlock();
651
652 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
653 if (unlikely(!new_blkg)) {
654 ret = -ENOMEM;
655 goto fail;
656 }
657
Gabriel Krisman Bertazif255c192020-10-22 16:58:42 -0400658 if (radix_tree_preload(GFP_KERNEL)) {
659 blkg_free(new_blkg);
660 ret = -ENOMEM;
661 goto fail;
662 }
663
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600664 rcu_read_lock();
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700665 spin_lock_irq(&q->queue_lock);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600666
667 blkg = blkg_lookup_check(pos, pol, q);
668 if (IS_ERR(blkg)) {
669 ret = PTR_ERR(blkg);
Gabriel Krisman Bertazi52abfcb2020-10-22 16:58:41 -0400670 blkg_free(new_blkg);
Gabriel Krisman Bertazif255c192020-10-22 16:58:42 -0400671 goto fail_preloaded;
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600672 }
673
674 if (blkg) {
675 blkg_free(new_blkg);
676 } else {
677 blkg = blkg_create(pos, q, new_blkg);
Kefeng Wang98d669b2019-06-05 22:24:27 +0800678 if (IS_ERR(blkg)) {
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600679 ret = PTR_ERR(blkg);
Gabriel Krisman Bertazif255c192020-10-22 16:58:42 -0400680 goto fail_preloaded;
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600681 }
682 }
683
Gabriel Krisman Bertazif255c192020-10-22 16:58:42 -0400684 radix_tree_preload_end();
685
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600686 if (pos == blkcg)
687 goto success;
688 }
689success:
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100690 ctx->bdev = bdev;
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700691 ctx->blkg = blkg;
Tejun Heo015d2542019-08-28 15:05:53 -0700692 ctx->body = input;
Tejun Heo726fa6942012-04-01 14:38:43 -0700693 return 0;
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600694
Gabriel Krisman Bertazif255c192020-10-22 16:58:42 -0400695fail_preloaded:
696 radix_tree_preload_end();
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600697fail_unlock:
Christoph Hellwig0d945c12018-11-15 12:17:28 -0700698 spin_unlock_irq(&q->queue_lock);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600699 rcu_read_unlock();
700fail:
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100701 blkdev_put_no_open(bdev);
Tahsin Erdogan457e490f2017-03-29 11:27:19 -0600702 /*
703 * If queue was bypassing, we should retry. Do so after a
704 * short msleep(). It isn't strictly necessary but queue
705 * can be bypassing for some time and it's always nice to
706 * avoid busy looping.
707 */
708 if (ret == -EBUSY) {
709 msleep(10);
710 ret = restart_syscall();
711 }
712 return ret;
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800713}
Pavel Begunkov89f3b6d2019-09-14 20:31:50 +0300714EXPORT_SYMBOL_GPL(blkg_conf_prep);
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800715
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700716/**
717 * blkg_conf_finish - finish up per-blkg config update
718 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
719 *
720 * Finish up after per-blkg config update. This function must be paired
721 * with blkg_conf_prep().
722 */
Tejun Heo829fdb52012-04-01 14:38:43 -0700723void blkg_conf_finish(struct blkg_conf_ctx *ctx)
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100724 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800725{
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100726 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
Tejun Heo3a8b31d2012-04-01 14:38:43 -0700727 rcu_read_unlock();
Christoph Hellwig22ae8ce2020-11-26 09:23:26 +0100728 blkdev_put_no_open(ctx->bdev);
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800729}
Pavel Begunkov89f3b6d2019-09-14 20:31:50 +0300730EXPORT_SYMBOL_GPL(blkg_conf_finish);
Gui Jianfeng34d0f172010-04-13 16:05:49 +0800731
Boris Burkovcd1fc4b2020-06-01 13:11:43 -0700732static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
733{
734 int i;
735
736 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
737 dst->bytes[i] = src->bytes[i];
738 dst->ios[i] = src->ios[i];
739 }
740}
741
742static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
743{
744 int i;
745
746 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
747 dst->bytes[i] += src->bytes[i];
748 dst->ios[i] += src->ios[i];
749 }
750}
751
752static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
753{
754 int i;
755
756 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
757 dst->bytes[i] -= src->bytes[i];
758 dst->ios[i] -= src->ios[i];
759 }
760}
761
762static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
763{
764 struct blkcg *blkcg = css_to_blkcg(css);
765 struct blkcg_gq *blkg;
766
Johannes Weinerdc265322021-04-29 22:56:23 -0700767 /* Root-level stats are sourced from system-wide IO stats */
768 if (!cgroup_parent(css->cgroup))
769 return;
770
Boris Burkovcd1fc4b2020-06-01 13:11:43 -0700771 rcu_read_lock();
772
773 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
774 struct blkcg_gq *parent = blkg->parent;
775 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
776 struct blkg_iostat cur, delta;
777 unsigned int seq;
778
779 /* fetch the current per-cpu values */
780 do {
781 seq = u64_stats_fetch_begin(&bisc->sync);
782 blkg_iostat_set(&cur, &bisc->cur);
783 } while (u64_stats_fetch_retry(&bisc->sync, seq));
784
785 /* propagate percpu delta to global */
786 u64_stats_update_begin(&blkg->iostat.sync);
787 blkg_iostat_set(&delta, &cur);
788 blkg_iostat_sub(&delta, &bisc->last);
789 blkg_iostat_add(&blkg->iostat.cur, &delta);
790 blkg_iostat_add(&bisc->last, &delta);
791 u64_stats_update_end(&blkg->iostat.sync);
792
Johannes Weinerdc265322021-04-29 22:56:23 -0700793 /* propagate global delta to parent (unless that's root) */
794 if (parent && parent->parent) {
Boris Burkovcd1fc4b2020-06-01 13:11:43 -0700795 u64_stats_update_begin(&parent->iostat.sync);
796 blkg_iostat_set(&delta, &blkg->iostat.cur);
797 blkg_iostat_sub(&delta, &blkg->iostat.last);
798 blkg_iostat_add(&parent->iostat.cur, &delta);
799 blkg_iostat_add(&blkg->iostat.last, &delta);
800 u64_stats_update_end(&parent->iostat.sync);
801 }
802 }
803
804 rcu_read_unlock();
805}
806
Boris Burkovef45fe42020-06-01 13:12:05 -0700807/*
Johannes Weinerdc265322021-04-29 22:56:23 -0700808 * We source root cgroup stats from the system-wide stats to avoid
809 * tracking the same information twice and incurring overhead when no
810 * cgroups are defined. For that reason, cgroup_rstat_flush in
811 * blkcg_print_stat does not actually fill out the iostat in the root
812 * cgroup's blkcg_gq.
Boris Burkovef45fe42020-06-01 13:12:05 -0700813 *
814 * However, we would like to re-use the printing code between the root and
815 * non-root cgroups to the extent possible. For that reason, we simulate
816 * flushing the root cgroup's stats by explicitly filling in the iostat
817 * with disk level statistics.
818 */
819static void blkcg_fill_root_iostats(void)
820{
821 struct class_dev_iter iter;
822 struct device *dev;
823
824 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
825 while ((dev = class_dev_iter_next(&iter))) {
Christoph Hellwig0d021292020-11-27 16:43:51 +0100826 struct block_device *bdev = dev_to_bdev(dev);
827 struct blkcg_gq *blkg =
828 blk_queue_root_blkg(bdev->bd_disk->queue);
Boris Burkovef45fe42020-06-01 13:12:05 -0700829 struct blkg_iostat tmp;
830 int cpu;
831
832 memset(&tmp, 0, sizeof(tmp));
833 for_each_possible_cpu(cpu) {
834 struct disk_stats *cpu_dkstats;
835
Christoph Hellwig0d021292020-11-27 16:43:51 +0100836 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
Boris Burkovef45fe42020-06-01 13:12:05 -0700837 tmp.ios[BLKG_IOSTAT_READ] +=
838 cpu_dkstats->ios[STAT_READ];
839 tmp.ios[BLKG_IOSTAT_WRITE] +=
840 cpu_dkstats->ios[STAT_WRITE];
841 tmp.ios[BLKG_IOSTAT_DISCARD] +=
842 cpu_dkstats->ios[STAT_DISCARD];
843 // convert sectors to bytes
844 tmp.bytes[BLKG_IOSTAT_READ] +=
845 cpu_dkstats->sectors[STAT_READ] << 9;
846 tmp.bytes[BLKG_IOSTAT_WRITE] +=
847 cpu_dkstats->sectors[STAT_WRITE] << 9;
848 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
849 cpu_dkstats->sectors[STAT_DISCARD] << 9;
850
851 u64_stats_update_begin(&blkg->iostat.sync);
852 blkg_iostat_set(&blkg->iostat.cur, &tmp);
853 u64_stats_update_end(&blkg->iostat.sync);
854 }
855 }
856}
857
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700858static int blkcg_print_stat(struct seq_file *sf, void *v)
859{
860 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
861 struct blkcg_gq *blkg;
862
Boris Burkovef45fe42020-06-01 13:12:05 -0700863 if (!seq_css(sf)->parent)
864 blkcg_fill_root_iostats();
865 else
866 cgroup_rstat_flush(blkcg->css.cgroup);
867
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700868 rcu_read_lock();
869
870 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
Tejun Heof7331642019-11-07 11:18:03 -0800871 struct blkg_iostat_set *bis = &blkg->iostat;
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700872 const char *dname;
Josef Bacik903d23f2018-07-03 11:14:52 -0400873 char *buf;
Tejun Heo636620b2018-07-18 04:47:41 -0700874 u64 rbytes, wbytes, rios, wios, dbytes, dios;
Josef Bacik903d23f2018-07-03 11:14:52 -0400875 size_t size = seq_get_buf(sf, &buf), off = 0;
876 int i;
877 bool has_stats = false;
Tejun Heof7331642019-11-07 11:18:03 -0800878 unsigned seq;
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700879
Tejun Heob0814362019-11-05 08:09:51 -0800880 spin_lock_irq(&blkg->q->queue_lock);
881
882 if (!blkg->online)
883 goto skip;
884
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700885 dname = blkg_dev_name(blkg);
886 if (!dname)
Tejun Heob0814362019-11-05 08:09:51 -0800887 goto skip;
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700888
Josef Bacik903d23f2018-07-03 11:14:52 -0400889 /*
890 * Hooray string manipulation, count is the size written NOT
891 * INCLUDING THE \0, so size is now count+1 less than what we
892 * had before, but we want to start writing the next bit from
893 * the \0 so we only add count to buf.
894 */
895 off += scnprintf(buf+off, size-off, "%s ", dname);
896
Tejun Heof7331642019-11-07 11:18:03 -0800897 do {
898 seq = u64_stats_fetch_begin(&bis->sync);
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700899
Tejun Heof7331642019-11-07 11:18:03 -0800900 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
901 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
902 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
903 rios = bis->cur.ios[BLKG_IOSTAT_READ];
904 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
905 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
906 } while (u64_stats_fetch_retry(&bis->sync, seq));
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700907
Josef Bacik903d23f2018-07-03 11:14:52 -0400908 if (rbytes || wbytes || rios || wios) {
909 has_stats = true;
910 off += scnprintf(buf+off, size-off,
Tejun Heo636620b2018-07-18 04:47:41 -0700911 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
912 rbytes, wbytes, rios, wios,
913 dbytes, dios);
Josef Bacik903d23f2018-07-03 11:14:52 -0400914 }
915
Tejun Heo07b0fde2019-07-16 07:58:31 -0700916 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
Josef Bacikd09d8df2018-07-03 11:14:55 -0400917 has_stats = true;
918 off += scnprintf(buf+off, size-off,
919 " use_delay=%d delay_nsec=%llu",
920 atomic_read(&blkg->use_delay),
921 (unsigned long long)atomic64_read(&blkg->delay_nsec));
922 }
923
Josef Bacik903d23f2018-07-03 11:14:52 -0400924 for (i = 0; i < BLKCG_MAX_POLS; i++) {
925 struct blkcg_policy *pol = blkcg_policy[i];
926 size_t written;
927
928 if (!blkg->pd[i] || !pol->pd_stat_fn)
929 continue;
930
931 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
932 if (written)
933 has_stats = true;
934 off += written;
935 }
Tejun Heo07b0fde2019-07-16 07:58:31 -0700936
Josef Bacik903d23f2018-07-03 11:14:52 -0400937 if (has_stats) {
Tejun Heof539da82019-06-13 15:30:38 -0700938 if (off < size - 1) {
939 off += scnprintf(buf+off, size-off, "\n");
940 seq_commit(sf, off);
941 } else {
942 seq_commit(sf, -1);
943 }
Josef Bacik903d23f2018-07-03 11:14:52 -0400944 }
Tejun Heob0814362019-11-05 08:09:51 -0800945 skip:
946 spin_unlock_irq(&blkg->q->queue_lock);
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700947 }
948
949 rcu_read_unlock();
950 return 0;
951}
952
Bart Van Asschee1f3b942016-06-14 17:04:32 +0200953static struct cftype blkcg_files[] = {
Tejun Heo2ee867dc2015-08-18 14:55:34 -0700954 {
955 .name = "stat",
956 .seq_show = blkcg_print_stat,
957 },
958 { } /* terminate */
959};
960
Bart Van Asschee1f3b942016-06-14 17:04:32 +0200961static struct cftype blkcg_legacy_files[] = {
Vivek Goyal31e4c282009-12-03 12:59:42 -0500962 {
Divyesh Shah84c124d2010-04-09 08:31:19 +0200963 .name = "reset_stats",
Tejun Heo3c798392012-04-16 13:57:25 -0700964 .write_u64 = blkcg_reset_stats,
Vivek Goyal22084192009-12-03 12:59:49 -0500965 },
Tejun Heo4baf6e32012-04-01 12:09:55 -0700966 { } /* terminate */
Vivek Goyal31e4c282009-12-03 12:59:42 -0500967};
968
Dennis Zhou (Facebook)59b57712018-08-31 16:22:43 -0400969/*
970 * blkcg destruction is a three-stage process.
971 *
972 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
973 * which offlines writeback. Here we tie the next stage of blkg destruction
974 * to the completion of writeback associated with the blkcg. This lets us
975 * avoid punting potentially large amounts of outstanding writeback to root
976 * while maintaining any ongoing policies. The next stage is triggered when
977 * the nr_cgwbs count goes to zero.
978 *
979 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
980 * and handles the destruction of blkgs. Here the css reference held by
981 * the blkg is put back eventually allowing blkcg_css_free() to be called.
982 * This work may occur in cgwb_release_workfn() on the cgwb_release
983 * workqueue. Any submitted ios that fail to get the blkg ref will be
984 * punted to the root_blkg.
985 *
986 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
987 * This finally frees the blkcg.
988 */
989
Tejun Heo9f13ef62012-03-05 13:15:21 -0800990/**
Tejun Heo92fb9742012-11-19 08:13:38 -0800991 * blkcg_css_offline - cgroup css_offline callback
Tejun Heoeb954192013-08-08 20:11:23 -0400992 * @css: css of interest
Tejun Heo9f13ef62012-03-05 13:15:21 -0800993 *
Dennis Zhou (Facebook)59b57712018-08-31 16:22:43 -0400994 * This function is called when @css is about to go away. Here the cgwbs are
995 * offlined first and only once writeback associated with the blkcg has
996 * finished do we start step 2 (see above).
Tejun Heo9f13ef62012-03-05 13:15:21 -0800997 */
Tejun Heoeb954192013-08-08 20:11:23 -0400998static void blkcg_css_offline(struct cgroup_subsys_state *css)
Vivek Goyal31e4c282009-12-03 12:59:42 -0500999{
Tejun Heoeb954192013-08-08 20:11:23 -04001000 struct blkcg *blkcg = css_to_blkcg(css);
Vivek Goyal31e4c282009-12-03 12:59:42 -05001001
Dennis Zhou (Facebook)59b57712018-08-31 16:22:43 -04001002 /* this prevents anyone from attaching or migrating to this blkcg */
1003 wb_blkcg_offline(blkcg);
1004
Tejun Heod866dbf2019-07-24 10:37:22 -07001005 /* put the base online pin allowing step 2 to be triggered */
1006 blkcg_unpin_online(blkcg);
Dennis Zhou (Facebook)59b57712018-08-31 16:22:43 -04001007}
1008
1009/**
1010 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1011 * @blkcg: blkcg of interest
1012 *
1013 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1014 * is nested inside q lock, this function performs reverse double lock dancing.
1015 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1016 * blkcg_css_free to eventually be called.
1017 *
1018 * This is the blkcg counterpart of ioc_release_fn().
1019 */
1020void blkcg_destroy_blkgs(struct blkcg *blkcg)
1021{
Baolin Wang6c635ca2021-01-28 13:58:15 +08001022 might_sleep();
1023
Tejun Heo9f13ef62012-03-05 13:15:21 -08001024 spin_lock_irq(&blkcg->lock);
Tejun Heo7ee9c562012-03-05 13:15:11 -08001025
Dennis Zhou (Facebook)6b065462018-08-31 16:22:42 -04001026 while (!hlist_empty(&blkcg->blkg_list)) {
1027 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1028 struct blkcg_gq, blkcg_node);
Tejun Heoc875f4d2012-03-05 13:15:22 -08001029 struct request_queue *q = blkg->q;
Vivek Goyalb1c35762009-12-03 12:59:47 -05001030
Baolin Wang6c635ca2021-01-28 13:58:15 +08001031 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1032 /*
1033 * Given that the system can accumulate a huge number
1034 * of blkgs in pathological cases, check to see if we
1035 * need to rescheduling to avoid softlockup.
1036 */
Tejun Heo9f13ef62012-03-05 13:15:21 -08001037 spin_unlock_irq(&blkcg->lock);
Baolin Wang6c635ca2021-01-28 13:58:15 +08001038 cond_resched();
Dan Carpentera5567932012-03-29 20:57:08 +02001039 spin_lock_irq(&blkcg->lock);
Baolin Wang6c635ca2021-01-28 13:58:15 +08001040 continue;
Jens Axboe0f3942a2010-05-03 14:28:55 +02001041 }
Baolin Wang6c635ca2021-01-28 13:58:15 +08001042
1043 blkg_destroy(blkg);
1044 spin_unlock(&q->queue_lock);
Tejun Heo9f13ef62012-03-05 13:15:21 -08001045 }
Jens Axboe0f3942a2010-05-03 14:28:55 +02001046
Tejun Heo9f13ef62012-03-05 13:15:21 -08001047 spin_unlock_irq(&blkcg->lock);
Tejun Heo7ee9c562012-03-05 13:15:11 -08001048}
1049
Tejun Heoeb954192013-08-08 20:11:23 -04001050static void blkcg_css_free(struct cgroup_subsys_state *css)
Tejun Heo7ee9c562012-03-05 13:15:11 -08001051{
Tejun Heoeb954192013-08-08 20:11:23 -04001052 struct blkcg *blkcg = css_to_blkcg(css);
Tejun Heobc915e62015-08-18 14:55:08 -07001053 int i;
Tejun Heo7ee9c562012-03-05 13:15:11 -08001054
Tejun Heo7876f932015-07-09 16:39:49 -04001055 mutex_lock(&blkcg_pol_mutex);
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001056
Tejun Heo7876f932015-07-09 16:39:49 -04001057 list_del(&blkcg->all_blkcgs_node);
Tejun Heo7876f932015-07-09 16:39:49 -04001058
Tejun Heobc915e62015-08-18 14:55:08 -07001059 for (i = 0; i < BLKCG_MAX_POLS; i++)
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001060 if (blkcg->cpd[i])
1061 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1062
1063 mutex_unlock(&blkcg_pol_mutex);
1064
Tejun Heobc915e62015-08-18 14:55:08 -07001065 kfree(blkcg);
Vivek Goyal31e4c282009-12-03 12:59:42 -05001066}
1067
Tejun Heoeb954192013-08-08 20:11:23 -04001068static struct cgroup_subsys_state *
1069blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
Vivek Goyal31e4c282009-12-03 12:59:42 -05001070{
Tejun Heo3c798392012-04-16 13:57:25 -07001071 struct blkcg *blkcg;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001072 struct cgroup_subsys_state *ret;
1073 int i;
Vivek Goyal31e4c282009-12-03 12:59:42 -05001074
Tejun Heo7876f932015-07-09 16:39:49 -04001075 mutex_lock(&blkcg_pol_mutex);
1076
Tejun Heoeb954192013-08-08 20:11:23 -04001077 if (!parent_css) {
Tejun Heo3c798392012-04-16 13:57:25 -07001078 blkcg = &blkcg_root;
Tejun Heobc915e62015-08-18 14:55:08 -07001079 } else {
1080 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1081 if (!blkcg) {
1082 ret = ERR_PTR(-ENOMEM);
weiping zhang4c18c9e2017-08-25 23:49:32 +08001083 goto unlock;
Tejun Heobc915e62015-08-18 14:55:08 -07001084 }
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001085 }
Vivek Goyal31e4c282009-12-03 12:59:42 -05001086
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001087 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1088 struct blkcg_policy *pol = blkcg_policy[i];
1089 struct blkcg_policy_data *cpd;
1090
1091 /*
1092 * If the policy hasn't been attached yet, wait for it
1093 * to be attached before doing anything else. Otherwise,
1094 * check if the policy requires any specific per-cgroup
1095 * data: if it does, allocate and initialize it.
1096 */
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001097 if (!pol || !pol->cpd_alloc_fn)
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001098 continue;
1099
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001100 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001101 if (!cpd) {
1102 ret = ERR_PTR(-ENOMEM);
1103 goto free_pd_blkcg;
1104 }
Tejun Heo81437642015-08-18 14:55:15 -07001105 blkcg->cpd[i] = cpd;
1106 cpd->blkcg = blkcg;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001107 cpd->plid = i;
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001108 if (pol->cpd_init_fn)
1109 pol->cpd_init_fn(cpd);
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001110 }
1111
Vivek Goyal31e4c282009-12-03 12:59:42 -05001112 spin_lock_init(&blkcg->lock);
Tejun Heod866dbf2019-07-24 10:37:22 -07001113 refcount_set(&blkcg->online_pin, 1);
Tejun Heoe00f4f42016-11-21 18:03:32 -05001114 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
Vivek Goyal31e4c282009-12-03 12:59:42 -05001115 INIT_HLIST_HEAD(&blkcg->blkg_list);
Tejun Heo52ebea72015-05-22 17:13:37 -04001116#ifdef CONFIG_CGROUP_WRITEBACK
1117 INIT_LIST_HEAD(&blkcg->cgwb_list);
1118#endif
Tejun Heo7876f932015-07-09 16:39:49 -04001119 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1120
1121 mutex_unlock(&blkcg_pol_mutex);
Vivek Goyal31e4c282009-12-03 12:59:42 -05001122 return &blkcg->css;
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001123
1124free_pd_blkcg:
1125 for (i--; i >= 0; i--)
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001126 if (blkcg->cpd[i])
1127 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
weiping zhang4c18c9e2017-08-25 23:49:32 +08001128
1129 if (blkcg != &blkcg_root)
1130 kfree(blkcg);
1131unlock:
Tejun Heo7876f932015-07-09 16:39:49 -04001132 mutex_unlock(&blkcg_pol_mutex);
Arianna Avanzinie48453c2015-06-05 23:38:42 +02001133 return ret;
Vivek Goyal31e4c282009-12-03 12:59:42 -05001134}
1135
Tejun Heo4308a432019-07-24 10:37:55 -07001136static int blkcg_css_online(struct cgroup_subsys_state *css)
1137{
1138 struct blkcg *blkcg = css_to_blkcg(css);
1139 struct blkcg *parent = blkcg_parent(blkcg);
1140
1141 /*
1142 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1143 * don't go offline while cgwbs are still active on them. Pin the
1144 * parent so that offline always happens towards the root.
1145 */
1146 if (parent)
1147 blkcg_pin_online(parent);
1148 return 0;
1149}
1150
Tejun Heo5efd6112012-03-05 13:15:12 -08001151/**
1152 * blkcg_init_queue - initialize blkcg part of request queue
1153 * @q: request_queue to initialize
1154 *
Christoph Hellwigc62b37d2020-07-01 10:59:43 +02001155 * Called from blk_alloc_queue(). Responsible for initializing blkcg
Tejun Heo5efd6112012-03-05 13:15:12 -08001156 * part of new request_queue @q.
1157 *
1158 * RETURNS:
1159 * 0 on success, -errno on failure.
1160 */
1161int blkcg_init_queue(struct request_queue *q)
1162{
Jens Axboed708f0d2017-03-29 11:25:48 -06001163 struct blkcg_gq *new_blkg, *blkg;
1164 bool preloaded;
Tejun Heoec13b1d2015-05-22 17:13:19 -04001165 int ret;
Tejun Heo5efd6112012-03-05 13:15:12 -08001166
Jens Axboed708f0d2017-03-29 11:25:48 -06001167 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1168 if (!new_blkg)
1169 return -ENOMEM;
1170
1171 preloaded = !radix_tree_preload(GFP_KERNEL);
1172
Jiang Biaobea54882018-04-19 12:04:26 +08001173 /* Make sure the root blkg exists. */
Tejun Heoec13b1d2015-05-22 17:13:19 -04001174 rcu_read_lock();
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001175 spin_lock_irq(&q->queue_lock);
Jens Axboed708f0d2017-03-29 11:25:48 -06001176 blkg = blkg_create(&blkcg_root, q, new_blkg);
Jiang Biao901932a2018-04-19 12:06:09 +08001177 if (IS_ERR(blkg))
1178 goto err_unlock;
1179 q->root_blkg = blkg;
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001180 spin_unlock_irq(&q->queue_lock);
Tejun Heoec13b1d2015-05-22 17:13:19 -04001181 rcu_read_unlock();
1182
Jens Axboed708f0d2017-03-29 11:25:48 -06001183 if (preloaded)
1184 radix_tree_preload_end();
1185
Tejun Heoec13b1d2015-05-22 17:13:19 -04001186 ret = blk_throtl_init(q);
Christoph Hellwig04be60b2018-11-14 17:02:12 +01001187 if (ret)
1188 goto err_destroy_all;
Yufen Yu27029b42020-08-10 22:21:16 -04001189
1190 ret = blk_iolatency_init(q);
1191 if (ret) {
1192 blk_throtl_exit(q);
1193 goto err_destroy_all;
1194 }
Christoph Hellwig04be60b2018-11-14 17:02:12 +01001195 return 0;
Jiang Biao901932a2018-04-19 12:06:09 +08001196
Christoph Hellwig04be60b2018-11-14 17:02:12 +01001197err_destroy_all:
Christoph Hellwig04be60b2018-11-14 17:02:12 +01001198 blkg_destroy_all(q);
Christoph Hellwig04be60b2018-11-14 17:02:12 +01001199 return ret;
Jiang Biao901932a2018-04-19 12:06:09 +08001200err_unlock:
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001201 spin_unlock_irq(&q->queue_lock);
Jiang Biao901932a2018-04-19 12:06:09 +08001202 rcu_read_unlock();
1203 if (preloaded)
1204 radix_tree_preload_end();
1205 return PTR_ERR(blkg);
Tejun Heo5efd6112012-03-05 13:15:12 -08001206}
1207
1208/**
Tejun Heo5efd6112012-03-05 13:15:12 -08001209 * blkcg_exit_queue - exit and release blkcg part of request_queue
1210 * @q: request_queue being released
1211 *
Marcos Paulo de Souza7585d502019-01-25 00:01:42 -02001212 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
Tejun Heo5efd6112012-03-05 13:15:12 -08001213 */
1214void blkcg_exit_queue(struct request_queue *q)
1215{
Tejun Heo3c96cb32012-04-13 13:11:34 -07001216 blkg_destroy_all(q);
Tejun Heo5efd6112012-03-05 13:15:12 -08001217 blk_throtl_exit(q);
1218}
1219
Vivek Goyal31e4c282009-12-03 12:59:42 -05001220/*
1221 * We cannot support shared io contexts, as we have no mean to support
1222 * two tasks with the same ioc in two different groups without major rework
1223 * of the main cic data structures. For now we allow a task to change
1224 * its cgroup only if it's the only owner of its ioc.
1225 */
Tejun Heo1f7dd3e52015-12-03 10:18:21 -05001226static int blkcg_can_attach(struct cgroup_taskset *tset)
Vivek Goyal31e4c282009-12-03 12:59:42 -05001227{
Tejun Heobb9d97b2011-12-12 18:12:21 -08001228 struct task_struct *task;
Tejun Heo1f7dd3e52015-12-03 10:18:21 -05001229 struct cgroup_subsys_state *dst_css;
Vivek Goyal31e4c282009-12-03 12:59:42 -05001230 struct io_context *ioc;
1231 int ret = 0;
1232
1233 /* task_lock() is needed to avoid races with exit_io_context() */
Tejun Heo1f7dd3e52015-12-03 10:18:21 -05001234 cgroup_taskset_for_each(task, dst_css, tset) {
Tejun Heobb9d97b2011-12-12 18:12:21 -08001235 task_lock(task);
1236 ioc = task->io_context;
1237 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1238 ret = -EINVAL;
1239 task_unlock(task);
1240 if (ret)
1241 break;
1242 }
Vivek Goyal31e4c282009-12-03 12:59:42 -05001243 return ret;
1244}
1245
Tejun Heo69d7fde2015-08-18 14:55:36 -07001246static void blkcg_bind(struct cgroup_subsys_state *root_css)
1247{
1248 int i;
1249
1250 mutex_lock(&blkcg_pol_mutex);
1251
1252 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1253 struct blkcg_policy *pol = blkcg_policy[i];
1254 struct blkcg *blkcg;
1255
1256 if (!pol || !pol->cpd_bind_fn)
1257 continue;
1258
1259 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1260 if (blkcg->cpd[pol->plid])
1261 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1262 }
1263 mutex_unlock(&blkcg_pol_mutex);
1264}
1265
Josef Bacikd09d8df2018-07-03 11:14:55 -04001266static void blkcg_exit(struct task_struct *tsk)
1267{
1268 if (tsk->throttle_queue)
1269 blk_put_queue(tsk->throttle_queue);
1270 tsk->throttle_queue = NULL;
1271}
1272
Tejun Heoc165b3e2015-08-18 14:55:29 -07001273struct cgroup_subsys io_cgrp_subsys = {
Tejun Heo92fb9742012-11-19 08:13:38 -08001274 .css_alloc = blkcg_css_alloc,
Tejun Heo4308a432019-07-24 10:37:55 -07001275 .css_online = blkcg_css_online,
Tejun Heo92fb9742012-11-19 08:13:38 -08001276 .css_offline = blkcg_css_offline,
1277 .css_free = blkcg_css_free,
Tejun Heo3c798392012-04-16 13:57:25 -07001278 .can_attach = blkcg_can_attach,
Tejun Heof7331642019-11-07 11:18:03 -08001279 .css_rstat_flush = blkcg_rstat_flush,
Tejun Heo69d7fde2015-08-18 14:55:36 -07001280 .bind = blkcg_bind,
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001281 .dfl_cftypes = blkcg_files,
Tejun Heo880f50e22015-08-18 14:55:30 -07001282 .legacy_cftypes = blkcg_legacy_files,
Tejun Heoc165b3e2015-08-18 14:55:29 -07001283 .legacy_name = "blkio",
Josef Bacikd09d8df2018-07-03 11:14:55 -04001284 .exit = blkcg_exit,
Tejun Heo1ced953b2014-07-08 18:02:57 -04001285#ifdef CONFIG_MEMCG
1286 /*
1287 * This ensures that, if available, memcg is automatically enabled
1288 * together on the default hierarchy so that the owner cgroup can
1289 * be retrieved from writeback pages.
1290 */
1291 .depends_on = 1 << memory_cgrp_id,
1292#endif
Tejun Heo676f7c82012-04-01 12:09:55 -07001293};
Tejun Heoc165b3e2015-08-18 14:55:29 -07001294EXPORT_SYMBOL_GPL(io_cgrp_subsys);
Tejun Heo676f7c82012-04-01 12:09:55 -07001295
Tejun Heo8bd435b2012-04-13 13:11:28 -07001296/**
Tejun Heoa2b16932012-04-13 13:11:33 -07001297 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1298 * @q: request_queue of interest
1299 * @pol: blkcg policy to activate
1300 *
1301 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1302 * bypass mode to populate its blkgs with policy_data for @pol.
1303 *
1304 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1305 * from IO path. Update of each blkg is protected by both queue and blkcg
1306 * locks so that holding either lock and testing blkcg_policy_enabled() is
1307 * always enough for dereferencing policy data.
1308 *
1309 * The caller is responsible for synchronizing [de]activations and policy
1310 * [un]registerations. Returns 0 on success, -errno on failure.
1311 */
1312int blkcg_activate_policy(struct request_queue *q,
Tejun Heo3c798392012-04-16 13:57:25 -07001313 const struct blkcg_policy *pol)
Tejun Heoa2b16932012-04-13 13:11:33 -07001314{
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001315 struct blkg_policy_data *pd_prealloc = NULL;
Tejun Heo9d179b82019-10-15 09:03:47 -07001316 struct blkcg_gq *blkg, *pinned_blkg = NULL;
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001317 int ret;
Tejun Heoa2b16932012-04-13 13:11:33 -07001318
1319 if (blkcg_policy_enabled(q, pol))
1320 return 0;
1321
Jens Axboe344e9ff2018-11-15 12:22:51 -07001322 if (queue_is_mq(q))
Jens Axboebd166ef2017-01-17 06:03:22 -07001323 blk_mq_freeze_queue(q);
Tejun Heo9d179b82019-10-15 09:03:47 -07001324retry:
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001325 spin_lock_irq(&q->queue_lock);
Tejun Heoa2b16932012-04-13 13:11:33 -07001326
Tejun Heo9d179b82019-10-15 09:03:47 -07001327 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
Tejun Heo71c81402019-06-13 15:30:40 -07001328 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001329 struct blkg_policy_data *pd;
Tejun Heoa2b16932012-04-13 13:11:33 -07001330
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001331 if (blkg->pd[pol->plid])
1332 continue;
1333
Tejun Heo9d179b82019-10-15 09:03:47 -07001334 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1335 if (blkg == pinned_blkg) {
1336 pd = pd_prealloc;
1337 pd_prealloc = NULL;
1338 } else {
1339 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1340 blkg->blkcg);
1341 }
1342
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001343 if (!pd) {
Tejun Heo9d179b82019-10-15 09:03:47 -07001344 /*
1345 * GFP_NOWAIT failed. Free the existing one and
1346 * prealloc for @blkg w/ GFP_KERNEL.
1347 */
1348 if (pinned_blkg)
1349 blkg_put(pinned_blkg);
1350 blkg_get(blkg);
1351 pinned_blkg = blkg;
1352
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001353 spin_unlock_irq(&q->queue_lock);
Tejun Heo9d179b82019-10-15 09:03:47 -07001354
1355 if (pd_prealloc)
1356 pol->pd_free_fn(pd_prealloc);
1357 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1358 blkg->blkcg);
1359 if (pd_prealloc)
1360 goto retry;
1361 else
1362 goto enomem;
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001363 }
Tejun Heoa2b16932012-04-13 13:11:33 -07001364
1365 blkg->pd[pol->plid] = pd;
1366 pd->blkg = blkg;
Tejun Heob276a872013-01-09 08:05:12 -08001367 pd->plid = pol->plid;
Tejun Heoa2b16932012-04-13 13:11:33 -07001368 }
1369
Tejun Heo9d179b82019-10-15 09:03:47 -07001370 /* all allocated, init in the same order */
1371 if (pol->pd_init_fn)
1372 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1373 pol->pd_init_fn(blkg->pd[pol->plid]);
1374
Tejun Heoa2b16932012-04-13 13:11:33 -07001375 __set_bit(pol->plid, q->blkcg_pols);
1376 ret = 0;
Tejun Heo4c55f4f2015-08-18 14:55:09 -07001377
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001378 spin_unlock_irq(&q->queue_lock);
Tejun Heo9d179b82019-10-15 09:03:47 -07001379out:
Jens Axboe344e9ff2018-11-15 12:22:51 -07001380 if (queue_is_mq(q))
Jens Axboebd166ef2017-01-17 06:03:22 -07001381 blk_mq_unfreeze_queue(q);
Tejun Heo9d179b82019-10-15 09:03:47 -07001382 if (pinned_blkg)
1383 blkg_put(pinned_blkg);
Tejun Heo001bea72015-08-18 14:55:11 -07001384 if (pd_prealloc)
1385 pol->pd_free_fn(pd_prealloc);
Tejun Heoa2b16932012-04-13 13:11:33 -07001386 return ret;
Tejun Heo9d179b82019-10-15 09:03:47 -07001387
1388enomem:
1389 /* alloc failed, nothing's initialized yet, free everything */
1390 spin_lock_irq(&q->queue_lock);
1391 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1392 if (blkg->pd[pol->plid]) {
1393 pol->pd_free_fn(blkg->pd[pol->plid]);
1394 blkg->pd[pol->plid] = NULL;
1395 }
1396 }
1397 spin_unlock_irq(&q->queue_lock);
1398 ret = -ENOMEM;
1399 goto out;
Tejun Heoa2b16932012-04-13 13:11:33 -07001400}
1401EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1402
1403/**
1404 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1405 * @q: request_queue of interest
1406 * @pol: blkcg policy to deactivate
1407 *
1408 * Deactivate @pol on @q. Follows the same synchronization rules as
1409 * blkcg_activate_policy().
1410 */
1411void blkcg_deactivate_policy(struct request_queue *q,
Tejun Heo3c798392012-04-16 13:57:25 -07001412 const struct blkcg_policy *pol)
Tejun Heoa2b16932012-04-13 13:11:33 -07001413{
Tejun Heo3c798392012-04-16 13:57:25 -07001414 struct blkcg_gq *blkg;
Tejun Heoa2b16932012-04-13 13:11:33 -07001415
1416 if (!blkcg_policy_enabled(q, pol))
1417 return;
1418
Jens Axboe344e9ff2018-11-15 12:22:51 -07001419 if (queue_is_mq(q))
Jens Axboebd166ef2017-01-17 06:03:22 -07001420 blk_mq_freeze_queue(q);
Jens Axboebd166ef2017-01-17 06:03:22 -07001421
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001422 spin_lock_irq(&q->queue_lock);
Tejun Heoa2b16932012-04-13 13:11:33 -07001423
1424 __clear_bit(pol->plid, q->blkcg_pols);
1425
1426 list_for_each_entry(blkg, &q->blkg_list, q_node) {
Tejun Heo001bea72015-08-18 14:55:11 -07001427 if (blkg->pd[pol->plid]) {
Dennis Zhou (Facebook)6b065462018-08-31 16:22:42 -04001428 if (pol->pd_offline_fn)
Tejun Heoa9520cd2015-08-18 14:55:14 -07001429 pol->pd_offline_fn(blkg->pd[pol->plid]);
Tejun Heo001bea72015-08-18 14:55:11 -07001430 pol->pd_free_fn(blkg->pd[pol->plid]);
1431 blkg->pd[pol->plid] = NULL;
1432 }
Tejun Heoa2b16932012-04-13 13:11:33 -07001433 }
1434
Christoph Hellwig0d945c12018-11-15 12:17:28 -07001435 spin_unlock_irq(&q->queue_lock);
Jens Axboebd166ef2017-01-17 06:03:22 -07001436
Jens Axboe344e9ff2018-11-15 12:22:51 -07001437 if (queue_is_mq(q))
Jens Axboebd166ef2017-01-17 06:03:22 -07001438 blk_mq_unfreeze_queue(q);
Tejun Heoa2b16932012-04-13 13:11:33 -07001439}
1440EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1441
1442/**
Tejun Heo3c798392012-04-16 13:57:25 -07001443 * blkcg_policy_register - register a blkcg policy
1444 * @pol: blkcg policy to register
Tejun Heo8bd435b2012-04-13 13:11:28 -07001445 *
Tejun Heo3c798392012-04-16 13:57:25 -07001446 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1447 * successful registration. Returns 0 on success and -errno on failure.
Tejun Heo8bd435b2012-04-13 13:11:28 -07001448 */
Jens Axboed5bf0292014-06-22 16:31:56 -06001449int blkcg_policy_register(struct blkcg_policy *pol)
Vivek Goyal3e252062009-12-04 10:36:42 -05001450{
Tejun Heo06b285b2015-07-09 16:39:50 -04001451 struct blkcg *blkcg;
Tejun Heo8bd435b2012-04-13 13:11:28 -07001452 int i, ret;
Tejun Heoe8989fa2012-03-05 13:15:20 -08001453
Tejun Heo838f13b2015-07-09 16:39:47 -04001454 mutex_lock(&blkcg_pol_register_mutex);
Tejun Heobc0d6502012-04-13 13:11:26 -07001455 mutex_lock(&blkcg_pol_mutex);
1456
Tejun Heo8bd435b2012-04-13 13:11:28 -07001457 /* find an empty slot */
1458 ret = -ENOSPC;
1459 for (i = 0; i < BLKCG_MAX_POLS; i++)
Tejun Heo3c798392012-04-16 13:57:25 -07001460 if (!blkcg_policy[i])
Tejun Heo8bd435b2012-04-13 13:11:28 -07001461 break;
Jens Axboe01c5f852018-09-11 10:59:53 -06001462 if (i >= BLKCG_MAX_POLS) {
1463 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
Tejun Heo838f13b2015-07-09 16:39:47 -04001464 goto err_unlock;
Jens Axboe01c5f852018-09-11 10:59:53 -06001465 }
Tejun Heo035d10b2012-03-05 13:15:04 -08001466
weiping zhange8401072017-10-17 23:56:21 +08001467 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1468 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1469 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1470 goto err_unlock;
1471
Tejun Heo06b285b2015-07-09 16:39:50 -04001472 /* register @pol */
Tejun Heo3c798392012-04-16 13:57:25 -07001473 pol->plid = i;
Tejun Heo06b285b2015-07-09 16:39:50 -04001474 blkcg_policy[pol->plid] = pol;
1475
1476 /* allocate and install cpd's */
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001477 if (pol->cpd_alloc_fn) {
Tejun Heo06b285b2015-07-09 16:39:50 -04001478 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1479 struct blkcg_policy_data *cpd;
1480
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001481 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
Bart Van Asschebbb427e2016-09-29 08:33:30 -07001482 if (!cpd)
Tejun Heo06b285b2015-07-09 16:39:50 -04001483 goto err_free_cpds;
Tejun Heo06b285b2015-07-09 16:39:50 -04001484
Tejun Heo81437642015-08-18 14:55:15 -07001485 blkcg->cpd[pol->plid] = cpd;
1486 cpd->blkcg = blkcg;
Tejun Heo06b285b2015-07-09 16:39:50 -04001487 cpd->plid = pol->plid;
Tejun Heo86a5bba2019-08-28 15:05:52 -07001488 if (pol->cpd_init_fn)
1489 pol->cpd_init_fn(cpd);
Tejun Heo06b285b2015-07-09 16:39:50 -04001490 }
1491 }
1492
Tejun Heo838f13b2015-07-09 16:39:47 -04001493 mutex_unlock(&blkcg_pol_mutex);
Tejun Heo8bd435b2012-04-13 13:11:28 -07001494
Tejun Heo8bd435b2012-04-13 13:11:28 -07001495 /* everything is in place, add intf files for the new policy */
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001496 if (pol->dfl_cftypes)
1497 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1498 pol->dfl_cftypes));
Tejun Heo880f50e22015-08-18 14:55:30 -07001499 if (pol->legacy_cftypes)
Tejun Heoc165b3e2015-08-18 14:55:29 -07001500 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
Tejun Heo880f50e22015-08-18 14:55:30 -07001501 pol->legacy_cftypes));
Tejun Heo838f13b2015-07-09 16:39:47 -04001502 mutex_unlock(&blkcg_pol_register_mutex);
1503 return 0;
1504
Tejun Heo06b285b2015-07-09 16:39:50 -04001505err_free_cpds:
weiping zhang58a9edc2017-10-10 22:53:46 +08001506 if (pol->cpd_free_fn) {
Tejun Heo06b285b2015-07-09 16:39:50 -04001507 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001508 if (blkcg->cpd[pol->plid]) {
1509 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1510 blkcg->cpd[pol->plid] = NULL;
1511 }
Tejun Heo06b285b2015-07-09 16:39:50 -04001512 }
1513 }
1514 blkcg_policy[pol->plid] = NULL;
Tejun Heo838f13b2015-07-09 16:39:47 -04001515err_unlock:
Tejun Heobc0d6502012-04-13 13:11:26 -07001516 mutex_unlock(&blkcg_pol_mutex);
Tejun Heo838f13b2015-07-09 16:39:47 -04001517 mutex_unlock(&blkcg_pol_register_mutex);
Tejun Heo8bd435b2012-04-13 13:11:28 -07001518 return ret;
Vivek Goyal3e252062009-12-04 10:36:42 -05001519}
Tejun Heo3c798392012-04-16 13:57:25 -07001520EXPORT_SYMBOL_GPL(blkcg_policy_register);
Vivek Goyal3e252062009-12-04 10:36:42 -05001521
Tejun Heo8bd435b2012-04-13 13:11:28 -07001522/**
Tejun Heo3c798392012-04-16 13:57:25 -07001523 * blkcg_policy_unregister - unregister a blkcg policy
1524 * @pol: blkcg policy to unregister
Tejun Heo8bd435b2012-04-13 13:11:28 -07001525 *
Tejun Heo3c798392012-04-16 13:57:25 -07001526 * Undo blkcg_policy_register(@pol). Might sleep.
Tejun Heo8bd435b2012-04-13 13:11:28 -07001527 */
Tejun Heo3c798392012-04-16 13:57:25 -07001528void blkcg_policy_unregister(struct blkcg_policy *pol)
Vivek Goyal3e252062009-12-04 10:36:42 -05001529{
Tejun Heo06b285b2015-07-09 16:39:50 -04001530 struct blkcg *blkcg;
1531
Tejun Heo838f13b2015-07-09 16:39:47 -04001532 mutex_lock(&blkcg_pol_register_mutex);
Tejun Heobc0d6502012-04-13 13:11:26 -07001533
Tejun Heo3c798392012-04-16 13:57:25 -07001534 if (WARN_ON(blkcg_policy[pol->plid] != pol))
Tejun Heo8bd435b2012-04-13 13:11:28 -07001535 goto out_unlock;
1536
1537 /* kill the intf files first */
Tejun Heo2ee867dc2015-08-18 14:55:34 -07001538 if (pol->dfl_cftypes)
1539 cgroup_rm_cftypes(pol->dfl_cftypes);
Tejun Heo880f50e22015-08-18 14:55:30 -07001540 if (pol->legacy_cftypes)
1541 cgroup_rm_cftypes(pol->legacy_cftypes);
Tejun Heo44ea53d2012-04-01 14:38:43 -07001542
Tejun Heo06b285b2015-07-09 16:39:50 -04001543 /* remove cpds and unregister */
Tejun Heo838f13b2015-07-09 16:39:47 -04001544 mutex_lock(&blkcg_pol_mutex);
Tejun Heo06b285b2015-07-09 16:39:50 -04001545
weiping zhang58a9edc2017-10-10 22:53:46 +08001546 if (pol->cpd_free_fn) {
Tejun Heo06b285b2015-07-09 16:39:50 -04001547 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
Tejun Heoe4a9bde2015-08-18 14:55:16 -07001548 if (blkcg->cpd[pol->plid]) {
1549 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1550 blkcg->cpd[pol->plid] = NULL;
1551 }
Tejun Heo06b285b2015-07-09 16:39:50 -04001552 }
1553 }
Tejun Heo3c798392012-04-16 13:57:25 -07001554 blkcg_policy[pol->plid] = NULL;
Tejun Heo06b285b2015-07-09 16:39:50 -04001555
Tejun Heobc0d6502012-04-13 13:11:26 -07001556 mutex_unlock(&blkcg_pol_mutex);
Tejun Heo838f13b2015-07-09 16:39:47 -04001557out_unlock:
1558 mutex_unlock(&blkcg_pol_register_mutex);
Vivek Goyal3e252062009-12-04 10:36:42 -05001559}
Tejun Heo3c798392012-04-16 13:57:25 -07001560EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
Josef Bacik903d23f2018-07-03 11:14:52 -04001561
Tejun Heod3f77df2019-06-27 13:39:52 -07001562bool __blkcg_punt_bio_submit(struct bio *bio)
1563{
1564 struct blkcg_gq *blkg = bio->bi_blkg;
1565
1566 /* consume the flag first */
1567 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1568
1569 /* never bounce for the root cgroup */
1570 if (!blkg->parent)
1571 return false;
1572
1573 spin_lock_bh(&blkg->async_bio_lock);
1574 bio_list_add(&blkg->async_bios, bio);
1575 spin_unlock_bh(&blkg->async_bio_lock);
1576
1577 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1578 return true;
1579}
1580
Josef Bacikd09d8df2018-07-03 11:14:55 -04001581/*
1582 * Scale the accumulated delay based on how long it has been since we updated
1583 * the delay. We only call this when we are adding delay, in case it's been a
1584 * while since we added delay, and when we are checking to see if we need to
1585 * delay a task, to account for any delays that may have occurred.
1586 */
1587static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1588{
1589 u64 old = atomic64_read(&blkg->delay_start);
1590
Tejun Heo54c52e12020-04-13 12:27:55 -04001591 /* negative use_delay means no scaling, see blkcg_set_delay() */
1592 if (atomic_read(&blkg->use_delay) < 0)
1593 return;
1594
Josef Bacikd09d8df2018-07-03 11:14:55 -04001595 /*
1596 * We only want to scale down every second. The idea here is that we
1597 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1598 * time window. We only want to throttle tasks for recent delay that
1599 * has occurred, in 1 second time windows since that's the maximum
1600 * things can be throttled. We save the current delay window in
1601 * blkg->last_delay so we know what amount is still left to be charged
1602 * to the blkg from this point onward. blkg->last_use keeps track of
1603 * the use_delay counter. The idea is if we're unthrottling the blkg we
1604 * are ok with whatever is happening now, and we can take away more of
1605 * the accumulated delay as we've already throttled enough that
1606 * everybody is happy with their IO latencies.
1607 */
1608 if (time_before64(old + NSEC_PER_SEC, now) &&
1609 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1610 u64 cur = atomic64_read(&blkg->delay_nsec);
1611 u64 sub = min_t(u64, blkg->last_delay, now - old);
1612 int cur_use = atomic_read(&blkg->use_delay);
1613
1614 /*
1615 * We've been unthrottled, subtract a larger chunk of our
1616 * accumulated delay.
1617 */
1618 if (cur_use < blkg->last_use)
1619 sub = max_t(u64, sub, blkg->last_delay >> 1);
1620
1621 /*
1622 * This shouldn't happen, but handle it anyway. Our delay_nsec
1623 * should only ever be growing except here where we subtract out
1624 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1625 * rather not end up with negative numbers.
1626 */
1627 if (unlikely(cur < sub)) {
1628 atomic64_set(&blkg->delay_nsec, 0);
1629 blkg->last_delay = 0;
1630 } else {
1631 atomic64_sub(sub, &blkg->delay_nsec);
1632 blkg->last_delay = cur - sub;
1633 }
1634 blkg->last_use = cur_use;
1635 }
1636}
1637
1638/*
1639 * This is called when we want to actually walk up the hierarchy and check to
1640 * see if we need to throttle, and then actually throttle if there is some
1641 * accumulated delay. This should only be called upon return to user space so
1642 * we're not holding some lock that would induce a priority inversion.
1643 */
1644static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1645{
Josef Bacikfd112c72019-07-09 14:41:29 -07001646 unsigned long pflags;
Tejun Heo5160a5a2020-09-01 14:52:52 -04001647 bool clamp;
Josef Bacikd09d8df2018-07-03 11:14:55 -04001648 u64 now = ktime_to_ns(ktime_get());
1649 u64 exp;
1650 u64 delay_nsec = 0;
1651 int tok;
1652
1653 while (blkg->parent) {
Tejun Heo5160a5a2020-09-01 14:52:52 -04001654 int use_delay = atomic_read(&blkg->use_delay);
1655
1656 if (use_delay) {
1657 u64 this_delay;
1658
Josef Bacikd09d8df2018-07-03 11:14:55 -04001659 blkcg_scale_delay(blkg, now);
Tejun Heo5160a5a2020-09-01 14:52:52 -04001660 this_delay = atomic64_read(&blkg->delay_nsec);
1661 if (this_delay > delay_nsec) {
1662 delay_nsec = this_delay;
1663 clamp = use_delay > 0;
1664 }
Josef Bacikd09d8df2018-07-03 11:14:55 -04001665 }
1666 blkg = blkg->parent;
1667 }
1668
1669 if (!delay_nsec)
1670 return;
1671
1672 /*
1673 * Let's not sleep for all eternity if we've amassed a huge delay.
1674 * Swapping or metadata IO can accumulate 10's of seconds worth of
1675 * delay, and we want userspace to be able to do _something_ so cap the
Tejun Heo5160a5a2020-09-01 14:52:52 -04001676 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1677 * tasks will be delayed for 0.25 second for every syscall. If
1678 * blkcg_set_delay() was used as indicated by negative use_delay, the
1679 * caller is responsible for regulating the range.
Josef Bacikd09d8df2018-07-03 11:14:55 -04001680 */
Tejun Heo5160a5a2020-09-01 14:52:52 -04001681 if (clamp)
1682 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
Josef Bacikd09d8df2018-07-03 11:14:55 -04001683
Josef Bacikfd112c72019-07-09 14:41:29 -07001684 if (use_memdelay)
1685 psi_memstall_enter(&pflags);
Josef Bacikd09d8df2018-07-03 11:14:55 -04001686
1687 exp = ktime_add_ns(now, delay_nsec);
1688 tok = io_schedule_prepare();
1689 do {
1690 __set_current_state(TASK_KILLABLE);
1691 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1692 break;
1693 } while (!fatal_signal_pending(current));
1694 io_schedule_finish(tok);
Josef Bacikfd112c72019-07-09 14:41:29 -07001695
1696 if (use_memdelay)
1697 psi_memstall_leave(&pflags);
Josef Bacikd09d8df2018-07-03 11:14:55 -04001698}
1699
1700/**
1701 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1702 *
1703 * This is only called if we've been marked with set_notify_resume(). Obviously
1704 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1705 * check to see if current->throttle_queue is set and if not this doesn't do
1706 * anything. This should only ever be called by the resume code, it's not meant
1707 * to be called by people willy-nilly as it will actually do the work to
1708 * throttle the task if it is setup for throttling.
1709 */
1710void blkcg_maybe_throttle_current(void)
1711{
1712 struct request_queue *q = current->throttle_queue;
1713 struct cgroup_subsys_state *css;
1714 struct blkcg *blkcg;
1715 struct blkcg_gq *blkg;
1716 bool use_memdelay = current->use_memdelay;
1717
1718 if (!q)
1719 return;
1720
1721 current->throttle_queue = NULL;
1722 current->use_memdelay = false;
1723
1724 rcu_read_lock();
1725 css = kthread_blkcg();
1726 if (css)
1727 blkcg = css_to_blkcg(css);
1728 else
1729 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1730
1731 if (!blkcg)
1732 goto out;
1733 blkg = blkg_lookup(blkcg, q);
1734 if (!blkg)
1735 goto out;
Dennis Zhou7754f662018-12-05 12:10:39 -05001736 if (!blkg_tryget(blkg))
Josef Bacikd09d8df2018-07-03 11:14:55 -04001737 goto out;
1738 rcu_read_unlock();
Josef Bacikd09d8df2018-07-03 11:14:55 -04001739
1740 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1741 blkg_put(blkg);
Josef Bacikcc7ecc252018-07-31 12:39:03 -04001742 blk_put_queue(q);
Josef Bacikd09d8df2018-07-03 11:14:55 -04001743 return;
1744out:
1745 rcu_read_unlock();
1746 blk_put_queue(q);
1747}
Josef Bacikd09d8df2018-07-03 11:14:55 -04001748
1749/**
1750 * blkcg_schedule_throttle - this task needs to check for throttling
Bart Van Assche537d71b2019-03-20 13:18:45 -07001751 * @q: the request queue IO was submitted on
1752 * @use_memdelay: do we charge this to memory delay for PSI
Josef Bacikd09d8df2018-07-03 11:14:55 -04001753 *
1754 * This is called by the IO controller when we know there's delay accumulated
1755 * for the blkg for this task. We do not pass the blkg because there are places
1756 * we call this that may not have that information, the swapping code for
1757 * instance will only have a request_queue at that point. This set's the
1758 * notify_resume for the task to check and see if it requires throttling before
1759 * returning to user space.
1760 *
1761 * We will only schedule once per syscall. You can call this over and over
1762 * again and it will only do the check once upon return to user space, and only
1763 * throttle once. If the task needs to be throttled again it'll need to be
1764 * re-set at the next time we see the task.
1765 */
1766void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1767{
1768 if (unlikely(current->flags & PF_KTHREAD))
1769 return;
1770
Chunguang Xu49d18222021-01-25 13:05:28 +08001771 if (current->throttle_queue != q) {
1772 if (!blk_get_queue(q))
1773 return;
Josef Bacikd09d8df2018-07-03 11:14:55 -04001774
Chunguang Xu49d18222021-01-25 13:05:28 +08001775 if (current->throttle_queue)
1776 blk_put_queue(current->throttle_queue);
1777 current->throttle_queue = q;
1778 }
1779
Josef Bacikd09d8df2018-07-03 11:14:55 -04001780 if (use_memdelay)
1781 current->use_memdelay = use_memdelay;
1782 set_notify_resume(current);
1783}
Josef Bacikd09d8df2018-07-03 11:14:55 -04001784
1785/**
1786 * blkcg_add_delay - add delay to this blkg
Bart Van Assche537d71b2019-03-20 13:18:45 -07001787 * @blkg: blkg of interest
1788 * @now: the current time in nanoseconds
1789 * @delta: how many nanoseconds of delay to add
Josef Bacikd09d8df2018-07-03 11:14:55 -04001790 *
1791 * Charge @delta to the blkg's current delay accumulation. This is used to
1792 * throttle tasks if an IO controller thinks we need more throttling.
1793 */
1794void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1795{
Tejun Heo54c52e12020-04-13 12:27:55 -04001796 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1797 return;
Josef Bacikd09d8df2018-07-03 11:14:55 -04001798 blkcg_scale_delay(blkg, now);
1799 atomic64_add(delta, &blkg->delay_nsec);
1800}
Josef Bacikd09d8df2018-07-03 11:14:55 -04001801
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001802/**
1803 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
Christoph Hellwig13c78632020-06-27 09:31:54 +02001804 * @bio: target bio
1805 * @css: target css
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001806 *
Christoph Hellwig13c78632020-06-27 09:31:54 +02001807 * As the failure mode here is to walk up the blkg tree, this ensure that the
1808 * blkg->parent pointers are always valid. This returns the blkg that it ended
1809 * up taking a reference on or %NULL if no reference was taken.
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001810 */
Christoph Hellwig13c78632020-06-27 09:31:54 +02001811static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1812 struct cgroup_subsys_state *css)
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001813{
Christoph Hellwig13c78632020-06-27 09:31:54 +02001814 struct blkcg_gq *blkg, *ret_blkg = NULL;
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001815
Christoph Hellwig13c78632020-06-27 09:31:54 +02001816 rcu_read_lock();
Christoph Hellwig309dca302021-01-24 11:02:34 +01001817 blkg = blkg_lookup_create(css_to_blkcg(css),
1818 bio->bi_bdev->bd_disk->queue);
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001819 while (blkg) {
1820 if (blkg_tryget(blkg)) {
1821 ret_blkg = blkg;
1822 break;
1823 }
1824 blkg = blkg->parent;
1825 }
Christoph Hellwig13c78632020-06-27 09:31:54 +02001826 rcu_read_unlock();
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001827
1828 return ret_blkg;
1829}
1830
1831/**
1832 * bio_associate_blkg_from_css - associate a bio with a specified css
1833 * @bio: target bio
1834 * @css: target css
1835 *
1836 * Associate @bio with the blkg found by combining the css's blkg and the
1837 * request_queue of the @bio. An association failure is handled by walking up
1838 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1839 * and q->root_blkg. This situation only happens when a cgroup is dying and
1840 * then the remaining bios will spill to the closest alive blkg.
1841 *
1842 * A reference will be taken on the blkg and will be released when @bio is
1843 * freed.
1844 */
1845void bio_associate_blkg_from_css(struct bio *bio,
1846 struct cgroup_subsys_state *css)
1847{
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001848 if (bio->bi_blkg)
1849 blkg_put(bio->bi_blkg);
1850
Christoph Hellwiga5b97522020-06-27 09:31:53 +02001851 if (css && css->parent) {
Christoph Hellwig13c78632020-06-27 09:31:54 +02001852 bio->bi_blkg = blkg_tryget_closest(bio, css);
Christoph Hellwiga5b97522020-06-27 09:31:53 +02001853 } else {
Christoph Hellwig309dca302021-01-24 11:02:34 +01001854 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg);
1855 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg;
Christoph Hellwiga5b97522020-06-27 09:31:53 +02001856 }
Christoph Hellwig28fc5912020-06-27 09:31:51 +02001857}
1858EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1859
1860/**
1861 * bio_associate_blkg - associate a bio with a blkg
1862 * @bio: target bio
1863 *
1864 * Associate @bio with the blkg found from the bio's css and request_queue.
1865 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1866 * already associated, the css is reused and association redone as the
1867 * request_queue may have changed.
1868 */
1869void bio_associate_blkg(struct bio *bio)
1870{
1871 struct cgroup_subsys_state *css;
1872
1873 rcu_read_lock();
1874
1875 if (bio->bi_blkg)
1876 css = &bio_blkcg(bio)->css;
1877 else
1878 css = blkcg_css();
1879
1880 bio_associate_blkg_from_css(bio, css);
1881
1882 rcu_read_unlock();
1883}
1884EXPORT_SYMBOL_GPL(bio_associate_blkg);
1885
1886/**
1887 * bio_clone_blkg_association - clone blkg association from src to dst bio
1888 * @dst: destination bio
1889 * @src: source bio
1890 */
1891void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1892{
1893 if (src->bi_blkg) {
1894 if (dst->bi_blkg)
1895 blkg_put(dst->bi_blkg);
1896 blkg_get(src->bi_blkg);
1897 dst->bi_blkg = src->bi_blkg;
1898 }
1899}
1900EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1901
Christoph Hellwigdb18a532020-06-27 09:31:58 +02001902static int blk_cgroup_io_type(struct bio *bio)
1903{
1904 if (op_is_discard(bio->bi_opf))
1905 return BLKG_IOSTAT_DISCARD;
1906 if (op_is_write(bio->bi_opf))
1907 return BLKG_IOSTAT_WRITE;
1908 return BLKG_IOSTAT_READ;
1909}
1910
1911void blk_cgroup_bio_start(struct bio *bio)
1912{
1913 int rwd = blk_cgroup_io_type(bio), cpu;
1914 struct blkg_iostat_set *bis;
1915
1916 cpu = get_cpu();
1917 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1918 u64_stats_update_begin(&bis->sync);
1919
1920 /*
1921 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1922 * bio and we would have already accounted for the size of the bio.
1923 */
1924 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1925 bio_set_flag(bio, BIO_CGROUP_ACCT);
Colin Ian King0b8cc252020-06-30 16:54:41 +01001926 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
Christoph Hellwigdb18a532020-06-27 09:31:58 +02001927 }
1928 bis->cur.ios[rwd]++;
1929
1930 u64_stats_update_end(&bis->sync);
1931 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1932 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1933 put_cpu();
1934}
1935
Tejun Heod3f77df2019-06-27 13:39:52 -07001936static int __init blkcg_init(void)
1937{
1938 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1939 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1940 WQ_UNBOUND | WQ_SYSFS, 0);
1941 if (!blkcg_punt_bio_wq)
1942 return -ENOMEM;
1943 return 0;
1944}
1945subsys_initcall(blkcg_init);
1946
Josef Bacik903d23f2018-07-03 11:14:52 -04001947module_param(blkcg_debug_stats, bool, 0644);
1948MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");